LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL

INPUT-STATE-OUTPUT SYSTEMS

TIMO REIS1 AND MANUEL SCHALLER1

A BSTRACT. We examine the minimization of a quadratic cost functional composed of the output and
arXiv:2401.11302v1 [math.OC] 20 Jan 2024

the final state of abstract infinite-dimensional evolution equations in view of existence of solutions
and optimality conditions. While the initial value is prescribed, we are minimizing over all inputs
within a specified convex subset of square integrable controls with values in a Hilbert space. The
considered class of infinite-dimensional systems is based on the system node approach introduced by
S TAFFANS [35]. Thus, our developed approach includes optimal control of a wide variety of linear
partial differential equations with boundary control and observation that are not well-posed in the sense
that the output continuously depends on the input and the initial value. We provide an application
of particular optimal control problems arising in energy-optimal control of port-Hamiltonian systems.
Last, we illustrate the our abstract theory by two examples including a non-well-posed heat equation
with Dirichlet boundary control and a wave equation on an L-shaped domain with boundary control of
the stress in normal direction.

Keywords. linear-quadratic optimal control, infinite-dimensional systems, partial differential equa-

tions, boundary control, input constraints
Mathematics subject classications (2020). 49J20, 49N10, 49K27, 93C25

1. T HE OBJECTIVE

In this work, we consider dynamic optimal control of abstract evolution equations. We aim to mini-
mize the cost functional for a specified linear system characterized by the input function u : [0, T ] →
U , the state x(t) ∈ X at time t ∈ [0, T ], and the output y : [0, T ] → Y , where U , X, and Y are
Hilbert spaces. The cost functional to be minimized is given by
1 T
Z
1
∥y(t) − yref (t)∥2Y dt + ∥F x(T ) − zf ∥2Z ,
2 0 2
where yref ∈ L2 ([0, T ]; Y ), zf ∈ Z (Z is another Hilbert space), and the bounded operator F : X →
Z are given. The minimization is performed while enforcing the constraint that the input function
u belongs to a specified closed and convex set Uad . Additionally, the system is initialized with a
predefined initial value x(0) = x0 .
Minimization is performed subject to a linear input-state-output system, typically denoted, in a
simple representation, as ẋ(t) = Ax(t)+Bu(t), y(t) = Cx(t)+Du(t). However, this representation
lacks generality, especially when handling boundary control and observation scenarios. Instead, we
adopt the system node framework as developed by S TAFFANS in [35] and a range of prior journal
publications. That is, we consider systems of the form
  h i 
ẋ(t) A&B x(t)
y(t)
= C&D u(t)
, x(0) = x0 , (1.1)

1
I NSTITUTE OF M ATHEMATICS , T ECHNISCHE U NIVERSIT ÄT I LMENAU , I LMENAU , G ERMANY
1
2 T. REIS AND M. SCHALLER

where A&B : X × U ⊃ dom(A&B) → X, C&D : X × U ⊃ dom(C&D) → Y are linear

operators, where detailed specifications regarding their properties will be outlined in the upcoming
section.
A central aspect is that domain of these operators is not necessarily a Cartesian product of subspaces
of X and U . This particular setup allows for the incorporation of boundary control, among other
considerations. We emphasize that we are not assuming well-posedness, which refers to the existence
of some constant c > 0, such that the solutions of (1.1) (the precise definition of the solution concept
will be provided in the following section) fulfill


∥y∥L2 ([0,T ];Y ) + ∥x(T )∥X ≤ c ∥u∥L2 ([0,T ];U ) + ∥x0 ∥X . (1.2)
This condition establishes existence and continuous dependence of the state and output upon the
initial value and input. While well-posedness is a valuable property in analysis, it poses certain chal-
lenges. Firstly, it can be difficult to verify in specific cases. Secondly, and perhaps more crucially, it
excludes a range of essential cases: For instance, systems like the heat equation with Dirichlet bound-
ary control and Neumann observation, which are prevalent in real-world applications, fall outside the
scope of well-posed systems. Thirdly, though well-posedness significantly simplifies the analysis of
the problem, the authors believe that it is not a natural assumption in optimal control. This is because
the formulation of the optimization problem itself ensures that the optimal control input corresponds
to an output that is square integrable, rather than a distribution. Nevertheless, since our results are
also novel for well-posed system, throughout this work we will briefly offer some observations on
how our presented theory simplifies for well-posed systems.
It is important to note that our presented theory also encompasses a quadratic penalization of the
input. That is, for some c > 0, the minimization of
1 T
Z
1
∥y(t) − yref (t)∥2Y + c∥u(t)∥2Y dt + ∥F x(T ) − zf ∥2Z . (1.3)
2 0 2
subject to (1.1) can be led back to to an optimal control problem with cost functional of type as in
(OCP) by artificially extending the output of (1.1). Namely, this can be performed by considering the
system
   A&B   
ẋ(t) x(t)
= C&D , (1.4)
h i
ye(t) √ u(t)
0 cI

and setting yeref := ( yref

0 ). Further note that, by respectively replacing C&D, yref and F with
√1 C&D, √1 yref and √1 F , the above optimal control problem is equivalent to one with c = 1.
c c c
Before delving into the material, we present a brief overview of existing results regarding linear-
quadratic optimal control of infinite-dimensional systems. A system node approach to optimal con-
trol, as pursued in [27,28], involves minimizing ∥u∥2L2 + ∥y∥2L2 on the positive and negative half-axis.
This was achieved by demonstrating that the value function, which maps the initial value to the op-
timal cost functional value, is quadratic. Based on this finding, a theory was developed that extends
algebraic Riccati equations to system nodes. While penalizing the input is also possible in our setup
(achieved through an artificial extension of the output, explained in detail in Section 3), the problem
addressed in this work differs for two primary reasons. Firstly, we consider a finite time horizon, and
secondly, we allow for input constraints. The latter aspect means that our problem may not neces-
sarily yield a quadratic value function. Consequently, our problem, in general, cannot be addressed
using the approaches outlined in [27, 28]. In this context of Riccati-based approaches, we also men-
tion the earlier works [8, 21] and the textbooks [4, 22, 23] for a thorough treatment of infinite-horizon
and unconstrained infinite-dimensional optimal control problems with well-posed dynamics. Further,
for linear quadratic optimal control of e.g., the particular case of a non-well-posed heat equation with
Dirichlet boundary control, we mention [24, Chapter 9] and [25, Section 9] where the smoothing
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 3

properties of the heat semigroup are used to deduce existence of solutions and optimality conditions.
The main difference to our approach is that we consider a more general setting, as we do not assume,
e.g., analyticity of the semigroup.
Another line of research in optimal control is based on variational theory [15, 25, 39]. This approach
is complementary to semigroup theory and in particular enables variational discretization techniques
such as finite elements. A central tool in these works is the derivation of a control-to-state map in
a suitable functional analytic setting (note that in the well-posed semigroup setting, such a mapping
can be directly obtained trough the variation of constants formula). In general, such a control-to-
state mapping has to be deduced and analyzed in a case-by-case scenario: We refer the reader, e.g.,
to [33] for a compact approach to linear parabolic equations, to [18] for wave equations with control
constraints, to [6] for fluid dynamical applications and to [5] for the treatment of Maxwell equations.
The main difference to our approach is that we present our theory for the general class of system
nodes. Therein, we define an unbounded control-to-output map in an abstract setting that allows for
existence of solutions or optimality conditions. Consequently, for particular applications, the only
remaining task is to verify that the problem under consideration can be formulated as a system node.
To this end, however, we refer to [29, 31], where system node formulations of dissipative heat, wave,
Maxwell’s and Oseen equations were presented.
Thus, the main novelty of this work is the abstract and operator-theoretic formulation of existence
theory and optimality conditions with very minor assumptions on the structure. In this way, we
include various classes of problems with unbounded input or observation, such as, e.g., a heat equation
with Dirichlet boundary control and Neumann observation.
Notation. Let X and Y denote Hilbert spaces, consistently assumed to be complex throughout this
work. The norm in X will be written as ∥ · ∥X or simply ∥ · ∥, if clear from context. The identity
mapping in X is denoted as IX (or just I, if context makes it clear).
The symbol X ∗ stands for the anti-dual of X, that is, it consists of all continuous and conjugate-
linear functionals. Correspondingly, ⟨·, ·⟩X ∗ ,X stands for the corresponding duality product. Further,
note that the Riesz map RX , sending x ∈ X to the functional ⟨x, ·⟩X is a linear isometric isomorphism
from X to X ∗ . If the spaces are clear from context, we may skip the subindices. Further, if not stated
else, a Hilbert space is canonically identified with its anti-dual. Note that, in this case, RX = IX .
The space of bounded linear operators from X to Y is denoted by L(X, Y ). As customary, we
abbreviate L(X) := L(X, X). The domain dom(A) of a potentially unbounded linear operator
A : X ⊃ dom(A) → Y is usually endowed with the graph norm, represented as ∥x∥dom(A) :=

1/2
∥x∥2X + ∥Ax∥2Y .
The adjoint of a densely defined linear operator A : X ⊃ dom(A) → Y is A∗ : Y ⊃ dom(A∗ ) →
X with domain

dom(A∗ ) = {y ∈ Y |∃ z ∈ X s.t. ∀ x ∈ dom(A) : ⟨y, Ax⟩Y = ⟨z, x⟩X } .

The vector z ∈ X in the above set is uniquely determined by y ∈ dom(A∗ ), and we define A∗ y = z.
A self-adjoint operator P : X ⊃ dom(P ) → X is called nonnegative, if ⟨x, P x⟩ ≥ 0 for all
x ∈ dom(P ). The operator square root of such a nonnegative operator, i.e., √ a self-adjoint and
nonnegative (i.e., also self-adjoint) operator whose square is P , is denoted by P .
We adopt the notation presented in the book by A DAMS [1] for Lebesgue and Sobolev spaces.
When referring to function spaces with values in a Hilbert space X, we indicate the additional notation
”; X” following the specification of the domain. For instance, the Lebesgue space of p-integrable X-
valued functions over the domain Ω is denoted as Lp (Ω; X).
4 T. REIS AND M. SCHALLER

For a finite time horizon T > 0, the spaces

2
([0, T ]; X) := v ∈ H 2 ([0, T ]; X) v(0) = dt
d

H0l v(0) = 0 , (1.5)
2 2 d

H0r ([0, T ]; X) := v ∈ H ([0, T ]; X) v(T ) = dt v(T ) = 0 (1.6)
play a crucial role throughout this work. Note that due to boundedness of the operators defined by
evaluation of the function and its derivative at zero, these spaces are closed and thus again are Hilbert
spaces when endowed with the usual norm in H 2 ([0, T ]; X).
Their dual spaces with respect to the pivot space L2 ([0, T ]; X) are denoted by
−2
H0l 2
([0, T ]; X) := H0r ([0, T ]; X)∗ , (1.7)
−2
H0r ([0, T ]; X) := 2
H0l ([0, T ]; X)∗ . (1.8)
d2 −2
The left second derivative dt2 l : L2 ([0, T ]; X) → H0l


([0, T ]; X) is defined via
d 2
d 2
⟨ dt 2 l v, w⟩H −2 ([0,T ];X),H 2 ([0,T ];X) := ⟨v, dt2 w⟩L2 ([0,T ];X)
0r
0l
(1.9)
∀ v ∈ L2 ([0, T ]; X), w ∈ H0r
2
([0, T ]; X).
d2


In a completely analogous manner, we can also contemplate the right second derivative dt2 r
:
−2
L2 ([0, T ]; X) → H0r ([0, T ]; X).
Moreover, we introduce the time reflection operator
RT : L2 ([0, T ]; X) → L2 ([0, T ]; X),
(1.10)
v(·) 7→ v(T − ·),
which is a self-adjoint isometry. Clearly, RT also restricts to isometries from H0r 2 ([0, T ]; X) to
2 2 2
H0l ([0, T ]; X), and also from H0l ([0, T ]; X) to H0r ([0, T ]; X). These are also denoted by RT for
sake of convenience. Moreover, by defining
⟨RT v, w⟩H −2 ([0,T ];X),H 2 := ⟨v, RT w⟩H −2 ([0,T ];X),H 2 ([0,T ];X)
0l 0r ([0,T ];X) 0r 0l
−2 2
∀v ∈ H0r ([0, T ]; X), w ∈ H0r ([0, T ]; X).
−2 −2
we see that RT extends to an isometry from H0r ([0, T ]; X) to H0l ([0, T ]; X). Its inverse is again
2
an extension of the time reflection operator on L ([0, T ]; X), and therefore also denoted by RT .

2. S YSTEM NODES AND SOLUTION OPERATORS

For Hilbert spaces X, U , and Y and linear operators A&B : dom(A&B) ⊂ X × U → X,

C&D : dom(C&D) ⊂ X × U → Y , we introduce the necessary fundamentals for systems of the
form (1.1). Here, in the infinite-dimensional case, the operators A&B and C&D are not assumed to
directly segregate into distinct components that correspond to the state and input and which act on X
and U , in contrast to the finite-dimensional scenario. This is primarily motivated by the application of
boundary control in partial differential equations. The autonomous dynamics (i.e, those with trivial
input u = 0) are determined by the so-called main operator A : dom(A) ⊂ X → X with dom(A) :=
{x ∈ X | ( x0 ) ∈ dom(A&B) } and Ax := A&B ( x0 ) for all x ∈ dom(A).

i node). A system node on the triple (X, U, Y ) of Hilbert spaces is a linear

Definition 2.1h (System
A&B
operator S = C&D with A&B : dom(A&B) ⊂ X ×U → X, C&D : dom(C&D) ⊂ X×U → Y
satisfying the following conditions:
(a) A&B is closed.
(b) C&D ∈ L(dom(A&B), Y ).
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 5

(c) For all u ∈ U , there exists some x ∈ X with ( ux ) ∈ dom(S).

(d) The main operator A generates a strongly continuous semigroup A(·) : R≥0 → L(X) on X.
Next, we define the term solution for the abstract evolution equation (1.1).
 A&B 
Definition 2.2 (Classical/generalized trajectories). Let T > 0, and let S = C&D be a system node
on (X, U, Y ).
A classical trajectory of (1.1) on [0, T ] is a triple
(x, u, y) ∈ C 1 ([0, T ]; X) × C([0, T ]; U ) × C([0, T ]; Y )
which for all t ∈ [0, T ] satisfies (1.1).
A generalized trajectory of (1.1) on [0, T ] is a limit of classical trajectories of (1.1) on [0, T ] in the
topology of C([0, T ]; X) × L2 ([0, T ]; U ) × L2 ([0, T ]; Y ).
Any operator A&B with the properties (a), (c) and (d) in Definition 2.1 can be regarded as a system
node on (X, U, {0}). Consequently, we may further speak of classical and generalized trajectories
(x, u) of
ẋ = A&B ( ux ) . (2.1)
We rephrase a solvability result from [35]. Here, besides demanding twice weak differentiability
with absolutely integrable derivative of the input, it is required that the pair consisting of the initial
state and initial input value lies in dom(A&B). For boundary control systems, this means that the
boundary value at t = 0 is consistent with the corresponding boundary value of the prescribed initial
state.
Proposition 2.3 (Existence of classical trajectories [35, Thm. 4.3.9]). Let S be a system node on
x0

(X, U, Y ), let T > 0, x0 ∈ X and u ∈ W 2,1 ([0, T ]; U ) with u(0) ∈ dom(S). Then there exists a
unique classical trajectory (x, u, y) of (1.1) with x(0) = x0 .
Whereas this result could be used to establish a control-to-state map for smooth controls, we require
to study the analogous map of square integrable controls, as the cost functional under consideration
is merely coercive in the L2 (0, T ; U )-norm. Thus, going towards a concept of a control to state map
in suitable spaces, we will now present more specific results regarding the existence and regularity of
the trajectories.
Initially, it is essential to highlight that the operator A&B can indeed be distinctly separated into
components corresponding to the state and the input, aligning with the conventional framework used
in numerous studies on infinite-dimensional systems, such as, for instance [40]. However, to establish
such a separation, it becomes necessary to define consider extrapolation spaces.
 A&B 
Remark 2.4 (System nodes). Let S = C&D be a system node on (X, U, Y ).
(a) For k ∈ N, the operator A extends to closed and densely defined operator A−k : X−k ⊃
dom(A−k ) = X−k+1 → X−k , where the Hilbert space X−k is the completion of X with respect
to the norm ∥x∥X−k := ∥(αI − A)−k x∥ for some α ∈ R such that αI − A is bijective. The semi-
group A(·) generated by A extends to a semigroup A−k (·) on X−k . The generator of this semi-
group is A−k . Moreover, for Xk := dom(Ak ), A restricts to Ak : Xk ⊃ dom(Ak ) = Xk+1 →
Xk , and the semigroup Ak (·) generated by Ak is the restriction of A(·) to Xk [40, Prop. 2.10.3 &
2.10.4].
(b) There exists an operator B ∈ L(U, X−1 ) such that [A−1 B] ∈ L(X × U, X−1 ) is an extension of
A&B. The domain of A&B (equally: the domain of S) satisfies
dom(A&B) = {( ux ) ∈ X × U |A−1 x + Bu ∈ X } ,
see [35, Def. 4.7.2 & Lem. 4.7.3].
6 T. REIS AND M. SCHALLER

(c) For k ∈ Z, we denote Xd,k as the space constructed as in (a), but now from A∗ . Then [40,
Prop. 2.10.2] yields Xd,k = X−k ∗ , where the latter is the dual of X
−k with respect to the pivot
space X.
In particular, the adjoint of B maps from Xd,1 = X−1 ∗ to U . Then, by using that A∗ generates

the adjoint semigroup A [40, Prop. 2.8.5], we obtain that, for Ad,1 (t) being the semigroup A∗
∗

restricted to Xd,1 (cf. (a)), it holds that for all t ≥ 0, B ∗ Ad,1 (t) is a bounded operator from Xd,1
to U .
In the following we develop a solution concept that is even more general than generalized tra-
jectories in Definition 2.2. To this end, we first notice that a generalized trajectory (x, u) of (2.1)
fulfills Z t
∀ t ∈ [0, T ] : x(t) = A−1 (t)x(0) + A−1 (t − τ )Bu(τ )dτ, (2.2)
0
where the latter has to be interpreted as an integral in the space X−1 with B ∈ L(U, X−1 ) as in
Remark 2.4 (b). The state satisfies
x ∈ C([0, T ]; X−1 ) ∩ W 1,1 ([0, T ]; X−2 ), (2.3)
 
x(t)
as shown in [35, Thm. 3.8.2]. Consequently, however, the output evaluation y(t) = C&D u(t) is
– at a glance – not necessarily well-defined for all t ∈ [0, T ]. However, it is shown in [35, Lem. 4.7.9]
that the second integral of ( ux ) is continuous as a mapping from [0, T ] to dom(A&B). Thus, the
output may – in the distributional sense – be defined as the left second derivative (as defined in
(1.9)) of C&D applied to the second integral of ( ux ). This can be used to show that (x, u, y)
is a generalizedtrajectory
 of (1.1) if, and only if, (x, u) is a generalized trajectory of (2.1) with
R· x(τ ) 2 ([0, T ]; Y ) and
C&D 0 (· − τ ) u(τ ) dτ ∈ H0l
Z ·  
d2 x(τ )


y = dt 2 l C&D (· − τ ) u(τ )
dτ (2.4)
0
cf. (4.7.6) in [35].
Considering now the above formula (2.4) in the distributional sense, allows us to define a solution
−2
concept with outputs in the space H0l ([0, T ]; Y ) defined in (1.7).
 A&B 
Definition 2.5 (Very generalized trajectory). Let T > 0, and let S = C&D be a system node on
(X, U, Y ). Then
−2
(x, u, y) ∈ C([0, T ]; X−1 ) × L2 ([0, T ]; U ) × H0l ([0, T ]; Y )
is a very generalized trajectory of (1.1) on [0, T ], if (2.2) and (2.4) hold.
The findings prior to the above definition imply that (x, u, y) is a generalized trajectory of (1.1) on
[0, T ], if, and only if, it is a very generalized trajectory of (1.1) on [0, T ] with x ∈ C([0, T ]; X) and
y ∈ L2 ([0, T ]; Y ).
Now, we will introduce a series of operators associated with trajectories of (1.1), which are of
essential importance for the addressed optimal control problem. To maintain clarity for both read-
ers and, admittedly, the authors, these operators are systematically presented in a tabular format in
Appendix A. We recommend the reader to have this table at hand while reading the article.
First, we define C ∈ L(dom(A), Y ) by Cx = C&D ( x0 ). The introduction of very generalized
trajectories gives rise to the introduction of the input-to-state map BT , the state-to-output map CT ,
and the input-to-output map DT . Namely for all x0 ∈ X, u ∈ L2 ([0, T ]; U ), there exist unique
−2
y ∈ H0l ([0, T ]; Y ) and x ∈ C([0, T ]; X−1 ) with x(0) = x0 (defined by (2.2) and (2.4)) such that
(x, u, y) is a very generalized trajectory of (1.1) on [0, T ]. Thus, we may define the bounded operators
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 7

BT : L2 ([0, T ]; U ) → X−1 , (2.5)

Z T
u 7→ A−1 (T − τ )Bu(τ )dτ,
0
−2
CT : X−1 → H0l ([0, T ]; Y ), (2.6)
Z ·
d2


x0 7→ dt2 l C (· − τ )A(τ )x0 dτ,
0
2 −2
DT : L ([0, T ]; U ) → H0l ([0, T ]; Y ), (2.7)
Z ·  
d2 Bτ u


u 7→ dt2 l
C&D (· − τ ) u(τ ) dτ
0
We note that, if im BT ⊂ X, then B is called an admissible control operator for A. Further, C is
called an admissible observation operator for A, if im CT ⊂ L2 ([0, T ]; Y ). Clearly, the latter two
properties are fulfilled, if the system is well-posed.
As the optimal control problem introduced in the previous section involves a weighting of the
terminal value x(T ), there arises a need for an operator which maps a pair consisting of an initial
value x0 and input u ∈ L2 ([0, T ]; U ) to F x(T ), where F ∈ L(X, Z) and a Hilbert space Z are
given. However, defining such an operator is not straightforward, especially if the control operator B
is not admissible for A. In fact, we impose an additional condition on F to facilitate a meaningful
definition for at least a certain class of F ∈ L(X, Z). The foundation for the subsequent definition is
as follows, where we use Proposition 2.3 to verify that BT u ∈ X for all u ∈ H0l 2 ([0, T ]; U ).

 A&B 
Proposition 2.6. Let T > 0, let S = C&D be a system node on (X, U, Y ). Let Z be a Hilbert
space, and let F ∈ L(X, Z), such that the set
 Z · 
BF := w ∈ Z B ∗ (· − τ )A(τ )∗ F ∗ w dτ ∈ H0l 2
([0, T ]; U ) (2.8)
0
is dense in Z. Then
( u0 ) ∈ X−1 × L2 ([0, T ]; U ) A−1 x0 + BT u ∈ X
 x
T̆F,T : →Z
(2.9)
( xu0 ) 7→ F Ax0 + F BT u
is closable with respect to u, that is, if x0 ∈ X−1 , and (u1n ) (u2n ) are sequences in L2 ([0, T ]; U )
which are both converging to the same limit, and, moreover
∀ n ∈ N, i ∈ {1, 2} : A−1 x0 + BT uin ∈ X,
∀ i ∈ {1, 2} : T̆F,T ( uxin


0
) → zi ∈ Z,
then z1 = z2 . Further, if x0 ∈ X−1 , such that there exists some û ∈ L2 ([0, T ]; U ), with A−1 x0 +
BT û ∈ X, then the set
u ∈ L2 ([0, T ]; U ) A−1 x0 + BT u ∈ X

(2.10)
is dense in L2 ([0, T ]; U ).

Proof. The claimed density of (2.10) follows, since Ax0 +BT (û+u) ∈ X for all u ∈ H0l 2 ([0, T ]; U ).

Thus, we remain to show closability.

Recall from Remark 2.4 that, for all t ≥ 0, B ∗ Ad,1 (t) is a bounded operator from Xd,1 to U , where
Ad,1 is the restriction of A∗ to Xd,1 . As a consequence, we have for all x ∈ X that the second integral
of t 7→ A(t)∗ x is continuous as an Xd,1 -valued function.
8 T. REIS AND M. SCHALLER

Assume that x0 ∈ X−1 , and (uin ), i = 1, 2 are sequences with the above properties. Then (un ) =
(u1n − u2n ) converges to zero, and we obtain for all w ∈ BF ,
⟨w, z1 − z2 ⟩Z = lim ⟨w, T̆F,T u0n ⟩Z


n→∞
= lim ⟨w, F BT un ⟩Z
n→∞
 Z T 
= lim F ∗ w, A(t)Bun (T − t)dt
n→∞ 0 X
 Z T Z t 
∗ d2


= lim F w, dt2 l
(t − τ )A(τ )Bdτ un (T − t)dt
n→∞ 0 0 X
Z T  Z t 
d2
F ∗ w,


= lim dt2 l
(t − τ )A(τ )Bdτ un (T − t) dt
n→∞ 0 0 X
Z T  Z t 
d2 ∗ ∗


= lim dt2 l
(t − τ )A(τ ) F wdτ, Bun (T − t) dt
n→∞ 0 0 Xd,1 ,X−1
Z T  Z t 
d2
B∗
(t − τ )A(τ )∗ F ∗ wdτ, un (T − t)


= lim dt2 l
dt
n→∞ 0 0 U
 Z · 
d 2
∗ ∗ ∗
= lim dt2 l
B (· − τ )A(τ ) F wdτ, un (T − ·) = 0.
n→∞ 0 L2 ([0,T ];U )
Density of BF in Z yields z1 = z2 , and the statement is proven. □

The above statement gives rise to a mapping arising from a closure of T̆F,T with respect to u. In
the following definition, we clarify this concept.
 A&B 
Definition 2.7 (F -terminal value map). Let T > 0, let S = C&D be a system node on (X, U, Y ).
Let Z be a Hilbert space, and let F ∈ L(X, Z), such that the set BF as in (2.8) is dense in Z. Then
the F -terminal value map
TF,T : X−1 × L2 ([0, T ]; U ) ⊃ dom(TF,T ) → Z,
is the closure of T̆F,T as in (2.9) with respect to u. That is, ( xu0 ) ∈ dom(TF,T ) if, and only if, there
exists sequence (un ) converging to u in L2 ([0, T ]; U ), with additionally, A−1 x0 + BT un ∈ X for all
n ∈ N, and T̆F,T ( uxn0 ) converges to some z ∈ Z. In this case we set

TF,T ( xu0 ) := z.

By Proposition 2.6, for fixed initial value, we may consider the F -terminal value map as a densely
defined mapping from L2 ([0, T ]; U ). A detailed discussion of the central density assumption on BF
will be provided in Remark 2.11.
Remark 2.8 (F -terminal value map).
(i) The F -terminal value map is not necessarily closed as a mapping on X−1 × L2 ([0, T ]; U ). Even
the mapping T̆F,T is in general not closable under the assumptions made in Proposition 2.6. To
achieve this, it can be seen that, for the set BF as in (2.8), a necessary and sufficient condition
for closability of T̆F,T is that,
{w ∈ BF |A(T )∗ F ∗ w ∈ dom(A∗ ) } (2.11)
is dense in Z. Though this is true, if BF is dense in Z and, additionally, by using [10, Chap. II,
Thm. 4.6], if A is an analytic semigroup, a density claim on (2.11) would however exclude a
variety of interesting cases. We note that, as the first argument of the F -terminal value map
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 9

stands for the initial value, it is fixed in our optimal control problem. Hence, there is actually no
need to presume additional properties which guarantee the “full closedness of TF,T ”.
(ii) For x0 ∈ X−1 such that ( xu0 ) for some u ∈ L2 ([0, T ]; U ), it is required that there exists some
û ∈ L2 ([0, T ]; U ) with A−1 x0 + BT û ∈ X. One can think about more general situations where
F x(T ) also makes sense even though some û with the above property does not exist (such as,
for instance, when we have that F extends to a bounded operator from X−1 to Z). One can think
about a more general definition of the operator T̆F,T in Proposition 2.6 by presuming that there
exists a Hilbert space X̃ with dense embeddings X ⊂ X̃ ⊂ X−1 , such that, in the definition of
T̆F,T in (2.9), X is replaced with X̃. To avoid further exaggerating the technical aspects of the
whole approach, we refrain from making this generalization here.
(iii) If B is an admissible control operator for A, then X × L2 ([0, T ]; U ) is contained in the domain
of the F -terminal.
To finalize the definitions necessary for the optimal control problem, summarized in Appendix A,
we need to further introduce F -input map and G-output map.
 A&B 
Definition 2.9 (F -input map and G-output map). Let T > 0, let S = C&D be a system node on
(X, U, Y ). Let Z be a Hilbert space, and let F ∈ L(X, Z), G ∈ L(Z, X).
(a) Let BF in (2.8) be dense in Z. Then the F -input map
IF,T : L2 ([0, T ]; U ) ⊃ dom(IF,T ) → Z × L2 ([0, T ]; Y )
is defined by
dom(IF,T ) = u ∈ L2 ([0, T ]; U ) ( u0 ) ∈ dom(TF,T ) ∧ DT u ∈ L2 ([0, T ]; Y ) ,

 
TF,T ( u0 )
IF,T x = .
DT u
(b) Further, the G-output map
OG,T : Z × L2 ([0, T ]; U ) ⊃ dom(OG,T ) → L2 ([0, T ]; Y )
is the mapping with
dom(OG,T ) = ( uz ) ∈ Z × L2 ([0, T ]; U ) CT Gz + DT u ∈ L2 ([0, T ]; Y ) ,


OG,T ( uz ) = CT Gz + DT u ∈ L2 ([0, T ]; Y ).
Now, we show under suitable assumptions, that the F -input map and the G-output map are closed
and densely defined.
 A&B 
Proposition 2.10. Let T > 0, let S = C&D be a system node on (X, U, Y ). Let Z be a Hilbert
space, and let F ∈ L(X, Z), G ∈ L(Z, X). Then the following holds:
(a) If the set BF as in (2.8) is dense in Z, then the F -input map IF,T is closed and densely defined.
(b) The G-output map OG,T is closed. If, further, the set
CG := z ∈ Z CT Gz ∈ L2 ([0, T ]; Y )

(2.12)
is dense in Z, then OG,T is densely defined.
10 T. REIS AND M. SCHALLER

Proof.
(a) Assume that (un ) ⊂ dom(IF,T ) converges in L2
([0, T ]; U ) to some u, and (IF,T un ) converges
in Z × L2 ([0, T ]; Y ) to ( yz ). Then (TF,T u0n ) converges in Z to z, and closedness of the
F -terminal value map with respect to u yields
TF,T ( u0 ) = z.
−2
By further using that H0l ([0, T ]; Y ) is continuously embedded in L2 ([0, T ]; Y ), we have that
−2
(DT un ) converges to y in H0l ([0, T ]; Y ), and boundedness of DT leads to convergence of y =
DT u. Altogether, we have
IF,T u = ( yz ) ,
which shows that IF,T is closed. Dense definition of IF,T holds, since by Proposition 2.3,
2
H0l ([0, T ]; U ) ⊂ dom(ĬF,T ).
(b) Assume that ( uznn ) is a sequence in dom(O
G,T ) that converges in Z × L2 ([0, T ]; U ) to ( uz ) ∈

Z × L2 ([0, T ]; U ), and, further, OG,T ( uznn ) converges in L2 ([0, T ]; Y ) to y ∈ L2 ([0, T ]; Y ).

Now boundedness of CT , G and DT together with the continuous embedding of L2 ([0, T ]; Y ) in
−2
H0l ([0, T ]; Y ) yields that
CT Gz + DT u = y,
and thus ( uz ) ∈ dom(OG,T ) with dom(OG,T ) ( uz ) = z. This shows that OG,T is a closed
operator.
To show the claim on dense definition, we first observe that, by Proposition 2.3, we have for all
2 ([0, T ]; U ) that ( 0 ) ∈ dom(O
u ∈ H0l u G,T ). Then density of CG immediately yields that OG,T is
densely defined.
□
Remark 2.11.
(i) We briefly recall that the purpose of this article is the treatment of the optimal control problem
as lined out in Section 1. A crucial ingredient in our approach to this problem will by the F ∗ -
output map, where F ∈ L(X, Z) is the operator that occurs in the weighting of the terminal
value. This will be elaborated in the forthcoming section.
(ii) The set BF in (2.8) is dense in Z, if one of the following conditions hold:
(a) B ∗ is an admissible observation operator for A∗ . Equivalently, by [40, Thm. 4.4.3], B is
an admissible control operator for A. In this case, the set specified in (2.8) even coincides
with Z.
(b) The set
{z ∈ Z |F ∗ x ∈ dom(A∗ ) }
is dense in X. This is for instance fulfilled in the case where F is the zero operator, or F is
injective and has closed range.
(iii) Likewise, the set CG in (2.12) is dense in Z, if one of the following conditions hold:
(a) C is an admissible observation operator for A. In this case, the set specified in (2.12) even
coincides with Z.
(b) The set
{z ∈ Z |Gx ∈ dom(A) }
is dense in X. This is for instance fulfilled in the case where G is the zero operator or it is
surjective.
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 11

(iv) If the system is well-posed, we have that BT ∈ L(L2 ([0, T ]; U ); X), CT ∈ L(X, L2 ([0, T ]; Y ))
and DT ∈ L(L2 ([0, T ]; U ); L2 ([0, T ]; Y )). Consequently, for F ∈ L(X, Z), G ∈ L(Z, X), we
have, in this case, that the F -input map and the G-output map are as well bounded, and they
moreover simplify to
 
F BT  
IF,T = , OF,T = CT G DT .
DT
Thus, we observe that the incorporation of non-well-posed systems significantly complicates
matters and demands a certain degree of technical finickiness.
 A&B 
In the following let S = C&D be a system node on (X, U, Y ). We now present a result on the
adjoint of the F -input mapand the
∗ G-output map showing that this can be constructed from the
∗ A&B
adjoint system node S = C&D . It indeed holds that the latter is a system node, if S itself is
a system node [35, Lem. 6.2.14]. We denote
 
∗ [A&B]d
S = d , (2.13)
[C&D]

and we consider the adjoint system

 
[A&B]d
  
ẋd (t) xd (t)
ud (t)
= yd (t)
, xd (0) = xd0 . (2.14)
[C&D]d

It is moreover shown in [35, Lem. 6.2.14] that the main operator of S ∗ is given by A∗ . The adjoint
system node will be employed to show that the adjoint F -input map can be constructed from the F ∗ -
output map associated with the system defined by S ∗ and additional “time flips”, which are expressed
by the time reflection operator as introduced in (1.10) and the subsequent lines.
Before the main result for the adjoints of IF,T and OG,T is presented, we advance an auxiliary
result (which generalizes [35, Lem. 6.2.16]) on an integration-by-parts like identity between the (very)
generalized solutions of a system node S and its adjoint. To this end, we recall from Remark 2.4 (c)
that Xd,1 = X−1 ∗ and that X ∗
d,−1 = X1 .

be a system node on (X, U, Y ), and let S ∗ as in

 A&B 
Lemma 2.12. Let T > 0, let S = C&D
(2.13) be the adjoint system node. Assume that (x, u, y) is a classical trajectory of (1.1) with
u ∈ H0l 2 ([0, T ]; U ) and x(0) = 0, and (x , y , u ) is a very generalized trajectory of (2.14). Then
d d d

⟨x(T ), xd (0)⟩X1 ,Xd,−1 + ⟨y, RT yd ⟩L2 ([0,T ];Y ) = ⟨u, RT ud ⟩H 2 ([0,T ];Y ),H −2 ([0,T ];Y ) . (2.15)
0l 0r

Proof. As in the proof of [35, Lem. 6.2.16], it can be shown that, for any classical trajectory (x, y, u)
of (1.1), and any classical trajectory (xd , yd , ud ) of (2.14), it holds that
d
∀ t ∈ [0, T ] : dt ⟨x(t), xd (T − t)⟩X + ⟨y(t), yd (T − t)⟩Y = ⟨u(t), ud (T − t)⟩U .
Now an integration over [0, T ] yields that classical trajectories fulfill
⟨x(T ), xd (0)⟩X + ⟨y, RT yd ⟩L2 ([0,T ];Y ) = ⟨x(0), xd (T )⟩X + ⟨u, RT ud ⟩L2 ([0,T ];Y ) . (2.16)
Now assume that (x, u, y) is a classical trajectory of (1.1) with u ∈ H0l 2 ([0, T ]; U ) and x(0) = 0,

and let (xd , yd , ud ) be a generalized solution for (2.14). Then, by density of dom(S ∗ ) in X × Y and
density of H 2 ([0, T ]; Y ) in L2 ([0, T ]; Y ), there exist sequences (yd,n ) in H 2 ([0, T ]; Y ), and (xd,0,n )
in Xd,1 , such that
· (yd,n ) converges in L2 ([0, T ]; Y ) to yd ,
· (xd,0,n )
converges in Xd,−1 to xd (0), and
x
· yd,0,n
d,n ∈ dom(S ∗ ) for all n ∈ N.
12 T. REIS AND M. SCHALLER

Let Bd,T , Cd,T , Dd,T be the mappings as in (2.5)-(2.7), but now associated to the system node
S ∗ . Then, by a combination of Proposition 2.3 with (2.2) and (2.4), we obtain that, for n ∈ N,
xd,n : [0, T ] → X and ud,n ∈ L2 ([0, T ]; U ) with
xd,n (t) = Bt ud,n + A(t)xd,0,n , t ∈ [0, T ],
ud,n = Cd,T xd,0,n + Dd,T yd,n ,
(xd,n , yd,n , ud,n ) is a classical trajectory of (2.14). Further, by boundedness of Bd,T , Cd,T , Cd,T and
the fact that A∗ extends to a strongly continuous semigroup on Xd,−1 (see Remark 2.4 (a)), we obtain
−2
that xd,n (T ) converges pointwise in Xd,−1 to xd (T ), and (ud,n ) converges in H0l ([0, T ]; Y ) to ud .
Now invoking that (2.16) holds for classical trajectories, we obtain
⟨x(T ), xd,n (0)⟩X1 ,Xd,−1 + ⟨y, RT yd,n ⟩L2 ([0,T ];Y )
= ⟨x(T ), xd,n (0)⟩X + ⟨y, RT yd,n ⟩L2 ([0,T ];Y )
= ⟨u, RT ud,n ⟩L2 ([0,T ];Y ) . = ⟨u, RT ud,n ⟩H 2 ([0,T ];Y ),H −2 ([0,T ];Y ) .
0l 0r

Now the result follows by taking the limit n → ∞. □

 A&B 
Proposition 2.13 (The adjoint of the F -input map). Let T > 0, and let S = C&D be a system node
on (X, U, Y ). Let Z be a Hilbert space, and let F ∈ L(X, Z), such that BF as specified in (2.8) is
dense in Z. Further, let Od,F ∗ ,T be the F ∗ -output map corresponding to S ∗ . Then
h i
I∗F,T = RT Od,F ∗ ,T I0 R0T . (2.17)

Proof. We show the equivalent statement

h i ∗
I 0
0 RT IF,T = RT Od,F ∗ ,T .
We first observe that, by Lemma 2.12, for all z ∈ Z, yd ∈ L2 ([0, T ]; Y ) and u ∈ H0l2 ([0, T ]; U ), we

have
Dh i E
I 0 z z

0 RT IF,T u, ( yd ) 2
= IF,T u, RT yd Z×L2 ([0,T ];Y )
Z×L ([0,T ];Y )
D  E
F BT u z

= DT u , R y
T d
Z×L2 ([0,T ];Y )
= ⟨F BT u, z⟩Z + ⟨DT u, RT yd ⟩L2 ([0,T ];Y )
= ⟨BT u, F ∗ z⟩X + ⟨DT u, RT yd ⟩L2 ([0,T ];Y )
= ⟨BT u, F ∗ z⟩X1 ,Xd,−1 + ⟨DT u, RT yd ⟩L2 ([0,T ];Y )
(2.15)
= u, RT Cd,T F ∗ z + Dd,T yd H 2 ([0,T ];U ),H −2 ([0,T ];U ) .


0l 0r
(2.18)
If ( yzd ) ∈ dom(RT Od,F ∗ ,T ) = dom(Od,F ∗ ,T ), then
RT Cd,T F ∗ z + Dd,T yd ∈ L2 ([0, T ]; U ),

and the last expression in the chain of equalities (2.18) can be considered as inner product in L2 ([0, T ]; U ).
In this case, we further have
RT Cd,T F ∗ z + Dd,T yd = RT Od,F ∗ ,T ( yzd ) .

Since (2.18) holds for all u in the 2 2

 H0l ([0, T ]; U ) of L ([0, T ]; U ), we have shown that
h densei subspace ∗
I 0
RT Od,F ∗ ,T is a restriction of 0 RT IF,T . Hence, to complete the proof, we have to prove that
h i ∗ 
I 0
dom 0 RT IF,T ⊂ dom (RT Od,F ∗ ,T ) .
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 13
h i ∗ 
I 0
Assume that ( yzd ) ∈ dom 0 RT I F,T
2 ([0, T ]; U )
. Then, by (2.18), we have for all for all u ∈ H0l
that
D h i ∗ E Dh i E
u, I0 R0T IF,T ( yzd ) 2 = I 0
0 RT IF,T u, ( z
yd )
L ([0,T ];U ) L2 ([0,T ];U )
(2.18)
= u, RT Cd,T F ∗ z + Dd,T yd H 2 ([0,T ];U ),H −2 ([0,T ];U ) .


0l 0r

This yields that RT Cd,T F ∗ z + Dd,T yd ∈ L2 ([0, T ]; U ), and thus

( yzd ) ∈ dom (RT Od,F ∗ ,T ) ,

which completes the proof. □

3. T HE OPTIMAL CONTROL PROBLEM

Having developed the operator-theoretic foundation, we are now prepared to analyze the optimal
control problem
1 T
Z
1
minimize ∥y(t) − yref (t)∥2Y dt + ∥F x(T ) − zf ∥2Z
2 0 2 (OCP)
  h i 
ẋ(t) A&B x(t)
subject to y(t)
= C&D u(t)
, x(0) = x0 , u ∈ Uad .

Hereby, Uad ⊂ L2 ([0, T ]; U ) is the set of admissible inputs which, as will be illustrated later, allows
to incorporate common control constraints of various types.
We first collect all the central assumptions on the problem (OCP).
Assumptions 3.1.
 A&B 
(a) S = C&D is a system node on the triple (X, U, Y ) of complex Hilbert spaces,
(b) F ∈ L(X, Z) for some Hilbert space Z, and, for the semigroup A : R≥0 → L(X) generated
by A, the set BF as in (2.8) is dense in Z.
(c) x0 ∈ X−1 , zf ∈ Z, and yref ∈ L2 ([0, T ]; Y ).
(d) Uad is a closed and convex subset of L2 ([0, T ]; U ).
(e) At least one of the following two conditions hold:
(i) Uad is bounded.
(ii) The cost functional is coercive. That is, there exists some c > 0, such that all generalized
solutions (x, u, y) with x(0) = 0 satisfy
∥y∥2L2 ([0,T ];U ) + ∥F x(T )∥Z ≥ c ∥u∥2L2 ([0,T ];U ) .
(f) For the mappings introduced in (2.6), (2.7), Definition 2.7 and Definition 2.9 (see also Table 1),
there exists an admissible input û ∈ Uad with
( xû0 ) ∈ dom(TF,T ) and CT x0 + DT û ∈ L2 ([0, T ]; Y ).

In the following, we discuss our assumptions.

Remark 3.2 (Our assumptions).
(a) We indeed permit initial values x0 in the extrapolation space X−1 . Of course, this includes
initialization with x0 ∈ X.
14 T. REIS AND M. SCHALLER

(b) A possible choice for set of admissible inputs is, for example, by means of (a.e.) pointwise control
constraints, i.e., for some closed and convex set Uad ⊂ U ,
Uad = u ∈ L2 ([0, T ]; U ) u(t) ∈ Uad for almost all t ∈ [0, T ] .


This for instance allows to incorporate box constraints on the input. However, in (OCP), also
time-varying constraints are possible, such as u(t) ∈ Uad (t) for almost all t ∈ [0, T ], where
Uad (t) is closed and convex for all t ∈ [0, T ].
(c) Condition (b) in Assumptions 3.1 guarantees, in view of Proposition 2.10, that the terminal cost
1
∥F x(T ) − zf ∥2Z (3.1)
2
is, in the weak sense, well-defined by ∥TF,T ( xu0 )−zf ∥2Z , if ( xû0 ) ∈ dom(TF,T ) and ∞ otherwise.
(d) Condition (e) in Assumptions 3.1 guarantees that any sequence of admissible inputs for which
the cost functional tends to the infimal value, is bounded in L2 ([0, T ]; U ) by means of the usual
argumentation via closed and bounded sublevel sets. As is standard in optimal control theory, this
gives rise to the existence of a weakly convergent subsequence, whose limit will be shown to be
the optimal control.
Note that coercivity is, for instance, fulfilled, if the cost functional is of type (1.3) for some c > 0.
(e) Condition (f) in Assumptions 3.1 is a crucial one, since it essentially expresses that there exists at
least one control with finite cost.
A sufficient criterion for this is the existence of some generalized trajectory (x̂, û, ŷ) with x̂(0) =
x0 and û ∈ Uad . This is, for instance, fulfilled if Uad is nonempty, yref ∈ L2 ([0, T ]; Y ), x0 , xf ∈
X, and the system is well-posed in the sense that there exists some c > 0, such that all classical
(and thus also generalized) trajectories fulfill (1.2).
Now we define what we mean by an optimal control.
Definition 3.3 (Optimal control). Consider an optimal control problem (OCP) with Assumptions 3.1.
Then, for the operators introduced in Section 2 (see also the table in Appendix A), the cost of the
input u ∈ Uad is given by
 ! 2
x0
 1 TF,T ( u ) − z f ( x0 ) ∈ dom(TF,T ) ∧
: u


CT x0 + DT û ∈ L2 ([0, T ]; Y ),

J (u) = 2 CT x0 + DT û − yref
Z×L2 ([0,T ];Y )
 ( x0 ) ∈
/ dom(TF,T ) ∨
: u

∞

/ L2 ([0, T ]; Y ).
CT x0 + DT û ∈
We call uopt ∈ Uad an optimal control for (OCP), if
J (uopt ) = inf J (u).
u∈Uad

Remark 3.4.
(i) If x0 ∈ X, then the cost functional can be rewritten to
 ! 2
0)−z
 1 F A(T )x 0 + T F,T ( u f ( u0 ) ∈ dom(TF,T ) ∧
:


OI,T ( xu0 ) − yref ( xu0 ) ∈ dom(OI,T ),

J (u) = 2 Z×L2 ([0,T ];Y )
 (0) ∈ / dom(TF,T ) ∨
: ux0

∞

(u)∈ / dom(OI,T ).
(ii) Assume that u, δu ∈ L2 ([0, T ]; U ), such that both J (u) < ∞ and J (u + δu) < ∞. Then the
definition of the F -input map yields
x0

TF,T ( xu0 ) − zf
     0


TF,T u+δu − zf TF,T δu
− = = IF,T δu.
CT x0 + DT u − yref CT x0 + DT (u + δu) − yref DT δu
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 15

In particular, for u ∈ L2 ([0, T ]; U ) with J (u) < ∞, we have

J (u + δu) < ∞ ⇔ δu ∈ dom(IF,T ).
By an expansion of J (u + δu), we obtain that
J (u + δu) − J (u)
TF,T ( xu0 ) − zf
  
1
= Re , IF,T δu + ∥IF,T δu∥2X×L2 ([0,T ];Y ) .
CT x0 + DT u − yref Z×L2 ([0,T ];Y ) 2
This will be later on used for our optimality condition based on the derivative of the cost func-
tional.

Now we collect some important properties of the cost functional.

Proposition 3.5. Consider an optimal control problem (OCP) with Assumptions 3.1, then the cost
functional
J : Uad → R≥0 ∪ {∞}
as specified in Definition 3.3 has the following properties:
(a) J is proper. That is, there exists some u ∈ Uad with J (u) < ∞.
(b) J is convex. That is, for all u1 , u2 ∈ Uad , λ ∈ [0, 1],
J (λu1 + (1 − λ)u2 ) ≤ λJ (u1 ) + (1 − λ)J (u2 ). (3.2)
(c) J is lower semicontinuous. That is, for all a ∈ R≥0 , the sublevel set
{u ∈ Uad |J (u) ≤ a } (3.3)
is closed subset of L2 ([0, T ]; U ).

Proof. Let û ∈ L2 ([0, T ]; U ) with J (û) < ∞ (which exists by Assumptions 3.1 (f)). Our proof is
mainly based on the identity
TF,T ( xu0 ) − zf TF,T ( xû0 ) − zf
   
= + IF,T (u − û), (3.4)
CT x0 + DT u − yref CT x0 + DT û − yref
which holds by Remark 3.4.
(a) This follows from J (û) < ∞.
(b) Let u1 , u2 ∈ Uad and λ ∈ [0, 1]. If at least one of J (u1 ), J (u2 ) is infinite, the inequality (3.2) is
trivially fulfilled. Now assume that J (u1 ), J (u2 ) ∈ R≥0 , and let λ ∈ [0, 1]. Then, by linearity
of the operators in Table 1,

TF,T ( ux01 ) − zf TF,T ( ux02 ) − zf

   
λ + (1 − λ)
CT x0 + DT u1 − yref CT x0 + DT u2 − yref
 x0

TF,T λu1 +(1−λ)u − zf
= 2
.
CT x0 + DT (λu1 + (1 − λ)u2 ) − yref
Now using convexity of the mapping ∥ · ∥2X×L2 ([0,T ];U ) (which is a simple consequence of the
Cauchy-Schwarz inequality), we can conclude that J is convex.
(c) Let a ∈ R≥0 , and let (un ) be a sequence in Uad with J (un ) < a for all n ∈ N, which converges
in L2 ([0, T ]; U ) to u ∈ L2 ([0, T ]; U ). Closedness of Uad leads to u ∈ Uad . Then it follows from
(3.4) that


IF,T (un − û) (3.5)
16 T. REIS AND M. SCHALLER

is a bounded sequence in Z × L2 ([0, T ]; Y ). Hence, by [3, Thm. 8.10], it has a weakly convergent
subsequence. It is therefore no loss of generality to assume that the sequence (3.5) itself is weakly
convergent. As a consequence, there exist z ∈ Z, y ∈ L2 ([0, T ]; Y ), such that weak convergence
!  !
û un − û
 x 
+
TF,T ( û0 )−zf
IF,T (un − û)
CT x0 +DT u−yref
!!
un

u

=
 x 
⇀ (3.6)
TF,T ( û0 )−zf
+ IF,T (un − û) ( yz )
CT x0 +DT u−yref

holds in L2 ([0, T ]; U ) × Z × L2 ([0, T ]; Y). The latter sequence evolves in the affine-linear space
!
û
 x 
+ G(IF,T ) (3.7)
TF,T ( û0 )−zf
CT x0 +DT u−yref

where G(IF,T ) stands for the graph of IF,T . Since G(IF,T ) is a closed space by closedness of the
operator IF,T (see Proposition 2.10) the affine-linear space (3.7) is closed and convex, whence it
is also weakly closed by the separation theorem [3, Thm. 8.12]. This yields that
  
z x0 
= TF,T ( û )−zf + IF,T (un − û).
y CT x0 +DT u−yref

Another use of the separation theorem now gives, by invoking (3.4), and convexity of the a-ball
in Z × L2 ([0, T ]; Y ) centered at zero,

J (u) = ∥( yz )∥2Z×L2 ([0,T ];Y )

 x  2
TF,T ( û0 )−zf
≤ lim inf + IF,T (un − û) = lim inf J (un ) ≤ a.
n→∞ CT x0 +DT u−yref Z×L2 ([0,T ];Y ) n→∞

As a consequence, u belongs to the sublevel set (3.3), which shows the desired result.
□

Next we show existence of optimal controls.

Theorem 3.6 (Existence of optimal controls). Consider an optimal control problem (OCP) with As-
sumptions 3.1. Then there exists an optimal control.

Proof. Here we follow the traditional argumentation for minimization of convex functions, such as,
for instance, presented in [9, Chap. II]: Let (un ) be a sequence in Uad , such that
lim J (un ) = Vopt := inf J (u).
n→∞ u∈Uad

Since it is assumed that Uad is bounded or J is coercive, we can conclude from each of these two
cases that (un ) is a bounded sequence in L2 ([0, T ]; U ). Consequently, (un ) has a weak limit uopt ∈
L2 ([0, T ]; U ). Now we show that uopt is indeed an optimal control. We first note that, by using
that Uad is closed and convex, we can conclude from the separation theorem that it is weakly closed,
i.e., uopt ∈ Uad . Convexity of the cost functional (as shown in Proposition 3.5 (b)) implies that each
sublevel set is convex. Combining this with lower semicontinuity of cost functional (which has been
proven in Proposition 3.5 (c)), we can again use the separation theorem to see that each sublevel set
is weakly closed. Consequently, we have
∀ε > 0 : uopt ∈ {u ∈ Uad |J (u) ≤ Vopt + ε } ,
and thus J (uopt ) = Vopt . In other words, uopt is an optimal control. □
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 17

The next result provides a characterization of optimality for controls.

Theorem 3.7. Consider an optimal control problem (OCP) with Assumptions 3.1. Then uopt ∈ Uad
is an optimal control, if, and only if,
∀ u ∈ Uad with u − uopt ∈ dom(IF,T ) :
TF,T ( uxopt
 0
 
) − zf
Re , IF,T (u − uopt ) ≥ 0. (3.8)
CT x0 + DT uopt − yref X×L2 ([0,T ];Y )

Proof. To prove the implication “⇒”, let uopt ∈ dom(IF,T ) ∩ Uad be an optimal control. Further, let
u ∈ Uad with δu := u − uopt ∈ dom(IF,T ), and λ ∈ [0, 1]. Convexity of Uad and dom(IF,T ) yields
uopt + λδu ∈ dom(IF,T ) ∩ Uad ,
and, by invoking Remark 3.4, and optimality of uopt , we have
0 ≤ λ1 J (uopt + λδu) − J (uopt )

TF,T ( uxopt
  
0
) − zf λ
= Re , IF,T δu + ∥IZ,T δu∥2Z×L2 ([0,T ];Y )
CT x0 + DT uopt − yref Z×L2 ([0,T ];Y ) 2
 x0  
TF,T ( uopt ) − zf
→ Re , IF,T δu .
λ↘0 CT x0 + DT uopt − yref Z×L2 ([0,T ];Y )
Now we show the converse implication: Assume that uopt ∈ Uad fulfills (3.8). Let u ∈ Uad with
J (u) < ∞. Now using Remark 3.4, we obtain for δu = uopt − u that δ ∈ dom(IF,T ) with
J (u) − J (uopt )
TF,T ( uxopt
  
0
) − zf 1
= Re , IF,T δu + ∥IF,T δu∥2Z×L2 ([0,T ];Y ) ≥ 0.
CT x0 + DT uopt − yref Z×L2 ([0,T ];Y ) 2
This shows that J (u) ≥ J (uopt ), i.e., the control uopt is optimal. Hence, the claim is proven. □

Next we present a sufficient criterion for uniqueness of optimal controls.

Theorem 3.8 (Uniqueness of optimal controls). Consider an optimal control problem (OCP) satisfy-
ing Assumptions 3.1 and, additionally, let the F -input map IF,T be injective. Then there exists exactly
one optimal control.

Proof. Existence of at least one optimal control has already been proven in Theorem 3.6. Now, under
the assumption that the F -input map is injective, and assume that uopt,1 , uopt,2 ∈ Uad are optimal
controls. Then, by Remark 3.4, we have uopt,1 − uopt,2 ∈ dom(IF,T ) and
0=J (uopt,1 ) − J (uopt,2 )
 x0  
TF,T ( uopt,2 ) − zf
= Re , IF,T (uopt,1 − uopt,2 )
CT x0 + DT uopt,2 − yref X×L2 ([0,T ];Y )
1
+ ∥IF,T (uopt,1 − uopt,2 )∥2Z×L2 ([0,T ];Y ) .
2
The first summand on the right hand side is nonnegative by Theorem 3.7, which means that both
summands vanish. This gives
IF,T (uopt,1 − uopt,2 ) = 0,
and injectivity of IF,T now leads to uopt,1 = uopt,2 , and the result is shown.
□
18 T. REIS AND M. SCHALLER

2 ([0, T ]; U ), then I F z(T ) ∈ Z × L2 ([0, T ]; Y ), where

If u − uopt ∈ H0l F,T (u − uopt ) = w
  h i 
ż(t) A&B z(t)
w(t)
= C&D u(t)−uopt (t) , z(0) = 0.
This will be used in the following result, which is a direct consequence of combining the Theorem 3.7
with Proposition 2.13. The proof is therefore omitted.
Corollary 3.9. Consider an optimal control problem (OCP) with Assumptions 3.1 and let uopt ∈ Uad
be an optimal control (which exists by Theorem 3.6). Then
2
∀ u ∈ Uad with u − uopt ∈ H0l ([0, T ]; U ) : Re ⟨u − uopt , ud ⟩H 2 ([0,T ];U ),H −2 ([0,T ];U ) ≥ 0,
0l 0r

where RT ud is an output of the dual system node in the very generalized sense, i.e.,
F ∗ TF,T ( uxopt
 0

) − zf
ud = RT [Cd,T , Dd,T ] . (3.9)
RT (CT x0 + DT uopt − yref )

By defining the state and output of the system driven by the optimal control
xopt (t) = A−1 (t)x0 + Bt uopt , t ∈ [0, T ],
(3.10)
yopt = CT x0 + DT uopt ,
and the adjoint state
µ(t) = AT −t (t)∗ F ∗ Bx̄,F,T uopt + Bd,T −t (yopt − yref ), t ∈ [0, T ], (3.11)
we obtain that ud in (3.9) solves, in a certain sense, the boundary value problem
  h i 
ẋopt (t) A&B xopt (t)
yopt (t)
= C&D uopt (t)
, xopt (0) = x0 ,
  h i   (3.12)
µ̇(t) −[A&B]d µ(t)
ud (t)
= [C&D]d yopt (t)−yref (t) , µ(T ) = F ∗ (F xopt (T ) − zf ).

Let us now treating the case where, loosely speaking, the constraints imposed by the set Uad of
admissible controls are not active anywhere for the optimal control uopt . First observe that, if uopt ±
δu ∈ Uad for some δu ∈ dom(IF,T ), we obtain from Theorem 3.7 that
TF,T ( uxopt
 0
 
) − zf
Re , IF,T δu = 0. (3.13)
CT x0 + DT uopt − yref Z×L2 ([0,T ];Y )
This will be used in the following, where we give a criterion for the variational inequality in Theo-
rem 3.7 becoming an equality.
Proposition 3.10. Consider an optimal control problem (OCP) with Assumptions 3.1, and let uopt ∈
Uad be an optimal control (which exists by Theorem 3.6). Assume that the the closure of the set
T := {u − uopt | u ∈ Uad ∩ dom(IF,T ) : uopt − u, u − uopt ∈ Uad } ⊂ dom(IF,T ) (3.14)
has a nonempty interior in L2 ([0, T ]; U ), i.e.,
int T ̸= ∅.
Then
TF,T ( uxopt
 0

) − zf
Od,F ∗ ,T
= 0.
RT CT x0 + DT uopt − yref

Proof. Assume that, for ε > 0, δu ∈ L2 ([0, T ]; U ), the ε-ball Uε (δu) centered at δu is contained in
int T . Since,by definition, T = −T , convexity of (Uad − uopt ) ∩ dom(I√N ,T ), yields
Uε (0) ⊂ int T .
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 19

Consequently, the space spanC T is dense in L2 ([0, T ]; U ).

The argumentation prior to this proposition that (3.13) holds for all δu ∈ T . Consequently, (3.13)
holds for all δu in the dense subspace spanC T . Hence, by multiplying with the imaginary unit, we
further obtain that, for all δu ∈ spanC T ,
TF,T ( uxopt
 0
 
) − zf
Im , IF,T δu
CT x0 + DT uopt − yref Z×L2 ([0,T ];Y )
TF,T ( uxopt
 0
 
) − zf
= ı Re , IF,T (ıδu) = 0.
CT x0 + DT uopt − yref Z×L2 ([0,T ];Y )
Hence, for all δu ∈ spanC T ,
TF,T ( uxopt
 0
 
) − zf
, IF,T δu = 0.
CT x0 + DT uopt − yref Z×L2 ([0,T ];Y )
This shows that
TF,T ( uxopt
 0

) − zf
∈ dom(I∗F,T ).
CT x0 + DT uopt − yref
with
TF,T ( uxopt
 0

) − zf
I∗F,T = 0.
CT x0 + DT uopt − yref
Then the result follows from the representation of I∗F,T in Proposition 2.13. □

Under the assumptions Proposition 3.10, we have that ud from Corollary 3.9 vanishes. That is, by
considering the state xopt and output yopt of the system driven by the optimal control (see (3.10)) and
the adjoint state µ as in (3.11), we are led to a boundary value problem
  h i 
ẋopt (t) A&B xopt (t)
yopt (t)
= C&D uopt (t)
, xopt (0) = x0 ,
  h i   (3.15)
µ̇(t) −[A&B]d µ(t) ∗
0
= [C&D] d yopt (t)−yref (t) , µ(T ) = F (F xopt (T ) − z f ).

Remark 3.11 (Regular systems and optimality Hamiltonians). By using  X, U and Y are Hilbert
that
 A&B
spaces, it has been shown in [35, Thm. 5.1.12] that system nodes C&D are “compatible” in the
sense that, loosely speaking, C&D splits into two parts - one corresponding to the state and one
corresponding to the input. More precisely, for
V := {x ∈ X |∃ u ∈ U s.t. ( ux ) ∈ dom(A&B) } (3.16)
which is a Hilbert space endowed with the norm
∥x∥V := inf ∥ ( ux ) ∥dom(A&B) ∃ u ∈ U s.t. ( ux ) ∈ dom(A&B) ,


we have that there exists some C̃ ∈ L(V, Y ), D ∈ L(U, Y ), such that C&D ( ux ) = C̃x + Du for
all ( ux ) ∈ dom(A&B). Unfortunately, C̃ is neither uniquely determined by C&D (for instance,
controlled boundary values can be put into D and in C̃), nor do we have
  h
[A&B]d
i
A∗ C̃ ∗
d = ∗ ∗
, (3.17)
[C&D] B D

in general. However, in case where the system is regular (that is, its transfer function has a limit
on the positive real axis, see [35, 36]), then C̃ and D can be chosen in a way that (3.17) holds,
see [36, Thm. 3.5]. In this case we obtain from (3.15) that
0 = B ∗ µ(t) + D∗ (Cxopt (t) + Duopt (t)) − Dyref (t)
20 T. REIS AND M. SCHALLER

= D∗ Cx(t) + B ∗ µ(t) + D∗ Duopt (t) − Dyref (t).

If, additionally, D has trivial nullspace and closed range, then D∗ D has a bounded inverse, and thus
we can solve the above relation for uopt (t), which gives
uopt (t) = −(D∗ D)−1 D∗ C̃xopt (t) − (D∗ D)−1 B ∗ µ(t) + (D∗ D)−1 Dyref (t).
Now we can eliminate the control and output in (3.15), to obtain, at least formally, the abstract
Hamiltonian system

A − B(D∗ D)−1 D∗ C̃ −B(D∗ D)−1 B ∗

    
ẋ(t) xopt (t)
=
µ̇(t) −C̃ ∗ D∗ (D∗ D)−1 DC̃ −A∗ + C̃ ∗ D(D∗ D)−1 B ∗ µ(t)
∗ −1
 
B(D D)
+ y (t), xopt (0) = x0 , µ(T ) = F ∗ (F xopt (T ) − zf ).
C̃ (I − D(D∗ D)−1 D∗ ) ref
∗

Operators of the aforementioned type are analyzed in [38] through the lens of spectral theory, with ap-
plications to algebraic Riccati equations that arise in linear-quadratic optimal control over an infinite
time horizon.
Remark 3.12 (Real spaces and operators). We assume that all spaces are complex. Real problems
can be addressed by complexifying the operators and spaces involved. Importantly, coercivity of the
cost functional is preserved under complexification.

4. T HE BENEFIT OF INPUT PENALIZATION

Here, we examine optimal control problems with an added quadratic control penalization and
elucidate the supplementary advantages. That is, we consider the problem
1 T
Z
1
minimize ∥y1 (t) − yref (t)∥2Y + ∥u(t)∥2U dt + ∥F x(T ) − zf ∥2Z
2 0 2
  h i 
ẋ(t) A&B x(t)
subject to y(t)
= C&D u(t)
, x(0) = x0 , u ∈ Uad .
As outlined in (1.4), we may equivalently reformulate this problem as
1 T
Z
1
minimize ∥yext (t) − yref,ext (t)∥2Yext dt + ∥F x(T ) − zf ∥2Z
2 0 2
  h A&B i  
ẋ(t) x(t)
subject to = C&D , x(0) = x0 , u ∈ Uad .
h i
yext (t) 0 I u(t)

with Yext = Y × U and yref,ext = ( yref

0 ). The coercivity condition of Assumption 3.1(e) is trivially
fulfilled by
∥yext ∥2L2 ([0,T ];Yext ) = ∥y∥2L2 ([0,T ];Y ) + ∥u∥2L2 ([0,T ];U ) ≥ ∥u∥2L2 ([0,T ];U ) .
As a consequence, we do not have to impose the
h assumption of boundedness of Uad in L2 ([0, T ]; U ).
h A&B i
i
Moreover, the adjoint system node of Sext = C&D on (X, U, Y × U ) reads
0 I
h i
∗ [A&B]d 0
Sext = [C&D]d I , (4.1)

 node, now on (X, Y × U, U ). The corresponding maps (2.5), (2.6) of the sys-
and it is again a system
A&B will now be denoted as in the previous sections, whereas those corresponding
tem node S = C&D
to Sext will be provided with the subscript ∗ext . We clearly have
BT,ext = BT , TF,T,ext = TF,T .
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 21

The state-to-output and input-to-state map of the system node S split into
CT,ext x0 = CT0x0


DT,ext u = ( DuT u ) .
Hence, for the input-to-state and output maps of the dual system nodes S ∗ and Sext
∗ are related by

Bd,F,T,ext ( yud ) = Bd,F,T yd Dd,I,T,ext ( yud ) = Dd,I,T yd + u.

In particular,
dom(Bd,F,T,ext ) = dom(Bd,F,T ) × L2 ([0, T ]; U ),
dom(Dd,F,T,ext ) = dom(Dd,F,T ) × L2 ([0, T ]; U ),
and the condition ud = 0 of Corollary 3.9 may be rewritten in terms of
TF,T,ext ( uxopt
 0

) − zf
0 = Od,F ∗ ,T,ext
RT (CT,ext x0 + DT,ext uopt − yref,ext )
TF,T ( uxopt0
!
)− zf
= Od,F ∗ ,T,ext
  
RT CT x0 +DT uopt
uopt − ( yref
0 )

TF,T ( uxopt
 0

) − zf
= Od,F ∗ ,T + RT uopt .
RT (CT x0 + DT uopt − yref )
Hence, we obtain a characterization of the control by means of the adjoint, i.e.,
x0
 
TF,T ( uopt ) −zf
uopt = − RT Od,F ∗ ,T .
RT (CT x0 +DT uopt −yref )
Note that the right-hand side corresponds to a backwards-in-time equation (due to the time flips) with
terminal value given by the derivative of the terminal cost, where the source term is the derivative of
the integrand of the stage cost.
By using that the adjoint system node is given by (4.1), and by considering the state xopt and output
yopt of the system driven by the optimal control (see (3.10)) and the adjoint state µ as in (3.11), we
obtain a boundary value problem
  h i 
ẋopt (t) A&B xopt (t)
yopt (t)
= C&D uopt (t)
, xopt (0) = x0 ,
  h i  
µ̇(t) −[A&B]d µ(t)
uopt (t) = −[C&D]d yopt (t)−yref (t) , µ(T ) = F ∗ (F xopt (T ) − zf ).

5. C ONSTRAINTS ON THE TERMINAL STATE

Here we show that also the incorporation of constraints on the terminal value are possible. To this
end, let Zc be a Hilbert space, Fc ∈ L(X, Zc ), zc ∈ Z, and assume that the set BFc as defined in (2.8)
is dense in Zc . Under Assumptions 3.1, we consider the optimal control problem
1 T
Z
1
minimize ∥y(t) − yref (t)∥2Y dt + ∥F x(T ) − zf ∥2Z
2 0 2
  h i 
ẋ(t) A&B x(t)
subject to y(t)
= C&D u(t)
, x(0) = x0 , u ∈ Uad , Fc x(T ) = zc ,
which is, in contrast to (OCP), provided with a terminal constraint whose mathematically precise
interpretation is - analogously to the procedure for the terminal penalization - given by
TFc ,T ( xu0 ) = zf . (5.1)
This leads to the additional assumption that there exists some admissible input û ∈ Uad with the
property as in Assumptions 3.1 (f) and, additionally, TFc ,T ( xu0 ) = zf . We further assume that the
operator u 7→ TFc ,T ( xu0 ) (equivalently, u 7→ TFc ,T ( u0 )) has closed range in Z, for reasons that will
22 T. REIS AND M. SCHALLER

be explained later. Our supplementary assumptions for the optimal control problem with terminal
state constraints are summarized below:
Assumptions 5.1. Zc is a Hilbert space Fc ∈ L(X, Zc ), zf ∈ Z, and the following holds for the
mappings in Table 1:
(a)
RFc := TFc ,T ( u0 ) u ∈ L2 ([0, T ]; U ) with ( u0 ) ∈ dom(TFc ,T )

(5.2)
is a closed subspace of Zc .
(b) For
x0 x0
U
g ad = {u ∈ Uad | ( u ) ∈ dom(TF,T ) ∧ TFc ,T ( u ) = zc } (5.3)
there exists an admissible input û ∈ U
g ad with

( xû0 ) ∈ dom(TF,T ) and CT x0 + DT û − yref ∈ L2 ([0, T ]; Y ).

Loosely speaking, the above condition means that there exists a state xc reachable from x0 in time
T with, further, F xc = zf . Closedness and linearity of the F -input-to-state map yields that the set
u ∈ L2 ([0, T ]; U ) ( xu0 ) ∈ dom(TFc ,T ) ∧ TFc ,T ( xu0 ) = zc


is closed and convex. Hence, the set U

g ad as in (5.3) is the intersection of two sets which are both closed
and convex and hence, it is closed and convex as well. This means, fortunately, that the inclusion of
terminal constraints of the aforementioned type is effectively encompassed by the class of optimal
control problems addressed in this article. In particular, uopt ∈ U g ad is an optimal control for (5.1), if

J (uopt ) = inf J (u).

u∈U
g ad

We may directly apply Theorem
3.7 to obtain that (3.8) holds for 2all u ∈ Uad with u − uopt ∈
g
0
dom(IF,T ) and TFc ,T u−uopt = 0. If, additionally u − uopt ∈ H0l ([0, T ]; U ), then Corollary 3.9
can be applied. In particular, since TFc ,T u−u0 opt vanishes, we obtain, by using Lemma 2.12, that

∀ λ ∈ Z : ⟨Cd,T Fc∗ λ, u − uopt ⟩H −2 ([0,T ];U ),H 2 ([0,T ];U ) = 0,

0r 0l

and the optimality condition in Corollary 3.9 means that we additionally have
2
∀u ∈ U
g ad with u − uopt ∈ H0l ([0, T ]; U ) : Re ⟨u − uopt , ud ⟩H 2 ([0,T ];U ),H −2 ([0,T ];U ) ≥ 0 (5.4)
0l 0r
−2
for all ud ∈ H0r ([0, T ]; U ) with
F ∗ TF,T ( uxopt ) − zf + Fc∗ λ
 0

ud = RT [Cd,T , Dd,T ] , λ ∈ Zc .
RT (CT x0 + DT uopt − yref )
In the case where the closure of T as defined in (3.14) possesses a nonempty relative interior in the
closed space
NFc := u ∈ L2 ([0, T ]; U ) TFc ,T ( u0 ) = 0 ,


the inequality (5.4) becomes an equation. In this scenario, for all δu ∈ dom(IF,T ) ∩ NFc , we have
that
TF,T ( uxopt
 0
 
) − zf
, IF,T δu = 0.
CT x0 + DT uopt − yref X×L2 ([0,T ];Y )
This shows that for all δu ∈ NFc ∩ H0l2 ([0, T ]; U ),

TF,T ( uxopt
  0
 
∗ ) − zf
[Cd,T F , Dd,T ] , δu = 0.
CT x0 + DT uopt − yref H −2 ([0,T ];U ),H 2 ([0,T ];U )
0r 0l
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 23

Now incorporating that RFc as in (5.2) is a closed subspace of Zc , this holds for the adjoint of the
closed and densely defined mapping u 7→ T∗Fc ,T as well. The latter is, by applying Proposition 2.13
with Y = {0}, given by the mapping λ 7→ RT Cd,T Fc∗ λ defined on the space of all λ ∈ Z and we
obtain that there exists some λ ∈ Zc with Cd,T Fc∗ λ ∈ L2 ([0, T ]; U ). Consequently, there exists some
λ ∈ Z, such that
TF,T ( uxopt
 0

∗ ) − zf
[Cd,T F , Dd,T ]
+ Cd,T Fc∗ λ = 0.
RT CT x0 + DT uopt − yref
By invoking that, for G = diag(F ∗ , Fc∗ )
[Cd,T F ∗ , Cd,T Fc∗ ] = Cd,T G,
we obtain that there exists some λ ∈ Z such that
 x0 
TF,T ( uopt )−zf

∈ dom Od,G,T
RT CT x0 +DT uopt −yref
with
λ
!
x0
TF,T ( uopt )
Od,G,T
= 0.
RT CT x0 +DT uopt −yref
This corresponds to the solution of the system
  h i 
ẋopt (t) A&B xopt (t)
y(t)
= C&D uopt (t)
, xopt (0) = x0 , Fc xopt (T ) = zc ,
  h i  
−[A&B]d µ(t)
µ̇(t)
0
= [C&D]d yopt (t)−yref (t) , µ(T ) = F ∗ (F xopt (T ) − zf ) + Fc∗ λ
for the unknowns xopt , uopt , y, µ and λ.
Let us finally give some remarks on the case where the full terminal state is prescribed.
Remark 5.2. Consider an optimal control problem with full end point constraint, i.e.,
1 T
Z
minimize ∥y(t) − yref (t)∥2Y dt
2 0
  h i 
ẋ(t) A&B x(t)
subject to y(t)
= C&D u(t)
, x(0) = x0 , u ∈ Uad , x(T ) = xf ∈ X.
It is clear that the penalization of the terminal state is obsolete in this case.
We adopt Assumptions 3.1 and, furthermore, Assumptions 5.1 (a). To ensure the validity of As-
sumption 5.1 (a), we must further assume that the space
im BT ∩ X
is a closed subspace of X. This indeed imposes a restriction, as it precludes the consideration of
various (in particular parabolic) problems. To extend the applicability of our theory to optimal control
problems with hard terminal state constraints and non-closed reachability space, additional efforts,
such as selecting alternative norms in the state space, are required. These endeavors are beyond the
scope of this article.
However, if the reachability space is closed, and, additionally, the closure of T as defined in (3.14)
possesses a nonempty relative interior in ker BI,T , the above optimal control problem leads to the
solution of the system
  h i 
ẋopt (t) A&B xopt (t)
yopt (t)
= C&D uopt (t)
, xopt (0) = x0 , xopt (T ) = xf ,
  h i  
µ̇(t) −[A&B]d µ(t)
0
= [C&D]d yopt (t)−yref (t) ,
24 T. REIS AND M. SCHALLER

6. C ONTROL OF PORT-H AMILTONIAN SYSTEMS WITH MINIMAL ENERGY SUPPLY

As a specific application of the theory presented thus far, we examine state transition of port-
Hamiltonian systems while supplying the system with the minimal amount of (physical) energy. This
optimal control problem has been suggested in [12, 32] for finite-dimensional port-Hamiltonian sys-
tems. In [30], an extension to infinite-dimensional systems was given, where, however, the considered
system class includes a bounded input operator with finite-dimensional control space and a governing
main operator which can be split into a dissipative and a skew-adjoint part.
We commence with a brief introduction to port-Hamiltonian system nodes, as developed in [29].
For sake of brevity, our setup is slightly simpler than that presented in [29]. The class nevertheless
covers a wide range of physical examples, such as Maxwell’s equations, advection-diffusion equations
and linear hyperbolic systems in one spatial variable, a wave equation in more spatial variables [11],
and Oseen’s equations [31].
Definition 6.1 (Port-Hamiltonian system). Let X, U be Hilbert spaces. Let
∗ ∗
 
M = F&G K&L : X × U ⊃ dom(M ) → X × U
be a dissipation node on (X, U ), that is,
(a) F&G : X ∗ × U ⊃ dom(F&G) → X with dom(F&G) = dom(M ) is closed;
(b) K&L ∈ L(dom(F&G), U ∗ );
′
for all u ∈ U , there exists some x′ ∈ X ∗ with xu ∈ dom(M );


(c)
(d) for the main operator F : X ⊃ dom(F) → X defined by
dom(F) := x′ ∈ X ∗ x0 ∈ dom(M )
′


′ −1
and F x′ := F&G x0 , there exists some λ > 0 such that λRX


− F has dense range, where RX
is the Riesz isomorphism on X.
Further, let H ∈ L(X, X ∗ ) be positive, self-dual and bijective. Then we call
  h i 
ẋ(t) F&G Hx(t)
y(t)
= −K&L u(t)
. (6.1)

a port-Hamiltonian system on (X, U ).

 F&G   H 0 
It follows from [29, Prop. 3.6] that −K&L 0 IU defines a system node, whence we call it a port-
Hamiltonian system node. Since H is assumed to be bounded, self-dual, positive and boundedly
invertible, the mapping
1/2
x 7→ ∥x∥H := ⟨x(t), Hx(t)⟩X,X ∗ (6.2)
is equivalent to the original norm ∥ · ∥X on X. It can be moreover seen that FH is a maximally dis-
sipative (see [10] for a definition) on X endowed with the norm ∥ · ∥H (and thus with the scalar
product ⟨·, H·⟩), whence, by [10, Chap. 2, Sec. 3b], it generates a contractive semigroup on X
equipped with ∥ · ∥H . Further, by [29, Prop. 3.11], for all T > 0, the generalized trajectories
(y, x, u) ∈ L2 ([0, T ]; U ∗ ) × C([0, T ]; X) × L2 ([0, T ]; U ) (and thus also the classical solutions)
fulfill the dissipation inequality
1
2 ⟨x(T ), Hx(T )⟩X,X− 12 ⟨x(0), Hx(0)⟩X,X ∗
∗
Z T D    E Z T
Hx(t) Hx(t)
= Re M u(t) , u(t) dt + Re⟨y(t), u(t)⟩U ∗ ,U dt.
0 X×U ∗ ,X×U 0
Z t
≤ Re⟨y(t), u(t)⟩U ∗ ,U dt. (6.3)
0
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 25

The physical interpretation of the latter is that 12 ⟨x(t), Hx(t)⟩X,X ∗ represents the stored energy at
time t ∈ [0, T ], while Re⟨y(t), u(t)⟩U ∗ ,U signifies the power supplied to the system. Consequently,
the term D    E
Hx(t) Hx(t)
− Re M u(t) , u(t) ∗
≥0
X×U ,X×U
stands for the power dissipated by the system at time t. Let us first consider the problem of minimizing
the the supplied energy while controlling the system from a given initial value x0 ∈ X to a prescribed
terminal state zf ∈ X, that is,
Z T
minimize Re⟨y(t), u(t)⟩U ∗ ,U dt
0
  h i 
ẋ(t) F&G Hx(t)
subject to y(t)
= −K&L u(t)
, x(0) = x0 ∈ X, u ∈ Uad , x(T ) = xc ∈ X.
The optimal control problem discussed here falls - at hand - outside the scope of the class addressed
in the previous section. Nevertheless, drawing inspiration from a technique highlighted in the finite-
dimensional scenario in [32], we propose a reformulation of the aforementioned optimal control prob-
lem, which enables a direct application of the theory presented earlier. Specifically, we create an
artificial output, the squared norm of which serves as a representation of dissipated power. The dis-
sipation inequality leads to the observation that minimizing the norm of this artificial output yields
the solution to the above energy minimization problem. The construction of this artificial output is
subject of the following result.
Proposition 6.2. Let h i
M= F&G
K&L
: X ∗ × U ⊃ dom(M ) → X × U ∗
be a dissipation node on (X, U ). Then there exists a Hilbert space W and R&S ∈ L(dom(F&G), W ),
such that
′

′

′

′

2
∀ xu ∈ dom(F&G) : − Re M xu , xu X×U ∗ ,X×U = 2 R&S xu W . (6.4)

Proof. Consider the mapping

Q: dom(F&G) → R,
x′ 7→ −2 Re M x′ x′




u u , u X×U ∗ ,X×U
.
Then Q is a bounded symmetric nonnegative quadratic form in the sense of [17, Chap. 6]. By using
Kato’s First Representation Theorem [17, Sec. VI.2, Thm. 2.1], there exists some self-adjoint and
nonnegative Q ∈ L(dom(F&G)), such that
′ ′ ′ ′






∀ xu ∈ dom(F&G) : Q xu = xu , Q xu dom(F&G) .
Now we set
W := (ker Q)⊥ ,
p
x′ x′ x′




R&S u := Q u ∀ u ∈ dom(F&G),
where the orthogonal complement has to be understood with respect to the inner product in dom(F&G).
This gives R&S ∈ L(dom(F&G), W ), and (6.4) holds. □

It can be seen that W and R&S can be chosen in a way that im(R&S) is dense in W . In this case,
W is uniquely determined up to isometry, and R&S is uniquely determined up to the multiplication
from the left with an isometric isomorphism.
One can readily observe that, for R&S as in Proposition 6.2,
 F&G 
H 0 
−K&L 0 IU
R&S
26 T. REIS AND M. SCHALLER

is a system node on (X, U, U ∗ × Z). Further, the dissipation inequality (6.3) gives rise to the prop-
erty that the generalized trajectories (x, u, y) of the port-Hamiltonian system (6.1) on [0, T ] fulfill
2 y
R&S ( Hxu ) ∈ L ([0, T ]; W ), and the classical (and thus also the generalized) trajectories (x, u, ( w ))
of the system
 ẋ(t)   F&G   
Hx(t)
y(t) = −K&L u(t)
(6.5)
w(t) R&S
on [0, T ] satisfy
Z T Z T
1 1 1
Re⟨y(t), u(t)⟩U ∗ ,U dt = ⟨x(T ), Hx(T )⟩X,X ∗ − ⟨x(0), Hx(0)⟩X,X ∗ + ∥w(t)∥2W dt,
0 2 2 2 0
(6.6)
which, in passing, shows that RH is an admissible output operator for the semigroup generated by
FH. We will call  
Hx(t)
w(t) = R&S u(t) (6.7)
a dissipation output for the port-Hamiltonian system (6.1).
Since the initial and terminal state are prescribed, the minimization of the supplied energy, by using
(6.6), corresponds to the minimization of the dissipated energy. Consequently, we may reformulate
the above optimal control problem to
1 T
Z
minimize ∥w(t)∥2W dt
2 0
  h i 
ẋ(t) F&G Hx(t)
subject to w(t)
= R&S u(t)
, x(0) = x0 ∈ X, u ∈ Uad , x(T ) = xc ∈ X,

which is now truly belonging to the class treated in Section 5. Recall from Remark 5.2 that the
reachability space has to be closed in order to guarantee that the findings in Section 5 are applicable.
The construction of the dissipation output also allows for appropriate reformulations of more gen-
eral energy-efficient (optimal) control problems, leading to ones discussed in earlier sections.
Remark 6.3. Let a port-Hamiltonian system (6.1) be given, and let R&S ∈ L(dom(F&G), W ) be
with the properties as in Proposition 6.2, i.e., w with (6.7) is a dissipation output for (6.1). Rather
than seeking for optimal controls that ensure a perfect landing at a predetermined terminal state, we
now only aim for a partial terminal condition on the state (or even none at all, which can be achieved
by choosing Z = {0}). Our objective is to penalize both energy wastage and a weighted deviation
from a desired terminal state xf . That is, we consider the optimal control problem
Z T
1
minimize Re⟨y(t), u(t)⟩U ∗ ,U dt + ∥F x(T ) − zf ∥2Z
2
0  h i 
ẋ(t) F&G Hx(t)
subject to y(t)
= −K&L u(t)
, x(0) = x0 , u ∈ Uad , Fc x(T ) = zc .

Now using (6.6), we obtain that this leads to the minimization of

1 T
Z
1 1
⟨x(T ), Hx(T )⟩X,X ∗ + ∥w(t)∥2W dt + ∥F x(T ) − zf ∥2Z .
2 2 0 2
To lead this back to a problem treated in earlier sections, we first observe that, for the Riesz iso-
morphism RX : X → ∗
√ X , we have that RX H is self-adjoint and positive, and thus it possesses an
operator square root RX H : X → X.
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 27

Now the above optimal control problem can be reformulated to

1 T
Z
1 h F i z
2
minimize ∥w(t)∥2W dt + √
R H
x(T ) − 0f
2 0 2 X Z×X
  h i 
ẋ(t) F&G Hx(t)
subject to w(t)
= R&S u(t)
, x(0) = x0 , u ∈ Uad , Fc x(T ) = zc ,

Last, we examine yet another category of energy-optimal control problems. For this purpose, let us
consider a system
 ẋ(t)   F&G   
Hx(t)
y(t) = −K&L u(t)
,
v(t) C&D
h i
F&G H 0

such that −K&L 0 IU is a port-Hamiltonian system node, and C&D ∈ L(dom(F&G), V ) for some
Hilbert space V and let vref ∈ L2 ([0, T ], V ). Now, let us delve into the optimal control problem that
seeks for a compromise between energy conservation and tracking of vref by v, i.e.,
Z T
1 1
minimize Re⟨y(t), u(t)⟩U ∗ ,U + ∥v(t) − vref (t)∥2V dt + ∥F x(T ) − zf ∥2Z
2 2
0ẋ(t)   F&G   
Hx(t)
subject to y(t) = −K&L u(t)
, x(0) = x0 , u ∈ Uad , Fc x(T ) = zc .
v(t) C&D
This leads, by an argumentation similar as before, to an optimal control problem
1 T  w(t)   0  2
Z
1 h F i zf
2
minimize v(t)
− vref (t) dt + √
RX H
x(T ) − 0
2 0 W ×V 2 Z×X
 ẋ(t)   F&G   
Hx(t)
subject to w(t) = R&S u(t)
, x(0) = x0 , u ∈ Uad , Fc x(T ) = zc .
v(t) C&D
Note that, by providing X with the norm ∥ · ∥H as in (6.2) and identifying the Hilbert spaces W , Z
with their respective anti-duals, the adjoint of the system node in the above optimal control problem
fulfills
 H 0  ∗ h IX 0 i F&G ∗ H 0 0
 F&G      
R&S 0 IU = 0 R−1 R&S 0 IW 0 ,
C&D
U
C&D 0 0 IV

where the second factor on the right hand side stands for the dual of
 F&G 
R&S ,
C&D
and RU : U → U∗
is the Riesz map. In particular, its domain is a dense subspace of X ∗ × W × V ,
and it maps to X × U ∗ . We would like to draw the reader’s attention, however, to the fact that the
adjoints of F and Fc with respect to the norm ∥ · ∥H must be considered in this context.

7. A PPLICATIONS

In this part, we illustrate the results through two examples involving boundary control. The first
example considers a advection-diffusion-reaction equation with Dirichlet control and Neumann ob-
servation, while the second one involves a wave equation on an L-shaped domain.
Indeed, both of the addressed problems are real, not complex. While our general theory is pri-
marily formulated for complex problems, it is equally applicable to real scenarios, as highlighted in
Remark 3.12. Specifically, it is important to emphasize that all the spaces involved are now considered
to be real.
28 T. REIS AND M. SCHALLER

7.1. Advection-diffusion-reaction equation with Dirichlet control and Neumann observation.

Consider the open unit interval [0, 1] ⊂ R, a heat diffusion coefficient a : [0, 1] → R, a convection
field b : [0, 1] → R and a reaction field c : [0, 1] → R. We inspect the advection-diffusion-reaction
equation
 
∂ ∂ ∂ ∂
∂t x(t, ξ) = ∂ξ a(ξ) ∂ξ x(t, ξ) + b(ξ) ∂ξ x(t, ξ) + c(ξ)x(t, ξ), t ≥ 0, ξ ∈ [0, 1],
with single input and output given by
∂
u(t) = x(t, 1), y(t) = a(0) ∂ξ x(t, 0), t ≥ 0.
We first gather our manageable prerequisites on the coefficients.
Assumptions 7.1.
(a) a ∈ L∞ ([0, 1]) is positive almost everywhere, and a−1 ∈ L∞ ([0, 1]);
(b) b, c ∈ L∞ ([0, 1]).

We note that the above system is not well-posed, cf. [34], such that in optimal control, this setup is
a delicate problem. We refer to [25, Section 9], where a Dirichlet optimal control problem in higher
dimension with square integrable controls and distributional output (in fact H −1 -valued Neumann
observation) was analysed. The stationary problem was thoroughly analyzed in [19].
We choose the state space X = L2 ([0, 1]), and, since our system is single-input-single-output, we
have U = Y = R. We note that the evaluation of x at one represents a bounded operator due to the
continuous embedding H 1 ([0, 1]) ,→ C([0, 1]), cf. [1].
By denoting the spatial derivative by a prime, and further setting, in coherence to the spaces intro-
duced in the first section,
1
([0, 1]) = v ∈ H 1 ([0, T ]) v(0) = 0

H0l
which due to its closedness is a Hilbert space equipped when with the standard inner product in
H 1 ([0, 1]). The system node corresponding to the advection-diffusion-reaction equation is defined by
 A&B
S = C&D with
dom A&B := ( ux ) ∈ H0l 1
([0, 1]) × R (ax′ )′ ∈ L2 ([0, 1]) ∧ x(1) = u

(7.1a)
and

′
∀ ( ux ) ∈ dom(A&B) : A&B ( ux ) = ax′ + bx′ + cx, C&D ( ux ) = (ax′ )(0). (7.1b)
The evaluation of ax′ is well-defined since ax′ ∈ H 1 ([0, 1]).
We first verify that this indeed defines a system node in the sense of Definition 2.1. In [29, Sec. 4.1],
a comparable equation was explored within a port-Hamiltonian framework. However, in that scenario,
the advection field was divergence-free, the reaction field vanished, and the complete Dirichlet and
Neumann traces were selected as the input and output, respectively. Consequently, both the main
operator and the system node itself were dissipative operators. Though dissipativity streamlined
certain steps, we are now able to present a shorter proof for a more general scenario.
To this end, we consider the “bilinear form associated to A&B”
1 1
q: H0l ([0, 1]) × H0l ([0, 1]) → R
(7.2)
(v, w) 7→ −⟨av ′ , w′ ⟩L2 ([0,1]) + ⟨bv ′ + cv, w⟩L2 ([0,1]) .
We see that q is continuous in the sense that
∃c > 0 : |q(v, w)| ≤ c∥v∥H 1 ([0,1]) ∥w∥H 1 ([0,1]) (7.3)
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 29

A crucial component for proving that the aforementioned operators constitute a system node is that
the parameter q ensures a specific, albeit weaker, form of coercivity.
Lemma 7.2. Under Assumptions 7.1, there exists some µ ∈ R and some α > 0 such that the form q
as in (7.2) fulfills
1
∀ v ∈ H0l ([0, 1]) : q(v, v) ≤ −α∥v∥2H 1 ([0,1]) + µ∥v∥2L2 ([0,1]) .

Proof. By using that a is bounded from below by a := ∥a−1 ∥−1

L∞ ([0,1]) > 0, we obtain that for all
1
v ∈ H0l ([0, 1]),
q(v, v) = −⟨av ′ , v ′ ⟩L2 ([0,1]) + ⟨bv ′ + cv, v⟩L2 ([0,1])
≤ −a∥v ′ ∥2L2 ([0,1]) + ∥b∥L∞ (0,1) ∥v ′ ∥L2 ([0,1]) ∥v∥L2 ([0,1]) + ∥c∥L∞ ([0,1]) ∥v∥2L2 ([0,1]) ,
and Young’s inequality implies
 
∥b∥L∞ (0,1) ∥v ′ ∥L2 ([0,1]) ∥v∥L2 ([0,1]) ≤ 1
2 a−1 ∥b∥2L∞ ([0,1]) ∥v∥2L2 ([0,1]) + a∥v ′ ∥2L2 ([0,1])
such that
 
q(v, v) ≤ − a2 ∥v ′ ∥2L2 ([0,1]) + ∥c∥L∞ ([0,1]) + 1
2a ∥b∥2L∞ ([0,1]) ∥v∥2L2 ([0,1]) .
Hence, choosing, e.g.,
 
a
µ= 2 + ∥c∥L∞ ([0,1]) + 1
2a ∥b∥2L∞ (0,1)

yields the claim for α = a2 . □

The above result is the main ingredient for showing the system node properties of the above intro-
duced operator A&B and C&D. We additionally that S is bijective which will be used later on.
 A&B 
Proposition 7.3. Under Assumptions 7.1, the operator S = C&D with A&B and C&D as in (7.1)
2
is a system node. Further, S : dom(S) → L ([0, 1]) × R is bijective.

Proof. We have to successively verify that S has the properties as indicated in Definition 2.1.
d): We show that A is the generator of a strongly continuous semigroup: By definition of the weak
derivative, we have Ax = z for x ∈ dom(A), z ∈ L2 ([0, 1]) if, and only if, x ∈ H01 ([0, 1]), and q as
in (7.2) fulfills
∀ φ ∈ H01 ([0, 1]) : ⟨z, φ⟩L2 ([0,1]) = q(x, φ).
Now using that q fulfills the weaker kind of coercivity in the sense of Lemma 7.2, A generates a
strongly continuous (in fact, even analytic) semigroup due to [37, §5.4 and §3.6, Theorem 3.6.1].
c): We show that for all u ∈ R, there exists some x ∈ L2 ([0, 1]) such that ( ux ) ∈ dom(S) =
dom(A&B).
First, with α and µ as in Lemma 7.2, we see that the form
q0 : H01 ([0, 1]) × H01 ([0, 1]) → R
(v, w) 7→ −⟨av ′ , w′ ⟩L2 ([0,1]) + ⟨bv ′ + (c − µ)v, w⟩L2 ([0,1])
fulfills q(v, v) > α∥v∥H 1 ([0,1]) for all v ∈ H01 ([0, 1]). Further, let xD ∈ H0l
1 ([0, 1]), such that

xD (1) = 1 (which exists by a simple linear interpolation). By the Lax-Milgram lemma, there exists
some x0 ∈ H01 ([0, 1]) such that
∀φ ∈ H01 ([0, 1]) : q0 (x0 , φ) = u⟨φ′ , ax′D ⟩L2 ([0,1]) − u⟨bx′D + (c − µ)xD , φ⟩L2 ([0,1])
30 T. REIS AND M. SCHALLER

1 ([0, 1]) fulfills x(1) = u by construction and

Clearly, x = x0 + uxD ∈ H0l
∀φ ∈ H01 ([0, 1]) : −⟨φ′ , ax′ ⟩L2 ([0,1]) = −⟨φ, bx′ + (c − µ)x⟩L2 ([0,1])
This however means that −bx′ − cx + µx ∈ L2 ([0, 1]) is the weak derivative of ax′ , which implies
that ( ux ) ∈ dom(A&B).
a): Assume that ( uxnn ) is a sequence such that ( uxnn ) → ( ux ) ∈ L2 ([0, 1]) × R and A&B ( uxnn ) →
z ∈ L2 ([0, 1]). By the already proven statement c), there exists some xD ∈ H0l 1 ([0, 1]), such that
xD xn −un xD

x−uxD


( 1 ) ∈ dom(A&B). Then 0 ∈ dom(A&B) for all n ∈ N, 0 ∈ dom(A&B),
the sequence (xn − un xD ) converges in L2 ([0, 1]) to x − uxD , and A&B xn −u0n xD converges

in L2 ([0, 1]) to z − A&B ( uxuD ). The latter means that A(xn − un xD ) converges in L2 ([0, 1]) to
z − A&B ( uxuD ). Since, by the already proven statement d), A generates a strongly continuous
semigroup, it is closed by [10, Chap. I, Thm. 1.4]. This yields
x − uxD ∈ dom(A) with A(x − uxD ) = z − A&B ( uxuD ) ,
and thus
( ux ) = + u ( x1D ) ∈ dom(A&B)
x−uxD


0
A&B ( ux ) = A&B x−ux + A&B ( uxuD ) = z − A&B ( uxuD ) + uA&B ( x1D ) = z,



with 0
D

which shows that A&B is closed.

b): Assume that the sequence ( uxnn ) is bounded in dom(A&B), that is,
( uxnn ) is bounded in L2 ([0, 1]) × R, (7.4)
(A&B ( uxnn )) 2
is bounded in L ([0, 1]). (7.5)
We first show that (xn ) is bounded in H 1 ([0, 1]). Again, by the already proven statement c), we can
choose some xD ∈ H0l 1 ([0, 1]) with ( xD ) ∈ dom(A&B). Then both sequences A(x − u x ) and
1 n n D
(xn − un xD ) are bounded in L2 ([0, 1]), whence the scalar sequence


⟨xn − un xD , A(xn − un xD )⟩L2 ([0,1])
is bounded. Now using that, by the definition of the weak derivative,
⟨xn − un xD , A(xn − un xD )⟩L2 ([0,1]) = q(xn − un xD , xn − un xD ),
whence (q(xn − un xD , xn − un xD )) is a bounded sequence. By combining this with boundedness
of (xn − un xD ) in L2 ([0, 1]) and Lemma 7.2, we deduce that (xn − un xD ) (and consequently (xn ))
forms a bounded sequence in H 1 ([0, 1]). As a result, the sequence (ax′n ) is bounded in L2 ([0, 1]).
Further combining this information with (7.5), we conclude that (ax′n ) is bounded in H 1 ([0, 1]). Since
point evaluation represents a bounded operator in H 1 ([0, 1]), we can then affirm that the sequence
(C&D ( uxnn )) = ((ax′n )(0))
is also bounded.
Now it remains to show bijectivity of S : dom(S) → L2 ([0, 1]) × R. To this end, let v ∈ L2 ([0, 1]),
w ∈ R. Then, by [14, Thm. 5.1], there exists a unique solution of the initial value problem
x
′
h i      
0 a−1 x
0 x(0)
x1 = −c −ba−1 x1 + v
, x1 (0)
= w0
in the sense of Carathéodory, which furthermore exists on [0, 1] by [14, Cor. 6.4]. This gives x′1 , x′ ∈
1 ([0, 1]). As a consequence, for x = ax′ , we have
L2 ([0, 1]), whence also x1 ∈ H 1 ([0, 1]), x ∈ H0l 1

(ax′ )′ + bx′ + cx = v, 1
x ∈ H0l ([0, 1]), (ax′ )(0) = w, (7.6)
which means that, for u = x(1), we have ( ux ) ∈ dom(S) with ( wv ) = S ( ux ).
On the other hand, to show injectivity of S, let ( ux ) ∈ ker S. Then, for x1 = ax′ , (7.6) holds with
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 31

v = 0 and w = 0. Since, by [14, Thm. 5.3], the solution of (7.6) is unique, we have x = 0, and thus
also u = x(1) = 0, and the result is proven.
□

We further identify the adjoint system node. To accomplish this, we introduce an operator-theoretic
auxiliary result.
Lemma 7.4. Let V, W be Hilbert spaces, and let F : V ⊃ dom(A) → W , G : W ⊃ dom(B) → V ,
are closed, bijective, and
∀ v ∈ dom(F ), w ∈ dom(G) : ⟨F v, w⟩W = ⟨v, Gw⟩V . (7.7)
Then F ∗ = G.

Proof. (7.7) means that dom(G) ⊂ dom(F ∗ ) with Gw = F ∗ w for all w ∈ dom(G). Hence, it
remains to show that dom(F ∗ ) ⊂ dom(G). Let w ∈ dom(F ∗ ). Then there exists some x ∈ V with
∀ v ∈ dom(F ) : ⟨F v, w⟩W = ⟨v, x⟩V .
By bijectivity of G, there exists some ŵ ∈ dom(G) with x = Gŵ, and thus,
∀ v ∈ dom(F ) : ⟨F v, w⟩W = ⟨v, x⟩V = ⟨v, Gŵ⟩V = ⟨F v, ŵ⟩W
Surjectivity of F now gives w = ŵ ∈ dom(G). □

Now we present our result on the adjoint of the system node (7.1). For sake of convenience, we
impose the additional property that b is Lipschitz continuous.
Proposition 7.5. Under Assumptions 7.1 and, additionally, b ∈ W 1,∞ ([0,h 1]), dthe
i adjoint of the
 A&B  ∗ [A&B]
operator S = C&D , with A&B and C&D as in (7.1), is given by S = [C&D]d with

(ax′d )′ − (bxd )′ ∈ L2 ([0, 1])

 
d xd 1
dom([A&B] ) := ( yd ) ∈ H ([0, 1]) × R (7.8a)
∧ xd (0) = yd ∧ xd (1) = 0
and
′
∀ ( xydd ) ∈ dom([A&B]d ) : [A&B]d ( xydd ) = ax′d − (bxd )′ + cxd ,
(7.8b)
[C&D]d ( xydd ) = −(ax′d )(1).
h i
[A&B]d e = dom([A&B]d ) and [A&B]d , [C&D]d
Proof. Consider the operator Se = [C&D]d with dom(S)
as in (7.8). Then, for all ( xyd ) ∈ dom(S),
d
e ( xy ) ∈ dom(S), we have
⟨( xydd ) , S ( ux )⟩L2 ([0,1])×R
= ⟨xd , (ax ) + bx + cx⟩L2 ([0,1]) + yd (ax′ )(0)
′ ′ ′

= −⟨x′d , ax′ ⟩L2 ([0,1]) + ⟨bxd , x′ ⟩L2 ([0,1]) + ⟨cxd , x⟩L2 ([0,1])
+ xd (0)(ax′ )(0) + xd (1)(ax′ )(1) −xd (0)(ax′ )(0)
| {z }
=0
= −⟨ax′d , x′ ⟩L2 ([0,1]) − ⟨(bxd )′ , x⟩L2 ([0,1]) + ⟨cxd , x⟩L2 ([0,1]) + bxd (1)x(1) − bxd (0)x(0)
| {z } | {z }
=0 =0
= ⟨(ax′d )′ , x⟩L2 ([0,1]) ′
− ⟨(bxd ) , x⟩L2 ([0,1]) + ⟨cxd , x⟩L2 ([0,1]) − (axd )(1) x(1) + (ax′d )(0)x(0)
′
|{z} | {z }
=u =0

= ⟨Se ( xydd ) , ( ux )⟩L2 ([0,1])×R .

32 T. REIS AND M. SCHALLER

Moreover, S is bijective by Proposition 7.3. The product rule for the weak derivative [3, Thm. 4.25]
yields (bxd )′ = b′ xd + bx′d ∈ L2 ([0, 1]), whence
e : [A&B]d ( xyd ) = ax′ ′ − bx′ + (c − b′ )xd ,
∀ ( xydd ) ∈ dom(S)


d d d

whereas the assumption b ∈ W 1,∞ ([0, 1]) gives rise to c − b′ ∈ L∞ ([0, 1]). As a consequence,
h i h i
Se := R01 I0 Se R01 I0
e

is a system node of a type that is subject of Proposition 7.3. In particular, we can conclude from
Proposition 7.3 that Se (and thus also S)
e is bijective. Altogether, we have that S and Se fulfill the
e
requirements of Lemma 7.4, and we can conclude that S ∗ = S. e
□

The optimal control problem. Having provided a formulation of the advection-diffusion-reaction

equation via system node, we will now consider the corresponding optimal control problem.
Let yref ∈ L2 ([0, T ]) and xf ∈ L2 ([0, 1]). We further consider a full terminal weight F =
1 ([0, 1]).
IL2 ([0,1]) , and we assume that the initial state fulfills x0 ∈ H0l
For α > 0 and the system node as in (7.1), we consider the optimal control problem
1 T
Z
1
minimize (y(t) − yref (t))2 + αu(t)2 dt + ∥x(T ) − xf ∥2L2 ([0,1])
2 0 2 (7.9)
  h i 
ẋ(t) A&B x(t) 2
subject to y(t)
= C&D u(t)
, x(0) = x0 , u ∈ Uad := L ([0, T ]).

This problem falls into the class treated in Section 4. To conclude existence of optimal controls and
optimality conditions, we verify the Assumptions 3.1.
 A&B 
(a) We have shown in Proposition 7.3 that C&D as in (7.1) is a system node.
(b) As N = I, the density assumption is fulfilled in view of Remark 2.11 (iib).
(c)–(d) These are satisfied by assumption.
(e) This is satisfied by construction, as the extended output is coercive with respect to the input.
(f) We choose û ≡ x0 (1). Let xD ∈ L2 ([0, 1]), such that ( x1D ) ∈ dom(A&B). For
f = û A&B ( x1D ) , z0 = x0 − ûxD ,
consider the solution z : [0, T ] → L2 ([0, 1]) of
ż(t) = Az(t) + f, z(0) = z0 . (7.10)
Since x0 ∈ H0l 1 ([0, 1]), the construction of x yields z ∈ H 1 ([0, 1]). By further using that,
D 0 0
for µ, α as in Lemma 7.2, the definition of the weak derivative gives

∀ v ∈ dom(A) : ⟨v, (−A + µI)v⟩

= ⟨av ′ , v ′ ⟩L2 ([0,1]) − ⟨bv ′ , v⟩L2 ([0,1]) − ⟨(c − µ)v, v⟩L1 ([0,1]),L∞ ([0,1]) , (7.11)
we obtain from (7.3) and Lemma 7.2 that
∀ v ∈ dom(A) : α∥v∥2H 1 ([0,1]) ≤ ⟨v, (−A + µI)v⟩ ≤ (c + µ)∥v∥2H 1 ([0,1])

with c > 0 as in (7.3). A consequence is that the domain of the fractional power (−A+µI)1/2
is the closure of dom(A) with respect to the norm in H 1 ([0, 1]). This gives rise to
z0 ∈ H01 ([0, 1]) = dom(−A + µI)1/2 .
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 33

Then [4, Part II-1, Thm. 3.1, p.143] yield that z ∈ L2 ([0, T ]; dom(A)). Consequently, x̂ :=
z + ûxD satisfies

= ( z0 ) + ûxD
∈ L2 ([0, T ]; dom(A&B)).
x̂



û û
(7.12)

We have x(0) = z0 + ûxD = x0 and, by invoking that, by (2.3), z ∈ W 1,1 ([0, T ]; X−2 ), we
have

x̂˙ = ż = Az + f = A(x̂ − ûxD ) + u A&B ( x1D )

= A&B x̂−ûx + û = A&B ( x1D ) = A&B

x̂


0
D
û ,

we obtain that û ∈ Uad with ( xû0 ) ∈ dom(TI,T ). Further, by using (7.12) together with
C&D ∈ L(dom(A&B), R), we also have

∈ L2 ([0, T ]).
x̂


y = A−1 x0 + DT û = C&D û

We now may conclude existence of an optimal control using Theorem 3.6 and its uniqueness, which
α > 0 by using the findings in Section 4.

Corollary 7.6. Let α > 0, yref ∈ L2 ([0, T ]), x0 ∈ H0l 1 ([0, 1]) and x ∈ L2 ([0, 1]). Then, under
f
2
Assumptions 7.1, X = L ([0, 1]), U = R, and A&B, C&D as in (7.1), the optimal control prob-
lem (7.9) has solution. That is, there exists an optimal control uopt ∈ L2 ([0, T ]; R2 ) in the sense of
Definition 3.3. Further, the optimal control is unique.

Further, optimality conditions may be expressed using the variational inequalities of Theorem 3.7
and Corollary 3.9. In particular, in view of the results in Section 4, the optimal control satisfies
  h i 
ẋopt (t) A&B xopt (t)
yopt (t)
= C&D uopt (t)
, xopt (0) = x0 ,
  h d
i 
µ̇(t) − [A&B] µ(t)
uopt (t) = −α−1 [C&D]d yopt (t)−yref (t) , µ(T ) = xopt (T ) − xf .

We provide now a brief numerical example illustrating the solution of this problem. We solve
the corresponding problem with an explicit Euler discretization and linear finite elements using FEn-
iCS [2] and dolfin-adjoint [26]. The optimal control problems aims at tracking yref as a Neumann
trace on the left side of the domain while actuating the Dirichlet trace on the right side. For the PDE,
we choose a(x) = 1, b(x) = −x, x0 = 0 and either c(x) = 0.1 (Figure 1) or c(x) = 5 (Figure 2).
For the cost functional, we set the reference trajectory yref (t) = sin(πt), the parameter α = 0.1 and
the time horizon T = 2.
The result is shown in Figure 1. To reach the desired Neumann trace at zero, the state at the
respective other side of the domain is chosen negative by means of the Dirichlet actuation, leading to
a positive Neumann trace at the other side. Then, after time t = 1, the Dirichlet boundary control is
chosen positive such that the Neumann trace on the opposite side becomes negative.
In Figure 2, we show the same quantities for a higher reaction coefficient. As this renders the
uncontrolled PDE unstable, the corresponding optimal state is smaller due to the penalization of the
control. Correspondingly, the optimal output has a higher disparity with the reference signal.
34 T. REIS AND M. SCHALLER

F IGURE 1. Left: Optimal state over time [0, T ] and space [0, L]. Right: Optimal
output over time.

F IGURE 2. Reaction coefficient c = 5. Left: Optimal state over time [0, T ] and
space [0, L]. Right: Optimal output over time.

7.2. Wave equation on an L-shaped domain. We consider a boundary controlled wave equation as
in [11, 20], which is given by

∂2 ∂
ρ(ξ) ∂t 2 w(ξ, t) = div(T (ξ) grad w(ξ, t)) − d(ξ) ∂t w(ξ, t), ξ ∈ Ω, t ≥ 0,
0 = w(ξ, t) ξ ∈ Γ0 , t ≥ 0,
u(ξ, t) = ν · (T (ξ) grad w(ξ, t)) ξ ∈ Γ1 , t ≥ 0, (7.13)
∂
y(ξ, t) = ∂t w(ξ, t), ξ ∈ Γ0 , t ≥ 0,
w(ξ, 0) = w0 (ξ), wt (ξ, 0) = v0 (ξ) ξ ∈ Ω,

where ν : ∂Ω → R2 is the unit outward normal vector of the L-shaped domain Ω ⊂ R2 . Its boundary
is decomposed into two parts, Γ0 and Γ1 , as depicted below.
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 35
Γ1

Ω = int(Ω1 ∪ Ω2 ),
Ω1 = (0, 1) × (0, 2), Γ0
Ω2 = (1, 2) × (0, 1), Γ0
Γ1 = (0, 1) × {2} ∪ {2} × (0, 1) ⊂ ∂Ω, Ω
Γ0 = ∂Ω \ Γ1 . Γ1

Γ0

F IGURE 3. L-shaped domain for the wave equation.

As usual, w(ξ, t) denotes the displacement of the wave at point ξ ∈ Ω and time t ≥ 0, u is the input
given by a force on the boundary part Γ1 and y is the output measured as the velocity at Γ1 . The
physical parameters are included via the Young’s modulus T (·) and the mass density ρ(·), which are
both assumed to be measurable, positive, and they have bounded inverses. The term with d can be
interpreted as an internal damping which is assumed to be a bounded nonnegative and measurable
function on Ω. Our assumptions are summarized below.
Assumptions 7.7.
(a) Ω, Γ0 , Γ1 are as in Fig. 3.
(b) T , ρ ∈ L∞ (Ω) are positive almost everywhere, with ρ−1 , T −1 ∈ L∞ (Ω);
(c) d ∈ L∞ (Ω) is nonnegative almost everywhere.
Now we show that the system can be formulated as one which is port-Hamiltonian in the sense of
Definition 6.1. The state will consist of the kinetic and potential energy variables, i.e.,
   ∂

p ρ ∂t w
x= q = −1
. (7.14)
T grad w

That is, p is the infinitesimal momentum, whereas q reflects the stress. As state space we choose
X := L2 (Ω) × L2 (Ω; R2 ), (7.15)
which is further identified with its anti-dual in the canonical manner. Further, let H : X → X with
   −1 
p
H q = ρT qp , (7.16)

and we have, by ρ > 0, T > 0, ρ, ρ−1 , T , T −1 ∈ L∞ (Ω) that H is self-adjoint with H, H −1 ∈

L(X). To introduce the corresponding dissipation node, we first mention that, by H(div; Ω) we
mean the space of square integrable functions with weak divergence in L2 (Ω). The trace operator is
denoted by γ : H 1 (Ω) → H 1/2 (∂Ω), and we have that γ is bounded and surjective due to the trace
theorem [13, Thm. 1.5.1.3]. For a relatively open subset of the boundary Γ ⊂ ∂Ω, γΓ : H 1 (Ω) →
H 1/2 (Γ) maps f ∈ H 1 (Ω) to the restriction of γf to Γ and we set
HΓ1 (Ω) = {f ∈ H 1 (Ω) | γΓ = 0}.
Then we obtain a bounded and surjective operator
1 1/2
γΓ : H∂Ω\Γ (Ω) → H0 (Γ),
1/2
where H0 (Γ) contains the elements of H 1/2 (Γ) which can be extended by zero to an element of
1/2
H 1/2 (∂Ω). Further, we consider H −1/2 (Γ) := H0 (Γ)∗ , and we may define the normal trace on
36 T. REIS AND M. SCHALLER

1/2
Γ ⊂ ∂Ω of x ∈ H(div, Ω) by w = γN,Γ x ∈ H0 (Γ)∗ , where
∀z ∈ H 1 (Ω) : ⟨w, γz⟩H −1/2 (Γ),H 1/2 (Γ) = ⟨div x, z⟩L2 (Ω) + ⟨x, grad z⟩L2 (Ω;R2 ) .
Now assume that Ω, Γ0 , Γ1 are as in Fig. 3. As input space, we choose
U = H −1/2 (Γ1 ), (7.17)
1/2
whence, consequently, U ∗ = H0 (Γ1 ). Then, for M = F&G
 
K&L with
dom(M ) = dom F&G
n v  o
:= ( F ) ∈ X × U v ∈ HΓ1 (Ω) , F ∈ H(div, Ω) ∧ γN,Γ1 F = u (7.18a)
u 0

and
 v
  v     v 
∀ (F ) ∈ dom(M ) : F&G ( F ) = div F−dv , K&L ( F ) = γΓ1 (v). (7.18b)
u u grad(v) u

the system introduced at the beginning of this section is represented by a port-Hamiltonian sys-
tem (6.1) with H as in (7.16). Here, we have replaced the clamping condition w(t)|Γ0 = 0 with
ρ−1 p(t)|Γ0 = 0 for t ≥ 0, achieved by taking the derivative with respect to time. To the best of
the authors’ knowledge, no criteria for the well-posedness of boundary-controlled wave equations on
two-dimensional spatial domains are currently available (in contrast to the one-dimensional spatial
case, as discussed in [16, Chap. 13]).
It has been shown in [11], that M is indeed a dissipative system node, and we can thus conclude that
it is a dissipation node. Further, it follows from [20, Thm. 3.2], that
Ml : X × U ⊃ dom(Ml ) = dom(M ) → X × U ∗ ,
 v   div F  (7.19)
Ml ( F
u
) = grad(v)
−γΓ1 (v)

is skew-adjoint. In particular, we have a dissipation output (6.7) with R&S ∈ L(dom(A&B), W ), for
W = X = L2 (Ω) and
 v   v  q
∀ (F u
) ∈ dom(M ) : R&S ( F
u
) = d v.
2 (7.20)
That is, in fact, R&S extends to a bounded operator on q whole X × U . This means that, actually,
v
R&S = [R S] with S = 0, and R ∈ L(X, W ) with ( F ) 7→ d2 v.
Reconstructing the displacement.
We revisit the system (7.13) whose indeterminate is the displacement w. The choice of the state x as
in (7.14), however, results in a partial loss of information about w - at first glance. Here, we briefly
discuss how to reconstruct w(t) ∈ HΓ10 (Ω) from the stress q(t) ∈ L2 (Ω) at time t ≥ 0.
Proposition 7.8. Under Assumptions 7.7, let q ∈ L2 (Ω). Then the following are equivalent:
(i) q ∈ T −1 grad HΓ10 (Ω). That is, there exists some w ∈ HΓ10 (Ω), such that
q = T −1 grad w (7.21)
(ii) For all ψ ∈ H(div, Ω) with div ψ = 0 and γN,Γ1 ψ = 0, it holds that
⟨ψ, T q⟩L2 (Ω;R2 ) = 0.
Moreover, in the case where the above statements are valid, then w with the above properties is the
unique solution of the variational problem
∀ φ ∈ HΓ10 (Ω) : ⟨grad φ, T −1 grad w⟩L2 (Ω;R2 ) = ⟨grad φ, q⟩L2 (Ω;R2 ) . (7.22)
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 37

Proof. To show that (i) implies (ii), assume that q = T −1 grad w for some w ∈ HΓ10 (Ω). By the
definition of the normal trace, any ψ ∈ H(div, Ω) and div ψ = 0, γN,Γ1 ψ = 0 fulfills
⟨ψ, T q⟩L2 (Ω) = ⟨ψ, grad w⟩L2 (Ω) =
− ⟨div ψ, grad w⟩L2 (Ω;R2 ) +⟨γN ψ , γw⟩H −1/2 (Γ ),H 1/2 (Γ ) = 0.
| {z } |{z} 1 0 1

=0 =0

Next, let us assume that (ii) holds. The Lax-Milgram lemma implies the existence of a unique w ∈
HΓ10 (Ω) satisfying (7.22). Since the solution w ∈ HΓ10 (Ω) to (7.21) is unique (provided it exists), our
remaining objective is to prove that w ∈ HΓ10 (Ω) with (7.22) indeed satisfies (7.21), a task which will
be accomplished in the following.
First, we note that, by (7.22),
∀ φ ∈ H01 (Ω) : ⟨grad φ, q − T −1 w⟩L2 (Ω;R2 ) =0,
and the definition of the weak divergence yields q − T −1 w ∈ H(div, Ω) with div(q − T −1 w) = 0.
Once again using (7.22), we obtain, for all φ ∈ HΓ10 (Ω) that

⟨γφ, γN (q − T −1 w)⟩H 1/2 (Γ −1/2 (Γ )

0 1 ),H 1

= ⟨φ, div(q − T −1 w)⟩L2 (Ω) + ⟨grad φ, q − T −1 w⟩L2 (Ω;R2 ) = 0,

| {z } | {z }
=0 (7.22)
= 0

whence we can choose ψ := q − T −1 w

in (ii) to obtain that
⟨q − T −1 grad w, T q − T −1 grad w ⟩L2 (Ω;R2 )

= ⟨ψ, T q⟩L2 (Ω;R2 ) − ⟨φ, T T −1 grad q ⟩L2 (Ω;R2 ) = 0.

| {z } | {z }
(ii) (i)⇒(ii)
=0 = 0

This indeed gives q = T −1 grad w. □

Coercivity of the form in (7.22) together with the Lax-Milgram lemma yields the existence of
a bounded operator
Fdisp : L2 (Ω; R2 ) → HΓ10 (Ω),
(7.23)
q 7→ w with (7.22).
Clearly, Fdisp is also bounded as an operator from L2 (Ω; R2 ) to L2 (Ω). Proposition 7.8 yields that
we always have that w = Fdisp q fulfills q = T −1 grad w whenever q ∈ T −1 grad HΓ10 (Ω). We give
a brief result on the range of the adjoint of Fdisp regarded as an operator from L2 (Ω; R2 ) to L2 (Ω).
Proposition 7.9. Under Assumptions 7.7, consider the operator Fdisp ∈ L(L2 (Ω; R2 ), L2 (Ω)) as in
(7.23). Then
∗
im Fdisp ⊂ {z ∈ H(div, Ω) |γN,Γ1 z = 0 } .
Proof. Let φ ∈ L2 (Ω). By Proposition 7.8, we have
∀ w ∈ HΓ10 (Ω) : w = Fdisp grad w.
Thus, for all w ∈ H01 (Ω), we have
∗
⟨φ, w⟩L2 (Ω) = ⟨φ, Fdisp grad w⟩L2 (Ω) = ⟨Fdisp φ, grad w⟩L2 (Ω) .
The definition of the weak divergence yields Fdisp ∗ φ ∈ H(div, Ω) with − div F ∗ φ = φ. Further,
disp
the definition of the normal trace gives rise to the fact that, for all w ∈ HΓ10 (Ω)
38 T. REIS AND M. SCHALLER

∗ ∗ ∗
⟨γN Fdisp φ, γw⟩H −1/2 (Γ 1/2 = ⟨div Fdisp φ, w⟩L2 (Ω) + ⟨Fdisp φ, grad w⟩L2 (Ω;R2 )
1 ),H0 (Γ1 )

= −⟨φ, w⟩L2 (Ω) + ⟨φ, Fdisp grad w⟩L2 (Ω;R2 ) = 0,

which shows the result. □
Now we show that, if the stress is initialized with an element of q ∈ T −1 grad H 1 (Ω) then the stress
Γ0
remains in that space.
Proposition 7.10. Under Assumptions 7.7, let a port-Hamiltonian system with operators as in (7.16)
and (7.18) be given, and let A be the corresponding semigroup on X = L2 (Ω) × L2 (Ω; R2 ). Then
the following holds:
(a) For all T > 0, the space L2 (Ω) × T −1 grad HΓ10 (Ω) ⊂ X is A(T )-invariant.
(b) The space of reachable states is contained in L2 (Ω) × T −1 grad HΓ10 (Ω), i.e.
∀ u ∈ L2 ([0, T ]; U ) s.t. ( u0 ) ∈ dom(TI,T ) : TI,T ( u0 ) ∈ L2 (Ω) × T −1 grad HΓ10 (Ω).
(c) For Fdisp,ext : X → X with
p
   
p
Fdisp,ext q = Fdisp q ,
it holds that the set BFdisp,ext (as defined in (2.8)) is dense in X. In particular, the Fdisp,ext -
∗
terminal value map, the Fdisp,ext -input map, and the Fdisp,ext -output map (cf. Table 1) are well-
defined.
Proof.
 
p(t)
(a) Assume that x0 = ( pq00 ) ∈ L2 (Ω) × T −1 grad HΓ10 (Ω), and denote q(t) = A(t)x0 ,
RT
t ∈ [0, T ]. By [10, Chap. II, Lem. 1.3], we have 0 A(t)x0 dt ∈ dom(FH) = H −1 dom(F)
with Z T
A(T )x0 − x0 = FH A(t)x0 dt.
0
In particular,
Z T
−1
q(t) = q0 + T grad ρ−1 p(t)dt.
0
RT
Then the claim follows, since 0 A(t)x0 dt ∈ dom(FH) implies ρ−1 p(t) ∈ HΓ10 (Ω).
(b) Let u ∈ L2 ([0, T ]; U ) such that ( u0 ) ∈ dom(TI,T ). The definition of the F -terminal value
map gives rise to the existence of a sequence 2 ([0, T ]; U ), such that (u ) converges
(un ) in H0l n
in L2 ([0, T ]; U ) to u, and (TI,T u0n ) = (BT un ) converges in X to BI,T u. Denote


 
pn (t)
qn (t)
:= BT (t)un , n ∈ N.
Then, by using that un ∈ H0l 2 ([0, T ]; U ), we have
  ρ−1 p (t)  
n
∀ t ∈ [0, T ] : T qn (t) ∈ dom(F&G)
un (t)

which particularly implies that ρ−1 pn (t) ∈ HΓ10 (Ω). Further,

 Z T   ρ−1 p (t)  
n

pn (T )
BT un qn (T ) = F&G T qn (t) dt,
0 un (t)
which again leads to
Z T
−1
qn (t) = T grad ρ−1 pn (t)dt ∈ T −1 grad HΓ10 (Ω).
0
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 39

Now Proposition 7.8 gives rise to ⟨ψ, T qn (T )⟩L2 (Ω;R2 ) = 0 for all ψ ∈ H(div, Ω) with
div ψ = 0 and γN,Γ1 ψ = 0. Now taking the limit n → ∞, we obtain ⟨ψ, T q(T )⟩L2 (Ω;R2 ) =
0, and another use of Proposition 7.8 yields that q(T ) ∈ T −1 grad HΓ10 (Ω).
(c) By using Proposition 7.9, we obtain that
∀(wv
) ∈ HΓ10 (Ω) × L2 (Ω) :
  
∗ v
v v
Fdisp,ext ( w ) = Fdisp w ∈ ( F
∗ ) ∈ X v ∈ HΓ10 (Ω) , F ∈ H(div, Ω) ∧ γN,Γ1 F = 0 .
In particular, the set
∗
v
) ∈ L2 (Ω) × L2 (Ω) Fdisp,ext

(F ∈ dom(F)
is dense in X. Now, utilizing the representation (7.19) of the dissipation node obtained by
taking the adjoint of the one with d = 0, we observe that dom(F) = dom(F∗ ) = dom((FH)∗ )
(now with respect to the standard inner product in L2 ), we find ourselves precisely in the
situation outlined in Remark 2.11 (iib). Leveraging the argumentation presented therein, we
can verify that the Fdisp,ext -input-to-state map is well-defined.
□
The optimal control problem.
We are considering an optimal control problem, as discussed in Remark 6.3, where no constraints
are imposed on the final state. We seek a control strategy that strikes a balance between minimizing
physical energy consumption and transferring the system to a final state close to a desired one. The
latter is represented by a terminal weight of the form 12 ∥Fdisp,ext x − zf ∥L2 (Ω)×L2 (Ω) , with Fdisp,ext
as in Proposition 7.10. Our objective is to achieve an energy-efficient control that (approximately)
realizes a given displacement profile wf ∈ L2 (Ω) at time T > 0 in a resting state, meaning that the
displacement velocity at time T > 0 is small. The corresponding optimal control problem is, under
Assumptions 7.7, and with X as in (7.15), H as in (7.16), U as in (7.17), F&G, K&L as in (7.18),
Z T
1
⟨y(t), u(t)⟩U ∗ ,U dt + ∥Fdisp,ext x − w0f ∥2L2 (Ω)×L2 (Ω)


minimize
2
0  h i 
ẋ(t) F&G Hx(t)
subject to y(t)
= −K&L u(t)
, x(0) = x0 , u ∈ Uad .

In particular, our terminal weight is given by 12 ∥p(T )∥2L2 (Ω) + ∥w(T ) − wf ∥2L2 (Ω) . Our controls

are supposed to be evolving pointwisely in the set of all elements H −1/2 (Γ1 ) which are represented
by square integrable functions that evolve in a box around zero. That is, for some u, u ∈ R with
u ≤ u, we consider Uad ⊂ L2 ([0, T ]; H −1/2 (Γ1 )) with
n o
Uad = u ∈ L2 ([0, T ]; H −1/2 (Γ1 ) u(t) ∈ Uad for a.e. t ∈ [0, T ] ,
(7.24)
where Uad = u ∈ L2 (Γ) u ≤ u(ξ) ≤ u for a.e. ξ ∈ Γ1 ⊂ H −1/2 (Γ1 )


According to Remark 6.3, this leads to an optimal control problem

 2
1 T x − w0f
Z 
2 1 Fdisp,ext
minimize ∥w(t)∥L2 (Ω) dt + √
2 0 2 Hx L2 (Ω)×L2 (Ω)×X
  h i 
ẋ(t) Hx(t)
subject to w(t)
= F&G
R&S u(t)
, x(0) = x0 , u ∈ Uad , .
The operator square root is simply given by the multiplication operator
√  −1/2 
H : X → X, ( pq ) 7→ ρT 1/2 qp
40 T. REIS AND M. SCHALLER

For the initial value, we choose one corresponding to an initial velocity in grad HΓ10 , together with
an initial force in H(div, Ω) with boundary trace in Uad (as defined in (7.24)), where we addition-
ally assume that the initial force is initiated by the stress that corresponds to a displacement. More
precisely, we assume
x0 = ( pq00 ) ∈ ρHΓ10 (Ω) × T −1 H(div, Ω)
with q0 ∈ T −1 grad HΓ10 (Ω) and γN,Γ1 T q0 ∈ Uad . (7.25)
To conclude existence of optimal controls and optimality conditions, we verify Assumptions 3.1.
 H 0
(a) As, by the argumentation above, F&G K&L 0 I is a port-Hamiltonian system node, we can
conclude from Proposition 6.2 that
h i
F&G  H 0 
R&S 0 IU

with R&S as in (7.20) is a system node.

(b) This is shown in Proposition (7.10) c.
(c) This is satisfied by assumption.
(d) Convexity of Uad of this set is clear. To show closedness of Uad , in view of Remark 3.2
it suffices to show that Uad as in (7.24) is a closed subset of H −1/2 (Γ1 ). Thus, consider a
sequence (un ) ∈ Uad such that un → u ∈ H −1/2 (Γ1 ). As Γ1 has a finite one-dimensional
Lebesgue measure |Γ1 |, we have
1
∀n ∈ N : ∥un ∥L2 (Γ1 ) ≤ |Γ1 | 2 max{|u|, |u|},
the sequence (un ) is bounded in L2 (Γ1 ), it is, by considering suitable subsequences, no loss
of generality to assume that (un ) is in L2 (Γ1 ) weakly convergent to some û ∈ L2 (Γ1 ). Due
to convexity and as Uad is closed in L2 (Γ1 ), we have û ∈ Uad by the separation theorem.
Further using the continuous embedding L2 (Γ1 ) ,→ H −1/2 (Γ1 ), the limits coincide, i.e.,
û = u. This shows that u ∈ Uad .
(e) As Γ1 has finite one-dimensional Lebesgue measure, Uad as in (7.24) is a bounded subset of
L2 (Γ1 ), and thus also a bounded subset of H −1/2 (Γ1 ). Now, by using that [0, T ] is a finite
interval, we verify that Uad as in (7.24) is a bounded subset of L2 ([0, T ]; H −1/2 (Γ1 )).
(f) The condition (7.25) yields that there exists some û0 ∈ Uad , such that
x0
 H 0 

û0 ∈ dom F&G 0 I .
Now we may choose the constant input û ≡ û0 ∈ Uad , and Proposition 2.3 gives rise to a
classical trajectory (x, u, y) in the sense of Definition 2.2. In particular, the final value is in
X, and the output fulfills y ∈ L2 ([0, T ]; Y ) for Y = U ∗ .
We now may conclude existence an optimal control using Theorem 3.6.
Corollary 7.11. Under Assumptions 7.7 on Ω ⊂ R2 , Γ0 , Γ1 ⊂ ∂Ω, ρ, T , d ∈ L∞ (Ω), and for T > 0,
wf ∈ L2 (Ω), and with X as in (7.15), H as in (7.16), U as in (7.17), F&G, K&L as in (7.18),
the optimal control problem (7.24) has a solution. In other words, there exists an optimal control
uopt ∈ L2 ([0, T ]; U ) in the sense of Definition 3.3.
We briefly comment on (possible) uniqueness of the optimal control.
Remark 7.12 (Uniqueness of the optimal control problem (7.24)). Under the assumptions in (7.11),
we briefly discuss possible uniqueness of the optimal control for the problem (7.24). By us-
ing the reformulation (7.25), uniqueness corresponds to the following condition: If, for some
y ∈ L2 ([0, T ]; U ∗ ), the wave equation (7.13) holds with w0 (·) = 0, v0 (·) = 0, w(·, T ) = 0,
∂ ∂ ∂
∂t w(·, T ) = 0, and d ∂t w(·, ·) ≡ 0, then u := γN,Γ1 ∂t w has to vanish constantly. This is, for
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 41

instance, guaranteed, if the wave equation on Ω with input formed by the Dirichlet boundary of the
∂
velocity on Γ0 and “distributed output” formed by d ∂t w(·, t) is observable. The latter issue has been
addressed in previous work, such as in [7], where conditions for observability are established. These
conditions are derived from geometric considerations involving supp(d) (the support of d, represent-
ing the location of damping), Ω and T , illustrated through the concept of “geodetic rays”.

Again, we provide a numerical example using FEniCS [2] and dolfin-adjoint [26]. To this end, we
choose linear finite elements for the momentum variable and linear vector-valued finite elements for
the stress. As a time-integrator, we choose the implicit midpoint rule. The parameters, including the
desired target profile and the chosen finite element grid are depicted in Figure 4.

x0 = 0, T ≡ 1, ρ ≡ 1, d ≡ 0.05,
T = 5, wf = dist(Γ0 ), αT = 10,
u = 1, u = −1

F IGURE 4. Parameters (left) and target displacement (right).

In Figure 5, we depict the part of the optimal control on the right boundary {2}×[0, 1] and the norm of
the optimal momentum. In both figures, we observe a swing-up behavior of the optimal displacement
and a small terminal momentum due to its penalization via ∥p(T )∥2L2 (Ω) in the terminal weight. The
swing-up behavior is necessary due to the control constraints limiting the force that can be applied at
the boundary.

F IGURE 5. Control on the right boundary (left) and kinetic energy over time (right)
42 T. REIS AND M. SCHALLER

Last, we provide snapshots of the displacement profile w(t) for different time instances t ∈ [0, T ]
in Figure 61. As in the right plot of Figure 5, the swing-up behavior is clearly visible. Further, as
can be seen in the in the last snapshot at the terminal time t = T = 5, the terminal displacement
w(T ) is approximating the piecewise reference signal wf depicted on the right plot of Figure 4. This
displacement is achieved with vanishing momentum variable, cf. the right plot of Figure 5.

F IGURE 6. Snapshots of the displacement variable w(t) for times t ∈

{1.5, 2.5, 3.5, 4.5, 5} (row-wise from top to bottom).

8. C ONCLUSION
In this work, we have proven existence of solutions and we have provided optimality conditions for
linear-quadratic optimal control with abstract (and not necessarily well-posed) infinite-dimensional
systems under control constraints and penalization of the terminal state and the L2 -norm of the output.
Further, terminal state constraints have been discussed. We have introduced a bunch of (not necessar-
ily bounded) solution using the theory of system nodes, which allowed us to show that under standard
1On the arXiv-page of this work, a video is available in the ancillary files section.
 LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS 43

assumptions (such as convexity and closedness of the set of admissible inputs), optimal controls do
exist. Further, we formulated an adjoint equation arising in the optimality condition by means of the
adjoint system node. We further provided applications to port-Hamiltonian system nodes for which
we have aimed for energy-optimal state transitions. Last, we have provided two numerical examples
by means of Dirichlet boundary controlled heat equation in one spatial variable and an energy effi-
cient control of a port-Hamiltonian system formed by a two-dimensional boundary-controlled wave
equation on an L-shaped domain.

R EFERENCES
[1] R. A. Adams and J. J. Fournier. Sobolev spaces. Elsevier, 2003.
[2] M. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg, C. Richardson, J. Ring, M. E. Rognes, and G. N.
Wells. The FEniCS project version 1.5. Archive of Numerical Software, 3(100), 2015.
[3] H. Alt. Linear Functional Analysis, An Application-Oriented Introduction. Universitext. Springer London, 2016.
[4] A. Bensoussan, M. C. Delfour, and S. K. Mitter. Representation and control of infinite dimensional systems, volume 1.
Birkhäuser Boston, 2007.
[5] V. Bommer and I. Yousept. Optimal control of the full time-dependent Maxwell equations. ESAIM: Mathematical
Modelling and Numerical Analysis, 50(1):237–261, 2016.
[6] M. Braack and B. Tews. Linear-quadratic optimal control for the Oseen equations with stabilized finite elements.
ESAIM: Control, Optimisation and Calculus of Variations, 18(4):987–1004, 2012.
[7] J. R. C. Bardos, G. Lebeau. Sharp sufficient conditions for the observation, control, and stabilization of waves from
the boundary. SIAM J. Control Optim., 30(5):1024–1065, 1992.
[8] R. Curtain and A. Pritchard. The infinite-dimensional Riccati equation for systems defined by evolution operators.
SIAM J. Control Optim., 14(5):951–983, 1976.
[9] I. Ekeland and R. Temam. Convex analysis and variational problems. SIAM Philadelphia, 1999.
[10] K.-J. Engel and R. Nagel. One-Parameter Semigroups for Linear Evolution Equations, volume 194 of Graduate Texts
in Mathematics. Springer, New York, 2000.
[11] B. Farkas, B. Jacob, T. Reis, and M. Schmitz. Operator splitting based dynamic iteration for linear infinite-dimensional
port-Hamiltonian systems, 2023. Submitted, preprint arXiv:2302.01195.
[12] T. Faulwasser, B. Maschke, F. Philipp, M. Schaller, and K. Worthmann. Optimal control of port-Hamiltonian descrip-
tor systems with minimal energy supply. SIAM Journal on Control and Optimization, 60(4):2132–2158, 2022.
[13] P. Grisvard. Elliptic problems in nonsmooth domains, volume 24 of Monographs and Studies in Mathematics. Pitman
Advanced Publishing Program, Boston, London, Melbourne, 1985.
[14] J. Hale. Ordinary Differential Equations. Robert E. Krieger Publishing Company, Malabar, Florida, 2nd edition, 1980.
[15] M. Hinze, R. Pinnau, M. Ulbrich, and S. Ulbrich. Optimization with PDE constraints, volume 23. Springer Dordrecht,
2008.
[16] B. Jacob and H. J. Zwart. Linear port-Hamiltonian systems on infinite-dimensional spaces, volume 223 of Operator
Theory: Advances and Applications. Springer Science & Business Media, Basel, 2012.
[17] T. Kato. Perturbation Theory for Linear Operators. Springer, Heidelberg, 2nd edition, 1980.
[18] A. Kröner, K. Kunisch, and B. Vexler. Semismooth Newton methods for optimal control of the wave equation with
control constraints. SIAM Journal on Control and Optimization, 49(2):830–858, 2011.
[19] K. Kunisch and B. Vexler. Constrained Dirichlet boundary control in L2 for a class of evolution equations. SIAM
Journal on Control and Optimization, 46(5):1726–1753, 2007.
[20] M. Kurula and H. Zwart. Linear wave systems on n-D spatial domains. International Journal of Control, 88(5):1063–
1077, 2015.
[21] I. Lasiecka and R. Triggiani. Differential and algebraic Riccati equations with application to boundary/point control
problems: continuous theory and approximation theory. Springer Heidelberg, 1991.
[22] I. Lasiecka and R. Triggiani. Control Theory for Partial Differential Equations: Volume 1, Abstract Parabolic Systems:
Continuous and approximation theories, volume 1. Cambridge University Press, 2000.
[23] I. Lasiecka and R. Triggiani. Control Theory for Partial Differential Equations: Volume 2, Abstract Hyperbolic-Like
Systems Over a Finite Time Horizon: Continuous and Approximation Theories, volume 2. Cambridge University
Press, 2000.
[24] X. Li and J. Yong. Optimal control theory for infinite dimensional systems. Birkhäuser Boston, 2012.
[25] J.-L. Lions. Optimal Control of Systems Governed by Partial Differential Equations. Springer Verlag Berlin Heidel-
berg, 1971.
44 T. REIS AND M. SCHALLER

[26] S. Mitusch, S. Funke, and J. Dokken. dolfin-adjoint 2018.1: automated adjoints for FEniCS and Firedrake. Journal of
Open Source Software, 4(38):1292, 2019.
[27] M. R. Opmeer and O. J. Staffans. Optimal control on the doubly infinite continuous time axis and coprime factoriza-
tions. SIAM Journal on Control and Optimization, 52(3):1958–2007, 2014.
[28] M. R. Opmeer and O. J. Staffans. Optimal control on the doubly infinite time axis for well-posed linear systems. SIAM
Journal on Control and Optimization, 57(3):1985–2015, 2019.
[29] F. Philipp, T. Reis, and M. Schaller. Infinite-dimensional port-Hamiltonian systems - a system node approach, 2023.
Submitted, preprint arXiv:2302.05168.
[30] F. Philipp, M. Schaller, T. Faulwasser, B. Maschke, and K. Worthmann. Minimizing the energy supply of infinite-
dimensional linear port-Hamiltonian systems. IFAC-PapersOnLine, 54(19):155–160, 2021.
[31] T. Reis and M. Schaller. Port-Hamiltonian formulation of Oseen flows, 2023. Submitted, preprint arXiv:2305.09618.
[32] M. Schaller, F. Philipp, T. Faulwasser, K. Worthmann, and B. Maschke. Control of port-Hamiltonian systems with
minimal energy supply. European Journal of Control, 62:33–40, 2023.
[33] A. Schiela. A concise proof for existence and uniqueness of solutions of linear parabolic PDEs in the context of
optimal control. Systems & Control Letters, 62(10):895–901, 2013.
[34] F. Schwenninger. Input-to-state stability for parabolic boundary control:linear and semilinear systems. In J. Kerner,
H. Laasri, and D. Mugnolo, editors, Control Theory of Infinite-Dimensional Systems, pages 83–116, Cham, 2020.
Springer International Publishing.
[35] O. J. Staffans. Well-posed linear systems, volume 103 of Encyclopedia of Mathematics and Its Applications. Cam-
bridge University Press, Cambridge, UK, 2005.
[36] O. J. Staffans and G. Weiss. Transfer functions of regular linear systems. Part III: Inversions and duality. Integral
Equations Operator Theory, 49:517–558, 2004.
[37] H. Tanabe. Equations of Evolution, volume 6 of Monographs and studies in mathematics. Pitman (Advanced Publish-
ing Program), London, 1979.
[38] C. Tretter and C. Wyss. Dichotomous Hamiltonians with unbounded entries and solutions of Riccati equations. Journal
of Evolution Equations, 14:121–153, 2014.
[39] F. Tröltzsch. Optimal control of partial differential equations: theory, methods, and applications, volume 112. Amer-
ican Mathematical Soc., 2010.
[40] M. Tucsnak and G. Weiss. Observation and Control for Operator Semigroups. Birkhäuser Advanced Texts Basler
Lehrbücher. Birkhäuser, Basel, 2009.
 A PPENDIX A. S UMMARY OF INVOLVED OPERATORS

LINEAR-QUADRATIC OPTIMAL CONTROL FOR INFINITE-DIMENSIONAL INPUT-STATE-OUTPUT SYSTEMS

Symbol Name Domain Target Space Argument Value Where defined?

semigroup x(T ) of (1.1)

A−1 (T ) X−1 X−1 x0 Rem. 2.4
on X−1 at T with u = 0

x(T ) of (1.1)
BT input-to-state map L2 ([0, T ]; U ) X−1 u eq. (2.5)
with x0 = 0

−2 y of (1.1)
CT state-to-output map X−1 H0l ([0, T ]; Y ) x0 eq. (2.6)
with u = 0

−2 y of (1.1)
DT input-to-output map L2 ([0, T ]; U ) H0l ([0, T ]; Y ) u eq. (2.7)
with x0 = 0

dom(TF,T ) ⊂
TF,T F -terminal value map Z (x0 , u) F x(T ) of (1.1) Def. 2.7
X−1 × L2 ([0, T ]; U )

dom(IF,T ) ⊂ (F x(T ), y) of (1.1)

IF,T F -input map Z × L2 ([0, T ]; Y ) u Def. 2.9
L2 ([0, T ]; U ) with x0 = 0

dom(OG,T ) ⊂ y of (1.1)
OG,T G-output map L2 ([0, T ]; Y ) (z, u) Def. 2.9
Z × L2 ([0, T ]; U ) with x0 = Gz

TABLE 1. Involved operators

45