Adaptivity in Model Order Reduction

with Proper Orthogonal Decomposition

Dissertation with the aim of achieving a doctoral degree
at the Faculty of Mathematics, Informatics and Natural Sciences
Department of Mathematics
University of Hamburg

submitted by
Carmen Gräßle

Hamburg
2019
Als Dissertation angenommen vom Fachbereich Mathematik der Universität Hamburg auf Grund
der Gutachten von:

Prof. Dr. Michael Hinze

Prof. Dr. Peter Benner
Prof. Dr. Stefan Volkwein

Datum der Disputation: 25.06.2019

Vorsitzender der Prüfungskommission: Prof. Dr. Armin Iske
Eidesstattliche Versicherung

Hiermit versichere ich an Eides statt, dass ich die vorliegende Dissertationsschrift selbst verfasst
habe und keine anderen als die angegebenen Quellen und Hilfsmittel benutzt habe.

I, the undersigned, declare upon oath that I have written the dissertation independently and have
not used further resources and aids than those stated in the dissertation.

Hamburg, 19. März 2019

 v

Abstract
This thesis is concerned with the approximation of dynamical systems and the optimal control
thereof using model order reduction based on proper orthogonal decomposition (POD). The ba-
sic idea of this approach is to replace high-fidelity models by low-order surrogates utilizing a
Galerkin projection based approach with a non-local basis.

A crucial challenge within snapshot based POD model order reduction for time-dependent sys-
tems lies in the input dependency. In an offline phase, the POD basis is computed from snapshot
data obtained by solving the high-fidelity model at several time instances. If a dynamical struc-
ture is not captured by the snapshots, this feature will be missing in the reduced-order solution.
In this sense, the accuracy of the POD surrogate solution is restricted by how well the snapshots
represent the underlying coherent structures.

If the process of snapshot sampling is restricted to uniform and static discretizations, this might
require very fine resolutions in the whole space-time domain in order to capture important
dynamical properties. As a consequence, large-scale systems arise which are computationally
expensive or even infeasible to solve.

The objective of this work is to introduce offline adaptation strategies which aim to identify
problem-specific key dynamics. In the first part, a snapshot location strategy for optimal control
problems is proposed which utilizes residual based a-posteriori error estimates in order to detect
suitable time instances for snapshot generation. In the second part, the inclusion of spatial
adaptivity in the offline phase for snapshot generation enables to resolve important structures
within the spatial domain. We consider the problem setting from an infinite-dimensional view
and investigate how POD in Hilbert spaces can be implemented in practice. It is an advantage of
this perspective that it only requires the snapshots to live in a common Hilbert space which leads
to a great flexibility concerning the actual discretization technique. Moreover, regarding optimal
control problems adaptive strategies are crucial in order to adjust the POD model according
to the current optimization iterate. In the context of model order reduction for incompressible
flows with space-adapted snapshots, two approaches are proposed in order to derive a stable
reduced-order model.
vi

Zusammenfassung
Diese Arbeit befasst sich mit der Approximation dynamischer Systeme und der Optimalsteuerung
dieser mittels Modellreduktion basierend auf Proper Orthogonal Decomposition (POD). Die
grundlegende Idee dieses Ansatzes ist, die hoch-dimensionalen Systeme durch niedrig-dimensionale
Modelle mithilfe eines Galerkin Projektions-basierten Ansatzes mit nicht-lokaler Basis zu erset-
zen.

Eine wesentliche Herausforderung innerhalb Snapshot-basierter POD Modellreduktion für zeit-

abhängige Systeme liegt in der Abhängigkeit von Eingangsdaten. In einer offline Phase wird eine
POD Basis bezüglich Snapshots berechnet, welche durch Lösen des hoch-dimensionalen Modells
an verschiedenen Zeitpunkten generiert werden. Falls eine dynamische Struktur nicht in den
Snapshots erfasst wird, bleibt diese Eigenschaft auch in der reduzierten Lösung fern. In diesem
Sinne ist die POD reduzierte Lösung dadurch eingeschränkt, wie gut die Snapshots die zugrunde
liegenden kohärenten Strukturen erkennen.

Falls der Prozess der Snapshot Generierung auf uniforme und statische Gitter eingeschränkt ist,
kann dies zu sehr feinen Auflösungen im gesamten Raum-Zeit Gebiet führen, damit wichtige dy-
namische Eigenschaften erfasst werden können. Folglich entstehen hoch-dimensionale Systeme,
welche rechnerisch aufwändig oder sogar nicht durchführbar sind.

Zielsetzung dieser Arbeit ist, offline adaptive Strategien einzuführen, welche das Ziel haben prob-
lemspezifische Schlüsseldynamiken zu identifizieren. Im ersten Teil wird eine Snapshot Platzier-
ungsstrategie für Optimalsteuerungsprobleme vorgeschlagen, welche residualbasierte a-posteriori
Fehlerschätzer verwendet um geeignete Zeitpunkte für die Snapshot Generierung zu bestimmen.
Im zweiten Teil ermöglicht die Einbindung von Ortsadaptivität in der offline Phase der Snapshot
Generierung wichtige Strukturen innerhalb des örtlichen Gebietes aufzulösen. Wir betrachten
die Problemstellung aus einer unendlich-dimensionalen Sichtweise und untersuchen wie POD in
Hilberträumen praktisch implementiert werden kann. Es ist ein Vorteil dieser Perspektive, dass
lediglich erfordert wird, dass die Snapshots in einem gemeinsamen Hilbertraum leben, was eine
große Flexibilität bezüglich der tatsächlichen Diskretisierungstechnik zur Folge hat. Außerdem
sind im Kontext von Optimalsteuerungsproblemen adaptive Strategien wesentlich um das POD
Modell an die aktuelle Optimierungsiterierte anzupassen. Im Zusammenhang der Modellreduk-
tion für inkompressible Strömungen mit ortsadaptiven Snapshots stellen wir zwei Methoden vor
um ein stabiles reduziertes Modell herzuleiten.
 vii

Publications
Some of the results of this thesis have already been published or submitted.

Section 3 contains results of

[9] A. Alla, C. Gräßle, and M. Hinze. A residual based snapshot location strategy for
POD in distributed optimal control of linear parabolic equations. IFAC-PapersOnLine,
49(8):13-18, 2016.

[10] A. Alla, C. Gräßle, and M. Hinze. Snapshot location for POD in control of a linear
heat equation. PAMM, 16(1):781-782, 2016.

[11] A. Alla, C. Gräßle, and M. Hinze. A-posteriori snapshot location for POD in optimal
control of linear parabolic equations. ESAIM:M2AN, 52(5):1847-1873, 2018.

Section 4 contains results of

[6] J. O. Alff, C. Gräßle, and M. Hinze. Adaptive trust-region POD for optimal control of
the Cahn-Hilliard equation. PAMM, 18(1):e201800453, 2018.

[84] C. Gräßle and M. Hinze. The combination of POD model reduction with adaptive
finite element methods in the context of phase field models. PAMM, 17(1):47-50, 2017.

[85] C. Gräßle, and M. Hinze. POD reduced-order modeling for evolution equations utilizing
arbitrary finite element discretizations. ACOM, 44(6):1941-1978, 2018.

[86] C. Gräßle, M. Hinze, J. Lang, and S. Ullmann. POD model order reduction with
space-adapted snapshots for incompressible flows. Accepted for publication in ACOM 2019,
preprint available [Link]

[87] C. Gräßle, M. Hinze, and N. Scharmacher. POD for optimal control of the Cahn-
Hilliard system using spatially adapted snapshots. In Numerical Mathematics and Ad-
vanced Applications ENUMATH 2017, 703–711. Springer International Publishing, 2019.

These publications build the fundament for this thesis and are therefore integrated.
viii

Acknowledgement
In the first place, I would like to express my deep gratitude to Prof. Dr. Michael Hinze for his
supervision. In particular, I would like to thank for his support, for the mathematical discussions
and valuable feedback, for giving interesting research directions and for sharing his expertise and
experience. Besides the mathematical context, I would like to thank for the opportunity to
participate in many conferences which enabled me to start building a scientific network. Next, I
would like to thank Prof. Dr. Peter Benner for his interest in my work, for reviewing my thesis and
for evaluating my defense. Moreover, my thanks go to Prof. Dr. Stefan Volkwein for reviewing
my thesis. In addition, I would like to thank for encouraging me and awaking my interest
in model order reduction and optimal control during my studies at the University Konstanz.
Further, I would like to thank Prof. Dr. Armin Iske for including me in his research team. I
gratefully acknowledge the financial support of the DFG SPP 1962 Priority Programme Non-
smooth and Complementarity-based Distributed Parameter Systems: Simulation and Hierarchical
Optimization. Next, my thanks go to Dr. Christian Kahle for sharing many codes which I could
use as a basis for my own projects and for helping me out with programming questions. Many
thanks for carefully proofreading parts of my thesis and the valuable feedback. I would like to
thank Prof. Dr. Alessandro Alla for the nice working atmosphere and his great enthusiasm about
mathematical research which is contagiously. Moreover, I would like to thank Dr. Sebastian
Ullmann for collaborating. I could learn a lot from him especially from his particular precise
and detailed working spirit. Many thanks to my colleague Nicolas Scharmacher for carefully
proofreading my thesis and his constructive feedback and remarks. Last, my thanks go to my
wonderful parents and sister and my amazing boyfriend.
CONTENTS ix

Contents
1 Introduction 2
1.1 Motivation and challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Novelty and scope of this work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Basic concepts 6
2.1 Functional analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Adaptive finite element discretization . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Optimal control and Lagrange calculus . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Model order reduction utilizing proper orthogonal decomposition . . . . . . . . . 17
2.4.1 Proper orthogonal decomposition . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.2 POD Galerkin reduced-order modeling . . . . . . . . . . . . . . . . . . . . 19
2.5 Phase field systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.1 Physical phenomenon and modeling . . . . . . . . . . . . . . . . . . . . . 22
2.5.2 Mathematical model and aspects . . . . . . . . . . . . . . . . . . . . . . . 23

3 Adaptivity with respect to time 28

3.1 Literature overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Optimal control problem with distributed control . . . . . . . . . . . . . . . . . . 30
3.3 Reformulation of the optimality system . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 A-posteriori error estimates for the time discretization . . . . . . . . . . . . . . . 37
3.5 Space-time mixed finite element discretization . . . . . . . . . . . . . . . . . . . . 38
3.6 Snapshot location in optimal control for POD model order reduction . . . . . . . 41
3.6.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.6.2 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.6.3 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.7 Located control and control constraints . . . . . . . . . . . . . . . . . . . . . . . 57
3.8 Snapshot location in optimal control for POD model order reduction with located
control and control constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.8.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.8.2 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.8.3 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.9 Further directions of research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4 Adaptivity with respect to space 73

4.1 Literature overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.2 POD with space-adapted snapshots . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.2.1 Abstract semilinear parabolic evolution problem . . . . . . . . . . . . . . 76
4.2.2 POD method with snapshots in different finite element spaces . . . . . . . 78
4.3 POD reduced-order modeling using space-adapted snapshots . . . . . . . . . . . 84
4.3.1 POD reduced-order modeling . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.3.2 Treatment of the nonlinearity . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3.3 Expressing the POD solution in the full spatial domain . . . . . . . . . . 88
4.4 Error analysis for the POD reduced-order model with space-adapted snapshots . 89
4.5 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.5.1 Linear heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.5.2 Cahn–Hilliard system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.5.3 Linear heat equation with non-nested grids . . . . . . . . . . . . . . . . . 105
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 106
4.6.1 Problem setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.6.2 POD reduced-order modeling . . . . . . . . . . . . . . . . . . . . . . . . . 109
CONTENTS 1

4.6.3 Velocity reduced-order model . . . . . . . . . . . . . . . . . . . . . . . . . 110

4.6.4 Velocity-pressure reduced-order model . . . . . . . . . . . . . . . . . . . . 112
4.6.5 Inhomogeneous Dirichlet and initial data . . . . . . . . . . . . . . . . . . 114
4.6.6 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.6.7 Further directions of research . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.7 POD model order reduction with space-adapted snapshots in optimal control . . 123
4.7.1 Optimal control of a convective Cahn–Hilliard system . . . . . . . . . . . 124
4.7.2 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

5 Summary, conclusions and outlook 131

6 Appendix 133
6.1 Snapshot location: uniform time mesh misses a signal . . . . . . . . . . . . . . . 133
2 1 INTRODUCTION

1 Introduction
1.1 Motivation and challenges
Many computational tasks in science and engineering are numerically expensive, challenging for
time-critical applications or problematic in regards to the storage needs.
We think of multi-query scenarios in which a model has to be solved many times for different
input data or problem settings. An example thereof is uncertainty quantification (UQ) where
an ensemble of simulations is required to estimate statistical quantities of the solution. Another
problem case is optimization with partial differential equation (PDE) constraints, where a system
of equations has to be solved repeatedly in order to find a minimum of a given cost functional.
An application example is shape optimization of an aircraft wing where the aim is to maximize
the lift-to-drag ratio, see [78], for instance. The need for good numerical predictions for different
material parameters or environmental influences is fundamental, since the financial costs and
engineering time for an experimental solution are infeasibly high.
A further class of problems comprises (near)-real-time scenarios, where a solution has to be avail-
able in a fraction of minutes or seconds. We think of parameter-dependent systems, for which
the actual parameter values are not known in advance. An example for such a time-critical
application is cancer treatment using thermal ablation, in which the doctor has to adapt the
heat delivery in the tissue immediately to the measured patient specific parameter data (see e.g.
[175]). A delayed reaction due to computational expenses can lead to serious health effects.
In addition, the issue of big data, handling of large-scale information sets as well as data storage
and analysis recently has become one of the major research trends. The extraction of dominant
structures, autonomous and intelligent learning from data and handling of storage issues require
advanced mathematical and numerical concepts.

All of these computational tasks have in common that they need an efficient numerical realiza-
tion. There are different approaches in order to decrease computational costs.
One option is to adapt the model fidelity by e.g. relaxing error tolerances, utilizing coarse mesh
resolutions or neglecting model properties. This can be carried out in a hierarchy of an in-
creasingly higher model accuracy when approaching a (local) solution. Examples are multigrid
methods (see e.g. [39]) and multilevel trust-region methods (see e.g. [179]). For a general review
on adaptive multilevel solution of nonlinear parabolic PDE systems we refer to [125].
Another approach of reducing computational complexity is to make use of adaptive discretization
schemes which are tailored for the quantity of interest or to the problem’s specifications. For
example, in the context of simulation and control of multiphase flows utilizing a diffuse interface
approach (see e.g. [34, 95, 114]), the use of h-adaptive finite element methods [4, 183] is crucial
for a numerical realization due to the steep transitions at the interfacial regions. The inclusion of
adaptive discretization techniques usually leads to a notable speed up in comparison to a static
discretization while ensuring accuracy. However, computations can still be quite costly.
A different strategy for tackling intensive numerical tasks is to apply model order reduction tech-
niques. The basic idea is to replace the high-dimensional problem by a low-order surrogate model
with the goal to speed up computations while obtaining solutions with a good approximation
quality. The usual procedure of reduced-order modeling can be decomposed as shown in Figure 1.

offline: online:
true solution
high-fidelity model low-order surrogate

Figure 1: Reduced-order modeling cascade

In an offline phase (training stage), solution data (so-called snapshots) is generated which ap-
proximates the true solution. The snapshots can be obtained by e.g. experimental physical mea-
1.1 Motivation and challenges 3

surements or by numerical solutions of the high-dimensional model at different time instances.

The latter is computationally intensive, since large-scale problems are solved. From this input
data, reduced spaces are constructed. Afterwards, in the online phase, the surrogate models
of low-dimension are utilized in order to compute approximate solutions in a cheap manner.
Commonly used model order reduction techniques are (amongst others) data fitting methods like
kriging [57, 31] or manifold learning [75, 174] which utilize interpolation or regression. Another
class of techniques is given by projection based methods. Examples are reduced basis methods,
rational interpolation [27] or balanced truncation [16]. In this thesis, we focus on model order
reduction utilizing proper orthogonal decomposition (POD).

For many problem settings, model order reduction utilizing POD has proven to be a powerful
tool in order to reduce time-dependent large-scale systems to surrogate models of low dimension
while preserving a good approximation quality. The range of applications of POD model order
reduction comprises a broad scope, including linear and nonlinear parabolic equations [120], op-
timal control of partial differential equations [3, 89, 103, 119, 185] and fluid dynamics [100, 128].
A general introduction to POD reduced-order modeling can be found in [105, 151, 186], for ex-
ample. The key idea of the POD technique is to apply a Galerkin ansatz, in which the ansatz
functions, i.e. the POD modes, contain information about the underlying dynamical system.
Following the approach of snapshot based POD in [166, 167], the system information is retrieved
from snapshots of the solution trajectory at several time instances, which are generated in a
simulation. Large speed up factors can be obtained due to the low dimension of the reduced
models. However, a primary challenge is the strong dependence of the reduced model on the
input data.

Let us further specify the issue of the input dependency of simulation based POD model order
reduction. The POD method is utilized in order to construct a basis which is in a certain sense a
best-approximation of the snapshot data. In particular, the POD modes are linear combinations
of the snapshots with specifically chosen weighting. In this way, the POD basis identifies key
dynamics which are captured in the snapshot set. If the snapshots do not catch an important
property of the true solution, then the POD model will not be able to reflect this - the POD
model will only be able to reflect what is given in the snapshot set. In other words, it is not
guaranteed in general that the solution to the reduced-order model is accurate in regards to the
true solution. Its accuracy depends on how well the snapshots reflect the true solution. One
possible way to address this issue is to consider online adaptivity concepts. The basic idea is
to utilize an estimator for the accuracy of the reduced-order model. If the reduced model is no
longer suitable for the considered dynamics, a reversion to the full-order model is performed and
the solution is added to the POD basis or a complete POD basis update is carried out. Some
approaches can even avoid a full-order solve. Related references are for example [17, 45, 83, 162].
In this thesis we follow a different approach.

The goal of this work is to address the issue of input dependency of simulation based POD model
order reduction by incorporating offline adaptivity concepts. The aim is to construct an adaptive,
efficient, reliable and autonomous reduced model in the following sense:

Adaptive: The adaptivity is related to problem specific structures. Either adaptive snapshot
location strategies are utilized which select suitable time instances for snapshot generation.
Or spatially adaptive schemes are used in order to resolve variations in space well.

Efficient: Very fine resolutions in the whole space-time domain are avoided by using adap-
tive strategies. This leads to a large speed up in the offline phase with only minor accuracy
loss.
4 1 INTRODUCTION

Reliable: Error estimates are provided in order to track the error of the reduced solution
with respect to the true solution. For a prescribed tolerance we can guarantee an according
POD reduced-order model accuracy.

Autonomous: We utilize adaptive methods based on error indicators which are related to
the problem by residuals, so that no user intervention is needed.

In [28, p. 249ff], adaptive discretization methods and POD reduced-order modeling are considered
as two different techniques for reducing computational complexity. In this thesis both approaches
are combined in order to exploit the respective advantages. On the one hand, in the offline phase
the use of adaptive discretization schemes for snapshot generation can remarkably reduce the
offline computation time in comparison to the use of a (very fine) uniform discretization. At
the same time a prescribed approximation quality can be ensured. On the other hand, in the
online phase, we expect to speed up the computations, since we solve reduced systems of low
dimensions.

1.2 Novelty and scope of this work

The achievements of this thesis consist in the incorporation of adaptivity concepts in simulation
based model order reduction utilizing POD and address the challenges arising from them. The
thesis is split into two major thematic areas: adaptivity with respect to time and adaptivity
with respect to space. In particular, the structure of the thesis and the specific contributions are
stated as follows:

In Section 2, we recall basic concepts and introduce notations concerning functional analy-
sis (Section 2.1), adaptive finite element discretization (Section 2.2), optimal control (Sec-
tion 2.3), model order reduction utilizing POD (Section 2.4) and phase field systems (Sec-
tion 2.5).

Section 3 is concerned with adaptivity with respect to time in optimal control. After giving
an overview of existing literature concerned with time adaptivity in the context of model
order reduction in Section 3.1, we introduce a linear quadratic optimal control problem in
Section 3.2 and recall known results concerning existence, uniqueness and regularity of the
solution. A reformulation of the optimality system into a biharmonic equation depending
on either the state or the adjoint state is presented in Section 3.3 and is based on [82]. In
Section 3.4 we give a summary of the a-posteriori error estimation from [82] for the time
discretization and specify the space-time finite element discretization for the biharmonic
equation in Section 3.5.
Having elaborated these foundations, we propose a snapshot location strategy in optimal
control for model order reduction utilizing POD in Section 3.6. The selection of suitable
time instances for snapshot generation in Section 3.6.1 is built upon the a-posteriori error
estimates from Section 3.4. In an adaptive cycle, a time grid is created by equidistributing
the local error contributions of the error indicator for the biharmonic equation on each
time interval. Therefore, the resulting adaptive time grid is related to the optimal solution.
A coarse spatial resolution is utilized in order to keep computations cheap. This is justi-
fied heuristically by a numerical observation in which spatial and temporal discretization
decouple for the considered setting. Moreover, at the same time an approximation of the
optimal control is produced which can be used as input control for snapshot generation.
Error estimates for the error between the true solution and the time-discrete POD solu-
tion to the optimal control problem are derived in Section 3.6.2. Numerical examples in
Section 3.6.3 emphasize the strength of the snapshot location method in the case of steep
transitions in the temporal component.
In Section 3.7, we focus on the case of optimal control problems with control constraints
1.2 Novelty and scope of this work 5

and located control. We derive an a-posteriori error estimate for the time-discrete solution
of the associated biharmonic equation depending on the adjoint variable in the spirit of
[82]. Moreover, we provide a snapshot location strategy for POD model reduction, derive
error estimates and perform numerical tests. Since a reformulation can only be made with
respect to the adjoint state due to the control constraints, we provide a post-processing
step based on [120] in order to adapt the time grid such that it is also a suitable time grid
for the state.

Section 4 focuses on adaptivity with respect to space. After presenting a survey of related
literature in Section 4.1, we investigate the combination of POD model order reduction with
spatially adapted snapshots in Section 4.2. The problem setting of an abstract semilinear
evolution problem is introduced in Section 4.2.1. Motivated from an infinite-dimensional
perspective, we perform the POD method in Section 4.2.2 which utilizes the snapshot
gramian. In this approach, the snapshots only need to fulfill the requirement that they
live in a common Hilbert space. We provide a numerical strategy how to implement the
inner product of finite element snapshots on adapted, non-nested grids. The numerical
realization is based on a collision detection, mesh intersection and integration on complex
polyhedra using Stoke’s formula. The POD reduced-order model is set up in Section 4.3.1,
the treatment of the nonlinearity is discussed in Section 4.3.2 and the expression of the so-
lution of the POD reduced system in the full spatial dimension is explained in Section 4.3.3.
An error analysis for the error between the true solution and the POD reduced solution
is proposed in Section 4.4. Numerical examples for the case of snapshots with nested and
non-nested discretizations are presented in Section 4.5.
We investigate POD model order reduction with space-adapted snapshots for incompressible
flows in Section 4.6. The problem setting of the Navier-Stokes equations is introduced in
Section 4.6.1. We identify the challenge arising in the case of space-adapted snapshots
in Section 4.6.2, which consists in the fact that the weak divergence-free property of the
snapshots holds only in the respective adapted finite element spaces and is not guaranteed
in a common sense. The first solution approach to derive a stable reduced-order model is
proposed in Section 4.6.3 and is a velocity reduced-order model in the spirit of [166, 167].
It utilizes an optimal projection onto a weakly divergence-free space, such that the POD
basis functions are weakly divergence-free in a common sense. The second approach is
proposed in Section 4.6.4 and is a velocity-pressure reduced-order model which utilizes
supremizer enrichment functions in the spirit of [22, 160]. The novelty of the work lies in
the consideration of space-adapted snapshots and the transfer of the forementioned known
concepts to this problem case. The treatment of inhomogeneous Dirichlet and initial data
is discussed in Section 4.6.5 and is based on the introduction of specific lifting functions.
A numerical study is carried out in Section 4.6.6 and shows that the two approaches lead
to stable reduced-order models.
In Section 4.7 we consider POD model order reduction with space-adapted snapshots in
optimal control for a convective Cahn–Hilliard system. The problem setting is described
in Section 4.7.1. The optimization aim is to control the velocity in the transport term.
We show the existence of an optimal control and present a numerical test example in
Section 4.7.2.
6 2 BASIC CONCEPTS

2 Basic concepts
This chapter recalls basic mathematical concepts which are well-known and relevant for the
thesis. It comprises definitions and theorems of functional analysis (Section 2.1), adaptive finite
element discretization (Section 2.2), optimal control (Section 2.3), model order reduction utilizing
proper orthogonal decomposition (Section 2.4) and phase field systems (Section 2.5). We do not
aim to provide a complete presentation of these areas, but focus on specific aspects which are of
particular relevance for the scope of this work. References are provided, but the notation might
be adapted to the notation within this thesis in order to ensure a consistent presentation.

2.1 Functional analysis

Function spaces

Let us start with the definition of the Lebesgue spaces of integrable functions.

Definition 2.1. (Lp (Ω) spaces, [101, Definition 1.11]). Let Ω be a Lebesgue measurable set on
Rd . We define for p ∈ [1, ∞) the seminorm
Z 1/p
p
kukLp (Ω) := |u(x)| dx
Ω

and
kukL∞ (Ω) := ess sup |u(x)| := inf{α ≥ 0 : µ(|u| > α) = 0}.
x∈Ω

Now, for p ∈ [1, ∞] we define the spaces

Lp (Ω) := {u : Ω → R Lebesgue measurable : kukLp (Ω) < ∞}.

These are not normed spaces since there exist measurable functions u : Ω → R, u 6= 0, with
kukLp (Ω) = 0. We use the equivalence relation

u∼v in Lp (Ω) :⇔ ku − vkLp (Ω) = 0 ⇔ u=v a.e.

to define Lp (Ω) = Lp (Ω)/ ∼ as the space of equivalence classes of a.e. identical functions,
equipped with the norm k · kLp . Finally we define

Lploc (Ω) := {u : Ω → R Lebesgue measurable : u ∈ Lp (K) for all K ⊂ Ω compact}

and set Lploc (Ω) := Lploc (Ω)/ ∼.

For the theory of partial differential equations, it is helpful to relax the concept of classical
differentiability to a weak sense.

Definition 2.2. (Weak partial derivative, [69, Section 5.2.1]). Suppose u, v ∈ L1loc (Ω) and α is
a multiindex. We say that v is the αth -weak partial derivative of u, written Dα u = v, provided
Z Z
uDα φ dx = (−1)|α| vφ dx
Ω Ω

for all test functions φ ∈ Cc∞ (Ω), i.e. infinitely differentiable functions with compact support in
Ω.

The Sobolev space of functions, which have weak derivatives of various orders living in Lp spaces,
is defined as follows.
2.1 Functional analysis 7

Definition 2.3. (W k,p (Ω) spaces, [101, Definition 1.14]). Let Ω ⊂ Rd be open. For k ∈ N0 ,
p ∈ [1, ∞], we define the Sobolev space W k,p (Ω) by

W k,p (Ω) = {u ∈ Lp (Ω) : u has weak derivatives Dα u ∈ Lp (Ω) for all |α| ≤ k}

equipped with the norm

 1/p

kDα ukpLp (Ω) 

X
kukW k,p (Ω) :=  , p ∈ [1, ∞),
|α|≤k
X
kukW k,∞ (Ω) := kDα ukL∞ (Ω) .
|α|≤k

For p = 2, we write H k (Ω) = W k,2 (Ω). We note that W 0,p (Ω) = Lp (Ω).

Definition 2.4. (W0k,p (Ω) spaces, [69, Section 5.2.2.]). We denote by W0k,p (Ω) the closure of
Cc∞ (Ω) in W k,p (Ω). We interpret W0k,p (Ω) as comprising those functions u ∈ W k,p (Ω) such that
“Dα u = 0 on ∂Ω” for all |α| ≤ k − 1.

It is common to write H0k (Ω) = W0k,2 (Ω).

In the context of time-dependent evolution problems, the concept of abstract function spaces,
i.e. function spaces on a bounded interval [a, b] ⊂ R with values in a Banach space, constitutes
a suitable tool for the description of weak solutions. In particular, we are concerned with the
Bochner space of integrable functions.

Definition 2.5. (Lp (a, b; X) spaces, [58, Chapter XVIII, §1.1, Definition 1]). Let X be a Banach
space and (a, b) ⊂ R an open set.
a) We denote by Lp (a, b; X), p ∈ [1, ∞), the space of (classes of ) functions: t → u(t) : (a, b) → X
such that

i) u is measurable for dt,

Z b 1/p
ii) kukLp (a,b;X) = ku(t)kpX dt < +∞.
a

b) We denote by L∞ (a, b; X) the space of (classes of ) functions u from (a, b) into X satisfying
i) and

ii)’ u is bounded almost everywhere over (a,b) and we set kukL∞ (a,b;X) = inf (M ).
ku(t)kX ≤M a.e.

Note that L2 (0, T ; Ω) and L2 (0, T ; L2 (Ω)) are isometric and isomorphic, i.e.

L2 (0, T ; Ω) ∼
= L2 (0, T ; L2 (Ω)).

Next, we consider mappings between function spaces. Let X and Y denote normed real vector
spaces and let L(X, Y ) denote the space of linear operators A : X → Y which are bounded in
the sense
kAxkY
kAk := sup = sup kAxkY < ∞.
x∈X\{0} kxkX x∈X,kxkX ≤1

Theorem 2.6. ([101, Theorem 1.2]). If Y is a Banach space then L(X, Y ) is a Banach space.

Definition 2.7. (Linear functionals, dual space and dual pairing, [101, Definition 1.5]).

i) Let X be a Banach space. A bounded linear operator u∗ : X → R, i.e. u∗ ∈ L(X; R) is

called a bounded linear functional on X.
8 2 BASIC CONCEPTS

ii) The space X ∗ := L(X, R) of linear functionals on X is called dual space of X and is (by
Theorem 2.6) a Banach space with the operator norm

ku∗ k := sup |u∗ (u)|.

x∈X,kxkX ≤1

iii) We use the notation

hu∗ , uiX ∗ ,X := u∗ (u).
h·, ·iX ∗ ,X is called the dual pairing of X ∗ and X.
We use the notation H −1 (Ω) in order to denote the dual space of H01 (Ω).

The following Riesz representation theorem allows to identify a dual space of a Hilbert space
with the Hilbert space itself.
Theorem 2.8. (Riesz representation theorem, [101, Theorem 1.4]). The dual space H ∗ of a
Hilbert space H is isometric to H itself. More precisely, for every v ∈ H the linear functional
u∗ defined by
hu∗ , uiH ∗ ,H := (v, u)H ∀u ∈ H
is in H ∗ with norm ku∗ kH ∗ = kvkH . Vice versa, for any u∗ ∈ H ∗ there exists a unique v ∈ H
such that
hu∗ , uiH ∗ ,H = (v, u)H ∀u ∈ H
and ku∗ kH ∗ = kvkH .
Consequently, the mapping H ∗ → H, u∗ 7→ v is an isometric linear isomorphism.
Definition 2.9. (Dual operator, [101, Definition 1.6]). Let X, Y be Banach spaces. Then for an
operator A ∈ L(X, Y ) the dual operator A∗ ∈ L(Y ∗ , X ∗ ) is defined by

hA∗ u, viX ∗ ,X = hu, AviY ∗ ,Y ∀u ∈ Y ∗ , v ∈ X.

The space (X ∗ )∗ is called the bidual space or double dual space.

Definition 2.10. (Reflexive space, [101, Definition 1.17]). A Banach space X is called reflexive
if the mapping u ∈ X 7→ h·, uiX ∗ ,X ∈ (X ∗ )∗ is surjective, i.e. for any u∗∗ ∈ (X ∗ )∗ there exists
u ∈ X with
hu∗∗ , u∗ i(X ∗ )∗ ,X ∗ = hu∗ , uiX ∗ ,X ∀u∗ ∈ X ∗ .
Remark 2.11. (Reflexive space, [101, Remark 1.8]). Lp is for 1 < p < ∞ reflexive, since we
have the isometric isomorphism (Lp )∗ = Lq with 1/p + 1/q = 1, and thus ((Lp )∗ )∗ = (Lq )∗ = Lp .
Moreover, any Hilbert space is reflexive by the Riesz representation theorem.
Definition 2.12. (Separable space, [156, II.3]). A metric space which has a countable dense
subset is said to be separable.
Theorem 2.13. (Separable Hilbert space, [156, Theorem II.7]). A Hilbert space H is separable
if and only if it has a countable orthonormal basis S. If there are N < ∞ elements in S, then
H is isomorphic to CN . If there are countably many elements in S, then H is isomorphic to `2 .
Definition 2.14. (Gelfand triple, [101, Definition 1.26]). Let H and V be separable Hilbert
spaces with continuous dense imbedding V ,→ H. We identify H with its dual H ∗ . Then, we
have the continuous and dense imbeddings

V ,→ H = H ∗ ,→ V ∗ ,

which is called Gelfand triple. Note that the imbedding H ,→ V ∗ is given by

v ∈ H 7→ (v, ·)H ∈ H ∗ ⊂ V ∗ .
2.1 Functional analysis 9

Often, the Hilbert space H in Definition 2.14 is called pivot space. Other names for the Gelfand
triple are evolution triplet or rigged Hilbert space. The inner product (u, v)H can also be written
as hu, viH ∗ ,H = u(v) since we identify H = H ∗ . If v ∈ V , then u(v) = hu, viV ∗ ,V also holds. In
conclusion, we have the identity (u, v)H = hu, viV ∗ ,V for u ∈ H and v ∈ V . Since V is a Hilbert
space one could also identify V with its dual V ∗ . However, it is not possible to identify H with
H ∗ and V with V ∗ at the same time, since then one would interpret u(v) also as (u, v)V .

For the description of dynamical systems it makes sense to consider a space which has a Hilbert
space structure, is continuous with respect to time t and has finite energy at each time instant.
Definition 2.15. (W (a, b; V ) spaces, [58, Chapter XVIII, §1.2, Definition 4, Proposition 6]).
Let V be a real, separable Hilbert space. Let a, b ∈ R ∪ {−∞, +∞}. We denote by W (a, b; V ) the
space
W (a, b; V ) = {u ∈ L2 (a, b; V ) : ut ∈ L2 (a, b; V ∗ )}.
It is an Hilbert space equipped with the norm
 1/2
kukW (a,b;V ) = kuk2L2 (a,b;V ) + kut k2L2 (a,b;V ∗ ) .

Concerning the regularity of elements in the space W (a, b; V ), we mention the following property.
Theorem 2.16. ([58, Chapter XVIII, §1.2, Theorem 1]). Let (V, H, V ∗ ) be a Gelfand triple. For
a, b ∈ R, every u ∈ W (a, b; V ) is almost everywhere equal to a continuous function of [a, b] in H.
Further, we have
W (a, b; V ) ,→ C 0 ([a, b]; H),
the space C 0 ([a, b]; H) being equipped with the norm of uniform convergence.

Domains, boundaries and integration formulas

In order to well-define boundary conditions and utilize integration rules according to the Green’s
formulas (Theorem 2.20), we introduce the Definition 2.17 of a Lipschitz boundary and recall
the trace Theorem 2.19.
Definition 2.17. (Lipschitz boundary, [101, Definition 1.13]). Let Ω ⊂ Rd be open and bounded.
We say that Ω has a Lipschitz boundary, if for any x ∈ ∂Ω there exists r > 0, l ∈ 1, . . . , d,
σ ∈ {−1, +1} and a function γ ∈ C 0,1 (Rd−1 ) such that
Ω ∩ B(x; r) = {y ∈ B(x; r) : σyl < γ(y1 , . . . , yl−1 , yl+1 , . . . , yd )},
where B(x; r) denotes the open ball around x with radius r.
Definition 2.18. (Normal derivative, [101, Definition 1.13]). If ∂Ω is a Lipschitz boundary we
define a.e. the unit outer normal field ~n : ∂Ω → Rd , where ~n(x), k~n(x)k2 = 1, is the outward
pointing unit normal of ∂Ω at x. We call the directional derivative
∂u
(x) := ~n(x) · ∇u(x), x ∈ ∂Ω
∂~n
the normal derivative of u.
Theorem 2.19. (Trace theorem, [101, Theorem 1.12]). Assume that Ω ⊂ Rd is open and
bounded with Lipschitz boundary. Then for all p ∈ [1, ∞] there exists a unique bounded linear
operator
T : W 1,p (Ω) → Lp (∂Ω)
such that
T u = u|∂Ω ∀u ∈ W 1,p (Ω) ∩ C(Ω̄).
Here, kT kW 1,p (Ω),Lp (∂Ω) depends only on Ω and p. T u is called the trace of u on ∂Ω.
10 2 BASIC CONCEPTS

The following theorem is frequently used in the variational theory of partial differential equations.
Theorem 2.20. (Green’s formulas, [69, C.1 Theorem 3]). Assume Ω is a bounded, open subset
of Rd and ∂Ω is C 1 . Let u, v ∈ C 2 (Ω̄). Then
Z Z
∂u
i) ∆u dx = ds,
Ω ∂Ω ∂~
n
Z Z Z
∂v
ii) Dv · Du dx = − u∆v dx + u ds,
Ω Ω ∂Ω n∂~
Z Z
∂v ∂u
iii) u∆v − v∆u dx = u −v ds.
Ω ∂Ω ∂~ n ∂~n

Lax Milgram and main theorem on monotone operators

The following Lax Milgram theorem delivers existence and uniqueness results for a class of partial
differential equations and uses the Riesz representation Theorem 2.8.
Theorem 2.21. (Lax Milgram, [69, §6.2.1, Theorem 1]). Let H be a real Hilbert space. Assume
that a : H × H → R is a bilinear mapping, for which there exist constants α, β > 0 such that
i) |a(u, v)| ≤ αkukH kvkH for u, v ∈ H (boundedness),

ii) βkuk2H ≤ a(u, u) for u ∈ H (H-coercivity).

Finally, let f : H → R be a bounded linear functional on H. Then there exists a unique element
u ∈ H such that a(u, v) = hf, viH ∗ ,H for all v ∈ H.
Let us proceed by stating theorems which are relevant for the theoretical analysis of some of
the considered problems within this thesis. The following definition of a monotone, coercive
and hemicontinuous operator is needed in order to formulate the main theorem on monotone
operators by Browder and Minty. This is used in Section 3.7 in order to prove existence of a
unique weak solution to the considered biharmonic equation.
Definition 2.22. (Monotone operator, [195, Definition 25.2.]). Let X be a real Banach space,
and let A : X → X ∗ be an operator. Then A is called monotone iff

hAu − Av, u − viX ∗ ,X ≥ 0 ∀u, v ∈ X.

Definition 2.23. (Coercive operator, [195, Definition 25.2.]). Let X be a real Banach space,
and let A : X → X ∗ be an operator. Then A is called coercive iff
hAu, uiX ∗ ,X
lim → +∞.
kukX →∞ kukX

Definition 2.24. (Hemicontinuous operator, [195, Definition 26.1]). Let X be a real Banach
space, and let A : X → X ∗ be an operator. Then A is said to be hemicontinuous iff the real
function
t 7→ hA(u + tv), wiX ∗ ,X
is continuous on [0, 1] for all u, v, w ∈ X.
Theorem 2.25. (Main theorem on monotone operators, Browder, Minty (1963), [195, Theorem
26.A]). Let A : X → X ∗ be a monotone, coercive and hemicontinuous operator on the real,
separable, reflexive Banach space X. Then, for each b ∈ X ∗ , equation Au = b, with u ∈ X has a
solution. The solution set is bounded, convex and closed. If the operator A is strictly monotone,
then the equation is uniquely solvable in X.
2.2 Adaptive finite element discretization 11

Spectral theory of operators

In the following, we provide the definition of a compact operator and the Hilbert-Schmidt theo-
rem, which are needed in the context of POD model order reduction.

Definition 2.26. (Precompact set, [156, VI.5]). Let (X, d) be a metric space and let A ⊂ X.
Then, A is called precompact (totally bounded) if every sequence in A has a subsequence that
converges in X.

Definition 2.27. (Compact operator, [156, VI.5]). Let X and Y be Banach spaces. An operator
A ∈ L(X, Y ) is called compact (or completely continuous) if A takes bounded sets in X into
precompact sets in Y . Equivalently, A is compact if and only if for every bounded sequence
{xn } ⊂ X, {Axn } has a subsequence convergent in Y .

Example 2.28. (Finite rank operators, [156, VI.5]). Suppose that the rangePof A is finite
n j
dimensional. That is, every vector in the range of A can be written Ax = j=0 αj y , for
some fixed family {y j }nj=0 in Y . If {xn } is any bounded sequence in X, the corresponding αjn
are bounded since A is bounded. The usual subsequence trick allows to extract a convergent
subsequence from {Axn } which proves that A is compact.

Theorem 2.29. (Hilbert-Schmidt theorem, [156, Th. VI.16]). Let A be a self-adjoint compact
operator on the Hilbert space X. Then, there is a complete orthonormal basis {ψn } for X so that
Aψn = λn ψn and λn → 0 as n → ∞.

Useful inequalities

The following Gronwall’s inequality is utilized in order to show energy estimates.

Lemma 2.30. (Gronwall’s inequality (differential form), [69, §B.2.j]).

i) Let η(·) be a non-negative, absolutely continuous function on [0, T ], which satisfies for a.e.
t the differential inequality
η 0 (t) ≤ φ(t)η(t) + ψ(t),
where φ(t) and ψ(t) are non-negative, summable functions on [0, T ]. Then
RT
 Z t 
φ(s)ds
η(t) ≤ e 0 η(0) + ψ(s)ds
0

for all 0 ≤ t ≤ T .

ii) In particular, if
η 0 ≤ φη on [0, T ] and η(0) = 0,
then η ≡ 0 on [0, T ].

Theorem 2.31. (Young’s inequality, [69, §B.2.c]). Let 1 < p, q < ∞, p1 + 1

q = 1. Then

ap bq
ab ≤ + (a, b > 0).
p q

2.2 Adaptive finite element discretization

The method of finite element discretization has proven to be a powerful tool for the numerical
solution of a broad class of partial differential equations. We focus on a continuous Galerkin
finite element approach. Starting from a variational formulation of the differential equation in a
Sobolev space V , a weak solution is searched in a finite-dimensional trial space Vh ⊂ V , where the
12 2 BASIC CONCEPTS

test functions live in the same subspace Vh ⊂ V . This is in contrary to Petrov-Galerkin schemes,
where different subspaces Vh , Wh ⊂ V are used as trial and test spaces. A possible construction
of such finite-dimensional subspaces Vh can be realized by the method of finite elements. For
a comprehensive presentation we refer to [54, 63, 81], for example. The basic idea of the finite
element approach is to divide the domain into finite element partitions (triangulation) and search
for polynomial approximations of the unknown function on each element. Let us introduce an
affine equivalent, admissible and regular triangulation in Definition 2.32, which is aligned with
the definition given in [4, Section 1.3.3.].

Definition 2.32. (Finite element triangulation). Let Ω be a polygonal domain with boundary
∂Ω. A finite element partition Th of Ω is a collection {T } of elements such that
S
(i) The elements form a partition of the domain, that is, Ω̄ = T ∈Th T̄ .

(ii) Each element is a simplex contained in Ω (affine equivalence).

(iii) The non-empty intersection of (the closure) of each distinct pair of elements is either a
single common vertex or a single common edge of both elements (admissibility).

(iv) There exists a constant c ≥ cT for all T ∈ Th , where cT = hT /ρT with hT being the
diameter of the simplex T and ρT is the diameter of the incircle (regularity).

The admissibility condition is often referred to as conformity and excludes the case of so-called
hanging nodes. The regularity condition means that the smallest angles of all elements is bounded
away from zero. Associated with a triangulation Th of the domain Ω, we define a finite element
space in Definition 2.33.

Definition 2.33. (Finite element space, [4, Section 1.3.4.]). Let p be a non-negative integer and
let Th be a finite element triangulation as given in Definition 2.32. The finite element subspace
of order p associated with the partition Th is defined by

Vh = {v ∈ C 0 (Ω̄) : ∀T ∈ Th , v|T ∈ Pp },

where Pp denotes the space of polynomials up to degree p.

A Lagrange or nodal basis for the finite-dimensional space Vh is given by the set of functions
{vi } with vi ∈ Vh and vi (Pj ) = δij , where δij denotes the Kronecker symbol and {Pj } is the set
of node points. In case of higher-order finite elements, there are also conditions imposed on the
derivatives of the nodal basis. The canonical basis functions have local support, which leads to
sparsity in the associated matrix representation of the discrete problem.

We note that an analogous definition of a triangulation can be done utilizing orthotopes instead
of simplices. The associated polynomial space which is then used to construct the finite element
space is the class Qp , which denotes the space of polynomials of separate degree p or less. We
will utilize such a discretization in Section 3.5.

There are three types of adaptive finite element techniques (see Figure 2):

h-adaptivity: the mesh is refined according to some error indicator (e.g. a-posteriori error
estimates, see [4, 37, 183]),

p-adaptivity: the polynomial degree of the nodal basis functions is increased according to
the smoothness of the solution (see e.g. [20]),

r-adaptivity: the node points are relocated, but the number of nodes is kept constant. This
approach is also known as moving mesh approach (see e.g. [107]).
2.2 Adaptive finite element discretization 13

Figure 2: From left to right: Initial mesh, uniform h-refinement, uniform p-refinement, possible
r-refinement

We provide details on the h-adaptive finite element approach, since this approach is frequently
used in Sections 3 and 4. The usual refinement procedure is visualized in Figure 3 and summarized
in a general Algorithm 1. Let us specify each step of the adaptive cycle.

solve → estimate → mark → refine

x
x
x
x
→ → →

Figure 3: Adaptive finite element cycle

We initialize the algorithm with an initial finite element mesh Thinit , which we also refer to as
macro-mesh. Each adapted triangulation is a refinement of the macro-mesh. We postulate that
no adapted triangulation is coarser than the initial mesh.
Associated with the current triangulation Th we construct a finite element space and solve the
problem using a Galerkin finite element approach.
In order to estimate the error between the finite element approximation and the (unknown)
solution, we utilize an error indicator based on a residual-type a-posteriori error estimation on
each element T ∈ Th .
We determine the triangles which need to be refined by the Dörfler marking strategy which is
introduced below in Definition 2.34.
As a refinement strategy, we use the newest vertex bisection [136], which has the advantage
that the resulting grids are nested. The smallest common mesh of two adapted meshes is their
overlay [49, 171]. Moreover, the newest vertex bisection conserves the shape regularity of the
triangulation.
We note that for time-dependent problems it is often necessary to introduce a coarsening.

Algorithm 1 General h-adaptive finite element algorithm.

Input: Initial finite element mesh Thinit , tolerance ε > 0, refinement parameter θ ∈ (0, 1),
problem data.
Output: Adapted mesh Th , discrete problem solution.
1: Set Th := Thinit .
P
2: while T ∈Th ηT ≥ ε do
3: Define Vh as the finite element space corresponding to Th .
4: Solve the discrete problem to find a solution in Vh .
5: Estimate the error contributions using an a-posteriori error estimate ηT for each element
T ∈ Th .
6: Mark time intervals according to the Dörfler criterion (Definition 2.34) with marking pa-
rameter θ.
7: Refine the marked intervals using bisection by newest vertex based on [136].
8: end while
14 2 BASIC CONCEPTS

Definition 2.34. (Dörfler marking, [60, strategy (M∗ ), Section 4.2]). Mark a set A ⊂ Th such
that X X
ηT ≥ (1 − θ) ηT ,
T ∈A T ∈Th

for a fixed given value θ ∈ (0, 1).

For the computation of error indicators, the jump across an edge plays an important role.

Definition 2.35. (Jumps, [183, Section 1.3.5]). With every edge E, we associate a unit vector
~nE . For interior edges E its orientation is arbitrary. For any piecewise continuous function ν
and any interior edge E, we denote by JE (ν) the jumps of ν across E in the direction of ~nE :

JE (ν)(x) = lim ν(x − t~nE ) − lim ν(x + t~nE ) for all x ∈ E.

t→0+ t→0+

Note that JE (ν) depends on the orientation of ~nE but that expressions of the form JE (~nE · ∇ν)
are independent thereof.

2.3 Optimal control and Lagrange calculus

A comprehensive introduction to the topic of optimization with PDE constraints is given in e.g.
[59, 101, 131, 176]. We recall some basic aspects here which are relevant for the scope of this
work. Generally spoken, we consider optimal control problems of the abstract form

min J(y, u) s.t. e(y, u) = 0 and u ∈ Uad , (2.1)

(y,u)∈Y ×U

where Y and U denote the state and control space, respectively, and J : Y × U → R is the
cost functional (objective functional). By e : Y × U → Z we denote an operator related to a
partial differential equation (state equation) and Uad ⊂ U is the admissible set of controls. Note
that various modifications of (2.1) are possible such as the consideration of state constraints or
inequality constraints which we do not address further within this scope. We make the following
assumptions.

Assumption 2.36. (i) Z is a Banach space and Y, U are reflexive Banach spaces.

(ii) J : Y × U → R is continuous, convex and bounded from below.

(iii) Uad ⊂ U is a closed, convex and non-empty subset.

Supposing that the state equation e(y, u) = 0 has a unique solution y ∈ Y for each control
u ∈ Uad , it is common to write y as a function of u, i.e. y = y(u). This enables to derive a
reduced problem
ˆ
min J(u) := J(y(u), u), (2.2)
u∈Uad

where the term reduced refers to the elimination of the state variable. We call a control ū ∈ Uad
a global solution to (2.2) with associated optimal state ȳ = y(ū) if

ˆ
J(ū) ˆ
≤ J(u) for all u ∈ Uad .

Further, ū ∈ Uad is called a local solution to (2.2) if there exists a neighborhood B(ū) of ū in
Uad such that
ˆ
J(ū) ˆ
≤ J(u) for all u ∈ B(ū).
In order to prove existence of an optimal control, the concepts of weak convergence, weak se-
quential compactness and lower semicontinuity are important.
2.3 Optimal control and Lagrange calculus 15

Definition 2.37. (Weak convergence, [101, Definition 1.16]). Let X be a Banach space. We say
that a sequence {uk } ⊂ X converges weakly to u ∈ X, written uk * u, if

hu∗ , uk iX ∗ ,X → hu∗ , uiX ∗ ,X as k → ∞ ∀u∗ ∈ X ∗ .

Theorem 2.38. (Weak sequential compactness, [101, Theorem 1.17]). Let X be a reflexive
Banach space. Then the following holds

i) Every bounded sequence {xn } ⊂ X contains a weakly convergent subsequence, i.e. there are
{xni } ⊂ {xn } and x ∈ X with xni * x.

ii) Every bounded, closed and convex subset A ⊂ X is weakly sequentially compact, i.e. every
sequence {xn } ⊂ A contains a weakly convergent subsequence {xni } ⊂ {xn } with xni * x,
where x ∈ A.

Theorem 2.39. (Lower semicontinuity of continuous, convex functionals, [101, Theorem 1.18]).
Let X be a Banach space. Then, any continuous, convex functional f : X → R is weakly lower
semicontinuous, i.e.
uk * u ⇒ lim inf f (uk ) ≥ f (u).
k→∞

Remark 2.40. For many optimal control problems, the following steps are carried out in order
to show existence of a solution.
ˆ
Show existence of an infimum j := inf J(u) due to Jˆ ≥ 0.
u∈Uad

ˆ k ) → j for k → ∞.
Find a minimizing sequence {uk } ⊂ Uad with J(u

Find a convergent subsequence {uki } such that uki * ū for some ū ∈ Uad .

Use Theorem 2.39 to show that ū is an optimal control.

In order to characterize a local optimal solution, we establish a first-order necessary optimal-

ity condition in Definition 2.43. For this, we need the specify the differentiability of the cost
functional J : Y × U → R and the operator e : Y × U → Z.

Definition 2.41. (Directionally, Gâteaux and Fréchet differentiability, [101, Definition 1.29]).
Let F : U ⊂ X → Y be an operator with Banach spaces X, Y and U 6= ∅ open.

(i) F is called directionally differentiable at u ∈ U if the limit

F (u + th) − F (u)
dF (u, h) = lim ∈Y
t→0+ t
exists for all h ∈ X. In this case, dF (u, h) is called directional derivative of F in the
direction h.

(ii) F is called Gâteaux differentiable at u ∈ U if F is directionally differentiable at u and

the directional derivative F 0 (u) : X 3 h 7→ dF (u, h) ∈ Y is bounded and linear, i.e.
F 0 (u) ∈ L(X, Y ).

(iii) F is called Fréchet differentiable at u ∈ U if F is Gâteaux differentiable at u and if the

following approximation condition holds:

kF (u + h) − F (u) − F 0 (u)hkY = o(khkX ) for khkX → 0.

Assumption 2.42. (i) J : Y × U → R and e : Y × U → Z are continuously Fréchet differen-

tiable.
16 2 BASIC CONCEPTS

(ii) For all u ∈ V in a neighborhood V ⊂ U of Uad , the state equation e(y, u) = 0 has a unique
solution y = y(u) ∈ Y .
(iii) ey (y(u), u) ∈ L(Y, Z) has a bounded inverse for all u ∈ V ⊃ Uad .
We recall that in (2.2) y is tied to u due to the state equation e(y, u) = 0. In particular, using
the implicit function theorem (see e.g. [194, Theorem 4B]), the state equation e(y, u) = 0 defines
locally y as a continuously Fréchet-differentiable function of u (using Assumption 2.42).
Definition 2.43. (First-order necessary optimality condition, [101, Theorem 1.48]). Let As-
sumptions 2.36 and 2.42 hold. If ū is a local solution of the reduced problem (2.2) then ū satisfies
the variational inequality
ū ∈ Uad and hJˆ0 (ū), u − ūiU ∗ ,U ≥ 0 ∀u ∈ Uad .
Note that for a convex problem setting, the first-order necessary optimality condition is sufficient.
In order to compute the derivative Jˆ0 of the reduced cost functional, two methods can be followed:
the sensitivity approach and the adjoint approach, see e.g. [101, Sections 1.6.1, 1.6.2] or [58,
Sections 8, 16]. Let us derive the adjoint gradient representation by a Lagrangian based view.
We introduce the Lagrange functional L : Y × U × Z ∗ → R by
L(y, u, p) = J(y, u) + hp, e(y, u)iZ ∗ ,Z .
Differentiation leads to
Jˆ0 (u) = Lu (y(u), u, p(u)) = Ju (y(u), u) + eu (y(u), u)∗ p(u),
where the adjoint state p = p(u) ∈ Z ∗ can be computed by solving the adjoint equation given as
ey (y(u), u)∗ p = −Jy (y(u), u).
For the numerical solution of an optimal control problem, we need to introduce a suitable dis-
cretization of the considered variables. There are two different approaches stated in Remark 2.44.
Finally, we enumerate possible optimization methods in Remark 2.45.
Remark 2.44. (Discretization concepts).
i) There are mainly two methodological approaches in order to introduce discrete concepts
in PDE constrained optimization: first-discretize-then-optimize and first-optimize-then-
discretize. If the same ansatz spaces are used for the state y and the adjoint state p, then
both approaches lead to the same discrete optimality system. However, in many situations
it makes sense to use a different ansatz space for the state y and adjoint state p, especially
if the regularity or the solution properties of the variables differ strongly. A discussion on
this issue is given in [101, Section 3.2.4].
ii) In [98] the concept of variational discretization is proposed. The idea is not to discretize
the space of admissible controls, but to implicitly derive the discretization based on the
first-order optimality condition and the discretization of the state space alone.
Remark 2.45. (Optimization methods). Depending on the problem setting and the desired
convergence rate, there are many different optimization methods available. We like to mention a
few:
Descent methods like the projected gradient method (see e.g. [101, Algorithm 2.3] or [59,
Algorithm 6]) with projected Armijo rule (see e.g. [101, Section [Link]] or [59, (5.14)]) or
Newton based methods (see e.g. [101, Algorithm 2.5] and [59, Algorithm 3]).
SQP methods (see e.g. [101, Algorithm 2.7] and [59, Algorithm 5]).
Primal-dual active set strategy (see e.g. [59, Algorithm 7]).
2.4 Model order reduction utilizing proper orthogonal decomposition 17

2.4 Model order reduction utilizing proper orthogonal decomposition

The major focus of this thesis lies in the inclusion of adaptivity concepts in the offline phase of
model order reduction utilizing proper orthogonal decomposition (POD). The goal of POD based
model order reduction is to replace computationally expensive high-fidelity models by low-order
surrogates in order to speed up computations and at the same time ensure a good approximation
quality. POD model order reduction is a Galerkin projection based method. As trial and test
spaces we use reduced spaces which are spanned by POD basis functions. These POD modes
have global support and contain information about characteristic properties of the solution. This
is in contrary to e.g. finite element methods, where the basis functions are uncorrelated to the
physical properties of the underlying dynamical system. The reflection of solution properties
within the reduced space is the key feature which enables to use only a small basis length, i.e.
reduced spaces of low dimension.
In this section, we recall known concepts and results concerning the POD method (Section 2.4.1)
and reduced-order modeling (Section 2.4.2).

2.4.1 Proper orthogonal decomposition

The method of proper orthogonal decomposition (POD) is also known as Karhunen-Loeve expan-
sion, empirical orthogonal functions (EOF) or principal component analysis (PCA) and goes back
to [146]. It is utilized in a broad spectrum of disciplines with different applications such as signal
analysis and pattern recognition [77], digital picture processing [159, Chapter 5], fluid dynamics
and coherent structures [105, 166, 167], data compression [15], meteorology and oceanography
[152], control of distributed parameter systems [80] and many more.

Let us assume we are given a set V ⊂ X. The principle aim within many model order reduction
approaches is to approximate the set V by some n-dimensional linear subspace Xn ⊂ X. A
natural question to ask is what is the best subspace Xn and how well can V be approximated
by Xn . A fundamental measure how well suited a low n-dimensional reduced space is for the
approximation of the solution space V is given by the Kolmogorov n-width according to [117].
Let us consider the deviation of V from Xn defined by
E(V; Xn ) := sup inf kx − ykX . (2.3)
x∈V y∈Xn

Then, the Kolmogorov n-width of V in X is defined as

dn (V, X) := inf E(V; Xn ),
Xn ⊂X

compare e.g. [150, Definition 1.1]. A small Kolmogorov n-width means that there exists a sub-
space Xn ⊂ X which approximates V well. This is a fundamental prerequisite for a successful
model order reduction.

In this thesis, our interest lies in approximating solution trajectories of the form
V = {y(t) : t ∈ [0, T ]} ⊂ X. (2.4)
As an L2 (0, T ; X)-analogon to (2.3), we define
` 2
Z T X
E(V; X` ) := y(t) − (y(t), ψi )X ψi dt = ky(t) − P ` y(t)k2L2 (0,T ;X)
0 i=1 X
P`
where P ` denotes an orthogonal projection P ` : X → X` , v 7→ i=1 (v, ψi )X ψi and {ψ1 , . . . , ψ` }
is an orthonormal basis of X` . The solution to
δ` (V, X) := inf E(V; X` )
X` ⊂X
18 2 BASIC CONCEPTS

is given by the POD space. This means that the POD space creates an optimal `-dimensional
subspace with respect to the mean of V, compare [153, Remark 6.3].

Let us now specify how a POD reduced-order space is constructed. The POD method in Hilbert
spaces is explained in [89, 121], for example. Here, we recall the main aspects. We specify X
as a real separable Hilbert space. The aim of the POD method is to describe the space V (2.4)
by means of few orthonormal functions {ψi }`i=1 ⊂ X with ` ≤ d := dimV ≤ ∞, such that the
error between the original trajectory y(t) and its projection onto the `-dimensional subspace
X` = span{ψ1 , . . . , ψ` } ⊂ X is minimized in a mean square sense

` 2
Z T X
min y(t) − (y(t), ψi )X ψi dt s.t. (ψi , ψj )X = δij for 1 ≤ i, j ≤ `, (2.5)
ψ1 ,...,ψ` ∈X 0 i=1 X

where δij denotes the Kronecker symbol.

In practice, the whole trajectory y(t) is not available, but snapshots of the solution at certain
instances. These snapshots can be measurement data of physical experiments or finite element
samples of the solution trajectory V on a given time grid, for example. In order to generate
information data of the underlying dynamical system, we follow the idea in [166, 167], where
numerical solutions at many time instances are computed and stored as snapshots of the solution
trajectory. From this snapshot data, a suitable orthonormal basis is computed in a least-squares
sense which extracts the dominant, coherent structures. Assume we are given the snapshots

y 0 , . . . , y n ∈ X. (2.6)

We determine a POD basis {ψ1 , . . . , ψ` } ⊂ X of rank ` by solving the minimization problem

n ` 2
X X
j j
min αj y − (y , ψi )X ψi s.t. (ψi , ψj )X = δij for 1 ≤ i, j ≤ `, (2.7)
ψ1 ,...,ψ` ∈X
j=0 i=1 X

where αj are non-negative weights. If we consider the case in which {y j }nj=0 are finite element
solutions of an evolution equation at time instances {tj }nj=0 , then the weights αj are often chosen
as trapezoidal weights, i.e.
∆t1 ∆tj+1 + ∆tj ∆tn
α0 = , αj = for j = 1, . . . , n − 1, αn = ,
2 2 2
where we denote by ∆tj = tj − tj−1 the length of the time intervals for j = 1, . . . , n. A solution
to (2.7) is called a rank-` POD basis. For the equality constrained minimization problem (2.7),
first-order necessary (and by convexity sufficient) optimality conditions can be derived. For this
purpose, we introduce the bounded linear operator Y : Rn+1 → X by
n
X √
Yφ = αj φj y j for φ = (φ0 , . . . , φn ) ∈ Rn+1 . (2.8)
j=0

Its Hilbert space adjoint Y ∗ : X → Rn+1 satisfies (Yφ, ψ)X = (φ, Y ∗ ψ)Rn+1 for φ ∈ Rn+1 and
ψ ∈ X and is given by
√
(ψ, α0 y 0 )X
 

Y ∗ψ =  ..
 for ψ ∈ X. (2.9)
 
.
√ n
(ψ, αn y )X
We introduce the operator
R := YY ∗ : X → X, (2.10)
2.4 Model order reduction utilizing proper orthogonal decomposition 19

whose action is given by

n
X
Rψ = αj (ψ, y j )X y j for ψ ∈ X.
j=0

It can be shown that the operator R is a linear, bounded, non-negative and self-adjoint operator.
Since the image R(X) = span{y 0 , . . . , y n } has finite dimension, the operator R is compact,
compare Example 2.28. Therefore the Hilbert-Schmidt Theorem 2.29 can be applied which
ensures the existence of a complete orthonormal basis {ψi } for X and a sequence of corresponding
non-negative eigenvalues {λi } with

Rψi = λi ψi with λ1 ≥ . . . ≥ λd > 0 and λi = 0 for all i > d. (2.11)

The following theorem states the necessary (and by convexity sufficient) optimality conditions
for problem (2.7) and presents the POD projection error. It is taken from [89, Theorem 2.7],
where it is formulated and proved for multiple snapshots.

Theorem 2.46. Let X be a separable real Hilbert space, y 0 , . . . , y n ∈ X and R : X → X

defined by (2.10). Suppose that {λi } and {ψi } denote the non-negative eigenvalues and associated
orthonormal eigenfunctions of R satisfying (2.11). Then, for every ` ∈ {1, . . . , d} the first `
eigenfunctions {ψi }`i=1 solve (2.7). Moreover, the value of the cost evaluated at the optimal
solution {ψi }`i=1 satisfies

n ` 2 d
X X X
j
αj y − (y j , ψi )X ψi = λi . (2.12)
j=0 i=1 X i=`+1

The basis {ψi }`i=1 can alternatively be computed via singular value decomposition (SVD). The
SVD of the operator Y is given by
d
X √
Y= σi αi (·, φi )Rn+1 ψi ,
i=1
√
where σ1 ≥ . . . ≥ σd > 0 is the ordered sequence of singular values of Y with σi = λi for
i = 1, . . . , d. For more details we refer to [156, Theorem VI.17], for instance.

In order to choose the dimension of the reduced space X` = span{ψ1 , . . . , ψ` }, we utilize the
following criterion. The information content of a POD basis of rank ` relatively to the amount
of the information content of all snapshots is given by the ratio of modeled information and total
information. It is defined by
P`
λi
Γ(`) := Pdi=1 .
i=1 λi
We will choose the POD basis length ` such that

`min = argmin{Γ(`) : Γ(`) > 1 − p},

for a given value p representing the loss of information.

2.4.2 POD Galerkin reduced-order modeling

After the computation of a POD reduced space X` ⊂ X, we utilize a Galerkin approach in
order to derive a reduced-order model. In particular, we start from a variational formulation of
the considered differential equation and search for a weak solution in the trial space X` ⊂ X,
20 2 BASIC CONCEPTS

where the test functions live in the same POD space X` . This leads to an `-dimensional system
of equations for the time-dependent expansion coefficients in the POD Galerkin ansatz for the
reduced solution.

POD reduced-order modeling for optimal control

Let us consider the optimal control problem (2.1). Analogously to the two structurally different
discretization concepts in Remark 2.44(i), we consider two concepts for reduced-order modeling
of optimal control problems given by
first-optimize-then-reduce vs. first-reduce-then-optimize
as illustrated in Figure 4. We start with a description of the latter approach, where we first re-
duce the optimal control problem and then derive the respective optimality system. A discussion
of modifications and other approaches is given afterwards.

optimal control problem

optimize reduce

KKT system of optimal reduced optimal

control problem control problem

reduce optimize

reduced KKT system of KKT system of reduced

optimal control problem optimal control problem

Figure 4: First-optimize-then-reduce vs. first-reduce-then-optimize

First-reduce-then-optimize. In order to reduce the optimal control problem (2.1), we introduce

the reduced state space Y` ⊂ Y . This space can be computed by a POD of a suitable snapshot
set (compare Remark 2.47). The reduced optimal control problem reads as

min J` (y` , u` ) s.t. e` (y` , u` ) = 0 and u` ∈ Uad (2.13)

(y` ,u` )∈Y` ×U

with reduced cost functional J` : Y` × U → R and e` : Y` × U → Z. Note that the control

is not reduced yet, but its reduction will be determined by the relation between the reduced
adjoint state and control given by the first-order optimality condition. Hence, the reduction of
the control is conservative (compare [101, Note 3.2]).
In the weak form of the state equation we search for a POD reduced state y` ∈ Y` and use
the POD space as test space. The reduced adjoint space p` then will also be in the same POD
space by construction, since the POD space is the test space of the state equation. Hence, both
reduced state and adjoint variables live in the same POD reduced space. If a solution to (2.1)
exists, then by construction there exists a solution to (2.13).
Remark 2.47. (Integrated snapshot set). If the snapshot set contains pure state information,
it is shown in [103, Proposition 4.7 and Remark 4.8(c)] that the error in the adjoint variable
depends on a POD projection error of the adjoint. This error can be avoided if one considers a
snapshot ensemble of both state and adjoint information. This approach is commonly used, see
e.g. [89]. In [116], for example, this is referred to as integrated snapshot set.
Discussion. It is also imaginable to follow the first-optimize-then-reduce approach, where the
first-order optimality system is derived and then all infinite-dimensional spaces are replaced by
2.5 Phase field systems 21

reduced-order spaces. The advantage of this approach is that different reduced spaces can be
used for the state and the adjoint state. Especially in the case in which the state and adjoint
regularities and dynamics are very different, this approach might better reflect the respective
properties. However, one has to be careful in this case: first, one has to prove that the solution
of the reduced optimality system is in fact a solution to a respective optimality problem. This
question is in general not easy to answer. Moreover, terms of the form
Z
ψiy ψjp dx
Ω

will appear, where ψiy denotes a POD basis function corresponding to state snapshots and ψjp is a
POD mode computed from adjoint snapshots. This potentially leads to non-symmetric matrices.
Moreover, if the POD reduced space for the state and for the adjoint state have different dimen-
sions, the corresponding matrices are rectangular. Thus, solvability has to be ensured. Finally,
gradient based optimization methods might not be applicable.

Treatment of nonlinearities

The POD method is a linear procedure, even if the underlying system is nonlinear. This is the
strength and weakness of the method. It is well known that the evaluation of nonlinearities in
the reduced-order model context is computationally expensive. To make this clear, let us assume
we are in a discrete setting and the nonlinear term has the form

ΨT W f (Ψη(t)),

where Ψ = [ψ1 | . . . | ψ` ] ∈ RN ×` is the matrix in which the POD modes are stored columnwise,
W ∈ RN ×N is a weighting matrix related to the utilized inner product in (2.7) and f describes
a nonlinear function. Hence, the treatment of the nonlinearity requires the expansion of the
reduced variable y` (t) = Ψη(t) ∈ RN in the full space, then the nonlinearity can be evaluated
and finally the result is projected back to the reduced space X` . Obviously, this means that the
reduced model is not fully independent of the high-order dimension N and an efficient simulation
cannot be guaranteed. Therefore, it is convenient to seek for hyper-reduction methods, i.e. for a
treatment of the nonlinearity where the model evaluation cost is related to a low dimension. A
possible remedy which is commonly used is given by (discrete) empirical interpolation methods,
for which we refer to [25] for EIM, [50] for DEIM and to [61] for Q-DEIM. Another option is
given by [12] which investigates nonlinear model reduction via dynamic mode decomposition.
Furthermore, in [188] nonlinear model reduction is realized by replacing the nonlinear term by
its interpolation into the finite element space. The treatment of the nonlinearity is done in [18]
by the missing point estimation method and in [141] by the best points interpolation method.

2.5 Phase field systems

The modeling concept of phase field systems is utilized in order to describe phase separation
and the development and evolution of micro structures over time. It is a vivid field of research
in many disciplines with a broad spectrum of applications. Examples of such are material sci-
ences, where it is important to know material properties at specific parameter settings and how
the process of aging or changes in the problem setting influences them, see e.g. [170]. Another
application field appears in image processing, where phase field models are used as tools for the
recovery of shapes from images, see e.g. [44].

In the scope of this thesis, a diffuse interface model of the Cahn–Hilliard system is used as an
example in which spatial adaptivity is fundamental in order to realize numerical implementations.
In this section, we describe the physical phenomenon, introduce the mathematical model and
recall known concepts. This enables us to refer to these concepts later.
22 2 BASIC CONCEPTS

2.5.1 Physical phenomenon and modeling

We are concerned with the Cahn–Hilliard system which is introduced in [43] as a model for phase
separation (spinodal decomposition) in binary systems. For a comprehensive study we refer to
e.g. [64].

Spinodal decomposition

In physical observations (see e.g. [21]) the following phenomenon has been discovered. Assume we
are given a mixture of two partially miscible substances, for example an alloy of two metals (e.g.
nickel and iron). At the initial situation, the temperature is greater than the critical temperature.
Then, the alloy is quenched (i.e. rapid cooling) to a temperature below the critical temperature.
What can be observed is shown in Figure 5 and splits into two stages:

(i) (ii)
−→ −−→

Figure 5: Phase separation (i) and aging (ii)

(i) Phase separation: in a rapid demixing phase the mixture decomposes into the pure mate-
rials and interfaces appear in between.

(ii) Aging: in a slow coarsening phase the length of the interfaces is reduced and patches melt
together to larger droplets. Topology changes appear (e.g. coalescence or droplet collisions).
An Ostwald-ripening type process can be observed.

Modeling approaches

There are three modeling approaches in order to describe the phase separation process:

Atomistic model : the behavior of single atoms or molecules in the mixture is described.
However, a considerable challenge lies in the numerical implementation for realistic scenar-
ios which usually exceeds the scope of usual computing machines due to the large amount
of atoms.

Sharp interface approach: the interface is modeled as a lower dimensional manifold. Thus,
jumps appear between pure material levels, see Figure 6. It is challenging to work with
this modeling approach in case the topology of the phase borders changes.

Diffuse interface approach: a coexistence of molecules from both materials is assumed at

the interfacial regions. Thus, the interface is not explicitly defined, but smeared out. A
smooth (needs to be specified mathematically) transition from one to the other material
phase is considered, see Figure 6. The advantage of this modeling approach is that topology
changes like droplet collision or coalescence can be handled in a natural way. However, in
practical simulations the interface width ε is usually larger than the real world thickness.

The relationship between the sharp interface approach and the diffuse interface approach is
investigated in the asymptotic analysis for vanishing ε → 0. For further study on this topic, we
refer to e.g. [46, 79, 137]. In this thesis, we concentrate our investigations on the diffuse interface
approach.
2.5 Phase field systems 23

1 1

0 0

−1 −1

Figure 6: Interface profile: sharp interface approach (left) versus diffuse interface approach (right)

2.5.2 Mathematical model and aspects

Let us consider a binary mixture comprising two substances A and B. In order to describe
the spatial distribution over time (0, T ] with fixed end time T > 0, a phase field variable ϕ is
introduced which fulfills ϕ(t, x) = +1 in the pure A-phase and ϕ(t, x) = −1 in the pure B-phase.
Values of ϕ between −1 and +1 represent the interfacial area between the two substances. Its
thickness is finite and of order O(ε) with 0 < ε  1. Of course, this setting can be extended
to the multi-component case, see e.g. [24, 34, 140]. In the following, we restrict ourselves to the
case of binary systems for the sake of clarity.

Ginzburg-Landau energy functional

A phenomenological theory which describes the dynamics of the phase separation is given by the
chemical thermodynamics using Gibb’s free energy F, see e.g. [147]. Particular mathematical
choices for F are discussed later (see (2.22),(2.23)). The Gibb’s free energy of a system describes
the difference between the enthalpy and entropy at the current temperature. A system will
approach an isothermal equilibrium state that minimizes F. In order to model capillarity (i.e.
surface energy of the interface which separates the bulk phases), the gradient term |∇ϕ| is
additionally taken into account. The resulting free energy E is given by
ε 1
E(ϕ) = |∇ϕ|2 + F(ϕ). (2.14)
2 ε
The Cahn–Hilliard model in order to describe the equilibrium state of phase separation subject
to mass conservation reads as
Z
min E(ϕ) s.t. ϕ dx = cvol |Ω|,
Ω

where E(·) denotes the Ginzburg-Landau energy functional defined by

Z Z Z
ε 2 1
E(ϕ) = E(ϕ)dx = |∇ϕ| dx + F(ϕ)dx. (2.15)
Ω 2 Ω ε Ω
The first term 2ε |∇ϕ|2 is large for large variations of ϕ and is minimal if ϕ is constant. This gives
rise to interfacial areas. It can be interpreted as a penalization of jumps and as a regularization
term, since it enforces ϕ to be H 1 (Ω). The second term 1ε F(ϕ) is large for ϕ ∈ / {−1, 1} and it is
minimal for ϕ ∈ {−1, 1}. Minimization of this term is responsible for phase separation.

Chemical potential

From a chemical point of view, the chemical potential µ describes how the (Gibb’s) free energy
changes if one of the components of the mixture changes. From a mathematical point of view,
the chemical potential can be regarded as first variation of E with respect to the phase field ϕ,
i.e.
δE 1
µ := = −ε∆ϕ + F 0 (ϕ).
δϕ ε
24 2 BASIC CONCEPTS

Cahn–Hilliard equation

There are different approaches to derive the Cahn–Hilliard equation. One option is to consider
thermodynamical concepts of mass flux and mass balance law which yields to a diffusion equa-
tion, see e.g. [64]. Another derivation can be carried out by computing the H −1 -gradient flow
of the Ginzburg–Landau free energy, see e.g. [33]. A related model is the Allen–Cahn equation,
which can be derived as L2 -gradient flow of the Ginzburg–Landau energy and does not conserve
mass [172].

The Cahn–Hilliard equation is a fourth-order parabolic nonlinear partial differential equation.

In a bounded and open domain Ω ⊂ Rd , d ∈ {1, 2, 3} with Lipschitz continuous boundary ∂Ω
the Cahn–Hilliard equation reads as
1
ϕt + ε∆2 ϕ − ∆F 0 (ϕ) = 0 in (0, T ] × Ω. (2.16)
ε
It is convenient to specify further problem parameters and consider the following model
bσ
ϕt + bσε∆2 ϕ − ∆F 0 (ϕ) = 0 in (0, T ] × Ω. (2.17)
ε
By b > 0 we denote the mobility which we assume to be constant. It is also possible to consider
a degenerate mobility, which is zero in pure phases, see e.g. [23, 66]. The parameter σ > 0
describes the surface tension and is a material dependent quantity.

Utilizing the definition of the chemical potential, we can decouple the Cahn–Hilliard equation
into two coupled partial differential equations of second-order. The coupled Cahn–Hilliard system
reads as

ϕt − b∆µ = 0 in (0, T ] × Ω, (2.18a)

σ
−σε∆ϕ + F 0 (ϕ) = µ in (0, T ] × Ω. (2.18b)
ε
Remark 2.48. For numerical implementations, the coupled form (2.18) of the Cahn–Hilliard
equation is advantageous, since it allows to use finite elements with piecewise continuous, linear
ansatz functions and leads to an approximate solution in H 1 (Ω). It circumvents the construction
of finite element subspaces in H 2 (Ω) which would be needed in the numerical consideration of
(2.16). The construction of such higher-order finite element spaces is more involved, since it
requires the basis functions to be continuously differentiable, i.e. in C 1 (Ω̄). A possible finite ele-
ment could be the HCT (Hsieh–Clough–Tocher, see e.g. [63, Chapter 8.2]) element, for instance,
which has 12 degrees of freedom, or the Agyris element (see e.g. [53, Chapter VII]) with 21
degrees of freedom. For the scope of this thesis, we focus on the numerical treatment of (2.18).

As boundary conditions we consider homogeneous Neumann conditions

∂ϕ ∂µ
=0 and =0 on [0, T ] × ∂Ω. (2.19)
∂~n ∂~n
These boundary conditions are natural and average conserving, i.e. they leave the spatial average
of ϕ invariant in time. Moreover, we impose an initial condition for the phase field

ϕ(0, ·) = ϕ0 in Ω. (2.20)
2.5 Phase field systems 25

Bulk free energies

We note that up to now we have assumed that F is differentiable. However, different choices for
the bulk free energy density F are possible. The free energy function F shall satisfy the following
properties. It is a symmetric function which should be minimal (i.e. F ≡ 0) at ϕ = ±1, and
F > 0 for ϕ 6= ±1. Moreover, it should hold true that lim F(ϕ) = +∞ for |ϕ| → ∞. It turns
out that we can split the free energies into a convex (F+ ) and a concave (F− ) part as follows

F(ϕ) = F+ (ϕ) + F− (ϕ). (2.21)

In particular, we consider three choices for the free energy F, see Figure 7.

0 0 0
-1 0 1 -1 0 1 -1 0 1

Figure 7: Free energies: double-well Fpol (left), double-obstacle F∞ (middle), Moreau–Yosida

relaxation Fs,r (right)

The smooth polynomial free energy Fpol (see e.g. [67]) is defined as
1 1 1 1
Fpol (ϕ) = (1 − ϕ2 )2 = ϕ4 + ( − ϕ2 ) . (2.22)
4 4
|{z} | 4 {z2 }
F+ F−

Another choice is a relaxed double-obstacle free energy

1 s
Fs,r (ϕ) = (1 − ϕ2 ) + (| max(0, ϕ − 1)|r + | min(0, ϕ + 1)|r ), (2.23)
|2 {z } |r {z }
F− F+

with relaxation parameter s  0 and degree r. The energy Fs,r is a smooth approximation of
the non-smooth double-obstacle free energy [35, 143]
(
1
(1 − ϕ2 ), if ϕ ∈ [−1, 1],
F∞ (ϕ) = 2
+∞, else,

using a Moreau–Yosida relaxation proposed in [96]. We note that only the double-obstacle free
energy F∞ enforces |ϕ| ≤ 1. Both polynomial and relaxed double-obstacle free energies allow ϕ
to have unphysical values |ϕ| > 1.

Properties of the Cahn–Hilliard system

We summarize some characteristic properties of the solution ϕ to the Cahn–Hilliard system.

Mass conservation: the total mass is preserved, i.e.

Z
d
ϕ(x)dx = 0.
dt Ω
Z
By redefining the pure phases we can assume ϕ(x) dx = 0 w.l.o.g.
Ω
26 2 BASIC CONCEPTS

Energy dissipation: it holds true

Z Z 
d σ 0 
σE(ϕ(t))dx = F (ϕ(t))ϕt (t) + σε∇ϕ(t)∇ϕt (t) dx = (µ(t), ϕt (t))L2 (Ω)
dt Ω Ω ε

= − b(∇µ(t), ∇µ(t))L2 (Ω) = −bk∇µ(t)k2L2 (Ω) ≤ 0,

i.e. the energy decreases in time.

Well-posedness: Existence and uniqueness analysis for different choices for the free energy
F is carried out in e.g. [1, 35, 36, 64, 66, 67]. For example, in the case of the double-well free
energy Fpol , global existence of a weak solution to (2.18) with ϕ ∈ L∞ (0, T ; H 1 (Ω)), ϕt ∈
L2 (0, T ; H 1 (Ω)∗ ), µ ∈ L2 (0, T ; H 1 (Ω)) is shown in [64], where also higher regularity results
are derived.

Weak form of the Cahn–Hilliard system

Let an initial phase field ϕ0 ∈ H 1 (Ω) be given with E(ϕ0 ) < ∞. The variational form of
the coupled Cahn–Hilliard system (2.18) reads as: find a phase field ϕ ∈ W (0, T ; H 1 (Ω)) with
ϕ(0, ·) = ϕ0 and a chemical potential µ ∈ L2 (0, T ; H 1 (Ω)) such that

d
(ϕ(t), v)L2 (Ω) + b(∇µ(t), ∇v)L2 (Ω) = 0 ∀v ∈ H 1 (Ω), (2.25a)
dt
σ
σε(∇ϕ(t), ∇v)L2 (Ω) + (F 0 (ϕ(t)), v)L2 (Ω) = (µ(t), v)L2 (Ω) ∀v ∈ H 1 (Ω), (2.25b)
ε
holds for almost all t ∈ (0, T ).

Time discretization

In order to find a numerical solution to (2.25), we first introduce a time discretization which
allows to use a different (adaptive) finite element space at each time instance for the spatial dis-
cretization. As discussed in [70, 71], explicit numerical schemes and most semi-implicit schemes
have restrictions on the time steps and force to take time step sizes which are (too) small. Only
a few of the fully implicit solvers are gradient stable in the sense that they have a discrete energy
that decreases from one time level to the next [108]. A comparison of different time discretization
schemes is given in [90].
We consider an unconditional gradient stable time discretization scheme which is based on a
convex-concave splitting of the free energy according to [68, 70, 71]. The basic idea is to utilize
(2.21) and treat the convex part implicitly and the concave part explicitly, i.e.
0 0
F+ (ϕj ) + F− (ϕj−1 ). (2.26)

We introduce a time grid 0 = t0 < t1 < · · · < tn = T with n ∈ N and time step sizes ∆tj =
tj − tj−1 for j = 1, . . . , n. The time-discrete scheme reads as: for given ϕ0 = ϕ0 ∈ H 1 (Ω) find
sequences ϕ1 , . . . , ϕn ∈ H 1 (Ω) and µ1 , . . . , µn ∈ H 1 (Ω) which fulfill the system
 j
ϕ − ϕj−1

,v + b(∇µj , ∇v)L2 (Ω) = 0 ∀v ∈ H 1 (Ω), (2.27a)
∆tj 2
L (Ω)
σ 0 j 0
σε(∇ϕj , ∇v)L2 (Ω) + (F+ (ϕ ) + F− (ϕj−1 ), v)L2 (Ω) = (µj , v)L2 (Ω) ∀v ∈ H 1 (Ω). (2.27b)
ε
2.5 Phase field systems 27

Space discretization

As mentioned in Remark 2.48, we utilize the coupled form of the Cahn–Hilliard equation in order
to use continuous, piecewise linear finite elements for the spatial discretization. In particular, we
use mixed finite elements with P1 − P1 nodal basis functions. For more details, we refer to e.g.
[65]. We utilize an adaptive discretization scheme following Algorithm 1 and the implementation
is based on [114]. Let us denote by Vh0 the finite element space associated with the initial
triangulation Thinit and ϕ0h is a suitable approximation of ϕ0 in Vh0 . For example, ϕ0h can be
computed as solution to
(ϕ0h , v)L2 (Ω) = (ϕ0 , v)L2 (Ω) ∀v ∈ Vh0 .
The fully discrete Cahn–Hilliard system reads as: for given ϕ0h ∈ Vh0 find sequences ϕ1h ∈
Vh1 , . . . , ϕnh ∈ Vhn and µ1h ∈ Vh1 , . . . , µnh ∈ Vhn which fulfill the system
!
ϕjh − ϕj−1
h
,v + b(∇µjh , ∇v)L2 (Ω) = 0 ∀v ∈ Vhj , (2.28a)
∆t 2
L (Ω)
σ 0 j
σε(∇ϕjh , ∇v)L2 (Ω) + (F+ 0
(ϕh ) + F− (ϕj−1 j
h ), v)L2 (Ω) = (µh , v)L2 (Ω) ∀v ∈ Vhj , (2.28b)
ε

where Vhj denotes the (adapted) finite element space for j = 1, . . . , n.

28 3 ADAPTIVITY WITH RESPECT TO TIME

3 Adaptivity with respect to time

Challenge. In order to construct a POD reduced-order model for an optimal control problem
which reflects the temporal dynamics of the true optimal solution well, a necessary prerequisite
is the representation of the important dynamical features within the snapshot ensemble. A naive
solution approach is to use a very fine temporal discretization for snapshot generation in order
to identify the key properties in the dynamics. However, this leads to very large computational
times, requires storage of data for many time instances and will probably contain redundant
information. Moreover, the snapshots are computed with respect to a reference control (e.g. zero
control), which might differ strongly from the optimal control. As a consequence, the resulting
dynamics of the snapshots might not display a correct temporal evolution with respect to the
true solution.

Solution idea. We provide an adaptive snapshot location method for POD reduced-order model-
ing of linear-quadratic optimal control problems. In order to compute an approximate solution
to the infinite-dimensional optimal control problem, we utilize a POD surrogate model which is
built from snapshots sampled at specifically chosen time instances. The selection of the snapshot
locations is determined by a residual based a-posteriori error estimation for a biharmonic equa-
tion which is equivalent to the optimality system. In this way, suitable time instances related
to the optimal solution are selected such that the dynamical evolution over time is captured
sufficiently well. At the same time an approximation of the optimal control is computed which
serves as input control for snapshot generation. In order to achieve fast offline computational
times, we utilize a coarse spatial discretization for the solution of the biharmonic equation.

Solution procedure. We reformulate the first-order optimality system of a linear-quadratic optimal

control problem as an elliptic equation which is second-order in time and fourth-order in space.
This reformulation can be done either with respect to the state or the adjoint state. As a
consequence the resulting biharmonic equation depends only on either of these variables, while
all other variables are eliminated from the system. For the resulting elliptic equation, standard
residual based a-posteriori error estimates can be employed in the spirit of [4] which has been
done in [82] for a specific optimal control problem of a linear heat equation. According to the
error indicator, an adaptive time grid is constructed which is appropriate with respect to the
optimal state or the optimal adjoint state depending on whether the elliptic equation depends
on the state or adjoint state.
The novelty of the approach in comparison to existing literature concerned with the reformulation
of the optimality system into a biharmonic equation consists in considering a specific control
structure (which we refer to as located control) and deriving a-posteriori error estimates in the
case of a control constrained optimization problem. In this situation, a reformulation of the
optimality system can only be made with respect to the adjoint variable.
In the framework of POD model order reduction, we exploit the construction of the adaptive time
grid in the offline phase in order to generate suitable snapshot locations. In particular, this time
grid is used to sample the snapshots and to compute the POD reduced-order optimal solution. In
order to enable a fast offline computation time, the adaptive time grid is computed with respect
to a coarse spatial resolution of the elliptic system. This is possible due to a heuristic observation
in our numerical test examples where spatial and temporal discretization decouple for the given
problem setting.
Another important aspect of the proposed methodology is that the solution of the elliptic system
delivers an approximation of the optimal control with respect to a coarse spatial resolution. This
control can be used as input control for snapshot generation.
We provide an error analysis for the proposed methodology which allows to estimate the error
between the true optimal solution and the solution to the POD reduced-order control problem.
3.1 Literature overview 29

Outline of the chapter

In Section 3.1 we give an overview of the existing literature concerned with snapshot location
strategies in POD reduced-order modeling for dynamical systems. We introduce the linear-
quadratic optimal control problem setting in Section 3.2 and reformulate the associated opti-
mality system as an elliptic partial differential equation in Section 3.3. We recall in Sections 3.4
and 3.5 results of [82] regarding the a-posteriori error estimation for the elliptic equation and
its discretization using a space-time finite element method. The snapshot location strategy for
POD model order reduction in optimal control is presented in Section 3.6. This section includes
the description and discussion of the proposed methodology, an error analysis for the state, ad-
joint state and control variable as well as a numerical example. Section 3.7 focuses on snapshot
location in optimal control in the case of located control and considers the inclusion of control
constraints. A reformulation into a biharmonic equation is investigated and a residual based
a-posteriori error estimation with respect to time is derived for the adjoint state. In Section 3.8
we propose a snapshot location strategy for this case and provide an error analysis for the POD
reduced-order solution. The section finishes with a presentation of numerical results. The pro-
posed snapshot location strategy leaves space for further directions of research for which we
sketch ideas in Section 3.9 for the consideration of a nonlinear state equation and the application
to the model predictive control context.

We note that this section is based on a collaboration with Alessandro Alla and Michael Hinze
and some of the results have been published in [9, 10, 11].

3.1 Literature overview

The following publications consider snapshot location in POD model order reduction, presented
in a chronological order.
In [118] a medical application of a hyperthermia process for the heating of tumors is considered.
Three different sampling strategies of snapshots for POD model order reduction are compared:
a uniform time sampling, a logarithmic distribution with early time sampling and a reversion
thereof leading to a late time sampling. Another snapshot sampling scheme is given in [163] in
the context of a groundwater flow model. Since the considered model reaches the steady state in
an exponential manner, an exponential function is utilized which locates many snapshots early
in time and less snapshots later in time.
In [123], POD model reduction for dynamical systems is considered. The choice of time instances
for snapshot generation is optimized in order to minimize the error between the solution trajec-
tory and the solution to the POD reduced problem. Starting with an initial uniform snapshot
grid, additional snapshots are placed in an optimal way. This leads to a minimization problem
subject to nonlinear equality constraints and inequality constraints. Existence of a solution as
well as first- and second-order optimality conditions are derived. A continuation of this approach
is given in [127], where a complex-valued Helmholtz equation on a frequency band is considered,
i.e. a setting of a nonlinear parameter-dependent elliptic system is studied. The optimal snap-
shot location of [123] is utilized in order to select additional parameters to an initial uniformly
distributed parameter set in an optimal way.
In [8], the location of the snapshots is assumed to be fixed. However, a time-adaptive procedure
is carried out in order to cluster the snapshots into snapshot sets on sub-intervals according to
the ratio of considered singular values versus the sum of all singular values. This leads to the
computation of a time local POD basis for each sub-interval.
A different approach is proposed in [106] for POD model order reduction of linear and semilinear
parabolic equations, where the snapshot locations are selected according to an error equidistri-
bution concept. The global discretization error is estimated by the difference of the solution to
an implicit Euler scheme and an implicit trapezoidal rule. For a fixed number of time instances,
30 3 ADAPTIVITY WITH RESPECT TO TIME

the snapshot locations are distributed such that the error indicator is approximately the same
in each partition. It is shown that the error equilibration of the full-order model is inherited by
the reduced-order model.
An online adaptive snapshot selection strategy called proper snapshot selection (PSS) is proposed
in [142]. The most significant snapshots are selected in an incremental procedure by measuring
and comparing the energy contained in the respective snapshots using the decay of the eigenval-
ues as indicator.
In [144], an on-the-fly adaptive snapshot selection method is introduced. Within the solution of
the full-order model, an incremental POD basis is computed (see e.g. [38, 73]). In particular, the
newly selected snapshots are added in order to update the singular value decomposition. The
choice of the snapshot locations is controlled by an error criterion which is based on a prediction
of the increase in the error of the reduced-order approximation.

To the best of the author’s knowledge, the publications [9, 10, 11] are the first contributions
concerning snapshot location for POD model order reduction in optimal control.

Apart from the computation of suitable time instances for snapshot generation, the method
proposed in this chapter also delivers an approximation of the optimal control which can be
used as input control for snapshot sampling. Since the dynamics of the state and adjoint state
computed from a reference control might differ strongly from the optimal trajectories, a proper
choice of the input control is crucial for the accuracy of the reduced-order model. Another option
is to update the POD model within the course of the optimization. Concerning this issue, we
like to mention the following references. In [3], an adaptive POD model (APOD) is proposed,
where new snapshots are computed corresponding to the current control iterate within each
optimization level. Thus, the POD model is updated according to the dynamics in the current
optimization stage.
Within the framework of trust-region POD [17, 162], the necessity of a POD basis update is
determined by the Carter condition which guarantees relative gradient accuracy.
In [122], the so-called optimality system POD (OSPOD) is proposed. The basic idea is to include
the equations which determine the POD basis into the optimization process. This makes the
optimal control problem more complicated, and thus numerically more intensive. However, it
turns out that the computation of a few (projected) gradient steps is sufficient for an initialization.
In [83], we consider two POD basis update strategies: a trust-region POD framework is compared
to an OSPOD approach combined with an inexact SQP method with a-posteriori error control
according to [177].

3.2 Optimal control problem with distributed control

First, we consider a standard linear-quadratic optimal control problem. This setting is used
to explain the reformulation of the associated first-order optimality system into a biharmonic
system and the snapshot location strategy in optimal control for POD model order reduction.
The inclusion of located control and control constraints is considered in Section 3.7. In addi-
tion, the applicability of the proposed method to other problem settings is outlined in Section 3.9.

The state equation

As governing equation (state equation) we consider a linear heat equation given by

yt − ∆y = u in (0, T ] × Ω, (3.1a)
y=0 on [0, T ] × ∂Ω, (3.1b)
y(0, ·) = y0 in Ω. (3.1c)
3.2 Optimal control problem with distributed control 31

The function y is in the following called state. We denote by T > 0 a fixed end time and
Ω ⊂ Rd , d ∈ {1, 2, 3}, is an open bounded domain with Lipschitz boundary ∂Ω (Definition 2.17).
The initial value is denoted by y0 . The source function u will in the following act as the control.
It enters the system (3.1) as distributed control on the right-hand side. We assume that we can
control within the whole spatial domain. A different kind of control which can only act on parts
of the domain (referred to as located control) as well as the inclusion of control constraints are
investigated in Section 3.7. As boundary condition we consider homogeneous Dirichlet conditions.
Note that we regard a consistent setting in the sense that the initial function complies with the
boundary condition. For the analysis and numerical investigation we consider the weak form
of (3.1) in the following. For this we make the regularity assumption for the initial condition
y0 ∈ L2 (Ω) and for the control function we assume u ∈ L2 (0,RT ; Ω). For a given u ∈ L2 (0, T ; Ω)
we define the functional û : [0, T ] → H −1 (Ω) as v 7→ û(t)v = Ω u(t, x)v(x)dx for v ∈ H01 (Ω). In
this way, we obtain an element û ∈ L2 (0, T ; H −1 (Ω)).
Definition 3.1 (Weak state solution). A function y ∈ W (0, T ; H01 (Ω)) is called weak solution to
(3.1) if y satisfies the initial condition (3.1c) and the variational equation
hyt (t), viH −1 (Ω),H01 (Ω) + (∇y(t), ∇v)L2 (Ω) = hû(t), viH −1 (Ω),H01 (Ω) (3.2)

for all v ∈ H01 (Ω) and almost everywhere in [0, T ].

Existence of a unique weak solution y ∈ W (0, T ; H01 (Ω)) to the linear heat equation (3.1) is
proved in e.g. [69, §7.1.2, Theorems 3 and 4]. Due to
W (0, T ; H01 (Ω)) ,→ C([0, T ]; L2 (Ω)), (3.3)
see Theorem 2.16, we have y ∈ C([0, T ]; L2 (Ω)) and therefore equality (3.1c) has a precise
meaning. The assignment of boundary values is considered in the sense of traces according to
Theorem 2.19.
Applying the differential form of Gronwall’s inequality (Lemma 2.30), one can derive the energy
estimate


max ky(t)kL2 (Ω) +kykL2 (0,T ;H01 (Ω)) +kyt kL2 (0,T ;H −1 (Ω)) ≤ c kûkL2 (0,T ;H −1 (Ω)) + ky0 kL2 (Ω) (3.4)
0≤t≤T

with a constant c > 0 which is independent of û and y0 , see e.g. [69, §7.1.2b]. Thus, equation
(3.1) is well-posed in the sense of Hadamard (parfaitement bien posé), see [91]. The operator
S : L2 (0, T ; H −1 (Ω)) × L2 (Ω) → W (0, T ; H01 (Ω)), (û, y0 ) 7→ y := S(û, y0 ) (3.5)
defines a weak solution operator associated with equation (3.1).
Note that due to the continuity of the functional û, we have
Z T Z T
kûk2L2 (0,T ;H −1 (Ω)) = kû(t)k2H −1 (Ω) dt ≤ ku(t)k2L2 (Ω) dt = kuk2L2 (0,T ;Ω) .
0 0

For the subsequent sections, we need the following higher regularity results, which are taken
from [69] and adapted to our notation.
Remark 3.2. (Improved regularity, [69, §7.1.3, Theorem 5]). Assume y0 ∈ H01 (Ω) and u ∈
L2 (0, T ; Ω). Then, the weak solution y to (3.1) fulfills
y ∈ L2 (0, T ; H 2 (Ω)) ∩ L∞ (0, T ; H01 (Ω)) with yt ∈ L2 (0, T ; Ω) (3.6)
and the energy estimate
 
ess sup ky(t)kH01 (Ω) + kykL2 (0,T ;H 2 (Ω)) + kyt kL2 (0,T ;Ω) ≤ c kukL2 (0,T ;Ω) + ky0 kH01 (Ω)
0≤t≤T

holds with a constant c > 0 which is independent of u and y0 .

32 3 ADAPTIVITY WITH RESPECT TO TIME

d k
Remark 3.3. (Higher regularity, [69, §7.1.3, Theorem 6]). Assume y0 ∈ H 2m+1 (Ω), dt ku ∈

L2 (0, T ; H 2m−2k (Ω)), k = 0, . . . , m. Suppose that the following mth -order compatibility conditions
hold:
dm−1
g0 := y0 ∈ H01 (Ω), g1 := u(0) − ∆g0 ∈ H01 (Ω), . . . , gm := u(0) − ∆gm−1 ∈ H01 (Ω).
dtm−1
Then, the weak solution y to (3.1) fulfills
dk
y ∈ L2 (0, T ; H 2m+2−2k (Ω)), k = 0, . . . , m + 1; (3.7)
dtk
and we have the estimate
m+1 m
!
X dk X dk
y ≤c u + ky0 kH 2m+1 (Ω) ,
dtk L2 (0,T ;H 2m+2−2k (Ω)) dtk L2 (0,T ;H 2m−2k (Ω))
k=0 k=0

the constant c > 0 independent of u and y0 .

The cost functional

The cost functional (objective functional) which shall be minimized is given by

1 α
J(y, u) := ky − yd k2L2 (0,T ;Ω) + kuk2L2 (0,T ;Ω) , (3.8)
2 2
where yd ∈ L2 (0, T ; Ω) is a given desired state and α > 0 denotes the regularization parameter.
Note that the cost functional is non-negative and quadratic. In particular, (y, u) 7→ J(y, u) is
continuous, (strictly) convex and hence weakly lower semicontinuous according to Theorem 2.39.
Quadratic cost functionals like (3.8) are known as tracking type costs.

The optimal control problem

The optimal control problem which we consider in the following is formulated as

min J(y, u) s.t. y = S(u, y0 ). (3.9)
(y,u)∈W (0,T ;H01 (Ω))×L2 (0,T ;Ω)

This is a linear-quadratic optimal control problem. The optimization goal is to steer the state
y, which fulfills the linear heat equation (3.1) in a weak sense, as close as possible to the given
desired state yd and at the same time keep the control costs minimal. The optimal control prob-
lem (3.9) has a unique solution, see e.g. [131, §III, 2.1, (2.10)]. The proof follows the usual steps
according to Remark 2.40. For the following, we proceed according to the first-optimize-then-
discretize approach as explained in Remark 2.44(i).

The optimality system

Following the usual Lagrange calculus as recalled in Section 2.3, we derive the first-order necessary
(and by convexity sufficient) optimality conditions associated with problem (3.9). They are given
by the state equation (3.1) together with the adjoint equation
−pt − ∆p = y − yd in [0, T ) × Ω, (3.10a)
p=0 on [0, T ] × ∂Ω, (3.10b)
p(T, ·) = 0 in Ω, (3.10c)
and the optimality condition (sometimes called gradient equation)
αu + p = 0 in [0, T ] × Ω. (3.11)
3.2 Optimal control problem with distributed control 33

Definition 3.4 (Weak adjoint solution). A function p ∈ W (0, T ; H01 (Ω)) is called weak solution
to (3.10) if p satisfies the final condition (3.10c) and the variational equation

− hpt (t), viH −1 (Ω),H01 (Ω) + (∇p(t), ∇v)L2 (Ω) = hr̂(t), viH −1 (Ω),H01 (Ω) (3.12)

for all v ∈ H01 (Ω) and Ralmost everywhere in [0, T ], where r̂ : [0, T ] → H −1 (Ω) is a functional
defined by v 7→ r̂(t)v = Ω (y(t, x) − yd (t, x))v(x)dx for v ∈ H01 (Ω).

While the state equation (3.1) is an initial boundary value problem and hence a forward
in time equation, the adjoint equation (3.10) runs backward in time. Utilizing a simple time
transformation with τ ∈ [0, T ] by p̃(τ, ·) := p(T − τ, ·), ỹ(τ, ·) := y(T − τ, ·) and ỹd (τ, ·) :=
yd (T − τ, ·), the following initial boundary value problem can be derived

p̃t − ∆p̃ = ỹ − ỹdτ in (0, T ] × Ω, (3.13a)

p̃ = 0 on [0, T ] × ∂Ω, (3.13b)
p̃(0, ·) = 0 in Ω. (3.13c)

The existence of a unique weak solution p ∈ W (0, T ; H01 (Ω)) to (3.13) as well as stability and
regularity results now follow analogue to the results for the state. Back transformation concludes
the argumentation. In particular, we have the following estimation


max kp(t)kL2 (Ω) +kpkL2 (0,T ;H01 (Ω)) +kpt kL2 (0,T ;H −1 (Ω)) ≤ c kukL2 (0,T ;Ω) + ky0 kL2 (Ω) + kyd kL2 (0,T ;Ω)
0≤t≤T
(3.14)
with a constant c > 0 which is independent of u, y0 and yd .

Fully discretized optimality system

For the temporal discretization, we use a fully implicit Euler scheme. To this end, we introduce
a time grid 0 = t0 < t1 < · · · < tn = T with n ∈ N and time step sizes ∆tj = tj − tj−1
for j = 1, . . . , n. We integrate the differential equations (3.2) and (3.12) from tj−1 to tj for
j = 1, . . . , n and approximate the integrals involving the state and adjoint variables by a right-
sided rectangle method. For the spatial discretization we introduce a partitioning Th of the
spatial domain Ω into regular disjoint simplices T such that Ω̄ = ∪T ∈Th T̄ holds true. We use
continuous, piecewise linear finite elements, i.e. we introduce the finite element space Vh as

Vh = {v ∈ C 0 (Ω̄) : v|T ∈ P1 (T ), T ∈ Th } ⊂ H 1 (Ω)

and define Vh,0 = Vh ∩ H01 (Ω). Then, the fully discretized state equation reads as follows: find
sequences yh0 , . . . , yhn ∈ Vh such that
!
yhj − yhj−1
Z tj
j 1
,v + (∇yh , ∇v)L2 (Ω) = (u(t), v)L2 (Ω) dt (3.15)
∆tj 2
∆tj tj−1
L (Ω)

and
(yh0 , v)L2 (Ω) = (y0 , v)L2 (Ω)
holds for all v ∈ Vh . Note that we do not discretize the control, but we make use of variational
discretization according to [98], see Remark 2.44(ii). The fully discretized adjoint equation reads
as: for given pnh = 0 find sequences pn−1 0
h , . . . , ph ∈ Vh such that
!
pj−1 − pjh tj
Z
1
h
,v + (∇pjh , ∇v)L2 (Ω) = (yhj−1 , v)L2 (Ω) − (yd (t), v)L2 (Ω) dt (3.16)
∆tj ∆tj tj−1
L2 (Ω)
34 3 ADAPTIVITY WITH RESPECT TO TIME

for all v ∈ Vh . Note that the time integral on the right-hand side can also be approximated by
the same rectangle method leading to the approximation
Z tj
1
(yd (t), v)L2 (Ω) dt ≈ (yd (tj ), v)L2 (Ω) .
∆tj tj−1

Of course, different integration schemes are possible. However, since yd is a given function, we
do not approximate the time integral at this stage in order to avoid the approximation error.

3.3 Reformulation of the optimality system

The optimality system (3.1)-(3.10)-(3.11) can be rewritten into a single elliptic boundary value
problem in the space-time domain. This biharmonic equation can be derived with respect to the
state variable y and with respect to the adjoint variable p, respectively.

Let us mention some publications which have considered different aspects related to this. In
[42] a linear-quadratic optimal control problem is transformed into an elliptic system. Numerical
experiments are presented in [138] for distributed and boundary control and the tests consider
1d- and 2d-spatial domains. Subsequent to this work, in [139] control constraints are added to
the optimal control problem leading to an optimality system with a pointwise projection for-
mula. The reformulation into a biharmonic system is derived. Due to the projection operator,
the elliptic system is non-differentiable. For this reason, the projection formula is regularized by
smoothed min/max-functions and the convergence of the associated solution for vanishing regu-
larization parameter is shown. In [193], the optimal control of a Burgers equation is considered.
Using the nonlinear Cole–Hopf transformation, the resulting optimality system of linear parabolic
equations is transformed into a biharmonic equation. Equivalence of the systems is proved and
regularization in case of additional control constraints is applied. In the related publication [115]
the Burgers equation is linearized and the associated optimality system is transformed into an
elliptic equation in space-time. Of particular importance for this chapter is [82]. In this work,
a-priori and a-posteriori error analysis is developed for the time discretization of the state and
adjoint biharmonic equation. The a-posteriori error estimate is residual based and gives rise to
an adaptive scheme which is related to the optimal state and adjoint state, respectively. This
time-adaptive scheme will be the key feature in the following snapshot location selection method
for POD in optimal control.
We proceed as follows: first, we state the biharmonic system in strong and weak form for the
state and the adjoint state, respectively. Well-posedness of the resulting elliptic systems is dis-
cussed. Finally, we note in which sense the biharmonic system is equivalent to the optimality
system. The presentation and notation is leaned upon [82].

We introduce the Hilbert space

H 2,1 (0, T ; Ω) = L2 (0, T ; H 2 (Ω) ∩ H01 (Ω)) ∩ H 1 (0, T ; L2 (Ω)),

equipped with inner product

T m 
∂2v ∂2w
Z Z X 
(v, w)H 2,1 (0,T ;Ω) = vw + vt wt + ∇v∇w + dxdt
0 Ω ∂xi ∂xj ∂xi ∂xj
i,j=1

and natural norm given by

 1/2
d 2
X ∂2v
kvkH 2,1 (0,T ;Ω) = kvk2L2 (0,T ;Ω) + kvt k2L2 (0,T ;Ω) + k∇vk2L2 (0,T ;Ω) +  .
∂xi ∂xj L2 (0,T ;Ω)
i,j=1
3.3 Reformulation of the optimality system 35

For the following, we need improved regularity results for the state and adjoint state. If
y0 ∈ H01 (Ω) and u ∈ L2 (0, T ; Ω) then we have by Remark 3.2 that the state variable ful-
fills y ∈ H 2,1 (0, T ; Ω). Same applies for the adjoint variable, i.e. it holds p ∈ H 2,1 (0, T ; Ω) if
yd ∈ L2 (0, T ; Ω).

Reformulation of the optimality system with respect to the state

Let us consider the following biharmonic equation in y in strong form

1 1
−ytt + ∆2 y + y = yd in (0, T ) × Ω, (3.17a)
α α
y=0 on [0, T ] × ∂Ω, (3.17b)
∆y = 0 on [0, T ] × ∂Ω, (3.17c)
(yt − ∆y)(T, ·) = 0 in Ω, (3.17d)
y(0, ·) = y0 in Ω. (3.17e)

We homogenize (3.17). For this, let the function f˜ ∈ H 2,1 (0, T ; Ω) fulfill the boundary conditions
(3.17b) and (3.17c) as well as the end time condition (3.17d) and the initial condition (3.17e).
We introduce the space

H02,1 (0, T ; Ω) = {v ∈ H 2,1 (0, T ; Ω) : v(0, ·) = 0 in Ω}

and set ỹ := y − f˜. Then, ỹ ∈ H02,1 (0, T ; Ω) is the solution to the biharmonic equation
1 1 1
−ỹtt + ∆2 ỹ + ỹ = yd − ∆2 f˜ + f˜tt − f˜ in (0, T ) × Ω, (3.18a)
α α α
ỹ = 0 on [0, T ] × ∂Ω, (3.18b)
∆ỹ = 0 on [0, T ] × ∂Ω, (3.18c)
(ỹt − ∆ỹ)(T, ·) = 0 in Ω, (3.18d)
ỹ(0, ·) = 0 in Ω. (3.18e)

We define a symmetric bilinear form AT by

AT : H02,1 (0, T ; Ω) × H02,1 (0, T ; Ω) → R,

Z TZ   Z
1
AT (ỹ, v) = ỹt vt + ỹv + ∆ỹ∆v dxdt + ∇ỹ(T )∇v(T )dx.
0 Ω α Ω

Furthermore, let us introduce a linear form L̃T by

L̃T : H02,1 (0, T ; Ω) → R,

Z T Z Z TZ Z
1 1˜ ˜ ˜
L̃T (v) = yd v dxdt + (− f v − ft vt − ∆f ∆v)dxdt − ∇f˜(T )∇v(T )dx.
0 Ω α 0 Ω α Ω
Then, the weak form of the homogenized biharmonic equation (3.18) is given as

AT (ỹ, v) = L̃T (v) ∀v ∈ H02,1 (0, T ; Ω). (3.19)

Existence of a unique solution ỹ ∈ H02,1 (0, T ; Ω) to (3.19) follows from the Lax Milgram Theo-
rem 2.21. For this, it is shown in [82, Theorem 2.6] that AT : H02,1 (0, T ; Ω) × H02,1 (0, T ; Ω) → R
is a bounded, coercive bilinear mapping. Note that L̃T is linear and bounded. Then, y =
ỹ + f˜ ∈ H 2,1 (0, T ; Ω) is the unique weak solution to (3.17). In particular, y ∈ H 2,1 (0, T ; Ω) with
y(0) = y0 ∈ H01 (Ω) is a weak solution to (3.17), if y satisfies

AT (y, v) = LT (v) ∀v ∈ H02,1 (0, T ; Ω), (3.20)

36 3 ADAPTIVITY WITH RESPECT TO TIME

where LT denotes the linear form defined by

Z T Z
1
LT : H02,1 (0, T ; Ω) → R, LT (v) = yd v dxdt.
0 Ω α

Note that it is also possible to homogenize (3.1). Then, a reformulation of the associated opti-
mality system leads automatically to a homogenized biharmonic equation.

Reformulation of the optimality system with respect to the adjoint state

Let us consider the following biharmonic equation in p in strong form

1
−ptt + ∆2 p + p = − (yd )t + ∆yd in (0, T ) × Ω, (3.21a)
α
p=0 on [0, T ] × ∂Ω, (3.21b)
∆p = yd on [0, T ] × ∂Ω, (3.21c)
(pt + ∆p)(0, ·) = yd (0, ·) − y0 in Ω, (3.21d)
p(T, ·) = 0 in Ω. (3.21e)

We introduce the space H̃02,1 (0, T ; Ω) = {v ∈ H 2,1 (0, T ; Ω) : v(T, ·) = 0 in Ω} and define a
symmetric bilinear form A0 by

A0 : H̃02,1 (0, T ; Ω) × H̃02,1 (0, T ; Ω) → R,

Z T Z   Z
1
A0 (p, v) = pt vt + pv + ∆p∆v dxdt + ∇p(0)∇v(0)dx.
0 Ω α Ω

Furthermore, let us introduce a linear form L0 by

L0 : H̃02,1 (0, T ; Ω) → R,
Z T Z Z Z T Z
(3.22)
L0 (v) = (−(yd )t + ∆yd )v dxdt − (yd (0) − y0 )v(0)dx + yd ∇v · ~n dsdt.
0 Ω Ω 0 ∂Ω

The weak formulation for (3.21) is given by

A0 (p, v) = L0 (v) ∀v ∈ H̃02,1 (0, T ; Ω). (3.23)

If yd ∈ W (0, T ; H 1 (Ω)) the existence of a unique weak solution p ∈ H̃02,1 (0, T ; Ω) to (3.21) fol-
lows from the Lax Milgram Theorem 2.21. For this, it is shown in [82, Theorem 2.6] that A0 :
H̃02,1 (0, T ; Ω)×H̃02,1 (0, T ; Ω) → R is a bounded coercive bilinear mapping and L0 : H̃02,1 (0, T ; Ω) →
R is linear and bounded.

Equivalence of the optimality system to a biharmonic equation

Theorem 3.5. ([82, Theorem 2.7]). Let (y, u) ∈ W (0, T ; H01 (Ω))×L2 (0, T ; Ω) denote the solution
to problem (3.9) with associated adjoint state p ∈ W (0, T ; H01 (Ω)). Assume that y0 ∈ H01 (Ω) ∩
H 3 (Ω), yd ∈ L2 (0, T ; Ω) and u ∈ L2 (0, T ; H 2 (Ω)) ∩ H 1 (0, T ; L2 (Ω)). Further, let the compatibility
assumption u(0) + ∆y0 ∈ H01 (Ω) hold. Then y satisfies (3.17) a.e. in space-time and is a weak
solution to (3.17).
If yd ∈ H 1 (0, T ; L2 (Ω)) ∩ L2 (0, T ; H 2 (Ω)) and the compatibility condition y(T ) − yd (T ) + ∆p(T ) ∈
H01 (Ω) is fulfilled, then p solves (3.21) a.e. in space-time and is a weak solution to (3.21).
3.4 A-posteriori error estimates for the time discretization 37

The proof uses a formal differentiation and insertion of the equations of the optimality sys-
tem (3.1)-(3.10)-(3.11). This is possible due to the higher regularity assumptions according
to Remark 3.3 with m = 1. In particular, the state y ∈ L2 (0, T ; H 4 (Ω)) ∩ H 2 (0, T ; L2 (Ω)) ∩
H 1 (0, T ; H 2 (Ω)) fulfills the biharmonic problem (3.17) a.e. in space-time. Then, y ∈ H 2,1 (0, T ; Ω)
is a weak solution to (3.17). Analogous arguments apply for the adjoint state p. For more details,
we refer to [82].

Since equation (3.17) has a unique weak solution and the weak solution y to (3.9) is a weak
solution to (3.17) according to Theorem 3.5, the other direction holds, i.e. the weak solution to
(3.17) is the optimal state y. The same argumentation holds for the adjoint state p.

We note that the biharmonic equation (3.17) is an equation only with respect to the state vari-
able y with given right-hand side data yd and y0 . The equation does not explicitly depend on
the control variable u or the adjoint variable p. Thus, the optimality system is reduced to an
equation depending on a single system variable. The respective observation holds true for the
biharmonic equation (3.21) which is an equation with respect to the adjoint variable p and does
not explicitly depend on the state y and the control u.

3.4 A-posteriori error estimates for the time discretization

This section gives a summary of the a-posteriori error analysis of [82, Section 3] and a description
of the space-time finite element discretization for the biharmonic equations. The residual based
a-posteriori error estimation is derived for the time discretization of the biharmonic equation for
the state and the adjoint state in the spirit of [4] while keeping the space variable continuous.

Let 0 = t0 < t1 < · · · < tn = T denote a time grid with n ∈ N and ∆tj = tj − tj−1 and
Ij = [tj−1 , tj ] for j = 1, . . . , n. We define the time-discrete space

Vk = {v ∈ H 2,1 (0, T ; Ω) : v|Ij ∈ P1 (Ij )}, V¯k = Vk ∩ H02,1 (0, T ; Ω).

The time-discrete problem consists in finding yk ∈ Vk with yk (0, ·) = y0 and

AT (yk , vk ) = LT (vk ) ∀vk ∈ V¯k . (3.24)

There exists a unique solution to the time-discrete problem (3.24) by construction and utilizing
Lax–Milgram.

Theorem 3.6. ([82, Theorem 3.3]). Let y ∈ H 2,1 (0, T ; Ω) and yk ∈ Vk denote the solutions to
(3.20) and (3.24), respectively. Then, we obtain the a-posteriori error estimation

ky − yk k2H 2,1 (0,T ;Ω) ≤ c ηy2 , (3.25)

with c > 0 and

n Z 2 n Z
X 1 1 X
ηy2 = ∆t2j yd + (yk )tt − yk − ∆2 yk + k∆yk k2L2 (∂Ω) .
Ij α α L2 (Ω) Ij
j=1 j=1

A proof of Theorem 3.6 is given in [82] and uses error estimates for Lagrange interpolation, the
trace inequality and Young’s inequality (Theorem 2.31).
Utilizing the residual based a-posteriori error estimation of Theorem 3.6 leads to a time-adaptive
grid which is tailored for the optimal state solution.
38 3 ADAPTIVITY WITH RESPECT TO TIME

Analogously, this can be carried out for the adjoint variable. We define the space
V˜k = Vk ∩ H̃02,1 (0, T ; Ω).
The time-discrete problem consists in finding pk ∈ V˜k with pk (T, ·) = 0 and
A0 (pk , vk ) = L0 (vk ) ∀vk ∈ V˜k . (3.26)
Theorem 3.7. ([82, Theorem 3.5]). Let p ∈ H̃02,1 (0, T ; Ω) and pk ∈ V˜k be the solutions to (3.23)
and (3.26), respectively. Then, we have
kp − pk k2H 2,1 (0,T ;Ω) ≤ c ηp2 , (3.27)
with c > 0 and
n Z 2 n Z
2
X
2 1 X
ηp = ∆tj −(yd )t + ∆yd + (pk )tt − pk − ∆2 pk + kyd − ∆pk k2L2 (∂Ω) .
Ij α L2 (Ω) Ij
j=1 j=1

Utilizing the residual based a-posteriori error estimation of Theorem 3.7 leads to an adaptive
time grid which is tailored for the optimal adjoint solution and by the optimality condition (3.11)
the adaptive time grid is tailored for the optimal control.

3.5 Space-time mixed finite element discretization

For the numerical solution of the biharmonic equations (3.17) and (3.21), we follow the description
in [82, Section 4] and adapt it to our needs. This approach follows the spirit of [55].
We consider a simultaneous space-time mixed finite element discretization scheme. In order
to use piecewise linear, continuous finite elements, we decouple the biharmonic equation. We
present the discretization for the biharmonic system (3.17) with respect to the state y. The
discretization for the biharmonic system (3.21) with respect to the adjoint state p follows along
the lines. We introduce an auxiliary variable w := −∆y. Then, (3.17) can be written as a
coupled system in y and w by
1 1
−ytt − ∆w + y = yd in (0, T ) × Ω, (3.28a)
α α
∆y + w = 0 in (0, T ) × Ω, (3.28b)
y=0 on [0, T ] × ∂Ω, (3.28c)
w=0 on [0, T ] × ∂Ω, (3.28d)
(yt − ∆y)(T, ·) = 0 in Ω, (3.28e)
y(0, ·) = y0 in Ω. (3.28f)
The use of the auxiliary variable allows to use continuous, piecewise linear finite elements and
avoids the construction of finite element subspaces in H 2 (Ω) which would be needed for a nu-
merical treatment of the fourth-order in space equation (3.17). Regarding this issue compare
Remark 2.48.

We define the function spaces

Y := {v ∈ H 1 (0, T ; H01 (Ω)) : v(0, ·) = 0 in Ω}, W := L2 (0, T ; H01 (Ω)).
Then, the mixed variational form for (3.28) reads as: find y ∈ H 1 (0, T ; H01 (Ω)) with y(0, ·) = y0
and w ∈ W such that
Z TZ Z Z TZ
1 1
yt vt + yv + ∇w∇v dxdt + ∇y(T )∇v(T ) dx = yd v dxdt ∀v ∈ Y, (3.29a)
0 Ω α Ω 0 Ω α
Z TZ
−∇y∇φ + wφ dxdt = 0 ∀φ ∈ W. (3.29b)
0 Ω
3.5 Space-time mixed finite element discretization 39

It is shown in [82, Section 4] that the mixed variational system (3.29) admits at most one solution
(y, w). Then, it is proved that the pair (y, w) with w = −∆y and y denoting the unique solution
to (3.17) is a solution to (3.29). As a consequence, the unique solution y to (3.17) coincides with
the mixed variational solution to (3.28).

We introduce a quasi-uniform partitioning Th of the space-time domain ΩT := (0, T ) × Ω into

disjoint regular orthotopes T , such that Ω̄T = ∪T ∈Th T¯ holds true. We define the finite element
space
Vh := {v ∈ C 0 (Ω̄T ) : ∀T ∈ Th , v|T ∈ Q1 }, (3.30)
where Q1 denotes the space of polynomials of separate degree up to 1. We set Yh := Vh ∩
H 1 (0, T ; H01 (Ω)), Yh,0 := Vh ∩ Y and Wh,0 = Vh ∩ W . Then the fully discrete approximation to
(3.29) reads as: find yh ∈ Yh , wh ∈ Wh,0 with yh (0) = y0 such that
Z T Z Z Z T Z
1 1
(yh )t vt + yh v + ∇wh ∇vdxdt + ∇yh (T )∇v(T )dx = yd v dxdt ∀v ∈ Yh,0 ,
0 Ω α Ω 0 Ω α
(3.31a)
Z T Z
−∇yh ∇φ + wh φ dxdt = 0 ∀φ ∈ Wh,0 .
0 Ω
(3.31b)

Now, we propose an adaptive finite element method tailored for the state variable which utilizes
the residual based a-posteriori error estimate (3.25) in order to adapt the time discretization to
the underlying temporal dynamics. We note that the refinement strategy is based on the error
indicator (3.25) which is discrete in time but continuous in space. In this way, we tailor the
discretization with respect to the time variable and neglect errors stemming from the spatial
variable. This is heuristically justified by the observation in the numerical tests (Section 3.6.3)
that for the considered problem setting the temporal discretization is insensitive to the spatial
discretization. For details on a-priori error estimations for the fully discretized space-time vari-
ational formulation, we refer to [82, Theorem 4.1 and Remark 4.2].
The adaptive finite element method is described in Algorithm 2. We follow the usual adaptivity
cycle
solve → estimate → mark → refine, (3.32)
compare Section 2.2. Since we aim to perform an adaptive cycle with respect to the time variable,
we choose a structured initial space-time mesh Thinit which is a Cartesian product of the spatial
and temporal domain. In particular, we choose an initial time grid 0 = t0 < · · · < tn = T with
n + 1 time points and discretize the spatial domain into d-orthotopes, i.e. intervals for d = 1,
rectangles for d = 2 and cuboids for d = 3, respectively. The advantage of this choice is that
hanging nodes like shown in Figure 8 (middle) are avoided and the bisection of a temporal interval
leads to a structured space-time finite element mesh by construction. Moreover, in such a space-
time discretization each space-time block corresponds to a time step of a semi-discretized problem
and meshes like shown in Figure 8 (right) are avoided. In particular, the mesh is partitioned in
space-time slabs (tj−1 , tj ) × Ω.
As finite element basis functions, we use Q1 nodal functions, i.e. polynomials in space-time
with separate degree up to 1 and local support on the space-time elements. We obtain the
adapted space-time grids by refining the initial space-time grid with respect to time, while the
discretization of the spatial domain stays the same.
Let us specify the adaptive cycle (3.32) in Algorithm 2. In each level of the adaptive cycle, the
variational system (3.31) is solved, for which we use a direct solver. The error estimation relies
on the error indicator (3.25). In particular, for each time interval Ij , j = 1, . . . , n we add all
residual contributions of each spatial domain and neglect the spatial discretization error.
In contrary to [82] where a bulk marking strategy is followed, we utilize the Dörfler criterion
40 3 ADAPTIVITY WITH RESPECT TO TIME

T T T

.
.
.
t2

t1
0 Ω 0 Ω 0 Ω

Figure 8: Space-time meshes for 1d spatial domains. Cartesian mesh without hanging nodes
(left), mesh with hanging nodes (middle), non-Cartesian mesh (right)

(Definition 2.34) as a marking strategy.

As a refinement procedure, we use bisection of those time intervals which are marked.
The adaptive cycle is repeated until either a maximal number of time discretization points is
reached or the error indicator ηy in (3.25) is below a prescribed tolerance.

Algorithm 2 Space-time finite element algorithm with time adaptivity for the state.
Input: Initial space-time mesh Thinit , tolerance ε > 0, refinement parameter θ ∈ (0, 1), desired
state yd , initial state y0 , regularization parameter α > 0, initial number of time discretization
points n, maximal number of time points nmax .
Output: Time-adapted space-time mesh Th , state solution yh .
1: Set Th = Thinit with n time points.
2: while n < nmax OR ηy ≥ ε do
3: Define Vh as the finite element space corresponding to Th according to (3.30).
4: Solve (3.31).
5: Estimate the error contributions from (3.25).
6: Mark time intervals according to the Dörfler criterion (Definition 2.34) with parameter θ.
7: Refine the marked time intervals using bisection.
8: end while

Remark 3.8. The space-time mixed finite element discretization for the adjoint variable follows
along the lines. We define the function spaces

Ỹ := {p ∈ H 1 (0, T ; H01 (Ω)) : p(T ) = 0 in Ω}, W̃ := L2 (0, T ; H 1 (Ω))

and introduce the auxiliary variable ϑ = −∆p. The resulting mixed finite element formulation
for the adjoint state reads as: find ph ∈ Ỹh,0 , ϑh ∈ Vh and ϑh |∂Ω = −ỹd such that
Z T Z Z
1
(ph )t vt + ph v + ∇ϑh ∇vdxdt + ∇ph (0)∇v(0)dx
0 Ω α Ω
Z TZ Z
= (−(yd )t + ∆yd )v dxdt − (yd (0) − y0 )v(0)dx ∀v ∈ Ỹh,0 , (3.33a)
0 Ω Ω
Z TZ
−∇ph ∇φ + ϑh φ dxdt = 0 ∀φ ∈ Wh,0 , (3.33b)
0 Ω

where Ỹh,0 = Vh ∩ Ỹ and ỹd ∈ Vh is an approximation of yd . For more details we refer to [82,
Section 4].

The space-time finite element algorithm with time adaptivity for the adjoint variable is summa-
rized in Algorithm 3.
3.6 Snapshot location in optimal control for POD model order reduction 41

Algorithm 3 Space-time finite element algorithm with time adaptivity for the adjoint state.
Input: Initial space-time mesh Thinit , tolerance ε > 0, refinement parameter θ ∈ (0, 1), desired
state yd , initial state y0 , regularization parameter α > 0, initial number of time discretization
points n, maximal number of time points nmax .
Output: Time adapted space-time mesh Th , adjoint solution ph .
1: Set Th = Thinit with n time points.
2: while n < nmax OR ηp ≥ ε do
3: Define Vh as the finite element space corresponding to Th according to (3.30).
4: Solve (3.33).
5: Estimate the error contributions from (3.27).
6: Mark time intervals according to the Dörfler criterion (Definition 2.34) with parameter θ.
7: Refine the marked time intervals using bisection.
8: end while

3.6 Snapshot location in optimal control for POD model order reduction
Let us recall the POD method from Section 2.4. The POD method is built on an optimal (with
respect to the chosen inner product) orthonormal least-squares approximation of a given solution
trajectory y(t), t ∈ [0, T ] by means of a few POD basis functions ψ1 , . . . , ψ` , see (2.5). In practice,
we do not have the whole solution trajectory at hand, but only a finite number of approximations
at some time instances. Hence, we consider the time-discrete POD problem (2.7), which can be
interpreted as a time-discrete approximation of (2.5) with an appropriate choice for the weights
{αj }nj=0 . A fundamental property of the POD basis is its strong dependence on the snapshots.
In [128] this is described in the principle “If it is not in the snapshots, it is not in the ROM” or
equivalently, “you can only describe what you’ve seen already” in [105, Section 3.3.1], meaning
that a POD reduced-order model can only ever be as good as the snapshots. If a dynamical
feature is missing in the snapshot ensemble, it will be missing in the POD reduced model, if the
reduced model is not updated somehow.
We focus on simulation based POD model order reduction for optimal control problems, i.e. we
generate snapshots by solving the state and the adjoint equation. In order to compute snapshots
which reflect the dynamics of the optimal time-continuous trajectories y(t) and p(t) respectively
u(t) sufficiently well, we need to decide on
(i) a suitable spatial discretization which resolves the dynamical distribution in the spatial
domain sufficiently well,

(ii) a suitable temporal grid at which snapshots are located in order to capture important
dynamical properties within the time domain,

(iii) a suitable input control for snapshot generation.

The first issue (i) will be investigated in Section 4. In this section, we focus on a uniform spatial
discretization. We address issues (ii) and (iii) in this section in the following way. In Section 3.6.1
the snapshot location methodology is presented and discussed. An error analysis of the proposed
approach is given in Section 3.6.2. In Section 3.6.3, a numerical example is presented.

3.6.1 Methodology
The strategy of finding suitable time locations in order to generate the snapshots is based on
(3.25) and (3.27), respectively. The basic idea is to solve either of the elliptic systems (3.17)
and (3.21) adaptively with respect to time using the a-posteriori error estimates according to
Theorem 3.6 and Theorem 3.7, respectively. The resulting time instances are used as locations
for snapshot generation. At the same time, we aim to keep the computational effort to find these
42 3 ADAPTIVITY WITH RESPECT TO TIME

snapshot locations low. In particular, the POD reduced-order modeling of the optimal control
problem including offline costs for snapshot location and POD basis computation shall give a
notable speed up in comparison to solving the full-order model. For this reason, a coarse spatial
resolution is used for solving the elliptic systems (3.17) and (3.21). This will be heuristically jus-
tified in the numerical tests where it turns out that spatial and temporal discretization decouple
for the considered problem setting and data.
Besides the computation of suitable time instances for snapshot generation, the proposed method
produces an approximation of the optimal control at the same time which serves as input control
for snapshot generation.
Moreover, it turns out that it is beneficial to use these snapshot locations also as time grid points
in the POD reduced optimization process, since they are related to the optimal solution.

We investigate three possible approaches:

(i) Snapshot location tailored for the state. We solve the elliptic system (3.17) adaptively
with respect to time using the error estimator (3.25) according to Algorithm 2. This
computation is done using a coarse spatial resolution in order to keep the computational
costs low. As a result, we get time points which are related to the optimal state variable.
At the same time an approximation of the optimal state is computed. The solution of
the adjoint equation (3.10) on a coarse spatial resolution with the approximate state as
right-hand side delivers an approximate adjoint state. From the optimality condition (3.11)
we get an approximation of the optimal control with respect to a coarse spatial resolution
which is utilized as input control for snapshot generation. A summary of this approach is
given by Algorithm 4.

(ii) Snapshot location tailored for the adjoint state / control. We solve the elliptic system (3.21)
adaptively with respect to time using the error estimator (3.27) on a coarse spatial grid.
As a result, we obtain time grid points which are related to the optimal adjoint state. At
the same time, an approximation of the optimal adjoint state is computed. According to
the optimality condition (3.11) we compute an approximation of the optimal control with
respect to a coarse spatial resolution. This control is then utilized as input control for
snapshot generation. A summary of this approach is given by Algorithm 5.

(iii) Snapshot location tailored for both state and adjoint state / control. We run both strategies
(i) and (ii) and then use the union of the time grid points as snapshot locations. At the
same time, we compute two approximations of the optimal control on a coarse spatial mesh
at different time points. From these, we compute a joint approximation as input control
for snapshot generation. This approach is summarized in Algorithm 6.

Note that by Algorithm 2 and Algorithm 3 we obtain the state and adjoint solution to the
biharmonic problem, respectively, which live in a space-time finite element space, i.e. yh ∈ Yh
and ph ∈ Ỹh,0 . In order to compute an approximation for the optimal control which lives for
each time instance in the finite element space Vh , we need to utilize an interpolation. This is
possible due to the specific construction of the triangulation into space-time slabs (tj−1 , tj ) × Ω.
In particular, we use the Lagrange interpolation π in order to interpolate the space-time discrete
functions yh ∈ Yh and ph ∈ Ỹh,0 for each time instance tj into the finite element space Vh , i.e.
π(yh (tj )), π(ph (tj )) ∈ Vh .

Discussion

We discuss the approaches (i), (ii), (iii) with regard to the suitability for the optimization inter-
est, the computational efforts and the heuristic flavor.
Suitability for the optimization interest. If one expects significant variations of the dynamics
3.6 Snapshot location in optimal control for POD model order reduction 43

Algorithm 4 Adaptive snapshot location selection for linear-quadratic optimal control problems
(tailored for the state).
Input: Coarse uniform spatial grid with resolution h+ .
Output: Snapshot locations {tj }nj=0 , approximation of the optimal control uh+ .
1: Call Algorithm 2 in order to solve (3.17) adaptively w.r.t. time using the error estimator (3.25)
with spatial resolution h+ . Obtain time instances {tj }nj=0 = I and the solution yh+ ∈ Yh .
2: Solve (3.10) on I with uniform spatial resolution h+ and right-hand side πyh+ to obtain ph+ .
1
3: Set uh+ = − α ph+ .

Algorithm 5 Adaptive snapshot location selection for linear-quadratic optimal control problems
(tailored for the adjoint state / control).
Input: Coarse uniform spatial grid with resolution h+ .
Output: Snapshot locations {tj }nj=0 , approximation of the optimal control uh+ .
1: Call Algorithm 3 in order to solve (3.21) adaptively w.r.t. time using the error estimator
(3.27) with spatial resolution h+ . Obtain time instances {tj }nj=0 = I and the solution ph+ .
1
2: Set uh+ = − α πph+ .

in time in either of the state or control variable while the other variable is expected to behave
mild, then an adaptive snapshot location strategy which is tailored for either the state or control
might be advantageous. However, if the dynamical behavior of both state and adjoint state
differs significantly over time and strong variations in the temporal domain are expected, the
resulting time grids from Algorithms 4 and 5 might also differ strongly. Thus, the adaptive time
grid tailored for the state variable constructed according to Algorithm 4 might not be a suitable
time grid for the adjoint variable and vice versa. Only Algorithm 6 ensures to capture both
dynamical evolutions of the state and adjoint state over time. However, if one expects a similar
temporal behavior of the state and adjoint variables, it might suffice to only consider one of the
Algorithms 4 and 5 depending on the quantity of interest and the application.
Computational efforts. Concerning the snapshot location computation, Algorithms 4 and 5 have
the same computational effort. However, the computation of the approximate control is cheaper
in Algorithm 5 since it only requires the solution of (3.11), while in Algorithm 4 it requires
additionally one solve of (3.10). The approach (iii) is the most expensive approach. The compu-
tational costs are the sum of the computational costs of Algorithms 4 and 5 plus the additional
costs for the construction of the approximate control uh+ according to step 3 in Algorithm 6.
Heuristic flavor. The efficiency of the proposed approaches is based on the assumption that

Algorithm 6 Adaptive snapshot location selection for linear-quadratic optimal control problems
(tailored for both state and adjoint state / control).
Input: Coarse uniform spatial grid with resolution h+ .
Output: Snapshot locations {tj }nj=0 , approximation of the optimal control uh+ .
1
1: Run Algorithm 4 and obtain snapshot locations {t1j }n j=0 and an approximation of the optimal
control u1h+ .
2
2: Run Algorithm 5 and obtain snapshot locations {t2j }n j=0 and an approximation of the optimal
control u2h+ .
3: Set {tj }n 1 n1 2 n2
j=0 = {tj }j=0 ∪ {tj }j=0 and set for j = 0, . . . , n
1 2

1 if tj ∈ {t1j }nj=0 and tj ∈
/ {t2j }nj=0 ,
uh+ (tj )

2 1
uh+ (tj ) = u2h+ (tj ) if tj ∈ {t2j }nj=0 and tj ∈
/ {t1j }nj=0 ,
1 n1 2 n2

1 1 2
2 (uh+ (tj ) + uh+ (tj )) if tj ∈ {tj }j=0 and tj ∈ {tj }j=0 .
44 3 ADAPTIVITY WITH RESPECT TO TIME

temporal and spatial discretization decouple for the considered problem setting and data. This
behavior is discovered in numerical experiments. In this way, the method has a certain heuristic
flavor. Based on this observation we use a coarse spatial resolution in the offline phase in or-
der to generate suitable snapshot locations. In general, the issue of when spatial and temporal
discretization decouple is not easy to answer. Related to this issue, we refer to [178, Chapter
3], where the relationship of temporal and spatial discretization is analyzed in the context of
unsteady internal flow. In [30] a combination of time adaptivity and spatial adaptivity is con-
sidered including a-priori and a-posteriori error analysis.

POD reduced-order modeling for the optimal control problem

After the computation of the snapshot locations and an approximate control uh+ using either of
the Algorithms 4 to 6, the usual POD offline phase begins. For this, we sample state and adjoint
snapshots at the computed snapshot locations with respect to a fine spatial resolution by solving
(3.15) with control input Iuh+ and (3.16), where I denotes a suitable interpolation of the control
uh+ from a coarse spatial grid onto the fine finite element grid.
In order to derive a POD reduction of the optimal control problem (3.9), we follow the first-
reduce-then-optimize approach, see Section 2.4.2. We utilize an integrated snapshot ensemble,
i.e. we take into account both state and adjoint snapshots, compare Remark 2.47. In this way,
the POD projection error for both state and adjoint variables is determined by the sum of the
neglected eigenvalues. Assume we have constructed a POD reduced space X` ⊂ H01 (Ω) according
to (2.7). Let us denote by ỹ0 a suitable approximation of the initial value y0 in the reduced space
X` . For example, it can be computed as the solution to
(ỹ0 , ψ)L2 (Ω) = (y0 , ψ)L2 (Ω) ∀ψ ∈ X` .
Note that it is also possible to utilize a homogenization with respect to the initial condition
(according to e.g. [89, Remark 3.3]) which then allows to construct a reduced POD model for
which the POD reduced state fulfills the initial condition exactly, see [89, Section 3.3]. The POD
reduced-order optimal control problem reads as follows:
min J` (y` , u` ) s.t. y` = S` (u` , ỹ0 ), (3.34)
(y` ,u` )∈W (0,T ;X` )×L2 (0,T ;Ω)

where the linear operator

S` : L2 (0, T ; H −1 (Ω)) × X` → W (0, T ; X` ), (û, ỹ0 ) 7→ y` := S` (û, ỹ0 )
is defined analogously to the solution operator S defined in (3.5). In particular, we replace the
space H01 (Ω) in (3.2) by X` . This leads to the following reduced-order state equation: for a given
initial condition y` (0) := ỹ0 determine a function y` ∈ W (0, T ; X` ) such that
d
(y` (t), ψ)L2 (Ω) + (∇y` (t), ∇ψ)L2 (Ω) = (u` (t), ψ)L2 (Ω) for t ∈ [0, T ], ψ ∈ X` . (3.35)
dt
Note that we use the notation u` for the control in order to indicate the affiliation of the control
to the reduced-order optimization problem. This helps to distinguish notationally between the
solution u to (3.9) and the solution u` to (3.34). However, note carefully that we do not impose a
POD reduction for the control space. In order to achieve a conservative reduction of the control
variable, the reduction of the control is induced by the reduction of the adjoint variable due to
the optimality condition
αu` + p` = 0 in [0, T ] × Ω, (3.36)
where p` denotes the solution to the following reduced-order adjoint equation: determine a
function p` ∈ W (0, T ; X` ) such that p` (T ) = 0 and
d
− (p` (t), ψ)L2 (Ω) + (∇p` (t), ∇ψ)L2 (Ω) = ((y` − yd )(t), ψ)L2 (Ω) for t ∈ [0, T ], ψ ∈ X` , (3.37a)
dt
3.6 Snapshot location in optimal control for POD model order reduction 45

Existence of a unique solution y` and p` to (3.35) and (3.37), respectively, as well as stability
results follow analogously to the infinite-dimensional case according to [69] and X` ⊂ H01 (Ω).

Now, the POD online phase begins in which (3.34) is solved. Since we already have a time grid
at hand which is related to the optimal control problem, it makes sense to use this time grid for
solving the problem (3.34). The POD reduced-order modeling for the linear-quadratic optimal
control problems using adaptive snapshot computation is summarized in Algorithm 7.

Algorithm 7 POD model order reduction for linear-quadratic optimal control problems using
adaptive snapshot locations.
Input: Snapshot locations {tj }nj=0 = I and input control uh+ from either of the Algorithms 4
to 6, fine spatial resolution h∗ , number of POD basis functions `
Output: y` , u`
1: Sample state and adjoint snapshots in a simulation of (3.15) and (3.16) with spatial resolution
h∗ and input control Iuh+ on I.
2: Compute a POD basis of rank ` according to (2.7) using an integrated snapshot ensemble.
3: Set up and solve the POD-ROM (3.34) on I to obtain y` , u` .

Remark 3.9. In principle, it is conceivable to derive a POD reduced-order formulation for the
biharmonic systems (3.17) and (3.21), respectively. However, this approach is not followed here,
since our motivation is to use the usual optimization algorithm, for example of Remark 2.45, and
utilize existing implementations. The snapshot location strategy is considered as an add-on with
the goal to deliver time grid points and an input control which are associated with the optimal
solution. For a space-time POD Galerkin approach we refer to [26, 187].

3.6.2 Error analysis

We want to know how well the time-discrete solution to the POD reduced-order optimization
problem (3.34) approximates the optimal solution to (3.9).
We use the following notation. Suppose that ` ∈ N is fixed. Let ū be the optimal control, i.e.
the solution to (3.9) and let ū` denote the optimal POD control, i.e. the solution to (3.34). Let
ȳ = y(ū) denote the associated state and p̄ = p(ȳ) = p(y(ū)) is the associated adjoint state.
Analogously, let ȳ` = y` (ū` ) denote the associated POD state and p̄` = p` (ȳ` ) = p` (y` (ū` )) the
associated adjoint state. We denote by yk and pk the solution to the time-discrete problem (3.24)
and (3.26), respectively. By ỹk (up ) and p̃k (up ) = p̃k (ỹk (up )) we denote time-discrete solutions
to (3.2) and (3.12), respectively, for a given particular control up . Moreover, we define the
time-discrete controls uk := − α1 pk and ûk := − α1 p̃k (yk ). We make the assumptions

kyk − ỹk (ûk )k < ε, (3.38a)

kpk − p̃k (ỹk (uk ))k < ε, (3.38b)

for 0 < ε  1. By (ȳk` , ū`k ) we denote the time-discrete solution to (3.34) and p̄`k is the associated
adjoint state.

For clarity, we indicate by k · k the k · kL2 (0,T ;X) -norm for either X = V = H01 (Ω) or X =
H = L2 (Ω). We define the sets H` and V` by H` = span{ψ1H , . . . , ψ`H } ⊂ V ⊂ H and V` =
span{ψ1V , . . . , ψ`V } ⊂ V , where the superscript indicates which space is used for X in (2.5). We
introduce the orthogonal projection P ` , which can either be PH ` : H → H ⊂ V or P ` : V →
` V
V` ⊂ V as follows
`
for f ∈ H, PH f minimizes inf kf − f` kH
f` ∈H`
46 3 ADAPTIVITY WITH RESPECT TO TIME

and
for f ∈ V, PV` f minimizes inf kf − f` kV .
f` ∈V`

These projections can be computed using the following expressions

`
X `
X
`
PH f = (f, ψiH )H ψiH and PV` f = (f, ψiV )V ψiV .
i=1 i=1

Error bound for the state

Let us start with the error estimation with respect to the state variable. We are interested in
estimating the quantity
kȳ − ȳk` k.
Using the null-trick and applying the triangle inequality we arrive at

kȳ − ȳk` k ≤ kȳ − yk k + kyk − ỹk (ûk )k + kỹk (ûk ) − ỹk` (ûk )k + kỹk` (ûk ) − ȳk` k (3.39)
| {z } | {z } | {z } | {z }
(3.39.i) ([Link]) ([Link]) ([Link])

Now, we analyze and estimate each term in (3.39).

If we compute the snapshot locations and the input control with either of the Algorithms 4 and 6,
then the term (3.39.i) can be estimated utilizing the a-posteriori error estimate (3.25) for the
state and concerns the snapshot location. Since V ,→ H there exists an embedding constant cv
such that
kvkH ≤ cv kvkV ∀v ∈ V.
In particular, it holds

kȳ − yk k2L2 (0,T ;H) ≤ cv kȳ − yk k2L2 (0,T ;V ) ≤ cv kȳ − yk k2H 2,1 (0,T ;Ω) ≤ c1 ηy2 , (3.40)

where c1 := cv · c with c denoting the constant from (3.25). The use of sufficiently enough time
instances ensures that the indicator ηy drops below a prescribed tolerance.

For the term ([Link]) we make use of the assumption (3.38a).

Let us now focus on the part ([Link]). For this, we make use of the decomposition

ỹk (ûk ) − ỹk` (ûk ) = ỹk (ûk ) − P ` ỹk (ûk ) + P ` ỹk (ûk ) − ỹk` (ûk ).

After the computation of the time-discrete solution yk of the biharmonic equation, we take the
control ûk = − α1 p̃k (yk ) as input control for snapshot generation. Since we utilize an integrated
snapshot set, i.e. we utilize both state and adjoint information, we consider the POD problem
` `
Z T !
X X X
2 2
min kỹk (t) − (ỹk (t), ψi )X ψi kX + kp̃k (t) − (p̃k (t), ψi )X ψi kX dt = λi ,
ψ1 ,...,ψ` ∈X 0 i=1 i=1 i>`

where we left out the dependency of ỹk = ỹk (ûk ) and p̃k = p̃k (yk ) for the sake of clarity. This
gives us the estimation X
kỹk (ûk ) − P ` ỹk (ûk )k2 ≤ λi . (3.41)
i>`

What is left, is to estimate the quantity − P ` ỹk (ûk ) ỹk` (ûk )

=: ϑ. For this, we follow [103,
Proposition 4.7, equation (4.19)] which gives the estimate
Z T
2
ess supkϑ(t)kH + kϑ(t)k2V dt ≤ kϑ(0)k2H + k(ỹk (ûk ))t − P ` (ỹk (ûk ))t k2L2 (0,T ;V ∗ ) . (3.42)
t∈[0,T ] 0
3.6 Snapshot location in optimal control for POD model order reduction 47

Finally, we consider the last part ([Link]). Deriving an energy estimate for the POD reduced
state equation following (3.4) leads to the estimate
kỹk` (ûk ) − ȳk` kL2 (0,T ;H) ≤ cv kỹk` (ûk ) − ȳk` kL2 (0,T ;V ) ≤ c2 kûk − ū`k kL2 (0,T ;Ω) ,

with c2 = cv · c and c according to (3.4). By ū`k we denote the time-discrete optimal control to
the POD reduced optimal control problem with POD basis computed from snapshots generated
from an input control ûk . In order to control the quantity
kûk − ū`k kL2 (0,T ;Ω) ≤ kûk − ūkL2 (0,T ;Ω) + kū − ū`k kL2 (0,T ;Ω)
we make use of the a-posteriori error estimation according to [177] which provides an upper
bound for the error between the (unknown) optimal control ū and an particular control up (here
up = ûk and up = ū`k ). It holds
1
kū − up kL2 (0,T ;Ω) ≤ kζp kL2 (0,T ;Ω) , (3.43)
α
according to [177, Theorem 3.1] adapted to this setting. The perturbation ζp ∈ L2 (0, T ; Ω) is
chosen such that αup + pp + ζp = 0 holds, where pp denotes the adjoint state associated with
the control up . For the construction of ζp we refer to [177, Proposition 3.2] which involves a full
evaluation of the adjoint and state equation (and not a solution of their POD approximations).

In summary, we can build the following theorem.

Theorem 3.10. Let ȳ denote the optimal state for (3.9) and let ȳk` denote the time-discrete
solution to (3.34). The POD basis is computed from an integrated snapshot set with input control
and snapshot locations computed according to either of the Algorithms 4 and 6. In (2.5) it is
either X = V or X = H. Then, it holds true
!1/2
√ X
kȳ − ȳk` kL2 (0,T ;X) ≤ c1 ηy + ε + λi + k(ỹk (ûk ))t − P ` (ỹk (ûk ))t kL2 (0,T ;V ∗ )
i>` (3.44)
c2
+ (kζ1 kL2 (0,T ;Ω) + kζ2 kL2 (0,T ;Ω) ),
α
where ζ1 , ζ2 ∈ L2 (0, T ; Ω) are chosen such that αûk + p̃k (ûk ) + ζ1 = 0 and αū`k + p̃k (ū`k ) + ζ2 = 0,
respectively, is fulfilled and c1 , c2 > 0 are constants.
Proof. Utilizing the decomposition (3.39) together with (3.40), (3.38a), (3.41), (3.42) and (3.43)
leads to the claim.

Remark 3.11. i) Note that the error estimation (3.44) in Theorem 3.10 depends on an er-
ror of time derivatives of the state. As discussed in [103, Remark 4.8(b)] this dependency
can be avoided by enriching the snapshot set with time derivatives (more specifically, finite
difference approximations of time derivatives), see also [102, 121].
Note that for a specific setting it is possible to avoid to include time derivatives into the
snapshot set. This is based on [165]. A comprehensive analysis on whether snapshot dif-
ference quotients are needed in POD is given in [109].
ii) Note that Theorem 3.10 is valid if the snapshot location strategy of either of the Algorithms 4
and 6, respectively, is used, i.e. the snapshot locations are related to the optimal state.
However, if Algorithm 5 is utilized (snapshot locations are only related to the adjoint state),
it is not guaranteed that the error contribution (3.39.i) is small, i.e. the selected time
grid points which are related to the optimal adjoint state might not be suitable for the
state variable. In Section 3.8 it is not possible to reformulate the optimality system as a
biharmonic equation for the state variable. For this case, a post-processing step is introduced
in Section 3.8.2.
48 3 ADAPTIVITY WITH RESPECT TO TIME

Error bound for the control

We are interested in estimating how well the optimal control ū is approximated by the time-
discrete POD reduced-order solution ū`k . For this, the a-posteriori error estimate according to
[177] can be utilized. Moreover, we recall the optimality condition (3.11) and a time discretization
of the optimality condition (3.36), i.e.

αū + p̄ = 0 in [0, T ] × Ω,
αū`k + p̄`k = 0 in [0, T ] × Ω.
Subtracting both equations leads to
1
ū − ū`k = − (p̄ − p̄`k ) in [0, T ] × Ω
α
and thus
1
kū − ū`k kL2 (0,T ;X) =
kp̄ − p̄`k kL2 (0,T ;X) in [0, T ] × Ω. (3.45)
α
Therefore, we now consider an error bound for the adjoint variable.

Error bound for the adjoint state

Analogously to the error analysis for the state, we derive an error estimation for the quantity

kp̄ − p̄`k k.

This is done along the lines to the error bound for the state, so we shorten the analysis to the
main points. Using the null-trick and applying the triangle inequality we arrive at

kp̄ − p̄`k k ≤ kp̄ − pk k + kpk − p̃k (uk )k + kp̃k (uk ) − p̃`k (uk )k + kp̃`k (uk ) − p̄`k k (3.46)
| {z } | {z } | {z } | {z }
(3.46.i) ([Link]) ([Link]) ([Link])

If we compute the snapshot locations and the input control with Algorithm 5, then the term
(3.46.i) can be estimated utilizing the a-posteriori error estimate (3.27) for the adjoint state and
concerns the snapshot location. This yields to

kp̄ − pk k2 ≤ cv kp̄ − pk k2H 2,1 (0,T ;Ω) ≤ c1 ηp2 . (3.47)

For term ([Link]) we utilize (3.38b). In order to estimate ([Link]), we make use of the decom-
position
p̃k (uk ) − p̃`k (uk ) = p̃k (uk ) − P ` p̃k (uk ) + P ` p̃k (uk ) − p̃`k (uk ).
In line with (3.41), it holds X
kp̃k (uk ) − P ` p̃k (uk )k2 ≤ λi . (3.48)
i>`

In order to estimate the quantity P ` p̃k (uk ) − p̃`k (uk ) =: θ, we get from [103, Proposition 4.7] the
estimation
Z T X
ess supkθ(t)k2H + kθ(s)k2V dt ≤ c λi + k(p̃k (uk ))t − P ` (p̃k (uk ))t k2L2 (0,T ;V ∗ )
t∈[0,T ] 0 i>` (3.49)

` 2
+kỹk (uk ) − ỹk (uk )kL2 (0,T ;H) .

Finally, the last term ([Link]) can be estimated in terms of the control according to the energy
estimate (3.14) for the POD reduced-order adjoint equation. Thus, it holds

kp̃`k (uk ) − p̄`k (ū`k )k ≤ ckuk − ū`k kL2 (0,T ;Ω) ,

3.6 Snapshot location in optimal control for POD model order reduction 49

with constant c > 0 from (3.14). Again, with the use of the a-posteriori error estimation according
to [177] this term can be controlled.
We summarize the findings in the following theorem.
Theorem 3.12. Let p̄ denote the optimal adjoint state for (3.9) and let p̄`k denote the time-
discrete adjoint solution to a POD approximation of (3.9). The POD basis is computed from
an integrated snapshot set with input control and snapshot locations computed according to Al-
gorithm 5. In (2.5) it is either X = V or X = H. Then, it holds true
!1/2
√ X
kp̄ − p̄`k kL2 (0,T ;X) ≤ c1 ηp + ε + c2 λi + c2 k(p̃k (uk ))t − P ` (p̃k (uk ))t k2L2 (0,T ;V ∗ )
i>` (3.50)
` 2 c3
+c2 kỹk (uk ) − ỹk (uk )kL2 (0,T ;H) + (kζ1 kL2 (0,T ;Ω) + kζ2 kL2 (0,T ;Ω) ),
α
where ζ1 , ζ2 ∈ L2 (0, T ; Ω) are chosen such that αuk + p̃k (uk ) + ζ1 = 0 and αū`k + p̃k (ū`k ) + ζ2 = 0,
respectively, is fulfilled and c1 , c2 , c3 > 0 are constants.
Proof. Utilizing the decomposition (3.46) together with (3.47), (3.38b), (3.48), (3.49) and (3.43)
lead to the claim.
Remark 3.13. i) Note that the error estimate (3.50) in Theorem 3.12 depends on the time
derivatives of the adjoint state. The same arguments of Remark 3.11(i) apply here.
ii) The estimate (3.50) in Theorem 3.12 is further dependent on the error between the state
snapshots with input control uk and the POD approximation with the same input control.
In order to estimate this, we can utilize the estimations for ([Link]). However, it is not
guaranteed that the snapshot locations generated by Algorithm 5 lead to snapshot locations
which are suitable for the state as well, compare Remark 3.11(ii).

3.6.3 Numerical results

In this numerical test, we construct an example in such a way that we know the analytical solu-
tion to the optimal control problem (3.9). Similar examples are presented in [9, 10]. All coding
is done in MATLAB R2018a. The time-adaptive solution of the biharmonic system is built upon
the codes from [82].

Problem setting

As spatial domain, we consider the one-dimensional interval Ω = (0, 1) ⊂ R. The end time is
T = 1 and the regularization parameter is α = 1. We introduce a forcing term f ∈ L2 (0, T ; Ω)
on the right-hand side of the state equation (4.40a) as follows
yt − ∆y = u + f in (0, T ] × Ω.
The remaining setting of the optimal control problem (3.9) stays the same. In order to reformulate
the resulting optimality system with respect to the state variable we need to assume higher
regularity for the forcing term, in particular f ∈ L2 (0, T ; H 1 (Ω)) ∩ H 1 (0, T ; L2 (Ω)). Then, in the
biharmonic system (3.17), the equations include terms involving the forcing term in the following
way
1 1
−ytt + ∆2 y + y = yd − ∆f − ft in (0, T ) × Ω, (3.51a)
α α
y=0 on [0, T ] × ∂Ω, (3.51b)
−∆y = f on [0, T ] × ∂Ω, (3.51c)
(yt − ∆y)(T, ·) = f (T, ·) in Ω, (3.51d)
y(0, ·) = y0 in Ω. (3.51e)
50 3 ADAPTIVITY WITH RESPECT TO TIME

The resulting elliptic problem for the adjoint variable includes the forcing term in the following
way
1
−ptt + ∆2 p + p = f − (yd )t + ∆yd in (0, T ) × Ω, (3.52a)
α
p=0 on [0, T ] × ∂Ω, (3.52b)
∆p = yd on [0, T ] × ∂Ω, (3.52c)
(pt + ∆p)(0, ·) = yd (0, ·) − y0 in Ω, (3.52d)
p(T, ·) = 0 in Ω. (3.52e)
Note that the residual based error indicators of Theorems 3.6 and 3.7 contain the terms involving
the forcing term accordingly.

We choose the optimal state as

 
t−1
ȳ(t, x) = sin(πx)atan .
ε
For small values of 0 < ε  1, the state variable ȳ develops a steep gradient for t → 1+ , compare
Figure 9, which shows the part which accounts for the steep gradient in the temporal domain.
For the numerical test, we use ε = 10−05 .

0 0

-0.5 -0.5

-1 -1

-1.5 -1.5
0 0.25 0.5 0.75 1 0.99 1

t−1


Figure 9: Time layer function atan ε (left) and zoom in (right)

The optimal control is chosen as

ū(t, x) = − sin(πx) sin(πt).
The analytical optimal state, its contour lines and the optimal control is shown in Figure 10.

Figure 10: Analytical optimal state ȳ (left), its contour lines (middle) and analytical optimal
control ū

Accordingly, the initial condition is

 
−1
y0 (x) = sin(πx)atan
ε
3.6 Snapshot location in optimal control for POD model order reduction 51

and the desired state is

   
t−1 2
yd (t, x) = sin(πx) atan + π cos(πt) − π sin(πt) .
ε
The forcing term is given as
   
ε t−1
f (t, x) = sin(πx) + π 2 atan + sin(πt) .
t − 2t + ε2 + 1
2 ε
In order to solve the optimal control problem (3.9) and the reduced-order optimal control prob-
lem (3.34), we use an implicit Euler method for temporal discretization, continuous piecewise
linear finite elements for spatial discretization and a gradient method with Armijo step size
determination, see Remark 2.45. As stopping criterion for the optimization, we use

k∇Jˆ` (u)k ≤ εr k∇Jˆ` (0)k + εa

or a maximal number maxiter of gradient iterations is reached, where we set εr = 10−08 as

relative tolerance, εa = 10−08 as absolute tolerance and maxiter = 20.

As time weights for computing POD basis functions in (2.7), we use trapezoidal weights
∆t1 ∆tj+1 + ∆tj ∆tn
α0 = , αj = for j = 1, . . . , n − 1, αn = .
2 2 2
Moreover, in (2.7) we use the X = L2 (Ω) for the inner product and norm.

The focus in this numerical test example lies on snapshot location. Hence we compare a uniform
placement of time instances for snapshot generation with an adaptive time discretization. So,
for a fixed number of time instances n + 1, we compare a uniform time grid with time step size
∆t = 1/n with an adaptive time grid with the same number of time points.
For the adaptive time discretization we compare the Algorithms 4 to 6. Due to the steep gra-
dient in the optimal state at the end of the time interval we expect that a time grid which is
adapted with respect to the state will improve the error in the state variable since we expect
that it locates the time grid points towards the end of the time interval where the layer is located.

Uniform snapshot location

First, we consider POD reduced-order modeling with a uniform time discretization. For now,
we consider an input control u = 0 for snapshot generation and focus on the influence of the
snapshot locations and time discretization on the approximation quality. Later, we investigate
other choices for the input control. We fix the POD basis length to ` = 1 and investigate different
lengths for the POD basis later. We utilize an integrated snapshot set (see Remark 2.47), i.e.
we compute a POD basis corresponding to a union of state and adjoint snapshots. This POD
basis is used in the POD Galerkin ansatz for both state and adjoint state. Figure 11 shows the
solution y` to (3.34), its contour lines and the associated control u` . We choose n = 27, i.e. we
have 28 time instances. We see that on a uniform grid, the layer in the optimal state at the end
of the time interval is overshot, i.e. the approximation is not satisfactory.

Adaptive snapshot location

Now, we use the adaptive time selection according to Algorithms 4 to 6 and produce an adaptive
time grid with the same number of time instances, i.e. n = 27, as in the uniform case. As initial
time grid for the time-adaptive selection we start with a uniform time grid with 9 time instances.
Then, according to the error indicator (3.25) and (3.27), respectively, suitable time instances are
52 3 ADAPTIVITY WITH RESPECT TO TIME

3000 0

2000
0.5 -0.5
1000

0 -1
1 1
1 1
0.5 0.5
0.5 0 0.5
0 0 0.5 1 0
0 0

Figure 11: POD optimal state computed on an equidistant time grid with n = 27 (left), its
contour lines (middle) and the POD optimal control (right)

selected. For the Dörfler criterion we use the parameter θ = 0.5.

The resulting time grid for the adaptivity with respect to the state according to Algorithm 4
is shown in Figure 12 (left and middle) for different spatial resolutions. At the end of the time
interval where the steep gradient is located, many time instances are located whereas before a
coarse resolution is sufficient.
The resulting time grid for the adaptivity with respect to the adjoint according to Algorithm 5
is shown in Figure 13 (left and middle) for different spatial resolutions. The location of the time
grid points is spread more since there are not as steep transitions as in the state variable.
The resulting time grid for the adaptivity addressing both state and adjoint state according to
Algorithm 6 is shown in the right plot of Figure 13. Note that this joint time-adaptive grid has
the same total number of time instances as in the pure state / adjoint state adaptive schemes in
order to achieve a fair comparison. For completeness, wee see a uniform space-time discretization
in the right plot of Figure 12.

We see for both state and adjoint variables in this example, that it does not matter whether a
fine spatial resolution or a coarse spatial resolution is used, the resulting time grid is the exact
same. This means that the temporal discretization in this case is insensitive against changes in
the spatial resolution, i.e. spatial and temporal discretization decouple. This observation justifies
the use of a coarse spatial resolution for the solution of the biharmonic system in the offline phase
in order to achieve cheap computations.

1 1 1

0.75 0.75 0.75

0.5 0.5 0.5

0.25 0.25 0.25

0 0 0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

Figure 12: Adaptive space-time grids according to Algorithm 4 with n = 27 and h+ = 1/100
(left) and h+ = 1/5 (middle), respectively, and the equidistant grid with n = 27 and h+ = 1/100
(right)
3.6 Snapshot location in optimal control for POD model order reduction 53

1 1 1

0.75 0.75 0.75

0.5 0.5 0.5

0.25 0.25 0.25

0 0 0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

Figure 13: Adaptive space-time grids according to Algorithm 5 with n = 27 and h+ = 1/100 (left)
and h+ = 1/5 (middle), respectively, and the adaptive space-time grid according to Algorithm 6
with n = 27 time points altogether (right)

POD model order reduction for optimal control using adaptive snapshot locations

Now we generate state and adjoint snapshots on the adaptive time mesh. Note that we now
compute the snapshots corresponding to an input control u = 0 (in contrary to the Algorithms 4
to 6) in order to numerically compare the results to the uniform case, where the only difference
between the models is the time discretization. A different input control is investigated later. We
also solve the POD reduced-order model optimization problem on the time-adaptive mesh.
The POD optimal state, its contour lines and the POD optimal control utilizing the state adaptive
grid according to Algorithm 4 are shown in Figure 14. Note that this looks more similar to the
analytical solution shown in Figure 10 than the POD solution on a uniform time mesh. Especially
note the different scaling. The layer is captured better since there are more time instances located
at the end of the time interval in comparison to a uniform discretization.

1 0

0
0.5 -0.5
-1

-2 -1
1 1
1 1
0.5 0.5
0.5 0 0.5
0 0 0.5 1 0
0 0

Figure 14: POD optimal state computed on the adaptive time grid with n = 27 (left), its contour
lines (middle) and the POD optimal control using Algorithm 4 with input control u = 0 (right)

The POD optimal state, its contour lines and the POD optimal control utilizing the adaptive
grid regarding the adjoint state according to Algorithm 5 are shown in Figure 15. We observe
that the results for the state solution look even worse than using a uniform time mesh, since the
time discretization at the end of the time interval is coarser in the case of the adjoint adapted
mesh than in the uniform case. In particular, the adaptive time grid for the adjoint state turns
out to be not suitable for the approximation of the state.
Finally, the POD optimal state, its contour lines and the POD optimal control utilizing the union
of the adaptive grids according to Algorithm 6 are shown in Figure 16. Note that the overlay
time mesh has the same number of time instances in total as in the pure state or pure adjoint
adaptive case, such that results are comparable.
We consider the relative L2 (0, T ; Ω)-error between the analytical solution and the POD solution
in the state and in the control variable given as
kȳ − y` ||L2 (0,T ;Ω) kū − u` ||L2 (0,T ;Ω)
erry = and erru =
kȳkL2 (0,T ;Ω) kūkL2 (0,T ;Ω)
54 3 ADAPTIVITY WITH RESPECT TO TIME

4000 0

3000

2000 0.5 -0.5

1000

0 -1
1 1
1 1
0.5 0.5
0.5 0 0.5
0 0 0.5 1 0
0 0

Figure 15: POD optimal state computed on the adaptive time grid with n = 27 (left), its contour
lines (middle) and the POD optimal control using Algorithm 5 with input control u = 0 (right)

0
600

400
0.5 -0.5
200

0 -1
1 1
1 1
0.5 0.5
0.5 0 0.5
0 0 0.5 1 0
0 0

Figure 16: POD optimal state computed on the adaptive time grid with n = 27 (left), its contour
lines (middle) and the POD optimal control using Algorithm 6 with input control u = 0 (right)

for different numbers of time instances n in Figure 17.

Concerning the error in the state variable, the advantage of using an adaptive time grid which
is tailored for the state is obvious. For n ≤ 45 it also delivers better results than the time grid
according to Algorithm 6, since we compare the approximation results for the same number of
time instances.

Regarding the error in the control variable, we observe that for all four approaches an enlarge-
ment of the number of time instances leads to a reduction in the error. The best results are
achieved in this example if we consider the uniform or the adjoint adaptive method.

10 4 10 0

10 2
10 -1

0
10

10 -2
— uniform
10 -2
— state adaptive
— adjoint adaptive
— state and adjoint adaptive
10 -4 10 -3
0 20 40 60 80 100 0 20 40 60 80 100

Figure 17: Relative L2 (0, T ; Ω)-errors between the analytical optimal solution and the POD
optimal solution depending on n. Left: error in the state, right: error in the control
3.6 Snapshot location in optimal control for POD model order reduction 55

10 0 0

10 -5
-0.5

10 -15 -1
— uniform
x state adaptive
-20
10 o adjoint adaptive
-1.5
1 2 3 4 5 0 0.25 0.5 0.75 1

Figure 18: Decay of normalized eigenvalues (left) and first POD mode ψ1 (x) using adaptive and
uniform time grids, respectively (right)

Influence of POD basis length on the approximation quality

In Figure 18 (left) we see the decay of the eigenvalues for a uniform and an adaptive time dis-
cretization using n = 27. The POD basis is computed corresponding to the input control u = 0
from a joint state and adjoint state snapshot set. Moreover, Figure 18 (right) shows the first
POD basis using a uniform mesh with n = 27 and the POD basis using an adaptive time grid
with the same number of time points. We see that in this example, increasing the number of
utilized POD basis functions will not lead to an improvement of the POD reduced-order accuracy
since λ2 ≤ 10−15 , i.e. no more information will be added. Moreover, the first POD basis function
computed with a uniform time discretization is numerically the same as the POD basis computed
corresponding to a state or adjoint adaptive time discretization. The same accounts for the POD
computations for the joint state and adjoint adaptive scheme, but is left out in the plots due to
visibility. The reason why only one POD basis function is sufficient is due to the fact that the
spatial structure of the optimal state and adjoint state is determined by the function sin(πx).
This choice is done by purpose since the focus of the numerical test lies on the snapshot location
and we want to eliminate other dependencies. More POD basis functions will be needed if the
spatial structure of the optimal state and/or the adjoint state cannot be determined by only one
function.

Moreover, Figure 18 even tells us more: since the uniform time discretization and the adap-
tive time discretization lead to the exact same POD basis, the location of the snapshots does
not influence the POD basis in this setting. The time locations play an important role in the
optimization algorithm. We provide an example in Section 6.1, in which a dynamical signal is
missed in case of a uniform snapshot location and the resulting snapshots are zero functions, such
that the corresponding POD basis is an arbitrary orthonormal basis. In contrary, the adaptive
snapshot location is able to detect the time signal and the corresponding POD basis captures
the spatial structure satisfactorily.

Influence of the input control

We now investigate the influence of the input control. In Figure 19 (left) we see the space-time
state solution yh+ computed according to the state adaptive scheme of Algorithm 4 on a coarse
spatial mesh with resolution h+ = 1/5. Then, we compute an associated adjoint solution by
solving (3.10) with right-hand side πyh+ , where π denotes the Lagrange interpolation which in-
terpolates the space-time discrete function yh+ ∈ Yh for each time instance into the finite element
space Vh . The corresponding control uh+ is then computed according to (3.11) on a coarse spatial
mesh which is shown in Figure 19 (middle). We interpolate this control onto a fine mesh with
resolution h∗ = 1/100, Figure 19 (right), and use this interpolation Iuh+ as input control for
snapshot generation.
56 3 ADAPTIVITY WITH RESPECT TO TIME

0 0 0

-1
-0.5 -0.5
-2

-3 -1 -1
1 1 1
1 1 1
0.5 0.5 0.5
0.5 0.5 0.5
0 0 0 0 0 0

Figure 19: Space-time state yh+ with n = 27 and h+ = 1/5 (left), corresponding control uh+
(middle) and interpolation Iuh+ of this control onto a fine mesh with h∗ = 1/100 (right)

Next, we consider Algorithm 5. The adjoint space-time solution ph+ with coarse spatial reso-
lution h+ = 1/5 as well as the with πph+ associated control uh+ are shown in Figure 20 (left
and middle). Finally, the interpolation of the control uh+ according to Algorithm 6 onto a fine
mesh is shown in Figure 20 (right). Note that this control constitutes a poor approximation
of the optimal control since its appearance is dominated by the part stemming from the state
adaptive scheme, but using only 19 time points (such that the joint time grid has 28 time points).

1 0 0

-50

0.5 -0.5 -100

-150

0 -1 -200
1 1 1
1 1 1
0.5 0.5 0.5
0.5 0.5 0.5
0 0 0 0 0 0

Figure 20: Space-time adjoint state ph+ with n = 27 and h+ = 1/5 (left), interpolation of
corresponding control Iuh+ onto a fine mesh with h∗ = 1/100 (middle) and interpolation of the
control associated with the union of the space-time adaptive state and adjoint state (right)

We compare the influence of the input control for snapshot generation in Figure 21. It turns
out, that for this problem setting, the input control has no influence on the approximation quality.

10 4 10 0

10 2
10 -1
uniform:
10 0
—: u = 0
state adaptive:
10 -2
—: u = 0, x: u = Iuh+
10 -2
adjoint adaptive:
—: u = 0, x: u = Iuh+
state and adjoint adaptive:
10 -4 10 -3
0 20 40 60 80 100 0 20 40 60 80 100 —: u = 0, x: u = Iuh+

Figure 21: Relative L2 (0, T ; Ω)-errors between the analytical optimal solution and the reduced-
order solution depending on n and comparing an input control u = 0 with an input control
u = Iuh+ for snapshot generation. Left: error in the state, right: error in the control
3.7 Located control and control constraints 57

Efficiency

We analyze the proposed methodology concerning computational efficiency versus accuracy in

Table 1. If an accuracy of order 10−03 of the numerical solution in comparison to the true solution
is aimed, then using the adaptive snapshot location tailored for the state (Algorithm 4) chooses
n = 27 and computes the adaptive time grid in 0.61 seconds. The generation of snapshots is
very fast with only 0.002 seconds. For the computation of the POD basis, we solve an eigenvalue
problem for Z̄ T Z̄ = D1/2 ZW ZD1/2 , where Z denotes the matrix in which the state and adjoint
snapshots are stored columnwise, D = diag(α0 , . . . , αn ) is a diagonal matrix containing the time
weights and W is a matrix related to the inner product (·, ·)X in (2.7), i.e. for this setting with
X = L2 (Ω) it is the mass matrix. This eigenvalue problem for Z̄ T Z̄ is the discrete counterpart
of a snapshot gramian which will be discussed in Section 4.2.2 in detail. Note that the dimension
of Z̄ T Z̄ only depends on the number of time instances n + 1, such that the solution of the
related eigenvalue problem is very cheap with only 0.09 seconds. Then, the online time for the
solution of the POD reduced optimal control problem is only 0.01 seconds. Thus, the major
time consumption lies in the construction of the time-adaptive grid. However, this effort pays
off if we consider a comparison using a uniform time discretization. In order to achieve the
same accuracy as in the adaptive reduced-order computation, we need in a full finite element
solution n = 30150 time instances, which leads to a computational time of 10.30 seconds. So,
the reduced-order model with adaptive snapshot location and including offline times leads to a
speed up of factor 14.5.
In comparison, using a uniform snapshot location for POD reduced-order modeling leads to the
following computational issues. Utilizing the snapshot gramian Z̄ T Z̄ for computing the POD
basis reaches its limits at n = 17500 and results in an accuracy of erry = 1.50 · 10−02 , where the
solution of the eigenvalue problem takes 70.18 seconds. For larger numbers of time instances, the
solution of the eigenvalue problem is not possible on our computing machines due to memory
problems. As an alternative, we can compute a POD basis from a singular value decomposition
of the snapshot matrix which reaches its memory limits for n = 22500 leading to an accuracy of
erry = 9.71 · 10−03 , where the time to compute the POD basis is 11.14 seconds. Another option
is to solve an eigenvalue problem for Z̄ Z̄ T = W 1/2 ZDZW 1/2 which is the discrete analogon to
the operator R in (2.10). The solution of an associated eigenvalue problem with n = 30150 takes
0.21 seconds and leads to a POD reduced-order accuracy of 5.77 · 10−03 . Note carefully that
this method involves the computation of an inverse weighting matrix W 1/2 which is feasible in
this setting due to the discretization level h∗ = 1/100. However, if we consider a finer spatial
resolution or a 2d or 3d setting, the inversion of a the matrix W might become intensive.
Moreover, we observe that the use of the adaptive snapshot grid for the solution of the reduced-
order model in the online phase brings a speed up in comparison to the full-order computation
(factor 1030) and the reduced-order solution (factor 51) in case of a uniform time grid.

3.7 Located control and control constraints

In many applications it is not possible to control the system everywhere in the spatial domain.
For this reason, the control is often considered as a variable which only depends on time and is
matched to the time-space domain via a control operator B. We refer to this as located control.
Let us consider the space U = L2 (0, T ; Rm ), m ∈ N, as control space. We introduce an admissible
set of controls by

Uad = {u ∈ L2 (0, T ; Rm ) : ua (t) ≤ u(t) ≤ ub (t) in Rm a.e. in [0, T ]}, (3.53)

with box constraints given by the lower and upper bounds ua , ub ∈ L∞ (0, T ; Rm ) with ua (t) ≤
ub (t) almost everywhere in (0, T ). Note that Uad is a non-empty, bounded, convex and closed
58 3 ADAPTIVITY WITH RESPECT TO TIME

uniform, FE uniform, ROM uniform, ROM adaptive, ROM

solve eigs(Z̄ Z̄ T ) solve eigs(Z̄ T Z̄) solve eigs(Z̄ T Z̄)
n 30150 30150 17500 27
erry 5.77 · 10−03 5.77 · 10−03 1.50 · 10−02 5.77 · 10−03
optimal control solution 10.30 sec 0.51 sec 0.40 sec 0.01 sec
snapshot generation – 0.75 sec 0.43 sec 0.002 sec
POD basis – 0.21 sec 70.18 sec 0.09 sec
time-adaptive grid – – – 0.61 sec

Table 1: Accuracy and computational times for the solution of the optimal control problem using
a uniform versus an adaptive time discretization according to Algorithm 4 comparing a full finite
element method with spatial discretization h∗ = 1/100 with a POD reduced-order model using
` = 1 POD basis function. The POD basis is computed solving a weighted eigenvalue problem
(eigs) for Z̄ Z̄ T or Z̄ T Z̄

subset of L∞ (0, T ; Rm ). We introduce a control operator as

m
X
2 −1
B : U → L (0, T ; H (Ω)), u 7→ Bu = ui χi ,
i=1

where χi ∈ H −1 (Ω), 1 ≤ i ≤ m, denote given shape functions. Thus, B is a linear and bounded
operator. The dual operator B ∗ : L2 (0, T ; H −1 (Ω))∗ → U ∗ fulfills
Z T
hBu, viL2 (0,T ;H −1 (Ω)),L2 (0,T ;H01 (Ω)) = hBu(t), v(t)iH −1 (Ω),H01 (Ω) dt
0
Xm Z T
= ui (t)hχi , v(t)iH −1 (Ω),H01 (Ω) dt
i=1 0
m Z
X T
= ui (t)(B ∗ v(t))i dt
i=1 0

= (u, B ∗ v)L2 (0,T ;Rm ) = hu, B ∗ viU

according to Definition 2.9. Thus, it holds (B ∗ v(t))i = hχi , v(t)iH −1 (Ω),H01 (Ω) . Note that we
identify L2 (0, T ; H −1 (Ω))∗ with L2 (0, T ; H01 (Ω)) and U ∗ with U . In the state equation (3.1) we
replace the distributed control by the located control as follows

yt − ∆y = Bu in (0, T ] × Ω, (3.54a)
y=0 on [0, T ] × ∂Ω, (3.54b)
y(0, ·) = y0 in Ω. (3.54c)

Existence of a unique weak solution y ∈ W (0, T ; H01 (Ω)) to (3.54) and the regularity results
follow according to the theory for (3.1). We introduce the weak solution operator associated
with equation (3.54) by

Sad : Uad × L2 (Ω) → W (0, T ; H01 (Ω)), (u, y0 ) 7→ y := Sad (u, y0 ). (3.55)

As cost functional we consider the quadratic objective

1 α
Jad (y, u) = ky − yd k2L2 (0,T ;Ω) + kuk2L2 (0,T ;Rm ) .
2 2
The optimal control problem reads as

min Jad (y, u) s.t. y = Sad (u, y0 ). (3.56)

(y,u)∈W (0,T ;H01 (Ω))×Uad
3.7 Located control and control constraints 59

Due to the control constraints, the gradient equation (3.11) in the optimality system is replaced
by the variational inequality
(αu + B ∗ p, v − u)U ≥ 0 for all v ∈ Uad . (3.57)
The resulting first-order necessary (and by convexity sufficient) optimality system is given by
the state equation (3.54) together with the adjoint equation
−pt − ∆p = y − yd in [0, T ) × Ω, (3.58a)
p=0 on [0, T ] × ∂Ω, (3.58b)
p(T, ·) = 0 in Ω, (3.58c)
and the variational inequality (3.57) which is equivalent to the projection formula
 
1 ∗
u(t) = P[ua (t),ub (t)] − (B p)(t) for almost all t ∈ [0, T ], (3.59)
α
where we define by
P[ua (t),ub (t)] : Rm → [ua (t), ub (t)]
for almost all t ∈ [0, T ] the componentwise orthogonal projection onto the admissible set of
control vectors. The projection can be computed as
    
1 ∗ 1
P[ua (t),ub (t)] − (B p)(t) = max ua (t), min ub (t), − hχi , p(t)iH −1 (Ω),H01 (Ω) ,
α α
which is meant componentwise for almost all t ∈ [0, T ]. Moreover, we denote by
P[ua ,ub ] {u} := P[ua (t),ub (t)] {u(t)} , for almost all t ∈ [0, T ]
the pointwise projection of a function u ∈ L∞ (0, T ; Rm ) onto the set of admissible controls
ua , ub ∈ L∞ (0, T ; Rm ).

Reformulation of the optimality system with respect to the adjoint state

We now can only reduce the optimality system to an elliptic system in p. The derivation of the
biharmonic system is given later in the proof of Theorem 3.16. The resulting biharmonic system
reads as
 
2 1 ∗
−ptt + ∆ p − B P[ua ,ub ] − B p = − (yd )t + ∆yd in (0, T ) × Ω, (3.60a)
α
p=0 on [0, T ] × ∂Ω, (3.60b)
∆p = yd on [0, T ] × ∂Ω, (3.60c)
p(T, ·) = 0 in Ω, (3.60d)
(pt + ∆p)(0, ·) = yd (0, ·) − y0 in Ω. (3.60e)
Note that a reduction with respect to the state variable is not possible in the standard way since
the projection operator P[ua ,ub ] is a non-smooth operator, thus non-differentiable in the classical
sense. One possible way to derive a biharmonic equation for the state variable is to consider
a regularization of the projection, see [139]. This option is not further discussed here and we
continue with the focus on the biharmonic equation with respect to the adjoint variable.
 
1 ∗
Lemma 3.14. Let us denote N (p) := −B P[ua ,ub ] − B p . It holds true
α
Z TZ
(N (p1 ) − N (p2 ))(p1 − p2 )dxdt ≥ 0 ∀p1 , p2 .
0 Ω
60 3 ADAPTIVITY WITH RESPECT TO TIME

Proof. Since the projection operator is monotone increasing, i.e.

u1 (t) ≤ u2 (t) ⇒ P[ua (t),ub (t)] {u1 (t)} ≤ P[ua (t),ub (t)] {u2 (t)} , t ∈ [0, T ]

for u1 , u2 ∈ L∞ (0, T ; Rm ) it is easy to see that

P[ua (t),ub (t)] {u1 (t)} − P[ua (t),ub (t)] {u2 (t)}, u1 (t) − u2 (t) Rm ≥ 0

We set u1 (t) := − α1 (B ∗ p1 )(t) and u2 (t) := − α1 (B ∗ p2 )(t) for p1 , p2 ∈ H̃02,1 (0, T ; Ω). Then it holds
     
1 ∗ 1 ∗ ∗ ∗
−P[ua (t),ub (t)] − (B p1 )(t) + P[ua (t),ub (t)] − (B p2 )(t) , (B p1 )(t) − (B p2 )(t) ≥ 0.
α α Rm

Hence by linearity of B ∗ it follows

   
1 ∗ 1 ∗
h−BP[ua (t),ub (t)] − (B p1 )(t) + BP[ua (t),ub (t)] − (B p2 )(t) , p1 (t) − p2 (t)iH −1 (Ω),H01 (Ω) ≥ 0
α α

which leads to the claim.

Thus, the quantity

 
1
−BP[ua ,ub ] − B ∗ p
α
is monotone in p and (3.60) is a semilinear elliptic problem with monotone nonlinearity.

In order to derive the weak formulation for (3.60) we define the operator Aad
0 as

2,1 2,1
Aad
0 : H̃0 (0, T ; Ω) × H̃0 (0, T ; Ω) → R,

Z T Z     Z
1 ∗
Aad
0 (p, v) = pt vt − B P[ua ,ub ] − B p v + ∆p∆v dxdt + ∇p(0)∇v(0)dx.
0 Ω α Ω

Then the weak form of equation (3.60) reads as

Aad
0 (p, v) = L0 (v) ∀v ∈ H̃02,1 (0, T ; Ω), (3.61)

where the linear form L0 is defined in (3.22).

Theorem 3.15. (Existence of a unique weak solution). For every L0 ∈ (H̃02,1 (0, T ; Ω))∗ the
equation (3.61) has a unique solution p ∈ H̃02,1 (0, T ; Ω).

Proof. In analogy to [139, Lemma 4.6] it can be shown that the operator Aad
0 is strongly mono-
tone, coercive and hemi-continuous. Then, the main theorem on monotone operators (Theo-
rem 2.25) ensures existence of a unique solution.

Equivalence of the optimality system to a biharmonic equation

Theorem 3.16. Let (y, u) ∈ W (0, T ; H01 (Ω)) × Uad denote the solution to problem (3.56) with
associated adjoint state p ∈ W (0, T ; H01 (Ω)). Assume that y0 ∈ H01 (Ω) and yd ∈ L2 (0, T ; H 2 (Ω))∩
H 1 (0, T ; L2 (Ω)). Further we assume the compatibility condition y(T ) − yd (T ) + ∆p(T ) ∈ H01 (Ω)
to hold true. Then p satisfies (3.60) a.e. in space-time and is a weak solution to (3.60).
3.7 Located control and control constraints 61

Proof. Due to y0 ∈ H01 (Ω) we follow by Remark 3.2 that y ∈ L2 (0, T ; H 2 (Ω)) ∩ H 1 (0, T ; L2 (Ω)).
Then, we can apply Remark 3.3 which leads to the regularity

p ∈ L2 (0, T ; H 4 (Ω)) ∩ H 1 (0, T ; H 2 (Ω)) ∩ H 2 (0, T ; L2 (Ω)).

Thus, we can take in (3.58a) the derivative with respect to time in order to get

−ptt − ∆pt = yt − (yd )t .

Then, we replace yt using (3.54a) and obtain

−ptt − ∆pt = ∆y + Bu − (yd )t .

We use the projection formula (3.59) in order to eliminate the control variable and get
 
1 ∗
−ptt − ∆pt = ∆y + BP[ua ,ub ] − B p − (yd )t .
α

Finally, we use the identity y = −pt −∆p+yd from (3.58a) in order to eliminate the state variable
and get  
1 ∗
−ptt − ∆pt = ∆(−pt − ∆p + yd ) + BP[ua ,ub ] − B p − (yd )t ,
α
which implies (3.60a). We get the initial condition (3.60e) by evaluating (3.58a) at t = 0. The
boundary condition (3.60c) is derived by evaluating (3.58a) on the boundary ∂Ω.
Thus, the adjoint function p ∈ L2 (0, T ; H 4 (Ω)) ∩ H 1 (0, T ; H 2 (Ω)) ∩ H 2 (0, T ; L2 (Ω)) fulfills the
biharmonic problem (3.60) a.e. in space-time. Then, p ∈ H̃02,1 (0, T ; Ω) is a weak solution to
(3.60).

Since the biharmonic equation (3.60) has a unique weak solution according to Theorem 3.15 and
the weak solution p to (3.58) is a weak solution to (3.60) according to Theorem 3.16, the other
direction holds, i.e. the weak solution to (3.60) is the optimal adjoint state p.

A-posteriori error estimate for the time discretization

We consider the same notation as in Section 3.4. The time-discrete problem for (3.61) reads as
follows: find pk ∈ V˜k with pk (T, ·) = 0 and

˜
0 (pk , vk ) = L0 (vk ) ∀vk ∈ Vk .
Aad (3.62)

The time-discrete problem has a unique solution by construction. We derive and prove a temporal
residual-type a-posteriori error estimate for the variable p in the spirit of [82]. For this, we need
the following results and properties.

Lemma 3.17. For α > 0 and v ∈ H̃02,1 (0, T ; Ω) it holds

 
2 2
kvkH 2,1 (0,T ;Ω) ≤ c kvt kL2 (0,T ;Ω) + k∆vkL2 (0,T ;Ω)

with constant c > 0.

Proof. We define the function g := −vt − ∆v ∈ L2 (0, T ; Ω). Then, v is the solution to

−vt − ∆v = g in [0, T ) × Ω,
v = 0 on [0, T ] × ∂Ω,
v(T, ·) = 0 in Ω,
62 3 ADAPTIVITY WITH RESPECT TO TIME

in a weak sense. This solution v depends continuously on the data (compare e.g. (3.14)), thus

kvk2H 2,1 (0,T ;Ω) ≤ ckgk2L2 (0,T ;Ω)

= ck − vt − ∆vk2L2 (0,T ;Ω)
 
≤ 2c kvt k2L2 (0,T ;Ω) + k∆vk2L2 (0,T ;Ω) ,

where we applied Young’s inequality, Theorem 2.31.

Lemma 3.18. It holds

Aad ad
0 (p, vk ) − A0 (pk , vk ) = 0 ∀vk ∈ V˜k .

Proof. With (3.61) and (3.62) we have

˜ 2,1
0 (p, vk ) − A0 (pk , vk ) = L0 (vk ) − L0 (vk ) = 0 ∀vk ∈ Vk ⊂ H̃0 (0, T ; Ω).
Aad ad

The following theorem states a temporal residual-type a-posteriori estimate for p which transfers
the result in Theorem 3.7 to the case of located control and control constrained optimal control.

Theorem 3.19. Let p ∈ H̃02,1 (0, T ; Ω) and pk ∈ V˜k denote the solutions to (3.61) and (3.62),
respectively. Then, we obtain the a-posteriori error estimation

kp − pk k2H 2,1 (0,T ;Ω) ≤ c ηad

2
, (3.63)

where c > 0 and

n Z   2
X 1
2
ηad = ∆t2j −(yd )t + ∆yd + (pk )tt + B PUad − B ∗ pk − ∆2 p k
Ij α L2 (Ω)
j=1
n Z
X
+ kyd − ∆pk k2L2 (∂Ω) .
j=1 Ij

Proof. We set ep := p − pk , N (p) := −B PUad − α1 B ∗ p and denote by πep ∈ V˜k the Lagrange


interpolation of ep . We make the following estimations

1
kp − pk k2H 2,1 (0,T ;Ω) ≤ k(p − pk )t k2L2 (0,T ;Ω) + k∆(p − pk )k2L2 (0,T ;Ω) (Lemma 3.17)
c
Z TZ
≤ (p − pk )t ept + ∆(p − pk )∆ep + (N (p) − N (pk ))ep dxdt
0 Ω
(Lemma 3.14)
Z T Z
≤ (p − pk )t ept + ∆(p − pk )∆ep + (N (p) − N (pk ))ep dxdt
0 Ω
+k∇(p − pk )(0)k2L2 (Ω)

= Aad p ad p
0 (p, e ) − A0 (pk , e )

= Aad p ad p ad p ad p
0 (p, e ) − A0 (pk , e ) − (A0 (p, πe ) − A0 (pk , πe )) (Lemma 3.18)

= Aad p p ad p p
0 (p, e − πe ) − A0 (pk , e − πe )

= L0 (ep − πep ) − Aad p p

0 (pk , e − πe )
3.7 Located control and control constraints 63

Z T Z Z
= (−(yd )t + ∆yd )(e − πe )dxdt − (yd (0) − y0 )(ep (0) − (πep )(0))dx
p p
0 ZΩ Z Ω
T Z
+ yd ∇(ep − πep ) · ~n dsdt − ∇pk (0)∇(ep − πep )(0)dx
Z0 T Z∂Ω Ω

− ((pk )t (e − πe )t + N (pk )(e − πep ) + ∆pk ∆(ep − πep ))dxdt

p p p

Z T Z0 Ω
= (−(yd )t + ∆yd + (pk )tt − N (pk ) − ∆2 pk )(ep − πep )dxdt
0 ZΩ Z
T
+ (yd − ∆pk )∇(ep − πep ) · ~n dsdt (integration by parts and
0 ∂Ω
Theorem 2.20).
Splitting the temporal integration into an integration over each time interval, using error es-
timates of the Lagrange interpolation π, the trace inequality and Young’s inequality (Theo-
rem 2.31), we get the claim.

Space-time mixed finite element discretization

We state the space-time mixed finite element variational form for (3.60) which is derived along the
lines to (3.33) and uses the same notations. It reads as: find ph ∈ Ỹ0h , ϑh ∈ V˜k and ϑh |∂Ω = −ỹd
such that
Z TZ   Z
1 ∗
(ph )t vt − BP[ua ,ub ] − B ph v + ∇ϑh ∇vdxdt + ∇ph (0)∇v(0)dx
0 Ω α Ω
Z TZ Z
= (−(yd )t + ∆yd )v dxdt − (yd (0) − y0 )v(0)dx ∀v ∈ Ỹ0h ,
0 Ω Ω
(3.64a)
Z T Z
−∇ph ∇φ + ϑh φ dxdt = 0 ∀φ ∈ W0h .
0 Ω
(3.64b)

The adaptive finite element cycle which utilizes the error estimate (3.63) in order to adapt the
time discretization according to the temporal dynamics of the adjoint state follows along the
lines of Section 3.5 and is summarized in Algorithm 8.

Algorithm 8 Space-time finite element algorithm with time adaptivity for the adjoint state in
case of control constraints and located control.
Input: Initial space-time mesh Thinit , tolerance ε > 0, refinement parameter θ ∈ (0, 1), desired
state yd , initial state y0 , regularization parameter α > 0, initial number of time discretization
points n, maximal number of time points nmax .
Output: Time-adapted space-time mesh Th , adjoint solution ph .
1: Th = Thinit with n time points
2: while n < nmax OR ηp ≥ ε do
3: Define V˜k as the finite element space corresponding to Th according to (3.30).
4: Solve (3.64).
5: Estimate the error contributions from (3.63).
6: Mark time intervals according to the Dörfler criterion (Definition 2.34) with parameter θ.
7: Refine the marked time intervals using bisection.
8: end while
64 3 ADAPTIVITY WITH RESPECT TO TIME

Remark 3.20. Note that (3.64) is a nonlinear non-smooth equation due to the projection such
that the solution can be done by a semi-smooth Newton method or a Newton method utilizing a
regularization of the projection formula, for example. In the numerical tests, we will use a fixed
point iteration in order to find an approximate solution.

3.8 Snapshot location in optimal control for POD model order reduction with located
control and control constraints
3.8.1 Methodology
The methodology of selecting appropriate time instances related to the optimal adjoint works
analogously to the approach (ii) in Section 3.6 tailored for the adjoint state with the differ-
ence that instead of the biharmonic system (3.21) with error estimate (3.27) we now consider
the equation (3.60) with error estimate (3.63). The approach is summarized in the following
Algorithm 9.

Algorithm 9 Adaptive snapshot selection for linear optimal control problems with located
control and control constraints.
Input: Coarse uniform spatial grid with resolution h+ .
Output: Snapshot locations {tj }nj=0 , approximation of the optimal control uh+ .
1: Call Algorithm 8 in order to solve (3.60) adaptively w.r.t. time using the error estimator
(3.63) with spatial resolution h+ . Obtain time instances {tj }nj=0 = I and the solution ph+ .
 1 ∗
2: Set uh+ = P[ua ,ub ] − α (B ph+ ) according to (3.59).

After the selection of suitable time instances for snapshot generation which are related to the
optimal adjoint state and thus related to the optimal control, we proceed with the usual POD
offline phase, in which state and adjoint state snapshots are sampled and a POD basis is com-
puted. Finally, the POD online phase is executed in order to compute a POD reduced-order
solution to the optimal control problem (3.56), which reads as
` `
min Jad (y` , u` ) s.t. y` = Sad (u` , ỹ0 ), (3.65)
(y` ,u` )∈W (0,T ;X` )×Uad

where the linear operator

`
Sad : Uad × X` → W (0, T ; X` ), (û, ỹ0 ) 7→ y` := S` (û, ỹ0 )

is defined analogously to the solution operator Sad defined in (3.55) associated with the reduced
state equation given by
d
(y` (t), ψ)L2 (Ω) + (∇y` (t), ∇ψ)L2 (Ω) = (Bu` (t), ψ)L2 (Ω) for t ∈ [0, T ], ψ ∈ X` . (3.66)
dt
The procedure is summed up in Algorithm 10.

Algorithm 10 POD model order reduction for linear-quadratic optimal control problems with
control constraints and located control using adaptive snapshot locations.
Input: Snapshot locations {tj }nj=0 = I and input control uh+ from Algorithm 9, fine spatial
resolution h∗ , number of POD basis functions `
Output: y` , u`
1: Sample state and adjoint snapshots in a simulation of (3.54) and (3.58) with spatial resolution
h∗ and input control Iuh+ on I.
2: Compute a POD basis of rank ` according to (2.7) using an integrated snapshot ensemble.
3: Set up and solve the POD-ROM (3.65) on I to obtain y` , u` .
3.8 Snapshot location in optimal control for POD model order reduction with located control and control
constraints 65

3.8.2 Error analysis

Analogously to the error analysis in Section 3.6.2, we derive error estimates for the reduced-order
solution computed with Algorithm 10. The notation follows along the lines. In particular, we
denote by (ȳ, ū) the solution to (3.56) with associated optimal adjoint
 1 ∗state p̄. The time-discrete
solution to (3.62) is denoted by pk and we define uk := P[ua ,ub ] − α B pk .

Error bound for the adjoint state

Let us start with estimating the error between the optimal adjoint state and the time-discrete
POD approximation. Using the null-trick and applying the triangle inequality we arrive at

kp̄ − p̄`k k ≤ kp̄ − pk k + kpk − p̃k (uk )k + kp̃k (uk ) − p̃`k (uk )k + kp̃`k (uk ) − p̄`k k. (3.67)
| {z } | {z } | {z } | {z }
(3.67.i) ([Link]) ([Link]) ([Link])

The first term (3.67.i) can be estimated utilizing the a-posteriori error estimate (3.63) for the
adjoint state and concerns the snapshot location. It holds true

kp̄ − pk k2 ≤ c1 ηad
2
. (3.68)

The estimation of the following two terms ([Link]) and ([Link]) are in line with the estimation
for ([Link]) and ([Link]). Finally, the last term ([Link]) can be estimated in terms of the control
variable using the a-posteriori error estimation from [177]. In particular, it holds

kp̃`k (uk ) − p̄`k (ū`k )k ≤ ckuk − ū`k kL2 (0,T ;Rm ) ≤ ckuk − ūkL2 (0,T ;Rm ) + ckū − ū`k kL2 (0,T ;Rm ) , (3.69)

with constant c > 0 from (3.14). The error between the (unknown) optimal control ū and a
particular control up (here up = uk and up = ū`k , respectively) can be estimated via
1
kū − up kL2 (0,T ;Rm ) ≤ kζp kL2 (0,T ;Rm ) , (3.70)
α
where the perturbation ζp ∈ L2 (0, T ; Rm ) is chosen such that

(αup + B ∗ pp + ζp , v − up )L2 (0,T ;Rm ) ≥ 0 for all v ∈ Uad

holds true where pp denotes the adjoint state associated with up .

In conclusion, we establish the following theorem.
Theorem 3.21. Let p̄ denote the optimal adjoint state associated with (3.56) and let p̄`k de-
note the time-discrete adjoint solution associated with (3.65). The POD basis is computed from
an integrated snapshot set with input control and snapshot locations computed according to the
Algorithm 9. In (2.5) it is either X = V or X = H. Then, it holds true
!1/2
√ X
kp̄ − p̄`k kL2 (0,T ;X) ≤ c1 ηad + ε + c2 λi + c2 k(p̃k (uk ))t − P ` (p̃k (uk ))t k2L2 (0,T ;V ∗ )
i>`
c3
+c2 kỹk (uk ) − ỹk` (uk )k2L2 (0,T ;H) +
(kζ1 kL2 (0,T ;Rm ) + kζ2 kL2 (0,T ;Rm ) ),
α
(3.71)
2 m
with constants c1 , c2 , c3 > 0 and ζ1 , ζ2 ∈ L (0, T ; R ) are chosen such that

(αuk + B ∗ p̃k (uk ) + ζ1 , v − uk )L2 (0,T ;Rm ) ≥ 0 and (αū`k + B ∗ p̃k (ū`k ) + ζ2 , v − ū`k )L2 (0,T ;Rm ) ≥ 0

for all v ∈ Uad , respectively, is fulfilled.

Proof. Utilizing the decomposition (3.67) together with (3.68), (3.38b), (3.48), (3.49) and (3.70)
lead to the claim.
66 3 ADAPTIVITY WITH RESPECT TO TIME

Remark 3.22. Note that in the error estimate (3.71) in Theorem 3.21, there exists a dependency
on the time derivatives, for which we refer to Remark 3.11(i). Moreover, the dependency on the
state kỹk (uk )− ỹk` (uk )k2L2 (0,T ;H) can be estimated analogously to the term ([Link]). This concerns
the estimation between the state snapshots and the POD state associated with the control uk ,
which is the input control in order to generate the snapshots.

Error bound for the control

The error regarding the control can be estimated using the a-posteriori error estimate from [177].
Moreover, we can estimate the error in the control variable with regard to the error in the adjoint
variable. In particular, due to the optimality condition (3.57) it holds

(αū + B ∗ p̄, v − ū)L2 (0,T ;Rm ) ≥ 0 ∀v ∈ Uad , (3.72a)

(αū`k + B ∗ p̄`k , v − ū`k )L2 (0,T ;Rm ) ≥ 0 ∀v ∈ Uad . (3.72b)

Taking v = ū`k as admissible test function in (3.72a) and v = ū as admissible test function in
(3.72b) and adding both inequalities leads to

(α(ū − ū`k ) + B ∗ (p̄ − p̄`k ), ū`k − ū)L2 (0,T ;Rm ) ≥ 0.

Thus, we can infer the estimation

1 ∗
kū − ū`k kL2 (0,T ;Rm ) ≤ kB (p̄ − p̄`k )kL2 (0,T ;Rm ) .
α
Error bound for the state

Regarding the error in the state variable kȳ − ȳk` k we cannot guarantee that the time grid con-
structed via Algorithm 9 is a suitable time grid for the state variable. Since it is not possible to
reformulate the optimality system into a biharmonic equation depending on the state and derive
corresponding error estimates, we introduce a post-processing step. In this post-processing step,
we undertake a further refinement of the adaptive time grid from Algorithm 8 in order to make
it suitable for the state. If we take X = V in (2.7), then in this situation, the following estimate
is proved in [121]:
Xn n 
X 
αj ky(tj ; u`k ) − yj` (u`k )k2H ≤ ∆t2j Cy ((1 + c2p )kytt (u`k )k2L2 (Ij ,H) + kyt (u`k )kL2 (Ij ;V ) )
j=0 j=1
(3.73a)
n d
!
X X
|hψi , y0 iV |2 + λi


+ Cy (3.73b)
j=1 i=`+1
n X
d
X λi
+ Cy (3.73c)
j=1 i=`+1
∆t2j

where Cy > 0 is a constant depending on T , but independent of the time grid {tj }nj=0 . We note
that y(tj ; u`k ) is the continuous solution of (3.54) at given time instances related to the suboptimal
control u`k . The constant cp is an upper bound of the projection operator. A similar estimate can
be carried out if X = H. Estimate (3.73) provides now a recipe for further refinement of the time
grid in order to approximate the state y within a prescribed tolerance. One option here consists
in equidistributing the error contributions of the term (3.73a), while the number of modes has to
be adapted to the time grid size according to the term (3.73c). Finally, the number ` of modes
should be chosen such that the term in (3.73b) remains within the prescribed tolerance.
3.8 Snapshot location in optimal control for POD model order reduction with located control and control
constraints 67

3.8.3 Numerical results

In order to validate the snapshot location strategy, we construct a numerical example such that
we know the analytical solution. Similar examples are presented in [11]. All coding is done
in MATLAB R2018a. The solution of the nonlinear, non-smooth biharmonic equation can be
done by a semi-smooth Newton method or by a regularization of the projection formula (Re-
mark 3.20). In this numerical test we compute an approximate solution to the biharmonic system
with a fixed point iteration and initialize the method with the adjoint state corresponding to the
unconstrained optimal control problem. In this way, only two iterations are needed for conver-
gence. Convergence of the fixed point iteration can be argued for large enough values of α.

Problem setting

The spatial domain is chosen to be Ω = (0, 1) ⊂ R. We fix the end time T = 1 and choose
the regularization parameter α = 1. As in Section 3.6.3, we introduce a forcing term f on the
right-hand side of the state equation (3.54) leading to a right-hand side f in (3.60a).

As optimal state we choose

ȳ(t, x) = x(x − 1) sin(πt)
and thus, the initial condition is y0 (x) = 0. The optimal adjoint state is set to
!
e(t−1)/ε − e−1/ε
p̄(t, x) = sin(πx) t − .
1 − e−1/ε

For small values of ε (we choose ε = 10−03 ), the adjoint state p̄ develops a layer towards t = 1,
which can be seen Figure 23. We choose m = 2 control shape functions given by

χ1 (x) = max{0, 1 − 16(x − 0.25)2 }, χ2 (x) = max{0, 1 − 16(x − 0.75)2 },

which are shown in Figure 24. This leads to

!
e(t−1)/ε − e−1/ε
Z Z  
32 8
(B ∗ p̄(t))1 = χ1 p̄(t)dx = (B ∗ p̄(t))2 = χ2 p̄(t)dx = 3
− 2 t− .
Ω Ω π π 1 − e−1/ε

We impose control constraints from below and above as ua (t) = −0.1 and ub (t) = +0.1. The
analytical optimal controls are given by
( !)
e(t−1)/ε − e−1/ε
 
1 32 8
ū1 (t) = ū2 (t) = P[−0.1,0.1] − − t−
α π3 π2 1 − e−1/ε

and are shown in Figure 24. The resulting source term f and the desired state yd are constructed
accordingly.

Uniform vs. adaptive snapshot location

In contrast to Algorithm 10, we choose as input controls u1 (t) = u2 (t) = 0 in order to create a
fair comparison between a uniform and adaptive snapshot location strategy which only depends
on the time discretization. We select ` = 1 POD mode and use the same time grid for snapshot
location and optimization.
For n = 40, the adaptive time grid computed according to Algorithm 9 as well as the associated
adjoint state and controls are shown in Figure 25.
We illustrate the POD optimal adjoint state, its contour lines and the POD state computed on a
uniform time mesh and on an adaptive time grid in Figures 26 and 27, respectively. The control
68 3 ADAPTIVITY WITH RESPECT TO TIME

Figure 22: Analytical adjoint state p̄ (left), its contour lines (middle) and analytical state ȳ
(right)

0.8

0.6

0.4

0.2

0
0 0.25 0.5 0.75 1

e(t−1)/ε −e−1/ε
Figure 23: Time layer function t − 1−e−1/ε

0 1 1

-0.025 0.75 0.75

-0.05 0.5 0.5

-0.075 0.25 0.25

-0.1 0 0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

Figure 24: Analytical controls ū1 = ū2 (left), control shape functions χ1 (middle) and χ2 (right)

1 0

1
0.75 -0.025

0.5 0.5 -0.05

0.25 0 -0.075
1
1
0.5
0 0.5 -0.1
0 0.25 0.5 0.75 1 0 0 0.25 0.5 0.75 1
0

Figure 25: Adaptive space-time grid according to Algorithm 9 with n = 40 and h+ = 1/5 (left),
space-time adjoint state ph+ (middle) and corresponding control (uh+ )1 = (uh+ )2 (right)
3.8 Snapshot location in optimal control for POD model order reduction with located control and control
constraints 69

functions are shown in Figure 28. We observe that the use of the uniform time mesh leads to
a reduced-order solution which lacks of an accurate representation of the steep transition at the
end of the time interval due to an insufficient number of time instances. In contrary, the use of
the adaptive time grid leads to a better gradient representation.

0.6 0

0.4 -0.1
0.5
0.2 -0.2

0 -0.3
1 1
1 1
0.5 0.5
0.5 0 0.5
0 0 0.5 1 0
0 0

Figure 26: POD optimal adjoint state computed on an equidistant grid with n = 40 (left), its
contour lines (middle) and the POD optimal state (right)

1 0

-0.1
0.5 0.5
-0.2

0 -0.3
1 1
1 1
0.5 0.5
0.5 0 0.5
0 0 0.5 1 0
0 0

Figure 27: POD optimal adjoint state computed on an equidistant grid with n = 40 (left), its
contour lines (middle) and the POD optimal state (right)

0 0

-0.025 -0.025

-0.05 -0.05

-0.075 -0.075

-0.1 -0.1
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

Figure 28: POD optimal controls u1 = u2 associated with a uniform (left) and an adaptive
(right) time grid for the adjoint state

The relative L2 (0, T ; Ω)- and L2 (0, T ; Rm )-errors between the analytical optimal adjoint and
control solution and the POD solution, respectively, depending on different numbers of time
instances are shown in Figure 29. The inclusion of time adaptivity leads to a notable improvement
of the errors in both adjoint state and the control variable.
Regarding the efficiency, we compare in Table 2 the offline and online computational times which
are spent in order to achieve an accuracy of 1.20 · 10−03 in the control variable. The use of an
adaptive time grid for snapshot generation and POD reduced-order optimization leads to a speed
up of factor 12 including all offline and online times in comparison to the use of a uniform time
grid. Note that for the computation of the POD basis, we solve in the uniform and the adaptive
approach an eigenvalue problem for the weighted matrix Z̄ T Z̄ which is of size (n + 1) × (n + 1).
70 3 ADAPTIVITY WITH RESPECT TO TIME

10 0 10 0

10 -1
10 -1

10 -2

10 -2
10 -3
— uniform
— adjoint adaptive
10 -3 10 -4
0 20 40 60 80 100 0 20 40 60 80 100

Figure 29: Relative L2 (0, T ; Ω)- and L2 (0, T ; Rm )-errors, respectively, between the analytical
optimal solution and the POD optimal solution depending on n. Left: error in the adjoint state,
right: error in the control

uniform, FE uniform, ROM adaptive, ROM

n 7550 7550 40
erru 1.2 · 10−03 1.2 · 10−03 1.2 · 10−03

optimal control solution 5.23 sec 0.58 sec 0.02 sec

snapshot generation – 0.18 sec 0.002 sec
POD basis – 6.34 sec 0.09 sec
time-adaptive grid – – 0.48 sec

Table 2: Solution of the optimal control problem using a uniform versus an adaptive time grid
comparing a full finite element method with h∗ = 1/100 with a POD reduced-order model with
` = 1 POD mode

In Section 3.8.2, we have seen that the adaptive time grid according to Algorithm 9 is related
to the optimal adjoint state and thus it is a suitable time grid for the adjoint state and control.
However, there is no guarantee that this time grid is also suitable with respect to the state. In
fact, Figure 30 shows that for this problem setting the error in the state is very similar in either of
the cases of a uniform or an adaptive time mesh, although the uniform time discretization leads
to slightly better results. For other problem settings, this difference might be more prominent.
We follow the post-processing step according to (3.73) in order to include further time instances
which are related to the state. Table 3 enumerates the improvement of the relative error in the
state variable for increasing numbers of inserted time instances. The computational times for
this post-processing step are very small (≤ 0.1 sec).

10 0
uniform
adjoint adaptive

10 -1

10 -2
— uniform
— adjoint adaptive
10 -3
0 20 40 60 80 100

Figure 30: Relative L2 (0, T ; Ω)-errors between the analytical optimal state and the POD optimal
state depending on n
3.9 Further directions of research 71

# erry time
0 4.63 · 10−02 –
5 4.40 · 10−02 0.03 sec
10 4.31 · 10−02 0.04 sec
20 3.90 · 10−02 0.07 sec
30 3.62 · 10−02 0.1 sec

Table 3: Post-processing: refinement of the adjoint adaptive grid with n = 40 according to (3.73)
by inserting # further time instances

3.9 Further directions of research

Various modifications and extensions of the considered optimal control problem can be investi-
gated which open new research questions. We think of the inclusion of an additional transport
term, for example, or the consideration of state constraints. In this outlook, we want to give
an outline to two examples of further research directions: the consideration of a nonlinear state
equation and the application of the method to the model predictive control (MPC) context.

Nonlinear state equation

We consider the optimal control problem of minimizing the cost functional

1 α
J(y, u) = ky − yd k2L2 (0,T ;Ω) + kuk2L2 (0,T ;Ω)
2 2
subject to the semilinear parabolic equation given by

yt − ∆y + N (y) = u in (0, T ] × Ω, (3.74a)

y=0 on [0, T ] × ∂Ω, (3.74b)
y(0, ·) = y0 in Ω. (3.74c)

Under certain monotonicity, boundedness and Lipschitz continuity assumptions on the nonlin-
earity, it is proved in e.g. [176, Theorem 5.7] that a solution to this optimal control problem
exists. The associated optimality system is given by the state equation (3.74) together with the
adjoint equation

−pt − ∆p + Ny (y)p = y − yd in [0, T ) × Ω, (3.75a)

p=0 on [0, T ] × ∂Ω, (3.75b)
p(T, ·) = 0 in Ω, (3.75c)

and the optimality condition

αu + p = 0 in [0, T ] × Ω. (3.76)
A formal reformulation of the optimality system into a biharmonic equation can now be done
only with respect to the state variable in the following way. We eliminate the control in (3.74a)
using (3.76) and we take the derivative with respect to time to get

1
ytt − (∆y)t + (N (y))t = − pt .
α
Utilizing (3.75a) one arrives at

1 1
ytt − (∆y)t + (N (y))t − y = (∆p − Ny (y)p − yd ).
α α
72 3 ADAPTIVITY WITH RESPECT TO TIME

In order to eliminate the adjoint variable, we use (3.74a) together with (3.76). Considering the
boundary conditions as well as the initial and end time condition, this leads to the nonlinear
biharmonic system
1 1
−ytt + ∆2 y + y − Ny (y)∆y + Ny (y)N (y) − ∆N (y) = yd in (0, T ) × Ω, (3.77a)
α α
y=0 on [0, T ] × ∂Ω, (3.77b)
∆y − N (y) = 0 on [0, T ] × ∂Ω, (3.77c)
y(0, ·) = y0 in Ω, (3.77d)
(yt − ∆y)(T ) + N (y(T )) = 0 in Ω. (3.77e)

It is now interesting to study the regularity of a solution to this equation depending on the non-
linearity. Moreover, since the original optimal control problem is non-convex, it is of interest to
study under which conditions a solution to the optimality system is a solution to the biharmonic
equation and vice versa.
Another approach is to consider a linearization of the optimal control problem. This is of inter-
est in the context of sequential quadratic programming (SQP), for example, where a sequence of
linearized problems is solved in order to find an approximate solution of the nonlinear problem.
This approach is followed in [115].

Adaptive time selection in model predictive control

We aim to investigate different time selection strategies in the context of model predictive control
(MPC). The principal idea in model predictive control is to split the time domain into smaller
time horizons and solve corresponding open-loop control problems successively, see e.g. [88]. This
approach, in particular, is applied to optimal control settings with very large or infinite (receding)
time horizons. A major advantage of the approach is the possibility to react to changes of the
problem data due to external influences.
A crucial challenge within the model predictive concept is the choice of the prediction horizon,
i.e. the time horizon length for each finite horizon open-loop problem. Since the length of the
prediction horizon strongly influences the quality and the computational times and thus the
efficiency of the method, a suitable (i.e. problem-specific) choice is advantageous. We utilize
the residual based time-adaptive cycle which solves the biharmonic equation iteratively and
adapts the time discretization according to an error indicator. In this way, dominant temporal
structures are recognized and determine the selection of appropriate time grid points and time
horizon lengths. Since the resulting solution to the biharmonic system is related to the optimal
solution, it can be used as a warm start in order to initialize the actual MPC iteration.
 73

4 Adaptivity with respect to space

Challenge. In order to build a POD surrogate model for an evolution problem which reflects the
spatial properties of the true solution well, the relevant spatial features need to be represented
within the snapshot set. In the offline phase, the use of adaptive finite elements for snapshot
generation is an efficient spatial discretization approach in order to capture important structures
and guarantee an approximation of the true solution within a prescribed tolerance. In particular,
many practical applications require the use of spatial adaptivity in order to make implementa-
tions feasible. As a consequence, the snapshots can have a varying number of degrees of freedom
and different locations of node points. This does not fit into the usual POD framework which
assumes static snapshots with the same number of degrees of freedom and location of the dis-
cretization points.

Solution idea. We provide an approach to use adaptive finite elements for snapshot genera-
tion in the offline phase of model order reduction utilizing proper orthogonal decomposition
for semilinear parabolic evolution problems. The main idea is to consider the setting from an
infinite-dimensional perspective. In particular, we assemble the snapshot gramian by evaluating
inner products of snapshots explicitly. In this way, the necessity to interpolate the snapshots onto
a reference space is avoided. The use of adaptivity in the offline phase for snapshot generation
leads to an acceleration of the offline computation time while providing a prescribed accuracy.
In the context of optimal control, the use of spatially adapted snapshots leads to a speed up
in the online phase whenever new snapshots have to be generated. Moreover, for an unsteady
incompressible flow problem we derive a stable POD reduced-order model by either using an op-
timal projection onto a weakly divergence-free space or by enriching the reduced velocity space
by pressure supremizer functions.

Solution procedure. Throughout the following, we assume to have a suitable time discretization
scheme at hand and concentrate on the inclusion of spatial adaptivity in the context of POD
based model order reduction. We start with introducing the problem setting of a general ab-
stract semilinear evolution equation and discretize in space using adaptive finite elements. This
is done in a generic framework comprising p-, h- and r-adaptive schemes. Next, we recall the
POD method in a Hilbert space setting. The continuous perspective only requires the snapshots
to belong to a common Hilbert space. This is the key point of the methodological approach. The
POD reduced-order model is constructed using the eigensystem of the snapshot gramian. This
matrix is set up explicitly. In the case of h-adaptive finite elements with nested, hierarchical
grids, the evaluation of the inner products can be done efficiently utilizing usual finite element
software packages. However, in the case of r-adapted meshes, the computation of an inner prod-
uct of snapshots can require the integration over cut finite elements, which is an involving task.
It is realized by a detection of the cut polygon and an integration over it based on a boundary
representation of the integral using Stoke’s formula. The analysis for the error between the true
(unknown) solution to the evolution problem and the POD solution is carried out and includes
not only the POD approximation error but also the finite element error depending on the adap-
tive discretization scheme. In order to illustrate the feasibility of the proposed approach, we
present the test case of the Cahn–Hilliard system utilizing h-adapted hierarchical meshes and
two settings of a linear heat equation using nested and non-nested grids.
In the context of POD model order reduction for unsteady incompressible flows governed by the
Navier–Stokes equations we propose two solution concepts in order to derive a stable reduced-
order model for the case of h-adapted snapshots. The first approach leads to a velocity reduced-
order model and is based on either projecting the snapshots or the POD modes onto a reference
velocity space such that a weak divergence-free property is fulfilled in a common sense. The
second approach yields a velocity-pressure reduced-order model. Stability is guaranteed by en-
riching the velocity reduced space with pressure supremizer functions. In particular, it is shown
74 4 ADAPTIVITY WITH RESPECT TO SPACE

that with this construction an inf-sup condition for the reduced model holds true. We treat the
inhomogeneous problem setting which requires the homogenization of the Navier–Stokes equa-
tions and a construction of a certain lifting function. Both reduced-order models are compared
in a numerical test case of a lid-driven cavity flow.
Finally, we consider the problem setting of an optimal control problem governed by the Cahn–
Hilliard system. The objective is to reach a desired phase distribution by controlling the velocity
which enters the state equation through a transport term. We show existence of an optimal
control and present a numerical test example. The novelty of the approach is the inclusion of
h-adaptive finite elements for snapshot generation in an optimal control framework.

Outline of the chapter

In Section 4.1, an overview of relevant literature considering space-adaptive schemes in the con-
text of model order reduction is given. We investigate the method of proper orthogonal de-
composition for space-adapted snapshots in Section 4.2. After introducing a semilinear abstract
parabolic evolution problem in Section 4.2.1, we recall the POD method in a Hilbert space frame-
work. The construction of the snapshot gramian for space-adapted snapshots is investigated in
Section 4.2.2 and a strategy for the numerical realization is presented. The method is analyzed
concerning the applicability and respective advantages and disadvantages are discussed. In Sec-
tion 4.3, a POD reduced-order model for the abstract semilinear evolution problem is derived
for the case of space-adapted snapshots. We further discuss the treatment of nonlinearities and
the expression of the POD solution in the full spatial domain. In Section 4.4 we provide an
analysis for the error between the true (unknown) solution and the POD reduced-order solu-
tion utilizing space-adapted snapshots. We present in Section 4.5 a numerical example of POD
reduced-order modeling for a linear heat equation and a Cahn–Hilliard model with a polynomial
and a Moreau–Yosida regularized free energy using nested hierarchical adapted meshes for the
snapshots. Moreover, we consider a linear heat equation using meshes with a fixed number but
different locations of node points leading to non-nested grids. The numerical results for the com-
putation of the snapshot gramian in this case using the proposed methodology of Section 4.2.2
is presented. In addition, the combination of POD model order reduction and h-adapted finite
elements is considered for incompressible flow problems governed by the Navier–Stokes equations
in Section 4.6. Two approaches of deriving a stable reduced-order model are proposed and com-
pared. In particular, we propose a velocity-ROM in Section 4.6.3 which is based on an optimal
projection onto a weakly divergence-free space. Further, a velocity-pressure reduced-order model
is constructed in Section 4.6.4 which guarantees stability by enriching the reduced velocity space
with pressure supremizer functions. The problem setting involving inhomogeneous Dirichlet and
initial data is addressed in Section 4.6.5. The reduced-order models are numerically compared
in the test case of a lid-driven cavity flow in Section 4.6.6. In Section 4.7, POD reduced-order
modeling is used in order to find an approximate solution to an optimal control problem governed
by a Cahn–Hilliard equation with transport. The existence of an optimal control is shown in
Section 4.7.1. Spatially h-adapted finite element data are used for the computation of a POD
basis. Finally, a numerical example is carried out in Section 4.7.2.

4.1 Literature overview

Let us give an overview of publications which consider spatial adaptivity in the context of model
order reduction, presented in a chronological order.
In [72], the use of dynamically adaptive meshes is combined with POD based reduced-order
modeling for an unstructured ocean model. The space-adapted snapshots are interpolated from
their own mesh onto a fixed reference mesh. Afterwards, the snapshots are vectors of the same
lengths and the usual POD procedure can be carried out. It is mentioned that a high-order
interpolation approach can be used in order to reduce the interpolation error. The numerical
4.1 Literature overview 75

examples comprise a flow past a cylinder and a gyre.

In [126, Section 2.4.3] an interpolation approach is proposed. The main idea is to interpolate
given space-adapted snapshot data (i.e. solution vectors and respective grids) by polynomials.
For the evaluation of the inner products in the snapshot gramian, an appropriate numerical
integration rule is used. This approach can be seen as a smoothing step, if e.g. noisy data is
provided. In a numerical test in a 1d setting, snapshots with different lengths are constructed by
cancelling out discretization points. A cubic spline interpolation is used and a Gauss–Legendre
quadrature rule is utilized for the numerical integration. For the considered problem setting, the
implementation turns out to be competitive with the standard approach on a static mesh.
In [99] POD based model reduction is considered for free boundary value problems governed by
the Stefan problem. Three different numerical approaches are compared in order to generate
snapshots: first a Landau-type transformation, second a control volume approach and third a
moving mesh approach. For the control volume approach a rescaling and interpolation is used
in order to construct discrete snapshots of the same lengths. In the moving mesh approach,
the location of the grid points is varied according to the evolution of the free boundary. The
grid points in the 1d setting are equidistributed and their movement is governed by an ordinary
differential equation. Using a projection onto a reference domain allows to compute a POD basis
in the usual manner.
In the context of model order reduction utilizing reduced basis methods for parametrized partial
differential equations, an adaptive wavelet discretization is used in [7, 169] for the snapshot com-
putations. A bound for the error between the reduced basis solution and the infinite-dimensional
exact solution is derived. Moreover, a new surrogate for the exact residual and its dual norm
is introduced. Convergence of the resulting adaptive greedy method is shown and numerical
experiments considering a time-periodic space-time convection-diffusion-reaction example are
presented.
In [45] an h-adaptive refinement approach for reduced-order models is proposed. It is an online
adaptive strategy which does not require any additional full-order solutions. The main idea is
to enrich the reduced space online by splitting the reduced basis vectors into separate vectors
and is based on a tree structure constructed via k-means clustering. The splitting is driven by
a dual weighted residual approach. In this way, a sequence of hierarchical reduced subspaces is
constructed and convergence to the full-order model is shown.
The combination of POD model order reduction with h-adaptive finite element snapshots is re-
alized in [181] by constructing common finite element spaces. Two options are considered: either
all snapshots are expressed in terms of a common finite element basis or pairs of snapshots are
expressed in terms of a common finite element basis of these pairs. Moreover, error estimates for
a parametrized elliptic boundary value problem are proved. The approach is illustrated numeri-
cally for a parametrized convection-diffusion equation and a viscous Burgers problem.
In [190, 191, 192], a reduced basis method is developed which is certified by a dual norm bound
of the residual relative to the infinite-dimensional exact solution. In the offline phase spatial
mesh adaptation is used together with a greedy parameter sampling strategy leading to a reli-
able online system. The online bound is independent of the complexity of the underlying mixed
finite element discretization and available for any parameter and not only for those which belong
to the training set.
We note that Section 4.1 to Section 4.5 is based on a collaboration with Michael Hinze and is
published in [84, 85]. Moreover, Section 4.6 is based on a collaboration with Michael Hinze,
Jens Lang and Sebastian Ullmann. A resulting manuscript [86] is accepted for publication in
Advances in Computational Mathematics in 2019. Finally, we note that Section 4.7 is based on
a collaboration with Michael Hinze and Nicolas Scharmacher, is published in [87] and a further
collaboration with Michael Hinze and Jan Oke Alff which is published in [6].
76 4 ADAPTIVITY WITH RESPECT TO SPACE

4.2 POD with space-adapted snapshots

This section introduces the problem setting of a general abstract semilinear parabolic evolution
problem. We first apply a discretization in time which then allows us to use a different finite
element space for spatial discretization at each time instance. Then, we focus on proper orthogo-
nal decomposition in a Hilbert space setting utilizing spatially adapted snapshots. The snapshot
gramian in set up explicitly by evaluating inner products of snapshots which live in a common
Hilbert space. A practical implementation is proposed which realizes this approach.

4.2.1 Abstract semilinear parabolic evolution problem

Problem setting

Let us specify the abstract semilinear parabolic evolution problem which we consider in the fol-
lowing. Let V and H be real separable Hilbert spaces such that there exists a dense and compact
embedding V ,→ H. The dual space H ∗ can be identified with H by the Riesz representation
Theorem 2.8 and the Gelfand triple (V, H, V ∗ ) is formed by

V ,→ H = H ∗ ,→ V ∗ , (4.1)

see Definition 2.14. Since V is continuously embedded in H, there exists a constant cv > 0 such
that
kvk2H ≤ cv kvk2V for all v ∈ V. (4.2)
For a given symmetric, V -elliptic bilinear form a : V × V → R, we assume boundedness, i.e.

∃β>0: |a(u, v)| ≤ βkukV kvkV for all u, v ∈ V (4.3)

and we assume that the Gårding’s inequality is fulfilled (coercivity over V with respect to H),
i.e.
∃ β1 > 0, β2 ≥ 0 : a(u, u) ≥ β1 kuk2V − β2 kuk2H for all u ∈ V.
Let A : V → V ∗ be the bounded linear operator associated with the bilinear form a, i.e. A ∈
L(V, V ∗ ) and
a(u, v) = hAu, viV ∗ ,V = hu, A∗ viV,V ∗ for all u, v ∈ V,
where h·, ·iV ∗ ,V denotes the dual pairing of V ∗ and V , see Definition 2.7, and A∗ is the dual
operator, see Definition 2.9. Moreover, we denote by N : V → V ∗ a nonlinear operator, for which
we assume local Lipschitz continuity. We are concerned with the following Cauchy problem for a
semilinear evolution problem. Let T > 0 be a fixed end time. For a given initial function y0 ∈ H
and external force f ∈ L2 (0, T ; V ∗ ) we consider the problem: find y ∈ W (0, T ; V ) ∩ L∞ (0, T ; Ω)
with
d
(y(t), v)H + a(y(t), v) + hN (y(t)), viV ∗ ,V = hf (t), viV ∗ ,V , (4.4)
dt
and
(y(0), v)H = (y0 , v)H (4.5)
d
for all v ∈ V and for almost all t ∈ (0, T ]. Note that it holds dt (y(t), v)H = hyt (t), viV ∗ ,V for
2 2 ∗
all y ∈ L (0, T ; V ) with yt ∈ L (0, T ; V ) and all v ∈ V in the sense of distributions in (0, T ),
see e.g. [58, Chapter XVIII, §1.2, Proposition 7]. For the following, we use (4.4)-(4.5) as the
governing dynamical model. We assume solvability in Assumption 4.1 and give references for a
detailed analysis in Remark 4.2.

Assumption 4.1. For given f ∈ L2 (0, T ; V ∗ ) and y0 ∈ H we assume the existence of a unique
solution y ∈ W (0, T ; V ) ∩ L∞ (0, T ; Ω) which fulfills the initial condition (4.5) and solves the
semilinear parabolic variational problem (4.4).
4.2 POD with space-adapted snapshots 77

Remark 4.2. For a detailed analysis of the existence of a unique weak solution to a semilinear
parabolic evolution problem of type (4.4)-(4.5) we refer to [176, Lemma 5.3 and 7.12], [155, Theo-
rem 3.1] and [48, Theorem 5.1], where an initial-boundary value problem is considered. Moreover
in [189, Theorem 4.1], existence of a unique solution to a Cauchy problem for a semilinear ab-
stract evolution equation is shown. In [145, Chapter 6] existence of solutions is discussed using
semigroup methods.

Time Discretization

In order to solve (4.4) numerically, we need to introduce a temporal and spatial discretization.
As a first step, we apply a time discretization which then enables us to use a different spatial
discretization for each time instance. For a given number n ∈ N of time steps let

0 = t0 < . . . < tn = T (4.6)

denote a time grid in the time interval [0, T ] with time step sizes

∆tj = tj − tj−1

for j = 1, . . . , n. This partition defines the time intervals Ij = [tj−1 , tj ] for j = 1, . . . , n. We

follow the implicit Euler method, i.e. we integrate the differential equation (4.4) from tj−1 to tj
for j = 1, . . . , n and approximate the integrals involving the solution variable y by a right-sided
rectangle method. Note that the integral on the right-hand side can also be approximated by
the same rectangle method leading to the approximation
Z tj
1
hf (t), viV ∗ ,V dt ≈ hf (tj ), viV ∗ ,V .
∆tj tj−1

Of course, other integration schemes are possible. However, since f is a given function, we leave
the right-hand side as it is in (4.7) at this stage in order to avoid a further error introduced by
the approximation of this integral. The resulting time-discrete system reads as follows: for a
given initial function y 0 := y0 ∈ H find solutions y 1 , . . . , y n ∈ V such that
tj
y j − y j−1
  Z
j j 1
,v + a(y , v) + hN (y ), viV ∗ ,V = hf (t), viV ∗ ,V dt for all v ∈ V. (4.7)
∆tj H ∆tj tj−1

Space Discretization

At each time level, we introduce a spatial discretization for which we use adaptive finite ele-
ments. There are mainly three categories of adaptivity concepts: h-, p- and r-refinement, recall
Section 2.2. All of these adaptivity schemes have the following structure in common. At each
time point tj for j = 0, . . . , n, we introduce a regular triangulation Tj of Ω̄ (Definition 2.32) and
define an Nj -dimensional conformal subspace Vhj of V by

Vhj := span{v1j , . . . , vN
j
j
}⊂V

N
with finite element basis functions {vij }i=1 j
. For Lagrangian polynomials it holds vij (Pkj ) = δik
Nj
for i, k = 1, . . . , Nj with node points {Pkj }k=1 of the underlying triangulation Tj . Higher-order
schemes require additionally derivative information.
At each time level j = 0, . . . , n, the utilized finite element spaces Vh0 , . . . , Vhn can differ both
in the underlying triangulation of the domain Ω̄ and in the polynomial degree. We apply a
finite element Galerkin scheme for (4.7). For this, we us denote by Vh0 the finite element space
78 4 ADAPTIVITY WITH RESPECT TO SPACE

associated with the initial triangulation Th0 . The fully discrete form reads as follows: find a
sequence yh0 ∈ Vh0 , . . . , yhn ∈ Vhn which fulfills
!
yhj − yhj−1 tj
Z
1
,v + a(yhj , v) + hN (yhj ), viV ∗ ,V = hf (t), viV ∗ ,V dt, for all v ∈ Vhj (4.8)
∆tj ∆tj tj−1
H

and
(yh0 , v)L2 (Ω) = (y0 , v)L2 (Ω) ∀v ∈ Vh0 . (4.9)

Note that for j = 1, . . . , n in (4.8) an exact evaluation of the inner product (yhj−1 , v)H of the
previous solution yhj−1 ∈ Vhj−1 and the test function v ∈ Vhj (i.e. an inner product of functions
which live in different finite element spaces) is possible since the functions live in the common
Hilbert space V ⊂ H. A possible implementation of an exact evaluation of such inner products
runs analogously to the evaluation of inner products of snapshots in different finite element spaces
which is explained later. Often, a Lagrange interpolation π j : C 0 (Ω̄) → Vhj is used in order to
interpolate the solution yhj−1 ∈ Vhj−1 ⊂ C 0 (Ω̄) of the previous time instance tj−1 to the new
finite element space Vhj . This usually simplifies the numerical implementation. However it might
introduce an interpolation error which we want to avoid at this stage.
Since for j = 0, . . . , n it holds yhj ∈ Vhj , we make the finite element Galerkin ansatz

Nj
yhj yij vij ∈ Vhj ⊂ V
X
= (4.10)
i=1

j T
with appropriate mode coefficient vectors yj = (y1j , . . . , yN j
) ∈ RNj .

4.2.2 POD method with snapshots in different finite element spaces

The aim of this section is to derive a low-order POD approximation for the infinite-dimensional
evolution problem (4.4) using space-adapted snapshots yh0 ∈ Vh0 , . . . , yhn ∈ Vhn with Vh0 , . . . , Vhn ⊂
V which are solutions to the fully discrete equations (4.8)-(4.9). The main conceptual challenge
is caused by the fact that the snapshots live in different finite element spaces for each time
instance. In order to understand this challenge, let us recall how the POD method works for
snapshots which all belong to the same finite element space, i.e. it holds true Vh0 = · · · = Vhn =: Vh
and yh0 , . . . , yhn ∈ Vh ⊂ V . Using the finite element Galerkin ansatz yhj = N j
P
i=1 yi vi ∈ Vh with
N degrees of freedom, the equation (4.8) can be written as a matrix-vector equation for the
coefficient vectors of the Galerkin ansatz as follows

(M + ∆tj A)yj = M yj−1 + ∆tj b(tj , yj ), (4.11)

where for k, l = 1, R t.j. . , N the following notations are used: Mkl = (vk , vl )H , Akl = a(vk , vl ) and
(b(tj , yj ))l = ∆t1 j tj−1 hf (t), vl iV ∗ ,V dt − hN (yhj ), vl iV ∗ ,V . Note that each finite element coefficient
vector yj has the same dimension N . For the model order reduction of (4.11), the Galerkin
finite element coefficient vectors {yj }nj=0 are taken as snapshots and a finite-dimensional POD is
carried out, e.g. by solving a truncated singular value decomposition for the finite element coeffi-
cient matrix [y0 | · · · | yn ] ∈ RN ×(n+1) . For more details, we refer to [186, Section 1], for example.

Now, the consideration of spatially adaptive finite elements in (4.8) leads to a matrix-vector
equation with time-dependent matrices and vectors as follows

(M j + ∆tj Aj )yj = M j yj−1 + ∆tj bj (tj , yj ), (4.12)

4.2 POD with space-adapted snapshots 79

j
where for k, l = 1, . . . , Nj the following notations are used: Mkl = (vkj , vlj )H , Ajkl = a(vkj , vlj ) and
t
(bj (tj , yj ))l = ∆t1 j tj−1 hf (t), vlj iV ∗ ,V dt − hN (yhj ), vlj iV ∗ ,V . For a clear presentation, we assume
R j

at this point that the coefficient vector yj−1 on the right-hand side of (4.11) is the finite coef-
ficient vector of the Lagrange interpolation of yhj−1 onto the finite element space Vhj . Since the
finite element nodal basis is different for each time instance, the corresponding mode coefficient
vectors {yj }nj=0 have different lengths or different meanings. In particular, in the case in which
the number of node points is the same, the location of the node points might be different, for
example. As a consequence, the consideration of a proper orthogonal decomposition on the level
of the finite element coefficient vectors is no longer possible. The snapshot matrix cannot be set
up directly and the usual POD procedure, which utilizes e.g. a singular value decomposition of
the snapshot matrix, cannot be carried out.

One intuitive option to handle this issue is to interpolate all snapshots onto a reference finite
element space. It is an advantage of this approach that the usual POD procedure can be followed
after the interpolation. However, a major disadvantage is that interpolation errors can occur
when interpolating the snapshots onto a reference finite element space. In particular, important
information about the spatial structure of the snapshots might get lost, for example, if the mesh
of the reference space is coarser or if the polynomial degree of the finite element ansatz functions
is lower. Let us discuss advantages and disadvantages of existing approaches related to this issue.
In order to avoid interpolation errors and ensure an exact representation of the snapshots
in the reference finite element space, a common finite element space for all snapshots is
constructed in [181] for the case of h-adapted snapshots which are generated using hier-
archical nested grids. The construction of a common finite element mesh in this case is
realized by a refinement of the initial mesh according to the adaptive cycle. In this way,
the overlay mesh of all computational meshes is generated, which is the smallest possible
common mesh. This overlay grid is then again a conformal triangulation of the domain
and the corresponding finite element space can be set up. The snapshots are expressed
with respect to the common finite element basis. This approach represents an efficient
handling with usual finite element implementation tools in the case of hierarchical nested
mesh structures. A possible disadvantage appears in the case in which the spatial struc-
tures differ strongly with time. Then, the common finite element space has a very large
dimension. As a consequence, storage problems might occur and/or solving a singular value
decomposition for the corresponding snapshot matrix might be computationally intensive.
For this reason, in [181] it is also considered to construct common finite element spaces
for pairs of snapshots. The presentation of this approach is restricted to h-adaptive finite
elements with hierarchical nested grids.
The construction of a common finite element space in the case of finite element snapshots
with non-nested meshes is quite involved. Let us think of scenarios in which we cannot
restrict the triangulation of the finite element model to nested hierarchical meshes: For
example, if we receive snapshot information from physical applications or measurements
or if black-box tools for mesh adaptation have to be used in order to work with complex
domain geometries. Another situation of non-nested spatial discretizations appears in
the case of r-adaptivity, where the location of node points is moved according to the
spatial structures. In such cases, the overlay of the snapshot meshes might lead to cut
simplices. A subtriangulation can be an option. However, this might also lead to a very
fine common mesh, such that the interpolated snapshots have a very large dimension and
the computation of a POD basis becomes very demanding. Moreover, as pointed out in
[134] in the context of a three-dimensional example, a subtetrahedralization of an arbitrary
polyhedron is challenging and additional vertices might have to be added.
In the case of p-adapted snapshots, the underlying triangulation is fixed such that the
80 4 ADAPTIVITY WITH RESPECT TO SPACE

mesh handling is not an issue. However, the number of node points and the corresponding
restrictions might change according to the polynomial degree of the finite element basis
functions. An interpolation of all snapshots onto the finite element space with highest
polynomial degree is an option and leads to an exact representation of the snapshot ensem-
ble without loss of information. However, it possibly contains an overhead of computation
in the situation in which only a very small number of snapshots has a higher polynomial
degree than the majority of the snapshots.
The goal of this section is to propose a POD reduced-order model that uses a snapshot gramian
in order to construct the POD surrogate model. This approach allows to avoid the necessity of
e.g. interpolating the snapshots into a reference or a common finite element space.
The crucial point of the approach is to consider the POD method from an infinite-dimensional
perspective, where a finite number of snapshots are used which live in a common Hilbert space
X. We can proceed along the lines of Section 2.4, POD in Hilbert spaces. As specified there, the
only assumption that is made on the snapshots is that they need to live in a common Hilbert
space, i.e. y 0 , . . . , y n ∈ X, compare (2.6). In the situation of spatially adapted snapshots this
holds true, since yh0 ∈ Vh0 , . . . , yhn ∈ Vhn with Vh0 , . . . , Vhn ⊂ V , i.e. yh0 , . . . , yhn ∈ V .

As explained in Section 2.4, the idea of the POD method is to describe the space

V := span{yh0 , . . . , yhn } ⊂ V ⊆ X

by means of few orthonormal functions {ψi }`i=1 ⊂ X, with ` ≤ d := dim V ≤ ∞, such that
the error between the snapshots {yhj }nj=0 and the projection of the snapshots onto the subspace
X` := span{ψ1 , ..., ψ` } ⊂ X is minimized in the following mean-square sense

n ` 2

yhj (yhj , ψi )X ψi
X X
min αj − s.t. (ψi , ψj )X = δij for 1 ≤ i, j ≤ `, (4.13)
ψ1 ,...,ψ` ∈X
j=0 i=1 X

compare (2.7), where X can denote V or H and {αj }nj=0 describe non-negative weights. For the
solution of (4.13) we refer to Theorem 2.46. Instead of considering a truncated eigenvalue value
problem for the operator R defined in (2.10) as
n
αj (ψ, yhj )X yhj
X
R := YY ∗ : X → X, ψ 7→ for ψ ∈ X,
j=0

we consider the snapshot gramian K := Y ∗ Y : Rn+1 → Rn+1 . We refer to (2.8) and (2.9) for the
definition of Y and its Hilbert space adjoint Y ∗ , respectively. The action of K is given by
 n 
X√ √
0 j
 αj ( α0 yh , yh )X φj 
 j=0 
 
.
φ 7→ Kφ :=  ..  for φ = (φ0 , . . . , φn ) ∈ Rn+1 .
 
 n 
X √ √ 
j
 αj ( αn yhn , yh )X φj 
j=0

The linear operator K can be represented as the symmetric matrix

√ √ √ √
α0 α0 (yh0 , yh0 )X . . . α0 αn (yh0 , yhn )X

 
 
.. ..
K=  ∈ R(n+1)×(n+1) . (4.14)
 
 . . 
 
√ √ n 0 √ √ n n
αn α0 (yh , yh )X . . . αn αn (yh , yh )X
4.2 POD with space-adapted snapshots 81

We consider the eigenvalue problem of K, i.e. we solve

Kφi = λi φi .
It can be shown that the eigenvalues {λi }di=1 of K coincide with the eigenvalues for R except for
possibly zero, see for example [121]. The corresponding orthonormal eigenvectors {φi }di=1 ⊂ Rn+1
for K are given by
√
(ψi , α0 yh0 )X
 
1 1  ..
φi = √ (Y ∗ ψi ) = √  ∈R
n+1
for i = 1, . . . , d.

λi λi .
√
(ψi , αn yhn )X
This approach is also known as method of snapshots or strobes according to [166, p. 567]. At
this point it is important to have a closer look at the snapshot gramian K. The dimension of
K is (n + 1) × (n + 1), where n denotes the number of snapshots. Thus, the matrix dimension
does not depend on any spatial dimension(s) of the snapshots, it only depends on the number
of utilized snapshots. In order to assemble the matrix K, inner products of snapshots have to
be computed. This is possible since yhi , yhj ∈ V ⊆ X for i, j = 0, . . . , n. Since K is symmetric, it
suffices to compute the upper (or lower) triangular matrix. A practical numerical realization of
the assembly of K is discussed later.
The computation of the snapshot gramian K and its eigenvalue decomposition is sufficient in order
to set up the POD reduced-order model. More details on that will be discussed in Section 4.3.
Finally, the POD modes {ψi }`i=1 can be written as a linear combination of finite element functions
n
1 1 X√
ψi = √ Yφi = √ αj (φi )j yhj ∈ V for i = 1, . . . , `, (4.15)
λi λi j=0

where (φi )j denotes the j-th component of φi ∈ Rn+1 .

To conclude, the snapshot gramian K and its eigenvalue decomposition is used in order to com-
pute the right singular vectors which sizes are independent of the spatial dimension(s) of the
space-adapted finite element snapshots. Thus, this perspective makes the approach independent
of the original spatial discretization method used for snapshot generation.
It remains to clarify how the snapshot gramian K defined in (4.14) can be set up practically.
The next paragraph proposes a possible numerical implementation.

Numerical realization of the POD method utilizing snapshots from different finite element spaces

Let us now turn our perspective to the numerical realization of computing the snapshot gramian
K defined in (4.14) for snapshots which live in different finite element spaces. For each time level
j = 0, . . . , n, the snapshots {yhj }nj=0 are taken from different finite element spaces {Vhj }nj=0 which
belong to a common Hilbert space V . In the fully discrete formulation of the POD method we
are given the evaluation of the snapshots on their corresponding grids, i.e. we are given the mode
coefficient vectors
y 0 ∈ RN 0 , . . . , y n ∈ RN n
of possibly different lengths with yj = (y1j , . . . yN
j T
j
) ∈ RNj , for j = 0, . . . , n. This is why we are
not able to set up the discrete counterpart to the operator R, which is an N × N matrix for
uniform spatial discretization with N degrees of freedom.
We turn our attention to the matrix K ∈ R(n+1)×(n+1) . This matrix dimension only depends on
the number of snapshots and can be computed for any underlying finite element discretization:
the ij−th component Kij , for i, j = 0, . . . , n, is given by
Ni Nj Nj
Ni X
√ √ j √ √ X j j √ √ X
yki ylj (vki , vlj )X .
X
i i i
Kij = αi αj (yh , yh )X = αi αj ( yk vk , yl vl )X = αi αj
k=1 l=1 k=1 l=1
82 4 ADAPTIVITY WITH RESPECT TO SPACE

Thus, for any i, j = 0, . . . , n, the evaluation of the inner product of two finite element snapshots
yhi and yhj is traced back to the evaluation of the inner products of finite element basis functions
(vki , vlj )X . Figure 31 visualizes finite element basis functions with potentially different grids and
polynomial degrees. We discover, then, that the calculation of the matrix K as well as the
determination of its eigenvectors can be done in the case of snapshots which live in different
finite element spaces. Thus, the eigenvectors {φi }di=1 of K are the right singular vectors of Y
and contain the space independent time information. This fact will be used in the following to
build up the reduced-order model. We note that the calculation of the matrix K can be done
for arbitrary finite element spaces, i.e. all kinds of adaptivity (h-, p- and r-adaptivity) can be
considered and we do not need the interpolation onto a reference finite element space. The
complexity of this methods lies in the computation of the inner products.

1 1

Ω Ω

Figure 31: 1d finite element basis functions on their corresponding grids. Left: both piecewise
linear finite element ansatz functions. Right: piecewise linear and cubic finite element ansatz
functions

The calculation of the matrix K ∈ R(n+1)×(n+1) is favorable in such a manner, since we assume
the temporal dimension n to be far smaller than the spatial dimension(s) {Nj }nj=0 . Due to the
symmetry of K it suffices to compute the entries on and upon the diagonal of the matrix. The
computations of the matrix entries of K can be done fully in parallel. Hence, if appropriate
hardware is available, setting up the matrix K can be done fast.

Numerical realization in the case of snapshots which lead to intersecting elements

In order to explain the practical realization of an inner product of type (yhi , yhj )X of space-adapted
finite element data, let us start with considering an example.

Example. We choose V = H 1 (Ω), H = L2 (Ω) and set X = H = L2 (Ω). The triangulations

of Ω̄ for each time level j = 0, . . . , n are denoted by {Thj }nj=0 and the finite element spaces are
defined by

Vhj = V (Thj ) = {v ∈ C 0 (Ω̄) : v|T ∈ Pr (T ), ∀T ∈ Thj } ⊂ X, j = 0, . . . , n,

where Pr denotes the space of polynomials of degree r ∈ N. The computation of the ij-th entry
√ √
Kij = αi αj (yhi , yhj )L2 (Ω) of the matrix K is calculated by
Z
√ √ √ √
αi αj (yhi , yhj )L2 (Ω) = αi αj yhi yhj dx
Ω
Nj
Ni X Z
√ √
yki ylj vki vlj dx
X
= αi αj
k=1 l=1 Ω
4.2 POD with space-adapted snapshots 83

 
Nj
Ni X Z
√ √
yki ylj  vki vlj dx .
X X X
= αi αj

k=1 l=1 T ∈Thi T̄ ∈T j T ∩T̄
h

Computations become easier when using nested grids. In this case, the intersection of two ar-
bitrary d-dimensional simplices coincides either with the smaller simplex, or is a common edge
simplex, or has no overlap. ♦

The example reveals that the setting of dealing with different finite element meshes causes the
possible challenge to deal with integrals of type
Z
yhi yhj dx (4.16)
P

over cut finite elements (convex polytopes) P = T ∩ T̄ with T ∈ Thi and T̄ ∈ Thj , which involves
the integration of functions defined on different (non-matching) meshes, compare Figure 32.

T̄

Figure 32: Two finite element meshes (left, middle) and the overlapping mesh (right)

The numerical realization is strongly based on [134], where a similar challenge is investigated in
the context of multimesh methods. There, the major challenges are identified as
(i) collision detection (find intersecting simplices),

(ii) mesh intersection (detect intersection interface),

(iii) integration on complex polyhedra.

In the numerical test, we make use of built-in FEniCS [13, 132] tools to handle the issues (i) and
(ii). In particular, for (i) we utilize the FEniCS collection of routines CollisionDetection which
(amongst others) comprises an edge-edge, triangle-triangle, tetrahedron-tetrahedron collision de-
tection and delivers the intersection coordinates for task (ii). For issue (iii) which addresses the
integration over the cut elements, a subtriangulation can be an option. But as pointed out in
[134] in the context of a three-dimensional example, a subtetrahedralization of an arbitrary poly-
hedron is challenging and additional vertices might have to be added. Moreover, the resulting
subtriangulation might lead to very fine meshes such that it might slow down the performance
or storage problems might occur. Therefore, an alternative approach is outlined which is based
on a boundary representation of the integrals, compare [129, section 2.4] or [135] for example.
Exemplarily, in the case of a two-dimensional domain with piecewise linear finite element dis-
cretization, the integrals of type (4.16) can be computed as respectively weighted sum of the
integrals Z Z Z Z Z Z
x21 dx, x22 dx, x1 x2 dx, x1 dx, x2 dx, 1dx, (4.17)
P P P P P P
with x = (x1 , x2 ) using Stoke’s formula
Z m Z
1 X bi
f (x)dx = f (x)dσi , (4.18)
P 2+q kai k Ei
i=1
84 4 ADAPTIVITY WITH RESPECT TO SPACE

where {Ei }m 2 T
i=1 denote the edges of the polytope P ⊂ R , ai x = bi is the hyperplane in which
Ei lies, and f is a polynomial of degree q. Note that the line integrals on the edges Ei can be
computed using standard Gauss quadrature, for example.

It is an advantage of this approach that arbitrary snapshot data can be handled. As long as
the snapshots live in the same Hilbert space, no further assumptions need to be set for the
snapshots. This allows a great flexibility concerning the actual discretization scheme, such that
h-, p- and r-adapted snapshots or a blend of them can be considered. Consequently, the gen-
eration of snapshot data is not a-priorily restricted to a certain discretization scheme and mesh
handling. Moreover, there is no necessity of interpolating the snapshots into a reference finite
element space. This means that interpolation errors can be avoided and an exact integration of
the polynomials over the spatial domain is possible. The number of degrees of freedom for each
snapshot is not touched and no additional degrees of freedom are introduced.

Nevertheless, the numerical implementation of integrals over cut finite elements is involving both
from the computational expenses and from the implementation point of view and might need
further packages additionally to the usual finite element software tools.

4.3 POD reduced-order modeling using space-adapted snapshots

In this section, we stay in the infinite-dimensional setting of the POD method and set up the
POD reduced-order model utilizing snapshots which belong to different finite element spaces.
This perspective allows us to determine the mode coefficients in the POD Galerkin ansatz for
different underlying finite element discretizations. We start with the derivation of the POD
reduced-order model for (4.4) in Section 4.3.1. The treatment of the nonlinearity is discussed in
Section 4.3.2. Finally, the expression of the POD solution in full spatial dimension is outlined in
Section 4.3.3.

4.3.1 POD reduced-order modeling

Suppose for given snapshots yh0 ∈ Vh0 , . . . , yhn ∈ Vhn we have computed the matrix K defined in
√ √
(4.14), with Kij = αi αj (yhi , yhj )X , for i, j = 0, . . . , n as well as its eigenvalue decomposition
according to the approach of the previous Section 4.2.2. Thus, we are given the ` largest eigen-
values {λi }`i=1 and corresponding eigenvectors {φi }`i=1 ⊂ Rn+1 of low rank `. The POD basis
{ψi }`i=1 can then be written as
n
1 1 X√
ψi = √ Yφi = √ αj (φi )j yhj ∈ V for i = 1, . . . , `,
λi λi j=0

according to (4.15) and thus it holds X` = span{ψ1 , . . . , ψ` } ⊂ V ⊆ X. This POD basis is

utilized in order to compute a reduced-order model for (4.4). For this reason, we make the POD
Galerkin ansatz
` `
X X 1
y` (t) = ηi (t)ψi = ηi (t) √ Yφi for all t ∈ (0, T ], (4.19)
i=1 i=1
λi

as an approximation for y(t), with the time-dependent coefficients

1
ηi (t) = (y` (t), ψi )X = (y` (t), √ Yφi )X
λi
for i = 1, . . . , `. Inserting y` into (4.4) and choosing X` ⊂ V as the test space leads to the
following problem: find y` ∈ X` such that
d
(y` (t), ψ)H + a(y` (t), ψ) + hN (y` (t)), ψiV ∗ ,V = hf (t), ψiV ∗ ,V (4.20)
dt
4.3 POD reduced-order modeling using space-adapted snapshots 85

and
(y` (0), ψ)H = (y0 , ψ)H
holds true for all ψ ∈ X` and for almost all t ∈ (0, T ]. Utilizing the ansatz (4.19), we can write
(4.20) as an `-dimensional system of ordinary differential equations for the POD mode coefficients
{ηi (t)}`i=1 ⊂ R as follows
`
X `
X
η̇j (t)(ψj , ψi )H + ηj (t)hAψj , ψi iV ∗ ,V + hN (y` (t)), ψi iV ∗ ,V = hf (t), ψi iV ∗ ,V (4.21)
j=1 j=1

for t ∈ (0, T ], i = 1, . . . , ` and

`
X
ηj (0)(ψj , ψi )H = (y0 , ψi )H
j=1

and i = 1, . . . , `. Note that (ψi , ψj )H = δij if we choose X = H.

Since the aim is to construct a POD reduced-order model which can be built and solved for
snapshots which live in different finite element spaces, we employ the results of Section 4.2.2 and
rewrite system (4.21) utilizing the identity (4.15). Then, the system (4.21) can be written as
` `
X 1 1 X 1 1
η̇j (t) √ p (Yφj , Yφi )H + ηj (t) √ p hAYφj , Yφi iV ∗ ,V
j=1
λi λj j=1
λi λj
1 1
+ √ hN (y ` (t)), Yφi iV ∗ ,V = √ hf (t), Yφi iV ∗ ,V
λi λi
(4.22)
for t ∈ (0, T ] and with initial condition
`
X 1 1 1
ηj (0) √ p (Yφj , Yφi )H = √ (y0 , Yφi )H .
j=1
λi λj λi

In order to write (4.22) in a compact matrix-vector form, let us introduce the diagonal matrix
D ∈ R`×` by  
1 1
D = diag √ , . . . , √ .
λ1 λ`
From the first ` eigenvectors {φi }`i=1 of K we build the matrix

Φ ∈ R(n+1)×` by Φ = [φ1 | . . . | φ` ].

Then, the system (4.22) can be written as the system

DΦT KΦD η̇(t) + DΦT Y ∗ AYΦD η(t) + DN (η(t)) = DF (t) (4.23)

with
DΦT KΦD η(0) = Dη̄0
for the vector-valued mapping η(t) = (η1 (t), . . . , η` (t))T : [0, T ] → R` . Note that the right-hand
side F (t) and the initial condition η̄0 can be computed as

(F (t))i = hf (t), Yφi iV ∗ ,V = (Y ∗ f (t), φi )Rn+1

and
(η̄0 )i = (y0 , Yφi )H = (Y ∗ y0 , φi )Rn+1 ,
86 4 ADAPTIVITY WITH RESPECT TO SPACE

for i = 1, . . . , `, respectively. The calculation can be done explicitly for any finite element
discretization. For P a given function w ∈ V (for example w = f (t) or w = y0 ) with finite element
discretization w = N Nw Nw
i=1 wi vi , nodal basis {vi }i=1 ⊂ V and appropriate mode coefficients {wi }i=1
w

we can compute
Nw Nj Nj
Nw X
(w, yhj )X j j
wi ykj (vi , vkj )X ,
X X X
∗
(Y w)j = =( wi vi , yk vk )X = for j = 0, . . . , n
i=1 k=1 i=1 k=1

where yhj ∈ Vhj denotes the j-th snapshot. Again, for any i = 1, . . . , Nw and k = 1, . . . , Nj , the
computation of the inner product (ϕi , vkj )X can be done as outlined in Section 4.2.2.

For linear evolution equations the POD reduced-order model (4.23) can be set up and solved
utilizing snapshots which belong to different finite element spaces, since the matrices DΦT KΦD
and DΦT Y ∗ AYΦD can be constructed explicitly. We discuss options in order to treat the non-
linear component N (η(t)) in the following Section 4.3.2.

Time-discrete reduced-order model

For simplicity, we choose for the reduced-order model the same temporal grid {tj }nj=0 introduced
in (4.6) which is utilized for the full-order solution. It is also possible to use a different time grid
for which we refer to [121]. The time-discrete POD reduced-order model reads as follows: find a
sequence y`0 , . . . , y`n ∈ X` satisfying
!
y`j − y`j−1
Z tj
j j 1
,ψ + a(y` , ψ) + hN (y` ), ψiV ,V =
∗ hf (t), ψiV ∗ ,V dt for all ψ ∈ X`
∆tj ∆tj tj−1
H
(4.24)
and
(y`0 , ψ)H = (y0 , ψ)H
In matrix-vector formulation this can also be written as
 j
η − η j−1

T
DΦ KΦD + DΦT Y ∗ AYΦD η j + DN (η j ) = DFj for j = 1, . . . , n (4.25)
∆tj
with
DΦT KΦD η 0 = Dη̄0 .

4.3.2 Treatment of the nonlinearity

Let us now consider the computation of the nonlinear term DN (η(t)) ∈ R` of the POD reduced-
order model (4.23). It holds true

(DN (η(t)))k = hN (y` (t)), ψk iV ∗ ,V

= hN ( `i=1 ηi (t)ψi ), ψk iV ∗ ,V
P

for k = 1, . . . , `. It is well-known that the evaluation of nonlinearities in the context of reduced-

order modeling is computationally expensive. To make this clear, let us assume we are given
a uniform finite element discretization with N degrees of freedom. Then, in the fully discrete
setting, the nonlinear term has the form

ΨT W N (Ψη(t)),

where Ψ = [ψ1 | · · · | ψ` ] ∈ RN ×` denotes the matrix in which contains the POD modes colum-
nwise and W ∈ RN ×N is a weighting matrix related to the utilized inner product. Hence, the
4.3 POD reduced-order modeling using space-adapted snapshots 87

treatment of the nonlinearity requires the expansion of y` (t) = Ψη(t) ∈ RN in the full space,
then the nonlinearity can be evaluated and finally the result is projected back to the POD space.
As a consequence, the reduced-order model is not fully independent of the high-order dimension,
which slows down the simulation time - often quite drastically. Therefore, it is convenient to seek
for so-called hyper-reduction methods, i.e. for a treatment of the nonlinearity where the model
evaluation cost is related to a low dimension. Possible hyper-reduction techniques are enumer-
ated in Section 2.4.2 in the paragraph treatment of nonlinearities. All the methods (except for
EIM) have in common that the snapshot data needs to be available in a common reference finite
element space. Since we want to circumvent this condition and set up the reduced-order model
for space-adapted snapshots without the necessity of interpolation into a reference space, we have
to go a different way.

We linearize and project the nonlinearity onto the POD space. For this approach, let us consider
for a given initial condition (y` (0), ψ)H = (y0 , ψ)H for all ψ ∈ X` the following linear reduced-
order system
d
(y` (t), ψ)H + a(y` (t), ψ) + hN (ȳ(t)), ψiV ∗ ,V = hf (t), ψiV ∗ ,V ∀ψ ∈ X` (4.26)
dt
where ȳ(t) ∈ V denotes a given state. This linear evolution problem (4.26) can be set up
and solved explicitly according to Section 4.3.1 without the necessity of interpolation of the
snapshots into a reference finite element space. In the numerical examples, we take the finite
element solution as given state in each time step, i.e. ȳ(tj ) = yhj .
Furthermore, the linearization of the reduced-order model (4.20) can be considered. For a given
initial condition (y` (0), ψ)H = (y0 , ψ)H for all ψ ∈ X` and a given state ȳ(t) ∈ V it reads as
follows
d
(y` (t), ψ)H + a(y` (t), ψ) + hN (ȳ(t)) + Ny (ȳ(t))(y` − ȳ)(t), ψiV ∗ ,V = hf (t), ψiV ∗ ,V (4.27)
dt
for all ψ ∈ X` where Ny denotes the Fréchet derivative. This linearized problem is of interest
e.g. in the context of optimal control, where it occurs in each iteration level within the sequential
quadratic programming (SQP) method, see [101, Section [Link]], for example. Choosing the
finite element solution as given state in each time instance leads to:
√
hN (yhj ), ψi iV ∗ ,V j
Pn
√1 k
= λi k=0 αk (φi )k hN (yh ), yh iV ∗ ,V ,

hNy (yhj )y` (tj ), ψi iV ∗ ,V hNy (yhj )( `k=1 ηk (tj )ψk ), ψi iV ∗ ,V ,

P
=
P`
√1 √1
Pn √ √
Pn j p r
= k=1 ηk (tj ) λk λi r=0 αp αr (φk )p (φi )r hNy (yh )yh , yh iV ∗ ,V ,
p=0
Pn √
hNy (yhj )yhj , ψi iV ∗ ,V = √1
λi
j j k
k=0 αk (φi )k hNy (yh )yh , yh iV ∗ ,V .

Finally, we approximate the nonlinearity DN (η j ) in (4.25) by

(DN (η j ))k ≈ hN (yhj ) + Ny (yhj )(y` (tj ) − yhj ), ψk iV ∗ ,V

which can be written as

DN (η j ) ≈ DΦT Nj + DΦT Njy ΦDη j − DΦT Njy

where
√ √
hN (yhj ), α0 yh0 iV ∗ ,V hNy (yhj )yhj , α0 yh0 iV ∗ ,V
   

Nj =  .. n+1
Njy =  .. n+1
∈R , ∈R
   
. .
j √ √
hN (yh ), αn yhn iV ∗ ,V hNy (yhj )yhj , αn yhn iV ∗ ,V
88 4 ADAPTIVITY WITH RESPECT TO SPACE

and
√ √ √ √
hNy (yhj ) α0 yh0 , α0 yh0 iV ∗ ,V . . . hNy (yhj ) αn yhn , α0 yh0 iV ∗ ,V
 

Njy =  .. .. (n+1)×(n+1)
∈R .
 
. .
j √ √ j √ √
hNy (yh ) α0 yh , αn yhn iV ∗ ,V
0 . . . hNy (yh ) αn yhn , αn yhn iV ∗ ,V
For weakly nonlinear systems this approximation may be sufficient, depending on the problem
setting and its goal. A great advantage of linearizing the semilinear partial differential equation
is that only linear equations need to be solved which leads to a further speed up. However, if
a more precise approximation is desired or necessary, we can think of approximations including
higher-order terms, like quadratic approximation, see e.g. [51, 157], or Taylor expansions, see
e.g. [74, 148, 149]. Nevertheless, the efficiency of higher-order approximations is limited due to
growing memory and computational costs.

4.3.3 Expressing the POD solution in the full spatial domain

After the computation of the solution sequence {η j }nj=0 to (4.25) with η j = (η1j , . . . , η`j ) ∈ R` , we
can write the reduced solution y`j as
` n
!
j
X j 1 X√
y` = ηi √ αk (φi )k yhk . (4.28)
i=1
λ i k=0

Let us recall the spatial discretization of the snapshots (4.10) given by

Nj
yhj ylj vlj
X
= for j = 0, . . . , n.
l=1

Now, we want to express the reduced solution y`j in the full spatial domain. For this, we assume
Mj
to be given a set of grid points {Qjr }r=1 for the reduced system at time level tj . The fully discrete
POD solution can be computed by evaluation
` n Nk
!
j j
X j 1 X√ X
k k j
y` (Qr ) = ηi √ αk (φi )k ( yl vl (Qr )) (4.29)
i=1
λi k=0 l=1

for j = 0, . . . , n and r = 1, . . . , Mj . This allows us to use any grid for expressing the POD
solution in the full spatial domain. For example, we can use the same node points at time level j
for the POD simulation as we have used for the snapshots, i.e. for j = 0, . . . , n it holds Mj = Nj
and Qjr = Pkj for all r, k = 1, . . . , Nj . Another option can be to choose
Nj
n [
M
{Pkj },
[
{Qjr }r=1
j
=
j=0 k=1

i.e. the common grid of all snapshots. Obviously, a special and probably the easiest case concern-
ing the implementation is to choose snapshots which are expressed with respect to a reference
finite element space. Then, expression (4.29) simplifies to
` n
!
j j
X j 1 X√ k
y` (Pr ) = ηi √ αk (φi )k y . (4.30)
i=1
λ i k=0

Note that the expression of the POD solution in the full spatial domain does not belong to the
POD online time (i.e. the solution of the reduced-order model). The evaluation of the reduced
solution (4.29) delivers the approximated solution data needed for the application at the chosen
time and space points.
4.4 Error analysis for the POD reduced-order model with space-adapted snapshots 89

4.4 Error analysis for the POD reduced-order model with space-adapted snapshots
In order to validate the accuracy of the reduced-order solution with respect to the true solution,
we are interested in analyzing the quantity
n
αj ky(tj ) − y`j k2
X
(4.31)
j=0

in a suitable norm, where y(tj ) denotes the true (unknown) solution to (4.4) at time instances
t0 , . . . , tn and y`j ∈ X` for j = 0, . . . , n is the solution to the reduced-order model (4.24). In par-
ticular, the reduced space X` = span{ψ1 , . . . , ψ` } is constructed from snapshots yh0 , . . . , yhn from
different finite element spaces, i.e. yhj ∈ Vhj for j = 0, . . . , n, according to Sections 4.2 and 4.3.
The weights α0 , . . . , αn are non-negative time weights.

Let us give references to similar error analysis studies. In [120] an a-priori estimate is proved for
the error between the true solution to a nonlinear parabolic equation at certain time instances
and the time-discrete reduced-order solution. In particular, [120, Theorems 7,8,9] consider an
estimation for a backward Euler–Galerkin scheme, a Crank–Nicolson scheme and a Forward
Euler–Galerkin scheme, respectively.
In [121, Theorem 4.7] an error estimation is proved which allows to use different time discretiza-
tions for the full-order and the reduced-order model.
Moreover, [186, Theorem 3.2.5] considers an error estimation in a finite-dimensional setting.
In order to derive an estimation for the quantity (4.31) we follow very closely the proofs in
[120, 121, 186]. Here, we compute the POD basis corresponding to the fully discrete snapshots
yh0 , . . . , yhn using adaptive finite element spaces, which is the main differentiator to the above
mentioned literature where the POD basis is computed from snapshots corresponding to the
solution trajectory at given time points. For this reason, an additional term associated with the
error for the spatial discretization will appear in the estimate.

We make use of the decomposition

y(tj ) − y`j = y(tj ) − yhj + yhj − P ` yhj + P ` yhj − P ` y(tj ) + P ` y(tj ) − y`j = ηj + %j + ζj + ϑj (4.32)

for j = 0, . . . , n, where ηj := y(tj )−yhj , %j := yhj −P ` yhj , ζj = P ` yhj −P ` y(tj ) and ϑj := P ` y(tj )−y`j .
The term ηj is the discretization error. We utilize the decomposition

ηj = y(tj ) − yhj = y(tj ) − y j + y j − yhj = Etj + Ehj

where y j denotes the solution to the time-discrete problem (4.7). By Etj := y(tj ) − y j we denote
the global time discretization error and Ehj := y j − yhj is the global spatial discretization error.
It is
kEhj kH ≤ max kEhj kH =: εh
j=0,...,n

and
kEtj kH ≤ max kEtj kH =: εt .
j=0,...,n

Since we use the implicit Euler method for time integration, it is εt = O(∆t) with ∆t :=
maxj=0,...,n ∆tj . Therefore, we can estimate
n n n
yhj k2H αj kEtj Ehj k2H
X X X
αj ky(tj ) − ≤ + ≤2 αj ((∆t)2 + ε2h ) ≤ 2T ((∆t)2 + ε2h ). (4.33)
j=0 j=0 j=0
90 4 ADAPTIVITY WITH RESPECT TO SPACE

Moreover, we have
n n n
αj kP ` yhj
X X X
αj kζj k2H = −P `
y(tj )k2H ≤ kP ` k2L(H) αj kηj k2H . (4.34)
j=0 j=0 j=0

The term %j is the projection error of the snapshot yhj projected onto the POD space V` . Using
(4.2), the weighted sum of all projection errors is given by the sum of the neglected eigenvalues
(2.12), i.e.

n n ` n ` d
αj kyhj (yhj , ψi )V ψi k2H αj kyhj j
X X X X X X
αj k%j k2H = − ≤ cv − 2
(yh , ψi )V ψi kV ≤ cv λi .
j=0 j=0 i=1 j=0 i=1 i=`+1
(4.35)
It remains to estimate the term ϑj which is the error between the projection of the true solution
y(tj ) at time instance tj onto the POD space V` and the time-discrete ROM solution y`j to (4.24).
With the use of the notation ∂ϑ ¯ j = (ϑj − ϑj−1 )/∆tj for j = 1, . . . , n, we get

y j − y`j−1
 
¯ j , ψ)H y(tj ) − y(tj−1 )
(∂ϑ = (P ` − ` , ψ)H
 ∆tj  ∆tj
y(tj ) − y(tj−1 )
= (P ` + N (y`j ) − fj , ψ)H + a(y`j , ψ)
 ∆t j 
y(tj ) − y(tj−1 ) y(tj ) − y(tj−1 )
= (P ` − + N (y`j ) − N (y(tj )), ψ)H + a(y`j − y(tj ), ψ)
∆tj ∆tj
= (zj + N (yj` ) − N (y(tj )), ψ)H + a(y`j − y(tj ), ψ)
 
y(tj ) − y(tj−1 ) y(tj ) − y(tj−1 )
for ψ ∈ V` with zj := P − . With the choice ψ = ϑj and the
∆tj ∆tj
use of the identity
2hu − v, ui = kuk2 − kvk2 + ku − vk2

we obtain
 
kϑj k2H ≤ kϑj−1 k2H + 2∆tj βky`j − y(tj )kH + kzj kH + kN (y`j ) − N (y(tj ))kH kϑj kH ,

where we have utilized (4.3). We assume that N is Lipschitz continuous, i.e. there exists L > 0
such that
kN (yj` ) − N (y(tj ))kH ≤ Lky`j − y(tj )kH for j = 0, . . . , n.

Applying Young’s inequality we find

kϑj k2H ≤ kϑj−1 k2H + ∆tj (c1 k%j k2H + c2 kϑj k2H + kzj k2H + c1 kηj k2H + c1 kζj k2H )

with the constants c1 := β + L and c2 := 5(β + L) + 1. Under the assumption that ∆t is

sufficiently small, we conclude

j
!
X
kϑj k2H ≤e 2c2 j∆t
kϑ1 k2H + ∆tk (kzk k2H + c1 k%k k2H + c1 kηk k2H + c1 kζk k2H ) . (4.36)
k=1

For more details on this, we refer to [121, 186]. We choose the initial condition for (4.24) such
that ϑ0 = P ` y(t0 ) − y0` = P ` g − y0` = 0.
4.4 Error analysis for the POD reduced-order model with space-adapted snapshots 91

Next, we estimate the term involving zk . It holds true

 
2 ` y(tk ) − y(tk−1 ) y(tk ) − y(tk−1 ) 2
kzk kH = kP − kH
 ∆tk  ∆tk
y(tk ) − y(tk−1 ) y(tk ) − y(tk−1 ) 2
= kP ` − P ` ẏ(tk ) + P ` ẏ(tk ) − ẏ(tk ) + ẏ(tk ) − kH
∆tk ∆tk
y(tk ) − y(tk−1 ) y(tk ) − y(tk−1 ) 2
≤ 2kP ` k2L(H) k − ẏ(tk )k2H + 2kP ` ẏ(tk ) − ẏ(tk )k2H + 2kẏ(tk ) − kH
∆tk ∆tk
2 `
≤ c3 kwk kH + 2kP ẏ(tk ) − ẏ(tk )kH 2

y(tk ) − y(tk−1 )
with c3 = 2 + 2kP ` k2L(H) and wk := ẏ(tk ) − , which can be estimated as
∆tk
j
X (∆t)2
∆tk kwk k2H ≤ kÿk2L2 (0,tj ;H) .
3
k=1

For more details on this, we refer to [121, 186].

Finally, we can summarize the estimation for the term involving ϑj by
n
!
X
kϑj k2H ≤ c4 αk (kP ` ẏ(tk ) − ẏ(tk )k2H + k%k k2H + kηk k2H + kζk k2H ) + (∆t)2 kÿk2L2 (0,T ;H)
k=0

with c4 := e2c2 T max{ c33 , 4, 2c1 } and thus it is

n
X n
X d
X
αj kϑj k2H ≤ c4 T ( αj kP ` ẏ(tj ) − ẏ(tj )k2H + (∆t)2 kÿk2L2 (0,T ;H) + λi
j=0 j=0 i=`+1
(4.37)
+2T (1 + kP ` k2L(H) )((∆t)2 + ε2h )).

Theorem 4.3. Let y(t0 ), . . . , y(tn ) denote the solution to problem (4.4) at the time instances
t0 , . . . , tn and let y`0 , . . . , y`n be the solution to (4.24). Let the nonlinear operator N be Lipschitz
continuous with Lipschitz constant L and let the maximal time step ∆t := maxj=0,...,n ∆tj be
sufficiently small. Furthermore, we assume ÿ(t) to be bounded on [0, T ]. We choose the ini-
tial condition for (4.24) such that P ` g = y`0 is fulfilled. Then, there exists a constant C =
C(T, cv , kP ` k2L(H) , β, L, kÿk2L2 (0,T,H) ) > 0 such that
 
n d n
αj ky(tj ) − y`j k2H ≤ C (∆t)2 + ε2h +
X X X
λi + αj kP ` ẏ(tj ) − ẏ(tj )k2H  , (4.38)
j=0 i=`+1 j=0

where cv , β are from (4.2),(4.3), the quantity εh := maxj=0,...,n ky j − yhj kH refers to the global
spatial discretization error and y j is the solution to (4.7) at time instance tj for j = 0, . . . , n.
Proof. Utilizing the decomposition (4.32) we infer
n n
αj ky(tj ) − y`j k2H ≤ 2
X X
αj (kηj k2H + k%j k2H + kζj k2H + kϑj k2H ).
j=0 j=0

Together with (4.33), (4.34), (4.35) and (4.37) this leads to the claim.

Remark 4.4. i) Note that (4.38) in Theorem 4.3 depends on an error of time derivatives. As
discussed in [103, Remark 4.8(b)] this dependency can be avoided by enriching the snapshot
set with time derivatives (more specifically, finite difference approximations of time deriva-
tives), see also [102, 121].
Note that for a specific setting it is possible to avoid to include time derivatives into the
snapshot set. This is based on [165]. A comprehensive analysis on whether snapshot dif-
ference quotients are needed in POD is given in [109].
92 4 ADAPTIVITY WITH RESPECT TO SPACE

ii) If we choose V = H 1 (Ω) and H = L2 (Ω) and utilize a static piecewise linear finite element
discretization with h being the diameter of the triangles, then εh = O(h2 ). It is possible to
control the spatial discretization error a-posteriori using residual based error estimations.

4.5 Numerical examples

We consider three numerical examples in order to quantify numerically how the inclusion of spa-
tial adaptivity in the offline phase for snapshot generation in POD model order reduction effects
the efficiency and accuracy of the reduced-order model in comparison to the use of a static and
uniform spatial discretization.

In the first two examples we focus on structured, hierarchical and nested meshes and utilize an h-
adaptive finite element concept which controls the spatial discretization error in the energy norm.
The first example (Section 4.5.1) is a linear heat equation and the second example (Section 4.5.2)
is concerned with a phase field system governed by a Cahn–Hilliard system with transport. In
both cases, we use conformal, piecewise linear and continuous finite elements and follow the usual
adaptive cycle given by
solve - estimate - mark - refine/coarsen,
recall Algorithm 1. As an error indicator we utilize at each time instance the jump of the normal
derivative across the edges given by

hE kJE (~nE · ∇yhj )k2L2 (E) ,

X
(4.39)
E∈E

where hE is the length of the edge, JE denotes the jump of the function across the edge E (see
Definition 2.35), ~nE is the outward normal derivative on the edge E, yhj is the finite element
solution at the current time and E denotes the collection of all (interior) edges in the current
triangulation. It is shown in [47] that the edge residuals dominate a-posteriori error estimates for
low-order finite element methods. For this reason, we neglect the further terms within residual
based error indicators. For reasons of efficiency, we note that at each time instance (except for
the first one) we perform only one refinement and coarsening step, since we expect only minor
dynamical changes from one time instance to another assuming that the time step size is small
enough.
These numerical test cases illustrate our approach explained in Sections 4.2 and 4.3 to set up
and solve a POD reduced-order model utilizing space-adapted snapshots in the specific case in
which a nested mesh structure is at hand. Thus, the computations benefit from the fact that
the intersection of two triangles coincides either with the smaller triangle or is a common edge
or has no overlap. We exploit this property in the numerical implementation. The results are
compared to the use of a uniform mesh, where the mesh size coincides with the fineness of the
smallest triangle in the adaptive meshes. The aim of this comparison is to investigate numerically
how the inclusion of spatial adaptivity for the snapshot discretization affects the accuracy and
efficiency of the POD reduced-order solution compared to using a static mesh without spatial
adaptivity. Furthermore, we compare the approach to the use of a common reference mesh fol-
lowing [181]. For this, the practical numerical concept works as follows. In a full-dimensional
simulation, h-adaptive snapshots are generated. At the same time we carry along a reference
grid with the simulation, which coincides with the computational grid at initial time and is only
refined (not coarsened) in the same manner as the computational grid. In this way, the reference
grid becomes the overlay of all computational grids at the end of the snapshot generation. Then,
the snapshots are expressed with respect to the finite element basis corresponding to the overlay
mesh and the usual POD procedure is carried out.
In the POD framework, we choose X = L2 (Ω) as Hilbert space in the setting of (2.7).
4.5 Numerical examples 93

The third example (Section 4.5.3) focuses on the realization of the numerical computation of the
correlation matrix K defined in (4.14), as described in Section 4.2.2 for non-nested meshes. In
this case, the overlap of two meshes leads to cut finite elements, which are convex polytopes of
more than three node points such that a boundary representation of the integrals is utilized.

All coding is done in C++, we utilize FEniCS [13, 132] for the solution of the differential equations
utilizing finite elements and ALBERTA [161] for dealing with hierarchical meshes. We run the
numerical tests on a compute server with 1536 GB RAM. We note that we use libraries from
[114] for the coupling of FEniCS with ALBERTA for mesh refinement and coarsening.

4.5.1 Linear heat equation

As an example for the evolution problem (4.4) we consider a linear heat equation. In the setting
of the Gelfand triple (4.1) we choose V = H01 (Ω) and H = L2 (Ω) as Hilbert spaces. For a given
source function f ∈ L2 (0, T ; V ∗ ) and initial condition g ∈ L2 (Ω), we consider the following linear
heat equation with homogeneous Dirichlet boundary conditions

yt − ∆y = f in (0, T ] × Ω, (4.40a)
y=0 on [0, T ] × ∂Ω, (4.40b)
y(0, ·) = g in Ω. (4.40c)

We can write (4.40) as an abstract evolution problem of type (4.4) by deriving a variational
formulation for (4.40) with the bilinear form a : V × V → R introduced by

a(u, v) = (∇u, ∇v)L2 (Ω) .

Problem setting

The spatial domain is chosen as the open unit square Ω = (0, 1) × (0, 1) ⊂ R2 and the end time
is set to T = 2. We construct an example in such a way that we know the analytical solution. It
is given by
ȳ(t, x) = r(t, x) · (s1 (t, x) − s2 (t, x))
with
1
50000·x1 ·(1−x1 )·(0.5+cos(t)·(x1 −0.5)−sin(t)·(x2 −0.5))4 · t+1 ·(1−(0.5+cos(t)·(x1 −0.5)−sin(t)·(x2 −0.5)))4 )
r(t, x) = 1+1000·(cos(t)·(x1 −0.5)−sin(t)·(x2 −0.5))2
,

10000·x2 ·(1−x2 )·(0.5+sin(t)·(x1 −0.5)+cos(t)·(x2 −0.5))2 ·(0.5−sin(t)·(x1 −0.5)−cos(t)·(x2 −0.5))2

s1 (t, x) = 1+100·((0.5+sin(t)·(x1 −0.5)+cos(t)·(x2 −0.5))−0.25)2
,

10000·x2 ·(1−x2 )·(0.5+sin(t)·(x1 −0.5)+cos(t)·(x2 −0.5))2 ·(0.5−sin(t)·(x1 −0.5)−cos(t)·(x2 −0.5))2

s2 (t, x) = (1+100·((0.5+sin(t)·(x1 −0.5)+cos(t)·(x2 −0.5))−0.75)2
,

where we use the notation x = (x1 , x2 ). The forcing term f and the initial condition g are chosen
accordingly. For the temporal discretization we introduce the uniform time grid by

tj = j∆t

for j = 0, . . . , 2000 with time step size ∆t = 0.001.

Finite element results

We start the adaptive finite element Algorithm 1 with an initial triangulation Thinit of the spatial
domain Ω which we choose as a structured criss-cross mesh with 20 node points in both x1 - and
x2 -direction. For the Dörfler marking in Definition 2.34, we set θ = 0.9 and utilize a coarsening
94 4 ADAPTIVITY WITH RESPECT TO SPACE

Figure 33: Surface plot (top) and top view (bottom) of the adaptive finite element solution of
(4.40) at t = t0 (left), t = T /2 (middle) and t = T (right)

strategy according to [114, Section 8.3] with tolerance θc = 0.06. A surface plot of the finite
element solution at three different time points is shown in Figure 33.
Due to the steep gradients in the neighborhood of the minimum and maximum, respectively,
the use of an adaptive finite element discretization is justified. A top view of the adaptive
finite element solutions as well as the adapted finite element meshes and the overlay mesh are
shown in Figure 34. The number of node points of the adaptive meshes varies between 4284
and 13076 points. The finest mesh has 42997 node points. In contrary, a uniform mesh with
the same discretization fineness as the smallest triangle in the adaptive grids with hmin = 0.0032
has 185437 node points. This clearly reveals the benefit of using adaptive meshes for snapshot
generation. In particular, the comparison of the computational times emphasizes the benefit of
adaptive snapshot sampling: the snapshot generation utilizing adaptive meshes takes on average
1053 seconds (17.5 minutes), whereas utilizing the uniform mesh takes around 14143 seconds
(3.9 hours). Therefore, the inclusion of spatial adaptivity leads to a speed up factor of 13.4 in
the offline phase (see Table 4) for snapshot generation.

Figure 34: From left to right: Adapted finite element meshes at t = t0 , t = T /2, t = T and the
overlay mesh

POD reduced-order results

We investigate the inclusion of spatial adaptivity for snapshot generation in comparison to the use
of a static uniform mesh. In Figure 35, the decay of the normalized eigenspectrum of the snapshot
gramian for uniform spatial discretization is shown in contrast to the normalized eigenvalues
using h-adaptive finite elements with nested hierarchical meshes. We note that about the first
40 eigenvalues computed corresponding to space-adapted snapshots coincide with eigenvalues
associated with snapshots from a simulation on a uniform mesh. Then, the methods deliver
4.5 Numerical examples 95

different results: for the uniform discretization the normalized eigenvalues fall below machine
precision and stagnate. In contrary, the normalized eigenvalues for the adaptive approach flatten
in the order around 10−10 . If the marking tolerance θ for the Dörfler marking is chosen larger,
i.e. θ = 0.98 (or smaller, i.e. θ = 0.5), the stagnation of the eigenvalues in the adaptive method
takes place at a higher (or lower) order (Figure 35, right).

10 0 10 0 10 0
adaptive adaptive adaptive
uniform uniform uniform
-5 -5 -5 =0.5
10 10 10 =0.98

10 -10 10 -10 10 -10

10 -15 10 -15 10 -15

10 -20 10 -20 10 -20

0 500 1000 1500 2000 0 50 100 150 200 0 50 100 150 200

Figure 35: Decay of the normalized eigenvalues utilizing an h-adaptive and a static finite element
discretization, respectively. Left: all eigenvalues, middle: first 200 largest eigenvalues, right: first
200 largest eigenvalues with different Dörfler marking tolerances θ for the adaptivity

Concerning dynamical systems, the magnitude of the eigenvalue corresponds to the dominant
characteristic properties of the underlying dynamical system: the larger the eigenvalue, the
more information is contained in the corresponding eigenfunction. Since all adaptive meshes are
contained in the uniform mesh, the difference in the amplitude of the eigenvalues is due to the
interpolation errors during coarsening. This is the price we have to pay in order to get a fast
snapshot generation utilizing adaptive finite elements. Moreover, the investigation of the decay
of the eigenvalues can be interpreted as an analyzing tool for adaptivity in the following sense:
using an adaptive mesh technique means that some parts of the domain are resolved coarsely
according to the utilized error estimation, i.e. information gets lost compared to a uniform fine
resolution. In the sense of a singular value analysis, this can be explained that adaptivity neglects
the noise which is indicated by the singular values on the uniform spatial mesh at those places
which are not resolved with the adaptive grid. We conclude that the overtones which get lost in
the adaptive computations lie in the same space which is not considered by POD when using the
adaptive finite element snapshots. This allows us to characterize the space which is not resolved
by adaptivity. From this point of view, adaptivity can be interpreted as a smoother.
Since the first few POD modes are the most important ones regarding the captured information
(coherent structures), we visualize ψ1 , ψ2 and ψ5 in Figure 36, which are computed from space-
adapted snapshots. Note that the POD bases are unique up to the sign. We can recognize the
initial condition in the first POD basis function. Then, the index of the POD basis corresponds
to the number of maxima and minima of the POD basis: ψ2 has two minima and two maxima etc.
This behavior is similar to the increasing oscillations in higher frequencies within trigonometric
approximations. The increasing number of oscillations is necessary in order to approximate the
transport of the steep gradients of the solution with increasing accuracy.
The solutions to the POD reduced-order model with dimensions ` = 10 and ` = 40, respectively,
are shown in Figure 37, where space-adapted snapshots are utilized. The visual comparison
reveals that less oscillations appear in the POD solution, if the number of utilized POD basis
functions is increased (until stagnation of the corresponding eigenvalues).
96 4 ADAPTIVITY WITH RESPECT TO SPACE

Figure 36: Surface plot (top) and top view (bottom) of the POD basis functions ψ1 (left), ψ2
(middle) and ψ5 (right)

Figure 37: Surface plot of the POD reduced-order solution utilizing ` = 10 (top) and ` = 40
(bottom) POD modes at t = t0 (left), t = T /2 (middle) and t = T (right)

In order to validate the approximation quality of the POD reduced-order solution we consider
the following relative errors

i) relative L2 (0, T ; Ω)-error between the finite element solution and the POD reduced-order
solution given by
kyh − y` kL2 (0,T ;Ω)
errF E−P OD =
kyh kL2 (0,T ;Ω)

ii) relative L2 (0, T ; Ω)-error between the true solution and the POD reduced-order solution
given by
kȳ − y` kL2 (0,T ;Ω)
errtrue−P OD =
kȳkL2 (0,T ;Ω)
4.5 Numerical examples 97

iii) POD projection error given by

` y k
pP
kyh − PH h L2 (0,T ;Ω) λi
errP ODproj = = pPi>`
kyh kL2 (0,T ;Ω) i λi

Figure 38 illustrates these errors (i)-(iii) utilizing static versus space-adapted snapshots depend-
ing on different POD basis lengths. Let us start with analyzing the decay of the relative error
in the case of uniform and static meshes (Figure 38, right). In this case, the POD reduced-
order solution converges to the finite element solution for an increasing number of utilized POD
modes. In fact, if we utilize ` = dim(span{yh1 , . . . , yhn }) POD modes, then the reduced space
coincides with the finite element space. The error between the finite element solution and the
POD reduced-order solution is very close to the POD projection error which is the lower bound
for the error errF E−P OD . However, the error between the POD reduced-order solution and the
true solution ȳ stagnates in the order of 10−03 for ` ≥ 40, since the finite element discretization
error then dominates the overall error according to Theorem 4.3. This is in accordance to the
decay of the eigenvalues shown in Figure 35.
Let us now look at the relative errors with regard to the space-adapted snapshots (Figure 38,
middle). We also observe a stagnation of the error between the true solution and the POD
reduced-order solution. In particular, the stagnation starts earlier and is at an order of 10−02 ,
i.e. larger than in the static case. This is due to the fact that the finite element discretization er-
ror is larger in the case of space-adapted solutions, since we are concerned with adapted meshes.
Moreover, the POD projection error stagnates in the case of space-adapted snapshots. This is in
accordance to the decay of the eigenvalues (compare Figure 35). It is related to the finite element
error in the snapshots. For an increasing number of utilized POD modes the POD reduced-order
solution might approximate spatial noise and artifacts which come from the adaptive finite ele-
ment meshes.

10 0 10 0 10 0

10 -2 10 -2 10 -2

10 -4 10 -4 10 -4
FE-POD adaptive
true-POD adaptive FE-POD adaptive FE-POD uniform
FE-POD uniform true-POD adaptive true-POD uniform
true-POD uniform POD proj POD proj
10 -6 10 -6 10 -6
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100

Figure 38: Relative L2 (0, T ; Ω)-error between the POD reduced-order solution and the finite
element solution / the true solution for space-adapted and static snapshots; POD projection
error. Left: errors for both static and space-adapted snapshots, middle: space-adapted snapshots,
right: static snapshots

Finally, of particular interest is the computational efficiency of the POD reduced-order modeling
utilizing adaptive finite element discretizations. In Table 4, the offline and online computational
times are listed for the case of static and adaptive finite element snapshot generation. We observe

For the snapshot generation we gain a speed up of factor 13.4 in the case of adaptive
discretization in comparison to using a static mesh.

The offline computational times in case of space-adapted snapshots using a construction of

an overlay mesh, interpolation of the snapshots onto this mesh, setting up and solving the
eigenvalue problem and then setting up the reduced-order model takes 606 seconds (10.1
98 4 ADAPTIVITY WITH RESPECT TO SPACE

minutes). In comparison, the construction of the overlay mesh and interpolation is not
needed for a static finite element discretization. However, setting up the eigenvalue problem
and the reduced-order model is more expensive, since it involves matrix multiplications with
larger dimensions due to the larger number of degrees of freedom in the uniform setting.
The offline computation time in case of static finite element discretization is 948 seconds
(15.8 minutes). Thus, the use of spatially adapted snapshots leads to a speed up of factor
1.5 in the offline time for computing the POD model.

The POD online simulation time utilizing ` = 20 POD modes is very fast with only 0.02
seconds and does not depend on any spatial dimension(s). This is 52650 times faster than
the adaptive finite element simulation and even 707150 times faster than the finite element
simulation with static meshes. This speed up even gains greater importance, if we think of
multi-query scenarios like optimal control, where multiple solutions of a dynamical system
are required.

process adaptive FE uniform FE speed up factor

fe simulation 1053 14143 13.4
offline times
→ construct overlay mesh and interpolate 361 –
→ set up and solve eigenvalue problem 73 128 1.7
→ set up ROM 172 820 4.7
POD simulation 0.02 0.02
speed up factor 52650 707150

Table 4: Computational times (sec) for the finite element and POD reduced simulations and
offline calculations utilizing ` = 20 POD modes

The last remark concerns the construction of the snapshot gramian K in (4.14). In this numerical
example, we utilize nested and hierarchical adaptive finite element mesh structures. Thus, the
interpolation of the snapshot data onto the overlay mesh does not lead to any information loss
and can be implemented efficiently. Thus, this approach is favorable. In comparison, we tested
an elementwise computation of K involving the explicit computation of inner products of pairs
of snapshots which leads to the same matrix but is more involving. We exploit the nested and
hierarchical structures, such that a boundary representation of the integral is not needed. Due
to symmetry of K we only calculate the entries on and above the diagonal of K, which are
P n+1 1 2
k=1 k = 2 ((n + 1) + n + 1) entries. For each entry the calculation time is around 0.03 seconds,
which leads to a computation time of around 60000 seconds (16.6 hours) for the matrix K. The
same effort is needed to build Y ∗ AY. In this case, the offline phase even takes longer than the
computational times for the static case. A parallelization can be applied in order to speed up
computations.

4.5.2 Cahn–Hilliard system

As a second example for the semilinear evolution problem (4.4) we consider a phase field system
described by the Cahn–Hilliard equations with an additional transport term given by

ϕt + y · ∇ϕ − b∆µ = 0 in (0, T ] × Ω, (4.41a)

σ
−σε∆ϕ + F 0 (ϕ) = µ in (0, T ] × Ω. (4.41b)
ε
We denote by ϕ the phase field, µ is the chemical potential, F is the free energy, y is a given
velocity field, b is the mobility, σ is the surface tension and ε is the interface parameter. For more
details on phase field systems, we recall Section 2.5. In particular, the Cahn–Hilliard system is a
4.5 Numerical examples 99

nonlinear system. As choices for the free energy F we consider the polynomial free energy Fpol
(2.22) and the relaxed double-obstacle free energy Fs,r (2.23). In order to derive a variational
form of type (4.4), we write (4.41) as a single fourth-order parabolic equation for the phase field
ϕ by eliminating µ. This leads to
σ 0
ϕt + y · ∇ϕ − b∆(−σε∆ϕ + F (ϕ)) = 0 in (0, T ] × Ω. (4.42)
ε
We equip the system with Neumann boundary conditions
∂ϕ ∂ σ
=0 and (−σε∆ϕ + F 0 (ϕ)) = 0 on [0, T ] × ∂Ω,
∂~n ∂~n ε
and an initial condition for the phase field

ϕ(0, ·) = ϕ0 in Ω. (4.43)

We choose V = H 2 (Ω) ∩ H01 (Ω) and H = L2 (Ω) and introduce the bilinear form a : V × V → R
by Z
a(ϕ, v) = bσε∆ϕ∆v + (y · ∇ϕ)vdx
Ω
and define the nonlinear operator N by
bσ
N (ϕ) = − ∆F 0 (ϕ).
ε
Then, the weak form of (4.42) has the form (4.4)-(4.5).

For the actual numerical solution, we utilize the coupled form of the Cahn–Hilliard equation,
which allows to use piecewise linear, continuous P1 − P1 finite elements, see Remark 2.48. For the
time stepping we use an implicit scheme with a convex-concave splitting for the treatment of the
nonlinearity. For more details we refer to the paragraph Time discretization in Section 2.5.2. For
the solution of the nonlinear equations, we utilize Newton’s method with tolerance tol= 10−08
and the maximal number of iterations is maxiter= 25.

The use of h-adaptive finite elements is fundamental, since we consider a diffuse interface ap-
proach. Many degrees of freedom are required at the interfacial regions in order to well reflect the
steep transitions, whereas in regions of pure phases, little numbers of degrees of freedom suffice.
The use of a uniform mesh for such problems would drastically enlarge the computational effort,
such that numerical simulations might be too slow for the requirements of the application or
cannot be performed as often as needed in a certain amount of time. Moreover, storage problems
might occur.

Problem setting

The choices for the problem data and parameters are listed in Table 5. In order to ensure time
step stability of the numerical scheme, we need to impose a CFL (Courant–Friedrichs–Lewy)-type
condition on each triangle T by
hT
max |y(x)| < .
x∈T ∆t
This condition means that the analytical transport of information is slower than the discrete
one. The numerical method is then able to transport information with velocity y. In order to
ensure an unconditionally stable system, we choose for simplicity a fixed time step size ∆t which
is small enough such that
hmin
∆t < ,
ymax
100 4 ADAPTIVITY WITH RESPECT TO SPACE

where ymax = 30 and hmin = 0.00625. The choice of the time step size ∆t = 2.5 · 10−05 fulfills
this requirement.

name value
Ω spatial domain (0, 2) × (0, 1) ⊂ R2
T end time 0.03
b constant mobility 2 · 10−05
ε interface parameter 0.02
σ surface tension 25.98
F nonlinear free energy Fpol (2.22) and Fs,r (2.23) with s = 104 , r = 3, respectively
y velocity y = (y1 , (
0) with
30 · (1 − ((x2 − 0.5)/0.35)2 ) if |x2 − 0.5| < 0.35
y1 (x) =
0 else
for x = (x1 , x2 ) ∈ Ω
ϕ0 initial condition circle with radius 0.35 and center (0.5, 0.5)

Table 5: Problem data and parameters of (4.41) and (4.43)

Finite element results

We start the adaptive finite element Algorithm 1 with an initial triangulation Thinit of the spatial
domain Ω which we choose as a structured criss-cross mesh with 21 node points in x1 -direction
and 11 node points in x2 -direction, see Figure 40 (left). For the Dörfler marking in Defini-
tion 2.34, we set θ = 0.5 and utilize a coarsening strategy according to [114, Section 8.3] with
tolerance θc = 0.01.

Figure 39 shows the finite element solution for the phase field ϕ and the corresponding adapted
meshes at three time instances. The initial condition (a circle) is transported horizontally through
the spatial domain and is deformed according to the given parabolic velocity profile. Note that
the use of the polynomial free energy leads to a violation of the restriction |ϕ| ≤ 1, whereas this
condition is complied better in the case of the relaxed double-obstacle free energy. Moreover, it
can be recognized that in the case of the relaxed double-obstacle free energy the phase transi-
tions are sharper. This leads to a smaller number of necessary degrees of freedom in the adapted
meshes compared to the case of the polynomial free energy. In fact, the number of degrees of
freedom in the adapted meshes varies between 8235 and 15795 in the case of the relaxed double-
obstacle free energy and between 10923 and 25756 in the case of the polynomial free energy.
The overlay of the adapted finite element meshes is displayed in Figure 40. It has 51168 degrees of
freedom in the case of the polynomial free energy and 48653 degrees of freedom in the case of the
relaxed double-obstacle free energy. In comparison, a uniform discretization with discretization
fineness as small as the smallest triangle in the adaptive meshes has 102881 degrees of freedom.
The computational times for the finite element simulation on adapted and uniform static spatial
meshes for different free energies are listed in Table 6. For the polynomial free energies only
3 Newton steps are needed on average for each time instance, whereas for the relaxed double-
obstacle free energy 5 Newton steps are needed on average. The finite element simulation time
using a static uniform spatial discretization takes in the case of the polynomial free energy 30447
seconds (8.5 hours) and 100772 seconds (27.9 hours) for the relaxed double-obstacle free energy.
Thus, the use of spatial adaptivity leads to an offline speed up of factor 4.5 and 8.5, respectively.
4.5 Numerical examples 101

t=0 t = T /2 t=T

Fpol

Fpol

Fs,r

Fs,r

Figure 39: Finite element solution and adapted meshes for the phase field ϕ at three time
instances comparing different free energies F

Figure 40: From left to right: macro mesh and overlay meshes for Fpol and Fs,r

POD reduced-order results

Since we use nested hierarchical meshes for the adaptive finite element solution, we compute
the POD basis by expressing all snapshots with respect to a common finite element basis. As
snapshots we choose the finite element solutions for the phase field ϕ. The resulting POD ba-
sis is then used in a POD Galerkin ansatz for both phase field ϕ and chemical potential µ.
The goal of this example is to investigate the influence of spatial adaptivity for snapshot genera-
tion in regard to the accuracy and efficiency of the reduced-order model for different free energies.

Figure 41 shows the decay of the normalized eigenspectrum for the phase field ϕ and the convex
part F+0 of the nonlinearity utilizing an adaptive versus a static finite element discretization.

Moreover, the use of the polynomial free energy is compared with the Moreau–Yosida relaxed
double-obstacle free energy. We note for the phase field ϕ in the case of using the polynomial free
energy that about the first 60 eigenvalues computed corresponding to space-adapted snapshots
coincide with the eigenvalues corresponding to static snapshots. Then, the eigenvalues corre-
sponding to the static discretization fall below machine precision whereas the eigenvalues for the
adaptive discretization stagnate in the order of around 10−12 . Similar observations apply for the
convex part of the nonlinearity.
In the case of the relaxed double-obstacle free energy, the decay of the eigenvalues is slower
such that more POD modes are needed than in the smooth case in order to achieve the same
approximation quality. This behavior is similar to the decay of the Fourier coefficients in the
102 4 ADAPTIVITY WITH RESPECT TO SPACE

context of trigonometric approximation, where the decay of the Fourier coefficients depends on
the smoothness of the approximated object.

10 0 10 0 10 0

10 -5 10 -5 10 -5

10 -10 10 -10 10 -10

10 -15 10 -15 10 -15

10 -20 10 -20 10 -20

0 250 500 750 1000 0 50 100 150 200 0 250 500 750 1000

Figure 41: Decay of the normalized eigenspectrum for the phase field ϕ (left: all eigenvalues,
middle: first 200 eigenvalues) and the convex part F+ 0 (right) for different free energies using

h-adaptive and static finite element discretizations, respectively

Since the first few POD modes capture the most important information, they are shown in Fig-
ure 42 related to space-adapted snapshots. Note that the POD modes are unique up to the sign.
Analogue to Section 4.5.1, we observe a periodicity in the POD basis functions corresponding to
its index number.

Fpol

Fs,r

Figure 42: POD basis functions ψ1 (left), ψ2 (middle) and ψ5 (right) for polynomial and relaxed
double-obstacle free energy computed from space-adapted snapshots of ϕ

In order to detach the nonlinear POD reduced-order model from the finite element dimension,
we apply the hyper-reduction technique DEIM [50] and choose the same number of DEIM basis
functions as POD modes. We visualize in Figure 43 the POD reduced-order solution utilizing
DEIM for the polynomial free energy and different POD basis lengths. It can be observed that
the oscillations due to the transport of the phase field are smoothened out with an increasing
number of POD modes.

However, we note that in the case of the relaxed double-obstacle free energy, the application of
DEIM is not always successful in the sense that the POD-DEIM reduced-order model does not
always converge. We observe in Figure 41 (right) that the decay of the eigenvalues for the convex
part of the relaxed double-obstacle free energy is very flat. The convex part F+ 0 is zero in the

interval [−1, 1], including the pure phase values ±1. For values smaller than −1 or larger than
+1, the function F+ 0 takes very small respectively very large values (compare Figure 44). This

turns out to be challenging for the application of DEIM. A normalization of the snapshots for
the nonlinearity could not resolve the stability issue in our experiments. In a series of test runs,
we figured out that neither increasing the exponent r to r = 4 or decreasing the parameter s
(s = 1000, s = 100) could solve this issue. Moreover, the use of a fully implicit time discretization
4.5 Numerical examples 103

t=0 t = T /2 t=T

Fpol
` = 30

Fpol
` = 50

Figure 43: POD reduced-order solution for the phase field ϕ utilizing DEIM for Fpol and space-
adapted snapshots at three time instances comparing different POD basis lengths

scheme could not alleviate the stability issues. In a numerical study in [85], we used a separate
POD Galerkin ansatz for the phase field ϕ and the chemical potential µ. Likewise, for the
resulting POD-DEIM reduced-order model we observe stability issues in the case of the relaxed
double-obstacle free energy. A future research direction could be to develop a hyper-reduction
method for such nonlinearities. We further note that in [182] it is proved that the POD-DEIM
reduced-order model of the Allen–Cahn equation is energy stable considering the double-well and
logarithmic potential and using discontinuous Galerkin discretization for the snapshots and an
energy stable average vector field method for the time discretization. A transfer of these ideas
to our problem setting of the Cahn–Hilliard model with a relaxed double-obstacle free energy
could be of interest.

1 100

0.5 50
X: -1
Y: 0
0 0

-0.5 -50
X: -1
Y: -1
-1 -100

-1 0 1 -1 0 1

0 of the polynomial free energy (left) and the relaxed double-obstacle

Figure 44: Convex part F+
free energy (right)

The accuracy of the POD-DEIM reduced-order solution using the polynomial free energy Fpol is
compared to the POD reduced-order solution using the relaxed double-obstacle free energy Fs,r
in case of space-adapted snapshots in Figure 45. In particular, we consider the relative error
between the finite element solution and the reduced-order solution given by
kϕh − ϕ` kL2 (0,T ;Ω)
errF E−P OD =
kϕh kL2 (0,T ;Ω)
As a reference, we plot the POD projection error on a uniform and adaptive mesh, respectively,
which is defined as
kϕh − `i=1 (ϕh , ψi )L2 (Ω) kL2 (0,T ;Ω)
P P
λi
errproj = = P i>` .
kϕh kL2 (0,T ;Ω) i=1 λi

In the case of a uniform static finite element discretization, we observe a convergence of the
POD reduced-order solution to the finite element solution. In the case of adaptive finite element
104 4 ADAPTIVITY WITH RESPECT TO SPACE

discretization for the snapshots we observe a stagnation of the relative error due to the dominance
of the finite element discretization error at this level. This is similar to the results of the previous
example, compare Figure 38.

10 0

10 -2

10 -4

10 -6

10 -8

0 20 40 60 80 100

Figure 45: Relative L2 (0, T ; Ω)-error in ϕ between the reduced-order solution and the finite
element solution for space-adapted snapshots for the polynomial free energy Fpol (POD-DEIM)
and the relaxed double-obstacle free energy Fs,r (POD), respectively. As a reference the POD
projection error is shown for adaptive and static discretizations

Table 6 summarizes the computational times for static and adaptive finite element snapshot
generation with reduced-order simulation times (dimension ` = 30) using the polynomial and the
relaxed double-obstacle free energy, respectively. We note that POD reduced-order models lead
to a speed up of 5.7 to 9.3 in comparison to the adaptive finite element simulations. Especially,
the inclusion of DEIM (we use the same number of DEIM basis functions as POD modes) in
the reduced-order model leads to immense speed up factors of around 30000 for both free energy
functions. Note that the speed up factors are of particular interest in the context of optimal
control problems. As a reference, we compare the results to a linearization of the equation
(recall Section 4.3.2). The resulting relative errors between the finite element solution and the
considered reduced-order solutions are provided.

process CPU times (sec)

Fpol Fs,r
static finite element simulation 30447 100772
adaptive finite element simulation 6722 11861
interpolation onto overlay mesh 690 566
POD basis computation 18 15 relative error
DEIM basis computation 12 11 Fpol Fs,r
1.75 · 10 −04 5.20 · 10−04
POD-ROM 718 1209
1.84 · 10 −04 3.67 · 10−03
POD-ROM-DEIM 0.22 0.38
2.85 · 10 −03 1.94 · 10−02
POD-ROM linearized 0.07 0.07
speed up adaptive FE vs POD-ROM 9.3 5.7
speed up adaptive FE vs POD-ROM-DEIM 30554 31213
speed up adaptive FE vs POD-ROM linearized 96028 98971

Table 6: Left: computational times (sec) for the finite element and reduced-order simulations
and offline calculations utilizing ` = 30 POD modes for different free energies. Right: relative
errors between the finite element solution and the reduced-order solution as specified in the table
left
4.5 Numerical examples 105

4.5.3 Linear heat equation with non-nested grids

We note that this example is published in [85, Example 6.3]. Let us consider again the linear heat
equation with homogeneous Dirichlet boundary data (4.40) from Section 4.5.1. The purpose of
this example is to confirm the numerical applicability of the strategy described in Section 4.2.2.
We set up the matrix K for snapshots given on non-nested meshes which requires the integration
over cut elements. The data is chosen as follows: as spatial domain we choose Ω = (0, 1)×(0, 1) ⊂
R2 , the time interval is [0, T ] = [0, 1], and we utilize a uniform temporal discretization with time
step size ∆t = 0.01. We construct an example such that the analytical solution is known. It is
given by
ȳ(t, x) = sin(πx1 ) · sin(πx2 ) · cos(2πtx1 ),

with x = (x1 , x2 ). The source term f and the initial condition g are chosen accordingly. The
initial condition is projected onto a finite element space with piecewise linear and continuous
node functions on a uniform spatial mesh which is shown in Figure 46 (left). Then, at each time
step, the mesh is disturbed by relocating each mesh node according to the assignment

x1 ← x1 + θ · x1 · (x1 − 1) · (∆t/10),
x2 ← x2 + θ · 0.5 · x2 · (x2 − 1) · (∆t/10),

where θ ∈ R+ is sufficiently small such that all coordinates of the inner node points fulfill
0 < x1 < 1 and 0 < x2 < 1. After relocating the mesh nodes, the heat equation is solved on
this mesh for the next time instance. For this, we use the Lagrange interpolation in order to
interpolate the finite element solution of the previous time step onto the new mesh. The disturbed
meshes at t = 0.5 and t = 1.0 as well as an overlap of two meshes are shown in Figure 46. Now
we have constructed finite element snapshots on non-nested meshes. For this reason, we follow
the strategy explained in Section 4.2.2 and compute the matrix K from (4.14) by evaluating the
inner products of the snapshots, which requires the integration over cut elements.

Figure 46: From left to right: Uniform mesh, disturbed meshes at t = 0.5 and t = 1.0 for θ = 10,
overlap of the mesh at t = 0 with the mesh at t = 1.0

We compute the eigenvalue decomposition of K for different values of θ and compare the results
with a uniform mesh (i.e. θ = 0) in Figure 47. We note that the eigenvalues of the disturbed mesh
are converging to the eigenvalues of the uniform mesh for θ → 0. As expected, the eigenvalue
spectrum depends only weakly on the underlying mesh given assuming that the mesh size is
sufficiently small. The POD reduced-order modeling then follows along the lines of Sections 4.5.1
and 4.5.2.
Our last remark concerns the computational complexity. Solving the heat equation takes 2.1
seconds on the disturbed meshes and 1.8 seconds on the uniform mesh. The computational time
for each entry of the matrix K is 0.022 seconds and computing the eigenvalue decomposition for
K takes 0.0056 seconds. Note that the cut element integration problem for each matrix entry
takes a fraction of time required to solve the finite element problem and can be performed in
parallel.
106 4 ADAPTIVITY WITH RESPECT TO SPACE

10 0 11.7

11.6
10 -5

11.5
λ i / trace

trace
10 -10
11.4

-15
10 uniform mesh 11.3
θ=1 disturbed mesh
θ=10 uniform mesh
10 -20 11.2
0 5 10 15 20 0 2 4 6 8 10
index i θ

Figure 47: Decay of the eigenvalues of the matrix K utilizing different meshes

4.6 POD model order reduction with space-adapted snapshots for incompressible flows

This section is concerned with POD model order reduction for incompressible flows in the case of
h-adapted finite element snapshots. In this case, the difficulty arises that a weak divergence-free
property only holds true in the respective finite element space at the current time level and
thus the challenge appears to derive a stable reduced-order model. We propose two solution
concepts. The first approach is a velocity reduced-order model (Section 4.6.3) and is based on
projecting the snapshots or the velocity POD basis onto a reference velocity space. This projec-
tion is constructed such that the resulting velocity POD modes are weakly divergence-free with
respect to a reference pressure space. The second approach is a velocity-pressure reduced-order
model (Section 4.6.4) and guarantees stability by enriching the velocity POD space with pressure
supremizer functions. In this way, it can be shown that the inf-sup stability condition from the
finite element model carries over to an inf-sup stability condition for the reduced-order model.
Moreover, we address inhomogeneous Dirichlet data by homogenization and constructing a cer-
tain lifting function (Section 4.6.5). As a numerical example, we present in Section 4.6.6 the
benchmark problem of a lid-driven cavity flow utilizing h-adapted finite element snapshots. A
comparison of the two proposed solution approaches is presented with regards to efficiency and
accuracy.

We note that this section is based on a collaboration with Michael Hinze, Jens Lang and Sebastian
Ullmann. A resulting manuscript [86] is accepted for publication in Advances in Computational
Mathematics in 2019.

Literature overview

Let us give an overview of publications which are related to model order reduction of incom-
pressible flows or address stability issues. Due to the amount of scientific contributions to this
issue, we only give a selection of publications here and also refer to the references therein.
The essential idea to identify coherent structures of spatial velocity correlations is given in [133].
In the pioneer work [166, 167, 168], the method of snapshots is introduced. The incompressibility
of the POD modes in case of a static discretization is then used in order to derive a reduced-order
model which only depends on the velocity and the pressure term vanishes. Different problem
settings of a channel flow and a Boussinesq approximation are investigated.
The publication [29] provides a review on proper orthogonal decomposition in the analysis of
turbulent flows with particular attention to the question what can be learnt about the original
partial differential equation.
In [111] reduced-order modeling for simulation and control of fluid flows is presented. Two nu-
merical experiments of a velocity tracking problem in cavity flow and a vorticity control problem
in a channel flow demonstrate the efficiency and feasibility of the reduced-order approximation.
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 107

Stability properties of POD reduced-order approximations for a compressible Navier–Stokes sys-

tem are the content of [110]. POD is interpreted as a filtering process. Moreover, it is suggested
to utilize the H 1 -norm in the POD formulation.
In [121], POD model order reduction is applied to a general equation in fluid dynamics including
the case of the Navier–Stokes equations in two dimensions. The convergence of the POD surro-
gate model is analyzed. The analysis comprises the case in which different time discretization
schemes are used for the full-order and the reduced-order simulations. Perturbation analysis is
carried out for ∆t → 0.
In [184], a parametrized steady incompressible Navier–Stokes system is considered. The reduced
basis method it utilized in order to provide real-time and certified approximations for deployed
systems. Rigorous reduced basis a-posteriori error bounds are provided.
In [41], POD and CVT (centroidal Voronoi tessellation) reduced-order modeling of the Navier–
Stokes equations in 2d are compared. For the numerical example of a T-cell problem, both
approaches seem to be equally effective.
A static unstructured mesh modeling is utilized in [62] in order to resolve both small- and large-
scale fluid flows in 2d and 3d settings. Starting from a uniform mesh, adaptive mesh refinement
is utilized in order to achieve a relatively fine mesh. This mesh is then fixed and the POD method
is carried out in the usual manner.
For a general review of challenges an perspectives in model order reduction for fluid dynamic
settings, we refer to [128]. General principles of reduced-order modeling for parametrized are
discussed, applications to steady and unsteady flow problems are provided and the question of
inf-sup stability is addressed.
Of particular importance for the following is the supremizer stabilization proposed in [160, 22].
A parametrized setting of the Stokes and the steady incompressible Navier–Stokes equations are
considered, respectively. Stability of the reduced-order models is guaranteed by enriching the
reduced basis velocity space with supremizer functions. The stability is first analyzed from a
theoretical point of view and then confirmed in numerical test examples.

4.6.1 Problem setting

Let Ω ⊂ R2 be a bounded domain with Lipschitz boundary ∂Ω, recall Definition 2.17, and let
T > 0 be a fixed end time. We consider an unsteady incompressible flow problem governed
by the Navier–Stokes equations which is given in strong form as follows: find a velocity field
y = (y1 , y2 ) and a pressure p which fulfill

yt + (y · ∇)y − Re−1 ∆y + ∇p = f in (0, T ] × Ω, (4.44a)

∇·y =0 in [0, T ] × Ω, (4.44b)
y=0 on [0, T ] × ∂Ω, (4.44c)
y(0, ·) = y0 in Ω. (4.44d)

The equation (4.44a) is called momentum equation, the (strong) divergence-free constraint (4.44b)
is referred to as continuity equation. By Re we denote the Reynolds number, f is a given body
force and y0 denotes the initial velocity field. Note that we consider a consistent setting in the
sense that the initial data fulfills the boundary condition (4.44c) and is solenoidal, i.e. it fulfills
the strong divergence-free property (4.44b). In the following, we have applications in mind of
laminar flow with low or moderate Reynolds numbers. For a general introduction to the deriva-
tion, mathematical analysis and numerical treatment of the Navier–Stokes equations we refer to
[56, 112, 173], for example.

We note that the divergence constraint (4.44b) encodes the incompressibility of the fluid, or
equivalently conservation of mass. The case of inhomogeneous Dirichlet and initial data is inves-
tigated in Section 4.6.5.
108 4 ADAPTIVITY WITH RESPECT TO SPACE

Weak form

For the ease of presentation, we use the same notation for scalar- and vector-valued functions,
meaning that all components of a vector-valued function belong to the corresponding scalar
function space. The same holds for scalar- and vector-valued operators. We introduce the space
of integrable solenoidal functions by

Hdiv (Ω) := {v ∈ L2 (Ω) : ∇ · v = 0, v · ~n|∂Ω = 0 in the sense of traces}. (4.45)

The pressure is determined only up to an additive constant such that an additional condition is
introduced, e.g. by imposing the integral mean value to vanish. For this reason we introduce the
space Z
L20 (Ω) = {q ∈ L2 (Ω) : qdx = 0}.
Ω

Moreover, we define the space

W∗1 (0, T ; H01 (Ω)) = {v ∈ L2 (0, T ; H01 (Ω)) : vt ∈ L1 (0, T ; H −1 (Ω))}.

The weak form of (4.44) reads as follows: for a given body force f ∈ L2 (0, T ; H −1 (Ω)) and
initial value y0 ∈ Hdiv (Ω) find a velocity y ∈ W∗1 (0, T ; H01 (Ω)) with y(0, ·) = y0 and a pressure
p ∈ L2 (0, T ; L20 (Ω)) such that

d
(y(t), v)L2 (Ω) + c(y(t), y(t), v) + a(y(t), v) + b(v, p(t)) = hf (t), viH −1 (Ω),H01 (Ω) ∀v ∈ H01 (Ω),
dt
(4.46a)
b(y(t), q) = 0 ∀q ∈ L20 (Ω),
(4.46b)

for almost all t ∈ [0, T ]. We use the notations a(u, v) = Re−1 (∇u, ∇v)L2 (Ω) , c(w, u, v) =
((w · ∇)u, v)L2 (Ω) and b(v, q) = −(q, ∇ · v)L2 (Ω) . For existence and uniqueness of a solution
to (4.46) we refer to e.g. [173, Chapter 3, Theorems 3.1 and 3.2].
In order to solve (4.46) numerically, we first discretize in time and then discretize in space which
allows to use a different finite element space at each time level.

Time discretization

For the temporal discretization of (4.46), we apply a fully implicit Euler scheme. Of course,
other time integration schemes are possible. For a given n ∈ N, we introduce a time grid
0 = t0 < t1 < · · · < tn = T and time step sizes ∆tj = tj − tj−1 for j = 1, . . . , n. The
time-discrete system reads as follows: for a given initial velocity y0 ∈ Hdiv (Ω) find sequences
y 1 , . . . , y n ∈ H01 (Ω) and p1 , . . . , pn ∈ L20 (Ω) such that

y j − y j−1
 
,v + c(y j , y j , v) + a(y j , v) + b(v, pj ) = hf (tj ), viH −1 (Ω),H01 (Ω) ∀v ∈ H01 (Ω),
∆tj L2 (Ω)
(4.47a)
b(y j , q) = 0 ∀q ∈ L20 (Ω).
(4.47b)

Note that we applied the box rule in order to approximate the right-hand side time integral in
(4.47a). An initial pressure field can be obtained from an additional pressure Poisson equation,
if required, see e.g. [94, Proposition 2.1, Corollary 2.1].
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 109

Fully discrete form

For the spatial discretization, we utilize h-adaptive finite elements based on LBB (Ladyzhenskaya–
Babuška–Brezzi [19, 40, 124]) stable P2 − P1 Taylor–Hood element pairs denoted by (Vhj , Qjh ) for
j = 1, . . . , n. Details on the construction of the adapted finite element spaces are given be-
low. Let Vh0 denote the finite element space associated with the initial mesh Thinit . For a given
initial condition (yh0 , v)L2 (Ω) = (y0 , v)L2 (Ω) for all v ∈ Vh0 with y0 ∈ Hdiv (Ω), let us consider
the fully discrete Navier–Stokes system which reads as follows: find yh1 ∈ Vh1 , . . . , yhn ∈ Vhn and
p1h ∈ Q1h , . . . , pnh ∈ Qnh such that
!
yhj − yhj−1
,v + c(yhj , yhj , v) + a(yhj , v) + b(v, pjh ) = hf (tj ), viH −1 (Ω),H01 (Ω) ∀v ∈ Vhj ,
∆tj 2
L (Ω)
(4.48a)
b(yhj , q) =0 ∀q ∈ Qjh .
(4.48b)
For existence of a unique solution to (4.48), we refer to [173, Chapter 3, §5 Scheme 5.1’]. Note
that in each time step the computation of the inner product of a velocity from the previous time
step yhj−1 ∈ Vhj−1 with a test function v ∈ Vhj is required. In [32] it is shown that for spatially
adapted mesh discretizations, the pressure diverges with order ∆t−1 for ∆t → 0. It is shown
that the discrete pressure is bounded if the L2 -projection of the velocity from the previous time
instance is weakly divergence-free with respect to test functions of the current finite element
velocity space. Suitable projections are proposed to circumvent the divergence of the pressure.
Alternatively, in [14, Section 6.2], an index-1 formulation is used in order to derive a robust
scheme.

For the construction of the adapted finite element spaces (Vhj , Qjh ), we follow the usual adaptive
cycle, recall Figure 3. In particular, we make the following choices in Algorithm 1: for the
solution of the nonlinear algebraic saddle point problem (4.48), we utilize Newton’s method. The
estimation of the spatial error on each triangle is based on a residual a-posteriori error estimator
in the spirit of [4]. In particular, we obtain on each triangle T of the current triangulation Thj
the error indicator
2
yhj − yhj−1
ηTj = h2T + yhj · ∇yhj − Re−1 ∆yhj + ∇pjh − f (tj )
∆tj
L2 (T )
!1
2
2 1 2
+ ∇ · yhj − Re−1 ∇yhj · ~n + pjh~n E ∀T ∈ Thj
X  
+ hE
2
L (T ) 2 L2 (E)
E∈∂T \∂Ω
(4.49)

for j = 1, . . . , n which assumes f (tj ) ∈ L2 (Ω). We denote by h2T the triangle area and hE is the
edge length. Since we do not expect drastic dynamical changes from one time step to another
for small time instances, we mark every triangle of the current triangulation and perform one
coarsening step. This ensures the possibility of reaching the initial mesh Thinit after a finite
number of time steps. Although this choice might lead to a finer triangulation than actually
postulated, we expect the advantage that only a small number of refinement steps are needed in
the next time iteration.

4.6.2 POD reduced-order modeling

Assume we are given a reduced velocity space V` and a reduced pressure space Q` . We discuss
the detailed construction later. In order to derive a reduced-order model, we replace the test
110 4 ADAPTIVITY WITH RESPECT TO SPACE

spaces in (4.47) by the reduced spaces V` and Q` , respectively. The reduced-order problem
reads as follows: for given initial condition (y`0 , ψ)L2 (Ω) = (y0 , ψ)L2 (Ω) with y0 ∈ Hdiv (Ω) find
y`1 , . . . , y`n ∈ V` and p1` , . . . , pn` ∈ Q` such that
!
y`j − y`j−1
,v + c(y`j , y`j , v) + a(y`j , v) + b(v, pj` ) = hf (tj ), viH −1 (Ω),H01 (Ω) ∀v ∈ V` ,
∆tj 2
L (Ω)
(4.50a)
b(y`j , q) =0 ∀q ∈ Q` .
(4.50b)

This is a system of algebraic equations for the expansion coefficients of the reduced solutions and
does not depend on the finite element dimension, see [168, Section 1]. Note that the stability of
the system (4.50) is not guaranteed for all pairs (V` , Q` ). In the following we propose two choices
for (V` , Q` ) which lead to stable reduced-order models.
The first approach (Section 4.6.3) is a velocity reduced-order model. It relies on a reference
pressure space paired with a velocity POD space, where the POD basis functions are weakly
divergence-free with respect to the reference pressure space. This enables a cancellation of the
pressure term in (4.50a) and the continuity equation (4.50b) is fulfilled by construction. The
stability of the resulting system is then given by [56, 173].
The second approach (Section 4.6.4) is a velocity-pressure reduced-order model. Separate POD
basis functions are computed for the velocity and for the pressure. Stability is achieved by en-
riching the velocity POD space with pressure supremizer functions in the spirit of [22, 160].

In the following we make use of a reference velocity space Ṽ ⊂ H01 (Ω) and a corresponding
reference pressure space Q̃ ⊂ L20 (Ω), such that the pair (Ṽ , Q̃) is inf-sup stable. These spaces
do not have to fulfill further assumptions and are therefore kept general. One option could
be to choose common finite element spaces Ṽ and Q̃ which contain all finite element spaces,
i.e. Vh1 , . . . , Vhn ⊂ Ṽ and Q1h , . . . , Qnh ⊂ Q̃. For example, if the meshes Th1 , . . . , Thn are nested,
hierarchical meshes, then the spaces Ṽ and Q̃ can be constructed associated with the overlay of
all meshes. However, it is also possible to choose Ṽ and Q̃ independently of the snapshot spaces.
This enables to decouple the reduced-order solution from the finite element discretization.

4.6.3 Velocity reduced-order model

In this section, we derive a POD reduced-order model which only depends on the velocity variable
while the pressure term and the continuity equation are canceled out. This approach is in
the spirit of [166, 167] and based on the fact that the POD modes are weakly divergence-free.
The challenge in the context of spatially adapted snapshots is that weakly divergence-free refers
to the test space which can be a different space at each time instance. More precisely, for
j = 1, . . . , n the solution yhj ∈ Vhj of (4.48) fulfills a weak divergence-free property with respect
to the corresponding pressure space Qjh , i.e.

b(yhj , q) = 0 ∀q ∈ Qjh .

However, in general yhj is not weakly divergence-free with respect to the other pressure spaces
Qih for i 6= j. Thus, for a linear combination of the snapshots it cannot be guaranteed that a
weak divergence-free property is fulfilled in a common sense. For this reason, the usual velocity
ROM does not work since the POD basis functions are a linear combination of the snapshots.
In order to tackle this issue, we construct a modified velocity POD basis which is weakly
divergence-free with respect to a reference pressure space. Two approaches of constructing a
suitable modified velocity POD basis are proposed. The first strategy is based on a projection
of the original snapshots meaning that we first project the snapshots such that they are weakly
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 111

divergence-free in a common sense and then compute a POD basis from the projected snapshots
(first-project-then-reduce). The second approach is based on projected POD modes, meaning
that we first compute a POD basis from the original snapshots and then project the POD basis
functions such that they are weakly divergence-free in a common sense (first-reduce-then-project).

Optimal projection onto a weakly divergence-free space

Our aim is to define weakly divergence-free approximations in a more general sense. We provide
a procedure that can be applied to problems with inhomogeneous Dirichlet data as well. To this
end, we introduce a Dirichlet lifting function g, which will be specified in Section 4.6.5. In the
context of homogeneous Dirichlet data, we set g ≡ 0.

For a given function u in a Hilbert space X we want to find a function ũ ∈ Ṽ ⊂ X such that
ũ + g is weakly divergence-free with respect to a space Q̃. To be more precise, we want to solve
the following equality constrained minimization problem:

Problem 4.5. For given u ∈ X and sufficiently smooth g, find ũ ∈ Ṽ ⊂ X which solves
1
min kv − uk2X subject to b(v + g, q) = 0 ∀q ∈ Q̃.
v∈Ṽ 2

Note that Problem 4.5 has a unique solution ũ ∈ Ṽ . We define the projection u 7→ Pg (u) = ũ.
In order to compute the solution to Problem 4.5, the usual Lagrange approach can be followed,
recall Section 2.3. The resulting optimality system is given by the following saddle point problem:

Problem 4.6. For given u ∈ X and sufficiently smooth g, find ũ ∈ Ṽ ⊂ X and λ ∈ Q̃ such that

(ũ, w)X + b(w, λ) = (u, w)X ∀w ∈ Ṽ ,

b(ũ, q) = −b(g, q) ∀q ∈ Q̃.

Note that the solution ũ ∈ Ṽ to Problem 4.6 is unique. If (Ṽ , Q̃) forms an inf-sup stable pair,
then λ ∈ Q̃ is unique. For the choice X = H01 (Ω) the Problem 4.6 is a Stokes problem, for
X = L2 (Ω), the Problem 4.6 is a weak form of the Leray projection.

a) First-project-then-reduce

The basic idea of this approach is to project the original snapshots, i.e. the velocity solution
of (4.48), in order to obtain functions which are weakly divergence-free with respect to the
reference pressure space Q̃. As a consequence, the resulting POD basis inherits the property of
weak-divergence by construction. In particular, for given snapshots yh1 ∈ Vh1 , . . . , yhn ∈ Vhn we
solve Problem 4.5 with g ≡ 0. Then, the projected snapshots ỹ 1 = P0 (yh1 ), . . . , ỹ n = P0 (yhn ) live
in
Ṽdiv := {v ∈ Ṽ : b(v, q) = 0 ∀q ∈ Q̃}.
From these projected snapshots ỹ 1 , . . . , ỹ n , we compute a POD basis according to (2.7) with basis
length `y . The resulting POD space V` := span(ψ1 , . . . , ψ`y ) fulfills the property V` ⊂ Ṽdiv ⊂ Ṽ .
Thus, b(v, q) = 0 holds true for all v ∈ V` and all q ∈ Q̃. Consequently, for this choice of V` in
(4.50), together with Q` = Q̃, the pressure term vanishes and the continuity equation is fulfilled
by construction. The resulting velocity reduced-order model reads as follows: for a given initial
condition (y`0 , ψ)L2 (Ω) = (y0 , ψ)L2 (Ω) with y0 ∈ Hdiv (Ω), find y`1 , . . . , y`n ∈ V` such that
!
y`j − y`j−1
,v + c(y`j , y`j , v) + a(y`j , v) = hf (tj ), viH −1 (Ω),H01 (Ω) ∀v ∈ V` . (4.51)
∆tj
L2 (Ω)
112 4 ADAPTIVITY WITH RESPECT TO SPACE

Obviously, the velocity reduced-order model (4.51) only depends on the velocity variable and the
pressure is eliminated. However, in the case of snapshot generation on static spatial meshes, it
is possible to reconstruct a reduced pressure afterwards by solving a discrete reduced pressure
Poisson equation, see [180], for example. A transfer of this concept to the case of space-adapted
snapshots is not carried out within the scope of this work.

Computational complexity

Concerning the computational complexity, we note that Problem 4.5 has to be solved for each
snapshot, which means the solution of a saddle point problem with reference spaces Ṽ and Q̃,
followed by the computation of a POD basis for the velocity field. Concerning the online compu-
tational costs, in each time step we solve a nonlinear algebraic system of equations with Newton’s
method. In each Newton step, we need to build a Jacobian matrix and a right-hand side and,
subsequently, solve a dense linear system using a direct method. Therefore, we find that solving
the velocity reduced-order model (4.51) is of order O(`3y ).

b) First-reduce-then-project

The basic idea of this approach is to compute a POD basis from the original velocity solutions
of (4.48) and project the resulting POD basis functions afterwards such that they fulfill a weak
divergence-free property. In particular, for given snapshots yh1 ∈ Vh1 , . . . , yhn ∈ Vhn , we define inter-
polated snapshots ŷ j = πṼ yhj ∈ Ṽ for j = 1, . . . , n where πṼ denotes the Lagrange interpolation
operator onto the velocity reference space Ṽ . From these interpolated snapshots, we compute
POD basis functions ψ̂1 , . . . , ψ̂`y ∈ Ṽ according to (2.7) with POD basis length `y . These POD
basis functions in general do not live in Ṽdiv . Thus, they are projected onto the space Ṽdiv by solv-
ing (4.5) with g ≡ 0. Then, the projected POD basis functions ψ1 = P0 (ψ̂1 ), . . . , ψ`y = P0 (ψ̂`y )
live in Ṽdiv . Choosing V` := span(ψ1 , . . . , ψ`y ) in (4.50) together with Q` = Q̃ leads to a velocity
reduced-order model of the form (4.51). Note that, in general, the reduced space constructed in
this approach does not coincide with the reduced space constructed according to a) first-project-
then-reduce.

Computational complexity

The computational complexity of the approach described in this subsection comprises the compu-
tation of a POD basis and, afterwards, the solution of Problem 4.5 for each POD basis function,
i.e. `y times. This makes the approach first-reduce-then-project cheaper than the approach first-
project-then-reduce, which requires n solutions of Problem 4.5, and `y ≤ n. Then, the costs of
setting up and solving the reduced-order model are equivalent.

4.6.4 Velocity-pressure reduced-order model

In this section, we derive a POD reduced-order model which depends on both velocity and pres-
sure variable. This approach is in the spirit of [22, 160] and is based on a construction of an
inf-sup stable pair of velocity and pressure reduced spaces. In particular, we compute separate
POD reduced spaces for the velocity and the pressure and enrich the velocity POD space with
suitable pressure supremizer functions in order to guarantee an inf-sup stability condition.

Construction of inf-sup stable POD spaces

From the velocity solutions yh1 ∈ Vh1 , . . . , yhn ∈ Vhn and pressure solutions p1h ∈ Q1h , . . . , pnh ∈ Qnh
of problem (4.48), we define interpolated velocity snapshots ŷ j = πṼ yhj ∈ Ṽ and interpolated
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 113

pressure snapshots p̂j = πQ̃ pjh ∈ Q̃ for j = 1, . . . , n, where πṼ and πQ̃ are the Lagrange in-
terpolation operators onto the reference velocity space Ṽ and the reference pressure space Q̃,
respectively. For the velocity POD, we compute `y POD basis functions according to (2.5). As a
result, we obtain velocity POD basis functions ψ1 , . . . , ψ`y ∈ span(ŷ 1 , . . . , ŷ n ). For the pressure,
we compute `p pressure POD basis functions according to (2.5). Thus, we obtain pressure POD
basis functions φ1 , . . . , φ`p ∈ span(p̂1 , . . . , p̂n ).

The goal is to construct a pair of reduced spaces (V` , Q` ) which ensures stability of the POD
reduced-order model (4.50). We introduce an inf-sup stability criterion as follows: there exists a
constant β` defined by
b(v, q)
β` := inf sup . (4.52)
q∈Q` v∈V` kvkH 1 (Ω) kqkL2 (Ω)
0
q6=0 v6=0

which needs to be greater than zero (β` > 0). We make the following choices. As reduces spaces
we take

V` := span(ψ1 , . . . , ψ`y , ψ̄1 , . . . , ψ̄`p ), (4.53a)

Q` := span(φ1 , . . . , φ`p ), (4.53b)

with supremizer functions ψ̄1 , . . . , ψ̄`p . In the following, we explain the computation of the
supremizer functions and show that this choice of reduced spaces (V` , Q` ) leads to a POD reduced-
order model which fulfills the inf-sup stability condition in the sense of (4.52).

a) Supremizer computation

We introduce a linear map T : L20 (Ω) → Ṽ as follows: for given q ∈ L20 (Ω), find Tq ∈ Ṽ such that

(Tq, v)H01 (Ω) = b(v, q) ∀v ∈ Ṽ . (4.54)

From the Riesz representation Theorem 2.8 follows

b(v, q) (Tq, v)H01 (Ω)

kTqkṼ = sup = sup ∀q ∈ L20 (Ω), (4.55)
v∈Ṽ kvkH01 (Ω) v∈Ṽ kvkH01 (Ω)
v6=0 v6=0

which means that

(Tq, v)H01 (Ω)
Tq = arg sup ∀q ∈ L20 (Ω). (4.56)
v∈Ṽ kvkH01 (Ω)
v6=0

Now, for each pressure POD function φj for j = 1, . . . , `p we compute a function ψ̄j ∈ Ṽ which
solves (4.54), i.e. ψ̄j = Tφj .

b) Inf-sup stability of the velocity-pressure reduced-order model

In order to show that the choice of reduced spaces (V` , Q` ) according to (4.53) is inf-sup stable
in the sense of (4.52), we relate the stability constant β` of the reduced-order model with the
stability constant β̃h of the reference spaces (Ṽ , Q̃) and proceed analogously to [22, Proposition
2] and [160, Lemma 3.1]:
114 4 ADAPTIVITY WITH RESPECT TO SPACE

b(v, q)
0 < β̃h = inf sup FE stability
q∈Q̃ v∈Ṽ kvkH01 (Ω) kqkL2 (Ω)
q6=0 v6=0
b(v, q)
≤ inf sup Q` ⊂ Q̃
q∈Q`
v∈Ṽ kvkH01 (Ω) kqkL2 (Ω)
q6=0 v6=0

b(Tq, q)
= inf (4.55), (4.56)
q∈Q` kTqkH01 (Ω) kqkL2 (Ω)
q6=0

b(v, q)
≤ inf sup
q∈Q` v∈{Tp:p∈Q } kvkH 1 (Ω) kqkL2 (Ω)
` 0
q6=0 v6=0
b(v, q)
≤ inf sup = β` . V` = span{ψ1 , . . . , ψ`y , ψ̄1 , . . . , ψ̄`p }
q∈Q` v∈V` kvkH01 (Ω) kqkL2 (Ω)
q6=0 v6=0

Consequently, the inf-sup stability of the finite element model carries over to an inf-sup stability
of the reduced-order model.

Computational complexity

Concerning the complexity of the construction of the velocity-pressure reduced-order model,

there are two major differences to the velocity reduced-order model. First of all, the velocity-
pressure reduced-order model requires an additional computation of pressure POD basis func-
tions, which is not needed for the pure velocity reduced-order model. Second, the construction
of the divergence-free POD modes in the velocity reduced-order approach requires the solution
of Problem 4.6, whose complexity depends on dim(Ṽ ) + dim(Q̃). In contrast, the construction of
the supremizer functions in the velocity-pressure reduced-order approach requires the solution of
(4.54), whose complexity depends on dim(Ṽ ). Concerning the online computational costs, the
solution of the velocity-pressure reduced-order model is of order O((`y + `p )3 ).

4.6.5 Inhomogeneous Dirichlet and initial data

So far, we considered the incompressible Navier–Stokes system with homogeneous Dirichlet data.
In the following, we extend the scope to problems involving inhomogeneous boundary conditions.
In the context of POD-Galerkin modeling based on adaptive finite element snapshots, the main
challenge is to find suitable continuous extensions of the Dirichlet data, which need to be sub-
tracted from the snapshots before a POD basis is computed. For the derivation of a velocity
POD-Galerkin model, we must ensure that these continuous extensions fulfill the correct weak
divergence-free property.

Inhomogeneous problem setting

We extend (4.44) to the case of inhomogeneous Dirichlet boundary conditions by introducing

Dirichlet boundary data yD : [0, T ] × ∂Ω → R. The resulting problem reads as follows: find a
velocity field y̆ and a pressure field p such that

y̆t + (y̆ · ∇)y̆ − Re−1 ∆y̆ + ∇p = f in (0, T ] × Ω, (4.57a)

∇ · y̆ = 0 in [0, T ] × Ω, (4.57b)
y̆ = yD on [0, T ] × ∂Ω, (4.57c)
y̆(0, ·) = y0 in Ω. (4.57d)
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 115

where ∇ · y0 = 0 in Ω and y0 = yD (0, ·) on ∂Ω. In the following, we derive a homogenized version

of this problem, which provides a foundation for the subsequent finite element discretization and
reduced-order modeling.

Homogenized problem setting

We assume that the boundary function yD is sufficiently smooth such that it can be continuously
extended by a function g : [0, T ] × Ω̄ → R with g(t, ·)|∂Ω = yD (t, ·) for t ∈ [0, T ]. The regularity
requirements on g are such that a unique weak solution exists, see [154]. We provide a concrete
choice of g by computation.
We homogenize (4.57) by subtracting the boundary function from the inhomogeneous velocity
solution such that the homogeneous velocity field is given by y = y̆ − g. Substituting y̆ in (4.57),
we obtain the following homogenized problem: find a velocity y and a pressure p such that

yt + (y · ∇)y + (g · ∇)y + (y · ∇)g− Re−1 ∆y + ∇p (4.58a)

−1
= f − (g · ∇)g + Re ∆g − gt in (0, T ] × Ω, (4.58b)
∇·y = −∇·g in [0, T ] × Ω, (4.58c)
y=0 on [0, T ] × ∂Ω, (4.58d)
y(0, ·) = y0 − g in Ω, (4.58e)

where ∇ · y0 = 0 in Ω and y0 = g(0, ·) on ∂Ω.

In order to derive a time-discrete weak form of the homogenized problem, we implement the
time integrals involving the Dirichlet data using the right-sided rectangle rule, which evaluates
the Dirichlet data at the new time instance. For ease of notation, we define f j := f (tj ) and
g j := g(tj ) for j = 0, . . . , n. Consequently, the time-discrete weak form of the homogenized
problem consists in finding sequences y 1 , . . . , y n ∈ H01 (Ω) and p1 , . . . , pn ∈ L20 (Ω), for given
y 0 = y0 − g 0 with y0 ∈ Hdiv (Ω), such that
 j
y − y j−1

,v + c(y j , y j , v) + c(g j , y j , v) + c(y j , g j , v) + a(y j , v) + b(v, pj ) (4.59a)
∆tj 2
L (Ω)
 j
g − g j−1

j j j j
= hf , viH −1 (Ω),H01 (Ω) − c(g , g , v) − a(g , v) − ,v ∀v ∈ H01 (Ω),
∆tj L2 (Ω)
(4.59b)
b(y j , q) = −b(g j , q) ∀q ∈ L20 (Ω).
(4.59c)

We apply an adaptive finite element method. Let us denote by Vh0 the velocity finite element space
associated with the initial mesh Th0 . For a given initial condition (yh0 , v)L2 (Ω) = (y0 − g 0 , v)L2 (Ω)
for all v ∈ Vh0 with y0 ∈ Hdiv (Ω) the fully discrete homogenized Navier–Stokes problem reads as
follows: find yh1 ∈ Vh1 , . . . , yhn ∈ Vhn and p1h ∈ Q1h , . . . , pnh ∈ Qnh such that
!
yhj − yhj−1
,v + c(yhj , yhj , v) + c(g j , yhj , v) + c(yhj , g j , v) + a(yhj , v) + b(v, pjh ) (4.60a)
∆tj 2
L (Ω)
 j
g − g j−1

= hf j , viH −1 (Ω),H01 (Ω) − c(g j , g j , v) − a(g j , v) − ,v ∀v ∈ Vhj ,
∆tj L2 (Ω)
(4.60b)
b(yhj , q) = −b(g j , q) ∀q ∈ Qjh .
(4.60c)
116 4 ADAPTIVITY WITH RESPECT TO SPACE

Lifting function

Based on (4.60), approximations to the inhomogeneous solutions y̆(tj ) of (4.57) are given by
ŷhj := yhj + g j for j = 0, . . . , n. Regardless of the choice of lifting functions g 0 , . . . , g n , we
can guarantee that ŷh0 fulfills the initial condition (4.57d) and ŷh1 , . . . , ŷhn fulfill the Dirichlet
condition (4.57c) by construction. Nevertheless, in order to solve (4.60) numerically, candidates
of g 0 , . . . , g n must be fixed, at least implicitly. Our approach to reduced-order modeling is not
tied to a particular choice. In the following, we provide suitable candidates which can be realized
without the need to modify usual finite element codes.
Note that for the velocity finite element spaces we have Vhj ⊂ H01 (Ω). Thus, for the context
of inhomogeneous Dirichlet conditions, we start by introducing the spaces VDj for j = 1, . . . , n,
which denote the spaces spanned by the union of the finite element basis functions of Vhj and
the finite element basis functions associated with the corresponding Dirichlet boundary nodes.
We assume that in (4.60) the integrals involving g j are approximated by a numerical quadrature
which consists of substituting the Lagrange interpolation of g j onto VDj and integrating the
resulting piecewise polynomials exactly. We assume that by a finite number of refinements of
any VD1 , . . . , VDn one can find a reference finite element space ṼD such that VD1 , . . . , VDn ⊂ ṼD .
Now, for j = 1, . . . , n, we define lifting functions g j as a sufficiently smooth continuous extension
of the Dirichlet data yD (tj ) into the domain Ω such that g j is zero at all nodes of the reference
finite element space Ṽ . This is equivalent to the standard approach of using an approximate
Dirichlet lifting given by a Lagrange interpolation of the Dirichlet data onto the finite element
space at the boundary and a subsequent continuous extension using the finite element space in
the interior, because we have
(
g j = yD (t) at all Dirichlet nodes of VDj ,
gj = 0 at all interior nodes of VDj .

A disadvantage of the standard approach is that it implies a Dirichlet lifting which satisfies the
boundary data only in an approximate sense. Our description, on the other hand, delivers an
output which is exact at the boundary. In particular, we have
 j j
ŷh = yh at all interior nodes of ṼD ,

ŷhj = g j at all Dirichlet nodes of ṼD ,
 j

yh = 0 at all Dirichlet nodes of ṼD .

This holds for all ṼD which fulfill our assumptions, without the need to specify a concrete
candidate of ṼD during the adaptive finite element simulation. When the adaptive finite element
simulation is finished and VD1 , . . . , VDn are available, some ṼD can be computed by refinement and
g j can be evaluated at all nodal points of ṼD . Therefore, we are even able to formulate a finite
element discretization of (4.59) on (Ṽ , Q̃) using the same g 0 , . . . , g n as in (4.60). Moreover, we
are able to solve (4.59) on subspaces of (Ṽ , Q̃) using the same g 0 , . . . , g n as in (4.60).

Remark 4.7. In principle, it is possible to impose a weak divergence-free constraint on the

homogenized velocity finite element solution by using lifting functions which are computed such
that b(g j , q) = 0 for all q ∈ Qjh for j = 0, . . . , n. But this would require the solution of an
additional stationary finite element problem for each g j . Moreover, this would not automatically
imply a weak divergence-free property with respect to a reference pressure space Q̃. An alternative
to the implicit choice of the Dirichlet lifting function is its explicit choice at the level of the strong
formulation (4.58). Disadvantages would be a possibly larger support of such a lifting function
and the effort of actually finding a suitable function. Also in this case, it would be attractive
to impose a strong divergence-free constraint on g j , because this implies a weakly divergence-free
homogenized velocity field yhj . Still, finding a suitable candidate may be challenging in general.
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 117

Velocity reduced-order model for the inhomogeneous setting

In the following, we derive a reduced-order model for the velocity field, based on the semi-
discretized problem (4.59). We introduce the projections PgV,Q according to Problem 4.5 by

Problem 4.8. For given u ∈ X, sufficiently smooth g and given spaces V and Q, find PgV,Q (u)
which solves
1
min kv − uk2X subject to b(v + g, q) = 0 ∀q ∈ Q.
v∈V 2

We use X = V = H01 (Ω) and Q = L20 (Ω) in the following. By definition of this projection onto
a divergence-free space, we have b(PgV,Q
j (0) + g j , q) = 0 for all q ∈ L20 (Ω) and j = 0, . . . , n. In
(4.59), we substitute g j = g j + PgV,Q
j (0) − PgV,Q
j (0) and reformulate the equations so that we can
set ỹ j = y j − PgV,Q
j (0). We obtain the following problem, which is equivalent to (4.59): for given
ỹ 0 = y0 − g̃ 0 with y0 ∈ Hdiv (Ω) and g̃ j = g j + PgV,Q j (0) for j = 0, . . . , n, find ỹ 1 , . . . , ỹ n ∈ H01 (Ω)
and p1 , . . . , pn ∈ L20 (Ω) such that
 j
ỹ − ỹ j−1

,v + c(ỹ j , ỹ j , v) + c(g̃ j , ỹ j , v) + c(ỹ j , g̃ j , v) + a(ỹ j , v) + b(v, pj ) (4.61a)
∆tj L2 (Ω)
 j
g̃ − g̃ j−1

j j j j
= hf , viH −1 (Ω),H01 (Ω) − c(g̃ , g̃ , v) − a(g̃ , v) − ,v ∀v ∈ H01 (Ω),
∆tj 2
L (Ω)
(4.61b)
b(ỹ j , q) = 0 ∀q ∈ L20 (Ω).
(4.61c)

One can show that (4.60) is a discretization of (4.61) by replacing (H01 (Ω), L20 (Ω)) with (Vhj , Qjh )
V j ,Qjh
for j = 1, . . . , n in (4.61). Then, for the resulting solution holds ỹ j = yhj − Pgjh if g̃ j =
Vhj ,Qjh
gj + P gj
(0) for j = 0, . . . , n. In this way, (4.60) can be used to obtain approximate solutions
of (4.61).

We base the model equation on a discretization of (4.61) using the pair of reference spaces
(Ṽ , Q̃) as test spaces together with g̃ j = g j + PgṼj ,Q̃ (0). The resulting solutions are approxima-
tions to the solutions of (4.61) using the original pair of spaces (H01 (Ω), L20 (Ω)). We have shown
that the solutions of (4.61) using the original pair of spaces (H01 (Ω), L20 (Ω)) are approximated
V j ,Qjh
by yhj − Pgjh (0) for j = 1, . . . , n resulting from (4.60). But these solutions are not weakly
divergence-free with respect to the reference pair of spaces (Ṽ , Q̃). Therefore, we have to modify
them.

V j ,Qjh
Following the velocity-ROM approach a) in Section 4.6.3 we substitute yhj − Pgjh (0) by their
approximations PgṼj ,Q̃ (yhj ) − PgṼj ,Q̃ (0) for j = 1, . . . , n. Using now PgṼj ,Q̃ (yhj ) − PgṼj ,Q̃ as snapshots
in a POD yields POD basis functions

ψi ∈ span(Pg1 (yh1 ) − Pg1 (0), . . . , Pgn (yhn ) − Pgn (0)) ⊂ Ṽdiv ∀i = 1, . . . , `y

for some `y ≤ n, which define a POD space V` := span(ψ1 , . . . , ψ`y ) ⊂ Ṽdiv .

In the time-discrete equation (4.61), we use the pair (V` , Q̃) as test and trial spaces. Consequently,
the continuity equation is fulfilled by construction. For the pressure term, we have b(v, pj ) = 0
for all v ∈ V` and all pj ∈ Q̃. The resulting reduced-order model is given by the following set
118 4 ADAPTIVITY WITH RESPECT TO SPACE

of equations: for y`0 = y0 − g̃ 0 with y0 ∈ Hdiv (Ω) and g̃ j = g j + PgṼj ,Q̃ (0) for j = 0, . . . , n, find
y`1 , . . . , y`n ∈ V` such that
!
y`j − y`j−1
,v + c(y`j , y`j , v) + c(g̃ j , y`j , v) + c(y`j , g̃ j , v) + a(y`j , v)
∆tj
L2 (Ω) (4.62)
 j
g̃ − g̃ j−1

j j j
= hf , viH −1 (Ω),H01 (Ω) − c(g̃ , g̃ , v) − a(g̃ , v) −j ,v ∀v ∈ V` .
∆tj L2 (Ω)

Remark 4.9. Concerning the computational complexity, we have to additionally consider the
projections PgṼj ,Q̃ (0) for j = 0, . . . , n in comparison to the homogeneous case. Therefore, the
solution of Problem 4.5 has to be computed n + 1 times additionally to the projections of the
homogeneous solutions yhj .
Following the velocity-ROM approach b) in Section 4.6.3, we first introduce a set of modified ho-
mogeneous solutions ŷ j −PgṼj ,Q̃ (0), for j = 1, . . . , n. These modified snapshots can be constructed
using e.g. a Lagrange interpolation of the original homogeneous solutions yhj onto the reference
j ,Qj
space Ṽ and an approximation of PgVj (0) for j = 1, . . . , n. From these modified snapshots,
we compute a POD basis ψ̂1 , . . . , ψ̂`y ∈ Ṽ . Note that these modes are in general not discrete
divergence-free. Thus, they are then projected onto the space Ṽdiv by solving Problem 4.6 with
g = 0. This leads to a divergence-free velocity POD space V` = span(ψ1 , . . . , ψ`y ) ⊂ Ṽdiv . Replac-
ing (H01 (Ω), L20 (Ω)) by the pair (V` , Q̃) in (4.61) leads to a reduced-order model of the form (4.62).

Velocity-pressure reduced-order model for the inhomogeneous setting

To derive a velocity-pressure reduced-order model of the homogenized problem (4.60), we require

a suitable inf-sup stable pair of reduced spaces. Since the homogenization does not alter the
bilinear form b(·, ·), the inf-sup stability criterion stays the same. Therefore, we compute a
pressure reduced space Q` and a velocity reduced space V` like in Section 4.6.4, but using Lagrange
interpolated velocity and pressure snapshots of (4.60) instead of (4.48). We derive a stable POD-
Galerkin model from the time-discrete problem (4.59) by using the pair (V` , Q` ) as test and trial
spaces.
We solve the reduced-order model for the POD approximations y`1 , . . . , y`n of the homogeneous
velocity fields and the POD approximations p1` , . . . , pn` of the pressure fields. Finally, y`j + g j
is a time-discrete reduced-order approximation of the velocity solution of the inhomogeneous
problem.

4.6.6 Numerical example

We consider the benchmark problem of a lid-driven cavity flow as a numerical example. The
implementation is done in C++ and we use FEniCS [13, 132] for the solution of the differential
equations using finite elements and ALBERTA [161] for dealing with hierarchical meshes. All
simulations are performed on a compute server with 512 GB RAM.

Problem setting

As spatial domain, we consider the open bounded unit square Ω = (0, 1) × (0, 1) ⊂ R2 and set the
end time T = 1. The governing equations are the Navier–Stokes equations with inhomogeneous
Dirichlet conditions given in (4.57). We utilize a regularized lid-velocity according to [76, 112],
which implies ∇ · y = 0 in the corners. An additional regularization in time allows for a smooth
t (t)y x (x) for
start up from y0 (x) = 0. In particular, the Dirichlet data is given as yD (t, x) = yD D
all (t, x) ∈ [0, 1] × ∂Ω, where
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 119

(
t 1 − 14 (1 − cos((0.1 − t)π/0.1))2 if t ∈ [0, 0.1),
yD (t) =
1 if t ∈ [0.1, 1],



 1 − 41 (1 − cos((0.1 − x1 )π/0.1))2 if x2 = 1, x1 ∈ [0, 0.1],

x
1 if x2 = 1, x1 ∈ (0.1, 0.9),
yD (x) = 1 2


 1 − 4 (1 − cos((x1 − 0.9)π/0.1)) if x2 = 1, x1 ∈ [0.9, 1],

0 otherwise.

As body force we choose f (t, x) = 0 and consider laminar flow with a Reynolds number Re = 100.

Finite element results

The initial triangulation Thinit is chosen as a criss-cross triangulation with 9 node points in both
x1 - and x2 -directions, see Figure 48 (left). For the time discretization, we utilize a uniform time
grid with time step size ∆t = 0.01 leading to n = 100 time points. In the h-adaptive finite
element Algorithm 1 we choose as stopping tolerance ε = 0.01 and as Dörfler marking parameter
θ = 0.1.

The resulting adaptive finite element velocity components and pressure solutions as well as the
adapted meshes are shown in Figure 49 for three different time instances. The maximal number
of node points in the adapted meshes is 3287. The overlay mesh (shown in Figure 48, right)
has 3729 node points. For comparison, an uniform discretization with fineness as small as the
smallest triangle in the adapted meshes would have 131585 node points, which leads to an out
of memory error with our computing machines.

Figure 48: Initial triangulation Thinit (left) and overlay of all adapted meshes (right)

POD reduced-order solutions

In order to compute a POD basis, we choose in (2.7) for the velocity POD the space X = H01 (Ω)
and for the pressure POD the space X = L2 (Ω). As time weights, we choose αj = ∆t, i.e.
we interpret the sum in (2.7) as a quadrature of the time integral in (2.5) with a right-sided
rectangle quadrature rule. This complies with the interpretation of the implicit Euler scheme as
a discontinuous Galerkin method.
For the methodologies described in Sections 4.6.3 and 4.6.4, we choose the reference velocity and
pressure spaces (Ṽ , Q̃) as the Taylor–Hood finite element pair associated with the overlay of the
adapted finite element meshes, which is shown in Figure 48 (right). In Problem 4.5, we choose
X = H01 (Ω) in order to enable a consistent setting.
The first three velocity POD modes (without divergence-free projection) as well as the pressure
POD modes and associated supremizer functions are displayed in Figure 50. Note that the POD
modes computed from weakly divergence-free projected snapshots as well as the projected POD
modes look very similar to the POD modes in Figure 50.
120 4 ADAPTIVITY WITH RESPECT TO SPACE

t6 t20 t100

mesh

y1,h

y2,h

ph

Figure 49: Adapted finite element meshes, velocity components yh = (y1,h , y2,h ) and pressure
solutions at time instances tn for n = 6, 20, 100
4.6 POD model order reduction with space-adapted snapshots for incompressible flows 121

i=1 i=2 i=3

ψ1,i

ψ2,i

φi

Tφi

Figure 50: First three POD modes i = 1, 2, 3 for the velocity ψi = (ψ1,i , ψ2,i ) and the pressure
φi , first three supremizer functions Tφi

We compare the approaches of a velocity reduced-order modeling and the velocity-pressure

reduced-order modeling based on supremizer stabilization concerning accuracy and efficiency.
In the velocity-pressure approach we choose for simplicity the same number of velocity POD
basis modes as pressure POD basis functions. A short-hand notation for the considered reduced-
order approximations is given in Table 7.
In order to validate the approximation quality of the POD reduced-order solution we consider
the following relative error for the velocity
kyh − y` kL2 (0,T ;Ω)
errF E−P OD = .
kyh kL2 (0,T ;Ω)
The results are shown in Figure 51. First of all, we observe that a velocity-pressure reduced-
order model without supremizer stabilization leads to an unstable model and unsatisfactory
approximation results. A naive approach, i.e. a velocity reduced-order model without utilizing a
divergence-free projection leads to better approximation results in this example. From `y = 1
122 4 ADAPTIVITY WITH RESPECT TO SPACE

10 1

10 0

unstable ROM
10 -1 naive ROM
div-free ROM (a)
div-free ROM (b)
stabilized ROM
10 -2

10 -3

0 5 10 15 20 25 30

Figure 51: Relative velocity errors of different reduced-order approximations according to Table 7
and different numbers of utilized velocity POD basis functions

method description
unstable ROM velocity-pressure reduced-order (4.50), but no supremizers
naive ROM velocity reduced-order model (4.51), but using Lagrange interpolations
instead of divergence-free projections of the FE solutions
div-free ROM (a) velocity reduced-order model according to approach a) in Section 4.6.3
div-free ROM (b) velocity reduced-order model according to approach b) in Section 4.6.3
stabilized ROM velocity-pressure reduced-order model according to Section 4.6.4

Table 7: Short-hand notation for the different reduced-order models

to `y = 8, the relative error between the finite element solution and the naive reduced-order
approximation decreases. However, for `y > 8, the error starts to increase, which is not what we
would like to achieve.

Now, let us compare the velocity reduced-order approaches of Section 4.6.3 with the velocity-
pressure reduced-order approach of Section 4.6.4. All of these methodologies lead to stable results
in the sense that the error between the POD reduced-order model and the finite element solution
decreases for an increasing number of utilized POD modes up to `y = 8. For `y > 8, the error
stagnates. Both velocity approaches (a) and (b) lead to very similar results. In comparison to
the velocity model, the accuracy of the velocity-pressure model is slightly better.

div-free ROM (b) stabilized ROM

FE solution 125.66 125.66
offline times
→ construction of overlay mesh 3.81 3.81
→ velocity POD 1.73 1.73
→ pressure POD – 0.05
→ div-free projection of POD modes 5.54 –
→ supremizers – 0.31
ROM solution 0.009 0.03

Table 8: Computation times (sec) for finite element solution, offline and online times using `y = 8
POD modes
4.7 POD model order reduction with space-adapted snapshots in optimal control 123

Concerning the computational efficiency, we compare the calculation times of the velocity reduced-
order approach (b) with the velocity-pressure reduced-order approach. The velocity reduced-
order approach (a) is more expensive than the approach (b), since Problem 4.5 has to be solved
for each snapshot, i.e. n times, whereas in the latter approach it only has to be solved for each
velocity POD mode, i.e. `y < n times. For comparison, we list the computational time for the
adaptive finite element solution in Table 8. It turns out that the divergence-free projection of the
POD modes is more expensive than the computation of the supremizer functions. Concerning
the reduced-order online time, the solution of the velocity model is very fast with only 0.009
seconds which leads to a speed up of factor 13962 compared to the adaptive full-order finite
element simulation. The solution of the velocity-pressure reduced order model takes 0.03 sec-
onds, so it is slower than the velocity reduced-order model. This is due to the fact that the
dimension of the velocity model depends solely on `y , whereas the complexity of the velocity-
pressure reduced-order model additionally depends on the number of supremizer functions and
the number of pressure POD basis functions. Recall that in this test run, we take for simplicity
the same number of POD modes for the pressure as for the velocity, i.e. `p = `y .
These factors motivate to utilize POD reduced-order modeling in a multi-query setting. We note
that a finite element simulation on a grid with the same fineness as the smallest triangle as in the
adapted meshes is not possible on our computing machines since it reaches the memory limit.

4.6.7 Further directions of research

In order to avoid the difficulty that the divergence constraint of the adaptive finite element solu-
tion only holds true in the current finite element discretization, an option could be to utilize finite
element methods which guarantee by construction that the discrete Navier-Stokes solutions are
divergence-free in the strong sense. In [113, 130], different approaches in order to tackle the lack
of pressure-robustness which stems from enforcing the divergence constraint only discretely are
compared. In particular, inf-sup stable conforming element pairs Vh ⊂ H01 (Ω), Qh ⊂ L20 (Ω) are
derived which satisfy ∇ · Vh ⊂ Qh . As a consequence, it can be proved that Vh,div ⊂ Hdiv . The
derivation of a POD reduced-order model for this setting as well as a comparison with respect
to efficiency and accuracy of the proposed methods of this section could be carried out in future
research.
Moreover, we aim to utilize the proposed methodology for POD reduced-order modeling of in-
compressible fluid flows in a multi-query setting like uncertainty quantification or optimal control.
There, we expect large speed up factors in the offline phase in comparison to the use of a static
uniform discretization as well as a speed up in the online phase whenever new full-order solutions
are required in order to update the reduced-order basis.

4.7 POD model order reduction with space-adapted snapshots in optimal control
In this section, we apply the techniques from the previous Sections 4.2 to 4.5 for the solution of
a multi-query scenario. We consider the optimal control of a convective Cahn–Hilliard system,
where the control enters through the velocity in the transport term. We prove the existence of
a solution to the considered optimal control problem. For the numerical solution, we replace
the full-order model by POD reduced-order surrogates which are computed from space-adapted
snapshots.

Literature overview

Let us give an overview of literature concerning optimal control of phase field models and the
reduced-order modeling thereof.
124 4 ADAPTIVITY WITH RESPECT TO SPACE

In [52] an optimal control problem of a coupled temperature phase field model with polynomial
free energy is considered. Existence of a solution to the considered optimal control problem is
proved and first-order optimality conditions are derived. In a numerical example, the shape of a
phase field is controlled. In [93] the setting of [52] is extended by introducing control constraints.
An efficient solver for the linearized state and adjoint equations is constructed which is based
on a projected Newton method. In [185] a POD reduced-order model for the problem setting
of [93] is considered. For the numerical solution of the optimal control system an augmented
Lagrangian–SQP method is used and the high-fidelity solutions on static meshes are replaced by
POD reduced-order approximations.
In [92] a phase field approach is used in order to control the shape evolution of single layer
islands on a crystalline surface. The control goal is to find an optimal electric field which steers
the initial phase shape as close as possible to a given desired shape while keeping the magnitude
of the electric field low. The control enters the Cahn–Hilliard model through (2.18b). A proof of
concept is given in a numerical test run which moves and transforms an initial shape in horizontal
direction.
An analysis of the optimal control of the Cahn–Hilliard model with distributed control is given
in [97]. Existence of a solution in case of a Moreau–Yosida relaxed double-obstacle free energy
is proved and a convergence analysis of optimal controls of the regularized problem to an opti-
mal control of the original problem with a double-obstacle free energy is carried out. Moreover,
C-stationarity conditions are derived.
In [158] the optimal control of a non-local Cahn–Hilliard model with degenerate mobility and
singular potential is considered in 3d. The control enters the system through the velocity in the
transport term. Existence of an optimal control is proved and first-order optimality conditions
are derived.
The optimal control problem of [97] with distributed control which enters the Cahn–Hilliard sys-
tem on the right hand side of (2.18a) is considered in [5] within a trust-region POD framework.
The numerical experiments are based on [85] and use spatially adapted finite element snapshots.
The optimization goal consists of deforming a circular shape into a rounded square. Different
choices for the snapshot set are considered such as a separate POD basis for the state and adjoint
states as well as a combined snapshot set of the system variables. Similar results are published
in a collaboration with Jan Oke Alff and Michael Hinze in [6].

Note that some of the results of this section are based on a collaboration with Michael Hinze
and Nicolas Scharmacher in [87].

4.7.1 Optimal control of a convective Cahn–Hilliard system

Let Ω ⊂ Rd , d ∈ {2, 3} denote an open bounded domain with Lipschitz boundary ∂Ω, see
Definition 2.17. Let T > 0 be a fixed end time. We consider the following optimal control
problem of a convective Cahn–Hilliard system: minimize the quadratic cost functional given by
β1 β2 γ
J(ϕ, u) = kϕ − ϕd k2L2 (0,T ;Ω) + kϕ − ϕT k2L2 (Ω) + kuk2U , (4.63)
2 2 2
subject to the Cahn–Hilliard system with transport

ϕt + (Bu) · ∇ϕ − b∆µ = 0 in (0, T ] × Ω, (4.64a)

σ
−σε∆ϕ + F 0 (ϕ) = µ in (0, T ] × Ω, (4.64b)
ε
and
u ∈ Uad . (4.65)
We denote by ϕ the phase field, µ is the chemical potential, F is the free energy function, b is
a constant mobility, σ is the surface tension and ε denotes the interface parameter. For more
4.7 POD model order reduction with space-adapted snapshots in optimal control 125

details on the Cahn–Hilliard system we refer to Section 2.5. In (4.63) we denote by β1 , β2 ≥ 0

given constants, γ > 0 is the penalty parameter, ϕd ∈ L2 (0, T ; Ω) is the desired phase field and
ϕT ∈ L2 (Ω) is the target phase pattern at final time.

Motivated by applications where it is only possible to control in certain prescribed areas within
the spatial domain, we consider located control which is a time-dependent quantity and matched
to the time-space domain via a control operator B. In particular, we consider the control space
U = L2 (0, T ; Rm ), with m ∈ N and introduce an admissible set of controls by

u ∈ Uad = {u ∈ U | ua (t) ≤ u(t) ≤ ub (t) in Rm a.e. in [0, T ]}

with ua , ub ∈ L∞ (0, T ; Rm ), ua (t) ≤ ub (t) almost everywhere in [0, T ]. The inequalities between
vectors are understood componentwise. We define the control operator as
m
X
B : U → L∞ (0, T ; L∞ (Ω)) ∩ L2 (0, T ; H 1 (Ω)), u 7→ Bu = ui χi ,
i=1

where χi ∈ L∞ (Ω) ∩ H 1 (Ω) for i = 1, . . . , m denote given shape functions. Note that we use
the same notation for scalar- and vector-valued functions, meaning that all components of a
vector-valued function belong to the corresponding scalar function space.

We equip (4.64) with homogeneous Neumann boundary conditions

∂ϕ ∂µ
=0 and =0 on [0, T ] × ∂Ω, (4.66)
∂~n ∂~n
and impose an initial condition for the phase field

ϕ(0, ·) = ϕ0 in Ω. (4.67)

The aim of the optimal control problem is to steer a given initial phase distribution ϕ0 to a given
desired phase pattern ϕd during the time (0, T ) and ϕT at end time T . This problem can also
be interpreted as an optimal control of a free boundary which is encoded through the phase field
variable.
We assume that the initial phase field ϕ0 is in H(0) 1 (Ω) = H 1 (Ω) ∩ L2 (Ω) with L2 (Ω) =
(0) (0)
1
{f ∈ L2 (Ω) : |Ω|
R
Ω f dx = g} with mean value g = 0 (w.l.o.g.). Moreover, we assume that
E0 = E(ϕ0 ) < ∞ where E denotes the Ginzburg–Landau free energy defined in (2.15).

Note that in (4.64) the convection models a possible flow of the phase field system at a given
velocity field Bu, for which we impose a no-flow condition on the boundary in order to preserve
mass conservation. In particular, we assume χi ∈ Hdiv (Ω) for i = 1, . . . , m with Hdiv introduced
in (4.45) as the space of solenoidal vector fields by

Hdiv (Ω) := {v ∈ L2 (Ω) : ∇ · v = 0, v · ~n|∂Ω = 0 in the sense of traces}.

Moreover, we will need the space

Hn1 (Ω) = {f ∈ H 1 (Ω) : f · ~n|∂Ω = 0 in the sense of traces}.

Note that in the context of multiphase flow, the transport term represents the coupling to the
Navier–Stokes system, see e.g. [2, 104].
126 4 ADAPTIVITY WITH RESPECT TO SPACE

The weak form of the state equation reads as follows: for a given initial phase field ϕ0 find a
phase field ϕ with ϕ(0, ·) = ϕ0 and a chemical potential µ such that
hϕt (t), viH −1 (Ω),H01 (Ω) + ((Bu(t)) · ∇ϕ(t), v)L2 (Ω) + b(∇µ(t), ∇v)L2 (Ω) = 0 ∀v ∈ H 1 (Ω),
(4.68a)
σ 0
σε(∇ϕ(t), ∇v)L2 (Ω) + (F (ϕ(t)), v)L2 (Ω) = (µ(t), v)L2 (Ω) ∀v ∈ H 1 (Ω),
ε
(4.68b)
holds for almost all t ∈ (0, T ).

It is shown in [1] for σ = b = ε = 1 that the state equation (4.64) is well-posed. This result is
obtained for a free energy with smooth convex part F+ ∈ C([−1, 1]) ∩ C 2 ((−1, 1)) and parts of
the theorem are stated in the following, where we adopted the notation to our setting. Moreover,
higher regularity solutions are provided in [1, Lemma 4.1.3]. Let us denote f ∈ L2loc (0, ∞; X)
with X being a Banach space, iff f ∈ L2 (0, T ; X) for every T > 0. The uniformly local variant
is denoted by L2uloc ([0, ∞); X) and comprises all strongly measurable functions f : [0, ∞) → X
with kf kL2 (0,∞;X) = supkf kL2 (t,t+1;X) < ∞.
uloc
t≥0

Theorem 4.10. (Well-posedness of the state equation, [1, Theorem 4.1.1]). Let Bu ∈ L2 (0, ∞; Hn1 (Ω))∩
L∞ (0, ∞; Hdiv (Ω)). Then for every ϕ0 ∈ H(0)
1 (Ω) with E < ∞ there exists a unique solution
0
ϕ ∈ BC([0, ∞); H(0) 1 (Ω)) of (4.64) with ϕ ∈ L2 (0, ∞; H −1 (Ω)), µ ∈ L2 ([0, ∞); H 1 (Ω)). This
t (0) uloc
solution satisfies
 
kϕk2L∞ (0,∞;H 1 (Ω)) + kϕt k2L2 (0,∞;H −1 (Ω)) + k∇µk2L2 (0,∞;Ω) ≤ C E0 + kBuk2L2 (0,∞;Ω) , (4.69)
(0)

and  
kϕk2L2 ([0,∞);W 2,p (Ω)) + kF 0 (ϕ)k2L2 ([0,∞);Lp (Ω)) ≤ Cp E0 + kBuk2L2 (0,∞;Ω) (4.70)
uloc uloc

where p = 6 if d = 3 and p ∈ (1, ∞) if d = 2. Here C, Cp are independent of Bu and ϕ0 .

Moreover, for every R, T > 0 the solution
−1
ϕ ∈ Y := L2 (0, T ; W 2,p (Ω)) ∩ H 1 (0, T ; H(0) (Ω))
depends continuously on
(ϕ0 , Bu) ∈ X := H 1 (Ω) × L1 (0, T ; Hdiv (Ω)) such that E0 + kBukL2 (0,∞;H 1 (Ω)) ≤ R
with respect to the weak topology on Y and the strong topology on X.
Due to Theorem 4.10, the optimal control problem (4.63)-(4.64)-(4.65) can be expressed as a
reduced optimal control problem given by
ˆ
min J(u) := J(ϕ(u), u) s.t. (ϕ(u), u) satisfies (4.64) with (4.66), (4.67). (4.71)
u∈Uad

Theorem 4.11. (Existence of an optimal control). Problem (4.71) admits a solution ū ∈ Uad .
Proof. Step 1: Boundedness of the state: Due to the energy estimates (4.69) and (4.70) it holds
 
kϕk2L2 (0,T ;H 2 (Ω)) + kϕt k2L2 (0,T ;H −1 (Ω)) ≤ c E0 + kBuk2L2 (0,T ;Ω) ,
(0)

such that we can find an upper bound M > 0, since B is linear and bounded.
Step 2: Find a candidate for an optimal control: Since the cost functional is bounded from below
(Jˆ ≥ 0) and Uad 6= ∅, the infimum
J ∗ := inf J(u)
ˆ = inf J(ϕ(u), u)
u∈Uad u∈Uad
4.7 POD model order reduction with space-adapted snapshots in optimal control 127

exists. Let {un }n∈N ⊂ Uad be a minimizing sequence, i.e.

lim J(ϕ(un ), un ) = J ∗ .
n→∞

Since Uad is closed, convex and bounded in L2 (0, T ; Rm ) ⊃ L∞ (0, T ; Rm ), it is weakly sequen-
tially compact according to Theorem 2.38. As a result, we can extract a subsequence (denoted
by the same name), which converges weakly in L∞ (0, T ; Rm ) to some ū ∈ Uad . Weak conver-
gence Bun * Bū in L2 (0, T ; H 1 (Ω)) follows from the linearity and boundedness of the operator B.

Step 3: Define associated sequence of states: Let {ϕn }n∈N denote the corresponding sequence of
states, i.e. ϕn := ϕ(un ) and let {µn }n∈N denote the corresponding sequence of chemical poten-
1 (Ω)) ∩ L2 (0, T ; H 2 (Ω)). Since
tials. From Theorem 4.10 it follows that ϕn ∈ Ỹ := W (0, T ; H(0)
Ỹ is reflexive, there exists a subsequence (denoted by the same name) that converges weakly
to some ϕ̄ ∈ Ỹ according to Theorem 2.38. Analogously, we obtain weak convergence for the
chemical potential to some µ̄ ∈ L2 (0, T ; H 1 (Ω)).

Step 4: Show (ϕ̄, ū) is admissible: It remains to show that the pair (ϕ̄, ū) is admissible, i.e.
ϕ̄ = ϕ(ū). We consider
hϕn,t (t), viH −1 (Ω),H01 (Ω) + ((Bun (t)) · ∇ϕn (t), v)L2 (Ω) + b(∇µn (t), ∇v)L2 (Ω) = 0, (4.72a)
σ
σε(∇ϕn (t), ∇v)L2 (Ω) + (F 0 (ϕn (t)), v)L2 (Ω) = (µn (t), v)L2 (Ω) ,
ε
(4.72b)

for all v ∈ H 1 (Ω) and for almost all t ∈ [0, T ]. While passing to the limit n → ∞ in the
weak formulation is clear for the linear terms, the nonlinear ones require further investigation.
Since W (0, T ) ∩ L2 (0, T ; H 2 (Ω)) is compactly embedded in L2 (0, T ; H 1 (Ω)) (see [164, Section 8,
Corollary 4]), the sequence {ϕn }n∈N converges strongly to ϕ̄ in L2 (0, T ; H 1 (Ω)). For the control
term we have for v ∈ H 1 (Ω) the splitting
Z T
(Bun (t) · ∇ϕn (t) − Bū(t) · ∇ϕ̄(t), v)L2 (Ω) dt =
0
Z T Z T
(Bun (t) · ∇(ϕn (t) − ϕ̄(t)), v)L2 (Ω) dt + (B(un (t) − ū(t)) · ∇ϕ̄(t), v)L2 (Ω) dt.
0 0

Due to ϕ̄ ∈ L2 (0, T ; H 2 (Ω))

it holds ∇ϕ̄ ∈ L2 (0, T ; H 1 (Ω))
and the product v·∇ϕ̄ ∈ L2 (0, T ; L2 (Ω))
gives rise to a continuous linear functional on L2 (0, T ; H 1 (Ω)). Hence, the right term vanishes
for n → ∞ by definition of weak convergence. For the left term we estimate
Z T
(Bun (t) · ∇(ϕn (t) − ϕ̄(t)), v)L2 (Ω) dt ≤ kBun kL2 (0,T ;H 1 (Ω)) kϕn − ϕ̄kL2 (0,T ;H 1 (Ω)) kvkH 1 (Ω) ,
0
0 we use
which vanishes for n → ∞ due to the convergence of {ϕn }n∈N . For the nonlinearity Fpol
the inequality from [101, Theorem 1.45]
|b3 − a3 | ≤ 3(|a|2 + |b|2 )|b − a|
for a, b ∈ R which leads to
0 0
|Fpol (ϕ) − Fpol (ψ)| = |ϕ3 − ϕ − ψ 3 + ψ| ≤ 3(|ϕ|2 + |ψ|2 ) |ϕ − ψ| + |ϕ − ψ|
for all ϕ, ψ ∈ R. As a consequence, we obtain the estimate
Z T
0 0
(Fpol (ϕn (t)) − Fpol (ϕ̄(t)), v)L2 (Ω) dt ≤
0
3(kϕ2n kL2 (0,T ;L2 (Ω)) + kϕ̄2 kL2 (0,T ;L2 (Ω)) + 1)kϕn − ϕ̄kL2 (0,T ;L2 (Ω)) kvkH 1 (Ω) ,
128 4 ADAPTIVITY WITH RESPECT TO SPACE

which gives the desired convergence due to the Sobolev imbedding L∞ (0, T ; H 1 (Ω)) ⊂ L4 (0, T ; L4 (Ω)).

Step 5: Optimality of ū: Since Jˆ is a continuous, convex functional we obtain from Theorem 2.39
that Jˆ is weakly lower semicontinuous, i.e.

un * ū ⇒ ˆ n ) ≥ J(ū).
lim inf J(u ˆ
n→∞

ˆ
Therefore, it follows J(ū) ˆ
= inf u∈Uad J(u) and ū is an optimal control.

The optimal control problem (4.71) is a non-convex programming problem so that different
minima might exist. Numerical solution methods will converge in a local minimum which is
close to the initial point. In order to compute a locally optimal solution to (4.71), we consider
the first-order necessary optimality condition given by the variational inequality

hJˆ0 (ū), u − ūiU ∗ ,U ≥ 0 ∀u ∈ Uad . (4.73)

Following the standard Lagrange technique from Section 2.3, we derive that (4.73) is equivalent
to Z TX m  Z 
γ ūi (t) + (χi (x) · ∇ϕ(t, x))p̄(t, x)dx (ui (t) − ūi (t))dt ≥ 0 (4.74)
0 i=1 Ω

for all u ∈ Uad where the function p̄ is a solution to the adjoint equations
σ 00
−pt − σε∆q + F (ϕ̄)q − Bu · ∇p = − β1 (ϕ̄ − ϕd ) in (0, T ] × Ω, (4.75a)
ε
−q − b∆p = 0 in (0, T ] × Ω, (4.75b)
∂n p = ∂n q = 0 on [0, T ] × ∂Ω, (4.75c)
p(T, ·) = − β2 (ϕ̄(T, ·) − ϕT ) in Ω. (4.75d)

The variable ϕ̄ in (4.75) denotes the solution to state equation (4.64) associated with an optimal
control ū. The weak form of the adjoint system is given as follows: find the adjoint states p and
q such that p(T, ·) = −β2 (ϕ̄(T, ·) − ϕT ) and
σ 00
−hpt (t), viH −1 (Ω),H01 (Ω) + (∇q(t), ∇v)L2 (Ω) + (F (ϕ̄(t))q(t), v)L2 (Ω) (4.76a)
ε
−(Bu(t) · ∇p(t), v)L2 (Ω) = (ϕ̄(t) − ϕd (t), v)L2 (Ω) ,
−(q(t), v)L2 (Ω) + b(∇p, ∇v)L2 (Ω) = (µ(t), v)L2 (Ω) , (4.76b)

holds for all v ∈ H 1 (Ω) and almost all t ∈ (0, T ).

Discretization

For the discretization, we first discretize the state equation in time using the convex-concave
splitting in (2.26) leading to an unconditional energy stable scheme. For the spatial discretiza-
tion of the state equation, we use h-adaptive finite elements according to Algorithm 1 where
the estimator is based on the gradient jump, compare (4.39). In order to make computations
efficient, we only refine once at each time instance and introduce an additional coarsening strat-
egy according to [114, Section 8.3]. For the discretization of the adjoint state, we follow the
first-discretize-then-optimize concept (compare Remark 2.44). In particular, we note that for
simplicity we use the same adapted meshes for the state as for the adjoint variable in order to
circumvent the handling of a two-mesh system. In our numerical example we observe that the
state and adjoint variable are related such that a refinement for the state is also suitable for
the adjoint state. It could also be of interest to investigate further adaptivity concepts for the
adjoint state, especially if the solution properties for the state and adjoint state differ strongly.
4.7 POD model order reduction with space-adapted snapshots in optimal control 129

4.7.2 Numerical results

Note that this example is published in [87].

Problem setting

The choices for the problem data and parameters are listed in Table 9.

name value
Ω spatial domain (0, 1) × (0, 1) ⊂ R2
T end time 0.0125
∆t time step size 2.5 · 10−05
b constant mobility 2 · 10−05
ε interface parameter 0.02
σ surface tension 25.98
F nonlinear free energy Fpol (2.22)
ϕ0 initial condition circle with radius 0.35 and center (0.5, 0.5)
γ regularization parameter 0.0001
m = 1, χ1 control shape χ1 (x) = (sin(πx1 )cos(πx2 ), −sin(πx2 )cos(πx1 ))T

Table 9: Problem data and parameters of the optimal control problem (4.71)

In order to fulfill the Courant–Friedrichs–Lewy (CFL) type condition, we impose the control
constraints ua = 0, ub = 50 and demand hmin > 0.00177. In the cost functional we use γ = 0.0001,
β1 = 20 and β2 = 20. The desired state is shown in Figure 52. The initial state ϕ0 coincides
with ϕd (0).
The optimization is initialized with an input control u = 0 ∈ Uad . We compute the POD basis
with respect to the L2 (Ω)-inner product for the snapshot ensemble formed by the desired phase
field ϕd , which is discretized using adaptive finite elements. For the solution of the optimal control
problem, we use a projected gradient method. Figure 53 shows the finite element solution and
the POD-DEIM solution for the phase field using ` = 10 and ` = 20 POD modes, respectively.
It turns out that a large number of POD modes is needed in order to smoothen out oscillations
due to the convection.

Figure 52: Desired phase field at t0 , t250 , t500 with adaptive meshes

Table 10 (left) summarizes the iteration history for the finite element projected gradient method
where we used the stopping criterion kJˆ0 (uk )k < 0.01 · kJˆ0 (u0 )k + 0.01. Table 10 (right) tab-
ulates the POD-DEIM optimization. Note that the value of the POD cost functional Jˆ` (uk )
stagnates due to the POD error. The value of the full-order cost functional at the POD solution
ˆ POD ) = 7.31 · 10−04 . If ` = 20 POD modes are used, the relative L2 (0, T ; Ω)-error between
is J(ū
the finite element and the POD-DEIM optimal solution is for the phase field errϕ = 7.38 · 10−03 .
In Table 11 the computational times for the uniform finite element, adaptive finite element, POD
and POD-DEIM optimization are listed. The offline costs for POD when using spatially adapted
snapshots are as follows: the interpolation of the snapshots takes 212 seconds, the POD basis
130 4 ADAPTIVITY WITH RESPECT TO SPACE

Figure 53: Finite element (top) and POD-DEIM reduced-order solution with ` = 10 (middle)
and ` = 20 (bottom) of the phase field ϕ at t = t0 , t250 , t500 with adaptive meshes

k ˆ k)
J(u kJˆ0 (uk )k sk k Jˆ` (uk ) kJˆ`0 (uk )kUh sk
0 8.61 · 100 2.85 · 100 1.0 0 8.77 · 100 2.81 · 100 1.0
1 6.48 · 10−01 2.32 · 100 0.25 1 7.98 · 10−01 2.41 · 100 0.25
2 1.90 · 10−02 4.56 · 10−01 0.25 2 5.79 · 10−02 3.67 · 10−01 0.25
3 3.82 · 10−03 1.93 · 10−01 0.25 3 5.02 · 10−02 1.63 · 10−01 0.25
4 1.18 · 10−03 8.45 · 10−02 0.25 4 4.76 · 10−02 7.45 · 10−02 0.25
5 6.80 · 10−04 3.67 · 10−02 5 4.76 · 10−02 3.48 · 10−02

Table 10: Iteration history of the finite element optimization (left) and the POD-DEIM opti-
mization (right) with ` = 20. The Armijo step size is denoted by sk

uniform FE adaptive FE POD POD-DEIM

optimization 36868 5805 675 0.3
→ solve each state eq. 1660 348 42 0.02
→ solve each adjoint eq. 761 121 16 0.01

Table 11: Computational times (sec) for the finite element, POD and POD-DEIM optimization

computation costs 40 seconds and the computations of a DEIM basis takes 30 seconds. In com-
parison, the use of uniformly discretized snapshots leads to the computational time of 243 seconds
for POD basis computation and 193 seconds for the DEIM computations.
The numerical tests show that the use of POD-DEIM reduced-order modeling combined with
spatially adapted snapshots leads to large speed up factors compared with a high-fidelity finite
element optimization.
 131

5 Summary, conclusions and outlook

This thesis is concerned with offline adaptivity strategies in model order reduction using proper
orthogonal decomposition (POD) for time-dependent systems and optimal control problems. In
order to construct suitable snapshots which identify key properties of the solution data, we pro-
pose two major strategies.

In the first part, we consider a snapshot location strategy for model order reduction in optimal
control. The selection of suitable time instances for snapshot generation is based on a reformu-
lation of the optimality system into a biharmonic equation and uses residual based a-posteriori
error estimates in order to compute a time-adaptive grid. We propose adaptive strategies which
are tailored for the state and adjoint state / control and ensure efficiency by a heuristic observa-
tion in the numerical experiments in which spatial and temporal discretization decouple. This
allows to use a coarse spatial grid in the offline phase. At the same time, we compute an ap-
proximation of the optimal control which can be used as input control for snapshot generation.
We further address the case of located control and control constraints. An error analysis for
the error between the true (unknown) optimal solution and the POD reduced-order solution is
carried out. It reveals that a time-adaptive grid which is suitable for the state might not be a
suitable time grid for the adjoint state / control and vice versa. In numerical experiments we
consider time-dependent systems which develop steep transitions within the temporal domain.
We demonstrate that the use of a time-adaptive grid leads to a good approximation quality of
the resulting reduced-order model with only a little number of time instances and is thus effi-
cient. Further, we observe that the adaptive snapshot grid is also a suitable time grid for the
reduced-order optimization, since we already detect the major dynamical structures related to
the optimal solution. Hence, the proposed methodology is not only an offline adaptive strategy,
but also provides benefits for the online solution.
As an outlook, we sketch ideas for future research questions in Section 3.9 comprising the case
of a nonlinear state equation and the application of the concept to a model predictive control
problem. Moreover, different variations of the problem setting like the inclusion of a transport
term or the consideration of state constraints could be addressed. Regarding the numerical ex-
periments it would be of interest to carry out computations for a 2d or 3d setting with complex
dynamical structures in time, where we expect the proposed snapshot location strategy and time
discretization to be advantageous.

In the second part, we include spatial adaptivity in the offline phase for snapshot generation in
model order reduction with POD. The goal is to identify dominant spatial structures by means
of adaptive finite element methods. This enables a fast snapshot sampling in comparison to
the use of a very fine uniform discretization while a prescribed accuracy of the finite element
solution is ensured by the use of residual based error indicators. The major challenge in the
context of POD model order reduction is that space-adapted snapshots have different numbers
and locations of discretization points which means in a fully discrete setting that the snapshots
are vectors of different lengths. In order to tackle this issue, we consider the POD method from
an infinite-dimensional perspective and compute an eigenvalue problem for the snapshot gramian
which is set up explicitly. It is an advantage of this approach that it allows snapshot discretiza-
tions stemming from r-, p- and h-adaptive schemes. In case of space-adapted snapshots with
non-nested meshes, the computation of the snapshot gramian involves the integration over cut
finite elements. For the numerical implementation, we propose a strategy which is based on an
integration over the boundary of the cut polyhedra utilizing Stoke’s formula. This is numerically
intensive, however a speed up can be achieved using parallelization. An error analysis for the
error between the true (unknown) solution and the POD reduced-order solution using space-
adapted snapshots is carried out and reveals a further dependency on the spatial discretization
error in comparison to existing error estimates. In numerical examples comprising a linear heat
132 5 SUMMARY, CONCLUSIONS AND OUTLOOK

equation and a nonlinear Cahn–Hilliard system, the combination of POD with space-adapted
snapshots in case of hierarchical, nested meshes and in case of non-nested meshes is presented.
We obtain large speed up factors in the offline times for space-adapted finite element solutions
in comparison with a very fine uniform discretization. Moreover, the POD reduced-order on-
line simulation time is very fast in comparison to the full-order model. At the same time we
can achieve a good accuracy. In the context of incompressible flows governed by the unsteady
Navier–Stokes equations, the challenge arises that due to the use of h-adapted finite elements
a weakly divergence-free property only holds in the current finite element space (and not in a
common sense). Hence, the POD modes do not automatically inherit a divergence-free prop-
erty like in the static case and the reduced-order model is not stable. We propose two solution
concepts in order to derive a stable reduced-order model. The first approach is based on a
velocity reduced-order modeling and utilizes a projection of the snapshots or the POD modes
such that the projected data is weakly divergence-free in a common sense. It is an advantage
of this approach that the reduced-order model is cheap to solve since it only depends on the
velocity. The second approach is a velocity-pressure reduced-order model, where the stability is
guaranteed by enriching the reduced velocity space with pressure supremizer functions. In this
way, it is shown that the inf-sup stability of the finite element model carries over to an inf-sup
stability of the reduced-order model. The advantage of this approach is that the model delivers
not only a velocity but also a pressure approximation, which is of interest in many applications.
The numerical experiment reveals that both approaches deliver equally accurate reduced-order
solutions.
Finally, we utilize h-adaptive finite element snapshots for the solution of an optimal control
problem subject to a Cahn–Hilliard system, where the control enters the system in the transport
term. In order to adapt the reduced-order model accuracy within the course of the optimiza-
tion, in [5, 6] a trust-region POD framework with space-adapted snapshots is used for a similar
optimal control problem. It is of future research interest to apply the developed strategies to
the simulation and optimal control of a coupled Cahn–Hilliard Navier–Stokes system describing
two-phase flow.
 133

6 Appendix
6.1 Snapshot location: uniform time mesh misses a signal
We construct an example in the following way. The optimal state is chosen as

y(t, x) = sin(πx) · g(t),

where g is the probability density function of the normal distribution given as

(t − 0.3)2
 
1
g(t) = √ exp −
2πε 2ε

with variance ε = 10−04 . The true optimal solution, its contour lines and the density function g
are shown in Section 6.1.
1 40

0.75 30

0.5 20

0.25 10

0 0
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

Figure 54: Optimal state (left), its contour lines (middle) and the density function g (right)

We choose the source function f (t, x) = 0 and construct the optimal control as

u(t, x) = sin(πx)(gt (t) + π 2 g(t)).

The corresponding desired state is given by

yd (t, x) = sin(πx)((1 + π 4 )g(t) − gtt (t)).

This academic example is constructed such that it shows a scenario in which a uniform dis-
cretization misses a dynamical structure. Thus, we choose a uniform time grid with n = 5,
i.e. ∆t = 1/n = 0.2. With this choice, it holds true y(tj , x) < 10−15 , p(tj , x) < 10−15 for
j = 0, . . . , n + 1 and all x ∈ [0, 1]. Therefore, both state and adjoint snapshot sets will only con-
tain zero vectors, i.e. a POD computes an arbitrary orthonormal basis and the resulting POD
state and adjoint solution will be zero. This reflects the snapshots correctly, however it does
not reflect the true solution at all. In particular, it misses the time-dependent signal at t = 0.3
determined by the normal distribution g. In contrary, the time-adaptive algorithm is able to
detect this signal with the same number of time instances. The space-time grid is shown in
Section 6.1. The resulting POD state solution is shown in the same figure (middle, right). We
observe that the time peak at 0.3 is detected. In comparison to the true solution, the scaling is
still not good, but this is due to the temporal resolution and can be improved by enlarging the
number of time instances. We see that the dynamic in the spatial domain which is determined
by the sin-function is well captured. This is in contrast to the uniform time discretization, in
which an arbitrary (POD) orthonormal basis will be chosen.
134 REFERENCES

1 1

2
0.75

0.5 1 0.5

0.25 0
1
1
0.5
0 0.5 0
0 0.25 0.5 0.75 1 0 0 0.5 1
0

Figure 55: Space-time grid with n = 5 and h+ = 1/5 (left), POD optimal state (middle) and its
contour lines (right)

References
[1] H. Abels. Diffuse interface models for two-phase flows of viscous incompressible fluids.
Lecture Notes, Max Planck Institute for Mathematics in the Sciences, (36), 2007.

[2] H. Abels, H. Garcke, and G. Grün. Thermodynamically consistent, frame indifferent diffuse
interface models for incompressible two-phase flows with different densities. Mathematical
Models and Methods in Applied Sciences, 22(3), 2012.

[3] K. Afanasiev and M. Hinze. Adaptive control of a wake flow using proper orthogonal
decomposition. Lecture Notes in Pure and Applied Mathematics, pages 317–332, 2001.

[4] M. Ainsworth and J. T. Oden. A posteriori error estimation in finite element analysis,
volume 37. John Wiley & Sons, 2011.

[5] J. O. Alff. Trust Region POD for Optimal Control of Cahn-Hilliard Systems. Master’s
thesis, Universität Hamburg, 2018.

[6] J. O. Alff, C. Gräßle, and M. Hinze. Adaptive trust-region POD for optimal control of the
Cahn-Hilliard equation. PAMM, 18(1):e201800453, 2018.

[7] M. Ali, K. Steih, and K. Urban. Reduced basis methods with adaptive snapshot compu-
tations. Advances in Computational Mathematics, 43(2):257–294, 2017.

[8] A. Alla and M. Falcone. A time-adaptive POD method for optimal control problems. IFAC
Proceedings Volumes, 46(26):245–250, 2013.

[9] A. Alla, C. Gräßle, and M. Hinze. A residual based snapshot location strategy for POD in
distributed optimal control of linear parabolic equations. IFAC-PapersOnLine, 49(8):13–
18, 2016.

[10] A. Alla, C. Gräßle, and M. Hinze. Snapshot location for POD in control of a linear heat
equation. PAMM, 16(1):781–782, 2016.

[11] A. Alla, C. Gräßle, and M. Hinze. A-posteriori snapshot location for POD in optimal
control of linear parabolic equations. ESAIM: M2AN, 52(5):1847–1873, 2018.

[12] A. Alla and J. N. Kutz. Nonlinear model order reduction via dynamic mode decomposition.
SIAM Journal on Scientific Computing, 39(5):B778–B796, 2017.

[13] M. S. 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):9–23, 2015.
REFERENCES 135

[14] R. Altmann and J. Heiland. Regularization and Rothe discretization of semi-explicit op-
erator DAEs. International Journal of Numerical Analysis and Modeling, 15(3):452–478,
2018.

[15] C. A. Andrews, J. M. Davies, and G. R. Schwarz. Adaptive data compression. Proceedings

of the IEEE, 55(3):267–277, 1967.

[16] A. C. Antoulas. Approximation of large-scale dynamical systems, volume 6. Siam, 2005.

[17] E. Arian, M. Fahl, and E. W. Sachs. Trust-region proper orthogonal decomposition for
flow control. Technical report, 2000.

[18] P. Astrid, S. Weiland, K. Willcox, and T. Backx. Missing point estimation in models
described by proper orthogonal decomposition. IEEE Transactions on Automatic Control,
53(10):2237–2251, 2008.

[19] I. Babuška. Error-bounds for finite element method. Numerische Mathematik, 16(4):322–
333, 1971.

[20] I. Babuska, B. A. Szabo, and I. N. Katz. The p-version of the finite element method. SIAM
Journal on Numerical Analysis, 18(3):515–545, 1981.

[21] A. Bailey, W. C. K. Poon, R. J. Christianson, A. B. Schofield, U. Gasser, V. Prasad,

S. Manley, P. N. Segre, L. Cipelletti, W. V. Meyer, et al. Spinodal decomposition in
a model colloid-polymer mixture in microgravity. Physical review letters, 99(20):205701,
2007.

[22] F. Ballarin, A. Manzoni, A. Quarteroni, and G. Rozza. Supremizer stabilization of POD–

Galerkin approximation of parametrized steady incompressible Navier–Stokes equations.
International Journal for Numerical Methods in Engineering, 102(5):1136–1161, 2015.

[23] L. Baňas, A. Novick-Cohen, and R. Nürnberg. The degenerate and non-degenerate deep
quench obstacle problem: a numerical comparison. Networks and Heterogeneous Media,
8:37–64, 2013.

[24] L. Baňas and R. Nürnberg. Numerical approximation of a non-smooth phase-field model

for multicomponent incompressible flow. ESAIM: Mathematical Modelling and Numerical
Analysis, 51(3):1089–1117, 2017.

[25] M. Barrault, Y. Maday, N. C. Nguyen, and A. T. Patera. An ‘empirical interpolation’

method: application to efficient reduced-basis discretization of partial differential equa-
tions. Comptes Rendus Mathematique, 339(9):667–672, 2004.

[26] M. Baumann, P. Benner, and J. Heiland. Space-Time Galerkin POD with Application in
Optimal Control of Semilinear Partial Differential Equations. SIAM Journal on Scientific
Computing, 40(3):A1611–A1641, 2018.

[27] C. A. Beattie and S. Gugercin. Model reduction by rational interpolation. Model Reduction
and Algorithms: Theory and Applications, P. Benner, A. Cohen, M. Ohlberger, and K.
Willcox, eds., Computational Science & Engineering, 15:297–334, 2014.

[28] P. Benner and R. Rannacher. Introduction to part III, Adaptivity and Model Reduction, in
Trends in PDE constrained optimization. International Series of Numerical Mathematics,
165, 2014.

[29] G. Berkooz, P. Holmes, and J. L. Lumley. The proper orthogonal decomposition in the
analysis of turbulent flows. Annual review of fluid mechanics, 25(1):539–575, 1993.
136 REFERENCES

[30] C. Bernardi and E. Süli. Time and space adaptivity for the second-order wave equation.
Mathematical Models and Methods in Applied Sciences, 15(02):199–225, 2005.

[31] A. Bertram and R. Zimmermann. Theoretical investigations of the new cokriging method
for variable-fidelity surrogate modeling: Well-posedness and maximum likelihood training.
Advances in Computational Mathematics, 44(6):1693–1716, 2018.

[32] M. Besier and W. Wollner. On the pressure approximation in nonstationary incompress-

ible flow simulations on dynamically varying spatial meshes. International Journal for
Numerical Methods in Fluids, 69(6):1045–1064, 2012.

[33] L. Blank, M. Butz, and H. Garcke. Solving the Cahn-Hilliard variational inequality with a
semi-smooth Newton method. ESAIM: Control, Optimisation and Calculus of Variations,
17(4):931–954, 2011.

[34] L. Blank, H. Garcke, M. Hassan Farshbaf-Shaker, and V. Styles. Relating phase field
and sharp interface approaches to structural topology optimization. ESAIM: Control,
Optimisation and Calculus of Variations, 20(4):1025–1058, 2014.

[35] J. F. Blowey and C. M. Elliott. The Cahn–Hilliard gradient theory for phase separation
with non-smooth free energy Part I: Mathematical analysis. European Journal of Applied
Mathematics, 2(3):233–280, 1991.

[36] J. F. Blowey and C. M. Elliott. The Cahn–Hilliard gradient theory for phase separation
with non-smooth free energy Part II: Numerical analysis. European Journal of Applied
Mathematics, 3(2):147–179, 1992.

[37] D. Braess. Finite elements: Theory, fast solvers, and applications in solid mechanics.
Cambridge University Press, 2007.

[38] M. Brand. Incremental singular value decomposition of uncertain data with missing values.
In European Conference on Computer Vision, pages 707–720. Springer, 2002.

[39] A. Brandt. Multi-level adaptive solutions to boundary-value problems. Mathematics of

computation, 31(138):333–390, 1977.

[40] F. Brezzi. On the existence, uniqueness and approximation of saddle-point problems aris-
ing from Lagrangian multipliers. Publications mathématiques et informatique de Rennes,
(S4):1–26, 1974.

[41] J. Burkardt, M. Gunzburger, and H.-C. Lee. POD and CVT-based reduced-order modeling
of Navier–Stokes flows. Computer Methods in Applied Mechanics and Engineering, 196(1-
3):337–355, 2006.

[42] G. Büttner. Ein Mehrgitterverfahren zur optimalen Steuerung parabolischer Probleme.

Dissertation, Technische Universität Berlin, 2004.

[43] J. W. Cahn and J. E. Hilliard. Free energy of a nonuniform system. I. Interfacial free
energy. The Journal of chemical physics, 28(2):258–267, 1958.

[44] I. Capuzzo Dolcetta, S. Finzi Vita, and R. March. Area-preserving curve-shortening flows:
from phase separation to image processing. Interfaces and Free Boundaries, 4(4):325–343,
2002.

[45] K. Carlberg. Adaptive h-refinement for reduced-order models. International Journal for
Numerical Methods in Engineering, 102(5):1192–1210, 2015.
REFERENCES 137

[46] J. Carr, M. E. Gurtin, and M. Slemrod. Structured phase transitions on a finite interval.
Archive for rational mechanics and analysis, 86(4):317–351, 1984.

[47] C. Carstensen and R. Verfürth. Edge residuals dominate a posteriori error estimates for
low order finite element methods. SIAM Journal on Numerical Analysis, 36(5):1571–1587,
1999.

[48] E. Casas. Pontryagin’s principle for state-constrained boundary control problems of semi-
linear parabolic equations. SIAM Journal on Control and Optimization, 35(4):1297–1327,
1997.

[49] J. M. Cascon, C. Kreuzer, R. H. Nochetto, and K. G. Siebert. Quasi-optimal conver-

gence rate for an adaptive finite element method. SIAM Journal on Numerical Analysis,
46(5):2524–2550, 2008.

[50] S. Chaturantabut and D. C. Sorensen. Nonlinear model reduction via discrete empirical
interpolation. SIAM Journal on Scientific Computing, 32(5):2737–2764, 2010.

[51] Y. Chen. Model order reduction for nonlinear systems. Master’s thesis, Massachusetts
Institute of Technology, 1999.

[52] Z. Chen and K.-H. Hoffmann. Numerical solutions of the optimal control problem governed
by a phase field model. In Estimation and Control of Distributed Parameter Systems, pages
79–97. Springer, 1991.

[53] P. G. Ciarlet and J.-L. Lions. Handbook of numerical analysis. Vol. 1. Amsterdam and
New York, North-Holland, 1990, 658, 1990.

[54] P. G. Ciarlet and J.-L. Lions. Handbook of numerical analysis. Vol. 2, Finite Element
Methods (part 1). Amsterdam and New York, North-Holland, 1990, 658, 1991.

[55] P. G. Ciarlet and P.-A. Raviart. A mixed finite element method for the biharmonic equa-
tion. Mathematical aspects of finite elements in partial differential equations, pages 125–145,
1974.

[56] P. Constantin and C. Foias. Navier-Stokes equations. The University of Chicago Press,
1988.

[57] N. Cressie. The origins of kriging. Mathematical geology, 22(3):239–252, 1990.

[58] R. Dautray and J.-L. Lions. Mathematical Analysis and Numerical Methods for Science
and Technology, Volume 5 Evolution Problems I. Springer, 1992.

[59] J. C. De los Reyes. Numerical PDE-constrained optimization. Springer, 2015.

[60] W. Dörfler. A convergent adaptive algorithm for Poisson’s equation. SIAM Journal on
Numerical Analysis, 33(3):1106–1124, 1996.

[61] Z. Drmac and S. Gugercin. A new selection operator for the discrete empirical interpolation
method - Improved a priori error bound and extensions. SIAM Journal on Scientific
Computing, 38(2):A631–A648, 2016.

[62] J. Du, F. Fang, C. C. Pain, I. Navon, J. Zhu, and D. A. Ham. POD reduced-order
unstructured mesh modeling applied to 2D and 3D fluid flow. Computers & Mathematics
with Applications, 65(3):362–379, 2013.

[63] G. Dziuk. Theorie und Numerik Partieller Differentialgleichungen. Walter de Gruyter,

2010.
138 REFERENCES

[64] C. M. Elliott. The Cahn-Hilliard model for the kinetics of phase separation. In Mathematical
models for phase change problems, pages 35–73. Springer, 1989.

[65] C. M. Elliott, D. A. French, and F. Milner. A second order splitting method for the
Cahn-Hilliard equation. Numerische Mathematik, 54(5):575–590, 1989.

[66] C. M. Elliott and H. Garcke. On the Cahn–Hilliard equation with degenerate mobility.
Siam Journal on Mathematical Analysis, 27(2):404–423, 1996.

[67] C. M. Elliott and Z. Songmu. On the Cahn-Hilliard equation. Archive for Rational Me-
chanics and Analysis, 96(4):339–357, 1986.

[68] C. M. Elliott and A. Stuart. The global dynamics of discrete semilinear parabolic equations.
SIAM Journal on Numerical Analysis, 30(6):1622–1663, 1993.

[69] L. C. Evans. Partial differential equations. American Mathematical Society, Providence,

R.I., 2010.

[70] D. J. Eyre. Unconditionally gradient stable time marching the Cahn-Hilliard equation.
MRS Online Proceedings Library Archive, 529, 1998.

[71] D. J. Eyre. An unconditionally stable one-step scheme for gradient systems. Technical
report, pages 1–15, 1998.

[72] F. Fang, C. Pain, I. Navon, M. Piggott, G. Gorman, P. Allison, and A. Goddard. Reduced-
order modelling of an adaptive mesh ocean model. International journal for numerical
methods in fluids, 59(8):827–851, 2009.

[73] H. Fareed, J. R. Singler, Y. Zhang, and J. Shen. Incremental proper orthogonal decompo-
sition for PDE simulation data. Computers & Mathematics with Applications, 75(6):1942–
1960, 2018.

[74] L. Feng, X. Zeng, C. Chiang, D. Zhou, and Q. Fang. Direct nonlinear order reduction with
variational analysis. In Design, Automation and Test in Europe Conference and Exhibition,
Proceedings, volume 2, pages 1316–1321. IEEE, 2004.

[75] T. Franz, R. Zimmermann, S. Görtz, and N. Karcher. Interpolation-based reduced-order

modelling for steady transonic flows via manifold learning. International Journal of Com-
putational Fluid Dynamics, 28(3-4):106–121, 2014.

[76] J. d. Frutos, V. John, and J. Novo. Projection methods for incompressible flow problems
with WENO finite difference schemes. Journal of Computational Physics, 309:368–386,
2016.

[77] K. Fukunaga. Introduction to statistical pattern recognition. Academic Press, New York,
1972.

[78] H. Garcke, C. Hecht, M. Hinze, C. Kahle, and K. F. Lam. Shape optimization for sur-
face functionals in Navier-Stokes flow using a phase field approach. Interfaces and Free
Boundaries, 18(2):219–261, 2016.

[79] H. Garcke, B. Nestler, and B. Stoth. On anisotropic order parameter models for multi-phase
systems and their sharp interface limits. Physica D: Nonlinear Phenomena, 115(1-2):87–
108, 1998.

[80] D. Gay and W. H. Ray. Identification and control of linear distributed parameter sys-
tems through the use of experimentally determined singular functions. IFAC Proceedings
Volumes, 20(1):173–179, 1987.
REFERENCES 139

[81] M. S. Gockenbach. Understanding and implementing the finite element method, volume 97.
SIAM, 2006.

[82] W. Gong, M. Hinze, and Z. Zhou. Space-time finite element approximation of parabolic
optimal control problems. Journal of Numerical Mathematics, 20(2):111–146, 2012.

[83] C. Gräßle, M. Gubisch, S. Metzdorf, S. Rogg, and S. Volkwein. POD basis updates for
nonlinear PDE control. at - Automatisierungstechnik, 65(5):298–307, 2017.

[84] C. Gräßle and M. Hinze. The combination of POD model reduction with adaptive finite
element methods in the context of phase field models. PAMM, 17(1):47–50, 2017.

[85] C. Gräßle and M. Hinze. POD reduced-order modeling for evolution equations utilizing ar-
bitrary finite element discretizations. Advances in Computational Mathematics, 44(6):1941–
1978, 2018.

[86] C. Gräßle, M. Hinze, J. Lang, and S. Ullmann. POD model order reduction with space-
adapted snapshots for incompressible flows. Accepted for publication in Advances in Com-
putational Mathematics, preprint available [Link] 2019.

[87] C. Gräßle, M. Hinze, and N. Scharmacher. POD for optimal control of the Cahn-Hilliard
system using spatially adapted snapshots. In Numerical Mathematics and Advanced Ap-
plications ENUMATH 2017, pages 703–711. Springer International Publishing, 2019.

[88] L. Grüne and J. Pannek. Nonlinear model predictive control. In Nonlinear Model Predictive
Control, pages 43–66. Springer, 2011.

[89] M. Gubisch and S. Volkwein. Proper orthogonal decomposition for linear-quadratic optimal
control. In P. Benner, A. Cohen, M. Ohlberger, and K. Willcox, editors, Model Reduction
and Approximation: Theory and Algorithms, 15:5–66, 2017.

[90] F. Guillén-González and G. Tierra. On linear schemes for a Cahn–Hilliard diffuse interface
model. Journal of Computational Physics, 234:140–171, 2013.

[91] J. Hadamard. Sur les problèmes aux dérivées partielles et leur signification physique.
Princeton University Bulletin, XIII(4), 1902.

[92] F. Haußer, S. Rasche, and A. Voigt. The influence of electric fields on nanostructures -
simulation and control. Mathematics and Computers in Simulation, 80(7):1449–1457, 2010.

[93] M. Heinkenschloss and E. Sachs. Numerical solution of a constrained control problem for a
phase field model. In Control and Estimation of Distributed Parameter Systems: Nonlinear
Phenomena, pages 171–187. Springer, 1994.

[94] J. Heywood and R. Rannacher. Finite Element Approximation of the Nonstationary

Navier–Stokes Problem. I. Regularity of Solutions and Second-Order Error Estimates for
Spatial Discretization. SIAM Journal on Numerical Analysis, 19(2):275–311, 1982.

[95] M. Hintermüller, M. Hinze, C. Kahle, and T. Keil. A goal-oriented dual-weighted adaptive

finite element approach for the optimal control of a nonsmooth Cahn–Hilliard–Navier–
Stokes system. Optimization and Engineering, 19(3):629–662, 2018.

[96] M. Hintermüller, M. Hinze, and M. H. Tber. An adaptive finite-element Moreau–Yosida-

based solver for a non-smooth Cahn–Hilliard problem. Optimization Methods and Software,
26(4-5):777–811, 2011.
140 REFERENCES

[97] M. Hintermüller and D. Wegner. Distributed optimal control of the Cahn–Hilliard system
including the case of a double-obstacle homogeneous free energy density. SIAM Journal
on Control and Optimization, 50(1):388–418, 2012.

[98] M. Hinze. A variational discretization concept in control constrained optimization: the

linear-quadratic case. Computational Optimization and Applications, 30(1):45–61, 2005.

[99] M. Hinze, J. Krenciszek, and R. Pinnau. Proper orthogonal decomposition for free bound-
ary value problems. Hamburger Beiträge zur Angewandten Mathematik, 2014.

[100] M. Hinze and K. Kunisch. Three control methods for time-dependent fluid flow. Flow,
Turbulence and Combustion, 65(3-4):273–298, 2000.

[101] M. Hinze, R. Pinnau, M. Ulbrich, and S. Ulbrich. Optimization with PDE constraints,
volume 23. Springer Science & Business Media, 2008.

[102] M. Hinze and S. Volkwein. Proper orthogonal decomposition surrogate models for nonlinear
dynamical systems: Error estimates and suboptimal control. In Dimension reduction of
large-scale systems, pages 261–306. Springer, 2005.

[103] M. Hinze and S. Volkwein. Error estimates for abstract linear–quadratic optimal con-
trol problems using proper orthogonal decomposition. Computational Optimization and
Applications, 39(3):319–345, 2008.

[104] P. C. Hohenberg and B. I. Halperin. Theory of dynamic critical phenomena. Reviews of

Modern Physics, 49(3):435–479, 1977.

[105] P. Holmes, J. L. Lumley, G. Berkooz, and C. W. Rowley. Turbulence, coherent structures,

dynamical systems and symmetry. Cambridge University Press, 2012.

[106] R. H. Hoppe and Z. Liu. Snapshot location by error equilibration in proper orthogonal
decomposition for linear and semilinear parabolic partial differential equations. Journal of
Numerical Mathematics, 22(1):1–32, 2014.

[107] W. Huang and R. D. Russell. Adaptive moving mesh methods, volume 174. Springer Science
& Business Media, 2010.

[108] A. Humphries and A. Stuart. Runge–Kutta methods for dissipative and gradient dynamical
systems. SIAM Journal on Numerical Analysis, 31(5):1452–1485, 1994.

[109] T. Iliescu and Z. Wang. Are the snapshot difference quotients needed in the proper or-
thogonal decomposition? SIAM Journal on Scientific Computing, 36(3):A1221–A1250,
2014.

[110] A. Iollo, S. Lanteri, and J.-A. Désidéri. Stability properties of POD–Galerkin approxima-
tions for the compressible Navier–Stokes equations. Theoretical and Computational Fluid
Dynamics, 13(6):377–396, 2000.

[111] K. Ito and S. S. Ravindran. A reduced-order method for simulation and control of fluid
flows. Journal of Computational Physics, 143(2):403–425, 1998.

[112] V. John. Finite element methods for incompressible flow problems, volume 51 of Springer
Series in Computational Mathematics. Springer, Cham, 2016.

[113] V. John, A. Linke, C. Merdon, M. Neilan, and L. G. Rebholz. On the divergence constraint
in mixed finite element methods for incompressible flows. SIAM Review, 59(3):492–544,
2017.
REFERENCES 141

[114] C. Kahle. Simulation and Control of Two-Phase Flow Using Diffuse Interface Models. PhD
thesis, Universität Hamburg, 2014.

[115] B. Karasözen and F. Yılmaz. Optimal boundary control of the unsteady Burgers equation
with simultaneous space-time discretization. Optimal Control Applications and Methods,
35(4):423–434, 2014.

[116] M. Kärcher and M. Grepl. A certified reduced basis methods for parametrized elliptic
optimal control problems. ESAIM: Control, Optimisation and Calculus of Variations,
20(2):416–441, 2014.

[117] A. Kolmogoroff. Über die beste Annäherung von Funktionen einer gegebenen Funktionen-
klasse. Annals of Mathematics, pages 107–110, 1936.

[118] M. E. Kowalski and J.-M. Jin. Model-order reduction of nonlinear models of electro-
magnetic phased-array hyperthermia. IEEE Transactions on Biomedical Engineering,
50(11):1243–1254, 2003.

[119] K. Kunisch and S. Volkwein. Control of the Burgers equation by a reduced-order approach
using proper orthogonal decomposition. Journal of optimization theory and applications,
102(2):345–371, 1999.

[120] K. Kunisch and S. Volkwein. Galerkin proper orthogonal decomposition methods for
parabolic problems. Numerische Mathematik, 90(1):117–148, 2001.

[121] K. Kunisch and S. Volkwein. Galerkin proper orthogonal decomposition methods for a
general equation in fluid dynamics. SIAM Journal on Numerical Analysis, 40(2):492–515,
2002.

[122] K. Kunisch and S. Volkwein. Proper orthogonal decomposition for optimality systems.
ESAIM: Mathematical Modelling and Numerical Analysis, 42(1):1–23, 2008.

[123] K. Kunisch and S. Volkwein. Optimal snapshot location for computing POD basis func-
tions. ESAIM: Mathematical Modelling and Numerical Analysis, 44(3):509–529, 2010.

[124] O. Ladyzhenskaya. The mathematical theory of viscous incompressible flow. Mathematics

and its Applications, volume 2, Gordon and Breach, Science Publishers, 1969.

[125] J. Lang. Adaptive multilevel solution of nonlinear parabolic PDE systems: theory, algo-
rithm, and applications, volume 16. Springer Science & Business Media, 2013.

[126] O. Lass. Reduced order modeling and parameter identification for coupled nonlinear PDE
systems. PhD thesis, Universität Konstanz, 2014.

[127] O. Lass and S. Volkwein. Adaptive POD basis computation for parametrized nonlinear
systems using optimal snapshot location. Computational Optimization and Applications,
58(3):645–677, 2014.

[128] T. Lassila, A. Manzoni, A. Quarteroni, and G. Rozza. Model order reduction in fluid
dynamics: challenges and perspectives. In Reduced Order Methods for modeling and com-
putational reduction, pages 235–273. Springer, 2014.

[129] Y. T. Lee and A. A. Requicha. Algorithms for computing the volume and other integral
properties of solids. I. Known methods and open issues. Communications of the ACM,
25(9):635–641, 1982.

[130] A. Linke. Divergence-free mixed finite elements for the incompressible Navier-Stokes Equa-
tion. PhD thesis, Friedrich-Alexander-Universität Erlangen-Nürnberg, 2007.
142 REFERENCES

[131] J. L. Lions. Optimal control of systems governed by partial differential equations, volume
170. Springer Berlin, 1971.

[132] A. Logg, K.-A. Mardal, and G. Wells. Automated solution of differential equations by the
finite element method: The FEniCS book, volume 84. Springer Science & Business Media,
2012.

[133] J. L. Lumley. The structure of inhomogeneous turbulence. Atmospheric Turbulence and

Wave Propagation, pages 166–178, 1967.

[134] A. Massing, M. G. Larson, and A. Logg. Efficient implementation of finite element methods
on nonmatching and overlapping meshes in three dimensions. SIAM Journal on Scientific
Computing, 35(1):C23–C47, 2013.

[135] B. Mirtich. Fast and accurate computation of polyhedral mass properties. Journal of
graphics tools, 1(2):31–50, 1996.

[136] W. F. Mitchell. Adaptive refinement for arbitrary finite element spaces with hierarchical
bases. Journal of Computational and Applied Mathematics, 36(1):65–78, 1991.

[137] L. Modica. The gradient theory of phase transitions and the minimal interface criterion.
Archive for Rational Mechanics and Analysis, 98(2):123–142, 1987.

[138] I. Neitzel, U. Prüfert, and T. Slawig. Solving time-dependent optimal control problems in
COMSOL multiphysics. In Proceedings of the COMSOL Conference Hannover, 2008.

[139] I. Neitzel, U. Prüfert, and T. Slawig. A smooth regularization of the projection formula
for constrained parabolic optimal control problems. Numerical functional analysis and
optimization, 32(12):1283–1315, 2011.

[140] B. Nestler and A. Choudhury. Phase-field modeling of multi-component systems. Current

opinion in solid state and Materials Science, 15(3):93–105, 2011.

[141] N. Nguyen, A. Patera, and J. Peraire. A ‘best points’ interpolation method for efficient
approximation of parametrized functions. International journal for numerical methods in
engineering, 73(4):521–543, 2008.

[142] P. Nigro, M. Anndif, Y. Teixeira, P. Pimenta, and P. Wriggers. An adaptive model order
reduction by proper snapshot selection for nonlinear dynamical problems. Computational
Mechanics, 57(4):537–554, 2016.

[143] Y. Oono and S. Puri. Study of phase-separation dynamics by use of cell dynamical systems.
I. Modeling. Physical Review A, 38(1):434, 1988.

[144] G. M. Oxberry, T. Kostova-Vassilevska, W. Arrighi, and K. Chand. Limited-memory

adaptive snapshot selection for proper orthogonal decomposition. International Journal
for Numerical Methods in Engineering, 109(2):198–217, 2017.

[145] A. Pazy. Semigroups of linear operators and applications to partial differential equations,
volume 44. Springer Science & Business Media, 1983.

[146] K. Pearson. LIII. On lines and planes of closest fit to systems of points in space. The
London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559–
572, 1901.

[147] A. D. Pelton. Thermodynamics and phase diagrams of materials. Materials science and
technology, 2006.
REFERENCES 143

[148] J. R. Phillips. Automated extraction of nonlinear circuit macromodels. In Custom Inte-

grated Circuits Conference, 2000. CICC. Proceedings of the IEEE 2000, pages 451–454.
IEEE, 2000.

[149] J. R. Phillips. Projection-based approaches for model reduction of weakly nonlinear, time-
varying systems. IEEE Transactions on computer-aided design of integrated circuits and
systems, 22(2):171–187, 2003.

[150] A. Pinkus. N-widths in Approximation Theory, volume 7. Springer Science & Business
Media, 2012.

[151] R. Pinnau. Model reduction via proper orthogonal decomposition. In Model Order Reduc-
tion: Theory, Research Aspects and Applications, pages 95–109. Springer, 2008.

[152] R. W. Preisendorfer. Principal component analysis in meteorology and oceanography.

Elsevier Science Publishers, 17:425, 1988.

[153] A. Quarteroni, A. Manzoni, and F. Negri. Reduced basis methods for partial differential
equations: an introduction, volume 92. Springer, 2015.

[154] J.-P. Raymond. Stokes and Navier–Stokes equations with nonhomogeneous boundary con-
ditions. Annales de l’Institut Henri Poincare – Non Linear Analysis, 24(6):921 – 951,
2007.

[155] J. P. Raymond and H. Zidani. Hamiltonian Pontryagin’s principles for control prob-
lems governed by semilinear parabolic equations. Applied Mathematics and Optimization,
39(2):143–177, 1999.

[156] M. Reed and B. Simon. Methods of modern mathematical physics, I: Functional analysis.
Academic Press, 1980.

[157] M. Rewienski and J. White. A trajectory piecewise-linear approach to model order reduc-
tion and fast simulation of nonlinear circuits and micromachined devices. IEEE Transac-
tions on computer-aided design of integrated circuits and systems, 22(2):155–170, 2003.

[158] E. Rocca and J. Sprekels. Optimal distributed control of a nonlocal convective Cahn–
Hilliard equation by the velocity in three dimensions. SIAM Journal on Control and
Optimization, 53(3):1654–1680, 2015.

[159] A. Rosenfeld and A. C. Kak. Digital picture processing. Academic press, 1976.

[160] G. Rozza and K. Veroy. On the stability of the reduced basis method for Stokes equa-
tions in parametrized domains. Computer methods in applied mechanics and engineering,
196(7):1244–1260, 2007.

[161] A. Schmidt and K. G. Siebert. Design of Adaptive Finite Element Software: The Finite
Element Toolbox ALBERTA. Lecture Notes in Computational Science and Engineering,
(42), 2005.

[162] M. Schu. Adaptive trust-region POD methods and their application in finance. PhD thesis,
University of Trier, 2012.

[163] A. J. Siade, M. Putti, and W. W.-G. Yeh. Snapshot selection for groundwater model
reduction using proper orthogonal decomposition. Water Resources Research, 46(8), 2010.

[164] J. Simon. Compact sets in the space Lp (0, t; B). Annali di Matematica Pura ed Applicata
(IV), 146(1):65–96, 1987.
144 REFERENCES

[165] J. R. Singler. New POD error expressions, error bounds, and asymptotic results for reduced
order models of parabolic PDEs. SIAM Journal on Numerical Analysis, 52(2):852–876,
2014.

[166] L. Sirovich. Turbulence and the dynamics of coherent structures. I. Coherent structures.
Quarterly of Applied Mathematics, 45(3):561–571, 1987.

[167] L. Sirovich. Turbulence and the dynamics of coherent structures. II. Symmetries and
transformations. Quarterly of Applied Mathematics, 45(3):573–582, 1987.

[168] L. Sirovich. Turbulence and the dynamics of coherent structures. III. Dynamics and scaling.
Quarterly of Applied Mathematics, 45(3):583–590, 1987.

[169] K. Steih. Reduced basis methods for time-periodic parametric partial differential equations.
PhD thesis, Universität Ulm, 2014.

[170] I. Steinbach. Phase-field models in materials science. Modelling and simulation in materials
science and engineering, 17(7):073001, 2009.

[171] R. Stevenson. Optimality of a standard adaptive finite element method. Foundations of

Computational Mathematics, 7(2):245–269, 2007.

[172] J. E. Taylor and J. W. Cahn. Linking anisotropic sharp and diffuse surface motion laws
via gradient flows. Journal of Statistical Physics, 77(1-2):183–197, 1994.

[173] R. Temam. Navier-Stokes equations, volume 2 of Studies in Mathematics and its Applica-
tions. North-Holland Publishing Co., Amsterdam-New York, 1979.

[174] J. B. Tenenbaum, V. De Silva, and J. C. Langford. A global geometric framework for

nonlinear dimensionality reduction. Science, 290(5500):2319–2323, 2000.

[175] Z. Tokoutsi, M. A. Grepl, K. Veroy-Grepl, M. Baragona, and R. Maessen. Real-Time

Optimization of Thermal Ablation Treatments. In Numerical Mathematics and Advanced
Applications ENUMATH 2017, pages 673–681. Springer International Publishing, 2019.

[176] F. Tröltzsch. Optimal control of partial differential equations: theory, methods, and appli-
cations, volume 112. American Mathematical Soc., 2010.

[177] F. Tröltzsch and S. Volkwein. POD a-posteriori error estimates for linear-quadratic optimal
control problems. Computational Optimization and Applications, 44(1):83, 2009.

[178] P. G. Tucker. Computation of unsteady internal flows: fundamental methods with case
studies. Springer Science & Business Media, 2012.

[179] S. Ulbrich and J. C. Ziems. Adaptive multilevel trust-region methods for time-dependent
PDE-constrained optimization. Portugaliae Mathematica, 74(1):37–68, 2017.

[180] S. Ullmann. POD-Galerkin Modeling for Incompressible Flows with Stochastic Boundary
Conditions. PhD thesis, Technische Universität Darmstadt, 2014.

[181] S. Ullmann, M. Rotkvic, and J. Lang. POD-Galerkin reduced-order modeling with adaptive
finite element snapshots. Journal of Computational Physics, 325:244–258, 2016.

[182] M. Uzunca and B. Karasözen. Energy stable model order reduction for the Allen-Cahn
equation. In Model Reduction of Parametrized Systems, pages 403–419. Springer, 2017.

[183] R. Verfürth. A posteriori error estimation techniques for finite element methods. Numerical
mathematics and scientific computation. OUP Oxford, 2013.
REFERENCES 145

[184] K. Veroy and A. Patera. Certified real-time solution of the parametrized steady incom-
pressible Navier-Stokes equations: rigorous reduced-basis a posteriori error bounds. Inter-
national Journal for Numerical Methods in Fluids, 47(8-9):773–788, 2005.

[185] S. Volkwein. Optimal control of a phase-field model using proper orthogonal decomposi-
tion. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte
Mathematik und Mechanik: Applied Mathematics and Mechanics, 81(2):83–97, 2001.

[186] S. Volkwein. Proper orthogonal decomposition: Theory and reduced-order modelling. Lec-
ture Notes, University of Konstanz, 2013.

[187] S. Volkwein and S. Weiland. An algorithm for Galerkin projections in both time and spatial
coordinates. Proc. 17th MTNS, 2006.

[188] Z. Wang. Nonlinear model reduction based on the finite element method with interpolated
coefficients: semilinear parabolic equations. Numerical Methods for Partial Differential
Equations, 31(6):1713–1741, 2015.

[189] A. Yagi. Abstract parabolic evolution equations and their applications. Springer Science &
Business Media, 2009.

[190] M. Yano. A minimum-residual mixed reduced basis method: Exact residual certification
and simultaneous finite-element reduced-basis refinement. ESAIM: Mathematical Modelling
and Numerical Analysis, 50(1):163–185, 2016.

[191] M. Yano. A reduced basis method with an exact solution certificate and spatio-parameter
adaptivity: application to linear elasticity, pages 55–76. Springer International Publishing,
2017.

[192] M. Yano. A reduced basis method for coercive equations with an exact solution certificate
and spatio-parameter adaptivity: energy-norm and output error bounds. SIAM Journal
on Scientific Computing, 40(1):A388–A420, 2018.

[193] F. N. Yilmaz. Space-time discretization of optimal control of Burgers equation using both
discretize-then-optimize and optimize-then-discretize approaches. PhD thesis, Middle East
Technical University, 2011.

[194] E. Zeidler. Nonlinear functional analysis and its applications, I: fixed-point theorems.
Springer Verlag, 1986.

[195] E. Zeidler. Nonlinear functional analysis and its applications, II/B: nonlinear monotone
operators. Springer Verlag, 1990.