Get the full ebook with Bonus Features for a Better Reading Experience on ebookgate.

com

Global Regularity and Uniqueness of Solutions in a

Surface Growth Model Using Rigorous a Posteriori
Methods 1st Edition Christian Nolde

https://ebookgate.com/product/global-regularity-and-
uniqueness-of-solutions-in-a-surface-growth-model-using-
rigorous-a-posteriori-methods-1st-edition-christian-nolde/

OR CLICK HERE

DOWLOAD NOW

Download more ebook instantly today at https://ebookgate.com

Instant digital products (PDF, ePub, MOBI) available
Download now and explore formats that suit you...

A posteriori error estimation techniques for finite

element methods 1st Edition Verfürth

https://ebookgate.com/product/a-posteriori-error-estimation-
techniques-for-finite-element-methods-1st-edition-verfurth/

ebookgate.com

Principles of electromagnetic methods in surface

geophysics 1st Edition Alexeev

https://ebookgate.com/product/principles-of-electromagnetic-methods-
in-surface-geophysics-1st-edition-alexeev/

ebookgate.com

A posteriori estimates for partial differential equations

Repin

https://ebookgate.com/product/a-posteriori-estimates-for-partial-
differential-equations-repin/

ebookgate.com

India s Emerging Financial Market A Flow of Funds Model

Routledge Studies in the Growth Economies of Asia 1st
Edition Tomoe Moore
https://ebookgate.com/product/india-s-emerging-financial-market-a-
flow-of-funds-model-routledge-studies-in-the-growth-economies-of-
asia-1st-edition-tomoe-moore/
ebookgate.com
Markedness and Economy in a Derivational Model of
Phonology Andrea Calabrese

https://ebookgate.com/product/markedness-and-economy-in-a-
derivational-model-of-phonology-andrea-calabrese/

ebookgate.com

Surface Wave Methods for Near Surface Site

Characterization 1st Edition Sebastiano Foti

https://ebookgate.com/product/surface-wave-methods-for-near-surface-
site-characterization-1st-edition-sebastiano-foti/

ebookgate.com

Naval Diplomacy in 21st Century A Model for the Post Cold

War Global Order 1st Edition Kevin Rowlands

https://ebookgate.com/product/naval-diplomacy-in-21st-century-a-model-
for-the-post-cold-war-global-order-1st-edition-kevin-rowlands/

ebookgate.com

Surface Modification of Biomaterials Methods Analysis and

Applications 1st Edition R. Williams

https://ebookgate.com/product/surface-modification-of-biomaterials-
methods-analysis-and-applications-1st-edition-r-williams/

ebookgate.com

Surface and Radiological Anatomy 3rd Edition A. Halim

https://ebookgate.com/product/surface-and-radiological-anatomy-3rd-
edition-a-halim/

ebookgate.com
 Augsburger Schriften

32 zur Mathematik, Physik

und Informatik

Global Regularity and

Uniqueness of Solutions in a
Surface Growth Model Using
Rigorous A-Posteriori Methods

Christian Nolde

λογος
 Christian Nolde

Global Regularity and

Uniqueness of Solutions in a
Surface Growth Model Using
Rigorous A-Posteriori Methods

λογος
Augsburger Schriften zur Mathematik, Physik und Informatik
Band 32

Edited by:
Professor Dr. B. Schmidt
Professor Dr. B. Aulbach
Professor Dr. F. Pukelsheim
Professor Dr. W. Reif
Professor Dr. D. Vollhardt

Bibliographic information published by the Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the

Deutsche Nationalbibliografie; detailed bibliographic data are
available in the Internet at http://dnb.d-nb.de .

c Copyright Logos Verlag Berlin GmbH 2017

All rights reserved.

ISBN 978-3-8325-4453-9
ISSN 1611-4256

Logos Verlag Berlin GmbH

Comeniushof, Gubener Str. 47,
10243 Berlin
Tel.: +49 030 42 85 10 90
Fax: +49 030 42 85 10 92
INTERNET: http://www.logos-verlag.de
 Global Regularity and
Uniqueness of Solutions in a
Surface Growth Model Using
Rigorous A-Posteriori Methods

Dissertation

zur Erlangung des akademischen Grades

Dr. rer. nat.

eingereicht an der

Mathematisch-Naturwissenschaftlich-Technischen Fakultät
der

UNIVERSITÄT AUGSBURG

von

Christian Nolde

Augsburg, Dezember 2016

Gutachter:

• Prof. Dr. Dirk Blömker, Universität Augsburg

• Prof. Dr. Malte Peter, Universität Augsburg

Datum der mündlichen Prüfung: 17.02.2017

 Contents

1 Introduction 1
1.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 General Setting and Notation . . . . . . . . . . . . . . . . . . 6
1.3 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 A-priori Analysis 9
2.1 Energy Estimate . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Time and Smallness Conditions . . . . . . . . . . . . . . . . . 12
2.3 ODE Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Verification Methods 17
3.1 First Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Second Method . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3 Third Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4 Numerical Comparison . . . . . . . . . . . . . . . . . . . . . . 21
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4 Eigenvalue Estimate 31
4.1 Setting and Theorem . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Proof of the Theorem . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Comparison with the Previous Estimate . . . . . . . . . . . . 35
4.4 Application to the Surface Growth Equation . . . . . . . . . . 36

5 Numerical Preparations 39
5.1 Norm of the Residual . . . . . . . . . . . . . . . . . . . . . . . 39
5.2 Rigorous Computable Bound for Method 3 . . . . . . . . . . . 41

iii
iv CONTENTS

5.3 Rigorous Computable Bound including the Eigenvalue Estimate 45

6 Simulations 53

A Basic Tools 67
A.1 General Inequalities . . . . . . . . . . . . . . . . . . . . . . . . 67
A.2 L2 Estimate of the SFG Equation . . . . . . . . . . . . . . . . 71
A.3 Scaling Property of the SFG Equation . . . . . . . . . . . . . 71

B Numerical Implementation 73
B.1 Approximation Method . . . . . . . . . . . . . . . . . . . . . . 73
B.2 Eigenvalue Estimate . . . . . . . . . . . . . . . . . . . . . . . 75
 CHAPTER 1

Introduction

The aim of this thesis is to set up rigorous a-posteriori methods to prove

the existence and uniqueness of a smooth global solution to a fourth order
semi-linear parabolic partial differential equation (PDE). More specific, we
consider the following surface growth equation for the height u(t, x) ∈ R at
time t > 0 over a point x ∈ [0, 2π]
ut = −uxxxx − (ux 2 )xx x ∈ [0, 2π], t ∈ [0, T ] (1.1)
with periodic boundary conditions and subject to a moving frame, which
yields the zero-average condition 02π u(x, t) dx = 0.
R

This equation, usually with additional noise term, was introduced as a

phenomenological model for the growth of vapor deposited amorphous sur-
faces [SP94]; [RLH00], and was also used to describe ion-sputtering processes,
where a surface is eroded by an ion-beam [CVG05]. The one dimensional
equation appears as a model for the boundaries of terraces in the epitaxy of
silicon [FV06]. A more detailed list of references can be found in [BR12] and
[BR15].
On the mathematical side, the existence of global weak solutions and lo-
cal unique and smooth solutions is shown in [BR09] and expanded in [BR12].
Unfortunately, due to a gap in the regularity requirements for the initial con-
dition, these weak solutions are neither unique nor smooth. It is also known,
that the complex-valued surface growth equation in fact does experience a
blowup in finite time for certain initial values, where the mass is transfered
faster to higher Fourier modes than it dissipates [BR15]. It remains if this is
also true for the real-valued function.

1
2

An exhaustive summary of the equation’s derivation and known mathe-

matical properties, in the stochastic as well as the deterministic case, is given
by Blömker and Romito in [BR15].

The famous sibling

The, by far, most famous problem for existence, uniqueness and regularity of
solutions to a semi-linear parabolic partial differential equation is the Navier-
Stokes Problem [Fef06]. This problem poses the question, if the motion of a
fluid is uniquely determined and smooth for all time, as long as the initial
data is smooth. Sometimes, it is also described as the problem whether
the water inside your bathtub might explode (which is a nice pun to the
mathematical “blowup” we are looking for). In the year 2000, the Clay
Mathematics Institute published their famous list of seven Millennium Prize
Problems including the Navier-Stokes Problem. Despite all this attention,
the problem remains unsolved until today. For more information about the
Navier-Stokes Problem and why it is so hard to solve, there is a detailed blog
post by Terrence Tao [07] where he analyzes, why methods, that work for
other problems, fail when it comes to Navier-Stokes.

Computer aided proofs

An increasingly popular method, especially for problems which are not (yet)
solvable in the classical way, are computer assisted proofs. These range from
“simple” brute force methods, as used in the proof for the famous four color
problem (see [AH89]), to extremely involved artificial intelligence programs
which use automated reasoning to proof or even deduce new theorems (e.g.
[HKM16]).
A very common type, and the one we will refer to in this thesis, is rigorous
numerics, where quantities calculated by the computer are set-valued rather
than a floating point number. These sets represent the error bound for the
true value and take into account machine precision, round-off and truncation
of values and arithmetic operations.
Unsurprisingly, the fact that computers are involved in mathematical
proofs was controversial in the beginning (and for some people maybe still
is), as these are usually not verifiable in the way proofs used to be, and
that they could suffer from bugs (errors) in software (the program itself,
compiler/interpreter, operating system) or hardware (memory, processor).
The answer to this is, that computer assisted proofs have to be verified
CHAPTER 1. INTRODUCTION 3

differently. In some sense they are like rigorous experiments, which have to
be replicated using different hardware and software to avoid bugs. There is
also the concept of “formal verification” to prove the correctness of algorithms
which is useful here. For the rigorous numerics, it might be the best to look
at them like a “really huge” calculation inside a proof. This is not uncommon
to mathematical proofs, just that in this case the calculations are done by
a computer rather than a human being and they just have to be verified
accordingly.
The Fields Medallist William P. Thurston wrote over 20 years ago in
[Thu94] about that topic:
The rapid advance of computers has helped dramatize this point,
because computers and people are very different. For instance,
when Appel and Haken completed a proof of the 4-color map
theorem using a massive automatic computation, it evoked much
controversy. I interpret the controversy as having little to do
with doubt people had as to the veracity of the theorem or the
correctness of the proof. Rather, it reflected a continuing desire
for human understanding of a proof, in addition to knowledge
that the theorem is true.
On a more everyday level, it is common for people first starting
to grapple with computers to make large-scale computations of
things they might have done on a smaller scale by hand. They
might print out a table of the first 10,000 primes, only to find that
their printout isn’t something they really wanted after all. They
discover by this kind of experience that what they really want
is usually not some collection of “answers” – what they want is
understanding.
Nowadays, rigorous methods for proving numerically the existence of solu-
tions for PDEs are an active research field. In addition to the approach we
will take, there are other methods based on topological arguments like the
Conley index, see [Mai+08]; [DLM07]; [Zgl10], for example. For solutions of
elliptic PDEs there are also methods using Brouwer’s fixed-point theorem,
as discussed in the review article [Plu08] and the references therein, or for a
small summary for fixed-point methods in rigorous numerics see [BL15].

Inspiration for our approach

In [Che+07]; [DR08]; [RMS13]; [RS08] Chernyshenko, Constantin, Dashti,
Marı́n-Rubio, Robinson, Sadowski and Titi applied the idea of rigorous nu-
4 1.1. OUTLINE

merics to the Navier-Stokes Problem. They set up the analytic framework to

show the existence of a smooth global solution using rigorous numerics, but
did not carry out the actual numeric implementation. A similar approach
using an integral equation based on the mild formulation was proposed in
[MP08]; [MP11]. The inspiration to establish similar methods for the sur-
face growth equation is the work of Robinson et al. [Che+07], but a similar
method was proposed by Morosi and Pizzocchero in [MP12].
The key ingredients for the analytic framework are a) a global regularity
result for small initial data and b) the existence of local unique and smooth
solutions. For Navier-Stokes a) is provided by Leray in [Ler34] and b) by
Koch and Tataru in [KT01]. The good news is, that both properties hold in
a similar way for the surface growth equation, which was shown by Blömker
and Romito in [BR09].
Now, the main idea of the method is the derivation of a scalar ODE for
the H 1 -norm of the difference between an arbitrary approximation, which
satisfies the boundary conditions, and the solution. The coefficients of this
ODE depend only on the numerical data (or any other approximation used).
As long as the solution of the ODE stays finite, one can rely on the contin-
uation property of unique local solutions, and thus have a smooth unique
solution up to a blowup time of the ODE. This is improved by the smallness
property and its corollary, a time property which proves global existence if
this solution does not experience a blowup until a given time T ∗ , or if the
solution is falling below a certain threshold. Please note, that this method
is only able to verify the existence of a smooth and unique global solution.
It is unable to verify a blowup.
As we limit ourselves to the scalar surface growth equation, the numerical
implementation should be a lot easier compared to 3D Navier-Stokes. More-
over, for the two-dimensional case the situation of energy estimates seems
even worse, as global Rexistence could only be established in H −1 using the
non-standard energy 02π eu(x) dx, see [Win11] for details. Nevertheless, we
believe that it should be possible to treat the 2D case using similar methods,
but the analysis becomes more delicate since in two dimensions H 1 is the
critical space (see [BR09]; [BR12]).

1.1 Outline
In Chapter 2, we will set up all necessary analytic results for the surface
growth equation. We will carry out the energy estimate to obtain the dif-
ferential inequality (2.1) governing the error between solution and approx-
imation. After that, we will formulate the smallness and time conditions
CHAPTER 1. INTRODUCTION 5

(Theorem 2.2 and Theorem 2.3) which allow us to prove global existence in
finite time, what is very important for practical purpose. Following, we pro-
pose two analytical methods: one, based on the standard Gronwall Lemma,
enforces a ‘small data’ hypothesis and adds little to standard analytical ex-
istence proofs. The second is based on an explicit analytical upper bound to
the ODE solution (Lemma 2.5 and Corollary 2.7).
Chapter 3 will now apply these results to the differential inequality (2.1)
(Theorem 3.1 and Theorem 3.2). We will also test Corollary 2.7 with restart-
ing, which means that one applies the analytical upper bound on a succession
of small intervals of length h > 0 to the numerical solution and then restarts
the argument (Theorem 3.3). We will also give a formal calculation which
indicates that the upper bound from the third method in the limit of step-
size to zero converges to the solution of the ODE. To finish the chapter, we
will make first (non rigorous) numerical experiments to compare the bounds
delivered by the methods. The simulations show that all methods work (i.e.
they are able to reach at least one of our two goals, smallness and time), but
Method 3 (Corollary 2.7 with restarting, Theorem 3.3) is superior as it is
not as sensitive to crucial quantities like kϕxx k∞ and the residual as its com-
petitors. Please note, that the results from Chapters 2 and 3 have already
been published as joint work with Dirk Blömker and James C. Robinson in
[BNR15].
In Chapter 4, we will limit the influence of kϕxx k∞ by improving our initial
energy estimate (2.1). Therefore we analyze the spectrum of the linearized
operator (here Lv = −vxxxx + (vx ϕx )xx , where ϕ is some given numerical
data) with a rigorous numerical method, which in the case of an unstable
linear operator yields substantially better results, at the price of a signifi-
cantly higher computational time (Theorem 4.1). This approach is similar
to [NH09]; [NKK12]; [Liu15] where this was proposed in a slightly different
context. An important aim for our estimate is, that it has to be efficiently
computable. Therefore, not every theoretical result might be suitable for us.
When we apply this new estimate to our ODE (2.1), we also free all con-
stants from Young inequalities, to further improve the estimate by solving
an optimization problem for given times, and obtain the new ODE bound.
In Chapter 5, we will apply Method 3 (Theorem 3.3) to our old (2.1) and
new bound with eigenvalue estimate (4.5). Further, we will do all calculations
and estimates to rigorously compute all quantities needed by our methods.
Therefore, we will define ϕ as the linear interpolation between the numerically
obtained grid points. This leads to the rigorous computable methods that
we will denote by Method 4 (Theorem 5.1) without eigenvalue estimate, and
Method 5 (Theorem 5.2) with eigenvalue estimate.
Chapter 6 will now deliver the rigorous (except interval arithmetic) sim-
6 1.2. GENERAL SETTING AND NOTATION

ulations of Methods 4 and 5. The results show that both methods work, but
Method 5 with the eigenvalue estimate is a huge improvement over Method
4. Initial values like sin(x) which are already far away from the scope of
analytic results are successfully handled by both methods. They even have
plenty of room for an additional error,which means we can be optimistic that
interval arithmetic should not change the outcome. As the methods are still
vulnerable to initial values where kϕxx k∞ is large (in this case large means
e.g. u0 = sin(3x)), the problem is not solved for all initial values.

1.2 General Setting and Notation

As solutions to our surface growth equation (1.1) are subject to periodic
boundary conditions on [0, 2π] with mean average, we are working on the
space  Z 2π 
H = u : R → R : 2π-periodic, u(x) dx = 0
0

with scalar product h·, ·i and norm

 Z 2π 1/2
2
kuk = |u(x)| dx .
0

We further define the Sobolev-spaces

Hk = {u ∈ H : ∂xk u ∈ L2 ([0, 2π])}.

Note, that by periodicity u ∈ H1 implies ux ∈ H. Moreover, we have

Poincare-inequality with optimal constant 1

kuk ≤ kux k for all u ∈ H1

and thus kux k is a norm on H1 , equivalent to the standard H 1 -Sobolev norm.

Furthermore, interpolation inequality (Theorem A.6)

kux k2 ≤ kuxx kkuk for all u ∈ H2

and Agmon’s inequality (Theorem A.5)

kuk2∞ ≤ kukkux k

also hold with constant 1.

We will use the abbreviations

u0 := u(x, 0)
CHAPTER 1. INTRODUCTION 7

for the initial value of solutions or approximations to our surface growth

model,
kf k−1 := kf kH−1
for the H−1 -norm and
kf k∞ := kf kL∞ ([0,2π])
for the L∞ ([0, 2π])-norm.

1.3 Acknowledgments
Firstly, I would like to express my sincere gratitude to my advisor Prof.
Dirk Blömker for the continuous support, guidance and advice he gave me
during my studies and the patience he had when necessary. Also, having the
possibility to discuss issues at any time is something that can not be taken
for granted and I am extremely thankful for. I really enjoyed and appreciated
every discussion we had, be it during coffee break or late at night on skype.
Further, I want to thank Prof. James C. Robinson for the invitation to
Warwick to discuss parts of this thesis.
I would also like to thank every member of the chair “Nichtlineare Ana-
lysis” for the nice time we had during the last years. It was a pleasure to
meet so many different people from all over the world. Our social activities,
discussions, joint teaching duties and seminars—I enjoyed every part of it.
Finally, I want to thank my family and friends for their support during
the last years.
8 1.3. ACKNOWLEDGMENTS
 CHAPTER 2

A-priori Analysis

The aim of this chapter is to establish upper bounds for the H1 -norm of the
error
d(x, t) := u(x, t) − ϕ(x, t),
where u is a solution to our surface growth equation (1.1) and ϕ is any
arbitrary, but sufficiently smooth approximation, that satisfies the boundary
conditions. Since we know ϕ, if we can control the H1 -norm of d then we
control the H1 -norm of u.
A very important property of the surface growth equation (1.1), and
basically the foundation of this work, is the existence of local solutions, which
are smooth in space and time. Their existence is given by the following
theorem from [BR09] (Theorem 3.1) (There is actually a much simpler proof,
that is sufficient for our setting, in [Hen81] using sectorial operators and
Banach’s fixed point theorem)

Theorem 2.1. Let u0 ∈ H 1 , then there exists a time τ (u0 ) > 0 such that
there is a unique solution u ∈ C 0 ([0, τ (u0 )), H 1 ) satisfying

1) if τ (u0 ) < ∞, then lim sup ku(t)kH 1 = ∞.

t→τ (u0 )

2) u is C ∞ in both, space and time, for all (t, x) ∈ (0, τ (u0 )) × [0, 2π].

Note, that the theorem implies that lack of blowup in H 1 is sufficient to

ensure that the solution exists for all time and is smooth. In particular, all of
the manipulations we make in what follows are valid until the blowup time.

9
10 2.1. ENERGY ESTIMATE

From now on, we consider the solutions with initial data in H1 whose
existence is guaranteed by Theorem 2.1, and approximations ϕ ∈ H4 in
space and H 1 in time.

2.1 Energy Estimate

In this section we prove the key estimate (2.1) on which the theorems of the
following sections are based.
As mentioned before, let u be a smooth solution with initial data in H1
whose existence is guaranteed by Theorem 2.1, and approximations ϕ ∈ H4
in space and H 1 in time.
If we use the surface growth equation (1.1) to find the evolution of d(x, t)
and define the residual of the approximation ϕ by

Res := ϕt + ϕxxxx + (ϕx 2 )xx ,

then we have
dt = −dxxxx − (ux 2 )xx + (ϕx 2 )xx − Res .
By replacing u with d + ϕ we obtain

dt = −dxxxx − (dx 2 )xx − 2(dx ϕx )xx − Res .

As d is sufficiently smooth for t > 0, the H1 -norm holds that

1
∂t kdx k2 = h∂t dx , dx i = −hdxx , dt i
2
= hdxx , dxxxx i + 2hdxx , (dx ϕx )xx i + hdxx , (dx 2 )xx i + hdxx , Resi,
| {z } | {z } | {z } | {z }
A B C D

where h·, ·i is the L2 scalar product.

Note that when using integration by parts, due to d ∈ H4 we know that
3
d ∈ Cper and thus, as we have at most three derivatives in the boundary
terms, these cancel out.
Now consider these terms separately. Integrating by parts we obtain

A = −kdxxx k2

(compare Theorem A.9).

Secondly,
Z 2π
B = −2 dxxx (dx ϕx )x dx
0
CHAPTER 2. A-PRIORI ANALYSIS 11

Z 2π Z 2π
= −2 dxxx dxx ϕx dx − 2 dxxx dx ϕxx dx
0 0
Z 2π Z 2π
=− (dxx 2 )x ϕx dx − 2 dxxx dx ϕxx dx
0 0
Z 2π Z 2π
= (dxx )2 ϕxx dx − 2 dxxx dx ϕxx dx
0 0

and so

|B| ≤ kdxx k2 kϕxx k∞ + 2kdxxx kkdx kkϕxx k∞

≤ 3kdxxx kkdx kkϕxx k∞
1
≤ kdxxx k2 + 9kdx k2 kϕxx k2∞ ,
4
using interpolation (Theorem A.6) and Young’s inequality (Theorem A.2).
For C we have
Z 2π Z 2π
2
C=− (dx )x dxxx dx = −2 dx dxx dxxx dx,
0 0

hence using Agmon’s inequality (Theorem A.5), interpolation, and Young’s

inequality,

|C| ≤ 2kdx kkdxx k∞ kdxxx k

1 3
≤ 2kdx kkdxx k 2 kdxxx k 2
5 7
≤ 2kdx k 4 kdxxx k 4
1 77
≤ kdxxx k2 + kdx k10 ,
4 4
and for the remaining term

1
|D| ≤ k Res k−1 kdxxx k ≤ kdxxx k2 + k Res k2−1 .
4
Combining these estimates and applying Poincaré inequality (Theorem A.3)
with the optimal constant ω = 1, we obtain

1 1 77
∂t kdx k2 ≤ − kdxxx k2 + kdx k10 + k Res k2−1 + 9kdx k2 kϕxx k2∞
2 4 4
7
7 1
 
10 2
≤ kdx k + 9kϕxx k∞ − kdx k2 + k Res k2−1 .
4 4
12 2.2. TIME AND SMALLNESS CONDITIONS

Thus
77 1
 
2
∂t kdx k ≤ kdx k10 + 18kϕxx k2∞ − kdx k2 + 2k Res k2−1 , (2.1)
2 2
which is a scalar differential inequality of type

ξ˙ ≤ bξ 5 + (a(t) − c)ξ + f (t), (2.2)

and by standard ODE comparison principles (Theorem A.8) a solution of the

equality in (2.1) provides an upper bound for kdx k2 .

2.2 Time and Smallness Conditions

We need two important properties of the surface growth model, which we
will prove now. These are for equations like Navier–Stokes well known facts,
namely: that smallness of the solution implies global uniqueness and that
solutions are actually small after some time by energy-type estimates. These
results go back to Leray ([Ler34]), more modern discussions can be found in
[CF88] (Theorem 9.3) and in a setting that parallels the treatment here in
[RS08]. For our model similar results for the critical H 1/2 -norm can be found
in [BR09]. But for our numerical evaluations, we need to derive the precise
values of constants in the H1 -norm, which were not determined before.
First, if the H1 -norm of a solution u at any time is smaller than some con-
stant ε0 , we have regularity after that time. If the H1 -norm is also bounded
up to this time, we have global regularity of u.

Theorem 2.2 (Smallness Condition). If for some t ∈ [0, T ] one has that
kux k is finite on [0, t] and
1
kux (t)k < =: ε0 ,
2
then we have global regularity (and thus uniqueness) of the solution u on
[0, ∞).

Proof. This is established by almost the same estimates derived for the parts
(A) and (C) in Section 2.1 and Young’s inequality with constant δ > 0. To
be more precise:
1 Z 2π
∂t kux k2 = −kuxxx k2 + uxx (ux 2 )xx dx
2 0
7 5
2
≤ −kuxxx k + 2kuxxx k 4 kux k 4
CHAPTER 2. A-PRIORI ANALYSIS 13

  8 −7 1

2 2
≤ −kuxxx k + 2 · δkuxxx k + δ · kux k10
7 8
  8 −7 1 
≤ −kuxxx k2 1 − 2δ − δ · kux k8 .
7 4
If the term in parentheses is positive, i.e.
−7
8 1

0 < 1 − 2δ − δ · kux k8
7 4
7
8

8
⇒ kux k < (1 − 2δ) · 4 · δ ,
7
then we obtain a global bound on kux k2 . The optimal choice for the constant
7
from Young inequality is δ = 16 as this maximizes the right hand side of the
inequality. With this value it follows, that if kux (t)k < 21 we have a negative
derivative and the norm decays over time and is therefore bounded.
The second property is that, based on the smallness condition, we can
determine a time T ∗ , only depending on the initial value u0 , such that
kux (T ∗ )k < ε0 .
Theorem 2.3 (Time Condition). If a solution u is regular up to time
1
T ∗ (u0 ) := ku0 k2 = 2ku0 k2 ,
2ε20
then we have global regularity of the solution u.
At the risk of laboring the point, we only need to verify regularity of a
solution starting at u0 up to time T ∗ (u0 ), and from that point on regularity
is automatic.
Proof. As an a-priori estimate we have (see Theorem A.9)
1
∂t kuk2 = −kuxx k2
2
and thus Z T Z T
1
kux (s)k2 ds ≤ kuxx (s)k2 ds ≤ ku0 k2 ,
0 0 2
where we used the Poincaré inequality (Theorem A.3) with constant ω = 1.
If we now assume that kux (s)k > ε0 for all s ∈ [0, T ], then
1 1
T ε20 < ku0 k2 or T < 2
ku0 k2 .
2 2ε0
This means, that if we stay bounded until time T ∗ := 2ε12 ku0 k2 , we know
0
that kux (t)k ≤ ε0 for at least one t ∈ [0, T ∗ ] and we have global regularity
by the smallness condition.
14 2.3. ODE ESTIMATES

2.3 ODE Estimates

We will now show several methods to bound solutions of ODEs of the type
(2.2). In this section we give the results for the scalar ODE, and present
applications in the next section.
Let us first state a lemma of Gronwall type, based on comparison princi-
ples for ODEs, for which we will only give the idea of a proof.

Lemma 2.4 (Gronwall). Let a, b ∈ L1 ([0, T ], R) and x ∈ W 1,1 ([0, T ], R) ∩

C 0 ([0, T ], R) such that

ẋ ≤ a(t)x + b(t) ∀t ∈ [0, T ].

Then for all t ∈ [0, T ]

Z t  Z t Z t 
x(t) ≤ exp a(s) ds x(0) + exp a(r) dr b(s) ds .
0 0 s

Idea of Proof. Consider the function

Z t Z t
u(t) = x(t) exp{− a(s)ds} with u0 (t) ≤ b(t) exp{− a(s)ds}.
0 0

Integrating and solving for x yields the result.

Lemma 2.5. Consider two functions x, u ∈ W 1,1 ([0, T ], R+ 0 +

0 ) ∩ C ([0, T ], R0 )
such that
ẋ ≤ c(t)xp + e(t) x(0) = x0
with p > 1, c ∈ L1 ([0, T ], R+ 1 +
0 ) and e ∈ L ([0, T ], R0 ), and let u be the solution
of Z T
u̇ = c(t)up u(0) = x0 + e(s) ds.
0

Then x(t) ≤ u(t) for all t ∈ [0, T ].

Proof. First note, that if e ≡ 0 on [0, T ] then by using the standard compar-
ison principle it follows that u(t) ≥ x(t) for all t ∈ [0, T ].
So now we assume that 0T e(s) ds > 0. For a contradiction, suppose
R

that there exists a time t∗ ∈ [0, T ] such that t∗ := inf {t > 0 : x(t) = u(t)}.
Because of the continuity of u(t) and x(t), and u(0) > x(0) due to our initial
assumption 0T e(s) ds > 0, it follows that t∗ > 0. From the definition
R

u(t) > x(t) for all t ∈ [0, t∗ ), and thus

Z t∗ Z t∗
∗ ∗
0 = u(t ) − x(t ) ≥ u(0) − x(0) − e(s) ds + c(s)(u(s)p − x(s)p ) ds
0 0
CHAPTER 2. A-PRIORI ANALYSIS 15

Z T Z t∗
= e(s) ds + c(s)(u(s)p − x(s)p ) ds,
t∗ 0

∗
which is strictly positive provided that 0t c(s) ds > 0.
R
R ∗
If 0t c(s) ds = 0, then as c ≥ 0 we obtain
Z t Z T
x(t) ≤ x(0) + e(s) ds ≤ x(0) + e(s) ds = u(t) ∀t ∈ [0, t∗ ],
0 0

and we can repeat the above argument on the interval [t∗ , T ] to obtain a
contradiction.

Theorem 2.6. Assume x ∈ W 1,1 ([0, T ], R+ 0 +

0 ) ∩ C ([0, T ], R0 ) such that

ẋ ≤ c(t)xp + e(t), x(0) = x0

with p > 1, c ∈ L1 ([0, T ], R+ 1 +

0 ) and e ∈ L ([0, T ], R0 ). Then for all t ∈ [0, T ],
as long as the right-hand side is finite,
 Z t   Z t p−1 Z t − 1
p−1
x(t) ≤ x0 + e(s) ds 1 − (p − 1) x0 + e(s) ds c(s) ds .
0 0 0

Proof. Given the setting of Lemma 2.5, we can solve for u(t). As du =
c(t)up dt, a straightforward calculation shows that
 Z t − 1
p−1
p−1
u(t) = u(0) 1 − (p − 1)u(0) c(s) ds
0

as long as the right-hand side is finite. Thus for all t ∈ [0, T ], as long as the
right-hand side is finite,
 Z T 
x(t) ≤ x0 + e(s) ds
0
  Z T p−1Z t − 1
p−1
× 1 − (p − 1) x0 + e(s) ds c(s) ds
0 0

This holds particularly when T = t.

We now extend this result to differential inequalities of the form

ẋ ≤ b(t)xp + a(t)x + f (t),

where p > 1, f, b ∈ L1 ([0, T ], R+ 1

0 ) and a ∈ L ([0, T ], R), as our inequality
(2.1) is of this type.
16 2.3. ODE ESTIMATES

Corollary 2.7. Assume x ∈ W 1,1 ([0, T ], R+ 0 +

0 ) ∩ C ([0, T ], R0 ) such that

ẋ ≤ b(t)xp + a(t)x + f (t),

with p > 1, b, f ∈ L1 ([0, T ], R+ 1

0 ) and a ∈ L ([0, T ], R). Then for all t ∈ [0, T ],
as long as the right-hand side is finite,
 Z t 
x(t) ≤ e A(t)
x0 + f˜(s) ds
0
  Z t p−1 Z t − 1
p−1
× 1 − (p − 1) · x0 + f˜(s) ds b̃(s) ds
0 0

where
Z t
b̃(t) = b(t)e(p−1)A(t) , f˜(t) = e−A(t) f (t), and A(t) = a(s) ds.
0
Rt
Proof. Consider the substitution y(t) = e−A(t) x(t) with A(t) = 0 a(s) ds. It
follows that

ẏ = −a(t)y + e−A(t) ẋ
≤ −a(t)y + e−A(t) (b(t)xp + a(t)x + f (t))
= b(t)e(p−1)A(t) y p + e−A(t) f (t)
| {z } | {z }
b̃(t) f˜(t)

with b̃(t) ≥ 0 and f˜(t) ≥ 0 for all t ∈ [0, T ]. Here we can apply Theorem 2.6
and obtain
 Z t   Z t p−1 Z t − 1
p−1
y(t) ≤ y0 + f˜ ds 1 − (p − 1) · y0 + f˜ ds b̃ ds .
0 0 0

Now substitute back with x(t) = eA(t) y(t).

 CHAPTER 3

Verification Methods

In this chapter, we outline three techniques for numerical verification. All

of them are based on the key estimate (2.1) for the difference d between an
arbitrary smooth approximation ϕ and a smooth local solution. The first
method is additionally based on the simple Gronwall Lemma 2.4, the second
on Corollary 2.7, and the third is similar to the second method, but restarts
the estimation after a series of short time-steps.

3.1 First Method

By assuming a bound for kdx (t)k, we can reduce (2.1) to a linear ODE which
has the right type for the simple Gronwall Lemma 2.4. With this lemma,
we can establish an, at least initially, better error estimate. The drawback
however is, that if our estimate exceeds the previously set bound, we can not
make assertions on kdx (t)k any longer.
Theorem 3.1 (Method 1). Let K ∗ = (2 · 77 )−1/8 . As long as
Z t
kdx (0)k2 eA(t) + 2 k Res(s)k2−1 e(A(t)−A(s)) ds ≤ K ∗ , (3.1)
0

we have
Z t
kdx (t)k2 ≤ kdx (0)k2 eA(t) + 2 k Res(s)k2−1 e(A(t)−A(s)) ds,
0
Rt
where A(t) = − 41 t + 18 0 kϕxx (τ )k2∞ dτ.

17
18 3.2. SECOND METHOD

Note, that the condition in (3.1) involves only the numerical approxima-
tion ϕ.
Proof. It follows from the inequality (2.1) that as long as kdx k8 ≤ (2 · 77 )−1
we obtain
1
∂t kdx k2 ≤ − kdx k2 + 2k Res k2−1 + 18kdx k2 kϕxx k2∞ .
4
Now we can apply Lemma 2.4 to deduce that
Z t
t
 
2 2 2
kdx (t)k ≤kdx (0)k exp − + 18 kϕxx (τ )k∞ dτ
4 0
Z t
t−s Z t
 
+ 2 k Res(s)k2−1 exp − + 18 kϕxx (τ )k2∞ dτ ds.
0 4 s

Again, please note that if the bound from this exceeds K ∗ , Theorem 3.1
makes no assertions on kdx k2 .

3.2 Second Method

This is the more sophisticated method based on direct application of Corol-
lary 2.7.
Theorem 3.2 (Method 2). As long as the right-hand side is finite, the fol-
lowing inequality holds for d(t):
 Z t 
kdx (t)k2 ≤ eA(t) kdx (0)k2 + f˜(s) ds
0
  Z t 4 Z t −1/4
× 1 − 4 kdx (0)k2 + f˜(s) ds b̃(s) ds
0 0
with
77 4A(t)
b̃(t) = e , f˜(t) = 2e−A(t) k Res(t)k2−1
2
and
t Zt
A(t) = − + 18kϕxx (s)k2∞ ds.
2 0
Again, the condition for regularity provided by the theorem depends only
on the numerical solution ϕ.
Proof. Apply Corollary 2.7 (CP-Type II) to our inequality (2.1). The corre-
sponding functions are
77 1
b(t) = , a(t) = 18kϕxx (t)k2∞ − , f (t) = 2k Res(t)k2−1 ,
2 2
which immediately give us the statement of the theorem.
CHAPTER 3. VERIFICATION METHODS
. 19

3.3 Third Method (Second Method with

Restarting)
The previous method can be further improved by introducing something
that can be best described as “restarting”. Instead of estimating over the
whole time interval [0, T ] at once, we estimate to some smaller t∗ and use the
resulting upper bound as the new initial value.

Theorem 3.3 (Method 3). Given any arbitrary partition {ti }0≤i≤n of the
interval [0, T ] with t0 = 0 and tn = T , then by Theorem 3.2 we have for all
1≤i≤n

z(0) := kdx (0)k2

 Z ti 
2
kdx (ti )k ≤ e A(ti )
z(ti−1 ) + f˜(s) ds
ti−1
  Z ti 4 Z ti −1/4
× 1 − 4 z(ti−1 ) + f˜(s) ds b̃(s) ds
ti−1 ti−1

=: z(ti )

as long as the right-hand side is finite, where for t ∈ (ti−1 , ti ]

77 4A(t)
b̃(t) = e , f˜(t) = 2e−A(t) k Res(t)k2−1
2
and
1 Z t
A(t) = − (t − ti−1 ) + 18kϕxx (s)k2∞ ds.
2 ti−1

Proof. Given some arbitrary partition {ti }0≤i≤n of the interval [0, T ] with
t0 = 0 and tn = T , we define our new method as follows.
First, we apply Theorem 3.2 to the interval [0, t1 ]. Here we start with

z(0) := kdx (0)k2

and thus obtain on the right-hand side of the interval

 Z t1 
kdx (t1 )k ≤ e2 A(t1 )
z(0) + f˜(s) ds
0
  Z t1 4 Z t1 −1/4
× 1 − 4 z(0) + f˜(s) ds b̃(s) ds
0 0
=: z(t1 ).
20 3.3. THIRD METHOD

This defines the upper bound for kdx (t1 )k2 as z(t1 ). In the next step, z(t1 ) is
taken as the new “initial value” when we apply Theorem 3.2 to the interval
[t1 , t2 ].
 Z t2 
2
kdx (t2 )k ≤ e A(t2 )
z(t1 ) + f˜(s) ds
t1
  Z t2 4 Z t2 −1/4
× (1 − 4 z(t1 ) + f˜(s) ds b̃(s) ds
t1 t1

where b̃(t), f˜(t) are defined as before, only A(t) for t ∈ (ti−1 , ti ] changes to
1 Z t
A(t) = − (t − ti−1 ) + 18kϕxx (s)k2∞ ds.
2 ti−1

This procedure is now repeated for every interval of the partition.

Lastly, we want to give an informal argument that this method converges
to a solution of the ODE as h → 0. Only informal, because a rigorous proof
would be extensive for adding little to the understanding.
Conjecture 3.4. Method 3 as given by Theorem 3.3, where the step-size h
is given by max1≤i<n (ti − ti−1 ) = h, converges for h → 0 to a solution of the
ODE (2.1) bounding kdx k.
Sketch of proof. Let z(t) be a smooth interpolation of the discrete points
z(ti ), i = 1, 2, . . . and h = tj+1 − tj . Then
z(tj+1 ) − z(tj )
∂t z(tj ) = + O(h)
h
Rt
Using tjj+1 g ds = g(tj )h + O(h2 ) and the abbreviations z(tj ) = zj , Aj =
A(tj ) and Resj = k Res(tj )k2−1 , we obtain from Theorem 3.3
1 eAj+1 (zj + hf˜j + O(h2 ))
 
∂t z(tj ) = − zj + O(h)
h (1 − 4[zj + hf˜j + O(h2 )]4 hb̃j + O(h2 ))1/4
Taylor 1 eAj+1 (zj + hf˜j )
 
= − zj + O(h)
h (1 − 4[zj + hf˜j ]4 hb̃j )1/4
77 4Aj 77
Using b̃j = 2
e = 2
and f˜j = 2e−Aj Resj = 2 Resj , as Aj = 0 yields
1 eAj+1 (zj + 2h Resj )
 
∂t z(tj ) = − zj + O(h)
h (1 − [zj + 2h Resj ]4 · 2h77 )1/4
q
 eAj+1 z + 2heAj+1 Resj −zj 4 1 − 77 2h[zj + 2h Resj ]4 
1 j
= q + O(h)
h 4
1 − 77 2h[zj + 2h Resj ]4
1
 q 
= eAj+1 zj + 2heAj+1 Resj −zj 4 1 − 77 2h[zj + 2h Resj ]4 + O(h).
h
CHAPTER 3. VERIFICATION METHODS 21

√
Now using 4
1 − x = 1 − 41 x + O(x2 ) (like before) and Aj+1 = O(h) leads to

1 1
∂t z(tj ) = 2eAj+1 Resj + (eAj+1 − 1)zj + zj 77 [zj + 2h Resj ]4 + O(h)
h 2
0 1 7 5
= 2 Resj +2A (tj )zj + 7 zj + O(h).
2

Recall that Resj = k Res(tj )k2−1 and A0 (tj ) = − 21 + 18kϕxx (tj )k2∞ , and we
recover that z solves (2.1) with equality in the limit h → 0.

3.4 Numerical Comparison

In this section, we will compare the three previously established methods
numerically. Note, that the simulations are not rigorous yet. We do not
implement interval arithmetic which would be necessary to take inevitable
rounding errors into account and further, we do not compute a strict upper
bound of the methods. For the latter see Chapter 5. However, as our aim
is to illustrate the general behavior and feasibility of the three methods, this
should be sufficient at this point. Moreover, adding interval arithmetic is
just a technical straight forward task, where we expect that the additional
error does not affect the picture that much if used with adaptive step-size.
Although our methods allow ϕ to be any arbitrary approximation, that
satisfies the boundary conditions, it should be a reasonable choice, i.e. close
to an expected solution, for the methods to be successful.
For our simulations, we calculate an approximate solution using a spectral
Galerkin scheme with N Fourier modes in space and a semi-implicit Euler
scheme with step-size h in time, yielding the values ϕ(t) for t = 0, h, 2h, ....
Further details about the numerical implementation can be found in Ap-
pendix B. To calculate the residual of ϕ, these values are interpolated piece-
wise linear in time.
There are two ways to show global regularity using the numerical methods
of the previous section:

• show that the solution exists until the time T ∗ (u0 ) (from Theorem 2.3),
since the solution is regular after this time; or

• show that kϕx (t)k + kdx (t)k < ε0 for some t > 0, since then Theorem
2.2 guarantees global regularity.

In all of our figures the maximum time is T ∗ , as defined by Theorem 2.3,

rounded to the first decimal digit +0.1 as a small padding, as long as no
22 3.4. NUMERICAL COMPARISON

0.2 0.025 0.025

0.02 0.02
0.15

0.015 0.015
0.1
0.01 0.01
threshold
0.05 method1
0.005 0.005

0 0 0
0 1 2 0 2 4 0 2 4
t t t
(a) Method 1 (b) Method 2 (c) Method 3

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 1 2 0 2 4 0 2 4
t t t
(d) Smallness Method 1 (e) Smallness Method 2 (f) Smallness Method 3

×10 -4
1.4
2.5
1.2

1 2

0.8
1.5
0.6
1
0.4
0.5
0.2

0 0
0 2 4 0 2 4
t t
(g) kϕxx k∞ (h) k Res k−1 (i) ϕ

Figure 3.1: Initial value u0 = sin(x), N = 128 Fourier modes and step-size
h = 10−5 . Methods 2 and 3 show global existence of a smooth solution as
they stay bounded until time T ∗ and even fulfill the smallness criterion before
time T ∗ . Method 1 fails as it hits its threshold at approximately t = 1.5 < T ∗
and also before the smallness criterion is reached.
CHAPTER 3. VERIFICATION METHODS 23

blowup occurred or threshold was hit. The values for T ∗ can be found in
Table 3.1.
In Figure 3.1 we have an initial value of u0 = sin(x), N = 128 Fourier
modes and a step-size of h = 10−5 . As Methods 2 and 3 stay bounded up to
time T ∗ , we would have shown global existence of a unique smooth solution
in a fully rigorous simulation. Method 1 fails because it hits its threshold at
approximately t = 1.57, which is smaller than T ∗ . In the “Smallness plots”
(Figures 3.1d, 3.1e and 3.1f) the gray area is the area around kϕx (t)k (the
solid blue line in the middle) with distance kdx (t)k, which is estimated by the
respective method. The red dashed line indicates the critical √ value ε0 = 21 of
the smallness criterion (Theorem 2.2). Note that kux (0)k = π  12 which
implies that we do not have any analytic results for this initial value. In
order to fulfill the smallness criterion, there has to be a time τ where the
upper border of the gray area is below the red dashed line.
To sum up, in this example we would have been able to show global
regularity for this initial value with Methods 2 and 3 by smallness and time
criteria, whereas both fail for Method 1. In Figure 3.2 we achieve global
existence also with Method 1 by decreasing the step-size to h = 10−6 . All
other parameters stay unchanged.
Figures 3.3, 3.4 and 3.5 have minimal alterations of the initial value with
u0 = sin(x) + 0.25 sin(2x), u0 = sin(x) + 0.225 sin(2x) and u0 = sin(x) +
0.2 sin(2x) respectively. We can see that these small alterations are sufficient
to change our results completely. In Figure 3.3 all three methods fail, in
Figure 3.4 only Method 1 fails and in Figure 3.5 all three methods succeed.
Note, that the bounds for all methods decline approximately at t ≈ 2
because ϕ is almost identical 0. In Figure 3.4, this decline is still too late
for Method 1 as it has already hit its threshold. A further small reduction
of the initial value in Figure 3.5 is now enough to keep also Method 1 alive.
Also kux (0)k  21 for these three initial values again.
Figure 3.3 is also interesting, as the numerical approximation alone does
get below the smallness threshold, but not if you take into account our error
bound (what is not too surprising if you remember that all approximations
tend to zero).
Table 3.1 is a collection of the actual values for the Figures 3.1 to 3.5.
It is worth to note, that, at least for the current dataset, if we have global
existence by one criterion, then also by the other. The smallness criterion is
also reached significantly earlier than the time criterion.
As the bounds from Methods 2 and 3 are in all of our simulations almost
identical, we use an artificial example to illustrate the difference between
these methods. In Figure 3.6 we do this by artificially setting a constant, rel-
atively large k Res k2−1 = 0.5 and an also constant, but smaller second deriva-
24 3.4. NUMERICAL COMPARISON

×10 -4 ×10 -4
2 2
0.2

0.15 1.5 1.5

threshold 1 1
0.1 method1

0.05 0.5 0.5

0 0 0
0 2 4 0 2 4 0 2 4
t t t
(a) Method 1 (b) Method 2 (c) Method 3

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 2 4 0 2 4 0 2 4
t t t
(d) Smallness Method 1 (e) Smallness Method 2 (f) Smallness Method 3

×10 -5
1.4
2.5
1.2

1 2

0.8
1.5
0.6
1
0.4
0.5
0.2

0 0
0 2 4 0 2 4
t t
(g) kϕxx k∞ (h) k Res k−1 (i) ϕ

Figure 3.2: Same setting as in Figure 3.1, but now with smaller step-size
h = 10−6 . Now Method 1 succeeds, too.
CHAPTER 3. VERIFICATION METHODS 25

0.2
0.2 0.2

0.15 threshold 0.15 0.15

method1

0.1 0.1 0.1

0.05 0.05 0.05

0 0 0
0 0.5 1 0 0.5 1 1.5 0 0.5 1 1.5
t t t
(a) Method 1 (b) Method 2 (c) Method 3

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 0.5 1 0 0.5 1 1.5 0 0.5 1 1.5
t t t
(d) Smallness Method 1 (e) Smallness Method 2 (f) Smallness Method 3

×10 -4
3
2

1.5
2

1
0.5

0 0
0 0.5 1 1.5 0 0.5 1 1.5
t t
(g) kϕxx k∞ (h) k Res k−1 (i) ϕ

Figure 3.3: Initial value u0 = sin(x) + 0.25 sin(2x), N = 128 Fourier modes
and step-size h = 10−6 . All three methods fail both, the time and smallness
criterion.
26 3.4. NUMERICAL COMPARISON

0.2 0.03 0.03

0.025 0.025
0.15
0.02 0.02

0.1 0.015 0.015

0.01 0.01
0.05
threshold
0.005 0.005
method1
0 0 0
0 0.5 1 1.5 0 2 4 0 2 4
t t t
(a) Method 1 (b) Method 2 (c) Method 3

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 0.5 1 1.5 0 2 4 0 2 4
t t t
(d) Smallness Method 1 (e) Smallness Method 2 (f) Smallness Method 3

×10 -4
2.5

1.5
2

1 1.5

1
0.5
0.5

0 0
0 2 4 0 2 4
t t
(g) kϕxx k∞ (h) k Res k−1 (i) ϕ

Figure 3.4: Initial value u0 = sin(x) + 0.225 sin(2x), N = 128 Fourier modes
and step-size h = 10−6 . Here only Method 1 fails.
CHAPTER 3. VERIFICATION METHODS 27

0.2 0.014 0.014

0.012 0.012
0.15
0.01 0.01

threshold 0.008 0.008

0.1 method1
0.006 0.006

0.004 0.004
0.05
0.002 0.002

0 0 0
0 2 4 0 2 4 0 2 4
t t t
(a) Method 1 (b) Method 2 (c) Method 3

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 2 4 0 2 4 0 2 4
t t t
(d) Smallness Method 1 (e) Smallness Method 2 (f) Smallness Method 3

×10 -4

2
1.5

1
1

0.5

0 0
0 2 4 0 2 4
t t
(g) kϕxx k∞ (h) k Res k−1 (i) ϕ

Figure 3.5: Initial value u0 = sin(x) + 0.2 sin(2x), N = 128 Fourier modes
and step-size h = 10−6 . All three methods succeed the time and smallness
criterion.
28 3.5. CONCLUSION

Smallness Time
u0 T∗ ≈ N h M1 M2 M3 M1 M2 M3
sin(x) 3.5 128 10−5 – 1.48 1.48 1.57 X X
sin(x) 3.5 128 10−6 1.46 1.19 1.19 X X X
sin(x) + 0.25 sin(2x) 3.7 128 10−6 – – – 0.98 1.41 1.41
sin(x) + 0.225 sin(2x) 3.7 128 10−6 – 1.53 1.53 1.32 X X
sin(x) + 0.2 sin(2x) 3.6 128 10−6 2.5 1.4 1.4 X X X

Table 3.1: Summary of the simulations of this chapter. All values are rounded
to fit into the table. A ”–” in the ”Smallness” columns means, that the
smallness criterion was not met by the respective method. Else, we give the
time when it was met. For the ”Time” columns this turns around. If the
time criterion was met, the respective method gets a ”X”, else the time of
the blowup / reaching the threshold is displayed.
0.2

0.15

0.1

0.05
Method2
Method3
0
0 0.05 0.1 0.15

Figure 3.6: Artificial example for fixed k Res k2−1 = 0.5 and kϕxx k2∞ = 0.1 to
compare Method 2 with Method 3.

tive kϕxx k2∞ = 0.1, without using any numerical approximation (If these
values are constant, we can calculate all integrals for the bounds exactly).
The simulation shows, that Method 3 has indeed a better handling of large
residuals. Usually, if we take a reasonable approximation ϕ (for interesting
initial values), we would have a significantly smaller k Res k2−1 compared to
kϕxx k2∞ which is the reason why we could not see a difference in the previous
simulations.

3.5 Conclusion
If we neglect that the simulations were not fully rigorous concerning the
residual and interval arithmetic, all three methods have shown that they are
capable of proving global existence of smooth solutions for the surface growth
CHAPTER 3. VERIFICATION METHODS 29

equation (1.1) for initial values which are not covered by a-priori results like
the smallness criterion (Theorem 2.2).
Also, the examples of Figure 3.3, 3.4 and 3.5 show, that all three methods
can react very sensitive to small alterations of the initial value. While this
might be obvious for Method 1 with its threshold, the exponential nature
of Methods 2 and 3 lead to a similar effect. In the end, Methods 2 and 3
look more promising than Method 1, because there were no examples for an
initial value where the results by Method 1 were better than the other two
methods (We have tested far more initial values than the few given here, but
listing all of them would have no further benefit for comparing the methods).
Further, we were able to show that Method 3 is in fact superior to Method
2 due to its better handling of the residual. This is not really visible in our
simulations because for a reasonable approximation ϕ, k Res k−1 is usually
several orders of magnitude smaller than kϕxx k∞ . Therefore, we are confident
that Method 3 and numerically improvements like more modes, smaller step-
size or even a different numerical method are sufficient to control the residual
as a reason for a blowup (Please note, that there are many other approaches
conceivable than the three methods presented, for example restarting the first
method or bounding the coefficients of the ODE and solving it directly are
valid options, too, which are not part of this work). The remaining quantity
that we have to control, and in fact the currently more influential one, is
kϕxx k∞ which we will take care of in the next chapter.
30 3.5. CONCLUSION
 CHAPTER 4

Eigenvalue Estimate

We will now improve the bound (2.1), especially with regard to kϕxx k which
enters the equation through the estimate of the (B) term. The aim is to find
a better estimate for the combined (A) and (B) terms (which is basically
the linearization of the nonlinear operator around the numerical approxi-
mation) by establishing an eigenvalue estimate that makes use of the finite
dimensional estimate and a rigorous error bound. Note, that there are many
theoretical results known (e.g. [Liu15]), but we have to find an estimate that
is also easily and accurately computable.

4.1 Setting and Theorem

Using the setting from Section 1.2, we consider the linear operators

Lϕ u = −∂x4 u − 2∂x2 (ϕx ux )

and
Aϕ u = −∂x4 u − 2∂x3 (ϕx u),
where ϕ ∈ H4 in space and H 1 in time is some arbitrary but fixed function
like before.
We are interested in bounding the quadratic form

λ = sup h∂x Lϕ u, ∂x ui (4.1)

kux k=1

31
32 4.1. SETTING AND THEOREM

in order to finally obtain a bound

h∂x Lϕ u, ∂x ui ≤ λkux k2 .

By substituting v = ux we immediately get

λ = sup hAϕ v, vi .
kvk=1

Define Hn as the 2n-dimensional subspace spanned by eix , . . . , einx and its

complex conjugates e−ix , . . . , e−inx (we can omit the constant mode due to
our solution space H). Denote by Pn the orthogonal projection onto Hn .
Finally, we set
λn := sup hPn Aϕ Pn u, ui, (4.2)
kuk=1

which is just the largest eigenvalue of the symmetric 2n×2n matrix 21 (Pn Aϕ +
A∗ Pn ) (see Chapter B.2).
Obviously, as the supremum is over a larger set, it holds that

λn ≤ λ.

Now we want to bound λ from above.

For our first estimate (2.1) we used integration by parts, interpolation,
and Poincare inequalities to obtain the following “worst case” estimate
Z
2
hAϕ u, ui = −kuxx k + 2 ϕx uuxxx dx
Z Z
= −kuxx k2 − 2 ϕxx uuxx dx − 2 ϕx ux uxx dx
Z Z
= −kuxx k2 − 2 ϕxx uuxx dx + ϕxx u2x dx
≤ −kuxx k2 + 2kϕxx k∞ kukkuxx k + kϕxx k∞ kux k2
≤ −kuxx k2 + 3kϕxx k∞ kukkuxx k
≤ − 12 kuxx k2 + 92 kϕxx k2∞ kuk2
≤ − 12 kuk2 + 92 kϕxx k2∞ kuk2 .

Thus
1 9
λ ≤ − + kϕxx k2∞ . (4.3)
2 2
Instead, the following theorem shows an improved estimate by analyzing the
quadratic form (4.1) separately for different mode ranges.
CHAPTER 4. EIGENVALUE ESTIMATE 33

Theorem 4.1. Let u be a solution to our surface growth equation (1.1),

ϕ ∈ Hn an arbitrary approximation and Hn , λ and λn be defined as in Section
4.1. Then, for
√ √
n≥ 2Cϕ = 2(2kϕxxx k∞ + 6kϕxx k∞ + 4kϕx k∞ )

it holds that

1 9kϕxx k2∞ − 2λn 1 4

 
λn ≤ λ ≤ λn + max 2Cϕ2 , 9kϕ xx k2
∞ + 2λ n − n .
2 n2 2

4.2 Proof of the Theorem

As a preparation, we split u = p + q, where p ∈ Hn and q ⊥ Hn . Thus

λ = sup hAϕ u, ui
kuk=1
 
= sup hAϕ p, pi + hAϕ p, qi + hAϕ q, pi + hAϕ q, qi .
kpk2 +kqk2 =1

Now, we will treat these scalar products separately, where we will denote
with “low modes” the parts only depending on p and with “high modes”
everything solely depending on q.

Low modes
First, notice that analogous to the brute force estimate

1 9
hAϕ p, pi ≤ − kpxx k2 + kϕxx k2∞ kpk2 .
2 2

Second, it holds by definition of λn , as p ∈ Hn

hAϕ p, pi ≤ λn kpk2 .

Thus in summary, we get

1 9
hAϕ p, pi ≤ (1 − ηn )λn kpk2 − ηn kpxx k2 + ηn kϕxx k2∞ kpk2
2 2

for some ηn ∈ [0, 1], that we will fix later.

34 4.2. PROOF OF THE THEOREM

Mixed terms
For the mixed terms we use
1 1
kpk ≤ kpx k ≤ kpxx k and kqk ≤ kqx k ≤ 2 kqxx k
n n
to obtain (any derivatives of p and q are still orthogonal in H)
Z
hAϕ p, qi = −2 (ϕx p)xxx q dx
Z
= 2 (ϕx p)xx qx dx
Z
= 2 (ϕxxx p + 2ϕxx px + ϕx pxx )qx dx
 
≤ 2kqx k · kϕxxx k∞ kpk + 2kϕxx k∞ kpx k + kϕx k∞ kpxx k
1
≤ Cϕ(1) kqxx kkpxx k
n

with Cϕ(1) = 2kϕxxx k∞ + 4kϕxx k∞ + 2kϕx k∞ ] and

Z
hAϕ q, pi = 2 (ϕx q)pxxx dx
Z
= −2 (ϕx q)x pxx dx
 
≤ 2kpxx k · kϕxx k∞ kqk + kϕx k∞ kqx k
1
≤ Cϕ(2) kqxx kkpxx k
n

with Cϕ(2) = 2kϕxx k∞ + 2kϕx k∞ . Further, we define

Cϕ = Cϕ(1) + Cϕ(2) = 2kϕxxx k∞ + 6kϕxx k∞ + 4kϕx k∞ .

High modes
Finally, for the high modes we use the rough “worst case” estimate (4.3),

1 9
hAϕ q, qi ≤ − kqxx k2 + kϕxx k2∞ kqk2 ,
2 2
but we will later apply the improved Poincare inequality

1
kqk ≤ kqx k ∀q ⊥ Hn . (4.4)
n
CHAPTER 4. EIGENVALUE ESTIMATE 35

Summary
Combining all estimates, we obtain (using ab ≤ 21 a2 + 12 b2 and eliminating
pxx terms)
hAϕ u, ui = hAϕ p, pi + hAϕ p, qi + hAϕ q, pi + hAϕ q, qi
1 9
≤ (1 − ηn )λn kpk2 − ηn kpxx k2 + ηn kϕxx k2∞ kpk2
2 2
1
+ Cϕ kqxx kkpxx k
n
1 9
− kqxx k2 + kϕxx k2∞ kqk2
2 2
9
≤ (1 − ηn )λn kpk2 + ηn kϕxx k2∞ kpk2
2
1 Cϕ2 9
 
+ 2
− 1 kqxx k2 + kϕxx k2∞ kqk2
2 n ηn 2
In order to apply the improved Poincare inequality (4.4) for q, we define
Cϕ2 √
ηn := 2 2
and thus n ≥ 2Cϕ to assert ηn ≤ 1
n
and obtain
9
 
hAϕ u, ui ≤ (1 − ηn )λn + ηn kϕxx k2∞ kpk2
2 
1 1

+ 9kϕxx k2∞ − n4 kqk2
2 2
which proves our main theorem
λ = sup hAϕ u, ui = sup hAϕ u, ui
kuk=1 kpk2 +kqk2 =1
9 1 1
   
≤ max (1 − ηn )λn + ηn kϕxx k2∞ , 9kϕxx k2∞ − n4
2 2 2
1 1
 
= λn + max ηn [9kϕxx k2∞ − 2λn ] , 9kϕxx k2∞ + 2λn − n4 .
2 2

4.3 Comparison with the Previous Estimate

Now, we want to give a rough comparison between the new and the old
estimate of the quadratic form. Recall the previous worst case estimate
(4.3),
3
λ ≤ − + 9kϕxx k2∞ ,
4
36 4.4. APPLICATION TO THE SURFACE GROWTH EQUATION

that we used for Methods 1 - 3 (and will use for Method 4 in the following
chapter). We can rewrite Theorem 4.1 in the following way:
1 9kϕxx k2∞ − 2λn 1
 
λ ≤ λn + max 2Cϕ2 2
, 9kϕxx k2∞ + 2λn − n4
2 n 2

9Cϕ2 kϕxx k2∞  2Cϕ2  9 2 1 4

= max + 1 − 2 λn , kϕxx k∞ + 2λn − n
n2 n 2 4
2 2
9

9C ϕ 9
 
2C ϕ 1 4

= kϕxx k2∞ + max − kϕ k
xx ∞
2
+ 1 − λ n , 2λ n − n
2 n2 2 n2 4
Cϕ2
1
and as n2 ≤ 2

9 1
 
≤ kϕxx k2∞ + max 0 , 2λn − n4 .
2 4
Thus, as long as the worst case estimate is finite, our new estimate is at least
less than roughly a half of it, as the max tends to zero with increasing n (as
λn is bounded by λ which is in itself bounded by the worst case estimate that
is independent of n) and − 34 is negligible for all interesting cases. In fact,
most of the time we have far better results than just a factor of 12 because
Cϕ2

n2  21 .

4.4 Application to the Surface Growth

Equation
Let us denote the eigenvalue bound from Theorem 4.1 with λ̃. If we want
to incorporate this result into our framework, we have to consider, that in
order to control the (C) and (D) terms, we need some part of the (A) term
of
1
∂t kdx k2 = hdxx , dxxxx + 2(dx ϕx )xx i + hdxx , (dx 2 )xx + Resi .
2 | {z } | {z }
A+B C+D

Therefore we split the first term into two parts

1
∂t kdx k2 = (1 − δ)hdxx , dxxxx + 2(dx ϕx )xx i + δhdxx , dxxxx + 2(dx ϕx )xx i
2
+ hdxx , (dx 2 )xx + Resi.
Now we can bound the first term with our new method and the remaining
parts like before in (2.1). If we do not fix the constants used in the Young
inequalities, we have
A = −kdxxx k2
CHAPTER 4. EIGENVALUE ESTIMATE 37

9
|B| ≤ εB kdxxx k2 + kdx k2 kϕxx k2∞
4εB
( 4 εC )−7
|C| ≤ εC kdxxx k2 + 7 kdx k10
4
1
|D| ≤ εD kdxxx k2 + k Res k2−1 ,
4εD
where we can set all ε{B,C,D} > 0 arbitrary small.
In this case, our differential inequality is

1 9 ( 4 εC )−7
∂t kdx k2 ≤ (1 − δ)λ̃kdx k2 + δkdx k2 kϕxx k2∞ + 7 kdx k10
2 4εB 4
1  
+ k Res k2−1 + δεB + εC + εD − δ kdxxx k2 ,
4εD
where ε{B,C,D} > 0 and δ ∈ (0, 1). By substituting ε{C,D} := δε{C,D} , this is
equivalent to

1 9 ( 4 δεC )−7
∂t kdx k2 ≤ (1 − δ)λ̃kdx k2 + δkdx k2 kϕxx k2∞ + 7 kdx k10
2 4εB 4
1  
+ k Res k2−1 + δ εB + εC + εD − 1 kdxxx k2 ,
4δεD
where ε{B,C,D} > 0 and δ ∈ (0, 1). Next, we set εB + εC + εD = 1 to remove
the last term, and therefore, we are now interested in the minimum of

1 9δ 77
∂t kdx k2 ≤ (1 − δ)λ̃kdx k2 + kdx k2 kϕxx k2∞ + 8 kdx k10
2 4εB 4 (δεC )7
(4.5)
1
+ k Res k2−1
4δεD
under the constraints ε{B,C,D} > 0, k∈{B,C,D} εk = 1, δ ∈ (0, 1).
P

Unfortunately, it is not so easy to determine a rigorous global minimum

for this problem. We first rewrite the function and condense all constants
x1
f (x) = (1 − x1 )c1 + c2 + (x1 x3 )−7 c3 + (x1 x4 )−1 c4
x2
with c2,3,4 ≥ 0 and the constraints of

0 < x1 < 1
x2,3,4 > 0
x2 + x3 + x4 = 1.
38 4.4. APPLICATION TO THE SURFACE GROWTH EQUATION

Now, we look at the function independent of x1 , by moving x1 to the constants

and proceed to solve the minimization only for x2 , x3 , x4

f (x) = c˜1 + x−1 −7 −1

2 c˜2 + x3 c˜3 + x4 c˜4
1 = x2 + x3 + x4
0 < x2 , x3 , x4 .

Lagrange multipliers now yield

−x−2
2 c˜2 λ
   
 −x−8 7c˜  λ
3 3  !  
 =  ,

 −x−2 c˜ λ

4 4 
x 2 + x3 + x 4 1

and thus as xk > 0 and c̃2,3,4 > 0, and therefore λ < 0, we have
q q q
x2 = − cλ˜2 , x3 = 8
− 7λc˜3 , x4 = − cλ˜4 ,
q q q
⇒1= − cλ˜2 + − cλ˜4 + 8 − 7λc˜3
√ √ √
= √1 ( c˜2 + c˜4 ) + √ 8
1 8
7c˜3 .
|λ| |λ|

This is an equation of type

ct4 + t − 1 = 0,

which we could solve by Ferrari’s method for quartic equations (see e.g.
[Neu65] or any other basic algebra book). Given the length of Ferrari’s
formula and the fact that we further had to take into account that now
all constants depend on x1 (which leads to another increase in complexity
because of the derivative), we will not pursuit this ansatz as its cost-benefit
ratio is not good enough. Luckily, we can not do anything wrong here that
breaks the rigorosity of our calculations, as valid parameter combinations
just might not be optimal. Therefore, we will just use MATLAB’s nonlinear
optimization solver to find a approximate local minimum and update it after
a given time interval (we could do this in every step, but given that the
step-size is small enough and the data is continuous, this is not necessary
and would just cost us lots of computational time).
 CHAPTER 5

Numerical Preparations

From our numerical algorithm we obtain discrete values of ϕ for times 0 =

t0 , t1 , . . . , tn . In order to calculate quantities like A(t), b̃(t), and f˜(t), which
occur in our bounds that we derived in Chapters 2 and 4, we have to define
ϕ in between these grid points. Again we have to make a trade off between
accuracy and feasibility, as there are multiple interpolation methods appli-
cable.
In our case, we will set ϕ as the piecewise linear interpolation of our
discrete values

ϕ(x, t0 + t) = (1 − ht )ϕ(x, t0 ) + ht ϕ(x, t1 )

= ht (ϕ(x, t1 ) − ϕ(x, t0 )) + ϕ(x, t0 ) t ∈ [0, h], h = t1 − t0 .

Following this ansatz, we have to derive computable quantities like k Res k−1
and A(t) for every time t ∈ [t0 , t1 ], or at least obtain upper bounds, using
only known values at time t0 or t1 .

5.1 Norm of the Residual

For the residual, we get by using the definition and applying the linear in-
terpolation

Res ϕ(x, t0 + t) = ∂t ϕ(x, t0 + t) + ∂x4 ϕ(x, t0 + t) + ∂x2 (∂x ϕ(x, t0 + t))2

= h1 (ϕ(x, t1 ) − ϕ(x, t0 ))

39
40 5.1. NORM OF THE RESIDUAL

+ ht ∂x4 (ϕ(x, t1 ) − ϕ(x, t0 )) + ∂x4 ϕ(x, t0 )

+ ∂x2 (∂x ϕ(x, t0 ))2
+ 2 ht ∂x2 (∂x (ϕ(x, t1 ) − ϕ(x, t0 )) · ∂x ϕ(x, t0 ))
2
+ ht 2 ∂x2 (∂x (ϕ(x, t1 ) − ϕ(x, t0 )))2
= h1 (ϕ(x, t1 ) − ϕ(x, t0 )) + ∂x4 ϕ(x, t0 ) + ∂x2 (∂x ϕ(x, t0 ))2
+ ht [∂x4 (ϕ(x, t1 ) − ϕ(x, t0 ))
+ 2∂x2 (∂x (ϕ(x, t1 ) − ϕ(x, t0 )) · ∂x ϕ(x, t0 ))]
t2 2
+ ∂ (∂ (ϕ(x, t1 )
h2 x x
− ϕ(x, t0 )))2 .

And therefore, we have for the spatial primitive

Z
Res ϕ(x, t0 + t) dx = h1 (Φ(x, t1 ) − Φ(x, t0 )) + ∂x3 ϕ(x, t0 ) + ∂x (∂x ϕ(x, t0 ))2
+ t h1 [∂x3 (ϕ(x, t1 ) − ϕ(x, t0 ))
+ 2∂x (∂x (ϕ(x, t1 ) − ϕ(x, t0 )) · ∂x ϕ(x, t0 ))]
+ t2 h12 ∂x (∂x (ϕ(x, t1 ) − ϕ(x, t0 )))2
=: a(x) + tb(x) + t2 c(x),

where we use the following abbreviations

a(x) = h1 (Φ(x, t1 ) − Φ(x, t0 )) + ∂x3 ϕ(x, t0 ) + ∂x (∂x ϕ(x, t0 ))2 ,

b(x) = h1 [∂x3 (ϕ(x, t1 ) − ϕ(x, t0 )) + 2∂x (∂x (ϕ(x, t1 ) − ϕ(x, t0 )) · ∂x ϕ(x, t0 ))],
c(x) = h12 ∂x (∂x (ϕ(x, t1 ) − ϕ(x, t0 )))2

and
Z
Φ(x) = ϕ(x) dx.

The square is now given by a simple multiplication

Z 2
Res ϕ(x, t0 + t) dx = a(x)2
+ t2a(x)b(x)
+ t2 (b(x)2 + 2a(x)c(x))
+ t3 2b(x)c(x)
+ t4 c(x)2 .
CHAPTER 5. NUMERICAL PREPARATIONS 41

Finally, we have
Z 2π
k Res ϕ(x, t0 + t)k2−1 = a(x)2 dx
0
Z 2π
+ 2t a(x)b(x) dx
0
Z 2π
+ t2 b(x)2 + 2a(x)c(x) dx (5.1)
0
Z 2π
+ 2t3 b(x)c(x) dx
0
Z 2π
+ t4 c(x)2 dx
0

or in a shorter form that we will use later

= C0 + tC1 + t2 C2 + t3 C3 + t4 C4 .

Accordingly, we have
Z h Z 2π
k Res ϕ(x, t0 + t)k2−1 dt = h a(x)2 dx
0 0
Z 2π
+ h2 a(x)b(x) dx
0
Z 2π
+ 31 h3 b(x)2 + 2a(x)c(x) dx (5.2)
0
Z 2π
+ 21 h4 b(x)c(x) dx
0
Z 2π
1 5
+ 5
h c(x)2 dx,
0

what we will also use later.

Note, that we can calculate the integrals of the right-hand side for both
equations (5.1) and (5.2) using the L2 -scalar product accurately and effi-
ciently as our simulation uses Fourier transformations.

5.2 Rigorous Computable Bound for

Method 3
Now we want to derive a computable upper bound for Method 3 as stated in
Theorem 3.3. First, we define
 Z h 
T := eA(h)
z(0) + f˜(s) ds
0
42 5.2. RIGOROUS COMPUTABLE BOUND FOR METHOD 3

Z h
= eA(h) z(0) + 2 eA(h)−A(s) k Res k2−1 ds.
0

Therefore, our bound for Method 3 on one step of size h is given by

77 Z h 4(A(s)−A(h))
 −1/4
4
T × 1 − 4T × e ds . (5.3)
2 0
To get an upper bound for (5.3), we need Ran upper bound on T itself, and
an upper bound for the exponential term 0h e4(A(s)−A(h)) ds. Therefore, we
need an upper bound for A(t), where we have
1 Z t
A(t) = − t + 18 kϕxx (t0 + s)k2L∞ ds
2 0
1 Z t
≤ − t + 18 ((1 − hs )kϕxx (t0 )kL∞ + hs kϕxx (t1 )kL∞ )2 ds
2 0
1 t2 t3

= − t + 18 (t − + 2 )kϕxx (t0 )k2∞
2 h 3h
 2 3 
t t
+2 − 2 kϕxx (t0 )k∞ kϕxx (t1 )k∞ (5.4)
2h 3h
t3

+ 2 kϕxx (t1 )k2∞ ,
3h
that we can use for
1 Z h
A(h) − A(s) = (s − h) + 18 kϕxx k2∞ dτ
2 s
1
≤ (s − h)
2
h2 + s2 h3 − s3 

+ 18 h − s − + kϕxx (t0 )k2∞
h 3h2
 2 (5.5)
h − s2 h3 + s3

+2 − kϕxx (t0 )k∞ kϕxx (t1 )k∞
2h 3h2
h3 − s3

2
+ kϕxx (t1 )k∞ .
3h2
Further, the upper bound on T consists of two parts, an upper bound
for eA(h) z(0), what we basically handled before, and an upper bound for
R h A(h)−A(s)
0 e k Res k2−1 ds. The first part can be treated by using (5.4),

1 h
 
A(h)
e z(0) ≤ exp − h + 18 kϕxx (t0 )k2∞
2 3 (5.6)
h h

2
+ kϕxx (t0 )k∞ kϕxx (t1 )k∞ + kϕxx (t1 )k∞ z(0)
3 3
Random documents with unrelated
content Scribd suggests to you:
The Project Gutenberg eBook of Practical
Phrenology Simplified
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.

Title: Practical Phrenology Simplified

Author: Theodore Foster

Release date: April 5, 2021 [eBook #64998]

Most recently updated: October 18, 2024

Language: English

Credits: Mike Stember and the Online Distributed Proofreading Team

at https://www.pgdp.net (This file was produced from
images generously made available by The Internet Archive)

*** START OF THE PROJECT GUTENBERG EBOOK PRACTICAL

PHRENOLOGY SIMPLIFIED ***
 PHRENOLOGICAL ANALYSIS
OF THE
CHARACTER
OF

By

Given
 NOTICE.
It is recommended to use in the annexed table the numerals, from
one to eight, commencing in the column headed Very Small. It will
then exhibit the relative developements of the organs in the head of
the individual examined.
 Predom- Very Rather Very
Large. Full. Moderate. Small.
inant Large. Small. Small.
Domestic Propensities.
Amativeness,
Philopro-
genitiveness,
Adhesiveness,
Inhabitiveness,
—
Concentrativeness,

Selfish Propensities.
Combativeness,
Destructiveness,
Secretiveness,
Alimentiveness,
Acquisitiveness,

Selfish Sentiments.
Firmness,
Self-esteem,
Love of Approbation,
Cautiousness,

Moral Sentiments.
Conscientiousness,
Veneration,
Benevolence,
Hope,
Marvellousness,

Intellectual Sentiments.
Identity,
Constructiveness,
Imitation,
Mirthfulness,

Perceptive Faculties.
Individuality,
Form,
Size,
Weight,
Colour,
Order,
Calculation,
Locality,
Eventuality,
Time,
Tune,
Language,

Reflective Faculties.
Causality,
Comparison,

Temperaments.
Lymphatic,
Sanguine,
Bilious,
Nervous,
 PRACTICAL

PHRENOLOGY
SIMPLIFIED.
BY
THEODORE FOSTER.
PHILADELPHIA:
ORRIN ROGERS, 67 SOUTH SECOND STREET.
1838.
Entered according to Act of Congress, A. D. 1838, by
Theodore Foster, in the Clerk’s Office of the District
Court for the Eastern District of Pennsylvania.
E. G. DORSEY, PRINTER,
LIBRARY STREET.
 CONTENTS.
Domestic Propensities, 1
Amativeness, ib.
Philoprogenitiveness, 5
Adhesiveness, 8
Inhabitiveness, 11
Concentrativeness, 13

Selfish Propensities, 16
Combativeness, ib.
Destructiveness, 19
Secretiveness, 22
Acquisitiveness, 26
Alimentiveness, 30

Selfish Sentiments, 34
Firmness, ib.
Self-esteem, 37
Love of Approbation, 43
Cautiousness, 46

Moral Sentiments, 49
Conscientiousness, ib.
Veneration, 53
Benevolence, 56
Hope, 59
Marvellousness, 62

Intellectual Sentiments, 65
Ideality, ib.
Constructiveness, 68
Imitation, 69
Mirthfulness, 72

Observing Faculties, 75
Individuality, ib.
Form, 78
Size, 80
Weight, 81
Colour, 83
Order, 85
Calculation, 87
Locality, 90
Eventuality, 92
Time, 95
Tune, 96
Language, 98

Reflective Faculties, 101

Causality, ib.
Comparison, 103

Temperaments, 107
 PREFACE.
The present volume is designed to exhibit the subject of Practical
Phrenology in as clear and as perspicuous a light as its nature will
admit. To this purpose the author has aimed to divest it of all
extraneous matter, and at the same time to avoid all unnecessary
conciseness. The learner will here find a comprehensive view of the
functions of each organ, with their different effects on the character
when in various stages of developement, and also when
compounded with each other.
The author presents few claims to originality. In a few instances he
has even adopted the language of others where it presented itself in
a felicitous manner—his aim being to make a good book rather than
to add to his own reputation.
It is but proper here to state that the work has passed through the
press without the benefit of the author’s personal inspection—an
affection of the eyes rendering this service impossible. But for this it
might have received many retouches, which, if they did not add
materially to its value, might have improved its appearance.
 DOMESTIC PROPENSITIES.
1. AMATIVENESS.
This organ produces the sexual passion, and imparts to its possessor
a desire for the happiness of the opposite sex. In society it does
much to promote general kindliness of feeling, and urbanity of
manners.
Predominant.—One in whom this organ predominates, will incline to
be libidinous, licentious and lustful. If his moral organs are very
large, particularly Firmness and Conscientiousness, he may restrain
the outward expression of this feeling; but it will, nevertheless, be
powerful, and at times overwhelming. If long deprived of the society
of the other sex, he will feel lonesome and disconsolate.
Large.—With large Amativeness and Adhesiveness, an individual will
be exceedingly attached to the society of the other sex; and will be
capable of readily ingratiating himself into their favour. If with these
organs large, and small Firmness and Conscientiousness, although
his love will be intense and fervid, yet he will be apt to be capricious
and fickle in his attachments. He will be inclined rather to seek the
favour of the sex generally, than to limit his regard to a single
object. If Adhesiveness, Inhabitiveness and Philoprogenitiveness are
large, he will be induced to marry early, but if Adhesiveness,
Philoprogenitiveness, and Conscientiousness are small, he will be
inclined to gratify this feeling without reference to the laws of
morality. With Self-esteem, Firmness, and Secretiveness large,
although he may love intensely, yet he will not allow his passion to
predominate over him; if disappointed, he will not be subdued, but
manifest to the spectator the appearance of unconcern. With such a
combination, he will, in all cases, feel much more intensely than his
expressions will imply. If Mirthfulness is large, and Conscientiousness
and Ideality small, he will be liable to joke, and be fond of licentious
allusions.
Moderate.—With a moderate developement of this organ, an
individual will take great pleasure in the society of ladies, whose
taste and feelings coincide with his own. If his Moral Sentiments and
Intellect are large, he will be averse to the society of the merely
volatile and frivolous. If Ideality and Love of Approbation are large,
he will be attracted by the company of the gay and fashionable. If
Ideality and Intellect generally, are large, he will be disgusted with
vulgarity and libidinous allusions. His passion will be deep, but not
lasting, and with a moderate amount of controlling organs he can
restrain it at will.
Small.—When this organ is small, an individual will be distant and
reserved towards females. If Adhesiveness is large, he may be
attached to the society of a select few; but the connexion will be of
a strictly Platonic character. He will be unable to feel the peculiar
pleasures of female society. If Adhesiveness and
Philoprogenitiveness are large, he may be disposed to a matrimonial
alliance; but if these organs are small, he will be decidedly averse to
such a connexion. If one with Moral Sentiments moderate, and
Destructiveness and Self-esteem large, under the influence of the
aforementioned combination, were to marry, the connexion would be
necessarily an unfortunate one; his attachment could not outlive the
vicissitudes attending the marriage state, and would inevitably
degenerate into disdain and aversion. He would, notwithstanding, be
a fond parent, though his affection would be capricious and ill
regulated. With Adhesiveness, Conscientiousness, Veneration, and
Benevolence large, an individual’s regard for the wife of his choice, if
fortunate in his selection, will increase with time; the strength of his
superior sentiments more than supplying the want of animal
passion.

2. PHILOPROGENITIVENESS.
The legitimate office of this organ is to produce love for one’s own
offspring. It produces, however, in the breast of its possessor an
affection for children indiscriminately; for the feeble and helpless; for
pets—as dogs, horses, cats, &c., and even for inanimate objects. It
has an influence in producing general kindliness of disposition. A
peculiarity of its character consists in its inspiring its possessor to
love with the fondest affection the child that is the most helpless,
and even the one that has caused the greatest solicitude and
brought down on its parent the deepest disgrace.
Predominant.—An individual in whom this organ predominates has a
constant hankering for the society of children. If without them
himself, he views the deprivation as a great misfortune, and if his
circumstances are favourable, will be likely to adopt one, for the
purpose of exhausting the energy of this feeling upon it. He will be
likewise much attached to pets, as horses and dogs.
Large.—Those who possess this organ large, betray it in every look
and motion when in company with children. They take the greatest
delight in their society, and enter into their little troubles and
enjoyments with the greatest zeal. They readily enlist their
confidence, and can easily control them. If deprived of their society,
they will exhaust their attachment upon some pet animal which they
will frequently fondle. When Combativeness, Destructiveness and
Philoprogenitiveness are large, an individual will punish children
severely when they annoy him, notwithstanding his great affection
for them. If Self-esteem and Combativeness are small, he will be
liable to humour his children and allow them improper indulgences.
With Combativeness and Destructiveness large, he will be apt to be
capricious in his feelings towards children, at one time humoring
them, and at another petulant and cross.
Moderate.—With Philoprogenitiveness moderate, a person will be
attached to his own children to a good degree, and may take some
interest in others after they begin to lose their infantile character.
This feeling, however, will not be durable. He will readily tire of
children when they annoy him. The death of a child will be a
poignant affliction to him, but it will be soon forgotten. If
Destructiveness and Self-esteem are large, he will be liable to punish
them with severity for trivial offences.
Small.—With this organ small, a person will be exceedingly annoyed
by children. If a parent, he will consign the care of them to menials.
In all his intercourse with company, he will betray a marked
indifference to their society. If ever induced to amuse them, his
awkwardness will betray itself to the most casual observer. If
Benevolence is large, he will take all needful care of them; but if
Secretiveness and Destructiveness are large, he will delight to
torment and teaze them.

3. ADHESIVENESS.
This organ furnishes the instinct of social attachment. Towards the
object of its regard it excites the purest feelings of affection. It is not
satisfied with loving, it must also be loved, and requires for its
healthy existence a constant exchange of pure and radiant affection.
It diffuses its influence over the whole character of the man, and
tends to render him kind, amiable, and affectionate. It leads to the
love of company, and of social intercourse. While it is the germ of
many virtues, it is to be feared; in the present state of society, it is
likewise productive of many vices.
Predominant.—With Adhesiveness predominant, an individual is pre-
eminently qualified to enjoy friendship, and will be miserable without
it. He will often feel the yearnings of affection coming over him with
all the intensity of a passion. His most vivid enjoyments are
experienced in the society of his friends. He readily recognises the
existence of a similar feeling in another, and, if circumstances are
favourable, they soon become intimate.
Large.—One having Adhesiveness large, is eminently social and
affectionate. With large Moral organs, will make great sacrifices to
render his friends service, and will esteem the pleasures of
friendship as one of the chief sources of enjoyment. With
Combativeness and Destructiveness large, and Self-esteem
moderate, will resent an aggression upon a friend which he would
not notice upon himself. If Self-esteem is large, with Combativeness
and Destructiveness large, he will easily get angry with his friends,
but will be readily conciliated. With Benevolence and Love of
Approbation large, is exceedingly liberal and forward among friends;
will do his utmost to please and gratify them; earnestly desire their
approbation; and will be exceedingly sensitive to their reproaches.
With Firmness and Conscientiousness small, will be capricious in his
attachments. With Secretiveness and Self-esteem large, he will not
fully express the feelings which he experiences, and will thus leave
the impression that his affection is less than it really is.
Moderate.—One having Adhesiveness moderate, may be strongly
attached to friends, but his friendships will be readily severed. He
may be companionable, and with large Benevolence will be generous
and good-hearted, but he will still lack that strong feeling of
sympathy without which friendship is but a name.
Small.—With Adhesiveness small, an individual will be unsocial, cold-
hearted and selfish. If his moral organs predominate over self-
esteem, he may be companionable, but he will be nearly wanting in
the attributes of character ascribed to this organ.

4. INHABITIVENESS.
This organ produces home-sickness, and causes a feeling of regret
to take possession of the mind when leaving a place in which one
has long resided. It is the first element of patriotism. It produces a
desire to locate and reside in a particular place, and adds much to
the strength of family attachments.
Predominant.—One in whom Inhabitiveness predominates, is pre-
eminently attached to any place with which he has become familiar.
It causes him much pain to leave it, and he returns to it with
eagerness.
Large.—One having inhabitiveness large, will experience the most
poignant sensations of regret at leaving a place with which he has
become familiar. Even a particular house, garden, office or room, has
for him peculiar gratifications. With large Locality, will take delight in
travelling, but will be constantly harassed by thoughts of home. This
is more especially the case if Concentrativeness is large. If Self-
esteem and Veneration are large, he will be eminently patriotic, and
will defend his country from aspersions with as much vigour as
himself. Veneration being large, he will experience the profoundest
feelings of respect and regard for the memory of the departed
worthies of its history; and with large Individuality, Eventuality, &c.,
he will take great delight in reading the history of his own country,
and of conversing upon its character and institutions. If long absent
from home, he is constantly curious, and eagerly seeks every means
of being informed concerning it. The peculiarities of the different
places in which he has resided often occur to him with feelings of
the most vivid pleasure.
Moderate.—One in whom Inhabitiveness is moderate, will not change
his residence without regret, yet soon becomes reconciled to a new
location. If long absent from his country, Self-esteem being small, he
will become expatriated in feeling, and identify himself wholly with
the country in which he resides.
Small.—When Inhabitiveness is small, the individual will be
constantly prompted to change his place of residence. Unless this
feeling is counteracted by the strength of other organs, he cannot
get familiarized with a place without becoming dissatisfied and
restless.

5. CONCENTRATIVENESS.
This organ imparts the power of continuity of thought. It also aids in
enabling its possessor to continue the action of the organs generally.
Predominant.—One having Concentrativeness predominant, with
Causality large, will be much subject to absence of mind. He will be
quite unable to attend to more than one thing at a time, and will be
generally prolix in conversation.
Large.—With large Concentrativeness, an individual will be much
disturbed if more than one thing claim attention at once; has a
strong inclination after taking up a subject to pursue it till he has
completed it. In conversation he will be much distracted if it is
desultory in its character. If a writer, his compositions will exhibit a
sustained unity of expression throughout.
Moderate.—One with Concentrativeness moderate, is inclined to
pursue a subject or train of thought, but can be easily diverted from
it. If Causality and Intellect generally are moderate, is neither
inclined to pursue a study to its termination, nor is he able to pass
rapidly to another. With nervous temperament he will possess great
versatility of attention.
Small.—With Concentrativeness small, an individual will be quite
unable to devote his attention for any length of time to a single
study or subject. In ordinary conversation, he will fly from one
subject to another, without order or arrangement. His friends, even if
strongly attached to them, will not be long thought of at a time. His
antipathies will be readily assuaged. He will possess great vivacity of
disposition.
 SELFISH PROPENSITIES.
1. COMBATIVENESS.
This organ gives the desire to oppose, resist and overcome. It
renders its possessor able to encounter difficulties, and to be bold
and strenuous in his opposition. If not properly regulated, it leads to
a desire to contradict and quarrel for the sake of opposition. It gives
vigour and zeal to the pugilist and warrior.
Predominant.—When this organ predominates, the individual will be
bold, disputatious and quarrelsome. In an encounter he will never be
satisfied till he has obtained the mastery. He will display great nerve
and determination in whatever he undertakes. With Self-esteem
large, and Conscientiousness and Benevolence small, he will be
extremely quarrelsome and overbearing.
Large.—With Self-esteem large, the individual will be pre-eminently
bold and enterprising. If Firmness is small, he will be wavering in his
views; but if Firmness is large, he will add perseverance to courage,
and never give up a point while a reasonable hope of success
remains. If with this combination, and Moral Sentiments small, he
will be litigious and quarrelsome. If Destructiveness is small, he will
be fond of disputing, but will avoid giving pain. If Destructiveness is
large, and Benevolence small, he will be vindictive and cruel, and will
ever be disposed to vindicate his own importance, regardless of
circumstances or the rights of others. If Love of Approbation,
Benevolence, Veneration and Conscientiousness are large, he will
avoid all low contentions, and will direct the action of this organ to
the maintenance of right and the enforcement of just opinion.
Moderate.—One with moderate Combativeness, will forbear in a
contention as long as Self-esteem will allow. If his temperament is
active, he may be irritable and passionate, but upon the whole, will
Welcome to Our Bookstore - The Ultimate Destination for Book Lovers
Are you passionate about books and eager to explore new worlds of
knowledge? At our website, we offer a vast collection of books that
cater to every interest and age group. From classic literature to
specialized publications, self-help books, and children’s stories, we
have it all! Each book is a gateway to new adventures, helping you
expand your knowledge and nourish your soul
Experience Convenient and Enjoyable Book Shopping Our website is more
than just an online bookstore—it’s a bridge connecting readers to the
timeless values of culture and wisdom. With a sleek and user-friendly
interface and a smart search system, you can find your favorite books
quickly and easily. Enjoy special promotions, fast home delivery, and
a seamless shopping experience that saves you time and enhances your
love for reading.
Let us accompany you on the journey of exploring knowledge and
personal growth!

ebookgate.com