Sie sind auf Seite 1von 273

Applied Partial Differential Equations

MAT1508/APM446 ∗

I. M. Sigal

Dept of Mathematics, Univ of Toronto


Wednesdays 5–6 (BA3008), Thursdays 5–6 (BA6183),
Fridays 2–3 (BA3008)

Contents
1 Examples of equations and general set-up 5
1.1 Important examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Abstract evolution equations . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Key classes of solutions of PDEs and symmetries 10


2.1 Key special solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Appendix: A dynamical systems perspective . . . . . . . . . . . . . . . . . 14
2.3 Symmetries and solutions of PDEs . . . . . . . . . . . . . . . . . . . . . . 16

3 Fourier transform and partial differential equations 26


3.1 Definitions and properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Application of Fourier transform to partial differential equations . . . . . . 28

4 Local Existence for Key Evolution Equations 31


4.1 Reduction to a fixed point problem . . . . . . . . . . . . . . . . . . . . . . 32
4.2 The contraction mapping principle . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Local existence for the nonlinear heat equation . . . . . . . . . . . . . . . . 35
4.4 Local existence for Hartree and Schrödinger equations . . . . . . . . . . . . 37
4.5 Classical solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

November 3, 2017

Thanks to Jordan Bell for careful reading these notes and flagging typos and sloppiness

1
2 Lectures on Applied PDEs, November 3, 2017

5 Global Existence 40
5.1 A priori estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2 Energy and Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.3 A priori estimates for the generalized NLS and Keller-Segel equations . . . 43

6 The implicit function theorem and applications 46


6.1 Gâteaux and derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.2 Appendix: The Fréchet derivative . . . . . . . . . . . . . . . . . . . . . . . 50
6.3 The implicit function theorem . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.4 Existence of breathers and constant mean curvature surfaces . . . . . . . . 52
6.5 Inverse function theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.6 Appendix I: Mean curvature of a hypersurface . . . . . . . . . . . . . . . . 56

7 Theory of bifurcation 60
7.1 Set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2 Key result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
7.4 Linearized stability/instability . . . . . . . . . . . . . . . . . . . . . . . . . 72

8 Linearized stability of systems with symmetry 75


8.1 Systems with symmetry and zero modes . . . . . . . . . . . . . . . . . . . 75
8.2 Linearized stability of kinks . . . . . . . . . . . . . . . . . . . . . . . . . . 76
8.3 Linearized stability of spheres, cylinders and planes . . . . . . . . . . . . . 79

9 Nonlinear stability 83
9.1 Stability: generalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
9.2 Asymptotic stability of kinks . . . . . . . . . . . . . . . . . . . . . . . . . . 84
9.3 Orthogonal Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
9.4 Evolution of Fluctuation ξ . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
9.5 Bound on a(t) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
9.6 Upper Bound on kξkH 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
9.7 Asymptotic Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

10 Elements of Variational Calculus 90


10.1 Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
10.2 The Gâteaux derivative for functionals . . . . . . . . . . . . . . . . . . . . 91
10.3 Critical points and connection to PDEs . . . . . . . . . . . . . . . . . . . . 92
10.4 Hessians and local minimizers. . . . . . . . . . . . . . . . . . . . . . . . . . 96
10.5 Convexity and uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
10.6 Constraints and Lagrange multipliers . . . . . . . . . . . . . . . . . . . . . 98
10.7 Minimization problem and spectrum. . . . . . . . . . . . . . . . . . . . . . 101
10.8 Tangent spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Lectures on Applied PDEs, November 3, 2017 3

10.9 Dual space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103


10.10Functionals on complex spaces . . . . . . . . . . . . . . . . . . . . . . . . . 103
10.11Appendix I: Area of a hypersurface . . . . . . . . . . . . . . . . . . . . . . 104

11 Bifurcation of surfaces of constant mean curvature 105


11.1 General properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
11.2 Bifurcation of new surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
11.3 Proof of the existence part . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
11.4 Proof of the linearized stability . . . . . . . . . . . . . . . . . . . . . . . . 110
11.5 Appendix: Volume preserving mean curvature flow . . . . . . . . . . . . . 113

12 Turing instability 114


12.1 Appendix: Derivation of Theorem 12.1 from Theorem 12.2 . . . . . . . . . 117
12.2 Pattern formation in reaction-diffusion equations . . . . . . . . . . . . . . . 118

13 Abrikosov lattices 120


13.1 Properties of the GLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
13.2 The linear problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
13.3 Appendix: Proof of Theorem 13.1 (todo) . . . . . . . . . . . . . . . . . . 127
13.4 Fixing the gauge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
13.5 Rescaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
13.6 Reformulation of the problem . . . . . . . . . . . . . . . . . . . . . . . . . 128
13.7 Reduction to a finite-dimensional problem . . . . . . . . . . . . . . . . . . 131
13.8 Proof of Theorem 13.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
13.9 Appendix: Proof of Proposition 13.6 . . . . . . . . . . . . . . . . . . . . . 135
13.10Appendix: Solving the equation (13.35b) . . . . . . . . . . . . . . . . . . . 136
13.11Appendix: Remarks about the automorphy factors . . . . . . . . . . . . . 136

14 Hamiltonian and gradient systems 137


14.1 Appendix: The gradient property of the VPF . . . . . . . . . . . . . . . . 138

15 Energy and orbital stability 139


15.1 Energy argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
15.2 Systems with symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
15.3 Orbital stability of kink solutions of the Allen - Cahn equation . . . . . . . 145

16 Harmonic, Schrödinger and wave maps maps 147


16.1 Harmonic maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
16.1.1 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
16.1.2 Topological invariants . . . . . . . . . . . . . . . . . . . . . . . . . 148
16.1.3 Complex representation . . . . . . . . . . . . . . . . . . . . . . . . 148
16.1.4 Bogomolny bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4 Lectures on Applied PDEs, November 3, 2017

16.1.5 Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150


16.1.6 Projective representation . . . . . . . . . . . . . . . . . . . . . . . . 151
16.1.7 Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
16.2 Schrödinger maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
16.3 Wave maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
16.3.1 Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

17 Mean curvature flow 155


17.1 Definition and general properties . . . . . . . . . . . . . . . . . . . . . . . 155
17.1.1 Special solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
17.2 Self-similar surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
17.3 Linearized stability of self-similar surfaces . . . . . . . . . . . . . . . . . . 162
17.4 F −stability of self-similar surfaces . . . . . . . . . . . . . . . . . . . . . . . 168

18 Keller-Segel Equations of Chemotaxis 169


18.1 Properties Keller-Segel equations . . . . . . . . . . . . . . . . . . . . . . . 171
18.2 Rescaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
18.3 M > 8π . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
18.4 Discussion of static solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 178
18.5 Blowup criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
18.6 Appendix: Gradient formulation . . . . . . . . . . . . . . . . . . . . . . . . 180
18.7 Appendix 2: Criterion of break-down in the dimension d ≥ 3 . . . . . . . . 181

19 Minimization: direct methods 182


19.1 General result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
19.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
19.3 Existence of ground state of nonlinear Schrödinger equation (without and
with potential) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

20 Interfaces, vortices, vortex lattices and harmonic maps 192


20.1 Allen-Cahn energy functional and interfaces . . . . . . . . . . . . . . . . . 192
20.2 Vortices in Superfluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
20.3 The Ginzburg-Landau Equations . . . . . . . . . . . . . . . . . . . . . . . 202
20.3.1 Appendix I. Relation to line bundles . . . . . . . . . . . . . . . . . 205
20.4 Harmonic maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206

21 PDEs of quantum mechanics and statistics 208


21.1 Hartree, Hartree - Fock and Gross-Pitaevski equations . . . . . . . . . . . 208
21.2 Appendix II: Heuristic derivation of the Hartree equation . . . . . . . . . . 213
21.3 Quantum statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
21.4 Self-consistent approximation . . . . . . . . . . . . . . . . . . . . . . . . . 214
21.5 Equilibrium states and entropy . . . . . . . . . . . . . . . . . . . . . . . . 217
Lectures on Applied PDEs, November 3, 2017 5

21.6 Local and global existence . . . . . . . . . . . . . . . . . . . . . . . . . . . 218


21.7 Existence of ground states and Gibbs states . . . . . . . . . . . . . . . . . 219
21.8 Appendix: Proof of Lemma 21.2 . . . . . . . . . . . . . . . . . . . . . . . . 220
21.9 Appendix: Hamiltonian formulation . . . . . . . . . . . . . . . . . . . . . . 220
21.10Appendix: Hilbert Space Approach . . . . . . . . . . . . . . . . . . . . . . 222

22 Existence of bubbles and Lyapunov - Schmidt decomposition 223


22.1 Appendix. Details of the Lyapunov-Schmidt reduction . . . . . . . . . . . 230

A Spaces and Operators: Review 235


A.1 Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
A.2 Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
A.3 Lp –spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
A.4 Supplement: Proofs of Hölder’s and Minkowski’s inequalities . . . . . . . . 239
A.5 Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
A.6 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
A.7 Linear operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
A.8 Special classes of operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

B Fourier transform 251

C Linear evolution and semigroups 255

D Local Existence for Evolution Equations 262


D.1 Linear problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
D.2 Reduction to a fixed point problem . . . . . . . . . . . . . . . . . . . . . . 264
D.3 Abstract result on local existence. . . . . . . . . . . . . . . . . . . . . . . . 265

E Elements of spectral theory 267


E.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
E.2 Location of the essential spectrum . . . . . . . . . . . . . . . . . . . . . . . 269
E.3 Perron-Frobenius Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269

1 Examples of equations and general set-up


In this course we will consider several key partial differential equations arising in physics,
material science, biology and geometry. We develop an existence theory for these equa-
tions, describe their key properties, isolate their most important solutions and study
stability or instability of these solutions.
6 Lectures on Applied PDEs, November 3, 2017

1.1 Important examples


Here are some examples of the equations we will be studying in the course:

The Allen-Cahn equation. This equation describes among other things motion of
interfaces between different compounds and is given by

∂u
= 2 ∆u + u − u3 . (1.1)
∂t

Here u = u(x, t) is a real function of x ∈ Rd and t ∈ R+ := [0, ∞) and ∆ is the Laplace


operator (the Laplacian):
d
X ∂ 2u
∆u := 2
.
j=1
∂x j

This simple looking equation has a rich structure and is an object of an intensive inves-
tigation in the last
 50 years. It has an important set of static solutions, called kinks,
χ (x) = tahn √x2
d
, or more generally,

(x − x0 ) · e
 
χ,x0 ,e (x) = tahn √ , (1.2)
2

for any e, x0 ∈ Rd . The kink χ (x) = tahn √x2



d
describes a flat interface at xd = 0.
(The kink (1.2) describes the flat interface which is shifted and rotated compared to
the previous one.) In fact, any slowly varying interface can be constructed by “gluing”
together rotated and shifted kinks.

The Ginzburg-Landau/Gross-Pitaevski/nonliear Schrödinger equation. or non-


linear Schrödinger equation. These describe dynamics of superfluids, Bose-Einstein con-
densates, nonlinear waves in plasmas and on surface of a fluid:

∂ψ
i = −∆ψ + λ|ψ|2 ψ (1.3)
∂t
for some real constant λ. This equation has remarkable special solutions - the vortices,
which for d = 2 are given by
ψ (n) (x) = f (n) (r)einθ , (1.4)
where (r, θ) are the polar coordinates of x ∈ R2 . Even more interestingly, the vortices can
arrange themselves into lattices, giving new solutions for which A.A. Abrikosov received
a few years back a Nobel prize in physics.
Lectures on Applied PDEs, November 3, 2017 7

s1 s2 nt s1

Figure 1: Portraits of vortices

The Keller - Segel equations. The next system of equations models the chemotaxis,
which is the directed movement of organisms in response to the concentration gradient of
an external chemical signal and is common in biology. The chemical signals either come
from external sources or are secreted by the organisms themselves. The latter situation
leads to aggregation of organisms and to the formation of patterns.
Consider organisms moving and interacting in a domain Ω ⊆ Rd , d = 1, 2 or 3.
Assuming that the organism population is large and the individuals are small relative to
the domain Ω, Keller and Segel derived a system of reaction-diffusion equations governing
the organism density ρ : Ω × R+ → R+ and chemical concentration c : Ω × R+ → R+ .
The equations are of the form

∂t ρ = Dρ ∆ρ − ∇ (f (ρ)∇c)
(1.5)
∂t c = Dc ∆c + αρ − βc.

Here Dρ , Dc , α, β are positive functions of x and t, ρ and c, and f (ρ) is a positive function
modelling chemotaxis (positive chemotaxis).
Assuming α and β are constants, this system has the simple homogeneous static
solution c = const and ρ = β/α =const. However this solution is unstable under small
perturbations. For initial conditions close to this static solution, ρ starts growing at
some point and becomes highly concentrated around this point. This is a chemotactic
aggregation which leads to formation of fruition bodies (colonies formed by Escherichia
coli under starvation conditions, or multicellular structures of ∼ 105 cells, by single cell
bacterivores, when challenged by adverse conditions). Similar phenomenon occurs in
dynamics of tumours.
Chemotaxis underlies many social activities of micro-organisms, e.g. social motility,
fruiting body development, quorum sensing and biofilm formation.
A simplified version of these equations is used to model stellar collapse and crime
patterns.

The Fisher-Kolmogorov-Petrovsky-Piskunov equation. This equation describes


propagation of mutations in population biology and flames in combustion theory. It is
8 Lectures on Applied PDEs, November 3, 2017

given by
∂u
= ∆u − (u − 1)u. (1.6)
∂t
It has interesting solutions, called traveling waves, which are of the form

u(x, t) = φ(x − vt), (1.7)

where ϕ(y) and v are independent of t and are called the traveling wave profile and the
traveling wave velocity.

PDEs of Quantum Physics.

1.2 Abstract evolution equations


We would like to address a general theory of evolution PDEs, i.e. equations of the form

∂t u = F (u), u|t=0 = u0 , (1.8)

where t → u(t) is a path in some space, F is a map defined on the same space, ∂t u = u̇ = ∂u ∂t
and we understand (??) as ∂t u(t) = F (u(t)). u0 is called the initial condition and (??),
the initial value problem.
The situation we will be mostly interested in is when F is a partial differential operator,
linear or nonlinear. An example of such an equation is the equation:
∂u
= ∆u + g(u). (1.9)
∂t
In this example, F (u) = ∆u + g(u). This is the celebrated nonlinear heat or reaction-
diffusion equation. The term ∆ is responsible for diffusion and the term, g(u), describes
the reaction in the system modelled.
Let us consider first several other examples of maps F :
1) F (u) = ∆u,
2) F (u) = f ◦ u for a given function f ,
∇u
3) F (u) = div( √ ).
1+|∇u|2

4) F (u) = ϕ ∗ u for given ϕ ∈ L1 (Rn ).


Note that the map in 1) is linear, while the map in 2) is linear if f is linear, and nonlinear
otherwise.
Sometimes (??) is called the dynamical system and F , the vector field (finite or infinite
dimensional). Though we define F on a vector space, it can be also defined on a manifold
(again, finite or infinite dimensional).
Lectures on Applied PDEs, November 3, 2017 9

To solve equation (??), we have to choose a space, say Y , to which the vector-function
t → u(t) belongs for all t considered. We have to make sure that F is defined on this
space (and maps it into another space, say, Z). Depending on the problem at hand, we
choose different spaces. For instance, for the examples 1) and 3) above, F maps the space
Y := C k (Ω) into the space C k−2 (Ω), and in the examples 4), Lp (Rn ) into itself. More
abstractly, Y could be some space with a norm, or a Banach space. For the definition
and review of Banach spaces see Subsection A.2 below.
We also have to choose a space, say X, to which the function t → u(t) belongs. If
u(t) ∈ Y for t in some interval I, then we can choose

• X = C(I, Y ), where, say, I := [0, T ].

Here X = C(I, Y ) is the space of once differentiable vector - functions u : t → u(t) on I


with values in Y and with the norm

kukX := sup( u(t) Y + ∂t u(t) Z ).
t∈I

If the map F (u) is linear, F (u) = Au, then the corresponding equation is also linear,

∂u
= Au and u|t=0 = u0 . (1.10)
∂t
An important example of the linear evolution equation is the linear (free) heat equation

∂u
= ∆u and u|t=0 = u0 , (1.11)
∂t

where u : Rnx × R+ t → R, Rt := {t ∈ R : t ≥ 0}, is an unknown function, and u0 : R → R


+ n

is a given initial condition.


For linear flows, there are general situations where we can show the existence of the
global (all t’s) flow:

• A is a bounded operator;

• A is either self-adjoint (A∗ = A) and bounded above, A ≤ C for some C < ∞


(hu, Aui ≤ Ckuk2 ) or anti-self-adjoint (A∗ = −A);

• A is a ‘constant coefficient pseudo-differential’ operator; more precisely, A = a(−i∇x ),


where a is some decent function.

Exercise 1.1. Review Appendix A.1 and do Exercise A.1.


10 Lectures on Applied PDEs, November 3, 2017

2 Key classes of solutions of PDEs and symmetries


2.1 Key special solutions
We consider key solutions for the dynamical system on some space Y of functions on Rn ,
given, as usual, by
∂u
= F (u). (2.1)
∂t
where F is a map on this space Y .

Static solutions. Static solutions are the solutions independent of time t. As the result
they satisfy the equation F (u) = 0. Here are some examples.
Allen-Cahn equation. First, we consider the Allen-Cahn equation, which plays a cen-
tral role in material science. It presents a basic model with many generalizations and
extensions. It is a reaction diffusion equation of the form

= 2 ∆u − g(u),
 ∂u
x ∈ Ω, t > 0,
∂t
∂u (2.2)
∂n
= 0, x ∈ ∂Ω,

with the initial condition u|t=0 = u0 (x), x ∈ Ω. Here Ω ⊆ Rn , u : Ω → R,  is a small


parameter and g(u) = u3 − u.
More generally, g : R → R is the derivative, g = G0 , of a double-well potential:
G(u) ≥ 0 and has two non-degenerate global minima with the minimum value 0. G(u) is
also called a bistable potential (see Figure 2). In our case, G(u) = 41 (u2 − 1)2 .

G u

a b x
a

Figure 2: Hill-valley-hill structure for −G, and the kink for u.

We consider static solutions of Allen-Cahn equation. They satisfy the static Allen-
Cahn equation,
− 2 ∆u + g(u) = 0. (2.3)
For complex u and g(u) = |u|2 u − u, (2.3) is the Ginzburg-Landau equation. It appeared
first in condensed matter physics. For this equation G is of the form G1 (|u|2 ), where G1
is the “half” of the double well potential.
Equation (2.3) has two trivial solutions u = ±1, which descibe pure phases.
Lectures on Applied PDEs, November 3, 2017 11

We are interested in the phase separation phenomena when the solutions are close to
+1 or −1 (pure phases) in most of the space with sharp transitional interfaces separating
+1 regions from −1 regions. We consider the simplest of planar interface.
We look for static solutions which very only in a direction, e, transversal to a plane,
x · e = 0, i.e. of the form
x·e
 
χ (x) = χ , (2.4)

for any e ∈ Rd , where χ is a function of single variable. By rotational symmetry and
scaling properties of (2.3), the function χ(s) satisfies the equation
χ00 = g(χ). (2.5)
This is the Newton’s equation with the potential −G(φ), where G0 (φ) = g(φ). The
potential −G(φ) has equilibria at φ = ±φ0 and at φ = 0. Hence the equation (2.3) has
the homogeneous solutions, φ = ±φ0 and φ = 0.
It is clear from this mechanical analogy that (2.3), besides these homogeneous solu-
tions, has the solutions which go asymptotically to ±φ0 as s → ±∞ (see Figure 2). To
show the existence of these solutions, we observe that by the conservation of mechanical
energy, we have
1 0 2
(φ ) − G(φ) = 0 (2.6)
2 p
(we take the mechanical energy to be 0), which can be solved for φ0 as φ0 = ± 2G(φ),
and then integrated to obtain φ. The solution with the + sign is called the kink and, with
the − sign, the anti-kink.
In our case, G(φ) = 14 (φ2 −1)2 , i.e. g(u) = u3 −u, and the integration can be performed
explicitly and gives
x·e
 
χ (x) = tahn √ , (2.7)
2
for any e, x0 ∈ Rd .
Example 2.1.
ωo (ϕ − ϕo )2 , ϕ ≥ 0,

G(ϕ) = and ωo ϕ2o = ωw ϕ2w . (2.8)
ωw (ϕ − ϕw )2 , ϕ ≤ 0,
Then equation (2.3) is piecewise linear, and we get

ϕw (1 − e√− ωw z ), z ≤ 0,

ϕk (z) = (2.9)
ϕo (1 − e ωo z ), z ≥ 0.
ϕ is a function consisting of a kink and an antikink glued together at a distance R.
Exercise 2.1. Show (2.9).
One can show existence of kink (and anti-kink) solutions by dynamical systems theory
and variational techniques (see Appendix 2.2 and Section 10, respectively).
12 Lectures on Applied PDEs, November 3, 2017

Lamellar phase. In this situation, layers of +1 and −1 phases (substances) coexist in


a periodic array. We conjecture that one can construct a periodic solution corresponding
to an array of kinks and antikinks (see below for a discussion of an approach to this
problem).

Stationary waves. Stationary or standing waves (or breathers) are solutions of the
form
Ψ(x, t) = e−iλt ϕ(x), (2.10)
where ϕ and λ are time-independent. The function ϕ is called the profile of the stationary
wave (2.10). As an example, consider the nonlinear Schrödinger or Gross - Pitaevskii
equation
i∂t Ψ = −∆Ψ + κ|Ψ|2 Ψ, (2.11)
for Ψ : Rd → C and Ψ|t=0 = Ψ0 . It has solutions of the form (2.10), where ϕ and λ ∈ R
satisfy the equation
− ∆ϕ + κ|ϕ|2 ϕ = λϕ (2.12)
(called the nonlinear eigenvalue problem).

Traveling waves. Traveling waves are solutions of the form

u(x, t) = φ(x − vt), (2.13)

where ϕ(y) and v are independent of t and are called the traveling wave profile and the
traveling wave velocity. Consider the reaction-diffusion equation

∂u
= ∆u − g(u). (2.14)
∂t
Substituting (2.58) into (2.14), we obtain the following equation for ϕ(y) and v

∆φ − g(φ) + v · ∇φ = 0. (2.15)

Consider the one-dimensional case. Then this equation can be rewritten as

φ00 = g(φ) − vφ0 . (2.16)

This is the Newton’s equation with the potential −G(φ), where G0 (φ) = g(φ), and the
friction term vφ0 .
Compute the change in mechanical energy:
1 0 2
(φ ) − G(φ) = (φ00 − g(φ))φ0 = −v(φ0 )2 < 0, for v > 0.

∂x (2.17)
2
Lectures on Applied PDEs, November 3, 2017 13

This implies that for v > 0 the particle descends to a (local) minimum of −G.
To be specific we consider the Fisher-Kolmogorov-Petrovsky-Piscunov (FKPP) equa-
tion, which appears in population biology and combustion theory. This is Eqn. (2.14)
with g(u) = u(u − 1):

∂u
= ∆u − (u − 1)u. (2.18)
∂t

For this equation, G = 13 u3 − 21 u2 , see Figure 3 (where the label G should be replaced by
−G).

vm
G
p vn
g

o oo ph
vm

Figure 3: Functions −G, g and φ.

Using the mechanical analogy described above, we can solve the equation (2.16) for
the nonlinearities described. At the remote past (time y = −∞), the particle leaves the
top of the hill in −G(φ) at φ = 1 and moves to the left toward the wall loosing the
altitude due to the dissipation. Hitting the wall particle turns around and moves toward
the opposite wall and so forth until it relaxes to the bottom of the well at φ = 0. The
front of the wave is at y at which φ(y) hits 0 the first time.
An analysis of the traveling wave equation (2.16) using elementary dynamical system
theory, done in Appendix 2.2 shows that the solution is monotonically decreasing for v > 2
(the over-damped case) and is oscillating for v < 2.
More generally we assume g(u) satisfies the conditions g(0) = g(1) = 0 together with
g 0 (0) < 0 and g 0 (1) > 0, see Figure 3 (where the label G should be replaced by −G).
Another type of the the nonlinearity g(u), which is used in combustion theory is shown
in Fig. 4.

Exercise 2.2. Investigate (and find if possible) the traveling waves in the reaction-diffusion
equation with 1) g(u) = (u − a)(u − 1)u with 0 < a < 1/2 (the a = 1/2 case gives the
Allen-Cahn equation) 2) g(u) as in Figure 4 (flame propagation for the ignition nonlin-
earity).
14 Lectures on Applied PDEs, November 3, 2017

g −G

T T
1 u 1

Figure 4: Ignition case

A more complicated situation takes place for the nonlinear Schrödinger or Gross -
Pitaevskii equation (2.11). A simple computation shows that (2.11) has the solutions of
the form
1 1 2
Ψ(x, t) := ei( 2 v·x− 4 |v| t−λt) φ(x − vt − x0 ) (2.19)
where e−iλt φ(x) is a stationary solution to (2.11). This solution is obviously a traveling
wave.
As another example we consider the celebrated KdV equation originated in the de-
scription of waves in shallow channels:

∂t u + u∂x u + ∂x3 u = 0. (2.20)

The traveling waves u(x, t) = φ(x − vt) for this equation satisfy the equation φ∂y φ +
∂y3 φ − v∂y φ = 0. Integrate this equation, to obtain 21 φ2 + ∂y2 φ − vφ = c, where c is a
constant of integration. To solve this equation, we interpret it as a Newton’s equation,
∂y2 φ = − 12 φ2 + vφ + c (in time y) and write out the conservation of energy, i.e. multiply
it by ∂y φ and integrate the result to obtain 16 φ3 + 12 (∂y φ)2 − 21 vφ2 − cφ = c0 , where c0 is
another q constant of integration. This is the first order ODE, which can be rewritten as
∂y φ = ± − 16 φ3 + 21 vφ2 + cφ + c0 and integrated as


Z
q = ±y + c00
1 3 1
− 6 φ + 2 vφ2 + cφ + c0

to obtain (not finished).


Exercise 2.3. Review Appendices A.2 and A.3 and do Exercise A.2.

2.2 Appendix: A dynamical systems perspective


We investigate the static equation (2.5) and the traveling wave equation (2.16) also from
the point of view of elementary dynamical system theory.
Rewrite (2.5) as
χ0 = ψ, ψ 0 = g(χ). (2.21)
Lectures on Applied PDEs, November 3, 2017 15

This dynamical system has equilibria (φ0 , 0), where φ0 solves g(φ) = 0. In our case,
φ0 = ±1, 0, and therefore the equilibria are (±1, 0) and (0, 0).
The linearized vector field at an equilibrium (φ0 , 0), where φ0 = ±1, 0, is
 
0 1
. (2.22)
g 0 (φ0 ) 0
p
The eigenvalue equation, λ2 − g 0 (φ0 ) = 0, gives λ = ± g 0 (φ0 ). At φ0 = 0 we have
g 0 (0) < 0, which gives two purely imaginary eigenvalues. Moreover, g 0 (±1) > 0, which
gives one positive and one negative eigenvalue. Hence the stationary point (0, 0) is a
stable equilibrium (a stable focus) and (±1, 0) are unstable equilibria (saddle points).
There is one trajectory going from (−1, 0) to (1, 0) and one from (1, 0) to (−1, 0). These
are exactly the kinks and anti-kinks (or fronts) mentioned above (see Figure 2).

0.25

0.2

0.15

0.1

0.05

0
Y

−0.05

−0.1

−0.15

−0.2

−0.25
Separatrix

−0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4


X

Figure 5: v > 2 there exists a separatrix. Plotted is arrows with direction dψ/dφ, at
X-coordinate φ and Y -coordinate ψ, note that dψ/dφ = ψ −1 φ(φ − 1) − v.

Now, we turn to the traveling wave equation (2.16). Rewrite (2.16) as

φ0 = ψ, ψ 0 = g(φ) − vψ. (2.23)

The equilibria (or static solutions or stationary points) of this system are (φ0 , 0) where φ0
solves g(φ) = 0. For g(φ) = φ(φ − 1) we have the equilibria (0, 0) and (1, 0). To determine
whether these equilibria are stable or not, we compute the linearized vector field (Jacobi
16 Lectures on Applied PDEs, November 3, 2017

 
0 1
derivative of the vector field on the r.h.s.) at (φ0 , 0). It is . The eigenvalue
g 0 (φ0 ) −v
equation, λ(λ + v) − g 0 (φ0 ) = 0, gives
r
−1 v2
λ= v± + g 0 (φ0 ). (2.24)
2 4
At φ0 = 0 we have g 0 (0) = −1, which gives two negative eigenvalues for v ≥ 2 and two
complex eigenvalues with negative real parts for v < 2. Hence the equilibrium (0, 0) is a
stable equilibrium. (0,0) is a stable focus for v > 2 and stable spiral for v < 2. For φ0 = 1
we find g 0 (1) = 1 and the point (1, 0) is a saddle point. For v > 2 there is a separatrix.
This is a bit subtle, see Figure 5 above.
The upshot of this is that for v ≥ 2 the solutions is monodically decreasing and for
v ≤ 2 , oscillating, see Figure 3 (where vn and vm denote the monotonic and oscillating
solutions, respectively.)

2.3 Symmetries and solutions of PDEs


It turns out that special solutions we considered above come from specific symmetries
of the equation we consider. We explore the relation between symmetries of evolution
equations and solutions of such equations in more depth. Transformations which map
solutions of an equation into solutions are called symmetries of this equation. Clearly, all
symmetries of a given equation form a group. This group is called the symmetry group of
the given equation. It is often a representation of an abstract group.
Let F be a map on some space Y . We consider the dynamical system (2.1), ∂u ∂t
= F (u).
If the map F is invariant under an invertible transformation T , i.e.

F (T u) = T F (u), (2.25)

then T is a symmetry: if u is a solution, then so is T u. We distinguish space - time


symmetries, i.e. symmetries of the form

u(x, t) 7→ u(g(x, t)), (2.26)

where g is a transformation of the space time Rn+1 , and gauge symmetries, i.e. symmetries
of the form
u(x, t) 7→ gu(x, t), (2.27)
where g is a transformation of the target space of u(x, t). Here are examples of space -
time symmetries:
Translation symmetry: Assume the additive group Rn acts on the space of u’s as: for
any h ∈ Rn ,
Th : u(x, t) 7→ u(x + h, t); (2.28)
Lectures on Applied PDEs, November 3, 2017 17

Rotation and reflection symmetry: for any R ∈ O(n) (including the reflections
f (x) → f (−x))
TR : u(x, t) 7→ u(Rx, t). (2.29)

Translations and rotations are called rigid motions. They are a part of Galilean group:

(R, h, v, τ )(R0 , h0 , v 0 , τ 0 ) = (R0 R, R0 h + h0 , R0 v + v 0 , τ + τ 0 ), (2.30)

with the representation

T(R,h,v,τ ) : u(x, t) 7→ u(Rx + h + vt, t + τ ). (2.31)

Examples of gauge symmetries will be special cases of the following general situation.
Assume V is a vector space and G, a group acting on V . Then for ψ : Rn × R+ → V , we
define
ψ(x, t) 7→ gψ(x, t), g ∈ G. (2.32)

The group G is called the gauge group and (2.32), the gauge transformation with the
corresponding gauge group G. Specifically, consider V = Cm , or = Rm . Then G = U (m)
or O(m), the unitary or orthogonal (rotation) group. In particular, for the gauge group
U (1), we have
ψ(x, t) 7→ eiα ψ(x, t), α ∈ R. (2.33)

Example. Consider the nonlinear Schrödinger or Gross - Pitaevskii equation (2.11).


Show that it has translational, rotational and gauge symmetries. What about NLH and
NLW?
Consider first solutions of the dynamical system (2.1), invariant under some subgroup
of its symmetry group:
Static solutions. Solutions invariant under time translations are called static solutions.
They are independent of time and form the most important class of special solutions. They
are also called equilibria and sometimes stationary solutions. Thus u∗ is an equilibrium
or static solution iff u∗ is independent of time and satisfy the equation F (u∗ ) = 0.
Examples: NLH, NLS, NLW.
Homogeneous solutions. Solutions invariant under spatial translations are independent
of x and are called the homogeneous solutions.
Spherically symmetric solutions. Solutions invariant under spatial rotations are called
spherically symmetric solutions. They are independent of angular variables and depend
only on two variables, |x| and t.
18 Lectures on Applied PDEs, November 3, 2017

Covariance. We generalize the above definition a bit and say that a map F is covariant
under an invertible transformation T , iff there is a real number a s.t.

F (T ϕ) = aT F (ϕ). (2.34)

In this case we can upgrade T to a symmetry by defining: if u is a solution, then so is

T̃ ut := T uat . (2.35)

Space-time symmetries. An example of a space-time symmetry is the Galilean sym-


metry of the nonlinear Schrödinger equation (NLS):
1 1 2
ψ(x, t) 7→ ei( 2 v·x− 4 |v| t) ψ(x − vt, t). (2.36)

Rotating solitons

Equivariant solutions. We begin with auxiliary definitions. We say two solutions are
equivalent if one can be obtained from the other by a gauge transformation. The group
of rigid motions is defined as a semi-direct product of the groups of translations and
rotations. We denote by Tg , g ∈ Grm , the action of the group, Grm , of rigid motions
on space of solutions. For the groups of translations, and rotations, is given (2.28) and
(2.29), respectively.
We say a solution u is equivariant (under a subgroup, G, of the group of rigid motions)
iff the action of G on u takes it into the (gauge-) equivalent function, i.e., for any g ∈ G,
there is γ = γ(g) s.t.
Tg u = Γγ u,
where Γγ is the action of for the gauge group, given in (2.33). Here are two important
examples coming from the Gross - Pitaevski equation, mentioned above in (2.37), and the
Ginzburg-Landau equations studied in detail later.
Vortices. We consider the static Gross - Pitaevskii (or Ginzburg-Landau) equation,
appearing in the condensed matter physics (the theory of superfluidity and Bose - Einstein
condensation). It is given by
∆Ψ = κ2 (|Ψ|2 − 1)Ψ, (2.37)
for Ψ : Rd → C. In the dimension d = 2 and for G, the group of rotations, O(2), these
equations have O(2)− equivariant solutions, which are labeled by integers n, which are
given by
Ψ(n) (x) = f (n) (r)einθ , (2.38)
where (r, θ) are the polar coordinates of x ∈ R2 . These solutions are called vortices. (n
labels the equivalence classes of the homomorphisms of S 1 into U (1).)
Lectures on Applied PDEs, November 3, 2017 19

Abrikosov lattices. If G is the subgroup of the group of lattice translations for a


lattice L, then we call the corresponding equivariant solution a lattice, or L-gauge-periodic,
solution. Explicitly,
Ts Ψ(x) = Γgs (x) Ψ(x), ∀s ∈ L, (2.39)
where gs : R2 → R is, in general, a multi-valued differentiable function, with differences
of values at the same point ∈ 2πZ, and satisfying

gs+t (x) − gs (x + t) − gt (x) ∈ 2πZ. (2.40)

The latter condition on gs can be derived by computing Ψ(x + s + t) in two different ways.
Note that for Abrikosov lattices, the physical quantity, |Ψ|2 , (the probability density
(of superfluid or Bose-Einstein condensed atoms) is doubly-periodic with respect to some
lattice L.
One can also show the converse: a state Ψ ∈ Hloc 1
(R2 ; C) for which the physical
2
quantity, |Ψ| , is doubly-periodic with respect to some lattice L (we call such a state the
(generalized) Abrikosov (vortex) lattice) is a L-gauge-periodic state.

The Ginzburg-Landau Equations. A more complicated example is presented by the


Ginzburg-Landau equations of superconductivity (and particle physics). These equations
describe equilibrium configurations of superconductors, and of the U (1) Higgs model from
particle physics, and are given by

∆A Ψ = κ2 (|Ψ|2 − 1)Ψ, (2.41a)

curl∗ curl A = Im(Ψ̄∇A Ψ). (2.41b)


for u := (Ψ, A) : R2 → C × R2 , ∇A = ∇ − iA, and ∆A = ∇2A , the covariant derivative
and covariant Laplacian, respectively.
These equations are discussed in detail in Section ??. Here we mention only vortex
and Abrikosov lattice solutions of these equations.
Origin. Superconductivity. The Ginzburg-Landau theory is a central part of theory of
superconductivity (its reviews can be found in any text on superconductivity, see eg.
[19, 18]). It gives a macroscopic description of a superconducting material in terms of
a complex-valued function Ψ(x) = an order parameter, so that |ψ(x)|2 gives the local
density of (Cooper pairs of) superconducting electrons, and the vector field A, so that
B(x) := curl A(x) = the magnetic field. The vector quantity J(x) := Im(Ψ̄∇A Ψ) =
the superconducting current. The parameter κ > 0 depends on the material properties
of the superconductor.
In the idealized situation of a superconductor occupying all space and homogeneous
in one direction, we are led to a problem on R2 and so may consider Ψ : R2 → C and
A : R2 → R2 . In this case, curl A := ∂1 A2 − ∂2 A1 is a scalar, and for a scalar function,
B(x) ∈ R, curl B = (∂2 B, −∂1 B) is a vector.
20 Lectures on Applied PDEs, November 3, 2017

Particle physics. In the Abelian-Higgs model, Ψ and A are the Higgs and U (1) gauge
(electro-magnetic) fields, respectively.
Geometrically, one can think of A as a connection on the principal U (1)-bundle Rn ×
U (1), n = 2, 3.
Symmetries. The Ginzburg-Landau equations (2.41) admit several symmetries (that
is, transformations which map solutions to solutions). Namely, besides the translation and
rotation symmetry, (2.28) - (2.29), it has
Gauge symmetry: for any sufficiently regular function η : R2 → R,
Γγ : (Ψ(x), A(x)) 7→ (eiη(x) Ψ(x), A(x) + ∇η(x)). (2.42)
Equivalence classes of solutions, gauge conditions.
Exercise 2.4. Prove that the above transformations are symmetries of the Ginzburg-
Landau equations (2.41).
Thus the set of all solutions of the Ginzburg-Landau equations can be split into equiv-
alence classes of solutions related by gauge transformations. A condition which pick a
subclass of each equivalence class is called the gauge condition. An example of a gauge
condition is div A = 0. This can be also arranged: if div A 6= 0 we can always find a gauge
η s.t. div(A + ∇η) = 0, namely, we take η solving the equation −∆η = div A.
One of the analytically interesting aspects of the Ginzburg-Landau theory is the fact
that, because of the gauge transformations, the symmetry group is infinite-dimensional.
Vortices. One of the most interesting mathematical and physical phenomenon con-
nected with the Ginzburg-Landau equations is the presence of vortices, which are equiv-
ariant solutions for the group of rotations, O(2):
Ψ(n) (x) = f (n) (r)einθ and A(n) (x) = a(n) (r)∇(nθ) , (2.43)
where (r, θ) are the polar coordinates of x ∈ R2 . They are labeled by the equivalence
classes of the homomorphisms of S 1 into U (1), i.e. by integers n.
For every n ∈ Z, the vortex exists and is unique, up to symmetry transformation, and
its profile has the following properties (see [7, ?] and references therein).
|∂ α (1 − f (n) (r))| ≤ ce−mκ r ,
|∂ α (1 − a(n) (r))| ≤ ce−r , (2.44)
(n) n n+2 (n) 2 4
f (r) = r + O(r ) and a (r) = r + O(r ), as r → 0.

Here mκ := min( 2κ, 1). The exponential decay rates at infinity for f (n) (r) and a(n) (r)
are called the coherence length and penetration depth, respectively.
To prove existence of vortices ones uses that the Ginzburg-Landau equations are Euler
- Lagrange equations for the energy functional
κ2
Z
E(Ψ, A) := |∇A Ψ|2 + | curl A|2 + (1 − |Ψ|2 )2 , (2.45)
R2 2
Lectures on Applied PDEs, November 3, 2017 21

which is the difference between the supperconducting and normal free energies in a state
(Ψ, A). One plugs the expressions (2.43) into this functional to obtain a functional in
f (n) (r) and a(n) (r) and then finds a minimizer of the latter (see [?]).
Thus, roughly speaking, a vortex is a spatially localized structure with a non-trivial
winding. In a superconductor, a vortex represents a localized defect where the normal
state intrudes, and magnetic flux penetrates.
Abrikosov lattices. If G is the group of lattice translations for a lattice L, then we call
the corresponding equivariant solution a lattice, or L-gauge-periodic, solution. Explicitly,

Ts (Ψ(x), A(x)) = Γgs (x) (Ψ(x), A(x)), ∀s ∈ L, (2.46)

where gs : R2 → R is, in general, a multi-valued differentiable function, with differences


of values at the same point ∈ 2πZ, and satisfying

gs+t (x) − gs (x + t) − gt (x) ∈ 2πZ. (2.47)

The latter condition on gs can be derived by computing Ψ(x + s + t) in two different ways.
Note that for Abrikosov lattices, all physical properties, the density of superconducting
pairs of electrons, ns := |Ψ|2 , the magnetic field, B := curl A, and the current density,
J := Im(Ψ̄∇A Ψ), are doubly-periodic with respect to some lattice L.
One can also show the converse: a state (Ψ, A) ∈ Hloc 1
(R2 ; C) × Hloc
1
(R2 ; R2 ) for which
the physical properties above are doubly-periodic with respect to some lattice L (we call
such a state the (generalized) Abrikosov (vortex) lattice) is a L-gauge-periodic state.

Solitons. Symmetries allow us to introduce special classes of solutions: stationary, ro-


tating and traveling waves, which sometimes grouped as solitons. We begin with some
definitions. The family {Ts , s ∈ R} is called a one-parameter group, iff each Ts is bounded
and
T0 = 1; Tt ◦ Ts = Tt+s . (2.48)
We define the generator, A, of the group Tλ as

Aϕ = ∂λ Tλ ϕ|λ=0 , (2.49)

for all ϕ’s for which the above derivative exists. The set of all such ϕ’s form the domain,
D(A), of A. We have ∂t Tt u0 = ATt u0 , t ≥ 0.
Consider the evolution equation (2.1). Let {Tλ } be a one-parameter generalized sym-
metry group of the equation (2.1) in the sense of (2.34). This means that for each λ, Tλ
is the symmetry in the sense of

F (Tλ ϕ) = b(λ)Tλ F (ϕ), (2.50)


22 Lectures on Applied PDEs, November 3, 2017

for some Ggauge −valued function b(λ), where Ggauge is the gauge group of F . Since {Tλ } is
a one-parameter group, we must have that b(λ) satisfies b(s + t) = b(s)b(t). We consider
solutions of the form
u = Tλ ϕ, (2.51)
where λ = λ(t), while ϕ is t−independent. We call such solutions, if they exist, solitary
waves. Substituting this into (2.1), using (2.50) and multiplying the resulting equation
by Tλ−1 , we obtain the equation for ϕ and λ:

∂λ
Aϕ = b(λ)F (ϕ), (2.52)
∂t
where A is the generator of the group Tλ . Since the r.h.s. of (2.52) is independent of t,
we see that ∂λ b(λ)−1 must be independent of t as well, say = −v, i.e. λ must solve the
R λ ∂t
equation 0 b(s)ds = −vt − λ0 , for some constant λ0 . In the simplest case of b(λ) = 1, it
should be of the form
λ = −λ0 − vt, (2.53)
for some constants λ0 and v, so that (2.52) becomes

F (ϕ) + vAϕ = 0. (2.54)

This is a time-independent equation for ϕ and v. The function ϕ is called the solitary
wave profile. We summarize this as

Theorem 2.1. Let {Tλ } be a one-parameter symmetry group of the equation (2.1) with
a generator A. Then (2.1) has a solitary wave solution of the form u = Tλ ϕ, where λ
depends on t, while ϕ is a t−independent function iff λ = −λ0 − vt, for some constants
λ0 and v, and ϕ and v satisfy (2.54).

Moving reference frame. We transform the equation to the moving frame by setting
u = Tλ w. Then w satisfies the equation

∂w
= F (w) + vAw. (2.55)
∂t
Now, if u = Tλ ϕ is a solitary wave for (2.1), then the solitary wave profile ϕ is a stationary
solution to the new equation (2.55). Thus one can apply general definitions and results
for the static solutions to solitary waves as well.
Assume that the space, Y , of solutions is a space of functions on Rn and let the
additive group Rm acts on the space of u’s: h → Th u, ∀h ∈ Rn and we assume that the
map F is covariant w.r.to this group, i.e. (2.35) holds. We consider several realizations
of the action Th of Rm .
Lectures on Applied PDEs, November 3, 2017 23

Stationary waves. The notion of stationary waves is related to the U (1)−gauge sym-
metry. Let Y be a space of complex functions. Let Y be a complex space and consider
the one-parameter group of transformations Tα f := eiα f, α ∈ R, on Y . The generator
of this transformation is Af = if , the solitary waves in this case are of the form

u(x, t) = ei(α0 +λt) ϕ(x), (2.56)

where ϕ(y) and λ satisfy the equation

F (ϕ) + iλϕ = 0, (2.57)

the nonlinear eigenvalue problem. Such solutions are called the stationary or standing
waves or breathers.

Traveling waves. The traveling waves arise from the translational symmetry. Let Th
be the group of translations,
(Th f )(x) := f (x + h).

Exercise 2.5. Show that the generator Aj for translations along the coordinate xj is
A j = ∂ xj .

For the translation invariance, the equations (2.51), (2.53) and (2.54) become

u(x, t) = ϕ(x − vt), (2.58)

where ϕ(y) and v satisfy the equation

F (ϕ) + v · ∇ϕ = 0. (2.59)

They are called the traveling waves. The function ϕ(y) is called the traveling wave profile.

Exercise 2.6. Show that the map F (u) := ∆u + |u|p−1 u is covariant under the shifts (we
say that F is translation invariant).

Rotating waves. The above construction can be generalized to other Lie groups. For
instance, consider an action of the group O(m) of rotations, or the group U (m). If Y be a
space of complex functions on Rn with values in Rm or Cm , then we can define an action
of O(n) on Y by
(TR f )(x) := f (R−1 x), R ∈ O(n),
or an action of O(m) on Y by

(TR f )(x) := Rf (x), R ∈ O(m).


24 Lectures on Applied PDEs, November 3, 2017

Similarly for an action of U (m). Note that the generator of TR f in the first case is −i
times the angular momentum of Quantum Mechanics. One can also combine various
transformations.

Shrinkers and expanders. The group R can be also realized on functions of n variables
as a scaling transformation:
(Tθ f )(x) := eαθ f (eθ x)
for some α.

Exercise 2.7. Show that the generator A is A = x · ∇ + α.

Exercise 2.8. Show that the map F (u) := ∆u + |u|p−1 u is covariant under the scaling
2
transformation above with α = p−1 .

Taking λ = eθ and we write the solitary waves in this case in the form

u(x, t) = λ(t)α ϕ(λ(t)x), (2.60)


∂λ
where v := ∂t
satisfy the equation

F (ϕ) + v · (y · ∇ + α)ϕ = 0. (2.61)

They are called variously the self-similar solution or shrinker or expanders, depending on
whether λ(t) grows or decreases. The function ϕ(y) is called the self-similar wave profile.
Passing to the variables y = λ(t)x is the same as using the frame of reference expanding
or shrinking together with the traveling wave. In this variable, the traveling wave is a
stationary solution to the new equation

∂w
= F (w) + v · (y · ∇ + α)w. (2.62)
∂t

As an example we consider the mean curvature flow (MCF) equation of a surface S. If


S is given by a graph, S = graph f , of function f : U → R, where U is an open set in Rn+1
0 0 0
i.e. S := {x = (x , f (x )) x ∈ U }, then the MCF equation reduces to (see Subsection
17.1)
∂f p
= − |∇f |2 + 1H(f ), (2.63)
∂t
 
∇f
where H(f ) := −div √ 2 , the mean curvature of S.
|∇f | +1

‘Spherically symmetric’ (equivariant) solutions:


Lectures on Applied PDEs, November 3, 2017 25

p
a) Sphere: upper hemisphere can be given
 as thegraph of the function f = R2 − |x0 |2 .
√ ∇f
Using the formula H(f ) := −div , it is easy to compute H(f ) = Rn and
|∇f |2 +1
p
therefore we get Ṙ = − Rn which implies R = R02 − 2nt. So this solution shrinks
to a point.
n−1
b) Cylinder: it is a sphere in in the variables (x1 , . . . , xn−1 ) H(x) = R
and Ṙ = − n−1
R
p
which implies R = R02 − 2(n − 1)t.

The MCF equation ((2.63)) is invariant under the rescaling transformation

Tλ : f (u, t) → λf (λ−1 u, λ−2 t), (2.64)

which corresponds to surfaces re-scaled as S(t) ≡ S λ(t) := λ(t)S (standing waves). This
leads to the soliton

f (u, t) = λφ(λ−1 u), λ depends on t. (2.65)


u
Substituting this into (2.63) and setting y = λ
we find
q
1 + |∇y φ|2 H(φ) = a(y∂y − 1)φ, (2.66)

where a = −λ̇λ. Since the l.h.s. is independent of time, we conclude that

a is time-independent. (2.67)

In the non-trivial case, a is a non-zero constant, which implies λ = λ0 − 2at. So
λ20
i) a > 0 ⇒ λ → 0 as t → T := 2a
⇒ (2.65) (S λ ) is a shrinker.

ii) a < 0 ⇒ λ → ∞ as t → ∞ ⇒ (2.65) (S λ ) is an expander.


p
Solution to (2.66): Sphere φ(y) = R2 − |y|2 , with R = na .
p

Combined symmetries. We consider solitons generated by two groups, one a space


group, say the group of translations, Th , and another, a gauge group, say the U (1)−gauge
group, Γα = eiα . (In other words, we allow the soliton to move inside the gauge-equivalent
class of solutions.) Hence we look solutions of the form

u = Γα Th ϕ, (2.68)

where h = h(t) and α = α(t), while ϕ is t−independent. To be specific, we consider the


nonlinear Schrödinger or Gross - Pitaevskii equation (2.11) and look for solution of the
26 Lectures on Applied PDEs, November 3, 2017

form (2.68). A simple computation gives α(t) = 12 v · x − 14 |v|2 t and h(t) = x − vt − x0 ,


which shows that (2.11) has the solutions of the form
1 1 2 t−λt)
u(x, t) := ei( 2 v·x− 4 |v| φ(x − vt − x0 ) (2.69)

where e−iλt φ(x) is a stationary solution to (2.11). This solution is obviously a traveling
wave. We have already encountered it above, in (2.19).
Put differently, the solution (2.19) comes from applying the Galilean symmetry trans-
formation (2.36) to the stationary solution e−iλt φ(x).

Manifolds of static solutions and traveling waves If the dynamical system (2.1)
with the symmetry group G has a static solution u∗ , then it has a manifold of static
solutions
M = {Tg u∗ : g ∈ G}.
Indeed, if ū is a static solution to (2.1), then, due to (2.35), T = Tg , so is Tg u∗ .
For example, since the Allen-Cahn equation (15.23) is translational invariant and has
(x−x0 )·e
the kink and anti-kink solutions, (2.4) and its negative, the functions χ±  (

2
) ≡
±χ( (x−x )·e
√ 0 ), are also solutions for any x0 ∈ R. These are kink and antikink solutions
2
centered at a. Thus the equation (15.23) has a family (manifold) of stationary solutions,
(x−x0 )·e
χ±
 (

2
) ≡ ±χ( (x−x
2
)·e
√ 0 ), x0 ∈ Rn , e ∈ S n−1 .

Exercise 2.9. Check the translation invariance of (15.23).

The same of course is true for traveling wave, as they can be reduced to static solution
when the equation is considered in the moving reference frame.

3 Fourier transform and partial differential equations


In this section, we describe one of the most powerful tools in analysis – the Fourier
transform. This transform allows us to analyze a fine structure of functions and to solve
differential equations. The Fourier transform takes functions of time to functions of
frequencies, functions of coordinates to functions of momenta, and vice versa.

3.1 Definitions and properties


Initially, we define the Fourier transform on the Schwartz space S(Rn ) = S:

S = {f ∈ C ∞ (Rn ) : hxiN |∂ α f (x)| is bounded ∀N and ∀α}, (3.1)


Lectures on Applied PDEs, November 3, 2017 27

where hxi = (1 + |x|2 )1/2 and α = (α1 , ..., αn ), with αj non-negative integers, ∂ α :=
Qn αj Pn ˆ
j=1 ∂xj and |α| = i=1 αj . On S, we define the Fourier transform F : f 7→ f by
Z
fˆ(k) := (2π)−n/2 f (x)e−ik·x dx. (3.2)

Define also the inverse Fourier transform of f (k) as


Z
ˇ
f (x) := (2π)−n/2
f (k)eix·k dk. (3.3)

Some key properties of the Fourier transform are collected in the following

Theorem 3.1. Assume f, g ∈ S(Rn ). Then we have:

(a) (−i∂)α f 7→ k α fˆ, and xα f 7→ (−i∂)α fˆ,

(b) f g 7→ (2π)−n/2 fˆ ∗ ĝ, and f ∗ g 7→ (2π)n/2 fˆĝ,

(c) (fˆ)ˇ= f = (fˇ)ˆ,

fˆĝ =
R R
(d) f g.

Properties (a) - (d) hold (possibly, with signs changed in (a)) also when ˆ is replaced by ˇ.

We give a formal proof of (a) and (b). For proofs of (c) and (d) see Appendix B.
Integrating by parts, we compute
Z
−n/2
−i(∂xj f )ˆ(k) = (2π) (−i)∂xj f (x)e−ik·x dx
Z
−n/2
= (2π) f (x)i∂xj e−ik·x dx

= kj fˆ(k).

Exercise 3.1. Prove the remaining relations in (a)

Now we prove the second relation in (b). Using e−ik·x = e−ik·(x−y) e−ik·y and changing
the variable of integration as x0 = x − y, we obtain
Z Z
[
f ∗ g (k) := (2π) −n/2
e−ik·x
( f (x − y)g(y)dy)dx
Z Z
= (2π)−n/2 ( e−ik·(x−y) f (x − y)dx) e−ik·y g(y)dy

= (2π)n/2 fˆ(k)ĝ(k).
28 Lectures on Applied PDEs, November 3, 2017

Exercise 3.2. Derive the first relation in (b) from the second one and (c).

A special case of Property (d) is the Planchrel theorem:

Z Z
ˆ2
|f | = |f |2 . (3.4)

The two key functions appearing often in applications as well as in theoretical research
are the gaussian e−x·Ax/2 and the power |x|−α . Fortunately, their transforms can be
computed explicitly (see Appendix B):
−1
F : e−x·Ax/2 7→ (det A)−1/2 e−k·A k/2 , (3.5)
cn,α |k|−n+α if α 6= n,

−α
F : |x| 7→ (3.6)
cn,n ln |k| if α = n.

The coefficients cn,α are given for α = 2 by

((2 − n)σn−1 )−1



for n 6= 2,
cn,2 = −1 −1 (3.7)
−(σn−1 ) = −(2π) for n = 2,

where σn is the volume of the n–dimensional unit sphere S n = {x ∈ Rn+1 : |x| = 1}.
Though it is easy to compute the Fourier transform of |x|−α , it is not easy to justify it.
Indeed, the function |x|−α is rather singular and definitely does not belong to S(Rn ).

3.2 Application of Fourier transform to partial differential equa-


tions
Our goal in this section is to apply the Fourier transform in order to solve elementary but
very basic partial differential equations (PDE’s).

The Poisson equation on Rn :


− ∆u = f, (3.8)
where u : Rn → R is an unknown function, f : Rn → R is a given function, and ∆ is the
Laplace operator (the Laplacian):
n
X ∂ 2u
∆u := .
j=1
∂x2j

The Poisson equation first appeared in the problem of determining the electric potential
u(x), created by a given charge distribution ρ(x) = f (x)/(4π). Since then, it came up in
various fields of mathematics, physics, engineering, chemistry, biology and economics.
Lectures on Applied PDEs, November 3, 2017 29

In order to solve the Poisson equation, we apply the Fourier transform to both sides
of (3.8) to obtain:
|k|2 û(k) = fˆ(k).
This equation can be easily solved: û = fˆ/|k|2 . We can now apply the inverse Fourier
transform to the last equality to get

u = ǧ ∗ f, where φ(k) = |k|−2 . (3.9)

But the inverse Fourier transform of φ(k) = |k|−2 is known:

[(2 − n)σn−1 ]−1 |x|−n+2 if n 6= 2



φ̌(x) =
[2π]−1 ln |x| if n = 2,

where σn is the volume of the unit–sphere Sn = {x ∈ Rn+1 : |x| = 1} in dimension n.


Explicitly, (3.9) can be written as

f (y)
Z
−1
u(x) = [(2 − n)σn−1 ] dy,
|x − y|n−2

for n 6= 2, and similarly for n = 2. In particular, for n = 3, we have the celebrated


Newton formula
1 f (y)
Z
u(x) = − dy.
4π |x − y|
Of course, the functions appearing in the above derivation are not necessarily from the
Schwartz space S and therefore these manipulations must be justified. We leave this as
an exercise, while proceeding in a similar fashion with other equations.

The heat equation on Rn :

∂u
= ∆u and u|t=0 = u0 , (3.10)
∂t
where u : Rnx × R+ t → R is an unknown function, and u0 : R → R is a given initial
n

condition. Problem (3.10) is called an initial value problem. It first appeared in the theory
of heat diffusion. In that case, u0 (x) is a given distribution of temperature in a body at
time t = 0, and u(x, t) is the unknown temperature–distribution at time t. Presently, this
equation appears in various fields of science, including mathematical modeling of stock
markets.
As before, we apply the Fourier transform to (3.10) and solve the resulting equation

∂ û
= −|k|2 û and û|t=0 = û0
∂t
30 Lectures on Applied PDEs, November 3, 2017

2 2
to get û = e−|k| t û0 . Applying the inverse Fourier transform, and using that (e−|k| t )ˇ =
2
(4πt)−n/2 e−|x| /(4t) , we obtain
u = g√2t ∗ u0 , (3.11)
2 /2
where gs (x) = s−n g(x/s) with g(x) = (2π)−n/2 e−|x| . In particular, u → u0 as t → 0, as
it should be.

The Schrödinger equation on Rn :


∂ψ
i = −∆ψ and ψ|t=0 = ψ0 . (3.12)
∂t
This is an initial value problem for the unknown function ψ : Rnx × R+
t 7→ C. Equation
(3.12) describes the motion of a free quantum particle. Proceeding as with the heat
equation, we obtain
ψ = g√2it ∗ ψ0 , (3.13)
2
where, recall, gs (x) = s−n ϕ(x/s) with g(x) = (2πt)−n/2 e−|x| /(2t) . Observe that this
formula can be obtained from (3.11) by performing the substitution t → t/i.

Exercise 3.3. Derive equation (3.13) using the Fourier transform.

The wave equation on Rn :

∂ 2u
= ∆u with u|t=0 = u0 and ∂t u|t=0 = u1 . (3.14)
∂t2
This is a second order equation in time and consequently, it has two initial conditions
u0 and u1 . The wave equation (3.14) describes various wave phenomena: propagation of
light and sound, oscillations of strings, etc. Proceeding as with the heat equation, we find

u = ∂t Wt ∗ u0 + Wt ∗ u1 , (3.15)

where Wt (x) is the inverse Fourier transform of the function sin(|k|t)/|k|. The latter can
be computed explicitly for n = 1, 2, 3:
 1
 2 χρ2 ≥0 for n = 1,
−1 −1
Wt (x) = (2π) ρ χρ2 ≥0 for n = 2,
(2π)−1 δ(ρ2 ) for n = 3,

where ρ2 := t2 − |x|2 , and χρ2 ≥0 stands for the characteristic function of the set {(x, t) ∈
R3+1 : ρ2 ≥ 0}, i.e.
1 if ρ2 ≥ 0,

χρ2 ≥0 =
0 otherwise,
Lectures on Applied PDEs, November 3, 2017 31

and δ(x) is the Dirac δ–function, a generalized function, or distribution.


Thus the dependence of W on x and t comes through the combination ρ2 = t2 − |x|2 ,
which is the Minkowski–distance in space–time, playing a crucial role in relativity. Observe
that χρ≥0 = χ|x|≤t and δ(ρ2 ) = (2t)−1 δ(t − |x|).
Exercise 3.4. Prove (3.15), and find Wt (x) for n = 1.
We examine closer the special case when n = 3 and u0 = 0. Then we get
1
Z
u = Wt ∗ u1 ≡ δ(|x − y| − t)u1 (y)dy
4πt
1
Z
= u1 (y)dS(y)
4πt S(x,t)
t
Z
= u1 (x + tz)dS(z),
4π S(0,1)

where S(x, t) = {y ∈ R3 : |y − x| = t} is a sphere of radius t centered at x. We see that


only the initial condition evaluated on the sphere S(x, t) matters in order to determine
the solution at time t and at position x. This is called the Huygens’ principle.

4 Local Existence for Key Evolution Equations


We address the problem of the short time existence of solutions for key evolution PDEs:
the reaction-diffusion (nonlinear heat), Hartree and nonlinear Schrödinger equations. In
all these case, the equation of interest can be written as

∂t u = Au + f (u), u|t=0 = u0 . (4.1)

where A is a linear operator: A = ∆ for the reaction-diffusion (nonlinear heat) equation


and A = i∆ for the Hartree and nonlinear Schrödinger equations, and f is a nonlinearity:
f (u) is a real function on R satisfying f (0) = f 0 (0) = 0 for the reaction-diffusion (nonlinear
heat) equation and f (u) = (v ∗ |u|2 )u, where v is a given function (potential), and f (u) =
|u|2 u for the Hartree and nonlinear Schrödinger equations, respectively. We will call f
the nonlinearity.
If f (u) = 0, then we arrive at the linear initial value problem

∂t u = Au, u|t=0 = u0 , (4.2)

which we tackled in Subsection 3.2. We have shown there that this equation has a unique
solution for every reasonable u0 . This solution is of the form u = g√2τ ∗ u0 , where τ = t, if
A = ∆, and τ = it, if A = i∆. We denote this solution by u(t) = etA u0 .
Thus, etA u0 := g√2τ ∗ u0 , where τ = t, if A = ∆, and τ = it, if A = i∆. The family
U (t) = etA , t ≥ 0, has the following properties (see Appendix C):
32 Lectures on Applied PDEs, November 3, 2017

(a) U (t) are bounded ∀t ≥ 0,

(b) U (0) = 1 and U (t + s) = U (t)U (s),

(c) ∂t U (t) = AU (t) = U (t)A.

Remarks. 1) A family of operators, U (t), satisfying (a) - (c) is called the evolution
semigroup or a propagator. Evolution semigroups are discussed in Appendix C.
2) The family Pt = et∆ , or sometimes its integral kernel, is called heat kernel and it
plays an important role in the theory of stochastic processes.
3) The semi-group Pt = et∆ has the following properties

(a) Pt is positivity improving, i.e. if f ≥ 0, then Pt f > 0,

(b) Pt 1 = 1.

Semi-groups with such properties are called stochastic semigroups. (The conditions f ≥ 0
and f > 0 can be stated for abstract vector spaces in terms of closed and open cones.)

4.1 Reduction to a fixed point problem


Duhamel principle. Consider the inhomogeneous linear initial value problem

∂t u = Au + f, u|t=0 = u0 , (4.3)

where A is either ∆ or i∆ and f = f (t) is a given function of x and t.

Proposition 4.1 (Duhamel principle). Provided the integral below exists, the solution,
u(t), of (4.3), is given by
Z t
tA
u(t) = e u0 + e(t−s)A f (s) ds. (4.4)
0

In opposite direction, the family u(t) given by (4.4), which is differentiable in t and is in
the domain of the operator A, satisfies (4.3).

Exercise 4.1. Prove this proposition. (Hint: For the first part, verify directly that (4.4)
satisfies (4.3). If v(t) := e−tA u(t) is well-defined, then one can prove the proposition by
showing that it satisfies the equation

∂t v = f, v|t=0 = u0

and then solving this equation by using the Fundamental Theorem of calculus.)
Lectures on Applied PDEs, November 3, 2017 33

Mild (weak) solutions. Consider the initial value problem (4.1). We apply (4.4) to
(4.1) to obtain Z t
tA
u(t) = e u0 + e(t−s)A f (u(s)) ds. (4.5)
0
If u(t) solves (4.1), then it also solves the equation (4.5). Conversely, if u(t) solves (4.5)
and is differentiable in t and twice differentiable in x, then it solves the equation (4.1).
If u(t) solves (4.5), but we do not know whether it is differentiable or not, we call u(t)
a mild (or weak) solution to (4.1).

Remark. There are several definitions of weak solutions depending on the methods
used. The above definition is adapted to the fixed point formulation of the local existence
problem for evolutions PDEs. Another common definition appears in the variational
approach (see Section 10 below) and states that u is a weak solution iff it satisfies the
equation Z Z Z Z
d
uA∗ g + f (u)g dd xdt

− u∂t gd xdt = (4.6)

for all g’s of compact support and which are differentiable in t and are in the domain of
(formally) adjoint operator A∗ .

Fixed point problem. Eq (4.5) can be written as the equation u = H(u), where
Z t
tA
H(u)(t) := e u0 + e(t−s)A f (u)(s) ds, (4.7)
0

called the fixed point equation or the fixed point problem. A solution of such an equation
is called a fixed point. In our next step we learn how to solve fixed point equations.

4.2 The contraction mapping principle


A key to dealing with a large class of equations is to reduce them to a fixed point problem,
H(u) = u, (4.8)
for some map H and then use one of several fixed point theorem stating existence and
uniqueness of solutions of the latter problem. (The equation (4.8) is called the fixed
point equation and its solution, a fixed point of the map H.) The most useful among these
theorems is the Banach contraction mapping principle which we now formulate and prove.
Let X be a Banach space. Denote by d(u, v) = ku − vk the distance between the
vectors u and v. We remark that actually all we need for the next theorem is that X is
a complete metric space (i. e. it does not have to have a norm). Let B be a closed set
in X. A map H : B → B is called a strict contraction if and only if there is a number
α ∈ (0, 1) s.t.
d(H(u), H(v)) ≤ α d(u, v), ∀u, v ∈ B.
34 Lectures on Applied PDEs, November 3, 2017

Theorem 4.2 (the contraction mapping principle). If H is a strict contraction in B,


then H has a unique fixed point in B.

Proof. We use the method of successive approximations to solve the equation u = H(u).
Pick some u0 ∈ B and define u1 = H(u0 ), . . . , un = H(un−1 ). Since H is a contraction,
un ∈ B.
We claim that {un } is a Cauchy sequence in X. Indeed, let n ≥ m, then

d(un , um ) ≤ αm d(un−m , u0 ).

Taking here n = m + 1, we find d(um+1 , um ) ≤ αm d(u1 , u0 ). Next, by the triangle


inequality (i.e. d(v, u) ≤ d(v, w) + d(w, u), ∀w ∈ X), we obtain

d(uk , u0 ) ≤ d(uk , uk−1 ) + d(uk−1 , uk−2 ) + · · · + d(u1 , u0 ).

Applying d(um+1 , um ) ≤ αm d(u1 , u0 ) to each term on the r.h.s. gives

αk−1 + αk−2 + . . . + 1 d(u1 , u0 )



d(uk , u0 ) ≤
1
≤ d(u1 , u0 ).
1−α

(If B is a bounded set in X, then we do not need the step above, since in this case kuk k
are uniformly bounded.) The last two inequalities imply that

αm
d(un , um ) ≤ d(u1 , u0 ) → 0 as m, n → ∞.
1−α

Thus {un } is a Cauchy sequence in X. Now since X is complete, un ∈ B and B is


closed, there is a u∗ ∈ B s.t. un → u∗ . Since d(H(un ), H(u∗ )) ≤ αd(un , u∗ ) → 0, we
have also that H(un ) → H(u∗ ). This and the equation un = H(un−1 ) imply that the
limit u∗ satisfies the equation u∗ = H(u∗ ). This demonstrates existence of a fixed point
in B, and we finish the proof by showing its uniqueness. Suppose that H(u∗ ) = u∗ , and
H(v∗ ) = v∗ . Then we have d(v∗ , u∗ ) = d(H(v∗ ), H(u∗ )) ≤ αd(v∗ , u∗ ). Since α ∈ (0, 1),
this gives d(v∗ , u∗ ) = 0 and so v∗ = u∗ .

Remark. As was mentioned above, there are many fixed point theorems and the Banach
contraction mapping principle is one of these, albeit the most useful one.
In the next two subsections, we use the contraction mapping principle to prove local
existence of solutions for the reaction-diffusion (nonlinear heat), Hartree and nonlinear
Schrödinger equations.
Lectures on Applied PDEs, November 3, 2017 35

4.3 Local existence for the nonlinear heat equation


We show local existence to the initial value problem for the nonlinear heat (or reaction-
diffusion) equation:
∂u
= ∆u + f (u), u|t=0 = u0 , (4.9)
∂t
on Rn . To fix ideas, for the nonlinearity f (u) we take f (u) = λ|u|p−1 u, with 1 < p < ∞,
and bounded initial conditions u0 on Rn . The special case of equation (4.9) without
nonlinearity first appeared in the theory of heat diffusion. In that case, u0 (x) is a
given distribution of temperature in a body at time t = 0, and u(x, t) is the unknown
temperature–distribution at time t. Presently, this equation appears in various fields of
science, including the theory of chemical reactions and mathematical modeling of stock
markets. Similar equations appear in the motion by mean curvature flow, vortex dynamics
in superconductors, surface diffusion and chemotaxis.
We will look for week solutions in space X := C([0, T ], Y ), with T specified later and
Y := L∞ (Rn ). The space L∞ is one of the standard Lp -spaces and is defined as (see
Appendix A.3 for more details)

L∞ (Ω) := {f : Ω → C | f is measurable, and ess sup |f | < ∞}, (4.10)

where, recall that ess sup |f | := inf{sup |g| : g = f a.e.}, with the norm defined as

kf k∞ := ess sup|f |. (4.11)

(Strictly speaking, elements of Lp (Ω) are equivalence classes of measurable functions: two
functions define the same elements of Lp (Ω) if they differ only on a set of measure 0.) We
often use the abbreviations Lp and kvkp for Lp (Ω) and kvkLp .

Theorem 4.3. Let u0 ∈ L∞ (Rn ) and T := p1 (1 − p1 )p−1 /(|λ|ku0 kp−1 ∞ ). Then the non-
p−1
linear heat equation (4.9) (with f (u) = λ|u| u, 1 < p < ∞) has a weak solution
u ∈ C([0, T ], L∞ ), satisfying kukC([0,T ],L∞ ) ≤ R, where R := ku1−0 k1∞ > ku0 k∞ , and unique
p
in the ball {u ∈ L∞ (Rn ) : kukC([0,T ],L∞ ) ≤ R}.

Proof. Using Duhamel’s principle, Eq (4.9) can be written as the fixed point equation
u = H(u), where Z t
t∆
H(u)(t) := e u0 + e(t−s)∆ f (u)(s) ds (4.12)
0

and we have written f (u)(s) for f (u(s)). Let Y := L∞ (Rn ) and X := C([0, T ], Y ), with
T specified later. The proof of existence and uniqueness will follow if we can show that
the map H has a unique fixed point in the ball

BR := {u ∈ X, kukX ≤ R},
36 Lectures on Applied PDEs, November 3, 2017

for some R > 0. We prove this statement via the contraction mapping principle.
We begin by proving that there is R > 0 s.t. H is a well-defined map from BR to BR .
First, we show the estimate t∆
e u ≤ kukY (4.13)
Y

We have shown
R above that the operator et∆ has the integral kernel pt (x, y),
R t > 0, i.e.
t∆
(e u)(x) = pt (x, y)u(y) dy, with the following properties: pt (x, y) > 0 and pt (x, y) dy =
1. Using these properties, we obtain the estimate (4.13).
Next, the elementary bound ||u|p−1 u| ≤ Rp , for |u| ≤ R, shows that, if t < T and
u ∈ BR , then
sup kf (w)kY = sup |f (w)| ≤ |λ|Rp (4.14)
kwkY ≤R |w|≤R

(remember that Y := L∞ (Rn )), which, together with the estimate (4.13), gives
Z t Z t
(t−s)A

e
≤ sup
f (u)(s) ds kf (u)(s)kY ds ≤ T |λ|Rp . (4.15)
0 t≤T X 0

Estimates (4.13) and (4.15) and the definition (4.12) of the map H imply that H : BR →
BR , provided
ku0 kY + T |λ|Rp ≤ R.
Now, we prove that H : BR → BR is a strict contraction. Recalling the definition f
and using the elementary bound

||u1 |p−1 u1 − |u2 |p−1 u2 | ≤ pRp−1 |u1 − u2 |,

for |u|, |u1 |, |u2 | ≤ R, we obtain, for u1 , u2 ∈ BR ,

sup kf (w1 ) − f (w2 )kY / kw1 − w2 kY ≤ |λ|pRp−1 , (4.16)


kw1 kY ,kw2 kY ≤R

which, together with the definitions of H and kukX and estimate (4.13), gives
Z t
kH(u1 ) − H(u2 )kX ≤ sup kf (u1 )(s) − f (u2 )(s)kY ds
t≤T 0
≤ T |λ|pRp−1 ku1 − u2 kX .

Therefore, if |λ|pRp−1 T < 1, then H is a strict contraction in BR . We see that the


inequalities
ku0 kY + |λ|T Rp ≤ R and pRp−1 |λ|T < 1
−1
are satisfied if we choose R so that pRp−1 = (R−ku0 kY )R−p , i.e. R := ku1−0 k1Y > ku0 kY ,
p

and T := p1 (1 − p1 )p−1 /(|λ|ku0 k∂−1


Y ). This gives the existence and uniqueness for the stated
R and T , as well as the estimate on the solution u.
Lectures on Applied PDEs, November 3, 2017 37

4.4 Local existence for Hartree and Schrödinger equations


In this subsection we prove the local existence of solutions for Hartree and Schrödinger
equations.

Local existence for the Hartree equation. In this subsection we show local existence
to the initial value problem for the Hartree equation
∂u
i = −∆u + (v ∗ |u|2 )u, u|t=0 = u0 , (4.17)
∂t
on Rn , where, as usual v∗f is the convolution of two functions, v∗g(x) := v(x−y)g(y)dy.
R

We assume that v is a ’nice’ function, i.e. sufficiently smooth and fast decaying at infinity.
Equation (4.17) arises in the problem in quantum physics of many-body systems.
Due to the physical interpretation of (4.17), we consider mild solutions in the space
L2 (Rn ). (See Appendix A.3 for a discussion of Lp − spaces.)
Theorem 4.4. Assume v ∈ L∞ . Then, for any u0 ∈ L2 (Rn ), the Hartree equation
−1
(4.17) has a unique solution u ∈ C([0, T ], L2 ), where T = 4 27ku0 k22 kvk∞ , satisfying
kukC([0,T ],L2 ) ≤ 23 ku0 k2 .
Proof. We proceed as in the proof of Theorem 4.3 for (4.9), but with Y = L2 (Rn ). Let
wt := eit∆ w. Passing to the Fourier transform and using the Plancherel theorem, we
2
obtain kwt k2 = kwbt k2 . Next, since wbt = e−i|k| t ŵ, we have kwbt k2 = kŵk2 = kwk2 , which
implies the estimate it∆
e w = kwk2 , (4.18)
2

uniformly in t. Next, using the elementary inequality kv ∗ gkp ≤ kvkp kgk1 , ∀p ≥ 1, we


obtain
(v ∗ |u|2 )u ≤ kvk∞ kuk32 .

2

Moreover, using the triangle inequality |(v∗|u1 |2 )u1 −(v∗|u2 |2 )u2 | ≤ |(v∗(|u1 |2 −|u2 |2 ))u1 |+
|(v ∗ |u2 |2 )(u1 − u2 )| and again the above inequality, we find
(v ∗ |u1 |2 )u1 − (v ∗ |u2 |2 )u2 ≤ 3kvk∞ (max kui k2 )2 ku1 − u2 k2 .

2 i

Now, proceeding exactly as in the proof of Theorem 4.3 for (4.9), we arrive at the state-
ment of the theorem.
Exercise 4.2. (a) Supply the details of the proof above. (b) Show the local existence
property for the Hartree equation, (4.17), in the Sobolev spaces H s (Rn ), s ≥ 0.
P
Exercise 4.3. Extend the above theorem from −∆ to an operator A = ij ∂xi aij ∂xj ,
where aij is a P smooth, symmetric real strictly positive definite matrix (in particular,
2
P
ij ξi aij ξj ≥ δ j |ξj | for some δ > 0).
38 Lectures on Applied PDEs, November 3, 2017

Local existence for nonlinear Schrödinger equation. Consider the initial value
problem for the nonlinear Schrödinger equation (or reaction-diffusion equation):
∂u
i = −∆u + λ|u|p−1 u, u|t=0 = u0 , (4.19)
∂t
with an initial condition u0 ∈ H s (Rn ). We assume p > 1 and s > 0. Equation (4.19)
arises in nonlinear optics, plasma physics, theory of water waves and in condensed matter
physics.
Proceeding as above, we find
Theorem 4.5. Let p be an odd integer. Then there are constants c, c0 > 0, s.t, for any
u0 ∈ H s (Rn ), with s > n/2, the nonlinear Schrödinger equation, (4.19) has a unique
solution u ∈ C([0, T ], H s ), where T = c/ku0 k2H s , satisfying kukC([0,T ],H s ) ≤ c0 ku0 kH s .

Exercise 4.4. Prove the above theorem. (Hint: Use Sobolev embedding theorems, e.g.
H s (Rn ) ⊂ L∞ (Rn ) for s > n/2, so that products of H s functions are again H s functions,
i.e. H s is an algebra.)
Techniques developed above apply to the generalized nonlinear Schrödinger equation
(gNLS) (cf. (4.19)),
i∂t ψ = −∆x ψ + g(|ψ|2 )ψ, (4.20)
where g(u) is a real function. We assume that f (ψ) = g(|ψ|2 )ψ satisfies the conditions
|f (k) (ψ)| . |ψ|p−k , ∀k = 0, ..., s + 1, (4.21)
with p < 1 + d4 .
Discussion. 1) Notice the difference in behaviour between the heat kernel Pt := et∆ and
the propagator Ut := eit∆ . The family et∆ is a one parameter semi-group, which is well
defined on the entire space, say Lp , only for t ≥ 0, while eit∆ is a one parameter group,
defined for all t.
While et∆ u Lp < kukLp , for u not identical 1, the characteristic property of the
propagator
R Ut is its unitarity. In particular it preserves the L2 − inner product, hu, vi :=
ūvdx: hu, Ut vi = hu, vi.
2) We can also show that the solutions u of the above equations depend continuously
on the initial condition u0 .

4.5 Classical solutions


One would like to show that mild (weak) solutions, i.e. solutions of Eq (4.5) in C([0, T ], Y )
are in fact classical ones of (4.1), i.e. the are once differentiable in t and twice in x in
appropriate sense. This is done by ‘bootstrapping’ estimates. We summarize the results
in
Lectures on Applied PDEs, November 3, 2017 39

Proposition 4.6. The local mild solutions of the NLH, Hartree and NLS equations (see
(4.1), with A and f described after this equation), i.e. solutions of Eq (4.5) in C([0, T ], Y ),
where Y is the corresponding Banach space, are in fact classical ones of (4.1).
Sketch of the proof of the NLH equation, (4.9). For the NLH equation, (4.9), Y := L∞ (Rn )
and we can start, for instance, by showing that u ∈ C([0, T ], W 1,∞ (Rn )), where, recall,
W s,p (Rn ) := {u ∈ Lp (Rn ) : ∂ α u ∀α, |α| ≤ s},
the Sobolev space. Differentiating (4.5), where A = ∆, we find
Z t
t∆
∂u(t) = ∂e u0 + ∂e(t−s)∆ f (u(s)) ds, (4.22)
0

where ∂ ≡ ∂xj for some j. Using R the explicit integral kernel representation (3.11) of
t∆/2 t∆/2
e u, we find (∂e f )(x) = ∂xj pt (x, y)f (y)dy, from which we estimate
k∂et∆/2 kL∞ →L∞ . t−1/2 , (4.23)
Using this estimate, one can show easily that ∂u ∈ C([0, T ], L∞ (Rn )) and therefore u ∈
C([0, T ], W 1,∞ (Rn )). Now, we can prove that u(t) is twice differentiable in x and once
differentiable in t. For example for x, we have
Z t
2 2 t∆
∂ u(t) = ∂ e u0 + ∂e(t−s)∆ ∂f (u(s)) ds, (4.24)
0

For the first term on the r.h.s. we use the estimate


k∂ 2 et∆/2 kL∞ →L∞ . t−1 , (4.25)
is obtained similarly to the one in (4.23) and for the term on the r.h.s., we estimate
Z t Z t
(t−s)∆
k ∂e ∂f (u(s)) dskL∞ ≤ k∂e(t−s)∆ ∂f (u(s))kL∞ ds (4.26)
0 0
Z t
. (t − s)−1/2 k∂f (u(s))kL∞ ds (4.27)
0
Z t
. (t − s)−1/2 kf 0 (u(s))kL∞ k∂u(s)kL∞ ds (4.28)
0
. sup kf 0 (u(s))kL∞ sup k∂u(s)kL∞ ds. (4.29)
s s

Now using that


k∂f (u(s))kL∞ . kf 0 (u(s))kL∞ k∂u(s)kL∞ (4.30)
. sup kf 0 (u(s))kL∞ sup k∂u(s)kL∞ (4.31)
s s

. sup ku(s)kLp−1
∞ sup k∂u(s)kL∞ , (4.32)
s s
40 Lectures on Applied PDEs, November 3, 2017

we conclude that
Z t
k ∂e(t−s)∆ ∂f (u(s)) dskL∞ →L∞
0
p−1
. kukC([0,T ],L∞ ) k∂u(s)kC([0,T ],W
1,∞ ) , (4.33)
which together with the relation (4.24) and the estimate (4.25) gives
k∂ 2 ukC([0,T ],L∞ ) . t−1 ku0 kL∞
+ kukp−1
C([0,T ],L∞ ) k∂u(s)kC([0,T ],W
1,∞ ) , (4.34)
as desired.
Exercise 4.5. Prove estimates (4.23) and (4.25).

5 Global Existence
5.1 A priori estimates
Let Y and Z be Banach spaces. Consider the abstract nonlinear evolution equation
∂t u = F (u), u|t=0 = u0 , (5.1)
where F is a map defined from an open set, U , in Y to Z. We assume
(LE) (Local existence) There is a positive, monotonically decreasing function f on R+ :=
(0, ∞), s.t. for any u0 ∈ U , (5.1) has a unique solution in C([0, T ], U ), with T =
f (ku0 kY ).
(AE) (A priori estimate) There is a positive function c on R+ := (0, ∞), s.t. c(s) ≥ s and
any solution u = u(t) of (5.1) satisfies the estimate
ku(t)kY ≤ c(ku0 kY ).

The estimate above is called a priori estimate. It allows one to iterate the local existence
argument and to obtain the existence on the infinite time interval, i.e. the global existence:
Proposition 5.1. Assume the (LE) and (AE). Then (D.2) has a global solution for every
initial condition u(0) = u0 ∈ Y .
Proof. Define T := f (c(ku0 kY )). Since f (ku0 kY ) ≥ f (c(ku0 kY )) = T , by Condition (LE),
(5.1) has a unique solution, u(t) ≡ u1 (t), on an interval [0, T ]. Now consider (5.1) with the
(1) (1)
new initial condition u0 = u( 43 T ). Since f (ku0 kY ) ≥ f (c(ku0 kY )) = T , it has a unique
solution, u(t) ≡ u2 (t), on the interval [ 43 T, 74 T ] and so on. This way we construct solutions
uj (t) on the intervals Ij , where Ij := [Tj , Tj +T ], with T1 = 0, Tj+1 = Tj + 43 T, j = 1, 2, . . . .
By the uniqueness uj (t) = uj+1 (t) on Ij ∩Ij+1 . Since the intervals Ij cover R, this gives the
solution, u(t), on R defined as u(t) ≡ uj (t) on the interval Ij for every j = 1, 2, . . . .
Lectures on Applied PDEs, November 3, 2017 41

Global existence of solutions of the Hartree equation. We demonstrate how


Proposition 5.1 works on the Hartree equation.

Proposition 5.2. Any solution to the Hartree equation (4.17) in C([0, ∞), H 2 )∩C 1 ([0, ∞), L2 )
satisfies
kψ(t)kL2 = kψ0 kL2 . (5.2)

Exercise 5.1. Prove this proposition.

Now, we use this conservation law and Propositions 4.6 and 5.1 to obtain

Theorem 5.3 (GWP). Assume v ∈ L∞ . Then, for any ψ0 ∈ L2 (Rn ), the Hartree
equation (4.17) has a unique mild solution ψ ∈ C([0, ∞), L2 ).

5.2 Energy and Entropy


For most of evolution equations appearing in natural sciences and especially in Physics
there are conserved quantities, like energy, and/or monotonic quantities, like the entropy.
Such quantities allow one to control the solutions at large times and, often, to prove the
global existence and stability.
In application to the equations we considered in the previous section, these are

1) Under the Hartree and (generalized) nonlinear Scrödinger evolutions (a) the L2 −norm
of solutions, Z
N (ψ) = |ψ|2 . (5.3)

(the ‘charge’, or the number of particles), and (b) the energy,

1
Z
E(ψ) = ( |∇ψ|2 + G(ψ)), (5.4)
2

where G(ψ) := 41 (v ∗ |ψ|2 )|ψ|2 and G(ψ) := p1 λ|ψ|p (for the generalized nonlinear
Schrödinger equation, G is an anti-derivative of g), respectively, are conserved,

E(ψ) = E(ψ0 ), N (ψ) = N (ψ0 ). (5.5)

2) Under the nonlinear heat (reaction-diffusion) evolution, (4.9), the energy

1
Z
E(u) = ( |∇u|2 + G(u)), (5.6)
2

where G(u) := p1 |u|p , decreases, as t increases: ∂t E(u) < 0.


42 Lectures on Applied PDEs, November 3, 2017

R
3) For the linear heat equation, ∂t u = ∆u, the entropy S(u) = u log u is decreasing:

Z
∂t S(u) = −4 |∇ u|2 < 0. (5.7)

• the (reduced) Keller-Segel equation of chemotaxis or nonlinear Fokker-Planck equa-


tion,
∂ρ
= ∆ρ − ∇ · (ρ∇c) ,
∂t (5.8)
0 = ∆c + ρ,

where ρ and c represent the organism density and chemical concentration (for the
Keller-Segel equation) or the mass density and gravitational potential (for the non-
linear Fokker-Planck equation). In this case, the ‘free energy’, defined by
 1
Z
− ρ(−∆)−1 ρ + ρ ln ρ dx,

F(ρ) = (5.9)
R2 2
R
is decreasing and the total mass, ρ, is conserved.
Exercise 5.2. Prove the above claims.
In the last subsection, we illustrated how this work on the Hartree equation on L2 , using
only the conservation of the L2 −norm. Using also energy, we can also show the global
existence of the Hartree equation on H 1 .
Proposition 5.4. We have
1
kψk2H1 ≤ 2E(ψ) + kvk∞ kψk42 + kψk2L2 . (5.10)
2
Proof. We estimate the term G(ψ) in (5.4), where G(ψ) := 14 (v ∗ |ψ|2 )|ψ|2 . Using the
R

elementary inequality kv ∗ gkp ≤ kvkp kgk1 , ∀p ≥ 1, we obtain


Z
| (v ∗ |ψ|2 )|ψ|2 | ≤ kv ∗ |ψ|2 k∞ kψk22 ≤ kvk∞ kψk42 . (5.11)

Combining this with (5.4) gives


1 1
Z
E(ψ) ≥ |∇ψ|2 − kvk∞ kψk42 (5.12)
2 4
1 1 1
= kψk2H1 − kvk∞ kψk42 − kψk2L2 . (5.13)
2 4 2
The last inequality gives (5.10).
Lectures on Applied PDEs, November 3, 2017 43

Using this result and the local existence of the Hartree equation on H 1 , one can show
the global existence of this equation on H 1 .
Similar proof could be also done for the generalized nonlinear Schrödinger equation
(gNLS). It is slightly more complicated. Instead of the trivial estimate (5.11), it uses the
more sophisticated Gagliardo-Nirenberg’s inequality
Z
 1
|ψ|p+1 p+1 ≤ CkψkaH1 kψk1−a
L2 , (5.14)

d p−1
where a = 2 p+1
. This proof can be found in the next subsection.

5.3 A priori estimates for the generalized NLS and Keller-Segel


equations
A priori estimates for solutions of the Keller-Segel equations. The proof of the
global existence for the Keller-Segel equations relies on a priori estimates of its entropy.
We start with computing the change in the entropy
√ 2
Z Z Z
∂t ρ ln ρ = −4 |∇ ρ| + ρ2 . (5.15)
| {z } | {z }
entropy dissipation entropy production

Depending on whether the entropy dissipation or the entropy production wins we expect
either dissipation of the solution or the collapse (blowup). The Nirenberg - Gagliardo
inequality,
kf k24 ≤ cgn k∇f k2 kf k2
shows that the dissipation wins if M c2gn ≤ 4.
To sharpen this result one uses that the free energy decreases together with the loga-
rithmic Hardy-Littlewood-Sobolev inequality,
1 1
Z Z
f ln f ≥ f (−∆)−1 f − C(M ),
(M/8π) 2
R
where M := f (the dimension n = 2) and C(M ) := M (1 + log π − log M ), which gives
the following lower bound on the free energy,
1 1
Z
F(ρ) ≥ ( − 1) ρ(−∆)−1 ρ − C(M ).
(M/8π) 2
Combining this inequality together with the fact that the free energy is decreases, F(ρ) ≤
F(ρ0 ), one finds the bound on the entropy
1
Z
(1 − M/8π) ρ ln ρ ≤ F(ρ0 ) − C(M ),

which is used to prove the global existence for M ≤ 8π.
44 Lectures on Applied PDEs, November 3, 2017

Global existence of solutions of the generalized NLS. We address the question


of global existence, i.e., existence for all t ≥ 0 of solutions of the generalized NLS (gNLS)
given in (4.20), which we reproduce here,
i∂t ψ = −∆x ψ + g(|ψ|2 )ψ. (5.16)
Here g(u) is a real function, s.t. f (ψ) = g(|ψ|2 )ψ satisfies the conditions (4.21). If G ≥ 0
(the non-focusing nonlinearity), then the conservation of energy, implies
k∇x ψ(t)kL2 ≤ E(ψ) = E(ψ0 ).
Using, in addition, the conservation of charge or the number of particles, N (ψ) = N (ψ0 ),
we find
kψ(t)kH 1 ≤ E(ψ0 ) + N (ψ0 ).
Hence ψ(t) is uniformly bounded in the Sobolev norm. Thus if we have the local existence
result of the type of Theorem 4.5 in the Sobolev space H 1 (Rn ), then we also have the
global one.
In the more difficult case G ≤ 0 (the focusing nonlinearity), we have to impose addi-
tional conditions on the growth of nonlinearity. We follow [20].
d+2
Theorem 1 (GWP). Assume (4.21) with 1 ≤ p < d
and s = 1. Then (5.16) has a
unique solution in H1 (Rd ) for all t ≥ 0.
We will derive this theorem from the following
2c
Proposition 5.5. Let p < 1 + d4 . Define the function J(h, n) := Cd,p (h + n + n 2−b ), where
b := d(p−1)
2
< 2 and c := p + 1 − d(p−1)
2
. Then
kψk2H1 ≤ J(H(ψ), N (ψ)). (5.17)
Proof. For f (ψ) = g(|ψ|2 )ψ satisfying (4.21) and F (ψ) = G(|ψ|2 ), we have
|F (ψ)| ≤ c|ψ|p+1 (5.18)
We have by the Gagliardo-Nirenberg’s inequality
Z
 1
|ψ|p+1 p+1 ≤ CkψkaH1 kψk1−a
L2 , (5.19)

d p−1
where a = 2 p+1
. Then
Z
|F (ψ)| ≤ CkψkbH1 kψkcL2 (5.20)
d(p−1) d(p−1)
where b := 2
and c := p + 1 − 2
. This gives
1 1 1
Z
E(ψ) ≥ |∇ψ|2 − CkψkbH1 kψkcL2 = kψk2H1 − CkψkbH1 kψkcL2 − kψk2L2 . (5.21)
2 2 2
Lectures on Applied PDEs, November 3, 2017 45

Now, since p < 1 + d4 , we have that b < 2 and therefore


1 2c
kψk2H1 − CkψkbH1 kψkcL2 ≥ −C 0 kψkL2−b
2 . (5.22)
4
The last two inequalities give (5.17).
Proof of Theorem 1. Thus if ψ is a solution to (5.16) with the initial condition ψ0 we have
by the conservation of energy and number of particles that

kψkH1 ≤ M0 , where M0 := J(H(ψ0 ), N (ψ0 )). (5.23)

Now the global existence follows from the local existence result of the type of Theorem
4.5 and Propositions 4.6 and 5.1.
We summarize some results on the relation between the global existence and functional
inequalities.

Relation between the global existence and functional inequalities


• the critical nonlinear Schrödinger equations vs the Nirenberg - Gagliardo inequality,
Z
 1
|ψ|p+1 p+1 ≤ CkψkaH1 kψk1−a
L2 , (5.24)

d p−1
where a = 2 p+1
;

• the fast diffusion equation, ∂t u = ∆um , 0 < m < 1, vs the Hardy-Littlewood-


Sobolev inequality,
f (x)f (y)
Z Z
dxdy ≤ kf k2p ;
|x − y|
• the (reduced) Keller-Segel or nonlinear Fokker-Planck equation,
∂ρ
= ∆ρ − ∇ · (ρ∇c) ,
∂t (5.25)
0 = ∆c + ρ,

with ρ and c describing the organism density and chemical concentration, vs the
logarithmic Hardy-Littlewood-Sobolev inequality,
2
Z Z Z
− f (x)f (y) log |x − y|dxdy ≤ − f log f + C(M ),
M
R
where M := f (the dimension n = 2).
46 Lectures on Applied PDEs, November 3, 2017

6 The implicit function theorem and applications


6.1 Gâteaux and derivative
Our goal is is to develop a differential calculus of maps between Banach spaces. Let X
and Y be Banach spaces and M , an open subset of X. We consider a map F : M → Y .
The map F is called Gâteaux differentiable at u ∈ M if and only if, for any ξ ∈ X, the
following limit exists

dF (u)ξ := F (u + λξ), (6.1)
∂λ λ=0

and gives a bounded, linear map dF (u) : X → Y . The map dF (u) is called the Gâteaux
derivative or sometimes, the variational derivative.

Remark. One can show that the existence of (6.1) for all ξ ∈ X implies that dF (u) :
X → Y is a bounded, linear map, but this is not easy.
The map F is called continuously differentiable, or C 1 (written F ∈ C 1 ) in an open
set U ⊂ M if and only if it is Gâteaux differentiable for very u in U , and u 7→ dF (u) is
a continuous map from U ⊂ X to the space L(X, Y ) of bounded operators from X to Y ;
i.e. if un → u in X, then dF (un ) → dF (u) in L(X, Y ).

Example 6.1 (formal). 1) If F (u) = Lu, where L is a linear map, then dF (u) = L
∂ ∂
(independently of u). Indeed, dF (u)ξ = ∂λ L(u + λξ)|λ=0 = ∂λ (Lu + λLξ)|λ=0 = Lξ.
1
Thus if L is bounded, then F is C .

2) If F (u) = f ◦ u (composition map), for a fixed C 1 –function f : R → R, and


u : Rn → R, then dF (u) is the multiplication operator by f 0 (u). Indeed, dF (u)ξ =

∂λ

F (u+λξ)|λ=0 = ∂λ f (u(x)+λξ(x))|λ=0 = f 0 (u)ξ. So if f 0 (u) is a bounded function,
say for some u ∈ Lp (Rn ), then F : Lp (Rn ) → Lp (Rn ) is differentiable at u.

Exercise 6.1. Compute dF (u) for F : Rn → Rm , and for the mean curvature map
!
∇u
F (u) = div p .
1 + |∇u|2

Exercise 6.2. Let Ω be a bounded domain in Rn with a smooth boundary and f ∈


C k+1 (R). Show that
(a) the map F (u) = f ◦ u satisfies F : C k (Ω) → C k (Ω), and F is C 1 with dF (u)ξ =
0
f (u)ξ and
(b) the map F (u) = −∆u + f ◦ u satisfies F : C k (Ω) → C k−2 (Ω), and F is C 1 with
dF (u)ξ = (−∆ + f 0 (u))ξ.
Lectures on Applied PDEs, November 3, 2017 47

As usual, the symbol o(kξk) stands for a map R : X → Y s.t.

kRk
→ 0, as kξk → 0.
kξk
The following properties of the Gâteaux derivative play an important role in applications.

Theorem 6.1. (a) (The fundamental theorem of calculus) If F is Gâteaux differentiable


in M , then we have
Z 1
F (u + ξ) − F (u) = dsdF (u + sξ)ξ. (6.2)
0

(b) (The chain rule) If F and G are Gâteaux differentiable at u ∈ M and F (u) ∈ F (M ),
respectively, then we have d(G ◦ F )(u) = dG(F (u))dF (u).

(c) Let K be a convex subset of a Banach space X (i.e. if u, v ∈ K, then su + (1 − s)v ∈


K, for all s ∈ [0, 1]). If F : K → K satisfies kdF (ψ)k ≤ α, ∀ψ ∈ K, then F is
Lipschitz:
kF (ψ) − F (ϕ)k ≤ αkψ − ϕk, ∀ψ, ϕ ∈ K.

(d) If F is C 1 at u ∈ X, then as kξk → 0

F (u + ξ) = F (u) + dF (u)ξ + o(kξk). (6.3)

Proof. For u, ξ ∈ X fixed, define the function g : [0, 1] → Y by

g(t) = F (u + tξ),

According to the definition of the Gâteaux derivative (6.1), we have

g 0 (t) = dF (u + tξ)ξ.

By Fundamental Theorem of Calculus,


Z 1
g(1) − g(0) = g 0 (t)dt. (6.4)
0

Since g 0 (t) = dF (u + tξ)ξ, this gives the first statement.


Exercise 6.3. Show the properties (b) and (c).
For (d), using (6.2) and writing
Z 1
F (u + ξ) − F (u) − dF (u)ξ = ds(dF (u + sξ) − dF (u))ξ, (6.5)
0
48 Lectures on Applied PDEs, November 3, 2017

we find

kF (u + ξ) − F (u) − dF (u)ξk
Z 1
≤ sup kdF (u + sξ) − dF (u)k kξk ds1.
0<s<1 0

Since dF (u) is continuous in u, we have kdF (u + sξ) − dF (u)k = o(1) in ξ, which gives
(6.3).

Remark 6.2. If F is C 1 , then it is easy to show that dF is linear. Indeed, set u = 0, for
simplicity of notation. (d) ( (6.3)) implies that F (ξ +η)−F (0) = dF (0)(ξ +η)+o(kξ +ηk).
On the other hand, F (ξ + η) − F (0) = F (ξ + η) − F (ξ) + F (ξ) − F (0) = dF (ξ)η +
o(kηk) + dF (0)ξ + o(kξk) = dF (0)ξ + dF (0)η + o(kξk) + o(kηk). These two relations imply
dF (0)(ξ + η) = dF (0)ξ + dF (0)η.

We conclude this section with some useful rigorous results about Gâteaux derivatives
of composition operators F (u) = f ◦ u, where f is a fixed function and u belongs to
the space of differentiable functions. Such operators appear often in applications. The
statements below are useful in this context. An important result in this direction is the
following.

Proposition 6.3. Let Ω ⊂ Rn . Let F (u) = f ◦ u with f ∈ C 2 (R) and obeying the
estimates

|f (k) (u)| ≤ c|u|p−k for k = 0, 1, 2, (6.6)

for some p ≥ 2. Then F : H r (Ω) → L2 (Ω) and is C 1 , provided r > n/2.

Proof. Let u ∈ H r (Ω). Then by the Sobelev embedding theorem (see Section A.6) u ∈
L2(p+1) (Ω) for r > n2 − 2p
n
, and kukL2p . kukH r . Hence, by (6.6), with k = 0, we have
Z Z
kF (u)k2L2 = 2
|f (u)| ≤ c |u|2p (6.7)
Ω Ω
= ckuk2p
L2p ≤ c kuk2p
0
Hr . (6.8)

This shows that F : H r (Ω) → L2 (Ω). To show that F has a Gâteaux derivative, we
compute formally dF (u)ξ = f 0 (u)ξ. By (6.6), with k = 1, and the Sobelev embedding
theorem, we have f 0 (u) ∈ L∞ (indeed, kf 0 (u)kL∞ . kukp−1 p−1
L∞ . kukH r , provided r > n/2)
and therefore dF (u) is a bounded operator.
To show that F ∈ C 1 , we use the mean value theorem to estimate

kdF (u0 ) − dF (u)kL2 ≤ kf 0 (u0 ) − f 0 (u)kL∞ (6.9)


≤ kf 00 (ū)kL∞ ku0 − ukL∞ , (6.10)
Lectures on Applied PDEs, November 3, 2017 49

where ū := θu0 + (1 − θ)u for some θ ∈ [0, 1]. Using this, (6.6), with k = 2 and the Sobolev
embedding theorem, we find, for r > n/2,

kdF (u0 ) − dF (u)kL2 . kukp−2 0


L∞ ku − ukH r . (6.11)

This shows that F ∈ C 1 .

Corollary 6.4. Assume f : C → C satisfies estimates (6.6) with 1 ≤ p < 2


(n−2r)+
,r > 0.
Define F (u) = −∆ + f (u). Then F : H r (Ω) → L2 (Ω) and is C 1 .

Exercise 6.4. Prove that, for r > n/2, under the conditions of Proposition 6.3, F = f ◦u
(a) maps H r (Ω) into H 1 (Ω) and (b) is Gâteaux differentiable, as a map from H r (Ω) into
L2 (Ω).

Furthermore, we have

Theorem 6.5. Let F (u) = f ◦ u, Ω ⊂ Rn be a bounded domain with smooth boundary,


and let r > n/2. If f ∈ C r+1 (R), then F : H r (Ω) → H r (Ω), and F is C 1 .

Exercise 6.5. Prove this theorem for n = 1 and r = 1.

For a complete proof of Theorem 6.5, see [12], page 221.

Example 6.2. For an exercise, we derive directly relation (6.3) for the map F (u) = g ◦ u.
We have
f (u + ξ) − f (u) = f 0 (u)ξ + R(ξ)
R1
where R(ξ) = 0 (f 0 (u + tξ) − f 0 (u))ξ dt. Now we estimate by the mean value theorem
Z 1
||R(ξ)||2 ≤ kf 00 (u + t̄ξ)ξ 2 k2 t dt
0

for some 0 ≤ t̄ ≤ t. Using the estimate |f 00 (u)| ≤ c|u|p−2 and the triangle and Hölder
inequalities we derive furthermore

Z 1
kR(ξ)k2 ≤ c k|u + t̄ξ|p−2 ξ 2 k2 t dt
0
k|u|p−2 ξ 2 k2 + kξ p k2

.
kukp2p kξk22p + kξkp2p .


Hence kR(ξ)k2 → 0 and therefore kf (u + ξ) − f (u) − f 0 (u)ξk2 → 0 as kξk → 0. This


shows (6.3) and consequently F is C 1 .
50 Lectures on Applied PDEs, November 3, 2017

6.2 Appendix: The Fréchet derivative


Though the Gâteaux derivative is straightforward to compute, for theoretical considera-
tions, one needs often a stronger notion of derivative: the Fréchet derivative. Before we
define the Fréchet derivative, let us remark that equation (6.1) is equivalent to

F (u + λξ) − F (u) = λdF (u)ξ + o(λ), (6.12)

were o(λ) is a vector in Y satisfying limλ→0 ko(λ)k/λ = 0. Notice that in general, o(λ)
depends on ξ.
The map F is called Fréchet differentiable at u ∈ X if and only if there exists a
bounded linear map dF (u) ∈ L(X, Y ) s.t. (6.3) holds as kξk → 0. The operator dF (u)
satisfying (6.3) is called the Fréchet derivative of F at the point u.
From this definition and equation (6.12), it is clear that if F is Fréchet differentiable at
u, with Fréchet derivative dF (u), then F is Gâteaux differentiable at u with the Gâteaux
derivative given by the same operator dF (u). In the opposite direction, if F ∈ C 1 , then
the preceding theorem implies

Theorem 6.6. If F is continuously Gâteaux differentiable at u ∈ X, with Gâteaux deriva-


tive dF (u), then F is Fréchet differentiable at u, and the Fréchet derivative is given by
the same operator dF (u).

For a detailed discussion of Fréchet and Gâteaux derivatives, we refer to [1, ?].
In everything that follows, by the derivative dF (u) we understand the Gâteaux deriva-
tive. We point out that in most of our applications, we deal with C 1 maps, in which case
the Fréchet and Gâteaux derivatives coincide, according to the last theorem.

6.3 The implicit function theorem


The implicit function theorem is a key theorem in analysis. Consider three Banach spaces
X, Y and Z and a map F : M × U → Z, where M and U are open sets in X and Y ,
respectively. We wish to solve the equation

F (µ, u) = 0

for u, i.e. we want to find a function u = g(µ) s.t. F (µ, g(µ)) = 0. For a map F (µ, u) de-
pending on two arguments, µ and u, we introduce the notion of partial Gâteaux derivative,
say du F (µ, u), by fixing u and taking the Gâteaux derivative in y.

Theorem 6.7 (The implicit function theorem). Assume

(1) F : M × U → Z is C 1 in y and C in µ;

(2) F (a, b) = 0, where a ∈ M and b ∈ U ;


Lectures on Applied PDEs, November 3, 2017 51

(3) du F (a, b) has a bounded inverse.


Then there is a neighborhood M 0 ⊂ M of a ∈ X and a map g : M 0 → Y such that
F (µ, g(µ)) = 0, ∀µ ∈ U 0 , and g(a) = b.
First we explain the idea of the proof. Without loss of generality, we can take (a, b) =
(0, 0). We want to solve F (µ, u) = 0 for y near (µ, u) = (0, 0). Expand F in u around 0
(see Theorem 6.1): F (µ, u) = F (µ, 0) + du F (µ, 0)u + R(µ, u), with R(µ, u) = o(||u||). So
we can rewrite this equation we want to solve as the following equation for u:
F (µ, 0) + du F (µ, 0)u + R(µ, u) = 0.
We will show in a moment that, since the operator L0 = du F (0, 0) is invertible, then so
is Lµ = du F (µ, 0) for sufficiently small kµk. Hence the equation above can be rewritten
as the fixed point equation,
u = −L−1
µ (F (µ, 0) + R(µ, u)) , (6.13)
or Hµ (u) = u, where Hµ (u) is the map given by
Hµ (u) := −L−1
µ (F (µ, 0) + R(µ, u)) . (6.14)
If we neglect the remainder term R(µ, u), then equation (6.13) yields for each given x the
corresponding u = G(µ) = −L−1 µ F (µ, 0). In the general case, the remainder is not zero,
but o(kuk) for u small, and we can use the fixed point argument to show existence of a
solution to (6.13). To do so,
We now show that the map Hµ is well defined and has a fixed point u = u(µ). First
we show that there is a δ1 such that Lµ is invertible for kµk ≤ δ1 . Write Lµ = L0 + Vµ ,
where Vµ := Lµ − L0 . Since Lµ is continuous in x by the conditions of the theorem we can
choose δ1 such that kVµ k ≤ (2kL0−1 k)−1 if kµk ≤ δ1 . Hence Lµ is invertible if kµk ≤ δ1 by
Theorem A.2 in Section A.7.
Denote by BZ (z, r) the open ball in Z of radius r centered at z ∈ Z.
Claim 6.1. ∃ > 0 and δ > 0 such that ∀µ ∈ BX (0, δ)

(i) Hµ : BY (0, ) → BY (0, ),

(ii) kdHµ (u)k ≤ 1/2 ∀u ∈ BY (0, ).


Given (i) and (ii), we see that for any µ ∈ BX (0, δ), Hµ is a contraction map and
therefore it has a unique fixed point in BY (0, ). Call this fixed point u = g(µ). It solves
the equation u = Hµ (u) which is equivalent to the equation F (µ, u) = 0. Thus, the
theorem follows with M 0 = BX (0, δ).
It remains to prove the claim. (i) First, we pick δ > 0 so that for m = 2kL−1 0 k and
given  > 0 and for any µ ∈ BX (0, δ), we have
52 Lectures on Applied PDEs, November 3, 2017

(a) kL−1
µ k ≤ m;

(b) kF (µ, 0)k ≤ /2m.


This is possible by the continuity of du F (µ, 0)−1 (due to the continuity of du F (µ, 0), see
above) and of F (µ, 0) in µ and the relation F (0, 0) = 0. Next, we pick  > 0 so that, for
any µ ∈ BX (0, δ) and any u ∈ BY (0, ),
(c) kR(µ, u)k = o(kuk) ≤ /2m;
(d) kdu F (µ, u) − du F (µ, 0)k ≤ 1/2m.
Again this is possible by a property of R(µ, u) and the continuity of du F (µ, u).
Now, inequalities (a)-(c) and the definition of Hµ imply
kHµ (u)k ≤ m (/2m + /2m) = 
∀µ ∈ BX (0, δ) and u ∈ BY (0, ). So (i) follows.
(ii) Using the definition, R(µ, u) := F (µ, u) − F (µ, 0) − du F (µ, 0)u, of the remainder
R(µ, u) and definition (6.14) of the map Hµ , we compute
du Hµ (u) = −du F (µ, 0)−1 du R(µ, u)
= −du F (µ, 0)−1 [du F (µ, u) − du F (µ, 0)] .
Applying inequalities (a) and (d), we conclude that kdu Hµ (u)k ≤ 1/2 for any µ ∈ BX (0, δ)
and any u ∈ BY (0, ). This proves (ii) and therefore completes the proof of the claim and
with it, the theorem. 

Remark. To prove that some linear operator is invertible is usually a difficult task.
It is significantly simplified if we know that the operator in question is self-adjoint (see
Appendix A.8 and [6] for the definition). Then it amounts to showing that 0 is not in
the spectrum of the operator and there are many techniques to find spectra of self-adjoint
operators (see Appendix E and [6]).

6.4 Existence of breathers and constant mean curvature sur-


faces
Existence of breathers. Consider the discrete nonlinear Schrödinger equation (DNLS)
∂ψ
i = −∆ψ − |ψ|2 ψ
∂t
in l2 (Zd ). Here Zd = {α = (α1 , . . . , αd ) | αj integers}, l2 (Zd ) = {ψ : Zd → C | |ψ(α)|2 <
P
x∈Zd
∞} and ∆ is the discrete Laplacian
X
(∆f )(α) = f (β) − 2df (α). (6.15)
|α−β|=1
Lectures on Applied PDEs, November 3, 2017 53

Remark. The discrete Laplacian can be also defined as follows. Let E := {ej = (0, ..., 1, ..., 0)|j =
1...d} be the basis of the elementary cell at the origin in Zd . Define ∆ := − div ·∇ where
X
(∇f )(α) = (f (α + e) − f (α))e
e∈E

and div (the divergence) is the adjoint negative of operator, div = −∇∗ , i.e., h∇f, ~g i =
−hf, div ~g i. Hence X
(div ~g )(α) = (ge (α) − ge (α − e))
e∈E

Example 6.3. For d = 1, (∇f )(α) = f (α + 1) − f (α), (div g)(α) = g(α) − g(α − 1) which
implies ∆f (α) = f (α + 1) + f (α − 1) − 2f (α).
Exercise 1. 1) Derive (6.15) from the formula ∆ := −div · ∇ and, separately, from the
result of Example 6.3 (applying the latter in each coordinate). 2) Show ∆ is a bounded
operator on l2 (Zd ).
Breathers are time-periodic solutions to DLNS of the form
ψ(α, t) = eiλt φ(α)
This implies that φ(α) satisfies
− ∆φ − |φ|2 φ + λφ = 0 (6.16)
Theorem 6.8. If  is sufficiently small and λ > 0, then (6.16) has a solution in l2 (Zd ).
Proof. We look for real solutions φ to (6.16). Define F (, φ) := −∆φ + λφ − |φ|2 φ. Note
that (6.16) ⇐⇒ F (, φ) = 0. Therefore, we have to solve F (, φ) = 0 for φ, provided
 sufficiently small. We want to use IFT. Note that 1) F (, φ) : R × l2 (Z√ d
) → l2 (Zd ) is
continuous in  and φ (since l2 ⊂ lp ∀p ≥ 2); 2) F (0, φ0 ) = 0, where φ0 = λ δα0 for any
α0 ; 3) F (, φ) is C 1 in φ: dφ F (, φ) = −∆ + λ + 3φ2 is a bounded operator on l2 (Zd ), and
is continuous in  and φ; 4) dφ F (0, φ0 ) = λ(1 − 3δα0 ) is invertible if λ > 0. By 1) − 4),
IFT is applicable and hence (6.16) has a unique solution for  sufficiently small.
Remark. One can extend the above theorem to large : let F (, φ ) = 0 for  small;
dφ F (, φ ) = −∆ + λ − 3φ2 is invertible if
minα |λ − 3φ2 |
λ∈
/ Ran 3φ and < . (6.17)
k∆k
Exercise 6.6. Show that if (6.17) holds, then dφ F (, φ ) = −∆ + λ − 3φ2 is invertible.
Hint: Write
−∆ + λ − 3φ2 = λ − 3φ2 1l − (λ − 3φ2 )−1 ∆ .
 

/ Ran 3φ , then the first factor on the r.h.s (the multiplication operator λ − 3φ2 ) is
If λ ∈
invertible. Show that the second factor is invertible as well.
54 Lectures on Applied PDEs, November 3, 2017

Existence of surfaces with prescribed mean curvature. Assume S is a hypersur-


face in Rn+1 , given as a graph of a function ψ : Ω ⊂ Rn → R, S = graph ψ (see Fig. 6).
We assume Ω is a bounded domain with smooth boundary.
The mean curvature of S at a point x0 = (x, ψ(x)) ∈ S is given, for x ∈ Ω, by (see

xn+1

S = graph

,
,
, ⌦
,

Figure 6: Graph of ψ

Appendix 6.6) !
∇ψ(x)
H(x0 ) ≡ H(ψ)(x) = div p . (6.18)
1 + |∇ψ(x)|2
Given a function h(x0 ) on Ω, is there a surface S = graph ψ which has mean curvature
h(x) at points x0 = (x, ψ(x))? This problem can be reformulated as finding a solution ψ
to the equation H(ψ) = h, i.e.
!
∇ψ(x)
div p = h(x), (6.19)
1 + |∇ψ(x)|2
R
on Ω. (6.19) implies that Ω h = 0. We sketch a proof that
R
• for any h ∈ H r−2 (Ω), r > n/2 + 1, sufficiently small and satisfying Ω h = 0, there
is a unique ψ ∈ H r (Ω), ψ ∂Ω = 0 solving equation (6.19).
To solve the equation H(ψ) = h, we apply the implicit (or inverse, see the next
subsection) function theorem to the equation F (h, ψ) := H(ψ) − h. So we need to show
that for some spaces X, Y and Z we have
1) F : M × U → Z, where M and U are a neighborhoods of 0 ∈ X and 0 ∈ Z,
Lectures on Applied PDEs, November 3, 2017 55

2) F is C 1 in u and C in h,
3) dF (0, 0) has a bounded inverse.
r r−2
For the spaces X, Y and Z we take the Sobolev spaces
R X = H0 (Ω), Y = Z = H⊥ (Ω),
s s
with r > n/2 + 1, whereR H⊥ (Ω) :=:= {u ∈ H (Ω) : Ω u = 0}.
s
(The subindex ⊥ refers
s
to
the condition h1, ui = Ω u = 0. Recall, that informally, H0 (Ω) := {u ∈ H (Ω) : u ∂Ω =
0}.)
To show 1), we first observe that by the Sobolev embedding theorem (see Theorem
A.1), ∇j ψ are continuous and bounded functions. (Here we use the condition r > n/2+1.)
Furthermore, differentiating through we have
∆ψ ∇ψ · ∇2 ψ∇ψ
H(ψ) = − .
(1 + |∇ψ|2 )1/2 (1 + |∇ψ|2 )3/2
Assuming for simplicity n = 1 and r = 2, we see that, since ∇j ψ are continuous and
bounded functions and ∆ψ, ∇2 ψ ∈ L2 , the function H(ψ) is in L2 , which shows 1).
In order to check 2), i.e. F ∈ C 1 , we remember from Exercise 6.1 #1 that
∇ξ (∇ψ · ∇ξ)∇ψ
 
dH(ψ)ξ = div − .
(1 + |∇ψ|2 )1/2 (1 + |∇ψ|2 )3/2
Exercise 6.7. Let n = 1 and r = 2. Show dH(ψ) : H r (Ω) → H r−2 (Ω) is bounded and
continuous in ψ (i.e. kdH(ψ 0 ) − dH(ψ)k → 0, as kψ 0 − ψk → 0).
Finally, we have to verify that 3) is satisfied, i.e. that dF (0, 0) = dH(0) has a bounded
inverse. Now dH(0) = ∆. For Ω = Rn , we have discussed the existence of ∆−1 in
Appendix B. For a general bounded domain, Ω, analysis is more subtle and involves
some non-trivial spectral theory (see Appendix E and e.g. [6, 8]). The upshot of it is
that ∆−1 : H⊥r−2 (Ω) → H0r (Ω) exists and is bounded. (On H r (Ω), the operator ∆ has
the eignevalue 0 with constant eigenfunctions and therefore is not invertible unless we
consider the orthogonal complement of the function 1.)
Modulo the proof of the fact above (we might do this in a later section), we
have thus shown that the conditions of the implicit function theorem are satisfied, and
therefore, for any sufficiently small h ∈ H r−2 (Ω), r > n/2 + 1, the equation F (ψ) = h has
a unique solution ψ ∈ H0r (Ω). In other words, there exists a surface S = graph ψ with
prescribed small mean curvature h.
If h is constant, then the corresponding surface is called the constant mean curvature
surface. For h = 0, this is a minimal surface.

6.5 Inverse function theorem


The implicit and inverse function theorems easily follow from one to the other. For the
inverse function theorem, we have to solve the equation
F (u) = h
56 Lectures on Applied PDEs, November 3, 2017

for u. Recall that a map G : X → Y is called the inverse of the map F : Y → X if and
only if G ◦ F = 1lX and F ◦ G = 1lY . Here, 1lZ denotes the identity on the space Z. We
write G = F −1 .
Recall that a linear map L is called invertible if and only if it has a bounded inverse.
If the map F is linear, F (y) = Ly, where L is a linear operator on Y , then the problem
we want to solve reads Ly = x, which is reduced to inverting the operator L. Thus the
inverse function theorem is a generalization to the nonlinear setting of the key problem
of linear analysis of inverting an operator.
Applying the implicit function theorem to the equation F (h, u) = 0, where F (h, u) :=
F (u) − h, we obtain the following theorem generalizing the corresponding theorem in
multivariable calculus:
Theorem 6.9 (The inverse function theorem). Let U be an open neighborhood of 0 ∈ X,
and let F : U → Y be a C 1 map s.t. dF (0) : Y → X has a bounded inverse (i.e.
dF (0) : Y → X is bijective). Then there is a neighborhood W of F (0) in Y and a unique
map G : V → X s.t. F (G(u)) = u, for all u ∈ W .

0. U V . F(0)

Y
X
G

Figure 7: Maps F and G

Note that the problem of proving of the existence of surfaces with prescribed mean
curvature fits nicely in the framework of the implicit function theorem. In this case we
have to show that F = H satisfies:

1) H : U → H⊥r−2 (Ω), where U is a neighborhood of 0 in H0r (Ω), r > n/2 + 1,


2) H is C 1 ,
3) dH(0) has a bounded inverse.

6.6 Appendix I: Mean curvature of a hypersurface


In this appendix we define the notion of the mean curvature of a hypersurface in Euclidean
spaces and derive convenient expressions for it. Let S be a smooth n–dimensional surface
in Rn+1 . Such a surface is called a (smooth) hypersurface. Hypersurfaces appear as graphs
or level sets of images of some functions.
Let U ⊂ Rn and f : U → R, be smooth. Then graph f := {(u, f (u)) | u ∈ U } is a
hypersurface in Rn+1 .
Lectures on Applied PDEs, November 3, 2017 57

Let ϕ : Rn+1 → R be a smooth function and let 0 ∈ Ran ϕ. Then the zero level set of
ϕ,

ϕ−1 (0) := {x ∈ Rn+1 | ϕ(x) = 0}


is a smooth hypersurface.
Finally a map ψ : U ⊂ Rn → S is called a local parametrization of S, provided (fill
in).
We give now the definition of different notions of curvatures at a point x0 ∈ S. We
begin with a simple but coordinate dependant definition. Pick a coordinate system s.t.
∇f (x00 ) = 0, where x0 = (x00 , xn+1
0 ). Then we define

• the principal curvatures at x0 as the eigenvalues of Hessf (x00 ),

• the Gauss curvature at x0 as det Hessf (x00 ),

• the mean curvature at x0 as h(x0 ) = Tr Hessf (x00 ) = ∆f (x00 ).

Lemma 1. Let S = graphf for some f s.t. f (x0 ) 6= 0. Then the mean curvature at x is
given by (6.18).

Proof. Consider first an arbitrary coordinate system and a function f : Ω → R, s.t.


S = graphf . We denote as before x = (x0 , xn+1 ) ∈ Rn+1 , x0 = (x1 , . . . , xn ) ∈ Ω ⊂ Rn . As
we have shown, the unit normal vector to S at x = (x0 , f (x0 )), ν(x), can be expressed as

(−∇f (x0 ), 1)
ν(x) = p . (6.20)
1 + |∇f (x0 )|2

Now for a given point x0 ∈ S, let x = (x0 , xn+1 ) ∈ Rn+1 be a special coordinate system s.t.
there is a domain U ⊂ Rn and a function f : U → R s.t. S = graphf and ∇f (x00 ) = 0.
Then we can express the normal vector ν(x) in terms of this function as

(−∇f (x0 ), 1)
ν(x) = p .
1 + |∇f (x0 )|2

Now compute

∆f (x0 ) |∇f (x0 )|2


div ν(x) = − + ,
(1 + |∇f (x0 )|2 )1/2 (1 + |∇f (x0 )|2 )3/2

and therefore we get div ν(x0 ) = −∆f (x00 ). By the definition of the mean curvature at
the point x0 , ∆f (x00 ) = −h(x0 ), and therefore div ν(x0 ) = −h(x0 ), which together with
(6.20) implies the lemma.
58 Lectures on Applied PDEs, November 3, 2017

Now we define the curvature in a coordinate-independent way. This way reveals the
geometrical meaning of this notion and on the way deals with some important concepts
in the theory of surfaces.
We define the tangent space, Tx S, to S at x as

Tx S := {ξ ∈ Rn+1 | ∃ C 1 path, γs , in S s.t γs |s=0 = x, and ∂s γs |s=0 = ξ}


(i.e., Tx S is the space of initial velocities of curves on S starting at x). Let ν(x) be
an ’outward’ unit normal to the surface S at a point x ∈ S. Denote by S n the unit
n–dimensional sphere. The map ν : S → S n , given by

ν : x → ν(x),
is called the Gauss map. The negative of its derivative

Ax := −∂ν(x)
at x is called the Weingarten map. The map Ax measures the rate of change in the
direction of ν(x) as it moves along S. By definition,

Ax : Tx S → Tν(x) S n .
TxS
nu(x) Tnu(x)Sn
nu(x)

Sn

Proposition 6.10. 1) Ran Ax ⊂ Tx S ⊂ Rn+1 ; 2) A∗x = Ax .

Before proceeding to the proof of this proposition, we give

Definition 6.1. The mean curvature of S at X is

H(x) := Tr Ax .

The Gauss curvature of S at x is

G(x) := det Ax .

The principle curvatures of S at x are the eigenvalues of Ax .


Lectures on Applied PDEs, November 3, 2017 59

Proof of Proposition 6.10. 1) Differentiating the relation hν(x), ν(x)i = 1, we find h∂ν(x), ν(x)i =
0. Thus, Ax ξ ⊥ ν(x), i.e., Ax ξ ∈ Tx S.
2) Let ϕst be a two dimensional parameterized surface in S, i.e., ϕst ∈ S, such that
ϕst |s=t=0 = x, ∂s ϕst |s=t=0 = ξ and ∂t ϕst |s=t=0 = η. Then

hξ, Ax ηi = −h∂s ϕst |s=t=0 , ∂t ν(ϕst )|s=t=0 i


= −∂t h∂s ϕst , ν(ϕst )i|s=t=0 + h∂t ∂s ϕst , ν(ϕst )i|s=t=0 .

Since ∂s ϕst ⊥ ν(ϕst ), we have finally

∂ 2ϕ
hξ, Ax ηi = h |s=t=0 , ν(x)i. (6.21)
∂s∂t
Similarly, we obtain

∂ 2ϕ
hAx ξ, ηi = h |s=t=0 , ν(x)i
∂s∂t
and therefore, hξ, Ax ηi = hAx ξ, ηi, i.e., A∗x = Ax .
Remark. Let ψ : U → S be a parametrization of S at x. Take ξ = ei and η = ej in
(6.21) where {ei } is an orthonormal basis in Rn ⊃ U . Then

∂ 2 ψ(u)
(Ax )ij = h , ν(x)i
∂ui ∂uj

where u = ψ −1 (x). The matrix Ax is called the 2nd fundamental form. (The 1st funda-
mental form of S is the metric on S induced by the Euclidean metric on Rn+1 .)
Proposition 6.11. In the level set representation, S = ϕ−1 (0), we have
∇ϕ(x) ∇ϕ(x)
 
ν(x) = − and H(x) = div .
|∇ϕ(x)| |∇ϕ(x)|
Proof. Let γs be a path on S starting at x with an initial velocity ξ: γs |s=0 = x and
∂s γs |s=0 = ξ. Differentiating ϕ(γs ) = 0 we find ∇ϕ · ξ = 0. Hence ∇ϕ ⊥ Tx S and
∇ϕ
therefore ν(x) is equal to |∇ϕ| up to a sign.
To prove the second formula we have by the definition of Ax
∇ϕ(γs ) ∇ϕ(x)
X  
Ax ξ = ∂s |s=0 = ∂xi ξi .
|∇ϕ(γs )| i
|∇ϕ(x)|
∂ 
xj ϕ(x)
Place the coordinate system at the point x with en+1 = ν(x). Then (Ax )ij = ∂xi |∇ϕ(x)|
 
∇ϕ(x)
and therefore H(x) = div |∇ϕ(x)| as claimed.
60 Lectures on Applied PDEs, November 3, 2017

7 Theory of bifurcation
7.1 Set-up
Consider a C 1 –map F : R × Y → Z, where Y and Z are Banach spaces. We would like
to find a function u = u(µ), implicitly defined by the equation
F (µ, u) = 0. (7.1)
Assume (7.1) has a branch of solutions (µ, ū(µ)) for all µ in some open interval. (The
“curve” (µ, u(µ)), µ ∈ [−, ], is called a branch of solutions if
F (µ, u(µ)) = 0, ∀µ ∈ [−, ]. (7.2)
The branch of solutions {(µ, ū(µ)) : µ ∈ R} is called the trivial branch.)
We are interested in a situation when a new branch of solutions emerges (or bifurcates)
from the existing one. A point (µ0 , u0 ) at which a new branch of solutions emerges from
(µ, ū(µ)) is called a bifurcation point.

u
nontrivial branch

bifurcation point trivial branch

Figure 8: Bifurcation

At a bifurcation point, the operator du F (µ, u0 ) is not invertible, since otherwise, the
implicit function theorem would apply and give a unique solution in a neighbourhood of
(µ0 , u0 ).
An important example is given by the case of a linear (in u) map F : R × Y → Y :
F (µ, u) = µLu − u,
where L is a linear operator on Y . Then (µ, 0) is the trivial branch of solutions. If µ−1 0 is
an eigenvalue of L, then in a neighbourhood of (µ0 , 0), the equation F (µ, u) = µLu−u = 0
has another branch of solution: (µ0 , au0 ), for any a ∈ C, where u0 is an eigenfunction of L
corresponding to the eigenvalue µ−10 . Since dF (µ, 0) = µL−1l, du F (µ0 , 0) is not invertible,
which verifies the conclusion above. The conclusion above is true for every eigenvalue of
L. (If 1/µ is an eigenvalue of L, then we call µ a characteristic value of L.)
Lectures on Applied PDEs, November 3, 2017 61

R u1 R u2 . . . . . . .

µ
...

(λ−1
1 ,0) (λ−1
2 ,0)
.......

Figure 9: Bifurcation:linear case

Example 7.1. Let Y = Z = L2 ([0, 2π]), L = −∆ with Dirichlet boundary condi-


tions: u(0) = u(2π) = 0. Recall that the domain of L satisfies D(L) = H2 ([0, 2π]) ⊂
C([0, 2π]) by the Sobolev embedding theorem, and therefore the boundary conditions
u(0) = u(2π) = 0 make sense.
To find the eigenvalues of L, we need to solve the equation −∆u = λu, i.e. u00 = −λu.
The solutions satisfying the Dirichlet boundary conditions are un = a sin( n2 x), where
a ∈ R, and the eigenvalues are given by λn = ( n2 )2 , n = 1, 2, 3, . . .. Thus besides the
trivial branch (µ, 0), µ ∈ R, the equation µ(−∆u) − u = 0 has the branches of solutions
(( n2 )2 , Run ), for n = 1, 2, 3, . . ..
Observe that in the the above example, F is linear in u, and as a result, the bifurcating
branches are straight lines. In general, if F is nonlinear, we expect the bifurcating branches
to be bent, as in the following example.
Example 7.2. Let Y = Z = R, and F (µ, u) = µu − u3 . Clearly we have F (µ, 0) = 0
∀µ ∈ R, so (µ, 0) is the trivial branch. We calculate the derivative du F (µ, u) = µ − 3u2 ,
so du F (µ, 0) = µ = 0 has the solution µ0 = 0, hence (0, 0) is a candidate for a bifurcation
point. On the other hand, we can solve the equation F (µ, u) = 0 explicitly, obtaining the

solutions (µ, 0) and u = ± µ. This shows that (0, 0) is indeed a bifurcation point (the
bifurcation here is called a pitchfork bifurcation because of the shape of its bifurcating
branch).
Below we will learn how to find out the qualitative behavior of the bifurcating branch
without actually solving for it. But before, let us find a necessary condition for a bifur-
cation to happen at a point (µ0 , u0 ). By the implicit function theorem, we know that
if du F (µ0 , u0 ) has a bounded inverse, then equation (7.1) has a unique solution in a
neighborhood of (µ0 , u0 ), which must be the trivial branch (µ0 , u0 ). Hence we have the
following
Proposition 7.1. If (µ0 , u0 ) is a bifurcation point, then du F (µ0 , u0 ) does not have a
bounded inverse.
Now, let us find a sufficient condition for a bifurcation to happen at a point (µ0 , u0 ).
We begin with following cautionary example.
62 Lectures on Applied PDEs, November 3, 2017

Example 7.3. For F : R × R2 → R2 given by F (µ, u1 , u2 ) = (u1 , u2 ) − µ(u1 − u32 , u2 + u31 ),


we find
du F (µ, u)(ξ1 , ξ2 ) = (ξ1 , ξ2 ) − µ(ξ1 − 3u22 ξ2 , ξ2 + 3u21 ξ1 ),
and therefore Du F (µ, 0) = (1 − µ)1l. Thus du F (1, 0) is not invertible. However, (1, 0) is
not a bifurcation point! Indeed, look at the two components of the equation F (µ, u) = 0.
Multiplying the first one by −u2 , the second one by u1 , we obtain

−(1 − µ)u1 u2 − µu42 = 0


(1 − µ)u1 u2 − µu41 = 0.

Adding the above two equations yields −µ(u41 + u42 ) = 0, so u1 = u2 = 0 (for µ 6= 0),
which shows that F (µ, u1 , u2 ) = 0 has only the trivial solution (u1 , u2 ) = (0, 0), ∀µ ∈ R
(if µ = 0, then this follows directly from the definition of F ). (1, 0) is therefore not a
bifurcation point.

7.2 Key result


We now give a sufficient condition for a bifurcation to take place. To minimize technical-
ities, we formulate and prove the next result for the special case of when du F (µ0 , u0 ) is a
self–adjoint operator (see a remark below for a discussion and Appendix A.8 and [6] for
the definition).
Theorem 7.2 (Krasnoselski). Assume that Y is dense in a Hilbert space Z and consider
a map F : R × Y → Z. Assume that
(i) F has a trivial branch of solutions (µ, ū(µ)) : µ ∈ R;

(ii) F is C 1 in u and C in µ:

(iii) du F (µ0 , u0 ) is a self–adjoint operator and has the eigenvalue 0 of odd multiplicity
(i.e. Null du F (µ0 , u0 ) is an odd dimensional subspace);

(iv) The restriction du F (µ0 , u0 ) (Null du F (µ0 ,u0 ))⊥ is an invertible operator;

(v) du F (µ, u) is C 1 in µ at the point (µ0 , u0 ) and there exists a v0 ∈ Null du F (µ0 , u0 )
such that
hv0 , ∂µ du F (µ0 , u0 )v0 i =
6 0.

Then (µ0 , u0 ) is a bifurcation point of (7.1).


We conduct the proof in two steps. One the first step, under very general conditions,
we reduce the problem of solving the equation (7.1) to the problem of solving an equation
in a few dimensions (the reduced or effective, or bifurcation equation). On the second
step, we solve the latter equation.
Lectures on Applied PDEs, November 3, 2017 63

On the first step, we use the powerful technique, called the Lyapunov-Schmidt decom-
position. We will use this technique repeatedly below.
Proof of the Krasnoselski theorem. Let L(µ) := du F (µ, u0 ), and denote by P the orthog-
onal projection onto the subspace Null L(µ0 ) and let P ⊥ := 1l − P , the orthogonal pro-
jection onto the orthogonal complement (Null L(µ0 ))⊥ of the subspace Null L(µ0 ). (For
the definition and properties of (orthogonal) projections, see Appendix A.8.) Note that
L(µ0 )P = P L(µ0 ) = 0.
By changing the unknown variables if necessary, we assume in what follows that ū(µ) =
0. We project u ∈ Y and the equation F (µ, u) = 0 onto the subspaces Ran P and Ran P ⊥ :
u = v + w, where v ∈ Ran P and w ∈ Ran P ⊥ , and

P F (µ, v + w) = 0, (7.3)
P ⊥ F (µ, v + w) = 0. (7.4)

We have thus two equations, (7.3) and (7.4), for two variables v and w. Observe that
since dim Ran P < ∞, variable v is finite–dimensional.
We now show equation (7.4) has a unique solution w. Define

F1 (µ, v, w) := P ⊥ F (µ, v + w) : X ×P ⊥ Y → P ⊥ Z.

where X := R × P Y . Then the problem (7.4) can be reformulated as solving the equation
F1 (µ, v, w) = 0 for w. Now observe that

(α) F1 is C 1 ,

(β) F1 (µ, 0, 0) = 0 for any µ,

(γ) dw F1 (µ0 , 0, 0) is invertible.

Indeed, (α) follows from the condition that F is C 1 , (β) results from the relation F1 (µ, 0, 0) =
P ⊥ F (µ, 0) = 0, and (γ) is due to the relation

dw F1 (µ0 , 0, 0) = P ⊥ du F (µ0 , 0)P ⊥

plus the fact that the r.h.s. is invertible as an operator from P ⊥ Y to P ⊥ Z.


Hence we can apply the implicit function theorem to the equation F1 (µ, v, w) = 0,
which shows thus that for any (µ, v) sufficiently close to (µ0 , 0), the equation F1 (µ, v, w) =
0, and therefore (7.4), has a unique solution. We denote this solution by w = w(µ, v).
We substitute the solution, w = w(µ, v), of (7.4) into (7.3) to obtain the equation

f (v, µ) = 0, (7.5)

with f (v, µ) := P F (µ, v + w(µ, v)).


64 Lectures on Applied PDEs, November 3, 2017

Below, we show that


w = o(kvk) + O(|µ − µ0 |kvk). (7.6)
This together with the relation u = v + w gives that the solution of the original problem
has the form

u = v + w(µ, v), with w(µ, v) = o(kvk) + O(|µ − µ0 |kvk). (7.7)

This completes the first step of the proof. The equation (7.5) is called the bifurcation
equation or branching equation. It describes the bifurcating branches, and usually, it is a
system of n = dim Null du F (µ0 , 0) algebraic equations for n + 1 variables µ and v.
Now, we proceed to the second step: solving equation (7.5). Recall the notation
L(µ) := du F (µ, 0). Assume for simplicity that dim Null L(µ0 ) = 1, i.e. that the eigenvalue
0 of L(µ0 ) is simple and show that equation (7.5) has a unique solution for µ as a function
of v ∈ R. Let v0 the zero eigenvector of the operator L(µ0 ), normalized as hv0 , v0 i = 1.
Then P u = hv0 , uiv0 . Since v = sv0 for some s ∈ R, equation (7.5) is equivalent to the
equation f (s, µ) = 0, where
(
1
hv0 , F (µ, sv0 + w(µ, sv0 ))i for s 6= 0,
f (s, µ) := s (7.8)
hv0 , L(µ)v0 i for s = 0.

Let u1 := s−1 u. Next, using the fact that F (µ, 0) = 0 for any µ, we expand the map
F (µ, u) around u = 0 as
F (µ, u) = L(µ)u + R(µ, u), (7.9)
with R(µ, u) = o(kuk) and, recall, L(µ) := du F (µ, 0). Using this expansion, we rewrite
f (s, µ) as
f (s, µ) = hv0 , L(µ)u1 i + hv0 , s−1 R(µ, su1 )i, (7.10)
Since s−1 kR(µ, su1 )k → 0 and, by (7.6), u1 − v0 = s−1 w → 0, as s → 0, we have that
f (s, µ) → hv0 , L(µ)v0 i, as s → 0. We compute the µ−derivative of f . Hence f (s, µ) is
continuous at s = 0. Next, we have
∂f
(s, µ) = hv0 , ∂µ L(µ)u1 i + hv0 , L(µ)∂µ u1 i + hv0 , s−1 (∂µ R(µ, su1 ) + sdu R(µ, su1 ))∂µ u1 i.
∂µ

Write u1 := s−1 u = v0 + w1 , where w1 := s−1 w. Below, we show that

w, ∂µ w = o(kvk) + O(|µ − µ0 |kvk). (7.11)

This implies kw1 k, k∂µ w1 k → 0, as s → 0 and |µ − µ0 | → 0. by the last relation and


s−1 k∂µ R(µ, su1 )k, kdu R(µ, su1 )k → 0, as s → 0, we have that
∂f
(0, µ) = hv0 , ∂µ du F (µ, 0)v0 i, (7.12)
∂µ
Lectures on Applied PDEs, November 3, 2017 65

which shows that f (s, µ) is differentiable in µ also at s = 0. Moreover, by condition (v) of


∂f
the Krasnoselski theorem, ∂µ (0, µ0 ) = hv0 , ∂µ du F (µ0 , 0)v0 i 6= 0. Therefore equation (7.8)
has a unique solution µ = µ(s), for µ as a function of s, if s is in a neighbourhood of
s = 0. This completes the second step.
We have shown above that the solution of the original problem has the form (7.7),
where v satisfies the equation (7.5). In the case when 0 is a simple eigenvalue of L(µ0 ),
we have shown that the equation (7.5) can be solved for µ as a function of v, µ = µ(v).
This give the bifurcation branch of solutions

(u = v + w(µ, v), µ = µ(v)), (7.13)

parametrized by v ∈ Null L(µ).


Proof of the estimate (7.11). To show (7.11), we use the expansion (7.9), to rewrite the
equation (7.4), as
L⊥ (µ)w +P ⊥ L(µ)v +P ⊥ R(µ, u) = 0, (7.14)
where L⊥ (µ) := Pµ⊥ L(µ)Pµ⊥ . Since the operator L⊥ (µ) : Pµ⊥ Y → Pµ⊥ Z is invertible, we
derive
w = −L⊥ (µ)−1Pµ⊥ (R(µ, u) + L(µ)v) .
The relations R(µ, u) = o(kuk) and L(µ)v = (L(µ) − L(µ0 ))v = O(|µ − µ0 | kvk) imply
w = o(kuk) + O(|µ − µ0 | kvk). Since u = v + w, this shows (7.11) for w. To prove (7.11)
for ∂µ w, we differentiate (7.14) w.r.to µ and then proceed with the resulting equation as
we did with (7.14).
Example. As in an example above, let F (µ, u) = µu − u3 . We have du F (µ, u) = µ − 3u2 ,
so 0 is an eigenvalue of du F (0, 0) of multiplicity 1. Next, ∂µ du F (0, 0) = 1, so the condition
(ii) is satisfied as well. Therefore (0, 0) is a bifurcation point.
We summarize the result of this step in the following
Theorem 7.3 (Reduction to effective equation). Assume conditions (i) - (iv) of Theorem
7.2 hold. Then (7.1) has a solution, u, iff (7.5) has a solution, v. Moreover, these
solutions are related by (7.7).

Connection to the spectral theory. The non-invertibility of du F (µ0 , 0) is equivalent


to the statement that
0 ∈ σ(du F (µ0 , 0)).
Here σ(A) denotes the spectrum of a linear operator A, i.e.

σ(A) := {z ∈ C : A − z1l is not invertible}.

Clearly, eigenvalues of A belong to σ(A). The multiplicity of an eigenvalue λ of A is


defined as the number of linearly independent eigenvectors with eigenvalue λ, i.e.
66 Lectures on Applied PDEs, November 3, 2017

• the multiplicity of an eigenvalue λ of A is dim Null(A − λ1l).

We say an eigenvalue λ of A is isolated iff there is a neighbourhood, U , of λ in C s.t.


U ∩ σ(A) = {λ}.
In general, the spectrum can also contain continuous pieces and it can take very
peculiar forms. For details on spectral theory see Appendix E and [6].
The spectral analysis simplifies considerably if an operator A acts on a Hilbert space
and is self–adjoint (see Appendix A.8 and [6] for the definition). For instance, the spec-
trum of a self–adjoint operator is real.
Self-adjoint operators are symmetric operators, i.e. operators obeying

hAu, vi = hu, Avi,

which in addition satisfy certain domain condition. (The latter condition is trivially
satisfied for bounded operators.) To show that a given operator is self-adjoint usually
requires some work. However, in the situations we will be dealing with, all symmetric
operators are self-adjoint. Thus in these situations to show that that a given operator is
self–adjoint it suffices to show that it is symmetric.

Exercise 7.1. Show that


(a) the multiplication operator Mf : u(x) → f (x)u(x) is symmetric provided the
function f is real and bounded and its spectrum is σ(Mf ) = Ran f ;
(b) the differentiation operator −i ∂x∂ j is symmetric and its spectrum is σ(−i ∂x∂ j ) = R;
(c) the identity operator has the spectrum consisting of one point σ(1l) = {1};
P ∂2
(d) the Laplacian ∆ := n1 ∂x 2 2
2 on L (Ω) with the domain D(∆) = H (Ω) has the
j
spectrum [0, ∞);
(e) the Laplacian ∆ on [−a, a]n with (i) Dirichlet boundary conditions (i.e. u = 0
on the boundary) or (ii) the periodic boundary conditions has only eigenvalues of finite
multiplicities; find these eigenvalues.
Hint: in cases (b) and (d) use the Fourier transform to reduce these problems to ones
of the type of (a).

We have the following result (see [6])

Proposition
7.4. Let A be a linear, self–adjoint operator with the eigenvalue 0. Then
A Null A is invertible iff 0 is an isolated eigenvalue of A.

This proposition implies that the condition (iv) of Theorem 7.2 above is equivalent to
the following condition

(iv’) 0 is an isolated eigenvalue of du F (µ0 , u0 ).


Lectures on Applied PDEs, November 3, 2017 67

How to determine that a given eigenvalue is isolated? There are two notable cases:
(a) elliptic differential operators on bounded and Schrödinger operators with confining
potentials have purely discrete spectrum and
(b) Schrödinger operators with potentials vanishing at infinity have the essential spec-
trum filling in the semi-axis [0, ∞) and therefore their negative spectrum if exists is
discrete.
In the latter case, one uses the variational calculus to find negative spectrum of
Schrödinger operators, see Section 10.7 and [6], for more details.

Corollary 7.5. Let F be a C 1 in u and C in λ map satisfying du F (λ, u0 ) = L−λ1l, where


L is a linear, self–adjoint operator. If λ0 is an isolated eigenvalue of L of odd multiplicity,
then (λ0 , u0 ) is a bifurcation point.

Proof. Clearly the first two conditions of the Krasnoselski’s theorem are satisfied. To
check the third condition, we compute ∂λ du F(λ, 0) = −1l, and hv0 , ∂λ du F(λ, u0 )v0 i =
−hv0 , v0 i = −kv0 k2 6= 0 ∀v0 ∈ Null (L − λ1l). Hence the third condition is satisfied as
well.

The proof of the Krasnoselski’s theorem (see (7.13)) implies the following

Corollary 7.6. Assume the operator du F (µ0 , 0) has a simple, isolated eigenvalue at 0
with an eigenfunction ϕ. Then the bifurcating at µ0 branch of solutions of (7.1) is of the
form
(u = sϕ + O(s2 ), µ = µ0 + O(s2 )), (7.15)
parametrized by s := hϕ, ui.

7.3 Applications
The nonlinear eigenvalue problem. Consider the nonlinear eigenvalue problem

Lu + f (u) = λu, with f ∈ C 1 and f (0) = f 0 (0) = 0. (7.16)

Then the corollary implies that if λ0 is an eigenvalue of a self–adjoint operator L of odd


multiplicity, then equation (7.16) has a nontrivial branch of solutions near the bifurcation
point (λ0 , 0).
We specify this example further by considering the following nonlinear eigenvalue
problem
(−∆ + V )u + f (u) = λu, (7.17)
where u ∈ H 2 (Rn , R), ∆ is the Laplacian on Rn and the function f : Rn → R satisfies
f (0) = 0, f 0 (0) = 0 and |f (u)| ≤ c|u|p with n2 − 2p
n
< 2.
68 Lectures on Applied PDEs, November 3, 2017

Exercise 7.2. Find the bifurcation points for the equation

∆u + u + u5 = 0, (7.18)

(a) for u ∈ H 2 ([−a, a]), with Dirichlet boundary conditions (i.e. u = 0 on the boundary),
2 2
and (b) for u ∈ Hsym ([−a, a]n ), n ≥ 2. Here Hsym ([−a, a]n ) denotes the subspace of the
Sobolev space H 2 ([−a, a]n ) consisting of functions symmetric with respect to permuta-
tions, i.e., ψ(x1 , . . . , xn ) = ψ(xπ(1) , . . . , xπ(n) ) for all permutations π. Hint: Here a is a
bifurcation parameter. You can break the solution in the following steps:
- Rescale (7.35) by passing from u(x) to v(x) = u(ax) to find an equivalent equation
for v depending on a but defined on the fixed domain [−1, 1]n ;
- find the trivial branch of solutions;
- find the candidate for the bifurcation point;
- for this candidate check the conditions of the bifurcation (Krasnoselski) theorem.

Lamellar phase. Problem (open? references?): Show the existence of periodic static
solutions (lamellar phases) of the Allen-Cahn equation, (2.2) (see Section 2.1). Static
solutions of the Allen-Cahn equation satisfy the static Allen-Cahn equation (see (2.3)),

− 2 ∆u + g(u) = 0 (7.19)

where g(u) = G0 (u), with G(u) a function of the form shown in Fig 2, or Fig 18 below.
G

u
−1 1

Figure 10: Double well potential.

(Recall that in lamellar phase, the layers of +1 and −1 phases (substances) coexist in
a periodic array.)
There are three ways to proving existence of such solutions. A variational approach
will be discussed in Section 20. Another way is to construct an approximate solutions
by gluing together kinks and antikinks (see Fig 11) and then using either the implicit
function theorem or its generalization due to the Lyapunov-Schmidt decomposition, to
prove the existence of the true solutions nearby.
The third approach which we discuss here is to use the bifurcation theory. The last two
approaches give the existence of the periodic solutions in two opposite regimes (the long
period large solutions and short period small solutions), while the variational approach
does not give much specific information about the solutions.
Lectures on Applied PDEs, November 3, 2017 69

+1 +1
 C  C 
 C  C 
 C  C or 
 C  C \
 C  C  \
-1  C  C -1  R

Figure 11: Periodic solution built out of kinks and antikinks.

We discuss briefly the bifurcation approach. This equation has three homogeneous
(x−independent) solutions u = 0, u = a and u = b, which give three solution branches
(, 0), (, a) and (, b), ∀ ≥ 0. Consider bifurcation from one of these branches. If we
define F (, u) := −2 ∆u + g(u), then

dF (, u) = −2 ∆ + g 0 (u). (7.20)

We are interested in u = 0, a, b. Since G has a strict local maximum at u = 0 and strict


local minima at u = a and u = b, we have

g 0 (0) = G00 (0) < 0, g 0 (a) = G00 (a) > 0, g 0 (b) = G00 (b) > 0.

Hence the linearized operator is strictly positive for v = a, b and the only branch which
might lead to bifurcations is (, 0).
If we consider the operator dF (, 0) on the entire space L2 (R), then σ(dF (, 0)) =
[g 0 (0), ∞) (see Exercise 7.1 above) and the bifurcation problem becomes intractable: 0 is
a part of the continuous spectrum of dF (, 0).
Since we are looking for the bifurcation of periodic solutions, we consider the operator
dF (, 0) on an interval, say [0, c] ⊂ R, with periodic boundary conditions. The spectrum
of dF (, 0) in this case is
2π 2
σ(dF (, 0)) = 2 k + g 0 (0), k = 0, 1, . . . .
c
Exercise 7.3. Find all bifurcation points (proving existence of bifurcating periodic solu-
tions) of the static Allen-Cahn equation

− ∆u + (u2 − τ )u = 0 (7.21)

on R, satisfying u(−x) = −u(x). (Hint: Since we are looking for periodic solutions,
we consider equation (7.21) for odd functions on an interval, say [−a/2, a/2] ⊂ R, with
periodic boundary conditions. Consider a as the bifurcation parameter. You can break
the solution in the following steps:
70 Lectures on Applied PDEs, November 3, 2017

rescale (7.21) by passing from u(x) to v(x) = u(ax) to find an equivalent equation
for v depending on a but defined on the fixed domain [−1/2, 1/2]; write this equation as
F (a, v) = 0 (not to confuse with (7.20));
find the trivial branch of solutions;
find the candidates for the bifurcation points;
for these candidates check the conditions of the bifurcation (Krasnoselski) theorem.)
Exercise 7.4. Prove existence of bifurcating periodic solutions of the static Allen-Cahn
equation (7.21) on R2 , satisfying u(−x, y) = −u(x, y) and u(x, −y) = −u(x, y). (Hint:
Here we are dealing with double periodic solutions, hence consider equation (7.21) on a
rectangle, say [0, a] × [0, b] ⊂ R2 , with periodic boundary conditions.)

The FitzHugh-Nagumo model. Consider the system of equations describing con-


duction of electric pulses along nerve exon and known as the FitzHugh-Nagumo model:
∂u
= ∆u − (u − a)(u − 1)u − v, (7.22)
∂t
∂v
= (u − v). (7.23)
∂t
Exercise 7.5. (open? references?) Prove existence of bifurcating static periodic solutions
of the FitzHugh-Nagumo equations, (7.22) – (7.23), on R.
Exercise 7.6. (open? references?) Consider bifurcation of traveling wave periodic solu-
tions (trains) of the FitzHugh-Nagumo equations, (7.22) – (7.23), on R.

Bifurcation of instantons. Let g(u) be given by g(u) = G0 (u), where G(u) is a func-
tion of the form shown in Fig 12.
G
u1

u0

−G
Figure 12: Function G and solution u1

We consider the equation


− ∆u + g(u) = 0, (7.24)
Lectures on Applied PDEs, November 3, 2017 71

on the strip [−a/2, a/2] × (−∞, ∞), with periodic boundary condition in x and L2 in y:
u(−a/2, y) = u(a/2, y) ∀y, (7.25)
u(x, ·) ∈ L2 (R) ∀x.
This equation describes tunnelling of a string (or line vortex) placed at the local minimum
of the potential G(u) (see Fig 12) through the barrier. The parameter a is inverse tem-
perature. Solutions of this equation are called instantons (for the physical background,
see [?]).
Equation (7.24) with boundary conditions (7.25) has the following branches of solu-
tions which are independent of x: (a, u0 ), ∀a, and (a, u1 ), ∀a, where u0 is a constant
function equal to the local minimizer of G(u) (see Fig 12) and u1 satisfies 0 ≤ u1 (y) ≤
α, u1 → 0 as y → ±∞ and u1 (0) = α (where α is the first positive zero of G(u), see Fig
12). To prove the existence of the solution u1 , we observe that it satisfies the equation

∂ 2u
− + g(u) = 0 on (−∞, ∞). (7.26)
∂y 2
The last equation is just Newton’s equation if we interpret y as being the time-variable,
and −G as a potential (whose derivative is a force). Then the existence of u1 is obvious
from the picture above.
The problem here is to find solutions bifurcating from the branches above as a varies.
Exercise 7.7. (a) Check whether any solutions bifurcate from (a, u0 ).
(b) Find the bifurcation points from the branch (a, u1 ). (Hint: Let F (a, u) := −∆u +
g(u). We can write La := dF (a, u1 ) as
La = −∂x2 + `,
where ` is the operator acting on the variable y and given by
` := −∂y2 + g 0 (u1 (y)).
Prove the following statements
– the eigenvalues, νi , of ` are of the form
νi = λi + µi , (7.27)
where λi and µi are the eigenvalues of the operators −∂x2 and ` acting on
L2per ([−a/2, a/2]) and L2 (R), respectively;
– the operator ` has the eigenvalue zero with the eigenfunction φ1 (y) := ∂y u1 (y);
– φ1 (y) has exactly one zero and hence by the Sturm theory the operator ` has
exactly one negative eigenvalue, say, λ0 , and this eigenvalue is simple (i.e. of
the multiplicity 1 or non-degenerate).
Find the eigenvalues of −∂x2 on L2per ([−a/2, a/2]) and use (7.27) to find the eigen-
values of La . Use these facts to solve the problem.
72 Lectures on Applied PDEs, November 3, 2017

7.4 Linearized stability/instability


In this section we study the linearized stability of stationary (and traveling wave) solu-
tions. The general theory is discussed in the next section. At the moment we give an
ad hoc definition of the linearized stability, which we will explain below. Consider the
abstract dynamical system, (2.1), which we reproduce here:
∂u
= F (u). (7.28)
∂t
Here F : U → Y , with U ⊂ X and X and Y Banach spaces. We say that a stationary
solution u∗ of (7.28) is linearly stable iff
σ(dF (u∗ )) ⊂ { Re z < 0 } (7.29)
and linearly unstable iff
σ(dF (u∗ ) ∩ { Re z > 0 } =
6 ∅.
Here, recall, dF (u) denote the Gâteaux derivative of the map F (u) at u.
To exhibit the origin of this definition, we write the equation (7.28) in the canonical
form as follows. First, we expand F (u) around u∗ and use that F (u∗ ) = 0 to obtain
F (u∗ + ξ) = dF (u∗ )ξ + f (ξ), (7.30)
where the term f (ξ) is defined by this expansion. According to (6.3) of Theorem 6.1, it is
of a higher order, f (ξ) = o(kξk). Substituting u = u∗ + ξ and this expansion into (7.28),
we find
∂ξ
= L∗ ξ + f (ξ), (7.31)
∂t
where L∗ := dF (u∗ ). This is the abstract dynamical system in the canonical form.
If the initial condition, u0 , for (7.28) is close to u∗ , then the initial condition, ξ0 =
u0 − u∗ , for (7.31) is close to 0. If u∗ is stable, then ξ remains small for all times. This
suggests that f (ξ) which is of a higher order in ξ does not have significant effect on the
behaviour of the solution and therefore can be omitted on the first step. This leads to
the linear equation
∂ξ
= L∗ ξ. (7.32)
∂t
In a very large variety of situations (e.g. when L∗ is self-adjoint), the spectrum of the
linear operator L∗ determines the long time behaviour of the latter equation (see Appendix
C), which leads to the definition above. (In all situations, the spectrum of the linerized
map is the first step in finding asymptotic behavior for the evolution equation in question
for initial conditions close to the static solution.)
Example 7.4. Let X = R2 and let L∗ be a real 2 × 2 matrix with eigenvalues λ1 , λ2 .
According to the nature of the spectrum of L∗ , we have the qualitative pictures for the flow
φt (u) = eL∗ t u shown in Figure 13.
Lectures on Applied PDEs, November 3, 2017 73

asymptotically stable unstable unstable

λ1 , λ2 < 0 λ1 ,λ2 > 0 λ1 < 0, λ2 > 0


asymptotically stable Lyapunov stable unstable

Reλ1 = Reλ2 < 0 Re λ1 = Reλ2 = 0 Reλ1 =Reλ2 > 0

Figure 13: Flow for Eqn. (7.32) and X = R2

Remark. If the spectrum of dF (u∗ ) is real, then the stability condition says that

dF (u∗ ) ≤ −θ, for some θ > 0. (7.33)

Change of stability at a bifurcation In this subsection we study the linearized


stability of the trivial and bifurcating solutions. We apply the definition above to the
bifurcation problem F (µ, u) = 0 coming out as the stationary equation (7.28) with F (u)
replaced by F (µ, u). As in the bifurcation Theorem 7.2 (Krasnoselski’s Theorem), we
assume that the equation
F (µ, u) = 0
has a family of (trivial, static) solutions (µ, ū(µ)) and that du F (µ, u) is self-adjoint.
Let µ∗ be a bifurcation point of this equation and u∗ ≡ ū(µ∗ ). According to Theorem
7.2, du F (µ∗ , u∗ ) has the isolated eigenvalue 0 of odd multiplicity. Since du F (µ, u) is self-
adjoint, its spectrum is real. We assume in addition that 0 is the largest spectral point
of du F (µ∗ , u∗ ) of the multiplicity one.

Theorem 7.7 (Change of stability at a bifurcation point). At the bifurcation point µ∗ ,


the stability of the trivial solution ū(µ) changes to the opposite: if ū(µ) is stable/unstable
for µ < µ∗ , then it is unstable/stable for µ < µ∗ and vice versa.

Proof. Denote Lµ := du F (µ, ū(µ)). Since Lµ∗ has has the isolated eigenvalue 0 of mul-
tiplicity one and since Lµ is differentiable in µ, by the perturbation theory, for |µ − µ∗ |
sufficiently small, we have (see [6])
74 Lectures on Applied PDEs, November 3, 2017

1) Lµ has an isolated eigenvalue, νµ , of the multiplicity one, s.t. νµ=µ∗ = 0;


2) νµ , as well as the corresponding eigenfunction, is differentiable in µ and satisfies
νµ=0 = 0;
3) if νµ∗ = 0 the largest spectral point of Lµ∗ , then νµ the largest spectral point of Lµ .
Let φµ be the eigenfunction of Lµ corresponding to the eigenvalue νµ .

Lemma 7.8.
∂µ νµ = hφµ , ∂µ Lµ φµ i.

Proof. We compute ∂µ νµ by differentiating the eigenvalue equation

Lµ φµ = νµ φµ ,

w.r. to µ, multiplying the result by φµ and using that φµ is normalized, to obtain

∂µ νµ = hφµ , ∂µ Lµ φµ i + ∂µ hφµ , Lµ φµ i + hφµ , Lµ pµ φµ i.

Now, since Lµ is self-adjoint, we have that

hφµ , Lµ ∂µ φµ i = hLµ φµ , ∂µ φµ i

1
= νµ hφµ , ∂µ φµ i = νµ ∂µ hφµ , φµ i = 0
2
and similarly for the other term. This gives the desired relation.

For µ = µ∗ , the r.h.s. is non zero by the condition (iii) of Theorem 7.2:

∂µ νµ∗ = hφ∗ , ∂µ Lµ∗ φ∗ i =


6 0,

where φ∗ ≡ φµ∗ and u∗ ≡ ū(µ∗ ). Hence, since νµ=µ∗ = 0, the largest eigenvalue νµ changes
the sign at µ = µ∗ . Specifically, if ∂µ νµ∗ > 0, then the largest eigenvalue of Lµ satisfies
νµ < 0 for µ < µ∗ and νµ > 0 for µ > µ∗ . (We assume that µ is sufficiently close to µ∗ .)
This implies that the (trivial, static) solution ū(µ) of F (µ, u) = 0 is stable for µ < µ∗ and
unstable for µ > µ∗ . Thus, as µ increases, at µ = µ∗ , the (trivial, static) solution ū(µ)
looses its stability. Similarly, for ∂µ νµ∗ < 0.

Assume ∂µ νµ∗ > 0. Does the equation F (µ, u) = 0 have stable solutions for µ > µ∗ ?
The answer is yes: it is the bifurcating solution, u(µ). One can show that the spectrum
of du F (µ, u(µ)) for µ > µ∗ , but close to µ∗ is strictly negative.

Exercise 7.8. Find values of λ for which the trivial solutions of the nonlinear eigenvalue
problem (7.17) are linearly stable and for which they are linearly unstable.
Lectures on Applied PDEs, November 3, 2017 75

Exercise 7.9. Find values of a for which the static solution, ū = 0 to the time-dependent
equation
∂t ut = ∆ut + ut + u5t , (7.34)
2
for ut ∈ Hsym ([−a, a]3 ), is linearly stable and for which it is linearly unstable. See Exercise
2
7.2 for the definition of ut ∈ Hsym ([−a, a]3 ). (Note that ū = 0 is the trivial solution of
the bifurcation problem in in Exercise 7.2.)
Find the bifurcation points for the equation

∆u + u + u5 = 0, (7.35)

2
for u ∈ Hsym ([−a, a]3 ), with Dirichlet boundary conditions (i.e. u = 0 on the boundary).
Here Hsym ([−a, a]n ) denotes the subspace of the Sobolev space H 2 ([−a, a]n ) consisting
2

of functions symmetric with respect to permutations, i.e., ψ(x1 , . . . , xn ) = ψ(xπ(1) , . . . ,


xπ(n) ) for all permutations π.

Exercise 7.10. Find values of the bifurcation parameter for which the trivial solution of
the static Allen-Cahn equation (7.21) in one and two dimensions (see Exercises 7.3 and
7.4) are linearly stable and for which they are linearly unstable.

Exercise 7.11. Find values of a for which the trivial solutions of problem (7.24) – (7.25)
(see Exercise 7.7) are linearly stable and for which they are linearly unstable.

8 Linearized stability of systems with symmetry


8.1 Systems with symmetry and zero modes
Often, the presence of symmetries implies that 0 is an eigenvalue of the linearized operator
and therefore the condition (7.29) is never satisfied (see Proposition 8.1 below). This leads
to modification of the notion of linearized stability.
Recall from Subsection 2.3, that we say that the dynamical system (7.28) has a sym-
metry group G if a representation T : g → Tg of G on the space of solutions satisfies

F (Tg u) = Tg F (u). (8.1)

(See Subsection 2.3 for examples.) We assume that G is a (matrix) Lie group and let G
denote the space spanned by the generators A := ∂s |s=0 TUs , where Us are various one-
parameter subgroups of G (see (2.48) - (2.49)). (In this case, TUs is a one-parameter
subgroup of the group {Tg : g ∈ G}.)
In the case of symmetry, inequality (7.29) never holds: the Gâteaux derivative dF (u∗ )
has always zero eigenvectors. Indeed, we have
76 Lectures on Applied PDEs, November 3, 2017

Proposition 8.1. If A ∈ G (one of generators of the algebra of the group G), then Au∗
is an eigenfunction of dF (u∗ ) with eigenvalue 0,
dF (u∗ )Au∗ = 0. (8.2)
Proof. If u∗ is a static solution to (7.28), then, due to (9.2), so is Tg u∗ , for any g ∈ G.
In particular, F (esA u∗ ) = 0, for any A ∈ G and s ∈ R. Differentiate this equation with
respect to s at s = 0 and use that ∂s F (esA u∗ ) s=0 = Au∗ to obtain

0 = ∂s F (esA u∗ ) s=0 = F (u∗ )Au∗ .
Hence (8.2) follows.
A specific example is given in the next subsection.
Thus in the case of symmetries, if Au∗ 6= 0 for some A ∈ G, then the assumption
(7.33) fails. We replace it by the following weaker assumption
• Assume for simplicity dF (u∗ ) is self adjoint. Then we say u∗ is linearly orbitally
stable iff
hξ, dF (u∗ )ξi ≤ −θkξk2 , for some θ > 0 and ∀ξ ⊥ Au∗ , ∀A ∈ G. (8.3)

If Au∗ 6= 0, or what is the same esA u∗ 6= u∗ , for some A ∈ G, then we say that u∗ breaks the
symmetry corresponding to the one-parameter subgroup {esA }. For a given u∗ , consider
the maximal subgroup, G∗ , of G, which is broken by u∗ , i.e. gu∗ 6= u∗ , ∀g ∈ G∗ . In what
follows, we use this subgroup instead of G.

8.2 Linearized stability of kinks


Consider the linearized stability of the kink solutions of the Allen - Cahn equation (see
(2.2))
∂u
= ∆u − g(u), (8.4)
∂t
where g(u) = u3 − u, or, more generally, g : R → R is the derivative, g = G0 , of a
double-well potential: G(u) ≥ 0 and has two non-degenerate global minima, say at ±1,
with the minimum value 0:
g(±1) = G0 (±1) = 0 and g 0 (±1) = G00 (±1) > 0. (8.5)
For g(u) = u3 − u, we take G(u) = 21 (u2 − 1)2 . G(u) is also called a bistable potential,
see Figure 18.

To keep the notation simple, assume we are in the dimension 1, i.e. u : R × R+ → R.


Recall from Section 2.1 that (8.4) has the kink static solution by χ(x), see Figure 19 (for
g(u) = u3 − u, we have ). (8.4) is translationally symmetric. Hence we expect that (7.33)
fails and we use the definition (8.3). We have
Lectures on Applied PDEs, November 3, 2017 77

u
−1 1

Figure 14: Double well potential

G u

a b x
a

Figure 15: Hill-valley-hill structure for −G, and the kink for u.

Theorem 8.2. The kink solutions of (8.4) are linearly orbitally stable w.r.to H 1 (R)−perturbations
in the sense of definition (8.3).

Proof. If we denote the negative of the r.h.s. of (8.4) by F (u),

F (u) := ∆u − g(u), (8.6)

then the linearization L := −dF (χ) of F at χ isis computed explicitly as

L := −∆ + g 0 (χ). (8.7)

The statement of the theorem follows from the theorem below.

Theorem 8.3. (a) The operator L is self-adjoint. (b) Let ρ be the second lowest eigenvalue
of L. Then, we have
hξ, Lξi ≥ ρkξk2L2 , ∀ξ⊥χ0 . (8.8)

Proof. It is straightforward to show that the operator L is symmetric. The self-adjointness


is a stronger properties and follows from observing that L is of a Schrödinger type and
standard results about Schödinger operators (see e.g. [6]).
To prove (b), we observe that the operator L has the following properties:

1. the spectrum of L in (−∞, min{g 0 (1), g 0 (−1)}) consists of isolated eigenvalues of


finite multiplicities, which might accumulate only at min{g 0 (1), g 0 (−1)};

2. 0 is the smallest eigenvalue of L, it is simple with the eigenfunction χ0 .


78 Lectures on Applied PDEs, November 3, 2017

(For the second statement, cf. Proposition 8.1.)


To derive the first property, given an operator L on L2 (Rn ), we define the number

µ(L) := lim inf{hξ, Lξi : ξ ∈ C0∞ (Rn /BR )},


R→∞

where BR is the ball of the radius R > 0 in Rn . It is a result in the spectral theory
(see e.g. [6]) that the spectrum of a Schrödinger type operator L in (−∞, µ) consists of
isolated eigenvalues of finite multiplicities, which might accumulate only at µ(L).
We show now that for (15.26),

µ(L) ≥ min{g 0 (1), g 0 (−1)}, (8.9)

which would prove the first statement. To prove this statement, we will use that |χ(x) ∓
1| ≤ |x|θ , for some θ > 0, as x → ±∞. (In fact, it is exponentially small.) We will not
prove the latter statement, but just mention that for g(u) = u3 − u, we have the explicit
solution  
x
χ(x) = tahn √ , (8.10)
2
which shows that χ(x) → ±1, as x → ±∞, exponentially.
Now, we compute, differentiating by parts
Z Z
hξ, Lξi = (|ξ | + g (χ)|ξ| ) ≥ g 0 (χ)|ξ|2 .
0 2 0 2

Since ξ ∈ C0∞ (Rn /BR ), we can write


Z Z −R Z ∞ Z −R Z ∞
0 2 0 2 0 2 0 2
g (χ)|ξ| = g (χ)|ξ| + g (χ)|ξ| = g (−1)|ξ| + g 0 (1)|ξ|2
−∞ R −∞ R

Z −R Z ∞
0 0 2
+ (g (χ) − g (−1))|ξ| + (g 0 (χ) − g 0 (1))|ξ|2 .
−∞ R

Since |χ(x) ∓ 1| ≤ |x|θ and g 0 (±1) > 0 (see (8.5)), we have


Z Z −R Z ∞
0 2 0 2 0
g (χ)|ξ| ≥ g (−1) |ξ| + g (1) |ξ|2
−∞ R

Z −R Z ∞
−θ 2
−cR ( |ξ| + |ξ|2 ),
−∞ R
R∞ R∞
which, due to ξ ∈ C0∞ (Rn /BR ), gives g 0 (χ)|ξ|2 ≥ min(g 0 (±1)) |ξ|2 − cR−θ |ξ|2 ),
R
−∞ −∞
which implies (8.9). This proves the first statement.
Lectures on Applied PDEs, November 3, 2017 79

To prove the second statement, we observe that, by Proposition 8.1 and the transla-
tional invariance of (8.4), 0 is an eigenvalue of L with eigenfunction ζ := ∂x χ,

Lχ0 = 0. (8.11)

For an exercise we prove this directly. Since χ solution to F (u) = 0 (see Section 2.1 and
(2.5)), where the map F is defined in (15.25), hence, for any a ∈ R, χa is also a solution
to F (χa ) = 0, where χa = ū(x + a). Differentiate this equation with respect to a:

∂a F (χa )|a=0 = ∂F (χa )∂a χ|a=0 = Lζ ,

where, recall L := ∂F (χ). Hence we have Lζ = 0 as claimed.


Now, since χ0 is positive, we have by the Perron - Frobenius theory (see Appendix
E.3) that 0 is the smallest eigenvalue of L and it is simple.
The above properties and a spectral theorem (see [6]) imply that

hLξ, ξi ≥ ρkξk2L2 (8.12)

for all ξ⊥χ0 , where ρ is the second lowest eigenvalue of L, which completes the proof.

8.3 Linearized stability of spheres, cylinders and planes


Below, all differential operations, e.g. ∇, ∆, are defined in the corresponding Euclidian
space (either Rn+1 or Rn ). We are interested in describing hypersurfaces, S, in Rn+1 (i.e.
surfaces of dimension n), whose mean curvature, H, is constant, i.e. satisfies the equation

H(x) = h. (8.13)

Such surfaces are called the constant mean curvature surfaces if h 6= 0 and the minimal
surfaces if h = 0. Recall from Appendix 6.6 that, if a surface S given in the level set
representation, S = {x0 ∈ Rn+1 : ϕ(x0 ) = 0}, then its mean curvature is given by

∇ϕ ∇ϕ
 
0 0
H(x ) ≡ H(ϕ)(x ) = div (H(x0 ) = div ν(x0 ) and ν(x0 ) = ), (8.14)
|∇ϕ| |∇ϕ|

where x = (x1 , . . . , xn ) and x0 = (x, xn+1 ). We give expressions for the mean curvatures
for hypersurface given as graphs over other surfaces.

1) Graph representation over a plane: S = graph of f : U → R, where U is an open set


in Rn ⊂ Rn+1 . In this case S is the zero level set of the function ϕ(x0 ) = xn+1 −f (x).
Using (8.14) with this function, we obtain
!
∇f
H(x) = div p . (8.15)
|∇f |2 + 1
80 Lectures on Applied PDEs, November 3, 2017

∂2f

Denote by Hess f the standard euclidean hessian, Hess f := ∂ui ∂uj
. Then we can
rewrite (8.16) as
1 ∇f Hess f ∇f
H(x) = p [∆f − ]. (8.16)
2
|∇f | + 1 |∇f |2 + 1

2) Graphs over sphere: S = graph of ρ : Sn → R+ , where Sn is the standard n−unit


sphere in Rn+1 centred at the origin. In this case S is the zero level set of the
function ϕ(x0 ) = |x0 | − ρ(x̂0 ), where x̂ = x/|x|. Using (8.14) with this function, we
obtain

2) Graphs over a cylinder: S = graph of ρ : C n → R+ , where C n is the standard n−unit


cylinder in Rn+1 with the axis along xn+1 −axis. In this case S is the zero level set
of the function ϕ(x0 ) = |x| − ρ(xn+1 , x̂), where x̂ = x/|x|. Using (8.14) with this
function, we obtain

Spherically and axi - symmetric (equivariant) solutions:

a) Sphere. The n−dimensional sphere of the radius R centred at the origin can be
the level set {ϕ(x) = 0}, where ϕ(x) := |x|2 − R2 , or as graph f ,
given either asp
where f (u) = R2 − |u|2 . (SR = RS n , where S n is the unit n-sphere.) Then we
have
∇ϕ
 
n
H(x) = div = div(x̂) = .
|∇ϕ| R

b) Cylinder. In the implicit function representation the cyliner is given by {ϕ = 0},


Pn−1 2  12
where ϕ := r − R, with r = i=1 xi .

Properties of (8.13):

• (8.13) is invariant under rigid motions of the surface, i.e. ψ 7→ Rψ + a, where


R ∈ O(n + 1), a ∈ Rn+1 and ψ = ψ(u) is a parametrization of St , is a symmetry of
(8.13).

• (8.13), with h = 0, is invariant under the scaling, for any λ > 0,

H(x) 7→ λH(λx). (8.17)

The first two properties come from the fact that the mean curvature is invariant under
translations, rotations and scaling. The invariance under rigid motions is obvious.

Hessians for spheres and cylinders. We finish this section with the discussion of the
normal hessians on the n−sphere and (n, k)−cylinder.
Lectures on Applied PDEs, November 3, 2017 81

Explicit expressions. We define the map


p nF (ϕ) := H(ϕ) − h and compute dF (ϕ).
1) For the n−sphere SR of radius R = a in R , we have
n n+1

sph 1 n
HessN
sph LR = − 2
∆Sn − 2 2 , (8.18)
R R
on L2 (Sn ), where ∆Sk is the Laplace-Beltrami operator onqthe standard n−sphere Sk .
2) For the n−cylinder CRn = SRn−k × Rk of radius R = n−k a
in Rn+1 , we have

1 n−k
Lcyl
R = −∆y − 2
∆Sn−k − 2 2 , (8.19)
R R
acting on L2 (C n ).

Spectra of Lsph a and Lcyl


a . We describep the spectra
p a of the normal hessians on the
n−sphere and (n, k)−cylinder, of the radii na and n−1 , respectively.
It is a standard fact that the operator −∆ = −∆Sn is a self-adjoint operator on
L2 (Sn ). Its spectrum is well known (see [17]):  it consists
  of the eigenvalues
 l(l + n −
n+l n+l−2
1), l = 0, 1, . . . , of the multiplicities m` = − . Moreover, the
n n
eigenfunctions corresponding to the eigenvalue l(l + n − 1) are the restrictions to the
sphere Sn of harmonic polynomials on Rn+1 of degree l and denoted by Ylm (the spherical
harmonics),

− ∆Ylm = l(l + n − 1)Ylm , l = 0, 1, 2, 3, . . . , m = 1, 2, . . . , ml . (8.20)

In particular, the first eigenvalue 0 has the only eigenfunction 1 and the second eigenvalue
n has the eigenfunctions ω 1 , · · · , ω n+1 .
Consequently, the operator Lsph R = − na (∆Sn + 2n) is self-adjoint and its spectrum
consists of the eigenvalues na (l(l + n − 1) − 2n) = a(l − 2) + na l(l − 1), l = 0, 1, . . . , of the
multiplicities m` . In particular, the first n + 2 eigenvectors of Lspha (those with l = 0, 1)
correspond to the non-positive eigenvalues,

Lsph 0 0
a ω = −2aω , Lsph j j
a ω = −aω , j = 1, . . . , n + 1, (8.21)

and are due to the scaling (l = 0) and translation (l = 1) symmetries.


We proceed to the cylindrical hessian Lcyla , given in (17.32). The variables in this
a
operator separate and we can analyze the operators −∆y − ay · ∇ and − n−k (∆Sn−k +
a
2(n − k)) separately. The operator − n−k (∆Sn−k + 2n) was already analyzed above. The
operator −∆y − ay · ∇ is the Ornstein - Uhlenbeck generator, which can be unitarily
mapped by the gauge transformation
a 2
v(y, w) → v(y, w)e− 4 |y|
82 Lectures on Applied PDEs, November 3, 2017

into the the harmonic oscillator Hamiltonian Hharm := −∆y + 14 a2 |y|2 − ka. Hence the
a 2
linear operator −∆y − ay · ∇ is self-adjoint on the Hilbert space L2 (R, e− 2 |y| dy). Since,
a
as was already mentioned, the operator − n−k (∆Sn−k + 2(n − k)) is self-adjoint on the
Hilbert space L (C ), we conclude that the linear operator Lcyl
2 n
a is self-adjoint on the
n−k − a2 |y|2
Hilbert
n P space L2
(R k
× S o, e dydw). Moreover, the spectrum of −∆y − ay · ∇ is
k
a 1 si : si = 0, 1, 2, 3, . . . , with the normalized eigenvectors denoted by φs,a (y), s =
(s1 , . . . , sk ),
X k
(−∆y − ay · ∇)φs,a = a si φs,a , si = 0, 1, 2, 3, . . . . (8.22)
1
a a
Using that we have shown that the spectrum of − n−k (∆Sn−k + 2(n − k)) is n−k (l(l +
a
P k
n−k −1)−2(n−k)) = n−k l(l +n−k −1)−2a, l = 0, 1, . . . , and denoting r = 1 si , si =
0, 1, 2, 3, . . ., we conclude the spectrum of the linear operator Lcyl a , for k = 1, is
 
cyl a
spec(La ) = (r − 2)a + `(` + n − 2) : r = 0, 1, 2, 3, . . . ; ` = 0, 1, 2, . . . , (8.23)
n−1
with the normalized eigenvectors given by φr,l,m,a (y, w) := φr,a (y)Ylm (w). This equation
shows that the non-positive eigenvalues of the operator Lcyl
a , for k = 1, are
a 4 1
• the eigenvalue −2a of the multiplicity 1 with the eigenfunction φ0,0,0,a (y) = ( 2π )
((r, l) = (0, 0)), due to scaling of the transverse sphere;
a 4 m 1
• the eigenvalue −a of the multiplicity n with the eigenfunctions φ0,1,m,a (y) = ( 2π ) w , m=
1, . . . , n ((r, l) = (1, 1)), due to transverse translations;
a 14 √
• the eigenvalue 0 of the multiplicity n with the eigenfunctions φ1,1,m,a (y) = ( 2π ) aywm , m =
1, . . . , n ((r, l) = (0, 1)), due to rotation of the cylinder;
a 14 √
• the eigenvalue −a of the multiplicity 1 with the eigenfunction φ1,0,0,a (y) = ( 2π ) ay
((r, l) = (1, 0));
a 4 1
• the eigenvalue 0 of the multiplicity 1 with the eigenfunction φ2,0,0,a (y) = ( 2π ) (1 −
2
ay ) ((k, l) = (2, 0)).
The last two eigenvalues are not of the broken symmetry origin and are not covered by
Theorem 17.8. They indicate instability of the cylindrical collapse
By the description of the spectra of the normal hessians of Lspha := HessNsph Va (ϕ) and
cyl N
La := Hesscyl Va (ϕ), we conclude that the spherical collapse is linearly stable while the
cylindrical one is not.
It is shown in [16] that indeed the spherical collapse is (nonlinearly) stable while the
cylindrical one is not. We will also show that the last two eigenvalues of Lcyla in the list
above are due to translations of the point of the neckpinch on the axis of the cylinder and
due to shape instability, respectively.
Lectures on Applied PDEs, November 3, 2017 83

9 Nonlinear stability
9.1 Stability: generalities
So far we studied mainly existence of static solutions and considered their linearized
stability (see Subsection 7.4). (Remember that at the same time this gives also the
existence and the linearized stability of of stationary or standing waves and traveling
wave solutions.)
Now, we address the full nonlinear stability, which is the next key step. The question
here is, starting with initial conditions close to a static solution, how the solutions of the
dynamical equation in question behave? This leads to the question of stability of static
solutions of general dynamical systems,
∂u
= F (u). (9.1)
∂t
Assume F : U → Y , where U is an open set in X and X and Y are Hilbert spaces, X ⊂ Y ,
densely. Let (9.1) have a static solution, u∗ , i.e., u∗ is time independent. (Such solutions
are also called equilibria and sometimes stationary solutions.) Thus u∗ is an equilibrium
or static solution iff u∗ is independent of time and satisfy the equation F (u∗ ) = 0.
We would like to understand the behavior of solutions to equation (9.1) for initial
conditions u0 near an equilibrium one u∗ . There are the following general scenarios
• The solutions stay in a neighborhood of u∗ (Lyapunov stability);
• The solutions converge to u∗ as t → +∞ (asymptotic stability);
• The solutions move away from u∗ as t → ∞ (instability).
More precisely, we say that a static solution of (9.1) is Lyapunov stable if for any
neighborhood, U of u∗ have another neighborhood, V ⊂ U , of u∗ such that if an initial
condition, u0 , is in V , then the solution, u, stays in U . Otherwise, u∗ is said to be unstable.
We say a static solution, u∗ , is asymptotically stable if ∃δ > 0 such that for any initial
condition u0 satisfying ku0 − u∗ k < δ, we have limt→∞ kφt (u0 ) − u∗ k = 0.
For systems with symmetry, the notion of stability/instability should be modified.
This is done below.
A weaker notion of stability - the linear stability - introduced in Subsection 7.4 is
usually is the first step in proving asymptotic stability and often gives the necessary
condition.

Stability in the presence of symmetry. Recall from Subsection 2.3, that we say
that the dynamical system (7.28) has a symmetry group G if a representation T : g → Tg
of G on the space of solutions satisfies
F (Tg u) = Tg F (u). (9.2)
84 Lectures on Applied PDEs, November 3, 2017

(See Subsection 2.3 for examples.) In this case, if the dynamical system (7.28) with the
symmetry group G has a static solution u∗ , then it has the manifold of static solutions

M∗ = {Tg u∗ : g ∈ G}.

Indeed, if u∗ is a static solution to (9.1), then, due to (9.2), so is Tg u∗ . Now, we have


to address the stability w.r.to this manifold. Clearly, we have to modify the definitionsat
the beginning of this section and the approach.
We say that a static solution u∗ (or more precisely, the manifold of static solutions M∗ )
is orbitally stable if any solution to the equation (7.28), starting in a small neighborhood
of u∗ stays in a small neighborhood of the manifold M∗ . In other words the solution ut
sticks very close to a possibly moving static solution Tg(t) u∗ .
Note that the tangent space, Tu∗ M∗ ⊂ X, to M∗ at u∗ is the vector space Gu∗ , where
G was defined above.
As an example we consider Eq. (8.4). This equation is translation invariant and has the
kink solution χ which breaks the translational symmetry. Hence if ū is a solution to (8.4)
then so is ūa and therefore the functions χa (x) := χ(x + a) ∀a ∈ R are also static solutions
to (8.4). Thus we have an entire manifold of stationary solutions Mkink = {χa : a ∈ R}.
(In the multidimensional case, we have also rotations.)
The energy method is particularly useful in proving the Lyapunov and orbital stability.
We demonstrate this in Subsection 15 after developing some variational calculus.

9.2 Asymptotic stability of kinks


(insert remarks about GWP) We are interested in the asymptotic stability of the
families of kink solutions of the Allen-Cahn equation (15.23) on R, discussed in Subsection
15.3. We recall this equation here:
 ∂u
= ∆u − g(u),
∂t (9.3)
u|t=0 = u0 (x),

where u : R → R and g : R → R is the derivative of a double-well potential (i.e. G(u) ≥ 0


and has two non-degenerate global minima with the minimum value 0, see Figure 2,
specifically, g(u) = u3 − u), and we set (say by rescaling) the parameter  to 1.
Recall that the equation (9.3) has a family (manifold) of stationary solutions, χ(x −
a), a ∈ R, where χ(z) is a smooth function satisfying χ(z) → ±1, as z → ±∞. Such
solutions are called the kinks. (χ(−x+a) are also solutions and they are called anti-kinks.)
As stated above, we are interested in the asymptotic stability of χ(x). Since χ(x) are
not in L2 , we consider (9.3) (in a weak sense) on the space X := χa + H 1 , where χa is
the translate of χ, with the notation fa (x) := f (x − a), a ∈ R, and H 1 is the standard
Sobolev space of order 1.
Lectures on Applied PDEs, November 3, 2017 85

Our result below concerns asymptotic stability of the family (manifold) of the kink
solutions
Mkink := {χ(x − a) | a ∈ R}. (9.4)
We assume that g is such that the initial value problem for (9.3) is locally well-posed
in X in H 2 (Rn ) for any initial condition u0 ∈ X, sufficiently close to Mkink . Moreover,
we assume that the nonlinearity
N (ξ) := g(χ + ξ) − g(χ) − g 0 (χ)ξ, (9.5)
satisfies the following conditions:
|N (ξ)| . |ξ|2 + |ξ|k , k > 2. (9.6)
This obviously holds for g(u) = u3 − u with k = 3. Then we have
Theorem 9.1. (Asymptotic stability of kinks) Under the assumptions above, the kink
manifold Mkink is asymptotically stable. More precisely, there are δ > 0 and a(t) s.t. if
dist(u0 , Mkink ) ≤ δ, then the solution u to (15.23) with u|t=0 = u0 satisfies
ρ
ku − χa(t) kH 1 . e− 4 t (9.7)
for all times t ≥ 0. Moreover, there exists a real number a∞ such that a(t) → a∞ ,
exponentially fast, as t → ∞.
Recall that (9.3) has the decreasing energy given by (see (15.24))
1
Z
E(u) = ( |∇u|2 + G(u)) , (9.8)
2
defined on χ + H 1 (R). To prove the theorem we use, instead the energy dissipation
property, the differential inequality for the Lyapunov functional Λ(ξ) := 21 hLξ, ξi, where,
recall, L is the hessian of the energy functions E(u) at χ: L := E 00 (χ), which is closely
related to the energy E(χ), via the Taylor expansion.

Remark. By Theorem 4.5 and remarks after it, we have the local existence for (9.3).
Indeed, by writing u = χ + ξ and plugging this into (9.3), we arrive at the equation for ξ
of the NLS type discussed in the remarks after Theorem 4.5.

9.3 Orthogonal Decomposition


Throughout, we will use the notation fa (x) = f (x − a). Similarly to the general case,
(15.15), we introduce a neighbourhood of Mkink :
Uδ = {u ∈ H 1 (Ω) : ∃a s.t. ku − χa kH 1 (Ω) ≤ δ}. (9.9)
The following statement and its proof are specialization of Proposition 15.4 and its proof
to the present situation.
86 Lectures on Applied PDEs, November 3, 2017

Proposition 9.2. There is δ > 0 s.t. any u ∈ Uδ can be decomposed as follows


Z
u = χa + ξa , with ξ(x)χ0 (x) dx = 0. (9.10)

Proof. We define the function G(a, u) = (u − χa )χ0a dx and rewrite the statement of the
R

proposition as solving the equation

G(a, χa ) = 0 (9.11)

for a as a function of u. To do this, we use the Implicit Function Theorem. Clearly, G is


C 1 in u because is it a linear functional in u, and 1 1
C in a since χa is C in a. Furthermore,
G(a, χa ) = 0. Finally, we show that ∂a G(a, u) u=χa 6= 0:
Z Z
0
∂a G(a, u) = − ∂a χa χa dx + (u − χa )∂a χa dx. (9.12)

Hence, using that ∂a χ = − 1 χ0 , we have, at u = χa ,


1
Z
∂a G(a, u) = (χ0a )2 dx. (9.13)

Therefore, the conditions of the Implicit Function Theorem are satisfied and it implies
that (9.11) has a unique solution for a as a function of u ∈ Uδ , which gives the statement
of the proposition.

9.4 Evolution of Fluctuation ξ


Recall from the previous section that we denote the negative of the r.h.s. of (9.3) by
F (u):
F (u) := −∆u + g(u). (9.14)
Then the linearization L := dF (χ) is computed explicitly as

L := −∆ + g 0 (χ). (9.15)

Proposition 9.3. Assume that for every t, a solution u of (9.3) is in Uδ , with δ > 0
given in Proposition 15.4. Then a(t) and ξ(x, t) given in that proposition (u = χa + ξa )
satisfy the following equations

− ∂t ξ = Lξ + N (ξ) + ȧχ0 + ȧ∂x ξ, (9.16)

where N (ξ) := g(χ + ξ) − g(χ) − g 0 (χ)ξ, and L := dF (χ), given explicitly by (9.15), and
Z Z  Z
ȧ(t) (χ ) dx + ∂x ξχ dx = N (ξ)χ0 dx.
0 2 0
(9.17)
Lectures on Applied PDEs, November 3, 2017 87

Proof. For any t decompose a solution u of (9.3) according to Proposition 15.4, u(x, t) =
χa(t) + ξa(t) (x, t). Substituting the decomposition u(x, t) = χa(t) + ξa(t) (x, t) into (9.3), we
see that our equation becomes ∂x χa ȧ + ∂t ξa + ∂x ξa ȧ = −F (χa + ξa ), where F is given in
(9.14). By changing the variables x → x + a in the last equation, we obtain

−∂t ξ − ∂x ξ ȧ − ∂x χa ȧ = F (χ + ξ). (9.18)

We will need the Taylor expansion of F ,

F (χ + ξ) = F (χ) + dF (χ)ξ + N (ξ), (9.19)

where N (ξ) := g(χ + ξ) − g(χ) − g 0 (χ)ξ, and the observation that, by the definition of F ,
F (χ) = 0 (see the sentence after (8.11)), to obtain

F (χ + ξ) = Lξ + N (ξ), (9.20)

where L := dF (χ). Equations (9.18) and (9.20) give (9.16).


We have shown in the proof of Proposition 15.6 that since the kink χ breaks the
translational symmetry, χ0 is a zero eigenfunction of the operator L:

Lχ0 = 0. (9.21)

Equation (9.16) contains two unknowns: a(t) and ξ(x, t). To derive a separate equation
for a(t) we project equation (9.16) onto Cχ0 by multiplying by χ0 and integrating over x,
to get
Z Z Z Z Z
ȧ(t) (χ ) dx + ȧ(t) ∂x ξχ dxn = ∂t ξχ dx + Lξχ dx + N (ξ)χ0 dx.
0 2 0 0 0
(9.22)

For the first term on the r.h.s. we have


Z Z
∂t ξχ dx = − ξ∂t (χ0 ) dx = 0.
0
(9.23)

For
R the second term on the r.h.s., we use (8.11) and the self-adjointness of L to obtain
0
Lξχ dx = 0. The last relation together with (9.23) implies (9.17).

9.5 Bound on a(t)


Proposition 9.4. We assume that kξkL2 (dx) ≤ 1/2. Then, if u satisfies (15.23) and a(t)
is defined by (15.16) then a(t) satisfies

|ȧ(t)| . kξk2L2 + kξkkH 1 , (9.24)

where k > 2 is given in (9.6).


88 Lectures on Applied PDEs, November 3, 2017

Proof. We use equation (9.17) and estimate the right-hand side of this equation. We have,
by the assumption (9.6) and a Sobolev embedding theorem,
Z
N (ξ)χ0 dx ≤ kN (ξ)kL1 kχ0 kL∞


. kξk2L2 + kξkkLk
. kξk2L2 + kξkkH 1 . (9.25)
We estimates the second term on the left-hand side of (9.17). Using integration by
parts, we have
Z Z
0
| ∂x ξχ dx| = |ξχ00 dx| (9.26)
≤ kξkL2 kχ00 kL2 . (9.27)
Combining this estimate with (9.17) and (9.25) gives (9.24).

9.6 Upper Bound on kξkH 1


Proposition 9.5. If u solves (9.3), and ξ is the fluctuation defined by (15.16), then
∂t hξ, Lξi + kLξk2L2 + 2ρhξ, Lξi ≤ Chξ, Lξi2 + Chξ, Lξik (9.28)
where ρ is the constant from Proposition 15.6.
Proof. We define the Lyapunov functional Λ(ξ) := 21 hLξ, ξi. We differentiate Λ(ξ) and
˙ Lξi. Using (9.16) on the r.h.s. of the
use the self-adjointness of L to find ∂t Λ(ξ) = hξ,
latter expression gives
∂t Λ(ξ) = −hLξ, Lξi + hN (ξ), Lξi − ȧ(t)hχ0 , Lξ i. (9.29)
Using that L is self-adjoint (see the proof of Proposition 15.6) and using (9.21), we have
that hχ0 , Lξi = 0. Therefore, using (9.29), we obtain
∂t Λ(ξ) + kLξk2L2 = −hLξ, N (ξ)i. (9.30)
Now we prove the following estimate of the terms on the right-hand side:
1
|hLξ, N (ξ)i| ≤ C(kξk4H 1 + kξk2k 2
H 1 ) + kLξkL2 . (9.31)
2
We start by using the bound on the nonlinearity, (9.6) to get
|hLξ, N (ξ)i| ≤ kN (ξ)kL2 kLξkL2 (9.32)
. (kξk2L4 + kξkkL2k )kLξkL2 (9.33)
1
≤ (kξk4L4 + kξk2k 2
L2k ) + δkLξkL2 . (9.34)

Lectures on Applied PDEs, November 3, 2017 89

We use the Sobolev embedding inequality to find that we have (9.31).


Finally, we prove the estimate
1
kLξk2L2 ≥ ρhLξ, ξi. (9.35)
2
To this end, we use the Cauchy-Schwarz inequality and the inequality (8.12) to find
1
ρkξk2 hLξ, ξi ≤ hLξ, ξi2 ≤ kLξk2 kξk2 , (9.36)
2
which implies (9.35).
Writing kLξk2L2 = 21 kLξk2L2 + 21 kLξk2L2 and using (9.35) gives kLξk2L2 ≥ ρΛ(ξ) +
1
2
kLξk2L2 . Substituting this and the inequality (9.31) into (9.30), we obtain
1
∂t Λ(ξ) + kLξk2L2 + ρΛ(ξ) . C(kξk4H 1 + kξk2k
H 1 ). (9.37)
2
Due to the inequality (15.27), this implies (9.28).
Proposition 9.6. If u solves (9.3) and u ∈ Uδ , where δ > 0 given in Proposition 15.4,
so that (15.16) hold and if ξ is the fluctuation defined by (15.16), then
ρ
kξkH 1 . e− 4 t kξ0 kH 1 , (9.38)
where ρ is the constant from Proposition 15.6.
ρ
Proof. We set X(t) := e 2 t Λ(ξ(t)) and use (9.28) to find Ẋ ≤ Ce−αρt X 2 , assuming for
simplicity that X . 1 and consequently dropping the term X k . Therefore,
Z t
X(t) ≤ X0 + C e−αρt X 2 (s)ds. (9.39)
0

Setting M = sups∈[0,t] Λ(ξ(s)), we have


ρ
M . e− 2 t M0 + CM 2 . (9.40)
ρ
Therefore, M ≤ e− 2 t M0 provided t is sufficiently large. Since M ≥ 21 hLξ, ξi & kξk2H 1 , we
have (9.38).
Now, assume u0 ∈ Uδ/2 , where δ > 0 given in Proposition 15.4. Then by the local
well-posedness (see the remark preceding Subsection 9.3), there is T > 0, s.t. the solution
u(t) satisfies u(t) ∈ Uδ for 0 ≤ t ≤ T . Then Proposition 15.4 and therefore Proposition
9.6 hold for this u(t) for 0 ≤ t ≤ T . This gives the bound (9.38), which shows that, in
fact, u(t) ∈ Uδ/2 . Continuing in this fashion, we find that the solution u(t) exists for all
t > 0 and satisfies u(t) ∈ Uδ and the bound (9.38). The bound (9.38) and the definition
of ξ in Proposition 15.4 imply (9.7).
It remains to prove the convergence of a(t).
90 Lectures on Applied PDEs, November 3, 2017

9.7 Asymptotic Behaviour


In the last subsection we proved that the solution u = χa + ξa approaches the manifold of
kinks Mk ; however, this does not imply the solution approaches a particular kink on the
manifold. We prove in this subsection that the limit does exist. More precisely, we prove
that a(t) → a∞ as t → ∞.
As above, we assume u0 ∈ Uδ/2 , where δ > 0 given in Proposition 15.4. Then, as was
shown above, the solution u(t) exists for all t > 0 and satisfies u(t) ∈ Uδ and the bound
(9.38), which, together with the definition of ξ in Proposition 15.4, implies (9.7).
In this case, Proposition 9.4 is applicable and gives the estimate (9.24) on a, which
together with (9.38) implies that a(t) satisfies
ρ
|ȧ(t)| . e− 2 t kξ0 k2H 1 . (9.41)
Rt
This shows that the function a(t) = a(0) + 0 ȧ(s) ds converges as t → ∞ and
Z ∞
ρ
|a(t) − a(∞)| ≤ |ȧ(s)| ds . e− 2 t kξ0 k2H 1 . (9.42)
t

We now complete the proof. The triangle inequality and the mean value theorem
imply
ku(t) − χa∞ kH 1 . u(t) − χa(t) H 1 + |a(t) − a∞ |,
which together with (9.7) and (9.42) completes the proof of Theorem 9.1.

10 Elements of Variational Calculus


In this section, we introduce an important structure into the manifold of partial differ-
ential equations - the variational structure. It distinguishes a class of partial differential
equations for which we can provide a uniform treatment of existence and stability prob-
lems.

10.1 Functionals
Functionals are maps which have R as the target space. More precisely, let X be a vector
space, and M ⊂ X, a not necessarily open subset of X. Then a functional is a map
E : M → R. Usually, X is a space of functions. If X has a basis, then functionals on X
can be represented as functions of an infinite number of coordinates along the basis. If X
is a finite–dimensional space (which we are not concerned with here), then a functional
on X is just a usual function of several variables. In the following list of examples, Ω is
a domain in Rn and G : R → R, u : Ω → R, or u : Ω → C:
R
1) E(u) = Ω G(u(x))dnx,
Lectures on Applied PDEs, November 3, 2017 91

1
|∇u|2 dnx,
R
2) E(u) = 2 Ω

( 1 |∇u|2
R
3) E(u) = Ω 2
+ G(u))dnx.

We also have to specify the spaces on which the functionals are defined. The (not
necessarily linear or vector) spaces are chosen according to the specific functional and the
problem at hand. Of course, we always try to choose the simplest possible space for a
given problem.
For instance consider example 1). Assume that the domain Ω is bounded, and that
the function G satisfies the estimate

|G(u)| ≤ C|u|p + C, (10.1)

for some constant C > 0. It is then natural to define E on the space Lp (Ω).
In example 2), we define E on the Sobolev–space H 1 (Ω), defined in Section A.6. In
example 3), if Ω is bounded, and G satisfies (10.1), then we define E on H 1 (Ω) ∩ Lp (Ω).

Exercise 10.1. For which p and n, the functional E(u) = Ω ( 21 |∇u|2 +G(u))dnx satisfying
R

(10.1) and with Ω bounded is defined on H 1 (Ω)?

An important example of functionals are linear functionals on a vector space X. This


is the special case of linear operators when the target space is Y = C. If X is a normed
space, then we consider bounded or continuous linear functionals, for which the following
norm is finite:
klk = sup |l(x)|. (10.2)
kxk=1

10.2 The Gâteaux derivative for functionals


Consider a functional E : M → R. If M is an open subset of a vector space X, then the
Gâteaux derivative dE(u), u ∈ M , is a functional on X defined, for every ξ ∈ X, as


dE(u)ξ = E(uλ ) λ=0 , (10.3)
∂λ
where uλ := u + λξ if the latter derivative exists. We say that E is C 1 in U ⊂ M if and
only if ∀ u ∈ U , dE(u) is a bounded linear functional.

Exercise 10.2. Show


that the definition
(10.3) is independent of the choice of the curve

uλ (as long as uλ λ=0 = u and ∂λ uλ ) λ=0 = ξ).

Remark 10.1. For the differentiability (C 1 ), it suffices to require that the map u →
dE(u)ξ, from U to R, is continuous, for every ξ ∈ X. Then the linearity of dE(u) follows
from this requirement, see Remark 6.2 of Subsection 6.1.
92 Lectures on Applied PDEs, November 3, 2017

Let us now consider a simple example. Let G be a real differentiable function on R


satisfying the estimate
|G(u)| + |G0 (u)| ≤ C|u|2 ,
where C is independent of u ∈ R, and let Ω ⊂ Rn . Then the functional
Z
V (u) := G(u(x))dnx

is defined on L2 (Ω). Using the definition, dV (u)ξ = ∂λ∂


V (u + λξ)|λ=0 , we compute its
Gâteaux derivative
 Z  Z
n
d G◦u d x ξ = G0 (u(x))ξ(x)dnx.
Ω Ω

Thus the Gâteaux derivative in this case is the linear functional standing on the r.h.s..

Exercise 10.3. Compute the Gâteaux derivatives in examples 1)–3).

10.3 Critical points and connection to PDEs


Let M be an open set in a real vector space. Given a C 1 –functional E : M → R, we say
that u∗ ∈ M is a critical point (CP) of E if and only if dE(u∗ ) = 0.

Exercise 10.4. Find the equations for the critical points in examples 1)–3) given at the
beginning of this section.

The equation dE(u∗ ) = 0 (or, in detail, dE(u∗ )ξ = 0, for every ξ ∈ X) for critical
points of E is sometimes called the Euler or Euler-Lagrange equation. Let us discuss very
briefly the Euler-Lagrange equations for the functional

1
Z
|∇u|2 + uf dn x,

E(u) = (10.4)
Rn 2

which is a modification of the Dirichlet functional in example 2). We consider this func-
tional on the Sobolev space H1 (Rn ). We compute


Z
dE(u)ξ = E(uλ )|λ=0 = (∇u · ∇ξ + f ξ) dn x,
∂λ Rn


where uλ ∈ H1 (Rn ) s.t. u0 = u and ∂λ uλ |λ=0 = ξ. Hence the Euler-Lagrange equation,
dE(u)ξ = 0, reads Z
(∇u · ∇ξ + f ξ) dn x = 0, ∀ξ ∈ H 1 (Rn ). (10.5)
Rn
Lectures on Applied PDEs, November 3, 2017 93

If u is twice differentiable, then we can integrate by parts to find


Z
dE(u)ξ = (−∆u + f )ξ.
Rn

Hence the Euler-Lagrange equation becomes Rn (−∆u + f )ξ = 0, ∀ξ ∈ H 1 (Rn ) and


R

therefore
∆u = f in Rn . (10.6)
This is the classical Poisson equation.
The above discussion motivates the following definition:

Definition 10.2. We say that u satisfying (10.5) is a (variational) weak solutions to the
equation (10.6) (see the remark around equation (4.6) in in Subsection D.2). Thus critical
points of E are (variational) weak solutions to the corresponding differential equations.

Discussion: Elliptic regularity. (to do)


If we consider the functional similar to the one above on a domain Ω ⊂ Rn , rather
than the entire Rn , i.e.
1
Z
|∇u|2 + uf dn x,

E(u) = (10.7)
Ω 2

then we have to be more careful. We first consider this functional on the Sobolev space
Hg1 (Ω) of functions in H 1 (Ω), which are equal to a given function g on ∂Ω. (Since the
boundary ∂Ω has n–dimensional Lebesgue-measure zero, we have to be careful about the
meaning of “u = g on ∂Ω”.) We compute


Z
dE(u)ξ = E(uλ )|λ=0 = (∇u · ∇ξ + f ξ) dn x,
∂λ Ω

where uλ ∈ Hg1 (Ω) s.t. u0 = u and ∂


u |
∂λ λ λ=0
= ξ. Integrating by parts, we find

∂u
Z Z
dE(u)ξ = (−∆u + f )ξ + ξ,
Ω ∂Ω ∂ν

where ν is the outward unit normal vector to ∂Ω.


Since our functions u are fixed (to g) on ∂Ω, we must take ξ vanishing on ∂Ω (so
that to have uλ = uR + λξ fixed at the boundary). Hence the Euler-Lagrange equation,
dE(u)ξ = 0, reads Ω (−∆u + f )ξ = 0, for every ξ ∈ H 1 (Ω) that vanishes on ∂Ω (i.e.
ξ ∈ H01 (Ω)), and therefore
∆u = f in Ω
(10.8)
u = g on ∂Ω.
This is the Dirichlet boundary value problem.
94 Lectures on Applied PDEs, November 3, 2017

If we consider the functional (10.4) on H 1 (Ω), then ξ is an arbitrary function in H 1 (Ω)


and varying its values in Ω and on ∂Ω independently, we see that the Euler-Lagrange
equation dE(u)ξ = 0, ∀ξ ∈ H1 (Ω), becomes

∆u = f in Ω
∂u
∂ν
= 0 on ∂Ω.

This is the Neumann boundary value problem. Solutions of these two problems are rather
different, so we see that the space on which a variational problem is considered plays an
important role.

Functionals from geometry. The first example is given by the length of a curve:
RT
4) L(γ) = 0 |γ 0 (t)|dt, where γ : [0, T ] → Rm ,

where L(γ) is the length of a curve γ given parametrically as γ : [0, T ] → Rm .

Exercise 10.5. Compute the Gâteaux derivative and find the Euler-Lagrange equation in
example 4).

As another example we consider the problem of minimal surfaces from differential


geometry. Let A(f ) be the area of the hypersurface S in Rn+1 given as a graph of a
function f : Ω → R, defined on an open subset Ω of the hyperplane x⊥ n+1 in R
n+1
:
S =graphf := {(x, f (x))|x ∈ Ω}. It is given by the formula
Z p
A(f ) = 1 + |∇f |2 dn x

and (see Appendix 10.11 to this section). We define A on C 2 (Ω). Critical points of the
functional A(f ) are called minimal surfaces. We have

Proposition 10.3. The equation for a minimal surface given locally as the graph of a
function f is given by
H(x0 ) = 0, (10.9)
where x0 = (x, f (x)) ∈ S =graphf and H(x0 ) is given by
!
∇f
H(x0 ) := div p . (10.10)
1 + |∇f |2

Exercise 10.6. Show that the Euler-Lagrange


R equation for A(f ) is is given by (10.9). (In
other words show that dA(f )ξ = − Ω H(x)ξ(x)dn x.)
Lectures on Applied PDEs, November 3, 2017 95

The expression on the l.h.s. of (10.9) is the mean curvature of S at x0 ∈ S.


Hence the equation for a minimal surface states that the mean curvature on it vanishes.
Sometimes this equation is taken for a definition of the minimal surface.
One can also define surface in Rn+1 locally via a parametrization S = Image(φ) ⊂
Rn+1 , where U is an open subset of Rn and ϕ : U → Rn (parametrization). This
is a parametric surface. Then the area A(φ) of S is given by the functional A(φ) =
R ∂φ ∂φ 2
U ∂x1
∧ ∂x2 d x.

Action and Lagrange functionals. We give now more examples of functionals:


RT
5) S(ϕ) = 0 [ 21 | ∂ϕ
∂t
|2 − V (ϕ)]dt, where ϕ : [0, T ] → Rm , and V : Rm → R;
RT R
6) S(φ) = 0 Ω [ 21 | ∂φ∂t
|2 − 21 |∇φ|2 − G(φ)]dnxdt, where φ : Ω × [0, T ] → R.

In these examples, the functionals come from expressions for an action.

Exercise 10.7. Compute the Gâteaux derivatives and find the Euler-Lagrange equation
in examples 5)–6).

The Euler-Lagrange equation in the 5) and 6) examples are Newton’s equation and
the classical relativistic field theory. In the second case,

∂ 2φ
− ∆φ − G0 (φ) = 0.
∂t2
∂φ
Let φ̇ := ∂t
. The functionals above are of the form
Z T 
S(φ) = L φ̇, φ dt, (10.11)
0

where L(φ̇, φ) := 21 |φ̇|2 − V (φ), a function of two variables φ̇ and φ, in the first example
and the functional
1 2 1
Z
|φ̇| − |∇φ|2 − G(φ) dnx,

L(φ̇, φ) :=
Ω 2 2
in the second one. φ : Ω×[0, T ] → R. L(φ̇, φ) is called the Lagrange function or functional.

Exercise 10.8. Show that the Euler-Lagrange equation for the functional (10.11) is given
by
 
−∂t (dφ̇ L φ̇, φ ) + dφ L φ̇, φ ) = 0. (10.12)

Exercise 10.9. Verify that (10.12) gives the same Euler-Lagrange equations for the ex-
amples 5) and 6) above, as the direct computation in Exercise 10.7.
96 Lectures on Applied PDEs, November 3, 2017

Functionals from image recognition. In conclusion, we give examples of functionals


from geometry and image recognition:

7) E(u) = Ω (|u − f |2 + λ|∇u|2 ) dn x, where u, f : Ω → R (f is fixed) and λ ≥ 0,


R

8) E(u, γ) = K\γ (|u − f |2 + λ|∇u|2 ) d2 x + µL(γ), where u, f : K → R (f is fixed),


R

λ, µ ≥ 0, and γ is a closed curve in K, i.e. γ : [0, 1] → K and γ(0) = γ(1).

Here E(u) is an error functional measuring how much a smooth function u differs from
a given function f , and E(u, γ) is the Mumford-Shah functional in image segmentation
(f : K → R is a given image, u is a piecewise smooth approximation of f , with jumps
across the curve γ are allowed, and γ is a segmentation of the image f , into two regions:
interior of γ and exterior of γ). Here L(γ) is the length of a curve γ given parametrically
as γ : [0, 1] → K ⊂ R2 .
The following elementary but important result connects the main problem of varia-
tional calculus to the problem of existence of solutions of differential equations

Theorem 10.4. Let F be a functional on an open subset M of a vector space X. If


u0 ∈ M is a minimizer of F , then u0 is a critical point of F .

Proof. Let ξ be an arbitrary vector from X, and λ sufficiently close to 0 so that there
is a curve uλ s.t. uλ=0 = u0 and du λ
|
dλ λ=0
= ξ. Then the function f (λ) := F (uλ ) has a
minimum at λ = 0, and therefore λ = 0 is a critical point of this function, f 0 (0) = 0.

This is equivalent to ∂λ F (uλ )|λ=0 = 0, which by the definition of the Gâteaux derivative
implies that dF (u0 )ξ = 0. This holds for every ξ ∈ X, and we conclude that dF (u0 ) = 0
as a functional on X.

A powerful method of proving existence of solutions of a large class of static equations


is showing that these equations are Euler - Lagrange equations of certain functionals and
then proving existence of minimizers (or saddle points) by the direct methods of variational
calculus. This will be done in Section 19 below.

10.4 Hessians and local minimizers.


Like in a finite dimensional case, to determine whether a critical point is a minimizer,
or maximizer, or a saddle point, we use the hessian operator. However, in the finite
dimensional context, this operator acts on an infinite dimensional space and handling it
is a more delicate mater.
We define the hessian of E at u as the linear operator defined by the quadratic form
given by the relation

hξ, Hess E(u)ηi = d2 E(u∗ )(ξ, η), (10.13)


Lectures on Applied PDEs, November 3, 2017 97

2 2
where d E(u)(ξ, η) is the hessian bilinear form defined as d E(u)(ξ, η) := d(dE(u)ξ)η =
∂t ∂s s=t=0 E(ust ), where ust := u + sξ + tη. Later on we give a more succinct definition
of the hessian in (14.8). Similarly to the finite dimensional case, we have the following
statement
Theorem 10.5. If u∗ is a minimizer of E, then Hess E(u∗ ) ≥ 0. If Hess E(u∗ ) ≥ θ, for
some θ > 0, then u∗ is a minimizer of E.

10.5 Convexity and uniqueness


Recall that a set A is said to be convex iff for every u, v ∈ A, we have λu + (1 − λ)v ∈ A
and a functional F : A → R is convex iff
F (λx + (1 − λ)y) ≤ λF (x) + (1 − λ)F (y)
and f is said to be strictly convex iff
F (λx + (1 − λ)y) < λF (x) + (1 − λ)F (y).
We have the following standard result
Proposition 10.6. Let A be a convex set. Then
• F : A → R be a convex C 1 –function iff F (x) − F (y) ≥ dF (y)(x − y)
and similarly, for the strictly convex functionals.
Proof. We give a proof in one direction. If F is convex, then so is f (s) := F (u + sv), ∀v ∈
X. Hence the function f 0 (s) := ∂s F (u + sv) is monotonically
R1 non-decreasing.
R 1 Let v :=
0 0 0
u − u. Then F (u ) − F (u) = F (u + (u − u)) − F (u) = 0 ds∂s F (u + sv) = 0 ds[∂s f (s) −
∂s f (0)] + ∂s f (0). By the monotonicity of ∂s f (s), we have that F (u0 ) ≥ F (u) + ∂s f (0).
Now, by the definition of dF (u), we have ∂s f (0) = dF (u)(u0 − u). The last two relations
give F (u0 ) ≥ F (u) + dF (u)(u0 − u).
Theorem 10.7. Let A be a convex set and F : A → R be A convex C 1 –function. We
have
(i) If a is a critical point of F , then a is a minimizer.
(ii) The set of all minimizers of F is a convex set.
(iii) If F is strictly convex, then it has at most one minimizer.
Proof. (i) and (iii) follow from Proposition 10.6 with y = a and the relation dF (a) = 0
and (ii), from the definition of the convexity.
Corollary 10.8. Weak solutions of the Laplace and Poisson equations with the Dirichlet
boundary conditions are unique.
For more details see [3, 9].
98 Lectures on Applied PDEs, November 3, 2017

10.6 Constraints and Lagrange multipliers


The problem we address now is concerns with minimizing or maximizing a functional,
E(u), under a constraint, J(u) = 0. Such problems appear naturally in applications:

Examples.
(1) Minimize the energy for a fixed entropy, or total mass, or total number of particles.
(2) Maximize output for a fixed cost.

Let U be an open set in a Banach space X and J : U → R, a functional on U . Wwe


consider minimization of a functional E(u), on a set, M , defined by side conditions or
constraints, i.e., M is of the form

M := {u ∈ X | J(u) = 0}, (10.14)

where J is a functional on X. Such a set can be thought of as an infinite dimensional


hypersurface in X. The restriction J(u) = 0 is called the constraint.
Problems of minimizing functionals on sets of the form (10.14) also come up in proving
existence of weak solutions of partial differential equations. This will be demonstrated
later.
Since M is not open in X, the definition of the Gâteaux derivative is more subtle, as
we cannot in general take pieces of straight lines uλ = u + λξ in the definition of dF (u).
For M is not open in X, to define the Gâteaux derivative, we take paths λ 7→ uλ , s.t.
u0 = u. Here λ ∈ [−, ] for some  = (u) sufficiently small. Then, for ξ ∈ X s.t. there is
an  > 0, and a differentiable path [−, ] 3 λ 7→ uλ ∈ M for which u0 = u, du λ
|
dλ λ=0
= ξ,
we define
d
dF (u)ξ := F (uλ )|λ=0 .

The set of all the ξ ∈ X s.t. there is an  > 0, and a differentiable path [−, ] 3 λ 7→
uλ ∈ M for which u0 = u, du λ
|
dλ λ=0
= ξ is called the tangent space to M at u, Tu M . Thus,
dF (u) : Tu M → R. By the definition, we have that dF (u) ∈ (Tu M )0 . Here X 0 is the space
of linear functionals on X, called the dual space to X. The dual space to Tu M is called
the cotangent space to M at u and is denoted by Tu∗ M := (Tu M )0
As before the points u ∈ M for which dF (u) = 0 are called the critical points of F .
Now we consider critical points of a functional E : X → R on the set (10.14) where
J is another functional on X. The key result here goes back to Lagrange and it is called
the method of Lagrange multipliers:

Theorem 10.9. Let E and J be C 1 functionals, and let M be defined by (10.14). Then
u0 is a critical point of E on M iff dE(u0 ) − λdJ(u0 ) = 0 on X for some λ ∈ R.

λ is called the Lagrange multiplier. It is determined by the condition that J(u0 ) = 0.


First we prove the following
Lectures on Applied PDEs, November 3, 2017 99

Proposition 10.10. Let J be a C 1 functional and let M is given by (10.14). Then


Tu M = Null J(u).

Proof. Let us be a differentiable path in M (i.e. a differentiable path in X satisfying


J(us ) = 0; existence of such paths is shown below in Proposition 10.11) starting at u0 , i.e.
us=0 = u0 and let du s
|
ds s=0
= ξ. Differentiating J(us ) = 0 in s at s = 0 gives dJ(u0 )ξ = 0,
which implies that ξ ∈ Null J(u0 ). We can also go in the opposite direction.
Proof of Theorem 10.9. The fact that u0 is a minimizer of E in M implies that s = 0 is a
s)
minimizer of E(us ) in s, which means that dE(u
ds
|s=0 = 0. The latter equation gives

dE(u0 )ξ = 0, ∀ξ ∈ Tu0 M = Null J(u0 ), (10.15)

(in the last step we used Proposition 10.10), i.e. dE(u0 ) is perpendicular to the subspace
dJ(u0 )⊥ := Null J(u0 ) and therefore dE(u0 ) is parallel to dJ(u0 ). In detail, let e ∈ X,
0 )η
e∈/ Null dJ(u0 ). We construct a projection P : X → Null dJ(u0 ) as P η := η − dJ(udJ(u0 )e
e∈
Null dJ(u0 ) for all η ∈ X. Let ξ := P η. Then

0 = dE(u0 )ξ = dE(u0 )η − λdJ(u0 )η (10.16)


dE(u0 )e
with λ = dJ(u 0 )e
, ∀η ∈ X. Hence dE(u0 ) − λdJ(u0 ) = 0 on X follows. (We showed that
if dE(u0 ) is perpendicular to dJ(u0 ), then dE(u0 ) is parallel to dJ(u0 )). Clearly, we can
reverse the steps to prove the opposite implication.
Example. Below Ω is a domain in Rn . Consider the Dirichlet functional
1
Z
E(u) = |∇u|2 dn x
2 Ω

on the M = {u ∈ Hg1 (Ω) : J(u) = 1}, where J(u) = p1 Ω |u|p dn x. To find the equation
R

of the critical points on the space M , we use the theorem. Since dE(u) = −∆u, and
dJ(u) = |u|p−2 u, the theorem implies that u satisfies

∆u + λ|u|p−2 u = 0 in Ω,
u = g on ∂Ω.

Compare this with the equation (19.5), with f = 0, for the problem on Hg1 (Ω).

Exercise 10.10. Find the equation for critical points of the following functionals:

(1) 21 Ω |∇u|2 on the space M = {u ∈ H01 (Ω) : Ω |u|2 = 1};


R R

(2) E(u) = Ω 21 |∇u|2 + G(u) dn x, on the space M = {u ∈ H 1 (Ω)| uf = a} for some


R  R

a ∈ R and f ∈ C(Ω);
100 Lectures on Applied PDEs, November 3, 2017

(3) the mean energy E(f ) = hf dn xdn p of a classical particle distribution f : Rn →


R

R +
R , wheren h(x, p) is a classical hamiltonian of the system and for entropy, S(f ) =
f log f d xdn p.
(4) E(u) = hu, Aui, where A is a self-adjoint operator on a complex Hilbert space H
and u ∈ M := {u ∈ D(A) | kuk = 1}.
(5) (Isoperimetric problem) the area of a closed n−dimensional surface (hypersurface)
in Rn for a fixed enclosed volume (see the next paragraph and Exercise 10.11).
There is a quantum analogue of the problem 3), with the classical particle distribution
f replaced by the density operator ρ, the mean energy given by E(ρ) = Tr(Hρ), where H
is a quantum hamiltonian, and the Boltzmann entropy S(f ) is given by the Boltzmann
entropy S(ρ) := − Tr(ρ log ρ).

Isoperimetric problem. The celebrated isoperimetric problem is differential geometry


requires to minimize the surface area of a closed surface for a given enclosed volume.
To formulate the problem formally, let S ⊂ Rn+1 be a closed hypersurface (i.e. an
n−dimensional surface). Denote its area and enclosed volume by A(S) and V (S), respec-
tively. The problem is to minimize A(S) for a fixed V (S) = c.
First one would like to find critical points for this problem, i.e. to find critical points
of A(S), given V (S) = c. This is what we do now.
We can find critical points for this problem as follows. Take a piece, S 0 , of S which
can be written as the graph, graphf , of a function f : Ω → R, defined on an open subset
Ω of the hyperplane Rn := {(x1 , . . . , xn , 0) : (x1 , . . . , xn ) ∈ Rn } in Rn+1 and satisfying
f ≥ 0 and f = 0 on ∂Ω. One can also find such a S 0 by translating and rotating S.
Then the area of S 0 is given by (see Appendix 10.11)
Z p
A(f ) = 1 + |∇f |2 dn x.

The enclosed volume between Ω and the graphf is equal to
Z
V (f ) = f dn x.

One canR show that the problem of finding critical points of A(f ) on the set M = {f ∈
H01 (Ω)| Ω f dn x = c} is locally equivalent to an original problem.
Exercise 10.11. Show that critical points of A(f ) on the set M = {f ∈ H01 (Ω)| Ω f dn x =
R

c} satisfy the Euler- Lagrange equation


∇f
div( p ) = h,
1 + |∇f |2
where h is the Lagrange multiplier.
A solution to this differential equation gives a surface of a constant mean curvature.
Lectures on Applied PDEs, November 3, 2017 101

10.7 Minimization problem and spectrum.


We would like to explain relation between the minimization (or in general critical point)
theory and the spectral theory of operators. Consider an operator A acting on a Banach
space X with a domain D(A) (i.e., A : D(A) → X). Recall that the spectrum, σ(A), of
an operator A is the set in C defined by

σ(A) := {z ∈ C : A − z1l is not invertible }. (10.17)

Clearly, eigenvalues of A belong to σ(A) (in fact, if λ is an eigenvalue, then Auλ = λuλ
for some nonzero uλ ∈ X (uλ is called an eigenvector), so (A − λ)uλ = 0, and A − λ is not
invertible). In general, the spectrum can also contain continuous pieces and it can take
very peculiar forms.
Now, let A be a self-adjoint operator on a complex Hilbert space H and define M :=
{u ∈ D(A) | kuk = 1}. Consider the functional E(u) = hu, Aui, where u ∈ M . In
Quantum Mechanics, if A is a quantum hamiltonian, then E is the expectation of the
energy in the state u.
There is a deep relation between existence of minimizers (or saddle points) for the
functional 12 hu, Aui on the set M (we can assume here that the operator A is bounded
below, i.e., hu, Aui ≥ −Ckuk2 for some C < ∞) and spectral properties of the operator
A. In particular, we have
Theorem 2. (i) inf u∈M hu, Aui = inf λ∈σ(A) λ;
(ii) hu, Aui has a minimizer on M if and only if inf λ∈σ(A) λ is an eigenvalue of A;
A proof of this theorem can be found in [6] where one can also find a theorem on a
relation between saddle points of the functional hu, Aui and eigenvalues of the operator
A greater than the smallest eigenvalue inf λ∈σ(A) λ. Here we mention only that the Euler-
Lagrange equation for the functional hu, Aui on the space M is

Au = λu

where λ is the Lagrange multiplier. This is the eigenvalue equation for A.


In the case when A is a Schrödinger operator, the number λ0 := inf λ∈σ(A) λ is called the
ground state energy of A and the theorem above expresses the variational characterization
of eigenvalues used frequently in Quantum Mechanics.
The above motivates the following definitions. We define the discrete spectrum of an
operator A as

σd = {λ ∈ C | λ is an isolated eigenvalue of A with finite multiplicity}.

The part of the spectrum which complements the discrete spectrum is called the essential
spectrum of an operator A:
σess (A) := σ(A) \ σd (A).
102 Lectures on Applied PDEs, November 3, 2017

(Some authors use the term ”continuous spectrum” rather than ”essential spectrum”.)
Hence we have σ(A) = σd (A) ∪ σess (A).
Examples of operators and their spectra are given in Appendices A.7 and E.
If A is a quantum hamiltonian (the Schrödinger operator), then the eigenfunctions
corresponding to discrete eigenvalues are called the bound states. They describe the
motion of quantum system localized essentially to a bounded domain of the physical
space. The essential spectrum is related to the scattering states, describing the system
broken into freely moving fragments.
A more difficult and very useful result is the following
Theorem 3. If inf u∈M hu, Aui < inf λ∈σess (A) λ, then inf u∈M hu, Aui is an eigenvalue of A.
(spectrum of ∆ on bounded domains)

10.8 Tangent spaces


Recall that we define the tangent space to M at u, Tu M , to be the set of all the ξ ∈ X
s.t. there is an  > 0, and a differentiable path [−, ] 3 λ 7→ uλ ∈ M for which
u0 = u, du λ
|
dλ λ=0
= ξ.
Exercise 10.12. Show that Tu X = X and therefore in particular, Tu H s (Ω) = H s (Ω).
Now, we describe tangent spaces to surfaces (10.14).
Proposition 10.11. If M = {u ∈ X : J(u) = 0}, where J is a C 1 functional on X, then
for u ∈ M , Tu M = Null dJ(u).
Proof. If u ∈ M and ξ ∈ Tu M , then, by the definition, there exists a path uλ ∈
C 1 ((−, ), M ) with ξ = ∂λ |λ=0 uλ and uλ=0 = u. Therefore,

0 = ∂λ J(uλ )|λ=0 = hdJ(u), ξi.

Hence ξ ∈ Null dJ(u).


Now, let u ∈ M and ξ ∈ Null dJ(u) and find uλ such that uλ=0 = u, ξ = ∂λ |λ=0 uλ and
J(uλ ) = 0. Let η ∈/ Null dJ(u) and define the path uλ = u + λξ + λaη, where a solves the
equation f (a, λ) = 0, where
1
f (a, λ) := J(u + λξ + λaη).
λ
We write J(u + v) = J(u) + dJ(u)v + Ru (v), where Ru (v) is defined by this equation, i.e.
Ru (v) := J(u + v) − J(u) − dJ(u)v. Now, we take v = λξ + λaη in this equation and
use J(u) = 0 and dJ(u)ξ = 0, to obtain f (a, λ) = adJ(u)η + λ1 Ru (λξ + λaη). Now, we
estimate
Ru (λξ + λaη) = J(u + λξ + λaη) − J(u) − dJ(u)(λξ + λaη)
Lectures on Applied PDEs, November 3, 2017 103

Z 1
= ds(dJ(u + s(λξ + λaη)) − dJ(u))(λξ + λaη) = o(λ).
0

Hence f (0, 0) = 0. Now, compute ∂a f (0, 0). We have

∂a Ru (λξ + λaη) = dJ(u + λξ + λaη)λη − dJ(u)λη = o(λ).

Hence ∂a f (a, λ) = dJ(u)η+o(λ) and, in particular, ∂a f (0, 0) = dJ(u)η 6= 0. Therefore, by


the implicit function theorem, the equation f (a, λ) = 0 has a unique solution for a = a(λ)
for λ sufficiently small and this solution satisfies a(λ) = oλ (1).
By the definition of f (a, λ) the family uλ = u + λξ + λaη, with a = a(λ) = oλ (1)
solving f (a, λ) = 0, has the properties mentioned above. Hence ξ ∈ Tu M .

10.9 Dual space


The set of all bounded linear functionals on X is called the dual space of X (or simply
the dual (or adjoint space) of X, and it is denoted as X 0 . Thus dE(u) ∈ X 0 . It is a vector
space with the norm (10.2). In fact, since C is complete, one can show that X 0 is always
a Banach space, whether X is complete or not.
If X is a space of functions, then X 0 can be identified with either a space of functions
or a space of distributions or a space of measures. Here are some examples of dual spaces:

1) (Lp )0 = Lq , where 1/p + 1/q = 1, if 1 ≤ p < ∞ (space of functions),

2) (L∞ )0 is a space of measures which is much larger than L1 ,

3) (Hs )0 = H−s (space of distributions if s > 0).

10.10 Functionals on complex spaces


We consider functionals on open subsets, M , of complex vector spaces, Z. For such
functionals, we define the complex Gâteaux derivatives as
¯
dE(ψ) ≡ dψ E(ψ) = (dψ1 − idψ2 )E(ψ) and dE(ψ) ≡ dψ̄ E(ψ) = (dψ1 + idψ2 )E(ψ), (10.18)

where ψ = ψ1 + iψ2 , and similarly the complex gradients ∂ψ E(ψ) and ∂¯ψ E(ψ). One way to
compute dψ E(ψ) and dψ̄ E(ψ) is to treat ψ and ψ̄ as independent functions and compute
the corresponding objects as partial Gâteaux derivatives.
To connect this to the real Banach theory considered above, we note that Z can be
written as Z = V + iV , for some real vector space. For example Z = H 1 (Rd , C) =
H 1 (Rd , R) + iH 1 (Rd , R). We associate with such a Z, a real space Ẑ := V ⊕ V . There is
one - to - one correspondence between Z and Ẑ:
~ := (φ1 , φ2 ), φ1 := Re φ, φ2 := Imφ.
φ ⇐⇒ φ
104 Lectures on Applied PDEs, November 3, 2017

Consider the map compl : Ẑ → Z, given by φ ~ := (φ1 , φ2 ) =⇒ φ = φ1 +iφ2 , and its inverse vect :
φ = φ1 + iφ2 =⇒ φ ~ := (φ1 , φ2 ). Using these maps identify a functional E(φ) on Z with
a functional E real (vect(φ)) = E(φ) on Ẑ and we can define the variational (or Gâteaux
or Fréchet) differentiability and derivative, ∂φ~ E(φ), and partial derivatives, ∂φ1 E(φ) and
∂φ2 E(φ), with respect the real, φ1 , and imaginary, φ2 , parts of the field φ for E(φ), by
~ After that we introduce the derivatives with respect φ and
computing them for E real (φ).
φ̄ as follows

∂φ E(φ) := ∂φ1 E(φ) − i∂φ2 E(φ), ∂φ̄ E(φ) := ∂φ1 E(φ) + i∂φ2 E(φ). (10.19)

Then, defining ∂φ~ E(φ) = ∂φ~ E real (vect(φ)), we have the following relations

∂φ̄ E(φ) = compl(∂φ~ E(φ)), ∂φ~ E(φ) = vect(∂φ̄ E(φ)). (10.20)

Here are some examples (below Ω ⊂ Rd , d = 2, 3):


6) The Ginzburg-Landau energy functional (ψ : Ω → C, a : Ω → Rd )
1 κ
Z n o
2 2 2 2
EΩ (ψ, a) := |∇a ψ| + (|ψ| − 1) + | curl a| , (10.21)
2 Ω 2
where ∇a = ∇ − ia, the covariant derivative.
7) The action functional for nonlinear Schrödinger equation (ψ : Ω × [0, T ] → C)

1 T
Z Z
S(ψ) = ˙ + |∇ψ|2 + G(|ψ|2 ))dd xdt.
(−Im(ψ ψ) (10.22)
2 0 Ω
Exercise 10.13. Compute the complex and real Gâteaux derivatives and the equations
for the critical pints in examples 6) – 7) above.
For complex vector spaces, we say that u0 ∈ M is a critical point (CP) of E if and
¯ 0 ) = 0.
only if dE(u

10.11 Appendix I: Area of a hypersurface


In this appendix we derive a standard formula for the area of a surface in Euclidean
spaces. Let S be a smooth n–dimensional surface in Rn+1 . Such a surface is called a
(smooth) hypersurface. Hypersurfaces appear as graphs or level sets of images of some
functions.
Let U ⊂ Rn and f : U → R, be smooth. Then graph f := {(u, f (u)) | u ∈ U } is a
hypersurface in Rn+1 .
Let ϕ : Rn+1 → R be a smooth function and let 0 ∈ Ran ϕ. Then the zero level set of
ϕ,

ϕ−1 (0) := {x ∈ Rn+1 | ϕ(x) = 0}


Lectures on Applied PDEs, November 3, 2017 105

is a smooth hypersurface.
Finally a map ψ : U ⊂ Rn → S is called a local parametrization of S, provided (fill
in).
Lemma 10.12. Let U ⊂ Rn be an open set and f : U → R. Then the area of the surface
S :=graphf is given by the formula, A(S) = A(f ), where
Z p
A(f ) := 1 + |∇f |2 dn u. (10.23)
U

Proof. Let x = (u, xn+1 ), where u = (x1 , . . . , xn ). The picture below shows the following
dn u
formula for the area element of S: ∆A = cos α
, where α is the angle between the xn+1 -axis
(the unit vector en+1 ) and the normal ν = ν(x) to S at a given point x ∈ S.
en+1

ν
J

J b ∆A
S J b
## 
bb

dn x0 = dx1 . . . dxn

Let ϕ(x) = xn+1 − f (u) so that S = {x ∈ U × R | ϕ(x) = 0}. Then on S:


∇ϕ(x)
ν(x) = ,
|∇ϕ(x)|
and therefore for x ∈ S
∇ϕ(x) · en+1 1
cos α = =p .
|∇ϕ(x)| 1 + |∇f |2
R
Since area(S) = S
dA, the result follows.

11 Bifurcation of surfaces of constant mean curva-


ture
11.1 surfaces of constant mean curvature
Let S be a (n + 1)−dimensional surface in Rn+2 (a hypersurface). Denote by H(x) its
mean curvature at a point x ∈ S. If S is described locally by a map (an immersion)
106 Lectures on Applied PDEs, November 3, 2017

4 ANTONIO ROS

Figure 3. (a) Doubly periodic surface with constant mean


curvature and genus 2 (modulo translations). These surfaces
were first constructed by Lawson [11] as conjugate surfaces
of minimal surfaces in the 3-sphere. (b) Periodic pattern
formed by a thin layer of liquid on a planar surface, [12]:
this self-assembled wetting phenomena may be modeled by
stable doubly periodic constant mean curvature surfaces.

convex curve in the (volume, area)-plane that projects injectively on, both
the volume and the area axes.
θ:U →R (where U , an open set in Rn+1 ), then we write H(x) = H(θ). We consider
n+2
Among self-assembled materials, mesoscopic wetting phenomena are par-
ticularly related to doubly periodic, stable, constant mean curvature sur-
the equation
faces, see [12]: under suitable conditions, a thin layer of liquid wetting an
hydrophobic planar surface produces a pattern as in Figure 3b, exhibit-
ing a periodic array of dry spots. As doubly periodic constant mean cur-
vature surfaces always have a horizontal mirror symmetry (assuming the
prescribed translations are horizontal), it follows that the corresponding
H(θ) = h (11.1)
periodic isoperimetric problems in R3 and in the halfspace {x3 ≥ 0} are
equivalent. As another consequence of Theorem 2 below, we conclude that
for some constant h. Surfaces satisfying this equation are called the constant mean cur-
the assumptions
vature surface.
(i) the pattern One of the reasons they are interesting is that they solve the isoperimetric
is doubly periodic,
(ii) the volume fraction of liquid is given, and
problem:
(iii) the energy of the system is just the area (per unit cell) of the liquid
surface,

• minimize surface area for a fixed enclosed volume.


predict the right experimental wetting phases: after reflection in the hori-
zontal plane we obtain either round spheres, or horizontal right cylinders,
or constant mean curvature doubly periodic surfaces with genus 2 (modulo
Critical points for this variational problem are surfaces of constant mean curvature.
translations), like in Figure 3.

Obvious solutions of the equation (11.1) are spheres of radius n+1 h


and cylinders of
n
radius h .
There is a vast number of constant mean curvature surfaces, e.g. rotationally sym-
metric surfaces in R3 , known as Delaunay surfaces, and various surfaces obtained by
perturbing several known surfaces, glued together ([?, ?]), see Fig. above.
However, it is probably only spheres which solve the isoperimetric minimization prob-
lem.
We consider solutions to (11.1) periodic along the xn+2 −axis of period a. In this case,
• Every round cylinder along the xn+2 −axis and every (half-)sphere are static solutions
of (11.1).
To see whether spheres or cylinders are stable, one can compare the surface area of the
cylinder and half-sphere of the same volume, namely which one is smaller. For a = 1, the
Lectures on Applied PDEs, November 3, 2017 107


surface area of a cylinder with volume V between the two planes is Acyl = 2 πV and the
surface area of a sphere with volume V is Asph = (36πV 2 )1/3 . Acyl ≤ Asph is equivalent to
81
V ≤ 4π . In particular, one expects that cylinders of radius R are stable when R is large
and as R decreases cylinders become unstable.
We are interested in rotationally symmetric, periodic static solutions. Our goal is
to show their existence and their (linear) stability and how the latter depends on the
parameter a.
The notion of linear stability is defined in at the beginning of Subsection 7.4. In
the present context, it should be applied to the volume preserving mean curvature flow.
Therefore we should consider the variations
R (perturbations) which preserve the enclosed
volume. Such variations, ξ, satisfy S ξ · ν = 0. Hence, we say that a CMC surface θ is
stable (in the sense of geometric analysis) iff dH(θ) ≥ 0 on the subspace
Z
Vθ = {ξ : S → R n+1
: ξ · ν = 0} (11.2)
S

and unstable, otherwise.

11.2 Bifurcation of new surfaces


We consider solutions of Eq. (11.1) which are surfaces of revolution around the xn+2 −axis,
periodic in the xn+2 −direction of period a. Such surfaces are obtained by rotating graphs
of functions ρ(xn+2 ) around the xn+2 −axis. (They are special cases of graphs, θρ : (ω, x) 7→
(ρ(ω, x)ω, x), over the unit round cylinder Can+1 = Sn ×[0, a] = {(ω, x) : ω ∈ Sn , x ∈ [0, a]},
with the graph functions ρ.) Let x = xn+2 . Eq. (11.1) has the cylindrical branch
(a, ρcyl ≡ nh ), ∀a > 0, of solutions (a cylinder is periodic along its axis with an arbitrary
period). This leads us to the following

Theorem 11.1. • At a = 2πk nh , a new branch of even/odd solutions of Eq. (11.1)


bifurcates from the cylindrical branch (a, ρcyl ).

• The bifurcating even solutions are are periodic surfaces of revolution of period a
given by the functions ρs (x) = nh + s cos(πx) + O(s2 ) (and similarly for the odd
solutions).

• For 0 < a < 2π nh , the cylindrical static solution, ρcyl , of Eq. (11.1) is unique in a
small even/odd neighbourhood of ρcyl and is stable. At a = 2π nh , it looses its stability
and is unstable for a > 2π nh .

• The latter branch is stable for 2π nh < a < 4π nh , at a = 4π nh , it looses its stability
and a new stable branch of solutions bifurcates at this point and so forth.
108 Lectures on Applied PDEs, November 3, 2017

In the next subsection we prove existence of bifurcating solutions. After that we prove
their stability. Before proceeding to the proofs, we give the expression of the mean
curvature H, of surfaces of revolution.
First, we rewrite out the expression for the mean curvature for the level set represen-
tation of S. Below, all differential operations, e.g. ∇, ∆, are defined in the corresponding
Euclidian space (Rn+2 ). First, we recall that, if a surface S given in the level set repre-
sentation, S = {x0 ∈ Rn+2 : ϕ(x0 ) = 0}, then we have (see Proposition 6.11)

∇ϕ ∇ϕ
 
0 0 0
ν(x ) = , H(x ) = div ν(x ) = div . (11.3)
|∇ϕ| |∇ϕ|

Using this expression, we can obtain an expression for the mean curvature of surfaces of
revolution. Let x0 = (x⊥ , x) ∈ Rn+2 , with x⊥ = (x1 , . . . , xn+1 ) ∈ Rn+1 . Then a surface
of revolution can be written as S = {x0 ∈ Rn+2 : ϕ(x0 ) := |x⊥ | − ρ(x) = 0}, where
2 12
r⊥ ≡ |x⊥ | = ( n+1 ⊥
:= x⊥ /|x⊥ | to
P
i=1 xi ) . We use this and (11.3) and the notation x̂

compute ν(ρ) = √(x̂ ,∂x ρ) 2 , which gives the following expression for the mean curvature
1+(∂x ρ)
0
H(x ), of S:
−1 ∂x2 ρ n
H(ρ) ≡ H(x0 ) = p 2
− . (11.4)
1 + (∂x ρ)2 1 + (∂x ρ) ρ
We define the map F (ρ, h, a) := H(θρ ) − h. Then the equation (11.1) can be rewritten
as

F (ρ, h, a) = 0. (11.5)

Note that the round cylinder of the radius R is given by the equation ρ = R.

11.3 Proof of the existence part


We consider a solution to (11.1) which are surfaces of revolution with the graph function
ρ = ρ(x). Then the mean curvature is given by the expression (11.4) and the equation
(11.1) can rewritten as F (ρ, h, a) = 0, with

−1 ∂x2 ρ n
F (ρ, h, a) := − − h, (11.6)
1 + (∂x ρ)2 1 + (∂x ρ)2
p
ρ

where ρ is periodic of a period a. Note the following properties of H(ρ):

• H(ρ = R) = n/R;

• H(ρλ )(x) = λ−1 H(ρ)(x/λ), where ρλ (x) := λρ(x/λ);

• dρ H(ρ)(ρ − x · ∇ρ) = −H(ρ) − x · ∇H(ρ).


Lectures on Applied PDEs, November 3, 2017 109

Proof. The first and second properties are obvious from the expression (11.6). They can
be also deduced from the general facts that H(θcyl ) = Rn and (see (17.9) and e.g. [16])
H(λθ) = λ−1 H(θ). (11.7)
Because of θ(ω, x) = (ρ(ω, x)ω, x), the rescaling θ → λθ induces the rescaling ρ → λρ(x/λ)
of ρ.
Differentiating the equation in the second property w.r.to λ at λ = 1 and using that
∂ρλ
∂λ λ=1
= ρ − x · ∇ρ, we arrive at the equation in the third property.
These properties imply the following properties of F :
n
• cylinder of radius R = R(h) := h
(ρcyl ) solves F (ρcyl , h, a) = 0, ∀a, h.
• F (ρλ , λ−1 h, λa)(x) = λ−1 F (ρ, h, a)(λ−1 x), where ρλ (x) := λρ(x/λ).
• dρ F (ρ∗ , h, a)(ρ∗ − x · ∇ρ∗ ) = −h, for any (ρ∗ , h, a) solving F (ρ, h, a) = 0.
The third property with ρ∗ = ρcyl (so that ρ∗ = R and ∇ρ∗ = 0) implies that dρ F (ρcyl , h, a)
2
has the eigenvalue − Rh = − hn with the eigenfunction equal identically to 1.
Using the second property we rescale the equation F (ρ, h, a) = 0, to eliminate one of
the parameters, e.g. by taking λ = h, or λ = a−1 , in the equation F (ρλ , λ−1 h, λa)(x) =
λ−1 F (ρ, h, a)(λ−1 x). (This might depend on which parameter physically we want to keep
fixed and which to vary.) This gives the new equation F (ρ0 , a0 )(x) = 0, 0 ≤ x ≤ ha, where
F (ρ0 , a0 )(x) := h−1 F (ρ0 , 1, a0 )(x), ρ0 (x) := hρ(x/h) and a0 := ha, or
F (ρ0 , h0 )(x) = 0, ρ0 (x + 1) = ρ0 (x)
where F (ρ0 , h0 )(x) := (h0 )−1 F (ρ0 , h0 , 1)(x), ρ0 (x) = a−1 ρ(ax) and h0 := ah. We choose the
second rescaling and drop the primes. Consequently, we have the equation
F (ρ, h) = 0, where F (ρ, h) := H(ρ) − h, (11.8)
1 ∂x2 ρ n
H(ρ) := p − 2
+ , (11.9)
1 + (∂x ρ)2 1 + (∂x ρ) ρ

on R with ρ periodic of the period 1:


ρ(x + 1) = ρ(x).
The latter equation has the family of cylindrical (homogeneous) solutions:
n
ρh = ,
h
which in the old variables become ρcyl = nah
.
To prove the bifurcation of the new solutions, we use the bifurcation theory, specifically,
the Krasnoselski theorem. We check the conditions of this theorem (below R(h) := nh ):
110 Lectures on Applied PDEs, November 3, 2017

(a) F is C 1 ;

(b) F (ρh , h) = 0, ∀h, where ρh = R(h);



(c) dρ F (ρh , h) has the eigenvalue 0 for h = π nk and this eigenvalue is of the multi-
plicity 2

(d) ∂h (dρ F (ρ, h) ρ=R(h) ) = −h−2 R(h)
n
= −h−1 6= 0.

Properties (a), (b) and (d) are straightforward. So, we check (c). We linearize F (ρ, h)
at ρh ≡ R ≡ R(h) := nh and let Lh := −dρ F (ρh , h). Then by direct computation
n
Lh = −∂x2 − ,
R2
acting on the space L2per (R) of periodic, locally L2 −functions of the period 1.

Proposition 11.2. The operator Lh is self-adjoint and its spectrum, σ(Lh ), is purely
discrete with the eigenvalues − Rn2 , of the multiplicity 1, and (2πk)2 − Rn2 , k = 1, 2, · · · ,
of the multiplicity 2, with the eigenfunctions 1, cos(2πkx) and sin(2πkx).

The self-adjointness and the form of the spectrum are obvious. We mention that the
fact that Lh has the eigenvalue − Rn2 is not accidental. As shown above, it is related to
the fact that ρh ≡ R breaks the scaling covariance of the equation.
Recall R ≡ R(h) := nh . If h satisfies (2πk)2 − Rn2 = 0, k √
= 1, 2, · · · , where R = nh , then
the operator Lh has a zero eigenvalue. These values, h = 2π nk, are the candidates for the
bifurcation points. This proves the property (c). To get eigenvalues of odd multiplicities
as required by the conditions of the Krasnoselski theorem, we restrict to either odd or
even functions, ρ(x) to get the multiplicity one.
Hence the Krasnoselski theorem is applicable and implies that the equation (11.8) has
non-trivial solution branches
√ of even/odd solutions bifurcating at from the trivial branch
(R(h), h), ∀h at h = 2π nk, k = 1, 2, . . . .
By Corollary 7.6 (of the proof of the Krasnoselski theorem), the first non-trivial even
solution branch is of the form
n
ρ= + s cos(πx) + ws , with ws = O(s2 ), (11.10)
h
and h depending on s. 

11.4 Proof of the linearized stability


In this section we show that (a) for 0 < a < 2π nh , the cylindrical static solution, ρh = nh ,
of Eq. (11.1) is stable, (b) at a = 2π nh , it looses its stability and (c) for a > 2π nh , but
sufficiently close to 2π nh , the new solution that bifurcated at a = 2π nh is stable.
Lectures on Applied PDEs, November 3, 2017 111

Again we pass to the rescaled problem (11.8). The notion of linear stability is defined
in at the beginning of Subsubsection 11.1. It translates in the present context as follows.
Let ρ∗ be a solution of F (ρ, h) = 0 and dρ F (ρ∗ , h) denote the Gâteaux derivative of
the map F (ρ, h) at ρ = ρ∗ . Recall that ρ describe surfaces of a fixed enclosed volume,
say V (ρ) = c > 0. Hence we have to define dρ F (ρ∗ , h) on the tangent space Tρ∗ {ρ ∈
2
Hper (R, R) : V (ρ) = c}, where Hper
2
(R, R) is the Sobolev space of order 2 of real periodic
functions on R of the period 1. R1
For surfaces of revolution we have by integrating over radial slices, V (ρ) = 0 2πρ2 (x)dx.
Since by Proposition 10.10, Tρ∗ {V (ρ) = c} = Null dV (ρ∗ ), we have that
Z 1
V∗ = {ξ ∈ Hper (R, R) :
2
ρ∗ ξdx = 0}. (11.11)
0

This subspace descends from (11.2).


We say that a solution ρ∗ of (11.8), i.e. a solution of F (ρ, h, a) = 0, is linearly stable
iff the spectrum of dρ F (ρ∗ , h) on V∗ lies in the half-plane { Re z > 0 } and we say that ρ∗
is linearly unstable iff σ(dρ F (ρ∗ , h)) V∗ ∩ { Re z < 0 } =
6 ∅. (Since in our case, dρ F (ρ∗ , h)

is self-adjoint, the above relations can be replaced by du F (ρ∗ , h)) V∗ ≥ 0 and the smallest

eigenvalue of du F (ρ∗ , h) V∗ is ≤ 0, respectively.)
Now, we apply this to ρ∗ = ρh . In this case space (11.11) becomes
Z 1
Vh = {ξ ∈ Hper (R, R) :
2
ξdx = 0}. (11.12)
0

The operator du F (ρh , h)) V =: Lh is self-adjoint. Hence its spectrum is real and therefore
h
the static solution ρh is linearly stable iff σ(Lh V ) ⊂ {λ ∈ R : λ > 0}, and unstable iff

h
σ(Lh V ) ∩ {λ ∈ R : λ < 0} =

6 ∅.
h
Proposition 11.2 implies that on the subspace Vh , the eigenvalues of Lh on even/odd
functions are (2πk)2 − Rn2 , k = 1, 2, · · · , of the multiplicity 1, with the eigenfunctions
cos(2πkx)/√ sin(2πkx). √
If 2π n > h, then σ(Lh ) on Vh is positive and for h = 2π n, the lowest eigenvalue
vanishes. This suggests that a new√stable stationary solution bifurcates at this point
from the old one. Moreover, for 2π n < h, the operator Lh has a negative eigenvalue
h2
2π 2 − R(h)
n 2
2 = 2π − n ) and therefore, while the old solution, the cylinder ρh , is stable for
√ √
2π n > h, it is unstable for 2π n < h. We expect that√the new bifurcating √ solution is
stable, at least for h not too large, more precisely, for 2π √n < h < 2π 32 n. √
To sum up, the cylinder ρh = R(h) is stable for h < π n, and unstable for h > π n.


Non-axi-symmetric surfaces. Let n = 1 (two dimensional surfaces in 3 dimensional


ambient space). We consider solutions to (11.1) which are graphs over the standard
112 Lectures on Applied PDEs, November 3, 2017

cylinder given by functions ρ depending also on the angle θ: ρ = ρ(x, θ). Then the mean
curvature is given by1

−∂x2 ρ − ρ−2 ∂θ2 ρ


 
1 1
H(ρ)(x, θ) = p + . (11.13)
1 + (∂x ρ)2 + (∂θ ρ/ρ)2 1 + (∂x ρ)2 + (∂θ ρ/ρ)2 ρ

We assume as before that ρ is periodic in x of a period a. Then equation (11.1) can


rewritten as

F (ρ, h, a) := H(ρ) − h = 0, ρ is periodic in x of a period a, (11.14)

where H(ρ) is the mean curvature given by the expression (11.13).

Exercise 11.1. (a) Show that the cylinder are still solutions to (11.14) for every a;
(b) Find the bifurcation points from the cylindrical branch;
(c) Find the values of the parameter a for which the cylindrical solution is linearly
stable/unstable.
(Hint: As in the axi-symmetric case, consider ρ’s even and odd in x separately. In
addition, one might want to separate the cases of ρ’s satisfying ρ(x, −θ) = ρ(x, θ) and
ρ(x, −θ) = −ρ(x, θ).)

Remark. Consider Eq. (11.8) on periodic ρ’s of the period 1. If ρ∗ is a non-constant


solution to (11.8), then dρ F (ρ∗ , h)∂x ρ∗ = 0, i.e. dρ F (ρ∗ , h) has the eigenvalue 0, with the
eigenfunction ∂x ρ∗ .
This property is related to the fact that ρ∗ breaks the translational invariance of the
equation: if ρy (x) := ρ∗ (x + y), then F (ρy , h)(x) = F (ρ, h)(x + y). Therefore, if ρ∗ solves
F (ρ, h) = 0, then ρλ satisfies F (ρy , h)(x) = 0. Differentiating the latter equation w.r.to y
∂ρy ∂ρy

at y = 0, we find dρ F (ρ∗ , h) ∂y y=0 = 0. Since ∂y y=0 = ∂x ρ∗ , the result follows.
Due to the third property of F (dρ F (ρcyl , h, a)(ρ∗ − x · ∇ρ∗ ) = −h, for any (ρ∗ , h, a)
solving F (ρ, h, a) = 0), to study the stability of ρ∗ , we have to consider dρ F (ρ∗ , h) on the
subspace
Z 1 Z 1
V∗∗ = {ξ ∈ 2
HNeum ([0, 1], R) : ξ(1 − x∂x ρ∗ ) = 0, ξ∂x ρ∗ = 0}. (11.15)
0 0
1
The unit normal vector is given by

(x̂⊥ − x̂⊥ ∂θ ρ/|x⊥ |, ∂x ρ)


ν(ρ)(x, θ) = p ,
1 + (∂x ρ)2 + (∂θ ρ/|x⊥ |)2

where x⊥ := (x1 , x2 ), x̂⊥ := x⊥ /|x⊥ | and x⊥ := (x2 , −x1 ), x̂⊥ := x⊥ /|x⊥ |, which gives expression (11.13)
for the mean curvature.
Lectures on Applied PDEs, November 3, 2017 113

11.5 Appendix: Volume preserving mean curvature flow


The volume preserving mean curvature flow (VPF) is a family {St ; t ≥ 0} of smooth
closed hypersurfaces in Rn+1 , given say by immersions ψ, satisfying the following evolution
equation:
∂t ψ N = H̄ − H, (11.16)
where ∂t ψ N denotes the normal velocity of St at time t, ∂t ψ N = h∂t ψ, νi, where ν is
the unit normal vector field on St , H = H(t) stands for the mean curvature of St and
H̄ = H̄(t) is the average of the mean curvature on St , i.e.,
R
Hdσ
H̄ := RSt , t ≥ 0. (11.17)
St

As an initial condition, we consider a simple hypersurface S0 in Rn+1 , with no boundary


in Ω (e.g. either entirely Ω or ∂S0 ⊂ ∂Ω) given by an immersion x0 .
Like the MCF, the VPF is invariant under rigid motions (translations and rotations)
and appropriate scaling. Moreover, it shrinks the area, A(ψ) = V (ψ), of the surfaces,
but, unlike the MCF, the VPF preserves the enclosed volume, Vencl (ψ). As the result,
it has stationary solutions - the Euclidean spheres (for closed surfaces) and cylinders for
surfaces with flat boundaries. We summarize these properties as follows
• (11.16) is invariant under rigid motions of the surface, i.e. ψ 7→ Rψ + a, where
R ∈ O(n + 1), a ∈ Rn+1 and ψ = ψ(u, t) is a parametrization of St , is a symmetry
of (11.16).
• (11.16) is invariant under the scaling x 7→ λx and t 7→ λ−2 t for any λ > 0.
• (11.16) is volume preserving.
d
(H̄ − H)2 dσ ≤ 0.
R
• (11.16) is area shrinking. Actually, dt
V (ψ(t)) = − St

• Static solutions of (11.16) are surfaces of constant mean curvature (CMC), H = h.


The first two statements are proven Ras for the MCF. The third and fourth ones follow
d
R
from dt Vencl (ψ(t)) = St h∂t ψ, νidσ = St (H̄ − H)dσ = 0 and (cf. Proposition 17.3 and its
proof)
d
Z Z Z
V (ψ(t)) = Hh∂t ψ, νidσ = (H̄H − H )dσ = − (H̄ − H)2 dσ ≤ 0.
2
dt St St St

The fifth one is obvious.


The third and forth properties suggests that, as t → ∞, St converges to a closed
surfaces with the smallest surface area for a given enclosed volume. The limiting surface
must be a static solution to the VPF and therefore, by the fifth property, it is a surface
of constant mean curvature (CMC). This leads to the isoperimetric problem:
114 Lectures on Applied PDEs, November 3, 2017

• minimize the area V (ψ) given the enclosed volume Vencl (ψ).

Namely, we have
Proposition 11.3. (i) Minimizers of the area V (ψ) for a given enclosed volume Vencl (ψ)
are critical points of the area functional V (ψ) on space of immersions with the given
enclosed volume Vencl (ψ).
(ii) The (Euler-Lagrange) equation for these critical points is exactly the CMC equation
H = h.
(iii) These critical points are critical points of the functional
Vh (ψ) := V (ψ) − hVencl (ψ),
where h is determined by c = Vencl (ψ) and vice versa.
Proof. The first and third statements are standard results (the third statement is follows
from the Lagrange multiplier theorem). The second statement follows from the relations
R √ n R
dV (ψ)ξ = U Hν · ξ gd u and dVencl (ψ)ξ = St hξ, νidσ and the definition of Vh (ψ).
Thus finding stationary solutions to the VPF is the same as finding closed CMC
surfaces. This leads to the following problems: (a) Find CMC surfaces and (b) determine
their stability w.r. to the VPF.
However, we expect that the VPF converges to a minimal CMC surface (i.e. one
solving the isoperimetric problem), not just to a CMC surface. For such we have an
additional characterization, which is a standard fact from the variational calculus (see
e.g. [?]):
Proposition 11.4. If ϕ minimizes the area V (ψ) for a given enclosed volume Vencl (ψ),
then HessN V (ϕ) ≥ 0 on Tϕ Xc .
The property of minimal CMC surfaces isolated in this proposition plays an important
role in their analysis and deserves the name. We say that a CMC surface θ is stable iff
HessN V (θ) ≥ 0 on Tθ Xc . (In dynamical systems and partial differential equations, this
notion is called (weak) linear or energetic stability.)

12 Turing instability
Let f be a continuously differentiable function f : Rn → Rn , n ≥ 2, and let Ω ⊆ Rn be
a bounded open set with smooth boundary. We consider the following reaction-diffusion
system with Neumann boundary conditions:
(
∂u
∂t
(x, t) = δ∆u(x, t) + f (u(x, t)) (x, t) ∈ Ω × (0, ∞)
∂u
(12.1)
∂ν
(x, t) = 0 (x, t) ∈ ∂Ω × (0, ∞),
Lectures on Applied PDEs, November 3, 2017 115

where δ is a positive definite n × n matrix (the diffusion tensor) and u : Ω × (0, ∞) → Rn .


Consider the associated kinetic system (the reaction equation)
∂u(t)
= f (u(t)). (12.2)
∂t
Notice that static solutions to (12.2) are also static, homogeneous (i.e. x−independent)
solutions of (12.1) and, if a static, homogeneous solution to (12.2) is stable, then it is also
a stable solution of (12.1) under homogeneous perturbations.
Turing proposed that pattern formation in biological systems could arise through
a process where a stable static solution of (12.2) became unstable in the presence of
diffusion, i.e., the static solution is not stable as a solution of (12.1). This is called Turing
or diffusion-driven instability.
Let u ≡ u∗ be a static homogeneous solution to (12.1), i.e. u∗ solves f (u∗ ) = 0.
u = u∗ is also a static solution of (12.2). If u(t) = u∗ is a stable equilibrium (static
solution) of (12.2), then u(x, t) ≡ u∗ is also a stable solution of (12.1) under homogeneous
perturbations.
The problem then is: if u = u∗ is a stable solution of (12.2), what conditions on f ,
d, and Ω determine the stability of u = u∗ as a solution of (12.1), and in the case where
u = u∗ becomes unstable, what can be said about the evolution of solutions which are
initially close to u = u∗ . We address this problem in this notes.
To simplify the notation, we take in what follows u∗ = 0. In this section we prove the
following
Theorem 12.1. Suppose n = 2 and δ is a diagonal matrix δ = diag (d1 , d2 ). Let fij be
∂fi
the ij entry of f 0 (u∗ ), i.e. fij = ∂x j
(u∗ ).

1. If f11 ≤ 0, then u∗ is a linearly stable solution of (12.1).


2. If f11 > 0, then there exists  such that u = 0 is linearly stable for d2 < d1 and
linearly unstable for d2 > d1 , provided σ(−∆) ∩ (αδ , βδ ) 6= ∅, for some 0 < αδ < βδ
.
First we observe that −∆ is a self-adjoint operator on L2 (Ω, C) (with the appropriate
boundary conditions) and, since, Ω is bounded, it has a purely discrete spectrum with
non-negative eigenvalues, which we denote by 0 ≤ λ1 ≤ λ2 ≤ . . . , marching off to ∞,
λn → ∞, as n → ∞. We define the n × n matrix

Lδ (λ) = −δλ + f 0 (u∗ ). (12.3)

We derive Theorem 12.1 from the following more general result


Theorem 12.2. Suppose n is even (n = 2) and u(t) = u∗ is a stable equilibrium (static
solution) of (12.2). Then the homogeneous static solution u∗ of (12.1) is a linearly stable,
if and only if det Lδ (λ) > 0 for all λ ∈ σ(−∆).
116 Lectures on Applied PDEs, November 3, 2017

Proof of Theorem 12.2. For the proof of the theorem, we will require a number of lemmas.
We begin investigating the linearization of the r.h.s. of (12.1) around u∗ , which is

Lδ = δ∆ + f 0 (u∗ ). (12.4)

We prove the following result on the spectrum of Ld .


Lemma 12.3. [
σ(Lδ ) = σ(Lδ (λ)). (12.5)
λ∈σ(−∆)

Proof. Let { φλ }λ∈σ(−∆) be a orthonormal basis of eigenfunctions of −∆. Then for any
u : Ω → Rn , there exist vectors cλ ∈ Rn such that
X
u= cλ φλ .
λ

We then have X
(Lδ − z)u = (Lδ (λ) − z) cλ φλ .
λ

Now, since Ω ⊆ Rm is a bounded open set with smooth boundary, Lδ has purely point
spectrum. Suppose that z ∈ σ(Ld ). Then there is non-zero u such that (Lδ − z)u = 0. For
each µ ∈ σ(−∆), we multiply by φ̄µ and integrate, to obtain, using the orthonormality of
the basis, Z
X
0= (Lδ (λ) − z) cλ φ̄µ φλ = (Lδ (µ) − z) cµ ,
λ Ω

which implies that either cµ = 0 or z is an eigenvalue of Lδ (λ). Since u is non-zero, not


all cµ are zero, and therefore z ∈ σ(Ld (µ)) for some µ.
Conversely, suppose that z ∈ σ(Lδ (λ)) for some λ. Fix non-zero c ∈ Rn such that
Lδ (λ)c = 0. Then if we set u = cφλ , (Lδ − z)u = 0. Therefore z ∈ σ(Ld ), and that
completes the proof.
Lemma 12.4. Suppose that λ ≥ 0. Then Tr Lδ (λ) < 0.
Proof. We then have
Tr Lδ (λ) = −λ Tr δ + Tr f 0 (0).
Since f 0 (u∗ ) is negative definite, we know that Tr f 0 (u∗ ) < 0. Since δ > 0 and therefore
Tr δ > 0, this implies the result.
For n = 2, the equation Tr Lδ (λ) < 0 implies
Corollary 12.5. One eigenvalue of Lδ (λ) is always negative. For n = 2, the other
eigenvalue is negative iff det Lδ (λ) > 0.
This corollary implies Theorem 12.2.
Lectures on Applied PDEs, November 3, 2017 117

12.1 Appendix: Derivation of Theorem 12.1 from Theorem 12.2


Suppose n = 2 and δ is a diagonal matrix δ = diag (d1 , d2 ). We have

det Lδ (λ) = d1 d2 λ2 − (d1 f22 + d2 f11 )λ + det f 0 (u∗ ) =: hδ (λ). (12.6)

This means that det Ld (λ) > 0 if and only if hδ (λ) > 0. Hence Lδ (λ) is negative definite
if and only if hδ (λ) > 0.
We we begin with studying the function hδ (λ) and prove the following result.

Lemma 12.6.

1. If f11 ≤ 0, then for all d > 0, hδ (λ) > 0 for all λ ≥ 0.

2. If f11 > 0, then there exists δ > d1 with the following properties:

(a) If d2 < δ, hd (λ) > 0 for all λ ≥ 0.


(b) If d2 = δ, there exists λ0 > 0, such that hd (λ) = d1 d2 (λ − λ0 )2 .
(c) If d2 > δ, there exist λ2 > λ1 > 0, such that hd (λ) = d1 d2 (λ − λ1 )(λ − λ2 ).

Proof. It is clear from the definition of hδ (λ), (12.6), that hδ (λ) has a unique global
minimum point, λmin , which can be easily calculated to be

f11 f22
λmin = + . (12.7)
2d1 2d2

We let hmin be the global minimum value. A number of properties of hδ (λ) can also be
determined using its discriminant, D, which is

D = (d1 f22 + d2 f11 )2 − 4d1 d2 det f 0 (0). (12.8)

Now suppose f11 ≤ 0. Since f11 + f22 < 0, we have f22 < −f11 . If f11 = 0, then
λmin < 0 for all d > 0. Now hd (0) = det f 0 (u∗ ) > 0, so it follows that for all λ > 0 > λmin ,
hδ (λ) ≥ hδ (0) > 0. So suppose f11 < 0. Fix d1 and set  = − ff22 11
d1 . Then if d2 > ,
then again λmin < 0 and the same argument gives hδ (λ) > 0 for all λ ≥ 0. Suppose then
that d2 < . We will show that hmin > 0, or equivalently that D < 0. Since d2 < ,
d22 f11
2
< d21 f22
2
. Also, since det f 0 (0) > 0, we have f12 f21 < f11 f22 . Therefore,

D = d21 f22
2
+ d22 f11
2
− 2d1 d2 f11 f22 + 4d1 d2 f12 f21
2 2
< 2d1 f22 + 2d1 d2 f11 f22
< 2d21 f22
2
− 2d21 f22
2
= 0.

That establishes (1).


118 Lectures on Applied PDEs, November 3, 2017

Now suppose f11 > 0. Then f22 < −f11 < 0. Set  = − ff22 11
d1 , so  > d1 . For d2 ≤ ,
one can check that λmin ≤ 0, and therefore the same argument as before shows that
hδ (λ) > 0 for all λ ≥ 0.
For d2 > , on the other hand, λmin > 0. As can be seen from above, the discriminant
Disc is a quadratic polynomial in d2 and D → ∞ as d2 → ∞. Now for d2 = , a simple
calculation shows that
D = −4d1 d2 det f 0 (0) < 0.
This means there exist a α >  such that D < 0 for  ≤ δ < α, D = 0 at d2 = α, and
D > 0 for d2 > α. One can calculate that
1 h p i
α = 2 2 −f12 f21 Tr f 0 (u∗ ) + Tr f 0 (u∗ ) − f12 f21 d1 . (12.9)
f11

This then implies (2), except for the assertion that λ0 , λ1 , λ2 > 0, but this follows from
the fact that λmin > 0 and hδ (0) > 0.
We now turn to the proof of Theorem 12.1.
Suppose first that f11 ≤ 0. Then we have seen that for all δ, hδ (λ) > 0 for all λ ≥ 0.
This means that for all λ ≥ 0, Lδ (λ) is negative definite. Now if λ ∈ σ(−∆), then λ ≥ 0,
so this means that the eigenvalues of Ld lie in the plane { Re z < 0 }, and therefore u∗ is
linearly stable.
Now suppose that f11 > 0 and let  be the following (from (12.9)):
1 h p 0 0
i
 = 2 2 −f12 f21 Tr f (u∗ ) + Tr f (u∗ ) − f12 f21 .
f11
Then if d2 < d1 , an argument similar to the one above, shows that u = 0 is stable. If
d2 > d1 , on the other hand, then there are λ ≥ 0, for which hδ (λ) < 0. If there is
λ ∈ σ(−∆) such that hδ (λ) < 0, then Lδ (λ) has an eigenvalue with positive real part,
and therefore so does Lδ , which means that u∗ is not linearly stable. 

12.2 Pattern formation in reaction-diffusion equations


(UNDER CONSTRUCTION) Recall that for n is even (n = 2), the homogeneous
static solution u∗ of (12.1) is a linearly stable, if and only if det Lδ (λ) > 0 for all λ ∈
σ(−∆).

Theorem 12.7. Suppose n is even (n = 2) and and u(t) = u∗ is a stable equilibrium


(static solution) of (12.2). Let δ∗ be the smallest δ at which det Lδ (λ1 ), where λ1 is the
smallest eigenvalue of −∆, vanishes. Then a non-homogeneous static solution to (12.2)
bifurcates at from the homogeneous one. At the bifurcation point, the solution u∗ looses
its stability, while the new non-homogeneous static solution is stable (the exchange of the
stability). The bifurcating solution is of the form ???.
Lectures on Applied PDEs, November 3, 2017 119

Proof. Define F (δ, u) = δ∆u + f (u). Then the static equation for (12.1) can be written
as
F (δ, u) = 0. (12.10)

Suppose n = 2 and δ is a diagonal matrix δ = diag (d1 , d2 ). Let d1 be fixed for simplicity’s
sake. Write d2 = αd1 . We have
1. (12.10) has a trivial branch: F (δ, 0) = 0 for all δ > 0.
2. There exists α ≥ , such that Lδ := du F (δ, 0) d2 =αd1 has a 0 eigenvalue for some δ.
The latter is s.t. 0 is an eigenvalue of Lδ (λ) for some λ ∈ σ(−∆).
3. The multiplicity of the zero eigenvalue of Lδ = the multiplicity of λ as an eigenvalue
of −∆.
4. Fix any non-zero u in the nullspace of Lδ and non-zero v in the nullspace of L∗ .
Then hv, dd2 ,u F (d1 , αd1 , 0)ui =
6 0.
We prove these properties. We know that for d2 > d1 , λmin > 0 and hmin < 0.
We also know that λmin → ∞ as d2 → ∞, so therefore there is a smallest d2 such that
det Lδ (λ) = hδ (λ) = 0 for some λ ∈ σ(−∆). Since hδ (λ) has two roots, it is possible for
there to be two such λ, but in the generic case, there will only be one and we assume that
from now on.
This means that 0 is an eigenvalue of Lδ (λ), which, since n = 2, has at most one
eigenvalue in {Re z ≥ 0}, so it has multiplicity 1.
Now 0 is an eigenvalue of Lδ if and only if it is an eigenvalue of Lδ (λ) for some λ ∈
σ(−∆). To show that the multiplicity of the 0 eigenvalue of Lδ is equal to the multiplicity
of λ as an eigenvalue of −∆, let φ1 , . . . , φn be the n eigenfunctions corresponding to λ.
Then the nullspace of Lδ is generated by { cφ1 , . . . , cφn }, where c is any non-zero vector
in the nullspace of Lδ (λ).
First we show that if c ∈ R2 is such that Lδ (λ)c = 0, then c2 6= 0. Suppose, on
the contrary, that c2 = 0. Then we have f21 c1 = 0, so f21 = 0. This implies that
0 < det f 0 (u∗ ) = f11 f22 , and since f11 > 0, this implies f22 > 0. But that would mean
Tr f 0 (u∗ ) = f11 + f22 > 0 and that contradiction proves c2 6= 0.
A similar proof shows that if c ∈ R2 is such thatLδ (λ)∗ c =0, then d2 6= 0.
0 0
Now, a simple calculation gives dd2 ,u F (δ, u) = , so dd2 ,u F (d1 , αd1 , u) =
  0 d2∆
0 0
. Next, u = a1 cφ1 + · · · + an cφn and v = b1 dφ1 + · · · + bn dφn . So
0 αd1 ∆

hv, dd2 ,u F (d1 , αd1 , 0)ui = c2 d2 ha1 φ1 + · · · + an φn , b1 φ1 + · · · + bn φn i


= c2 d2 (ā1 b1 + · · · + ān bn ),

which can at least be made to not be zero. The four properties above imply the theorem.
120 Lectures on Applied PDEs, November 3, 2017

Literature. [?] found conditions on f 0 (0) and d for u = 0 to be unstable or a linearly


stable. [?] have Ω ⊂ R. They find conditions for when Turing instability can and cannot
occur, see also [?]. For a textbook exposition see [13].

13 Abrikosov lattices
(See also Subsection 20.3, which repeats and expands some definitions and
results.) In this section we consider the Ginzburg-Landau equations of superconductivity
(and particle physics):
∆A Ψ = κ2 (|Ψ|2 − 1)Ψ, (13.1a)

curl∗ curl A = Im(Ψ̄∇A Ψ). (13.1b)


for (Ψ, A) : R2 → C × R2 , ∇A = ∇ − iA, and ∆A = ∇2A , the covariant derivative and
covariant Laplacian, respectively. The equations are discussed in detail in Section 2.3.
Recall that these equations describe equilibrium configurations of superconductors,
and of the U (1) Higgs model from particle physics. In the superconductivity, the complex-
valued function Ψ(x) is called the order parameter, with |Ψ(x)|2 giving the local density
of (Cooper pairs of) superconducting electrons, and A is a vector field called the magnetic
potential, so that B(x) := curl A(x) = the magnetic field. The vector quantity J(Ψ, A) :=
Im(Ψ̄∇A Ψ) = the superconducting current.
In particle physics, Ψ and A are the Higgs and U (1) gauge (electro-magnetic) fields,
respectively.
Geometrically, one can think of A as a connection on the principal U (1)-bundle Rn ×
U (1), n = 2, 3.
We prove existence of the Abrikosov lattices defined in Section 2.3.
In the idealized situation of a superconductor occupying all space and homogeneous
in one direction, we are led to a problem on R2 and so may consider ψ : R2 → C and
A : R2 → R2 . In this case, curl A := ∂1 A2 − ∂2 A1 is a scalar, and for a scalar function,
B(x) ∈ R, curl B = (∂2 B, −∂1 B) is a vector.
After Abrikosov, we look for solutions (Ψ, A) defined on all of R2 , whose all physical
properties, namely the density of superconducting pairs of electrons, ns := |Ψ|2 , the
magnetic field, B := curl A, and the current density, J := Im(Ψ̄∇A Ψ), are doubly-periodic
with respect to some lattice L. (For us, a lattice is the set

L = {m1 ω1 + m2 ω2 : m1 , m2 ∈ Z} ⊂ R2

for some basis vectors {ω1 , ω2 } in R2 ; note that L forms a group under addition.) We call
such states (L−)Abrikosov lattice states.
We formulate a simplified version of the main result of this section. The general result
can be found in [?].
Lectures on Applied PDEs, November 3, 2017 121


Theorem 13.1. For every lattice L satisfying 0 < κ2 −b  1, where b = |Ω| , the equations
(13.1) have an L−Abrikosov lattice solution in a neighbourhood of the branch of normal
solutions.

Before proceeding to the proof we present some general discussion of properties of the
equations (13.1) and of some general notions.

13.1 Properties of the GLE


Symmetries. As we already mentioned in Section 2.3, the Ginzburg-Landau equa-
tions (13.1) admit several symmetries, that is, transformations which map solutions to
solutions.
Gauge symmetry: for any sufficiently regular function χ : R2 → R,

Tχgauge : (Ψ(x), A(x)) 7→ (eiχ(x) Ψ(x), A(x) + ∇χ(x)); (13.2)

Translation symmetry: for any h ∈ R2 ,

Thtransl : (Ψ(x), A(x)) 7→ (Ψ(x + h), A(x + h)); (13.3)

Rotation and reflection symmetry: for any R ∈ O(2) (including the reflections
f (x) → f (−x))

TRrot : (Ψ(x), A(x)) 7→ (Ψ(Rx), R−1 A(Rx)). (13.4)

Exercise 13.1. Prove that the above transformations are symmetries of the Ginzburg-
Landau equations (13.1), i.e. if (Ψ, A) is a solution to (13.1), then so is (Ψ, A), for T be
any of the above transformations.

Thus the set of all solutions of the Ginzburg-Landau equations can be split into equiv-
alence classes of solutions related by gauge transformations. A condition which pick a
subclass of each equivalence class is called the gauge condition. An example of a gauge
condition is div A = 0. This can be also arranged: if div A 6= 0 we can always find a gauge
η s.t. div(A + ∇η) = 0, namely, we take η solving the equation −∆η = div A.
One of the analytically interesting aspects of the Ginzburg-Landau theory is the fact
that, because of the gauge transformations, the symmetry group is infinite-dimensional.

L-equivariant states. The following proposition will play a key role in the proof of
Theorem 13.1:
122 Lectures on Applied PDEs, November 3, 2017

Proposition 13.2. A pair u = (Ψ, A) is an Abrikosov (vortex) lattice iff there are, in
general, multi-valued differentiable functions, ηs : R2 → R, s ∈ L, s.t.

Ψ(x + s) = eiηs (x) Ψ(x) and A(x + s) = A(x) + ∇ηs (x), ∀s ∈ L, (13.5)

or in the short hand, Tstransl u = Tηgauge


s
u. (I.e. u = (Ψ, A) is mapped by lattice translations
into a gauge equivalent pair.)

Proof. If state (Ψ, A) satisfies (13.5), then all associated physical quantities are L−periodic,
i.e. (Ψ, A) is an Abrikosov lattice. In the opposite direction, if (Ψ, A) is an Abrikosov
lattice, then curl A(x) is periodic w.r.to L, and therefore A(x + s) = A(x) + ∇ηs (x),
for some functions ηs (x). Next, we write Ψ(x) = |Ψ(x)|eiφ(x) . Since |Ψ(x)| and J(x) =
|Ψ(x)|2 (∇φ(x) − A(x)) are periodic w.r.to L, we have that ∇φ(x + s) = ∇φ(x) + ∇η̃s (x),
which implies that φ(x + s) = φ(x) + ηs (x), where ηs (x) = η̃s (x) + cs , for some constants
cs .
We call a pair satisfying (13.5) the lattice, or η-gauge-periodic state. In terminology of
Section 2.3 it is an equivariant state w.r. to the group of lattice translations for a lattice
L.
Since Tstrans is a commutative group, we see that the family of functions ηs has the
important cocycle property

ηs+t (x) − ηs (x + t) − ηt (x) ∈ 2πZ. (13.6)

This can be seen by evaluating the effect of translation by s + t in two different ways. We
call ηs (x) ≡ η(s, x) the gauge exponent. (In algebraic geometry it is called the automorphy
exponent.)

Automorphy factors. We denote by Ω and |Ω| the fundamenta lattice cell and its
area, respectively. We list some important properties of ηs :

• If (Ψ, A) satisfies (13.5) with ηs (x), then Tχgauge (Ψ, A) satisfies (13.5) with ηs (x) →
ηs0 (x), where
ηs0 (x) = ηs (x) + χ(x + s) − χ(x). (13.7)

• The functions ηs (x) = 2b s ∧ x + cs , where b satisfies b|Ω| ∈ 2πZ and cs are numbers
satisfying cs+t − cs − ct − 21 bs ∧ t ∈ 2πZ, satisfies (13.6).

• By the cocycle condition (13.6), for any basis {ν1 , ν2 } in L, the quantity
1
c(η) = (ην (x + ν1 ) − ην2 (x) − ην1 (x + ν2 ) + ην1 (x)) (13.8)
2π 2
is independent of x and of the choice of the basis {ν1 , ν2 } and is an integer.
Lectures on Applied PDEs, November 3, 2017 123

• Every exponential ηs satisfying the cocycle condition (13.6) is equivalent to the


exponent
b
s ∧ x + cs , (13.9)
2
for b and cs satisfying b|Ω| = 2πc(gs ) and

1
cs+t − cs − ct − bs ∧ t ∈ 2πZ. (13.10)
2

• The condition (13.6) implies the magnetic flux quantization:

1
Z
curl A = deg Ψ = c(η). (13.11)
2π Ω

Indeed, the first and second statements are straightforward. For the third statement,
by the relation (13.6), ην2 (x + ν1 ) + ην1 (x) − ην1 +ν2 (x) ∈ 2πZ and ην1 (x + ν2 ) + ην2 (x) −
ην1 +ν2 (x) ∈ 2πZ. Subtracting the second relation from the first shows that c(η) is inde-
pendent of x and is an integer. The quantity c(η) is called the Chern number.
For the fourth property, see e.g. [?, ?, ?, ?], though in these papers it is formulated
differently. In the present formulation this property was shown by A. Weil and generalized
in [?].
To prove the fifth statement,
R we Rnote that by Stokes’ theorem, the magnetic flux
through a lattice cell Ω is Ω curl A = ∂Ω A, is given by
Z 1 
ν1 · (A(aν1 + ν2 ) − A(aν1 )) − ν2 · (A(aν2 + ν1 ) − A(aν2 )) da
0
Z 1
 
= ν1 · ∇ην2 (aν1 ) − ν2 · ∇ην1 (aν2 ) da,
0
R
which, by (13.6), gives Ω
curl A = ην2 (ν1 ) − ην2 (0) − ην1 (ν2 ) + ην1 (0) ∈ 2πZ.

Flux quantization. The important property (13.11) of lattice states tells that the
magnetic flux through a lattice cell is quantized:
Z
curl A = 2πn (13.12)

for some integer n. We give another proof of this statement. If |Ψ| > 0 on the boundary
of the cell, we can write Ψ = |Ψ|eiθ and 0 ≤ θ < 2π. The periodicity
R of |Ψ|
H and J Hensure
the periodicity of ∇θ − A and therefore by Green’s theorem, Ω curl A = ∂Ω A = ∂Ω ∇θ
and this function is equal to 2πn since Ψ is single-valued.
124 Lectures on Applied PDEs, November 3, 2017

Equation (13.12) R then implies the relation between the average magnetic flux, b, per
1
lattice cell, b = |Ω| Ω curl A, and the area of a fundamental cell

2πn
b= . (13.13)
|Ω|

We note that due to the reflection symmetry of the problem we can assume that b ≥ 0.
From now on we assume that

(Ψ, A) satisfy (13.5) with ηs given by (13.9) - (13.10) and (13.13). (13.14)

Normal solutions and bifurcation. The GLE have two basic homogeneous (x−independent)
solutions. First we mention the homogeneous solution, describing the perfect supercon-
ductor solution, namely uS := (ΨS ≡ 1, AS ≡ 0). Without this solution there would be
no superconductors.
More important for our analysis are the normal (or non-superconducting) solutions,

uN ≡ ub := (ΨN ≡ 0, AN ≡ Ab ),

where Ab is a vector potential with constant magnetic field curl Ab =: b = constant.


Below we use the bifurcation theory developed above to show existence of the Abrikosov
lattice solutions. The branch of ‘trivial’ solutions they bifurcate from is exactly the nor-
mal branch displayed above. b is the bifurcation parameter. Interestingly, it does not
enter the GLE explicitly, but only through the space of solutions.
As we remember from the bifurcation theorem, Theorem 7.2, the key step in the
bifurcation analysis is solving the linearized problem, namely finding the values of the
bifurcation parameter, b (the candidates for the bifurcation points), for which this problem
has a non-trivial solution. In the next subsection we solve the linearized problem, the
nonlinear analysis is carried out in Appendix 13.3.

Lattices. By identifying R2 with C, any lattice L can be given a basis {ν1 , ν2 } such
that the complex number τ = νν21 satisfies Imτ > 0. τ will be called the shape parameter
of the lattice. By rotating L, if necessary, we can bring it to the form

Lω = r(Z + τ Z), where ω = (τ, r), with τ ∈ C, Imτ > 0, and r > 0. (13.15)

By the quantization condition (13.13),


r
2πn
r := . (13.16)
Imτ b
Lectures on Applied PDEs, November 3, 2017 125

13.2 The linear problem


To find candidates for bifurcation points we have solve the linear problem:

− ∆Ab ψ = κ2 ψ, (13.17)

for ψ satisfying the gauge - periodic boundary condition (see (13.14))


b
ψ(x + s) = ei( 2 x∧s+cs ) ψ(x), ∀s ∈ Lω . (13.18)

This quasiperiodic boundary condition is consistent with the fact that ψ is a single valued
function if and only if the magnetic flux, b|Ω|, through the fundamental cell Ω is quantized:
b|Ω| = 2πn, for some integer n.
We consider this eigenvalue problem above on the Sobolev space of order two, H 2 (Ω, C),
whose elements satisfy the quasiperiodic boundary condition (13.18). We identify R2 with
C in the usual way as x = (x1 , x2 ) ↔ z = x1 + ix2 . The key result here is
Theorem 13.3. Let b be determined by the quantization condition b = 2πn/|Ω|. The
smallest |Ω| for which the problem (13.17) - (13.18) has a non-trivial solution is given by
b = κ2 . In this case the space solutions is of the dimension n.

Remark. The value b = κ2 is called the second critical magnetic field and is denoted
Hc2 , so that Hc2 = κ2 .
Proof of Theorem 13.3. We consider the operator −∆Ab on the space L2 (Ω, C), with the
domain consisting of functions from H 2 (Ω, C) satisfying the boundary conditions (13.18).
By a standard result, it is self-adjoint. Spectral information about −∆Ab can be obtained
by introducing the complexified covariant derivatives (harmonic oscillator annihilation
and creation operators), ∂¯A and ∂¯A∗ = −∂A , with
1
∂¯A := ((∇A )1 + i(∇Ab )2 ). (13.19)
2
Remembering the definition of ∇A , we compute

∂¯A = ∂¯ + iĀc , (13.20)

where ∂¯ := 21 (∂x1 + i∂x2 ) and Ac := A1 − iA2 . One can verify that these operators satisfy
the following relations:
1
[∂¯A , ∂¯A∗ ] = curl A; (13.21)
4
− ∆A − curl A = 4∂¯A∗ ∂¯A . (13.22)

(As for the harmonic oscillator (see [6]), this, together with the relation curl Ab = b, gives
σ(−∆Ab ) = { (2k + 1)b : k = 0, 1, 2, . . . }.)
126 Lectures on Applied PDEs, November 3, 2017

Exercise 13.2. Prove the relations (13.21) and (13.22).


The second property and the relation curl Ab = b imply
Lemma 13.4.
Null(−∆Ab − b) = Null ∂¯Ab . (13.23)
Proof. if ∂¯Ab φ = 0, then by the second property, (−∆Ab − b)φ = 0. If (−∆Ab − b)φ = 0, on
the other hand (−∆Ab − b)φ = 0, then by the second property, ∂¯A∗ b ∂¯Ab φ = 0. Multiplying
this by φ, we have 0 = hφ, ∂¯A∗ b ∂¯Ab φi = h∂¯Ab φ, ∂¯Ab φi = k∂¯Ab φk2 , which implies ∂¯Ab φ =
0.
To find Null ∂¯Ab , we solve for h satisfying h−1 ∂¯Ab h = ∂. ¯ A simple calculation gives
− 2b (ix1 x2 −x22 )
h=e . (This solution is highly non-unique.) Hence we have
b 2 b 2
e 2 (ix1 x2 −x2 ) ∂¯Ab e− 2 (ix1 x2 −x2 ) = ∂.
¯
b
This immediately proves that Ψ ∈ Null ∂¯Ab if and only if ξ(x) = e 2 (ix1 x2 −x2 ) ψ(x) satisfies
2

¯ = 0.
∂ξ
For z 0 = 1r (x1 + ix2 ), where r is given in (13.16), and z = x1 + ix2 , we define θ(z 0 ) by
the relation
b 2 2 1
ψ0 (z) = e− 4 (|z| −z ) θ(z 0 ), z 0 := z, (13.24)
r
By the above, the function θ is entire and, due to the periodicity conditions on φ, satisfies

θ(z + 1) = θ(z), (13.25a)


1
θ(z + τ ) = e−2πin(z+ 2 τ ) θ(z). (13.25b)
(θ is the theta function, see [?]).
To complete the proof, we now need to show that the space of the analytic functions
which satisfy these relations form a vector space of dimension n. By the first relation, θ
has the absolutely convergent Fourier expansion

X
θ(z) = ck e2πkiz . (13.26)
k=−∞

The second relation, on the other hand, leads to a relation for the coefficients of the
expansion. Namely, we have

ck+n = einπτ e2kiπτ ck

and that means such functions are determined solely by the values of c0 , . . . , cn−1 and
therefore form an n-dimensional vector space. This completes the proof of Theorem 13.3.
Lectures on Applied PDEs, November 3, 2017 127

Corollary 13.5. Let b = κ2 . Then the solutions of the linear problem (13.17) - (13.18)
are of the form(13.24) - (13.25).
Abrikosov considered the case n = 1. In this case, the space (13.23) is one-dimensional
and spanned by the function
∞ √
i
X
ψ := e 2 x2 (x1 +ix2 ) ck eik 2πImτ (x1 +ix2 )
, (13.27)
k=−∞
k−1
Y
ikπτ
ck = ce ei2mπτ . (13.28)
m=1

This is the leading approximation to the Abrikosov lattice solution. The normalization
coefficient c cannot be found from the linear theory and is obtained by taking into account
nonlinear terms by perturbation theory.
The rest of the proof of Theorem 13.1 is given in Appendix 13.3.

13.3 Appendix: Proof of Theorem 13.1 (todo)


13.4 Fixing the gauge
The gauge symmetry allows one to fix solutions to be of a desired form. Let Ab (x) = 2b Jx,
where J is the symplectic matrix
 
0 −1
J= . (13.29)
1 0
We will use the following proposition
Proposition 13.6. Let (Ψ, A0 ) be an L-lattice state, i.e. it satisfies (2.46) and let b be
the average magnetic flux per cell. Then there is a L-lattice state (Ψ, A), that is gauge-
equivalent to a translation of (Ψ, A), such that
b b
Ψ(x + s) = ei( 2 x·Js+cs ) Ψ(x) and A(x + s) = A(x) + Js, ∀s ∈ L, (13.30)
2
where cs satisfy (13.10). Moreover, we have for α := A − Ab ,
(i) α(x + s) = α(x) for all s ∈ L;
R
(ii) α has mean zero: Ω α = 0;
(iii) α is divergence-free: div α = 0.
A proof of this proposition is given in an appendix at the end of this section.
From now on we assume that (Ψ, A) satisfy (13.30).
128 Lectures on Applied PDEs, November 3, 2017

13.5 Rescaling
Suppose, that we have a Lω -lattice state (Ψ, A), where ω = (τ, r). We now define the
rescaled fields (ψ, a) to be

(ψ(x), a(x)) := (rΨ(rx), rA(rx)).

Let Lτ be the lattice spanned by 1 and τ , i. e. Lτ := Z + τ Z, with Ωτ being a primitive


cell of that lattice. We note that |Ωτ | = 2πn. We summarize the effects of the rescaling
above:

(A) (ψ, a) is a Lτ -lattice state.


(B) Ψ and A solve the Ginzburg-Landau equations if and only if ψ and a solve

(−∆a − λ)ψ = −κ2 |ψ|2 ψ, (13.31a)

curl∗ curl a = Im(ψ̄∇a ψ) (13.31b)


for λ = κ2 r2 . The latter equations are valid on Ωτ with the boundary conditions
given in the next statement.
(C) If (Ψ, A) is of the form described in Proposition 13.6, then (ψ, a) satisfies
kn
ψ(x + t) = ei 2
x·Jt
ψ(x), (13.32)

where k = k(τ ) := r2 b/n = and, with a = an + α, an (x) := kn
Imτ
, 2
Jx,
Z
τ
α(x + t) = α(x), ∀t ∈ L , and α = 0, div α = 0. (13.33)
Ωτ

Our problem then is, for each n = 1, 2, . . ., find (ψ, a), solving the rescaled Ginzburg-
Landau equations (13.31) and satisfying (i).

13.6 Reformulation of the problem


In this section we reduce two equations (13.31) for ψ and a to a single equation for ψ.
We introduce the spaces Ln (τ ) := L2 (Ωτ , C) and L~ (τ ) := {a ∈ L2 (Ωτ , R2 ) | hai =
0, div a = 0, in the distributional sense} and the Sobolev spaces of order two: Hn (τ ) and
H~ (τ ), whose elements, ψ and α, satisfy the quasiperiodic boundary condition (13.32)
and (13.33), respectively. Define the operators

Ln := −∆an and M := curl∗ curl, (13.34)

on the spaces Ln (τ ) and L~ (τ ), with the domains being . Their properties that will be
used below are summarized in the following propositions:
Lectures on Applied PDEs, November 3, 2017 129

Proposition 13.7. Ln is a self-adjoint operator on Hn (τ ) with spectrum σ(Ln ) = { (2m+



1)kn : m = 0, 1, 2, . . . } and dimC Null(Ln − kn) = n, where recall k = k(τ ) = Imτ .

Proposition 13.8. M is a strictly positive operator on H~ (τ ) with discrete spectrum.

The proofs of these results are standard and, for the convenience of the reader, are
given below.

Proof of Proposition 13.8. The fact that M is positive follows immediately from its def-
inition. We note that its being strictly positive is the result of restricting its domain to
elements having mean zero.

Proof of Proposition 13.7. First, we note that Ln is clearly a positive self-adjoint operator.
To see that it has discrete spectrum, we first note that the inclusion H 2 ,→ L2 is compact
for bounded domains in R2 with Lipschitz boundary (which certainly includes lattice
cells). Then for any z in the resolvent set of Ln , (Ln − z)−1 : L2 → H 2 is bounded and
therefore (Ln − z)−1 : L2 → L2 is compact. In fact, the operator Ln is unitary equivalent
(by rescaling) to k times the operator L, whose the spectrum was found explicitly in the
previous section. This completes the proof of Proposition 13.7.

Our goal now is to solve the equation (13.31b) for α and, substituting the solution
into the equation (13.31a), find an equation containing only ψ. First we reformulate the
equations (13.31), by substituting a = an + α, to obtain

(Ln − λ)ψ + 2iα · ∇an ψ + |α|2 ψ + κ2 |ψ|2 ψ = 0, (13.35a)

(M + |ψ|2 )α − Im(ψ̄∇an ψ) = 0. (13.35b)


In Appendix 13.10 at the end of this section, we solve the second equation, (13.35b), for
α in terms of ψ. We write the result as α = α(ψ), where

α(ψ) = (M + |ψ|2 )−1 Im(ψ̄∇an ψ). (13.36)

We collect the elementary properties of the map α in the following proposition, where we


identify Hn (τ ) with a real Banach space using ψ ↔ ψ := (Re ψ, Imψ).

Proposition 13.9. The unique solution, α(ψ), of (13.35b) maps Hn (τ ) to H~ (τ ) and


has the following properties:

(a) α(·) is analytic as a map between real Banach spaces.

(b) α(0) = 0.

(c) For any δ ∈ R, α(eiδ ψ) = α(ψ).


130 Lectures on Applied PDEs, November 3, 2017

Proof. The only statement that does not follow immediately from the definition of α is
(a). It is clear that Im(ψ̄∇an ψ) is real-analytic as it is a polynomial in ψ and ∇ψ, and
their complex conjugates. We also note that (M − z)−1 is complex-analytic in z on the
resolvent set of M , and therefore, (M + |ψ|2 )−1 is analytic. (a) now follows.
Now we substitute the expression (13.36) for α into (13.35a) to get a single equation

F (λ, ψ) = 0, (13.37)

where the map F : R × Hn (τ ) → Ln (τ ) is defined as

F (λ, ψ) = (Ln − λ)ψ + 2iα(ψ) · ∇an ψ + |α(ψ)|2 ψ + κ2 |ψ|2 ψ. (13.38)

For a map F (λ, ψ), we denote by ∂ψ F (λ, φ) its Gâteaux derivative in ψ at φ. The
following proposition lists some properties of F .

Proposition 13.10.

(a) F is analytic as a map between real Banach spaces,

(b) for all λ, F (λ, 0) = 0,

(c) for all λ, ∂ψ F (λ, 0) = Ln − λ,

(d) for all δ ∈ R, F (λ, eiδ ψ) = eiδ F (λ, ψ).

(e) for all ψ, hψ, F (λ, ψ)i ∈ R.

Proof. The first property follows from the definition of F and the corresponding ana-
lyticity of a(ψ). (b) through (d) are straightforward calculations. For (e), we calculate
that
Z
n
hψ, F (λ, ψ)i = hψ, (L − λ)ψi + 2i ψ̄α(ψ) · ∇ψ
Ωτ
Z Z Z
2 2 2 2
+2 n
(α(ψ) · a )|ψ| + |α(ψ)| |ψ| + κ |ψ|4 .
Ωτ Ωτ Ωτ

The final three terms are clearly real and so is the first because Ln − λ is self-adjoint. For
the second term we calculate the complex conjugate and see that
Z Z Z
2i ψ̄α(ψ) · ∇ψ = −2i ψα(ψ) · ∇ψ̄ = 2i (∇ψ · α(ψ))ψ̄,
Ωτ Ωτ Ωτ

where we have integrated by parts and used the fact that the boundary terms vanish due
to the periodicity of the integrand and that div α(ψ) = 0. Thus this term is also real and
(e) is established.
Lectures on Applied PDEs, November 3, 2017 131

13.7 Reduction to a finite-dimensional problem


In this section we reduce the problem of solving the equation F (λ, ψ) = 0 to a finite
dimensional problem. We address the latter in the next section. We use the standard
method of Lyapunov-Schmidt reduction. Let X := Hn (τ ) and Y := Ln (τ ) and let
K = Null(Ln − kn). We let P be the Riesz projection onto K, that is,
1
I
P := − (Ln − z)−1 dz, (13.39)
2πi γ
where γ ⊆ C is a contour around n that contains no other points of the spectrum of
Ln . This is possible since n is an isolated eigenvalue of Ln . P is a bounded, orthogonal
projection, and if we let Z := Null P , then Y = K ⊕ Z. We also let Q := I − P , and so
Q is a projection onto Z.
The equation F (λ, ψ) = 0 is therefore equivalent to the pair of equations
P F (λ, P ψ + Qψ) = 0, (13.40)
QF (λ, P ψ + Qψ) = 0. (13.41)
We will now solve (13.41) for w = Qψ in terms of λ and v = P ψ. To do this, we
introduce the map G : R × K × Z → Z to be G(λ, v, w) := QF (λ, v + w). Applying
the Implicit Function Theorem to G, we obtain a function w : R × K → Z, defined on a
neighbourhood of (n, 0), such that w = w(λ, v) is a unique solution to G(λ, v, w) = 0, for
(λ, v) in that neighbourhood. This solution has the following properties
w(λ, v) real-analytic in (λ, v); (13.42)
kwk = O(kvk3 ) and k∂λ wk = O(kvk3 ), (13.43)
where the norms are in the space Hn (τ ). The last property follows from (13.36) and
(21.31) together with (13.41). (Indeed, by (13.36), we have kα(ψ)kH 2 . kψk2H 2 and
therefore combining (21.31) with (13.41) and using (as above) that the product of Hn (τ )
functions is again a Hn (τ ) function (and the norms are bounded correspondingly), one
concludes that (13.43) holds.)
We substitute the solution w = w(λ, v) into (13.40) and see that the latter equation
in a neighbourhood of (n, 0) is equivalent to the equation (the bifurcation equation)
γ(λ, v) := P F (λ, v + w(λ, v)) = 0. (13.44)
Note that γ : R × K → C. We have shown that in a neighbourhood of (n, 0) in R × X,
(λ, ψ) solves F (λ, ψ) = 0 if and only if (λ, v), with v = P ψ, solves (13.44). Moreover, the
solution ψ of F (λ, ψ) = 0 can be reconstructed from the solution v of (13.44) according
to the formula
ψ = v + w(λ, v). (13.45)
Finally we note that w and γ inherit the symmetry of the original equation:
132 Lectures on Applied PDEs, November 3, 2017

Lemma 13.11. For every δ ∈ R, w(λ, eiδ v) = eiδ w(λ, v) and γ(λ, eiδ v) = eiδ γ(λ, v).

Proof. We first check that w(λ, eiδ v) = eiδ w(λ, v). We note that by definition of w,

G(λ, eiδ v, w(λ, eiδ v)) = 0,

but by the symmetry of F , we also have G(λ, eiδ v, eiδ w(λ, v)) = eiδ G(λ, v, w(λ, v)) = 0.
The uniqueness of w then implies that w(λ, eiδ v) = eiδ w(λ, v). We can now verify that

γ(λ, eiδ v) = P F (λ, eiδ v + w(λ, eiδ v))


= eiδ P F (λ, v + w(λ, v))i = eiδ γ(λ, v).

Solving the bifurcation equation (13.44) is a subtle problem, unless n = 1. In the


latter case, this is done in the next section.

13.8 Proof of Theorem 13.1


In this section we look at the case n = 1, and look for solutions near the trivial solution.

Recall that by Theorem 13.3, the nullspace of the operator Ln −kn, where k = Imτ , acting
on Hn (τ ) is a one-dimensional complex subspace for n = 1. We denote a0 = an=1 and
drop the (super)index n = 1 from the notation Ln . We begin with the following result
which gives the existence and uniqueness of the Abrikosov lattices.

Theorem 13.12. For every τ there exist  > 0 and a branch, (λs , ψs , αs ), s ∈ [0, ),
of nontrivial solutions of the rescaled Ginzburg-Landau equations (13.31), unique modulo
the global gauge symmetry (apart from the trivial solution (1, 0, a0 )) in a sufficiently small
neighbourhood of (1, 0, a0 ) in R × H (τ ) × H~ (τ ), and such that

2
λs = k + gλ (s ),

ψs = sψ0 + sgψ (s2 ), (13.46)
 2
as = a0 + ga (s ),

where k = Imτ2π
, (L − k)ψ0 = 0, gψ is orthogonal to Null(L − k), gλ : [0, ) → R, gψ :
[0, ) → H (τ ), and gα : [0, ) → H~ (τ ) are real-analytic functions such that gλ (0) = 0,
gψ (0) = 0, gα (0) = 0.

Proof. The proof of this theorem is a slight modification of a standard result from bifur-
cation theory. Our goal is to solve the equation (13.44) for λ. Since the projection P ,
defined there, is rank one and self-adjoint, we have
1
Pψ = hψ0 , ψiψ0 , with ψ0 ∈ Null ∂ψ F (λ0 , 0). (13.47)
kψ0 k2
Lectures on Applied PDEs, November 3, 2017 133

We can therefore view the function γ in the bifurcation equation (13.44) as a map γ :
R × C → C, where
γ(λ, s) = hψ0 , F (λ, sψ0 + w(λ, sψ0 )i. (13.48)
We now show that γ(λ, s) ∈ R. Since the projection Q is self-adjoint, Qw(λ, v) = w(λ, v),
w(λ, v) solves QF (λ, v + w) = 0 and v = sψ0 , we have

hw(λ, sψ0 ), F (λ, sψ0 + w(λ, sψ0 ))i = hw(λ, sψ0 ), QF (λ, sψ0 + w(λ, sψ0 ))i = 0.

Therefore, for s 6= 0,

hψ0 , F (λ, sψ0 + Φ(λ, sψ0 ))i = s−1 hsψ0 + w(λ, sψ0 ), F (λ, sψ0 + w(λ, sψ0 ))i,

and this is real by property (e) of Proposition 13.10. Thus, since by Lemma 13.11,
γ(λ, s) = ei arg s γ(λ, |s|), it therefore suffices to solve the equation

γ0 (λ, s) = 0 (13.49)

for the restriction γ0 : R × R → R of the function γ to R × R, i.e. for real s. Since by


(13.43), w(λ, sψ0 ) = O(s2 ) and therefore (13.49) has the trivial branch of solutions s ≡ 0
for all λ. Hence we factorize γ0 (λ, s) as γ0 (λ, s) = sγ1 (λ, s), i.e. we define the function

γ1 (λ, s) := s−1 γ0 (λ, s), if s > 0, and = 0 if s = 0, (13.50)

and solve the equation γ1 (λ, s) = 0 for λ. The definition of the function γ1 (λ, s) and
(13.42) imply that it has the following properties: γ1 (λ, s) is real-analytic, γ1 (λ, −s) =
γ1 (λ, s), γ1 (1, 0) = 0 and, by (21.31) and (13.43), ∂λ γ1 (1, 0) = −kψ0 k2 6= 0. Hence by a
standard application √ √ of the Implicit Function Theorem, there is  > 0 and a real-analytic
function φλ : (− , √
e ) → R such that φeλ (0) = 1 and λ = φeλ (|s|) solves the equation
γ(λ, s) = 0 with |s| < .
We also note that because of the symmetry, φeλ (−|s|) = φeλ (|s|), φeλ is an even real-
analytic function,
√ and therefore must in fact be a function solely of s2 . We therefore set
φλ (s) = φeλ ( s) for s ∈ [0, ), and so φλ is real-analytic.
We now define gψ : [0, ) → H (τ ) to be
( √
√1 w(φλ (s), sψ0 ) s 6= 0,
s
gψ (s) = (13.51)
0 s = 0,

2
It is easily
√ check that gψ is real-analytic and satisfies sgψ (s ) = w(φλ (s), sψ0 ) for any
e
s ∈ [0, ).
Now, we know that there is a neighbourhood of (1, 0) in R × H (τ ) such that in this
neighbourhood F (λ, ψ) = 0 if and only if γ(λ, s) = 0 where P ψ = sψ0 . By taking a smaller
neighbourhood if necessary, we have proven that F (λ, ψ) = 0 in this neighbourhood if and
134 Lectures on Applied PDEs, November 3, 2017

only if either s = 0 or λ = φλ (s2 ). If s = 0, we have ψ = sψ0 + w(φeλ (s), sψ0 ) = 0, which


gives the trivial solution. In the other case, ψ = sψ0 + w(φeλ (s), sψ0 ) = sψ0 + sgψ (s2 ).
If we now also define gλ (s) = 1 − φλ (s), then the above gives us a neighbourhood of
(1, 0) in R × H (τ ) such that the only non-trivial solutions of the equation (13.37) are
given by the first two equations in (13.46). We now define the function g̃a (s) = α(ψs ),
where, recall, α(ψ) is defined in (13.36) and ψs = sψ0 + sgψ (s2 ) was found above. This
function is real-analytic and satisfies g̃a (−s) = α(−ψs ) = g̃a (s), and therefore
√ is really a
2 2 2
function of s , ga (s ). Define as = a0 + ga (s ). Then (λs , ψs , αs ), s ∈ [0, ), solves the
rescaled Ginzburg-Landau equations (13.31).

For ψ0 a non-zero element in the nullspace Null(L − k), we define the function of τ as

h|ψ0 |4 i
β(τ ) := , (13.52)
h|ψ0 |2 i2

which we call this function the Abrikosov function.


Simple computations give the following expression for the derivative gλ0 (0) at 0 of the
function gλ (s2 ) defined in (13.46)
  
0 2π 2 1 1
gλ (0) = κ − β(τ ) + h|ψ0 |2 i. (13.53)
Imτ 2 2

Note that the definition λ = κ2 r2 (n = 1), the first equation (13.46) and the relation
(13.53) imply that for (κ2 − 21 )β(τ ) + 21 ≥ 0, the bifurcated solution exists for b ≤ κ2 , and
for (κ2 − 12 )β(τ ) + 12 < 0, it exists for b > κ2 . Thus Theorem 13.12, after rescaling to the
original variables, implies Theorem 13.1.

Remark. The proof of Theorem 13.12 gives in fact the following abstract result.

Theorem 13.13. Let X and Y be complex Hilbert spaces, with X a dense subset of Y ,
and consider a map F : R × X → Y that is analytic as a map between real Banach spaces.
Suppose that for some λ0 ∈ R, the following conditions are satisfied:

(1) F (λ, 0) = 0 for all λ ∈ R,

(2) ∂ψ F (λ0 , 0) is self-adjoint and has an isolated eigenvalue at 0 of (geometric) multiplic-


ity 1,

(3) For non-zero ψ0 ∈ Null ∂ψ F (λ0 , 0), hψ0 , ∂λ,ψ F (λ0 , 0)ψ0 i =
6 0,

(4) For all α ∈ R, F (λ, eiα ψ) = eiα F (λ, ψ).

(5) For all ψ ∈ X, hψ, F (λ, ψ)i ∈ R.


Lectures on Applied PDEs, November 3, 2017 135

Then (λ0 , 0) is a bifurcation point of the equation F (λ,√ ψ) = 0, in the sense that there is
a family of non-trivial solutions, (λs , ψs ), for s ∈ [0, ), unique modulo the global gauge
symmetry (apart from the trivial solution (1, 0)) in a neighbourhood of (λ0 , 0) in R × X.
Moreover, this family has the form
(
λ = φλ (s2 ),
ψ = sψ0 + sφψ (s2 ).
Here ψ0 ∈ Null ∂ψ F (λ0 , 0), and φλ : [0, ) → R and φψ : [0, ) → X are unique real-
analytic functions, such that φλ (0) = λ0 , φψ (0) = 0.

13.9 Appendix: Proof of Proposition 13.6


Proof of Proposition 13.6. We begin by defining the function B : R → R to be
1 r
Z
B(ζ) = curl A0 (ξ, ζ) dξ.
r 0

It is clear that b = rτ12 0 B(ζ) dζ. A calculation shows that B(ζ + rτ2 ) = B(ζ).
R 2

We now define P = (P1 , P2 ) : R2 → R2 to be


Z x2 Z x1
τ ∧x
P (x) = (bx2 − B(ζ) dζ, curl A0 (ξ, x2 ) dξ + B(x2 )).
0
τ1
τ
x2 τ2
2

A calculation shows that P is doubly-periodic with respect to L.


We now define η 0 : R2 → R to be
Z x1 Z x2
0 b 0
η (x) = x1 x2 − A1 (ξ, 0) dξ − A02 (x1 , ζ) − P2 (x1 , ζ) dζ.
2 0 0

η 0 satisfies ∇η 0 = −A0 + A0 + P and let η 00 be a doubly-periodic solution of the equation


∆η 00 = − div P . Also let C = (C1 , C2 ) be given by
1
Z
C=− (P + ∇η 00 ) dx,
|Ω| Ω
where Ω is any fundamental cell, and set η 000 = C1 x1 + C2 x2 and define η = η 0 + η 00 + η 000 .
We claim that the pair (eη Ψ0 , A0 +∇η) satisfies (13.30) and (i) - (iii) of the proposition.
Let α := A − Ab = A0 + ∇η − Ab . We first note that α = P + ∇η 00 + C and by
00
the
R above, R it is periodic. We also calculate that div α = div P + ∆η = 0. Finally

α = Ω (P + ∇η − C) = 0.
b b b
We note  since α(x) and L−periodic, A (x) satisfies A (x + s) = A (x) +
 that,
b −s2 −s2
2
and = ∇(s ∧ x), we have that A = Ab + α satisfies
s1 s1
b
A(x + s) = A(x) + ∇(s ∧ x), ∀s ∈ L.
2
136 Lectures on Applied PDEs, November 3, 2017

Next, if (Ψ0 , A0 ) satisfies condition (2.46) with the exponent gs0 (x), then (eη Ψ0 , A0 +∇η)
satisfies condition (2.46) with the exponent gs (x) := gs0 (x) + η(x + s) − η(x). We have
shown above that for our choice of η, ∇gs (x) = 2b ∇(t ∧ x). Therefore gs (x) = 2b s ∧ x + cs
for some constant cs . To establish (13.30), we need to have it so that cs = 0 for s = r, rτ .
First let l be such that r ∧ l = − cbr and rτ ∧ l = − crτ b
. This l exists as it is the solution
to the matrix equation
    cr 
0 r l1 −b
= ,
−rτ2 rτ1 l2 − crτ
b

and the determinant of the matrix is just r2 τ2 , which is non-zero because (r, 0) and rτ
form a basis of the lattice. Let ζ(x) = 2b l ∧ x. A straightforward calculation then shows
that Ψ(x) := eiζ(x) Ψ0 (x + l) satisfies (13.30) and that α(x) := α0 (x + l) + ∇ζ(x) still
satisfies (i) through (iii). This proves the proposition.

13.10 Appendix: Solving the equation (13.35b)


(under construction)

13.11 Appendix: Remarks about the automorphy factors


Given a family ηs of functions satisfying (13.6), equivariant functions u = (Ψ, A) for gs
are identified with sections of the vector bundle

R2 × (C × R2 )/L,

with the base manifold R2 /L = Ω and the projection p : [(x, u)] → [x], where [(x, u)]
and [x] are the equivalence classes of (x, u) and x, under the action of the group L on
R2 × (C × R2 ) and on R2 , given by

s : (x, u) → (x + s, Tηgauge
s (x)
u) and s : x → x + s,

respectively. This implies, in particular, that

• For a family ηs of functions satisfying (13.6), there exists a continuous pair (Ψ, A)
satisfying (13.5) with this family.

Remarks. 1) Relation (13.6) for Abrikosov lattices was isolated in [?], where it played
an important role. This condition is well known in algebraic geometry and number theory
where eiηs (x) is called the automorphy factor (see e.g. [?]). However, there the associated
vector potential (connection on the corresponding principal bundle) A is not considered
there.
Lectures on Applied PDEs, November 3, 2017 137

0
2) In algebraic geometry and number theory, the automorphy factors eiηs (x) and eiηs (x)
satisfying ηs0 (x) = ηs (x) + χ(x + s) − χ(x), for some χ(x), are said to be equivalent. A
function Ψ satisfying Tstrans Ψ = eiηs Ψ is called eiηs −theta function.
3) The special form (13.9) is related to a general construction of line bundles over
the complex torus using symplectic form ω(z, w) to construct automorphy factors, e.g.
ηs (z) = bω(z, s) + cs , where b|C/L| = 2πn and cs satisfies cs+t − cs − ct − 2b ω(s, t) ∈ 2πZ.
The Chern number, c(η), is expressed in terms of ω as

c(γ) = bω(ν1 , ν2 ), (13.54)

where {ν1 , ν2 } is a basis of L.


4) The exponentials ηs satisfying the cocycle condition (13.6) are classified by the
irreducible representation of the group of lattice translations. This follows from the fact
that cs ’s satisfying (13.10) are classified by the irreducible representation of the group of
lattice translations.

14 Hamiltonian and gradient systems


We will not give a formal definition of a Hamiltonian system (see Appendix ?? for this),
but define it informally as an evolution equation which is the Euler-Lagrange equation
for a functional of the form (10.11). Consequently, it is an equation of the form (10.12)
for some Lagrange functional L.
A key fact about Hamiltonian systems is that they have a conserved functional, called
the energy. Indeed, define the energy functional by
 
E(φ) := dφ̇ L φ̇, φ φ̇ − L φ̇, φ ). (14.1)

This functional is a constant of motion, i.e. the energy is conserved:

∂t E(φ) = 0. (14.2)

We call such systems conservative.


Exercise 14.1. Prove (14.2). Verify it for the examples 4) and 5) above.
To define the gradient evolution equations (or systems), we first define the gradient of
a functional. Let X be a real space with a real inner product, which we denote g, so that
g(v, w) is a bilinear functional satisfying all the conditions of the inner product. Let M
be an open subset of X and E : M → R and differentiable functional on M . We define
the gradient of E at u w.r. to the inner product g as an element gradg E(u) ∈ X defined
by the equation

g(gradg E(u), v) = dE(u)v. (14.3)


138 Lectures on Applied PDEs, November 3, 2017

(By the Riesz representation theorem gradg E(u) is well defined, provided X is a hilbert
space; we will not go into further details on this and refer to [?].) Thus the gradient
defines the map, or the vector field, gradg E : u ∈ X → gradg E(u) ∈ X. Hence we can
define the evolution equation

∂t u = − gradg E(u). (14.4)

Such equations are called the gradient systems or gradient flows. If u satisfies (14.4), then
differentiating E(u) w.r.to to t and using definition (14.3), we find ∂t E(u) = dE(u)∂t u =
hgradg E(u), ∂t ui, which together with (14.4), gives

∂t E(u) = −g(gradg E(u), gradg E(u)) < 0. (14.5)

Thus for gradient systems, the energy decays with time. We call such systems dissipative.
The examples of the hamiltonian systems are the Hartree and nonlinear Schrödinger
equations, (4.17) and (4.19), or the generalized nonlinear Schrödinger equation (4.20).
The energy functionals in all three cases are of the form
Z
|∇x ψ|2 + G(|ψ|2 ) dx,

E(ψ) = (14.6)
Rd

where G(u) is 14 (v ∗ u)|ψ|2 , p1 λup/2 and an anti-derivative of g, g(u) = G0 (u), respectively.


An example of the gradient flow is the nonlinear heat (reaction-diffusion) evolution,
(4.9). The energy functional in this case is

1
Z
E(u) = ( |∇u|2 + F (u)), (14.7)
2

where F (u) := p1 |u|p , and the inner product g is the standard L2 −inner product.

Remark. With the definition of the gradient in (14.3), we can give a more succinct
definition of the hessian of E at u (compared to (10.13)) as the linear operator given by

Hess E(u) := d gradg E(u). (14.8)

14.1 Appendix: The gradient property of the VPF


Furthermore, we have the following important property

• (11.16) is a gradient flow for the area functional on closed surfaces with given en-
closed volume.
Lectures on Applied PDEs, November 3, 2017 139

Proof. To prove the last statement we use the formula (??), which we rewrite in the
present notation as

Z Z
dV (ψ)ξ = Hν · ξ = Hν · ξ gdn u, (14.9)
ψ U

and the fact that ξ is a vector field for deformations with a fixed enclosed volume, i.e.
it is a tangent vector field to the manifold Xc := {ψ : U → Rn+1 , Vencl (ψ) = c} (we do
not specify the topology here, we could R take the topology of differentiable or Sobolev
functions), and therefore it satisfies ψ ξ · ν = 0. Indeed, let ψs be a family of constant
enclosed volume surfaces deforming ψ and let ξ be the corresponding vector field at s = 0,
i.e. ξ = ∂s ψs s=0 . Since Vencl (ψs ) = c, we have dVencl (ψ)ξ = ∂s Vencl (ψs ) s=0 = 0. Hence

the result follows from the formula
Z
dVencl (ψ)ξ = ξ · ν. (14.10)
ψ

This formula can be proven by either considering an infinitesimal change in the enclosed
volume
R under the1 variation of S or by using that, by the divergence theorem, Vencl (ψ) =
1
R
n+1 Ω
div x = n+1 S x · ν, where Ω is domain enclosed by the surface S, described by
the immersion ψ, and then differentiating
R the latter integral, see Appendix ?? (to be
added). Conversely, if f satisfies S f = 0, then there is a volume preserving normal
variation with the vector field f ν. This implies that
Z
Tθ Xc = {ξ : S → R , ξ · ν = 0}.
n+1
S

¯
Now, for anR arbitrary normal R (f − f )ν,
vector field η = f ν, the vector field ξ :=
where f¯ := |S| S f , satisfies S ξ · ν = 0 and therefore we have dV (ψ)ξ = ψ Hν · ξ =
1
R

H(f − f¯) = ψ (H − H)f . This shows that the L2 −gradient of V (ψ) on closed surfaces
R R
ψ
with given enclosed volume is H − H.

15 Energy and orbital stability


15.1 Energy argument
The energy conservation or dissipation restricts severely possible dynamics of our system.
We use this in investigating stability of static solutions, i.e. in understanding the behavior
of solutions to equation (9.1) near an equilibrium point u∗ . We assume that (9.1) is a
such that there exist a functional E(u) (called energy, or entropy or Lyapunov functional),
defined on an open set U in a Hilbert space X, with the following properties:

(i) E(u) is non-increasing under the evolution equation (9.1) (a stronger statement:
∂t E(u(t)) ≤ 0, for any solution u(t) of (9.1))
140 Lectures on Applied PDEs, November 3, 2017

(ii) Static solutions to (9.1) are critical points of E(u).

These conditions are satisfied for the hamiltonian (conservative) and gradient (dissipative)
equations, two key classes of evolution PDE appearing in applications and in mathematics.
These classes are discussed in Subsection 14 and Appendix ??.
On a basis of energy considerations one expects the following behavior
1. if u∗ is a strict local minimizer then u∗ is a stable solution,

2. if u∗ is either a saddle point or a maximizer then u∗ is an unstable solution.


We will turn Possibility #1 into a rigorous statement below. The Possibility #2 is not
always true. Sometimes conservation laws present in the equations can lead to stability
in the cases of the second type. We will study this case below as well.
An important role below is played by the hessian operator

Hess E(u) := d grad E(u),

introduced in (14.8) above. With Theorem 10.5 in mind, we say that E is coercive at u∗
if the following inequality satisfies

Hess E(u∗ ) ≥ θ, for some θ > 0. (15.1)

In what follows we use the notation E 00 ≡ Hess E. We make the following assumptions:
• (a) E(u) is is C 2 , (b) E is coercive at u∗ .
s/2
Since E is coercive at u∗ , we can define the family (scale) of the Hilbert spaces Ys := L∗ Y ,
where L∗ := E 00 (u∗ ). We denote the norm in Ys by k · ks . Clearly, E 00 (u∗ ) : Ys → Ys−2 .
We make a technical assumption that E 00 (u) maps Y1 into Y−1 and is continuous in u.
Theorem 15.1. Under the assumptions (i), (ii), (a) and (b), the static solution u∗ of
the evolution equation (9.1) is Lyapunov stable.
The proof of this theorem is based on the following important energy estimate
Theorem 15.2. Under the assumptions above, there is δ > 0 s.t. for any u satisfying
ku − u∗ k1 ≤ δ, we have the estimate

θku − u∗ k2 ≤ ku − u∗ k21 ≤ 3(E(u) − E(u∗ )). (15.2)

Proof. Using that E ∈ C2 and dE(u∗ ) = 0 and writing u = u∗ + ξ, we expand E(u) around
u∗ to the third order as
1
E(u) = E(u∗ ) + hξ, E 00 (u∗ )ξi + R(ξ) , (15.3)
2
Lectures on Applied PDEs, November 3, 2017 141

where R(ξ) is the remainder term defined by this expression. By the assumption that E
is C2 and the technical assumption, it satisfies the estimate R(ξ) = o(kξk21 ). The vector
ξ is called a fluctuation of u around u∗ . Hence, since hξ, E 00 (u∗ )ξi = kξk21 , we have, for
kξk1 = ku − u∗ k1 ≤ δ,
1 1
E(u) − E(u∗ ) ≥ kξk21 − kξk21 ≥ kξk21 , (15.4)
2 3
which gives (15.2).
Proof of Theorem 15.1. We write the solution ut of (9.1) as ut = u∗ + ξt . Let kξ0 k =
ku0 − u∗ k ≤ δ/2, where u0 is the initial condition, u|t=0 = u0 , and δ > 0 is the same as in
Theorem 15.2. The there is T > 0 s.t. kξt k ≤ δ, for 0 ≤ t ≤ T . Then by Theorem 15.2,
kut − u∗ k21 ≤ 3(E(ut ) − E(u∗ )), which, together with the assumption (i), that E(ut ) is a
non-increasing function of time, E(ut ) ≤ E(u0 ), gives

kut − u∗ k21 ≤ 3(E(u0 ) − E(u∗ )), (15.5)

for 0 ≤ t ≤ T .
Next, by equation (15.3) and estimate the estimate R(ξ) = O(kξk21 ), we have
1 3 3
|E(u0 ) − E(u∗ )| ≤ hξ0 , E 00 (u∗ )ξ0 i + kξk21 ≤ hξ0 , E 00 (u∗ )ξ0 i ≤ kξ0 k21 , (15.6)
2 4 4
where ξ0 := u0 − u∗ and we used again that hξ, E 00 (u∗ )ξi = kξk21 , for kξk1 sufficiently small.
Combining the last two estimates and remembering the definition ξ0 := u0 − u∗ , we find
9
kut − u∗ k21 ≤ ku0 − u∗ k21 , (15.7)
4
for 0 ≤ t ≤ T .
Estimate (15.7) implies the Lyapunov stability of u∗ . Indeed, for any  > 0, we take
δ := 31 . Then kut − u∗ k1 ≤  for ku0 − u∗ k1 ≤ δ, which is the precise statement of the
Lyapunov stability.
2nd proof of Theorem 15.1: Lyapunov functional. We introduce the functional (Lyapunov
functional) Λ(ξ) := 21 hξ, E 00 (u∗ )ξi. Using that, by the assumption, kξk2 ≤ 1θ Λ(ξ) and
R(ξ) ≥ Ckξk3 , for some 0 < C < ∞, we have

E(u) − E(u∗ ) ≥ Λ(ξ) − µΛ(ξ)3/2 , (15.8)

where µ := C/θ3/2 . Denote λ = Λ(ξ)1/2 and α = E(u) − E(u∗ ) and define the function
F (λ) by
F (λ) = α − λ2 + µλ3 .
We have shown that λ satisfies F (λ) ≥ 0. Consider the graph of F (x), see Figure 16.
For sufficiently small α (54αµ2 < 1), the function F (λ) positive in two disjoint intervals
142 Lectures on Applied PDEs, November 3, 2017

x
A

Figure 16: Observe


√ the√first crossing √
of the graph of F and the x-axis is for sufficiently
small α at λ∗ ≈ 2α, ( 2α ≤ λ∗ ≤ 2 α when 52C 2 α < 1) in the graph the y-axis is in
units of α.

[0, λ∗ ) and (λ√∗∗ , ∞), where λ∗∗ > λ∗ > 0 are the zeros of F (λ), λ > 0. For α sufficiently
small, λ∗ ≈ α and λ∗∗ ≈ 1/µ. Then if F (λ) ≥ 0 and λ < λ∗∗ , we must have λ ≤ λ∗ .
Taking δ := 1θ λ2∗∗ , we conclude that if kξk2 ≤ 1θ Λ(ξ) < δ, then

1 1 1
kξk2 ≤ Λ(ξ) ≤ λ2∗ ≈ α. (15.9)
θ θ θ

By (15.8), the function λt := Λ(ξt )1/2 satisfies F (λt ) ≥ 0, where the function F is
defined there. By the (15.9) and the continuity of Λ(ξt ), if kξ0 k2 ≤ 1θ Λ(ξ0 ) < δ, then
kξt k2 ≤ 1θ Λ(ξt ) ≤ 1θ λ∗ (t)2 ≈ 1θ (E(u) − E(u∗ )). Here ξ0 is the initial condition, ξ|t=0 = ξ0 .
(If λt starts at t = 0 at λ0 ≤ λ∗ (0), then since f (λt ) > 0, λt must stay in the interval
[0, λ∗ (t)].)
Next, by the assumption (i), E(ut ) is a non-increasing function of time: E(ut ) ≤ E(u0 ),
where u0 is the initial condition, u|t=0 = u0 . The last two results give

4
kξk ≤ (E(u0 ) − E(u∗ )), (15.10)
δ
which shows that by making u0 to be close to u∗ , we can make ut to be arbitrary close to
u∗ . Hence the static solution u∗ is Lyapunov stable. (needs cleaning)

15.2 Systems with symmetry


For systems with symmetry, the notion of stability/instability should be modified. Recall
from Subsection 2.3, that we say that the dynamical system (9.1) has a symmetry group
G if a representation T : g → Tg of G on the space of solutions satisfies

F (Tg u) = Tg F (u). (15.11)


Lectures on Applied PDEs, November 3, 2017 143

As was discussed in Subsection 9.1, in this case, if the dynamical system (9.1) with the
symmetry group G has a static solution u∗ , then it has the manifold of static solutions

M∗ = {Tg u∗ : g ∈ G}.

Indeed, if u∗ is a static solution to (9.1), then, due to (15.11), so is Tg u∗ . Now, we have


to address the stability w.r.to this manifold. Clearly, we have to modify the definitionsat
the beginning of this section and the approach.
We say that a static solution u∗ (or more precisely, the manifold of static solutions M∗ )
is orbitally stable if any solution to the equation (9.1), starting in a small neighborhood
of u∗ stays in a small neighborhood of the manifold M∗ . In other words the solution ut
sticks very close to a possibly moving static solution Tg(t) u∗ .
As far as the energy is concerned, the symmetry requirement means that

E(Tg u) = E(u). (15.12)

This implies that, if u∗ is a critical point of E(u), then so is Tg u∗ . (This can be proved
independently by differentiating the equality E(Tg (u∗ + sξ)) = E(u∗ + sξ) w.r. to s.)
Note that the tangent space, Tu∗ M∗ ⊂ X, to M∗ at u∗ is the vector space τ (g)u∗ ,
where τ (g) is the representation of the Lie algebra g of G acting on X. (This can be seen
by differentiating Tg(s) u∗ , where Tg(s) is any one-parameter subgroup of G, w.r. to s.)
As we know from Subsection 8, in the case of symmetry, the inequality (15.1) never
holds: the hessian E 00 (u∗ ) has always zero eigenvectors (see Proposition 8.1). Indeed, we
have Thus in the case of symmetries, if Au∗ 6= 0 for some A ∈ τ (g), then the assumption
(b) fails. We replace it by the following weaker assumption

(b’) E is coercive at u∗ in the sense that

hξ, E 00 (u∗ )ξi ≥ θkξk2 , for some θ > 0 and ∀ξ ⊥ Tu∗ M∗ . (15.13)

If Au∗ 6= 0, or what is the same esA u∗ 6= u∗ , for some A ∈ τ (g), then we say that u∗
breaks the symmetry corresponding to the one-parameter subgroup {esA }. For a given u∗ ,
consider the maximal subgroup, G∗ , of G, which is broken by u∗ , i.e. gu∗ 6= u∗ , ∀g ∈ G∗ .
In what follows, we use this subgroup instead of G.
Now, we have the following strengthening of Theorem 15.1:

Theorem 15.3. Under the assumptions (i), (ii), (a) and (b’), the static solution u∗ of
the evolution equation (9.1) is orbitally stable.

Proof. In the case of symmetry the argument above should be modified. In a neighborhood
of the manifold M∗ we decompose u as

u = Tg u∗ + ξ , hζa , ξi = 0 , (15.14)
144 Lectures on Applied PDEs, November 3, 2017

u
xi

zeta
M

chi

Figure 17: Manifold decomposition

for some symmetry transformation g. (This is a nonlinear, orthogonal decomposition


with respect to the manifold M∗ , see Figure 17.) Note that a in (15.14) is determined
by u uniquely. This follows from the implicit function theorem applied to the function
F (g, u) := hu − Tg u∗ , Au∗ i.
We introduce a neighbourhood of M∗ :

Uδ = {u ∈ X : ∃g s.t. ku − Tg u∗ kX ≤ δ}. (15.15)

Proposition 15.4. There is δ > 0 s.t. any u ∈ Uδ can be decomposed as follows

u = Tg u∗ + ξ, with g = g(u) and ξ ⊥ TTg u∗ M∗ . (15.16)

The latter condition can be written as

hξ, Au∗ i = 0, ∀A ∈ g. (15.17)

Proof. Assume for simplicity that G is a one-parameter (one-dimensional) group, parametrized


by a ∈ R and write Ta for Tg(a) . Let A be the corresponding generator. We define the
function G(a, u) = hu − Ta u∗ , ATa u∗ i and rewrite the statement of the proposition as
solving the equation
G(a, u) = 0 (15.18)
for a as a function of u. To do this, we use the Implicit Function Theorem. Clearly,
G is C 1 in u because is it a linear functional in u, and C 1 in 1
a since Ta u∗ is C in a.
Furthermore, G(a, Ta u∗ ) = 0. Finally, we show that ∂a G(a, u) u=Ta u∗ 6= 0:

∂a G(a, u) = −h∂a Ta u∗ , ATa u∗ i + hu − Ta u∗ , ∂a ATa u∗ i. (15.19)

Hence, using that ∂a Ta u∗ = ATa u∗ , we have, at u = Ta u∗ ,

∂a G(a, u)|u=Ta u∗ = −kATa u∗ k2 = −kAu∗ k2 . (15.20)

By the definition of G∗ , we have Au∗ 6= 0. Therefore, the conditions of the Implicit


Function Theorem are satisfied and it implies that (15.18) has a unique solution for a as
a function of u ∈ Uδ , which gives the statement of the proposition.
Lectures on Applied PDEs, November 3, 2017 145

The importance of this proposition lies in the fact that for ξ ⊥ Tu∗ M∗ , we have the
estimate (15.13).
Proceeding as in the proof of Theorem 15.1 and using (15.13) instead of (15.1) and
using that E(Tg u∗ ) = E(u∗ ), we find as in (15.2),
2
ku − Tg u∗ k ≤ (E(u) − E(u∗ )), (15.21)
θ
where g = g(u) is the same as in (15.16). However, as little contemplation shows,
ku − Tg u∗ k = dist(u, M∗ ), which gives the estimate
2
dist(u, M∗ ) ≤ (E(u) − E(u∗ )). (15.22)
θ
Continuing as in the proof of Theorem 15.1, we obtain the conclusion of our theorem.
We apply the analysis above on the problem of stability of kinks in the Allen - Cahn
equation. This will demonstrate the concepts introduced above.

15.3 Orbital stability of kink solutions of the Allen - Cahn equa-


tion
Consider the question of stability of the kink solutions of the Allen - Cahn equation (see
(2.2))

∂u
= ∆u − g(u), (15.23)
∂t
where g(u) = u3 − u, or, more generally, g : R → R is the derivative, g = G0 , of a double-
well potential: G(u) ≥ 0 and has two non-degenerate global minima with the minimum
value 0. For g(u) = u3 − u, we take G(u) = 21 (u2 − 1)2 . G(u) is also called a bistable
potential, see Figure 18.

u
−1 1

Figure 18: Double well potential

To keep the notation simple, assume we are in the dimension 1, i.e. u : R × R+ → R.


Recall from Section 2.1 that (15.23) has the kink static solution by χ(x), see Figure 19;
146 Lectures on Applied PDEs, November 3, 2017

G u

a b x
a

Figure 19: Hill-valley-hill structure for −G, and the kink for u.

and Eq. (15.23) is translation invariant and therefore the functions χa (x) := χ(x + a)
∀a ∈ R are also static solutions to (15.23). Thus we have an entire manifold of stationary
solutions Mkink = {χa : a ∈ R}. (In the multidimensional case, we have also rotations.)
We now use the results above to derive orbital stability of the kinks. We have

Theorem 15.5. The kink solutions of (15.23) are orbitally stable w.r.to H 1 (R)−perturbations.

Proof. It is not difficult to see that (15.23) is an L2 −gradient system (in the sense of the
definition given in Subsection 14) with the energy functional (cf. (14.7) or (19.4))

1
Z
E(u) = ( |∇u|2 + G(u)) , (15.24)
2

defined on χ + H 1 (R). Consequently, we have the properties (i) and (ii) of Subsection
15.1.
Because of the translational symmetry mentioned above, we use Theorem 15.3. Ac-
cording to this theorem, to prove the theorem, it suffices to show that the energy (15.24)
satisfies Conditions (a) and (b’) on the space H 1 (R). Verifying Conditions (a) is straight-
forward. Conditions (b’) requires more work.
Let L be the hessian of the energy functions E(u) at χ: L := E 00 (χ). If we denote the
negative of the r.h.s. of (15.23) by F (u),

F (u) := −∆u + g(u). (15.25)

then the hessian L := E 00 (χ) of the energy functions E(u) at χ is given by L := dF (χ)
and is computed explicitly as
L := −∆ + g 0 (χ). (15.26)
Below we will use the following lower bound on the operator L:
Proposition 15.6. Say g 0 (±1) > 0. Let ρ be the second lowest eigenvalue of L. We have

1
hξ, Lξi ≥ ρkξk2H 1 , ∀ξ⊥χ0 . (15.27)
2
Lectures on Applied PDEs, November 3, 2017 147

Proof. By Theorem 8.3, we have that

hLξ, ξi ≥ ρkξk2L2 (15.28)

for all ξ⊥χ0 , where ρ is the second lowest eigenvalue of L. We combine this estimate
together with δ > 0 times the elementary estimate

hLξ, ξi ≥ k∇ξk2L2 − Ckξk2L2 ,

for C = − min−1≤u≤1 g 0 (u), to obtain

hLξ, ξi ≥ δk∇ξk2L2 + (ρ − δC)kξk2L2 ,

and optimizing over δ (say taking δ = ρ/2C), completes the proof.

This proposition implies Condition (b’) for the energy (15.24). Since satisfies Condi-
tion (a) (E(u) is is C 3 ), the statement of the theorem follows.

16 Harmonic, Schrödinger and wave maps maps


16.1 Harmonic maps
In the simplest context the harmonic map or the stationary sigma model is described a
map, φ, from a d−dimensional euclidean space, Rd to the space Rn+1 , satisfying |φ| = 1,
i.e. φ maps Rd to S n embedded in Rn+1 in the standard way, and with the energy given
by
1
Z X
E(φ) = ∂i φ · ∂i φdd x. (16.1)
2 Rd i

(a · b denotes the dot product in Rn+1 .)


Critical points of (16.1) satisfy the Euler-Lagrange (stationary) equations

φ ∧ ∆φ = 0. (16.2)

To derive this equation, we observe that the variations of φ are along maps Rd → the
tangent space Tφ S 3 = φ⊥ and that the orthogonal projection, Pφ⊥ , onto Tφ S 3 = φ⊥ is
given by Pφ⊥ ξ = φ ∧ ξ, which leads to (16.2). (To obtain this expression we use that any
ξ ∈ φ⊥ can be written as ξ = Pφ⊥ η, where η is an arbitrary vector.)
148 Lectures on Applied PDEs, November 3, 2017

16.1.1 Symmetries
Gauge transformation: for any sufficiently regular function R ∈ O(n + 1),
Tρgauge : φ(x) 7→ R−1 φ(x); (16.3)
Translation transformation: for any h ∈ Rd ,
Thtransl : φ(x) 7→ φ(x + h); (16.4)
Rotation and reflection transformation: for any R ∈ O(d) (including the reflections
f (x) → f (−x))
TRrot : φ(x) 7→ φ(Rx). (16.5)
Scaling symmetry:
φ(x) → φ(λx).

16.1.2 Topological invariants


For a map, φ, to have finite energy, it should converge to a constant at infinity. In this
case, φ can be extended to a continuous map from S d to S n taking the point at infinity
to the limit of φ(x) at the spatial infinity. (In other words, we pass to the one-point
compactification of Rd .) Then one can define the degree, deg φ, as the homotopy class of
φ as a map from S d to S n , i.e. a member of the homotopy group πd (S n ).
This degree is conserved under the dynamics generated by the (rescaled) Landau-
Lifshitz (Schrödinger map) equation (16.15).
In the particle physics one takes d = 2, 3 and n = 2, and in the condensed matter
physics one takes d = n = 2, i.e. φ : Rd → S 2 , d = 2, 3. In both cases, the degree of φ is
an integer (the degree for maps from S d to S 2 ), i.e. πd (S 2 ) = Z, for d = 2, 3.
Let d = n = 2. Then the degree is given by
1 1
Z Z
2
degφ = φ · (∂1 φ ∧ ∂2 φ)d x (= φ∗ (dS)). (16.6)
8π R2 4π S 2
By Kronecker’s integral formula that deg φ is an integer and is equal to the Brouwer
degree of m. In the present case, degm is called the skyrmion number.

16.1.3 Complex representation


For computations as well as for theoretical studies it is convenient to pass to the projective,
or complex, representation of the harmonic map functional and equation. To this end, we
perform the stereographic projection, φ → w, of S 2 to the complex plane C, as

2 Re w −2Imw 1 − |w|2 
φ= , , . (16.7)
1 + |w|2 1 + |w|2 1 + |w|2
Lectures on Applied PDEs, November 3, 2017 149

Here we identified R2 with the complex plane C, by z := x1 + ix2 . In the new variables
the energy is given by
Z ¯ 2
|∂w|2 + |∂w|
E(w) := dz, (16.8)
(1 + |w|2 )2

where we used the complex derivatives ∂ := ∂1 − i∂2 and ∂¯ := ∂1 + i∂2 , while the degree
is given by

1 ∂w∂¯w̄ − ∂w∂
¯ w̄
Z
deg φ ≡ deg w = dz. (16.9)
4π (1 + |w|2 )2

The Euler - Lagrange equation, dw̄ E(w) = 0, where dw̄ := 21 (dw1 + idw2 ), for w = w1 + iw2 ,
is given by
2w ¯ 2 ).
∆w = (|∂w|2 + |∂w| (16.10)
1 + |w|2

16.1.4 Bogomolny bounds


Let d = n = 2. A key point here is that one has the Bogomolnyi-type inequality

E(φ) ≥ 4π|degφ|.

Indeed, we have the Bogomolnyi-type identity


1
Z X
E(φ) = ±4πdegφ + |∂i φ ± ij φ ∧ ∂j φ|2 d2 x,
4 R2 i

where ij is the Levi-Cevita antisymmetric symbol with 12 = −21 = 1, 11 = 22 =
0, which implies the inequality above. Moreover, the last relation yields that in every
homotopy class solutions of the self-dual/ anti-dual equations,

∂i φ ± ij φ ∧ ∂j φ = 0 (16.11)

are the minimizers of E(φ). (φ∧ in the second term gives a complex structure, see Murray.)
In the complex representation, the Bogomolnyi identity becomes

|∂ # w|2
Z
E(w) = ±2π deg φ + dz, (16.12)
(1 + |w|2 )2

where ∂ # stands for either ∂¯ or ∂. Eqs (16.12) implies that E(w) is minimized by w
¯ = 0 or ∂w = 0 (the Cauchy - Riemann equations), i.e. w is either
satisfying either ∂w
a holomorphic or anti-holomorphic function. They are mapped into each other by the
complex conjugation.
150 Lectures on Applied PDEs, November 3, 2017

16.1.5 Solutions
Ground states vs excitations: Usually, the ground states are stationary solutions corre-
sponding to energy minimizers. In translationally invariant problems like ours they have
infinite energy and are either homogeneous (x−independent) or lattice gauge-periodic
solutions, while finite energy localized solutions are interpreted as excitations.
We have the following solutions:

(a) Ferromagnetic states: φfm = constant.

(b) Skyrmions: These are solutions with F (m) < ∞ and deg =1. They can be classified
by Tρrot φ = TRgauge
ρ
φ (equivariance) for every ρ ∈ O(2) and some group homomor-
phisms Rρ : O(2) → O(2). The existence is given in Theorem 16.1 below.

(c) Skyrmion lattices: These are solutions periodic w.r.to a lattice L. They can be
classified by Tstrans φ = TRgauge
s
φ, for every s ∈ L and for some group homomorphisms
Rs : L → O(2).

Since the gauge symmetry here is global, the maps Rρ : O(2) → O(2) and Rs : L → O(2)
are group homomorphisms, not co-cycles.

Topological invariants for skyrmions and skyrmion lattices. Denote by ΩL ' T2


a fundamental cell of L. One has the following results

(a) Skyrmions: the topological invariant is given by the degree (16.6);

(b) Skyrmion lattices: the topological invariant is given by


1
Z
c1 (φ) = φ · (∂1 φ ∧ ∂2 φ)d2 x.
8π T2

Here c1 (φ) is the first Chern number of the map φ : T2 → S 2 .

Self-dual and anti-self-dual solutions. For any degφ = k, minimizers of satisfy self-
dual/ anti-dual equations (16.11), which are the first order equations. These equations
have explicit solutions (harmonic or anti-harmonic maps), φstat
k :

Theorem 16.1. For any degφ = k, minimizers of E(φ) are given by

φstat
k (ρ, α) = (Uk (ρ), kα),

where (ρ, α) are the polar coordinates in R2 and (θ, ϕ) are the spherical coordinates in S 2
and
Uk (ρ) = 2arctan ρk .
Lectures on Applied PDEs, November 3, 2017 151

Proof. The solutions above can be found by using the projective representation. As we
found above, in this representation the minimizers of the energy are either holomorphic
or anti-holomorphic functions. They are mapped into each other by the complex con-
jugation. We consider holomorphic solutions. One can show that they are of the form
w(x) = Q(z)
P (z)
, where P (z) and Q(z) are polynomials ?? with no common factors. The
degree of denominator = deg φ (the harmonic maps of degree n.).

16.1.6 Projective representation


(to do)

16.1.7 Generalization
More generally, the harmonic map is a map from a d−dimensional euclidean space, Rd ,
with the euclidean metric η = (δij ), to a Riemannian manifold, N , with a metric (gab ),
which is a critical point of the energy functional, given by
1
Z
E(φ) := h∂j φ, ∂ j φi. (16.13)
2 Rd

Here h∂j φ, ∂ j φi is the Riemann scalar product in N , which, in local coordinates, is


gab ∂j φa ∂ j φb , as usual, ∂j = ∂x∂ j and ∂ i := η ij ∂j , and we assume the summation over
repeated indices i, j = 1, . . . , d, a, b, c = 1, . . . , dim(N ), i, j = 1, . . . , d. With this nota-
tion, we can write energy the energy functional as,
1
Z
E(φ) := gab ∂i φa ∂ i φb .
2 Rd

The Euler-Lagrange equation for critical points of E(φ) is

∆φa + Γabc (φ)∂i φb ∂ i φc = 0, (16.14)

where ∆ = ∂j ∂ j and Γabc (φ) is the Christoffel symbols on N .


In the most important case N = G/H, where G is a compact Lie group and H is its
subgroup, specifically, G = SO(n + 1) and H = SO(n), so that N = S n .
If we consider S n to be embedded in Rn+1 in the standard way, then φ can be thought
of as a map from Rd+1 to Rn+1 , satisfying |φ| = 1.

16.2 Schrödinger maps


The Schrödinger map ((rescaled) Landau-Lifshitz) equation

∂t φ = −φ ∧ ∆φ. (16.15)
152 Lectures on Applied PDEs, November 3, 2017

Stationary solutions to this equations are exactly the harmonic maps whose equation is
given in (16.2).
The energy for (16.15) is given by (16.1). It is conserverd:
Z
∂t F (φ) = (−∆φ) · ∂t φ = 0.

By multiplying (16.15) externally by φ and using the identity a∧(b∧c) = (a·c)b−(a·b)c,


we obtain its equivalent form

φ ∧ ∂t φ = −∆φ − |∇φ|2 φ. (16.16)

The map φ∧ : ξ → φ ∧ ξ gives a complex structure on Tφ S 2 , see Murray. In particular,


(φ∧ )2 = −1.
This can be generalized to arbitrary surfaces by replacing φ∧ by νφ∧ , where νφ is the
outward normal at the point φ.
The Schrödinger map equation in complex and projective representations (to do)
Hamiltonian structure (to do)

16.3 Wave maps


The sigma model, or wave map equation, plays an important role in particle and condensed
matter physics. In the former, it is used as a model for strong interactions and in the
latter, in the theory of two-dimensional magnetism. In both cases, the model describes the
low-energy excitations around broken symmetry ground state. Since uniform changes in
the (classical) vacuum space costs nothing, the model involves only space-time derivatives
of the excitation fields.
The sigma model is a classical field theory with a field, φ, which is a map from
a (d + 1)−dimensional Minkowski space-time, Md+1 , with the Minkowski metric η =
diag(1, −1, . . . , −1), S n embedded in Rn+1 in the standard way, i.e. φ maps Md+1 to the
space Rn+1 and satisfies |φ| = 1, and with the action given by

1
Z X
S(φ) = ∂µ φ · ∂ µ φdd+1 x. (16.17)
2 Md+1 µ

where a · b denotes the dot product in Rn+1 , x = (x0 , x1 , . . . , xd ) and, as usual, ∂µ = ∂x∂ µ
and ∂ µ := η µν ∂ν , and we assume the summation over repeated indices µ = 1, . . . , d + 1.
Critical points of S(φ) satisfy the Euler-Lagrange equation

φ ∧ ∂µ ∂ µ φ = 0. (16.18)

The system (16.18) is Hamiltonian, and in particular has conserved energy.


Lectures on Applied PDEs, November 3, 2017 153

By multiplying (16.18) externally by φ and using the identity a∧(b∧c) = (a·c)b−(a·b)c,


we obtain its equivalent form

∂t2 φ = −∆φ − |∇φ|2 φ + |∂t φ|2 φ. (16.19)

Action (16.17) has the conserved current


1
j := φ · (∂α φ ∧ ∂β φ)dxα ∧ dxβ , (16.20)

1 µαβ
(or in the co-ordinate form, j µ := 8π  φ · (∂α φ ∧ ∂β φ)). (what is the symmetry it
corresponds to?) Indeed, one can easily check that dj = 0 (or ∂µ j µ = 0). Since j is
conserved, we can write
j = da, (16.21)
(j µ = µαβ ∂α aβ ) for some one-form (gauge field) a. deg φ given in (16.6) can be written
in terms of the conserved current as the charge
1
Z
degφ = d2 xj 0 , (16.22)
2π R2
where
1
j 0 := φ · (∂1 φ ∧ ∂2 φ)dx1 ∧ dx2 . (16.23)

The curvature of the gauge field a is Fµν := ∂µ aν − ∂ν aµ and the electric and magnetic
fields are B = F12 and Ei = F0i . Writing j µ = (ρ, ~j) we see that each field caries the
charge Z
q(φ) := ρ(x) ≡ deg φ (16.24)
R2
0 1
(it is independent of x ∈ S ) and the magnetic flux
Z
Φ(φ) := B(x) (16.25)
R2

Due to the definitions B = F12 and Ei = F0i , (16.21) implies that ρ = B and j i = ij Ej
and therefore q = Φ. Since q(φ) ≡ deg φ, we conclude that the charge, q, and the
magnetic flux, φ, are quantized. Hence we have an elementary point charge and this
point unit charge caries a point unit magnetic flux. That is the particles are charge-flux
tube composites with respect to the fictitious field a.
In the case of d = 3, deg φ is given by the Hopf invariant, which is the linking number
of pre-images of two distinct point in S 2 . There is also explicit formula for deg φ using
the conserved current
1 1
Z Z
3
deg φ = µ
d xaµ j ≡ d3 xa ∧ j. (16.26)
4π R3 4π R3
154 Lectures on Applied PDEs, November 3, 2017

Note that the integral on the r.h.s. is gauge invariant (aµ → aµ + ∂µ χ) and is independent
of the choice of a particular gauge field a.
We consider (16.26) as a constraint for the action (16.17). Then the Lagrange multi-
plier theory leads to the action
1 θ
Z
Sθ (φ) := [h∂µ φ, ∂ µ φi + a(φ) ∧ j(φ)], (16.27)
2 M2+1 π
where a(φ) is found by integrating (16.21). Since a(φ) is non-local, then so is Sθ (φ). To
get rid of non-locality, we think of the integral on the r.h.s. of (16.26) as an interaction
action for the fields φ and a and add to it an action of the gauge field a,
1 1
Z Z
3 αβγ
Sgauge (a) := − d x aα ∂β aγ ≡ d3 xa ∧ da (16.28)
2π M2+1 4π M2+1
(the Chern-Simmons action),
R chosen so as to produce (16.21): the Euler - Lagrange
equation for d3 xaµ j µ − 12 M2+1 d3 xαβγ aα ∂β aγ in a is (16.21). Now, considering φ and
R

a as independent fields, the total action is


1 θ 1
Z
S(φ, a) := [h∂µ φ, ∂ µ φi + (a ∧ j(Φ) − a ∧ da)], (16.29)
2 M2+1 π 2
where j(φ) is given by (16.20). This is a local equation, whose critical point equations in
φ and a are (16.18) and (16.21), respectively.

∗∗∗∗∗∗∗
d3 xaµ j µ − µ2 d3 xαβγ aα ∂β aγ is
R R
The field equation for q
qj µ = µµαβ ∂α aγ , (16.30)
(cf. (16.21)). Integrating the µ = 0 component of (16.30), one finds that a point charge of
magnitude q caries point flux −q/µ. That is the particles are charge-flux tube composites
with respect to the fictitious field a.
In d = 2 one can extend the action (16.17) by adding the generalized Wess-Zumino-
Novikov-Witten term
1
Z
SgW ZN W := hab (Φ)αβ ∂α Φa ∂β Φb .
2

Problems:
1) Consider dynamics of several solitons carrying charge and magnetic flux.
2) Quantization of sigma model and anyon statistics? (See Dunne’s lectures. Poor
man quantization: replacing solitons by particles.)
Hamiltonian structure for the wave maps and conservation of energy.
Connection to fractional (anyon) statistics.
Lectures on Applied PDEs, November 3, 2017 155

16.3.1 Generalization
More generally, the wave map, or the sigma model, is a map from a (d + 1)−dimensional
Minkowski space, Md+1 , with the Minkowski metric, to a Riemannian manifold, N , with
a metric (gab ), and the action functional, given as

1
Z
S(Φ) := h∂µ Φ, ∂ µ Φi. (16.31)
2 Md+1

Here h∂µ Φ, ∂ µ Φi is the Riemann scalar product in N , which, in local coordinates, is


gab ∂µ Φa ∂ µ Φb , as usual, ∂µ = ∂x∂ µ and ∂ µ := η µν ∂ν , and we assume the summation over
repeated indices µ = 1, . . . , d + 1, a, b, c = 1, . . . , dim(N ), i, j = 1, . . . , d. Critical points
of S(Φ) satisfy the Euler-Lagrange equation

∂µ ∂ µ Φa + Γabc (Φ)∂µ Φb ∂ µ Φc = 0, (16.32)

where Γabc (Φ) is the Christoffel symbols on N . Solutions of this system of nonlinear PDEs,
i.e. critical point of the action functional (16.31), are called wave maps.
The system (16.32) is Hamiltonian, and in particular has conserved energy,

1
Z
E(Φ) := gab ∂i Φa ∂ i Φb .
2 Rd

17 Mean curvature flow


17.1 Definition and general properties
The mean curvature flow (starting with a hypersurface S0 in Rn+1 ) is the family of hy-
persurfaces S(t) given by local parametrizations x(·, t) : U → Rn+1 , where U is an open
set in Rn (or immersions x(·, t)), which satisfy the evolution equation
 ∂x
= −H(x)ν(x)
∂t (17.1)
x|t=0 = x0

where x0 parametrizes (or is an immersion of) S0 , H(x) and ν(x) are mean curvature
and the outward unit normal vector at x ∈ S(t), respectively. The terms used above are
explained in Appendix ??, for more details and extensions, see [16].
In this lecture we describe some general properties of the mean curvature flow, (17.1).
We begin with writing out (17.1) for various explicit representations for surfaces St .

Mean curvature flow (MCF) for level sets and graphs. We rewrite out (17.1) for
the level set and graph representation of S. Below, all differential operations, e.g. ∇, ∆,
are defined in the corresponding Euclidian space (either Rn+1 or Rn ).
156 Lectures on Applied PDEs, November 3, 2017

1) Level set representation S = {ϕ(x, t) = 0}. Then, by Proposition 6.11 of Appendix


??, we have
∇ϕ ∇ϕ
 
ν(x) = , H(x) = div . (17.2)
|∇ϕ| |∇ϕ|
 
∇ϕ ∇ϕ
We compute 0 = dϕ dt
= ∇ x ϕ · ∂x
∂t
+ ∂ϕ
∂t
and therefore ∂ϕ
∂t
= ∇ϕ · |∇ϕ|
div |∇ϕ|
, which
gives
∇ϕ
 
∂ϕ
= |∇ϕ| div . (17.3)
∂t |∇ϕ|

2) Graph representation: S = graph of f . In this case S is the zero level set of the
function ϕ(x) = xn+1 − f (u), where u = (x1 , . . . , xn ) and x = (u, xn+1 ), and using
(17.3) with this function, we obtain
!
∂f p ∇f
= |∇f |2 + 1 div p . (17.4)
∂t |∇f |2 + 1

∂2f

Denote by Hess f the standard euclidean hessian, Hess f := ∂ui ∂uj
. Then we can
rewrite (17.4) as
∂f ∇f Hess f ∇f
= ∆f − . (17.5)
∂t |∇f |2 + 1

Exercise 17.1. Using (17.2), find the graph representation of the mean curvature,
!
∇f
H(x) = div p . (17.6)
1 + |∇f |2

Different form of the mean curvature flow. Multiplying the equation (17.1) in
Rn+1 by ν(x), we obtain the equation
∂x
ν(x) · = −H(x). (17.7)
∂t
In opposite direction we have
Proposition 17.1. If x satisfies (17.7), then there is a (time-dependent) reparametriza-
tion ϕ of S, s.t. x ◦ ϕ satisfies (17.1).
Proof. Denote ( ∂x
∂t
)T := ∂x
∂t
− (ν · ∂x
∂t
)ν (the projection of ∂x
∂t
onto Tx S) and let ϕ satisfy the
−1 ∂x T ∂ ∂
ODE ϕ̇ = −(dx) ( ∂t ◦ ϕ) (parametrized by u ∈ U ). Then ∂t (x ◦ ϕ) = ( ∂t x) ◦ ϕ + dx ϕ̇.
Substituting ϕ̇ = −(dx)−1 ( ∂x∂t
◦ ϕ) T
into this, we obtain ∂
∂t
(x ◦ ϕ) = (ν · ( ∂
∂t
x) ◦ ϕ)ν =
−H.
Lectures on Applied PDEs, November 3, 2017 157

Thus the MCF in the form (17.7) is invariant under reparametrization, while the form
(17.1) is obtained by fixing a specific parametrization (fixing the gauge).
Now we describe some general properties of MCF. We summarize these properties as
follows
• (17.1) is invariant under rigid motions of the surface, i.e. ψ 7→ Rψ + a, where
R ∈ O(n + 1), a ∈ Rn+1 and ψ = ψ(u, t) is a parametrization of St , is a symmetry
of (17.1).
• (17.1) is invariant under the scaling, for any λ > 0,

S(t) → λS(λ−2 t) ⇔ x→ λx, t → λ−2 t. (17.8)

d
H 2 dσ ≤ 0.
R
• (17.1) is area shrinking. Actually, dt
V (ψ(t)) = − St

• Static solutions of (17.1) are minimal surfaces, i.e. critical points of the area func-
tional, in particular they satisfy the equation H = 0.
The first two properties come from the fact that the mean curvature is invariant under
translations, rotations and scaling. The invariance under rigid motions is obvious. The
remaining properties are proven below.
Proposition 17.2. (17.1) is invariant under the scaling x 7→ λx and t 7→ λ−2 t for any
λ > 0.
Proof. To prove the invariance under scaling, we first observe that, as follows from the
level set representation of the mean curvature, (17.2),

H(λx) = λ−1 H(x). (17.9)

Now, let τ := λ−2 t, and xλ (t) := λx(τ ). Then

∂t xλ = λλ−2 (∂τ x)(τ ) = −λ−1 H(x(τ ))ν(x(τ )).

On the other hand, ν λ ≡ ν(xλ (τ )) = ν ≡ ν(x(τ )) and H λ ≡ H(xλ (τ )) = λ−1 H(x(τ )) ≡


λ−1 H (by (17.9)), which gives

∂t xλ = −H λ ν λ . (17.10)

For S = graph f , f (u) → λf (λ−1 u) ⇔ ψ(u) → λψ (λ−1 u) , where ψ(u) := (u, f (u)).
Indeed,

ψ(u) := (u, f (u)) → u, λf (λ−1 u) = λ λ−1 u, f (λ−1 u) =: λψ λ−1 u .


  

Exercise 17.2. Check (17.9) using the graph representation of the mean curvature, (17.6).
158 Lectures on Applied PDEs, November 3, 2017

17.1.1 Special solutions


Static solutions. Static solutions satisfy the equation H(x) = 0 on S, which, as shown
below, is the equation for critical points of the volume functional V (ψ) (the Euler -
Lagrange equation for V (ψ)). Thus, static solutions of the MCF (17.1) are minimal
surfaces.

Spherically and axi - symmetric (equivariant) solutions:


a) Sphere. The n−dimensional sphere of the radius R(t) centred at the origin can
be given either by the immersion x(t) = R(t)x̂, where x̂ = x/kxk, or as the level
2 2
p {ϕ(x, t) = 0}, where nϕ(x, t) := n|x| − R(t) , or as graph f , where f (u, t) =
set
R(t)2 − |u|2 . (SR = RS , where S is the unit n-sphere.) Then we have

∇ϕ
 
n
H(x) = div = div(x̂) =
|∇ϕ| R
p
and therefore we get Ṙ = − Rn which implies R = R02 − 2nt. So this solution
shrinks to a point.

b) Cylinder. Define the n−dimensional cylinder of the radius R(t) with the axis
along the xn+1 −axis by the immersion x(t) = (R(t)x̂0 , x00 ), where x̂ = x0 /kx0 k
and x = (x0 , x00 ) ∈ Rn × R1 . Then H(x) = n−1
R
and Ṙ = − n−1R
which implies
p
2
R = R0 − 2(n − 1)t. (In the implicit function representation the cyliner is given
Pn−1 2  21
by {ϕ = 0}, where ϕ := r − R, with r = i=1 xi .)
Exercise 17.3. Show that the above expressions give solutions to the mean curvature flow
equation, (17.1).

Motion of torus (H. M. Soner and P. E. Souganidis). (to be done)

Volume functional. Recall that locally the surface volume functional can be written
as
√ n
Z
V (S ∩ U ) = gd u,
U
where g := det(gij ) for a metric gij . Recall, from Lemma 10.12 that for a surface S given
by a graph of a function f : U → R, the surface area is given by the formula
Z Z p
V (f ) ≡ 1= 1 + |∇f |2 dn u. (17.11)
S U

Proposition 17.3. • If a closed surface S evolves according to (17.1), then ∂t V (S) =


2
R
− S H < 0 and consequently, its area, V (S), decreases.
Lectures on Applied PDEs, November 3, 2017 159

• Critical points of V (S) are static solution to the MCF (17.1).

Proof. We prove the proposition for a surface S given by a graph of a function f : U → R,


the general case can be derived from this result. In this case, the surface area is given by
the formula (17.11). We compute, using that ξ|∂U = 0,

∇f
Z
dV (f )ξ = ∂s |s=0 V (f + sξ) = p · ∇ξdn u,
1 + |∇f |2
U

 
∇f
which, after integrating by parts and using the equation H(x) = div √ 2
(see
1+|∇f |
(17.6)), gives
Z
dV (f )ξ = Hξdn u. (17.12)
U

Using (17.12) we find

dn u
Z Z p
n
dV (f )ξ = Hξd u = 1 + |∇f |2 Hξ p .
U U 1 + |∇f |2

This proves the second statement. To prove the first one, we compute ∂t V (f ) = dV (f )∂t f ,
which due to the previous relation gives
Z Z p Z
∂t V (f ) = n
H∂t f d u = − 1 + |∇f |2 H d u = − H 2 ,
2 n
U U S

as stated in the first statement.

17.2 Self-similar surfaces


Recalling (2.51) - (2.54), or (17.8), we consider solutions of the MCF of the form S(t) ≡
S λ(t) := λ(t)S (standing waves), or x(u, t) = λ(t)y(u), where λ(t) > 0. Plugging this into
(17.1) and using H(λy) = λ−1 H(y), gives λ̇y = −λ−1 H(y)ν(y), or λλ̇y = −H(y)ν(y).
Multiplying this by ν(y), we obtain (cf. (2.54))

H(y) = ahν, yi, and λλ̇ = −a. (17.13)

p is independent of t, then so should be λλ̇ = −a. Solving the last equation, we


Since H(y)
find λ = λ20 − 2at.
λ20
i) a > 0 ⇒ λ → 0 as t → T := 2a
⇒ S λ is a shrinker.

ii) a < 0 ⇒ λ → ∞ as t → ∞ ⇒ S λ is an expander.


160 Lectures on Applied PDEs, November 3, 2017

Recall, that static solutions of the MCF (17.1) satisfy the equation H(x) = 0 on S and
therefore are minimal surfaces. Hence the minimal surfaces are special cases of self-similar
ones corresponding to a = 0. In fact, a = 0 separates two types of evolution: contracting
a > 0 ( λ decreasing) and expanding a < 0 ( λ increasing). (Remember that a = −λ∂t λ is
the negative of the speed of scaling λ.) (We see that scaling solitons generalize the notion
of the minimal surface.)
The equation (17.13) has the solutions: a is time-independent and x is one of the
following

a) Sphere x = Rx̂, where R = na .


p

b) Cylinder x = (Rx̂0 , x00 ), where x = (x0 , x00 ) ∈ Rn+1 × R1 , where R = na .


p

As stated in the following theorem, these solutions are robust.

Theorem 17.4 (Huisken). Let S satisfy H = ax · ν and H ≥ 0.pWe have


(i) If n ≥ 2, and S is compact, then S is a sphere of radius na .
q
n−1
(ii) If n = 2 and S is a surface of revolution, then S is the cylinder of radius a
.

Now we consider self-similar surfaces in the graph representation. Let

f (u, t) = λχ(λ−1 u), λ depends on t. (17.14)

Substituting this into (17.4) and setting y = λ−1 u and a = −λ̇λ, we find
q
1 + |∇y χ|2 H(χ) = a(y∂y − 1)χ. (17.15)

Translation solitons. These are solutions of the MCF of the form S(t) ≡ S + h(t)
(traveling waves), or x(u, t) = y(u) + h(t), where h(t) ∈ Rn+1 . Plugging this into (17.1)
and using H(y + h) = H(y), gives ḣ = −H(y)ν(y), or

H(y) = v · ν(y), and ḣ = v. (17.16)

Since H(y) is independent of t, then so should be ḣ = v. (more to come)

Breathers. MCF periodic in t.

Self-similar surfaces and rescaled MCF. Minimal surfaces are static solutions of the
MCF (17.1). Are more general the self-similar surfaces static solutions of some equation?
The answer is yes, the self-similar surfaces are static solutions of the rescaled MCF,

∂τ ϕ = −(H(ϕ) − aϕ · ν(ϕ))ν(ϕ), (17.17)


Lectures on Applied PDEs, November 3, 2017 161

where a = −λ̇λ, which is obtained by rescaling the surface ψ and time t as


Z t
−1
ϕ(u, τ ) := λ (t)ψ(u, t), τ = λ−2 (s)ds, (17.18)
0

and then reparametrizing the obtained surface S resc (τ ) := λ(t)S(t) as in the proof of
Proposition 17.1. Indeed, since λH = H λ (or λH(ψ) = H(λ−1 ψ)), we find

∂τ ϕ = λ2 ∂t ϕ = λ2 (−λ̇λ−2 ψ + λ−1 ∂t ψ)

= −λ̇λϕ − λH(ψ)ν(ψ) = aϕ − H(ϕ)ν(ϕ).


Then the mean curvature flow equation (17.1) implies

∂τ ϕ = −H(ϕ)ν(ϕ) + aϕ, (17.19)

which after the reparametrization gives (17.17). This is another analogy with minimal
surfaces.
For static solutions, a =const. Then solving the equation λ̇λ = −a, we obtain the
parabolic scaling:
p 1
λ= 2a(T − t) and τ (t) = − ln(T − t), (17.20)
2a
where T := λ20 /2T , which was already discussed in connection with the scaling solitons.
Now, we know that minimal surfaces are critical points of the volume functional V (ψ)
(by (??), the equation H(x) = 0 on S is the Euler - Lagrange equation for V (ψ)). Are
self-similar surfaces critical points of some modification of the volume functional? (Recall
that, because of reparametrization (see Proposition 17.1), it suffices to look only at normal
variations, ψs , of the immersion ψ, i.e. generated by vector fields η, directed along the
normal ν: η = f ν.) The answer to this question is yes and is given in the following
a 2
Proposition 17.5. Let ρ(x) = e− 2 |x| and Va (ϕ) := S λ ρ. For a surface S given locally
R

by an immersion ψ and normal variations, η = f ν, we have


Z
dVa (ϕ)η = (H − aϕ · ν)ν · η ρdn u, (17.21)
U
p a 2
Proof. The definition of Va (ϕ) gives Va (ϕ) = S λ ρ(ϕ) g(ϕ), where ρ(ϕ) = e− 2 |ϕ| . We
R
√ √
have dρ(ϕ)η = −aρ(ϕ)η. Using this formula and the equation d gη = H gν · η proven
above (see (??)) and the fact that we are dealing with normal variations, η = f ν, we
obtain (17.21).

The equation (17.17) and Proposition 17.5 imply


162 Lectures on Applied PDEs, November 3, 2017

Corollary 4. Assume a in (17.17) is constant (i.e. the rescaling (17.18) is parabolic,


(17.20)). Then the modified area functional Va (ϕ) is momotonically decaying under the
rescaled flow (17.17), more precisely,
Z
∂τ Va (ϕ) = − ρ|H − aν · ϕ|2 . (17.22)

The relation (17.22) is the Huisken monotonicity formula (earlier results of this type
were obtained by Giga and Kohn and by Struwe).

Most interesting minimal surfaces are not just critical points of the volume functional
V (ψ) but are minimizers for it. What about self-similar surfaces? We see that
(i) inf Va (ϕ) = 0 and, for compact minimal surfaces, Va (ϕ) is minimized by any sequence
shrinking to a point.
(ii) Va (ϕ) is unbounded from above. This is clear for a < 0. To see this for a < 0, we
construct a sequence of surfaces lying inside a fixed ball in Rn+1 and folding tighter
and tighter.
Thus self-similar surfaces are neither minimizers nor maximizers of Va (ϕ). We conjecture
that they are saddle points satisfying min-max principle: supV inf ϕ:V (ϕ)=V Va (ϕ). One can
try use this principle (say in the form of the mountain pass lemma) to find solutions of
(17.13).
Remark 17.6. For minimal surfaces (strict) stability implies the the surface is a (strict)
local minimizer of the volume functional V (ψ). As is already suggested by the discussion
above, this is not so for self-similar surfaces, they are saddle points possibly satisfying
some min-max principle.

17.3 Linearized stability of self-similar surfaces


Normal hessians. Recall that the hessian of a functional E(ϕ) is defined as Hess E(ϕ) :=
d grad Va (ϕ). Note that unlike the Gâteaux derivative, d, the gradient grad and therefore
the hessian, Hess, depends on the Riemann metric on the space on which E(ϕ) is defined.
Since the tangential variations lead to reparametrization of the surface, in what follows
we are dealing with normal variations, η = f ν. (In physics terms, specifying normal vari-
ations is called fixing the gauge.) We use the following notation for a linear operator, A,
on normal vector fields on S: AN f = A(f ν). For instance, HessN E(ϕ)f = Hess E(ϕ)(f ν)
and
dN F (ϕ)f = dF (ϕ)(f ν). (17.23)
We consider the hessian of the modified volume functional Va (ϕ), at a self-similar ϕ
(i.e. H(ϕ) = aϕ · ν) and in the normal direction (i.e. for normal variations, η = f ν) in
Lectures on Applied PDEs, November 3, 2017 163

R
the Riemann metric h(ξ, η) := S λ ξηρ. In what follows, we call this hessian the normal
hessian and denote it by HessN Va (ϕ).
Before we proceed, we mention the following important property of Va (ϕ): the equation
H(ϕ) − aϕ · ν(ϕ) = 0 breaks the scaling and translational symmetry. Indeed, using the
relations

H(λϕ) = λ−1 H(ϕ), ν(λϕ) = ν(ϕ), ∀λ ∈ R+ , (17.24)


H(ϕ + h) = H(ϕ), ν(ϕ + h) = ν(ϕ), ∀h ∈ Rn+1 , (17.25)
H(gϕ) = H(ϕ), ν(gϕ) = gν(ϕ), ∀g ∈ O(n + 1), (17.26)

using that g ∈ O(n + 1) are isometries in Rn+1 and using the notation Ha (ϕ) := H(ϕ) −
aϕ · ν(ϕ), we obtain

Hλ−2 a (λϕ) = λ−1 Ha (ϕ), ∀λ ∈ R+ , (17.27)


Ha (ϕ + h) + ah · ν(ϕ) = Ha (ϕ), ∀h ∈ Rn+1 , (17.28)
Ha (gϕ) = Ha (ϕ), ∀g ∈ O(n + 1). (17.29)

We want to address the spectrum of the normal hessian, Hess⊥ Va (ϕ). First, we note
that the tangential variations lead to zero modes of the full hessian, Hess Va (ϕ). Indeed,
we have
Proposition 17.7. The full hessian, Hess Va (ϕ), of the modified volume functional Va (ϕ),
has the eigenvalue 0 with the eigenfunctions which are tangential vector fields on S
Proof. We consider a family αs of diffeormorphisms of U , with α0 = 1 and ∂s ϕ◦αs |s=0 = ξ,
a tangential vector field, reparametrizing the immersion ϕ, and define the family ϕ ◦ αs
of variations of ϕ. Then ϕ ◦ αs satisfies again the soliton equation, Ha (ϕ ◦ αs ) = 0.
Differentiating the latter equation w.r.to s at s = 0 and using that ∂s ϕ ◦ αs |s=0 = ξ, is a
tangential vector field, we obtain

dHa (ϕ)ξ = 0, (17.30)

which proves the proposition.


Theorem 17.8. The hessian, HessN Va (ϕ), of the modified volume functional Va (ϕ), at a
self-similar ϕ (i.e. H(ϕ) = aϕ·ν) and in the normal direction (i.e. for normal variations,
η = f ν), has
• the eigenvalue −2a with the eigenfunction ϕ · ν(ϕ),

• the eigenvalue −a with the eigenfunctions ν j (ϕ), j = 1, . . . , n + 1, and,

• the eigenvalue 0 with the eigenfunctions σj ϕ · ν(ϕ), j = 1, . . . , 21 n(n − 1), where σj


are generators of the Lie algebra of SO(n + 1), unless ϕ is a sphere.
164 Lectures on Applied PDEs, November 3, 2017

Proof. If an immersion ϕ satisfies the soliton equation H(ϕ) = aϕ · ν(ϕ), then by (17.27),
we have Hλ−2 a (λϕ) = 0 for any λ > 0. Differentiating this equation w.r.to λ at λ = 1, we
obtain dHa (ϕ)ϕ = −2aϕ · ν(ϕ).
Now, choosing ξ to be equal to the tangential projection, ϕT , of ϕ, and subtracting the
equation (17.30) from the last equation, we find dHa (ϕ)(ϕ · ν(ϕ))ν(ϕ) = −aϕ · ν(ϕ). Since
by the definition (17.23), dHa (ϕ)f ν(ϕ) = dN Ha (ϕ)f , this proves the first statement.
To prove the second statement, we observe that the soliton equation implies, by
(17.28), that Ha (ϕ + sh) + ash · ν(ϕ) = 0 and any constant vector field h. Differen-
tiating this equation w.r.to s at s = 0, we obtain dHa (ϕ)h = −ah · ν(ϕ). Now, choosing
ξ to be equal to the tangential projection, hT , of h, and subtracting the equation (17.30)
from the last equation, we find dHa (ϕ)(h · ν(ϕ))ν(ϕ) = −ah · ν(ϕ), which together with
(17.23) gives the second statement.
Finally, to prove the third statement, we differentiate the equation Ha (g(s)ϕ) = 0,
where g(s) is a one-parameter subgroup of O(n + 1), w.r.to s at s = 0, to obtain
dHa (ϕ)σϕ = 0, where σ denotes the generator of g(s). Now, choosing ξ in (17.30) to
be equal to the tangential projection, (σϕ)T , of σϕ, and subtracting the equation (17.30)
from the last equation, we find dHa (ϕ)(σϕ · ν(ϕ))ν(ϕ) = 0, which together with (17.23)
gives the third statement.
Remark 17.9. a) For a 6= 0, the soliton equation, ϕ · ν(ϕ) = a−1 H(ϕ), and Proposi-
tion 17.13 imply that the mean curvature H is an eigenfunction of HessN Va (ϕ) with the
eigenvalue −2a.
b) Strictly speaking, if the self-similar surface is not compact, then ϕ · ν(ϕ) and
ν (ϕ), j = 1, . . . , n + 1, generalized eigenfunctions of HessN Va (ϕ). In the second case,
j

the Schnol-Simon theorem (see Appendix ?? or [6]) implies that the points −2a and −a
belong to the essential spectrum of HessN Va (ϕ).
c) We show below that the normal hessian, HessN sph Va (ϕ), on the sphere of the radius
, given in (17.31), has no other eigenvalues below 2a
pa
n n
, besides −2a and −a. A similar
statement, but with n replaced by n − 1, we have for the cylinder.
We call the eigenfunction ϕ · ν(ϕ), ν j (ϕ) and σj ϕ · ν(ϕ), j = 1, . . . , n + 1, the scaling,
translational and rotational modes. They originate from the normal projections, (λϕ)N
and (sh)N , of scaling, translation and rotation variations.

Linearized stability. Given a self-similar surface ϕ, we consider for example the man-
ifold of surfaces obtained from ϕ by symmetry transformations,
Mϕ := {λgϕ + z : (λ, z, g) ∈ R+ × Rn+1 × SO(n + 1)}.
By the spectral theorem above, it has unstable and central manifolds corresponding to
the eigenvalues −2a, −a and 0. Hence, we can expect only the dynamical stability in the
transverse direction.
Lectures on Applied PDEs, November 3, 2017 165

Definition 17.10 (Linearized stability of self-similar surfaces). We say that a self-similar


surface φ, with a > 0, is linearly stable (for the lack of a better term) iff the normal hessian
satisfies HessN Va (ϕ) > 0 on the subspace span {ϕ · ν(ϕ), ν i (ϕ), i = 1, . . . , n + 1, σj ϕ ·
⊥ ⊥
ν(ϕ), j = 1, . . . , 21 d(d−1)} (i.e. on span {scaling, translational, rotational modes} ).
(I.e. the only unstable motions allowed are scaling, translations and rotations.)
In what follows a > 0. Theorem 17.8 implies that if φ is not spherically symmetric,
then 0 is an eigenvalue of HessN Va (ϕ) of multiplicity at least n + 1. This gives the first
statement in the following corollary, while
Corollary 17.11. If a self-similar surface with a > 0 satisfies HessN Va (ϕ) > 0 on the
subspace span {ϕ · ν(ϕ), ν j (ϕ), j = 1, . . . , n + 1}⊥ , then it a sphere or a cylinder.
Theorem 17.12. For a self-similar surface with a > 0, HessN Va (ϕ) ≥ −2a iff H(ϕ) > 0.
To prove this theorem we will use the Perron-Frobenius theory (see Appendix E.3)
and its extension as given in [?]. We begin with
Definition 17.13. We say that a linear operator on L2 (S) has a positivity improving
property iff either A, or e−A , or (A + µ)−1 , for some µ ∈ R, takes non-negative functions
into positive ones.
Proposition 17.14. The normal hessian, HessN Va (ϕ), has a positivity improving prop-
erty.
Proof. By standard elliptic/parabolic theory, e∆ and (−∆ + µ)−1 , for any µ > 0, has
strictly positive integra kernel and therefore is positivity improving. To lift this result to
HessN Va (ϕ) := −∆ − |W |2 − a1 + ϕ · ∇ we proceed exactly as in [?].
Since, the operator HessN Va (ϕ) is bounded from below and has a positivity improving
property, it satisfies the assumptions of the Perron-Frobenius theory (see Appendix E.3)
and its extension in [?] to the case when the positive solution in question is not an
eigenfunction. The latter theory, Proposition 17.13 and Remark 17.9 imply the statement
of Theorem 17.12.
Corollary 17.15. Let ϕ be a self-similar surface. We have
(a) For a < 0 (ϕ is an expander), H(ϕ) changes the sign.
(b) For a = 0, inf HessN Va (ϕ) < 0.
(c) For a > 0, if ϕ is an entire graph over Rn and is weakly stable, then ϕ is a
hyperplane.
Indeed, (a) follows directly from Theorems 17.8 and 17.12, while (b) follows from the
fact that for a = 0, 0 is an eigenvalue of the multiplicity n + 2 and therefore, by the
Perron-Frobenius theory, it is not the lowest eigenvalue of HessN Va (ϕ). (c) is based on
the fact that entire graphs over Rn cannot have strictly postive mean curvature. (check,
references)
166 Lectures on Applied PDEs, November 3, 2017

Hessians for spheres and cylinders. We finish this section with the discussion of the
normal hessians on the n−sphere and (n, k)−cylinder.

in Rn+1 , we have
pn
Explicit expressions. 1) For the n−sphere SRn of radius R = a

a
HessN
sph Va (ϕ) = − ∆Sn − 2a, (17.31)
n
on L2 (Sn ), where ∆Sk is the Laplace-Beltrami operator onqthe standard n−sphere Sk .
2) For the n−cylinder CRn = SRn−k × Rk of radius R = n−k a
in Rn+1 , we have

a
HessN
cyl Va (ϕ) = −∆y − ay · ∇y − ∆ n−k − 2a, (17.32)
n−k S

acting on L2 (C n ).

Spectra of Lsph a := HessN cyl N


sph Va (ϕ) and La := Hesscyl Va (ϕ). We describe the spectra
pa p a
of the normal hessians on the n−sphere and (n, k)−cylinder, of the radii n and n−1 ,
respectively.
It is a standard fact that the operator −∆ = −∆Sn is a self-adjoint operator on
2 n
L (S ). Its spectrum is well known (see [?, 16]):  it consists
  of the eigenvalues
 l(l + n −
n+l n+l−2
1), l = 0, 1, . . . , of the multiplicities m` = − . Moreover, the
n n
eigenfunctions corresponding to the eigenvalue l(l + n − 1) are the restrictions to the
sphere Sn of harmonic polynomials on Rn+1 of degree l and denoted by Ylm (the spherical
harmonics),

− ∆Ylm = l(l + n − 1)Ylm , l = 0, 1, 2, 3, . . . , m = 1, 2, . . . , ml . (17.33)

In particular, the first eigenvalue 0 has the only eigenfunction 1 and the second eigenvalue
n has the eigenfunctions ω 1 , · · · , ω n+1 .
Consequently, the operator Lsph a := HessN a
sph Va (ϕ) = − n (∆Sn +2n) is self-adjoint and its
spectrum consists of the eigenvalues na (l(l+n−1)−2n) = a(l−2)+ na l(l−1), l = 0, 1, . . . , of
the multiplicities m` . In particular, the first n+2 eigenvectors of Lspha (those with l = 0, 1)
correspond to the non-positive eigenvalues,

Lsph 0 0
a ω = −2aω , Lsph j j
a ω = −aω , j = 1, . . . , n + 1, (17.34)

and are due to the scaling (l = 0) and translation (l = 1) symmetries.


We proceed to the cylindrical hessian Lcyla := HessNcyl Va (ϕ), given in (17.32). The
variables in this operator separate and we can analyze the operators −∆y − ay · ∇ and
a a
− n−k (∆Sn−k + 2(n − k)) separately. The operator − n−k (∆Sn−k + 2n) was already analyzed
Lectures on Applied PDEs, November 3, 2017 167

above. The operator −∆y − ay · ∇ is the Ornstein - Uhlenbeck generator, which can be
unitarily mapped by the gauge transformation
a 2
v(y, w) → v(y, w)e− 4 |y|

into the the harmonic oscillator Hamiltonian Hharm := −∆y + 14 a2 |y|2 − ka. Hence the
a 2
linear operator −∆y − ay · ∇ is self-adjoint on the Hilbert space L2 (R, e− 2 |y| dy). Since,
a
as was already mentioned, the operator − n−k (∆Sn−k + 2(n − k)) is self-adjoint on the
Hilbert space L (C ), we conclude that the linear operator Lcyl
2 n
a is self-adjoint on the
n−k − a2 |y|2
Hilbert
n P space L2
(R k
× S o, e dydw). Moreover, the spectrum of −∆y − ay · ∇ is
a k1 si : si = 0, 1, 2, 3, . . . , with the normalized eigenvectors denoted by φs,a (y), s =
(s1 , . . . , sk ),
X k
(−∆y − ay · ∇)φs,a = a si φs,a , si = 0, 1, 2, 3, . . . . (17.35)
1
a a
Using that we have shown that the spectrum of − n−k (∆Sn−k + 2(n − k)) is n−k (l(l +
a
P k
n−k −1)−2(n−k)) = n−k l(l +n−k −1)−2a, l = 0, 1, . . . , and denoting r = 1 si , si =
0, 1, 2, 3, . . ., we conclude the spectrum of the linear operator Lcyl a , for k = 1, is
 
cyl a
spec(La ) = (r − 2)a + `(` + n − 2) : r = 0, 1, 2, 3, . . . ; ` = 0, 1, 2, . . . , (17.36)
n−1

with the normalized eigenvectors given by φr,l,m,a (y, w) := φr,a (y)Ylm (w). This equation
shows that the non-positive eigenvalues of the operator Lcyl
a , for k = 1, are

a 4 1
• the eigenvalue −2a of the multiplicity 1 with the eigenfunction φ0,0,0,a (y) = ( 2π )
((r, l) = (0, 0)), due to scaling of the transverse sphere;
a 4 m 1
• the eigenvalue −a of the multiplicity n with the eigenfunctions φ0,1,m,a (y) = ( 2π ) w , m=
1, . . . , n ((r, l) = (1, 1)), due to transverse translations;
a 14 √
• the eigenvalue 0 of the multiplicity n with the eigenfunctions φ1,1,m,a (y) = ( 2π ) aywm , m =
1, . . . , n ((r, l) = (0, 1)), due to rotation of the cylinder;
a 14 √
• the eigenvalue −a of the multiplicity 1 with the eigenfunction φ1,0,0,a (y) = ( 2π ) ay
((r, l) = (1, 0));
a 4 1
• the eigenvalue 0 of the multiplicity 1 with the eigenfunction φ2,0,0,a (y) = ( 2π ) (1 −
2
ay ) ((k, l) = (2, 0)).

The last two eigenvalues are not of the broken symmetry origin and are not covered by
Theorem 17.8. They indicate instability of the cylindrical collapse
168 Lectures on Applied PDEs, November 3, 2017

By the description of the spectra of the normal hessians of Lspha := HessNsph Va (ϕ) and
cyl N
La := Hesscyl Va (ϕ), we conclude that the spherical collapse is linearly stable while the
cylindrical one is not.
It is shown in [16] that indeed the spherical collapse is (nonlinearly) stable while the
cylindrical one is not. We will also show that the last two eigenvalues of Lcyla in the list
above are due to translations of the point of the neckpinch on the axis of the cylinder and
due to shape instability, respectively.

17.4 F −stability of self-similar surfaces


Another notion of stability was introduced by analogy with minimal surfaces in [?]:
Definition 17.16 (F −stability of self-similar surfaces, [?]). We say that a self-similar
surface φ, with a > 0, is F −stable iff the normal hessian satisfies HessN Va (ϕ) ≥ 0 on the
subspace span {ϕ · ν(ϕ), ν j (ϕ), j = 1, . . . , n + 1}⊥ .
Remark 17.17. a) The F −stability, at least in the compact case, says that the HessN Va (ϕ)
has the smallest possible negative subspace, i.e ϕ has the smallest possible Morse index.
b) The reason the F −stability works in the non-compact case is that, due to the sep-
aration of variables for the cylinder = (compact surface) ×Rk , the orthogonality to the
negative eigenfunctions of the compact factor removes the entire branches of the essential
spectrum. This might not work for warped cylinders.
c) Remark 17.9(c) shows that the spheres and cylinders are F −stable. However, it is
shown in [?, 16] that cylinders are dynamically unstable.
The following statement follows from Remark 17.9(c) and the definition of the F −stability
(see also the first part of Remark 17.17(5)):
q−stable surfaces in R
n+1
There are no smooth, embedded self-similar (a > 0), F close
to S k × Rn−k , where S k is the round k−sphere of radius ka .
A slight but a key improvement of this result is a hard theorem:
Theorem 17.18 (Colding - Minicozzi). The only smooth, complete, embedded self-similar
Rn+1 of polynomial growth are S k × Rn−k , where S k is
(a > 0), F −stable surfaces in q
k
the round k−sphere of radius a
.
Remark 17.19. The difference between this theorem and (trivial) Corollary 17.11 is
that the latter requires a slightly stronger condition HessN Va (ϕ) > 0 on the subspace
span {ϕ · ν(ϕ), ν j (ϕ), j = 1, . . . , n + 1}⊥ , then the F −stability.
Theorem 17.12 implies that if it is F −stable, then H(ϕ) > 0.
Theorem 17.20 (Colding- Minicozzi, Huisken). The only smooth, complete, embedded
self-similar surfaces in Rn+1 , with a > 0, polynomial
q growth and H(ϕ) > 0, are S k ×Rn−k ,
k
where S k is the round k−sphere of radius a
.
Lectures on Applied PDEs, November 3, 2017 169

(Compare with surfaces of constant mean curvature)


Theorems 17.12 and 17.20 imply Theorem 17.18.
There is a considerable literature on stable minimal surfaces. Much of it related to
the Bernstein conjecture:
The only entire minimal graphs are linear functions.
It was shown it is true for n ≤ 7:
(a) Bernstein for n = 2;

(b) De Georgi for n = 3;

(c) Almgren for n = 4; .

(d) Simons for n = 5, 6, 7.


(b) and (d) used in part results of Fleming on minimal cones. These results were extended
by Schoen, Simon and Yau.
Bombieri, De Georgi and Giusti constructed a contra example for n > 7.

18 Keller-Segel Equations of Chemotaxis


Chemotaxis is the directed movement of organisms in response to the concentration gra-
dient of an external chemical signal and is common in biology. The chemical signals can
come from external sources or they can be secreted by the organisms themselves. The
latter situation leads to aggregation of organisms and to the formation of patterns.
Chemotaxis is believed to underly many social activities of micro-organisms, e.g. social
motility, fruiting body development, quorum sensing and biofilm formation. A classical
example is the dynamics and the aggregation of Escherichia coli colonies under starvation
conditions [?]. Another example is the Dictyostelium amoeba , where single cell bacteri-
vores, when challenged by adverse conditions, form multicellular structures of ∼ 105 cells
[?, ?]. Also, endothelial cells of humans react to vascular endothelial growth factor to
form blood vessels through aggregation [?].
Consider organisms moving and interacting in a domain Ω ⊆ Rd , d = 1, 2 or 3.
Assuming that the organism population is large and the individuals are small relative to
the domain Ω, Keller and Segel derived a system of reaction-diffusion equations governing
the organism density ρ : Ω × R+ → R+ and chemical concentration c : Ω × R+ → R+ .
The equations are of the form

∂t ρ = Dρ ∆ρ − ∇ (f (ρ)∇c)
(18.1)
∂t c = Dc ∆c + αρ − βc.

Here Dρ , Dc , α, β are positive functions of x and t, ρ and c, and f (ρ) is a positive


function modelling chemotaxis (positive chemotaxis). Assuming a closed system, one is
170 Lectures on Applied PDEs, November 3, 2017

led to impose no flux boundary conditions on ρ and c:

∂ν ρ = 0 and ∂ν c = 0 on ∂Ω, (18.2)

where ∂ν g is the normal derivative of g.


The equations (18.1) have the family of homogeneous static solutions, (ρ∗ , c∗ ) : αρ∗ −
βc∗ = 0. A simple calculation shows that these solutions are linearly unstable: the
linearized operator has two spectral bands, one positive and the other negative (they are
separated by the gap). Due to the positive chemotaxis in the system, one expects that
the system evolves to a non-uniform state describing organism aggregation (the organisms
secrete the chemical and move towards areas of higher chemical concentration). One refers
to this process as (chemotactic) collapse.
In this case, the density may become infinite and form a Dirac delta singularity. More
precisely, we say that a solution ρ(x, t) undergoes collapse at a point x0 ∈ Rd in finite
time T < ∞ if it exists for 0 ≤ t < T and
Z
lim ρ(x, t) dx = 0, ∀ > 0.
t↑T |x−x0 |≥

Since the total mass is conserved, this implies that limt↑T ρ(x0 , t) = ∞.
The common approximations made in the literature for system (18.1) is based on the
fact that, in practically all situation, the coefficients in (18.1) are constant and satisfy

Dρ α β
 :=  1, α̃ := = O (1) and β̃ :=  1. (18.3)
Dc Dc Dc

The first of these conditions says that the chemical diffuses much faster than the organisms
do. As a result of this relation, one drops the ∂t c term in (18.1) (after rescaling time
t → t/Dρ , this term becomes ∂t c). Furthermore, one takes f (ρ) to be a linear function
f (ρ) = Kρ and neglects the term βc in (18.1) compared with αρ, as one expects that it
would not effect the blow-up process where ρ  1 (it is also small due to the last relation
in (18.3)). These approximations, after rescaling, lead to the system

∂ρ
= ∆ρ − ∇ · (ρ∇c) ,
∂t (18.4)
0 = ∆c + ρ,

with ρ and c satisfying the no-flux Neumann boundary conditions. Equation (18.4) is the
simplest model of positive chemotaxis considered in the literature. This is the equation
studied in these lectures. We note that Eq. (18.4) in three dimensions also appear in
the context of stellar collapse (see [?, ?, ?, ?]); similar equations—the Smoluchowski or
nonlinear Fokker-Planck equations—model non-Newtonian complex fluids (see [?, ?, ?, ?].
Lectures on Applied PDEs, November 3, 2017 171

18.1 Properties Keller-Segel equations


Equations (18.4) have the following properties

• (18.4) preserves positivity: if the initial condition ρ0 (x) is positive, then so is ρ(x, t)
(by the maximum principle).

• (18.4) is scaling invariant, that is, if a pair ρ(x, t) and c(x, t) is a solution to (18.4),
then for any λ > 0 so is the pair
   
1 1 1 1 1
ρ x, t and c x, t . (18.5)
λ2 λ λ2 λ λ2

• With the no flux boundary conditions, the total number of organisms in Ω is con-
served: Z Z
ρ(x, t) dx = ρ(x, 0) dx.
Ω Ω

• (18.4) is a gradient system, ∂t ρ = −grad F(ρ), with the energy


 1
Z
− ρ(−∆)−1 ρ + ρ ln ρ dx.

F(ρ) = (18.6)
R2 2

and the metric hv, wiJ := hv, J −1 wiL2 , where J := −∇ · ρ∇ > 0. In particular,

• the functional F(ρ) decreases under the evolution and


R
• its critical points under the constraint that ρ = M = const are static solutions of
(18.4).
(The first term of F(ρ) can be thought of as the internal energy of the system and
the remaining terms are the entropy.)

• For d = 2, the system (18.4) has the


R radially symmetric static solution, R(r), where
r = |x|, with the total mass, M = R(|x|)d2 x = 8π, it is given explicitly by
8
R(r) := . (18.7)
(1 + r2 )2

prove that ∂t ρ = −grad F(ρ), we compute the formal Gâteaux derivative dF(ρ)φ =
R To −1
(−∆ ρ + ln ρ)φ, and therefore the gradient in the metric hv, wiJ := hv, J −1 wiL2 is

grad F(ρ) = −∇ · ρ∇(−∆−1 ρ + ln ρ) = ∇ · ρ∇∆−1 ρ − ∆ρ,

which is the negative of the r.h.s. of the first equation in (18.4) with c = −∆−1 ρ. Hence
the equation (18.4) can be written as ∂t ρ = −grad F(ρ).
172 Lectures on Applied PDEs, November 3, 2017

Under this scaling, the total mass changes as


 
1 1
Z Z
(d−2)
ρ x, t = λ ρ (x, t) .
λ2 λ
Thus one does not expect collapse for d = 1 and that collapse is possible for d ≥ 2 with
critical collapse for d = 2 and supercritical collapse for d > 2. (Equation (18.4) in d = 2
is said to be L1 −critical, etc.)
By the scaling invariance, for d = 2, the system (18.4) has the one-parameter family
of radially symmetric static solutions, Rλ (r) = λ12 R( λ1 r).

Existence vs blowup dichotomy. The dimension of the interest for us is the critical
dimension d = 2. Recall that Rfor d = 2, system (18.4) has a radially symmetric static
solution (18.7). We note that R2 R dx = 8π. This mass turns out to be the threshold
separating a regular behavior and a breakdown of the solution. It is shown that for the
initial condition ρ0 ≥ 0,
R
• (Blanchet, Dolbeault, Perthame) If the initial total mass satisfies
R M := ρ dx ≤
R2 0
8π, then the solution to (18.4) exists globally and, for M := R2 ρ0 dx < 8π, con-
verges to 0, as t → ∞;
R
• (Biler) If the initial total mass satisfies M := R2 ρ0 dx > 8π, then the solution to
(18.4) blows up in finite time.
We sketch some key ideas in proving the first statement. To prove the global existence,
we have to control some appropriate
R positive quantity, like a Sobolev norm. In the present
case, this is the entropy, ρ ln ρ.
If the blowup takes place along the family,R Rλ (r) = λ2 R(λr),
R of static solutions
R with
λ → ∞, then the entropy would blow up, ρλ ln ρλ = ρ ln ρ + 2 ln λ ρ → ∞ as
λ → ∞, for any ρ with finite entropy and the total mass.
Hence, if the entropy stays bounded during the evolution, this would indicate that ρ
does not blow up by the compression as in ρλ , λ → ∞. To check this, we start with
computing the change in the entropy
√ 2
Z Z Z
∂t ρ ln ρ = −4 |∇ ρ| + ρ2 . (18.8)
| {z } | {z }
entropy dissipation entropy production

Depending on whether the entropy dissipation or the entropy production wins we expect
either dissipation of the solution or the collapse (blowup). The Nirenberg - Gagliardo
inequality,
kf k24 ≤ cgn k∇f k2 kf k2
shows that the dissipation wins if M c2gn ≤ 4.
Lectures on Applied PDEs, November 3, 2017 173

To sharpen this result one uses that the free energy decreases together with the loga-
rithmic Hardy-Littlewood-Sobolev inequality,
1 1
Z Z
f ln f ≥ f (−∆)−1 f − C(M ),
(M/8π) 2
R
where M := f (the dimension n = 2) and C(M ) := M (1 + log π − log M ), which gives
the following lower bound on the free energy,
1 1
Z
F(ρ) ≥ ( − 1) ρ(−∆)−1 ρ − C(M ).
(M/8π) 2
Combining this inequality together with the fact that the free energy is decreases, F(ρ) ≤
F(ρ0 ), one finds the bound on the entropy
1
Z
(1 − M/8π) ρ ln ρ ≤ F(ρ0 ) − C(M ),

which is used to prove the global existence for M ≤ 8π.
To obtain the control of ρ one uses, instead of F(ρ) ≤ F(ρ0 ), its quantatative version
(the free energy production or generalized Fisher information)
Z
∂t F(ρ) = − ρ|∇ ln ρ − ∇c|2 (18.9)

(this can be thought of as an entropy monotonicity formula).


Remark (Hardy-Littlewood-Sobolev inequality.). The logarithmic Hardy-Littlewood-Sobolev
inequality follows from the generalized Hardy-Littlewood-Sobolev inequality,
Z
−λ


|t − s| f (s)g(t)dsdt . kf kL` kgkLm
(18.10)
s<t

provided 1` + m1 + λ = 2, λ < 1, `, m > 1. Inequality (18.10) follows from the generalized


Young inequality kh ∗ gkr ≤ khkw,p kf kq , p1 + 1q = 1r + 1, 1 < p, q, r < ∞, where

khkw,p = sup(tp µ{x : |h(x)| > t})1/p < ∞. (18.11)


t

Indeed, we have k|t|−λ kw, 1 < ∞ and


λ
Z
(|t|−λ ∗ f )g(t)dt ≤ k|t|−λ ∗ f km0 kgkm ≤ k|t|−λ k 1 kf k` kgkm .

w, (18.12)
λ

The equation (18.11) defines the weak Lp space, Lpw := {h is measurable and khkw,p <
∞}. Indeed, we have khkw,p ≤ khkp , so that Lp ⊂ Lpw . This follows from the following
expression Z ∞
khkp = − tp dmf (t), where mf (t) := µ(x : |f (x)| > t). (18.13)
0
174 Lectures on Applied PDEs, November 3, 2017

Virial relation. To see how the critical mass M∗ = 8π enters here, we consider the
second moment of mass Z
W := |x|2 ρ(x, t) dx.
R2

We have the following virial relation

1
∂t W = 4M (1 − M ). (18.14)

If M > 8π, then the right hand side is constant and negative, and hence, W becomes
negative in finite time. When this happens we have a contradiction since W is by definition
always positive (recall that if ρ0 ≥ 0 then ρ(t) ≥ 0 by the maximum principle). Thus, if
M > 8π, then the solution ρ exists only for a finite time (t < t∗ , where t∗ is the point of
time when W vanishes).
This result tells us nothing about how the solution break down in finite time. The
latter process is investigated in the subsequent sections.

18.2 Rescaling
Recall that (18.4) has the manifold of static solutions Mstat := {λ−2 R(|x + h|/λ) | λ >
0, h ∈ R2 }. Assuming this manifold is stable, one can slide along it either in the direction
λ → ∞ (dissipation) or in the direction λ → 0 (collapse).
Our goal is to understand which scenario takes place. The first key step is to pass
to the reference frame collapsing with the solution, by introducing the adaptive (blowup)
variables,
Z t
2 −1
u(y, τ ) = λ ρ(x, t), c(x, t) = v(y, τ ), where y = λ x and τ = λ−2 (s) ds,
0
(18.15)
where λ : [0, T ) → [0, ∞), T > 0, is a positive differentiable function (compression or
dilatation parameter). The advantage of passing to blowup variables is
(a) if the solution ρ blows up at a finite time T , we expect λ to adjust in such a way
that λ(t) → 0, while u stays bounded, as t ↑ T , and similarly for the dispersion;
(b) in the case of the blowup in a finite time T , we expect that τ → ∞, as t ↑ T , and
therefore the blowup time, T , gets eliminated from consideration (it is mapped to ∞).
Writing (18.4) in blowup variables, we find the equation for the rescaled mass function

∂τ u = ∆u − ∇ · (u∇v) − a∇ · (yu), (18.16)

where a := −λ̇λ and, recall −∆v = u. Now, the blowup problem for (18.4) is mapped
into the problem of asymptotic dynamics of solitons for the equation (18.16). One can
Lectures on Applied PDEs, November 3, 2017 175

forget about λ and consider (18.16) as an equation for u and a and then find λ, given
λ(0) = λ0 , according to the formula
Z t
2 2
λ (t) = λ0 − 2 a(s) ds. (18.17)
0

The equation (18.16) retains all but the symmetry properties


R of (18.4): it is (a) posi-
tivity preserving (in fact, improving), (b) mass conserving ( u = ρ = const) and (c) is a
gradient system, ∂t u = −grad Fa (u), with the energy
 1 a
Z
− u(−∆)−1 u + u ln u − |y|2 u dx.

Fa (u) = (18.18)
R2 2 2
and the metric hv, wiJ := hv, J −1 wiL2 , where J := −∇ · u∇ > 0.

18.3 M > 8π
We are interested in behaviour of solutions for initial condition with M > 8π. We begin
with considering the linearized stability of the static solution R(|x|) of (18.4). ∗ ∗ ∗ Note
that the linearized analysis does not differentiate between M > 8π and M < 8π and
therefore, between the blowup and stability. It would have to be supplemented by addi-
tional information. (Extending the virial relation (??) to the equation (18.16) points out
resc 2
R
at a possible connection between M and a. Namely, let W := R2
|y| Ru(y, τ2) dy. Then
2
R
proceeding as in the derivation of (??)nd using R2 |y| ∇ · (yu) dy = −2 R2 |y| u(y, τ ) dy,
1
we find ∂t W resc = 4M (1 − 8π M ) + 2aW resc .)
Linearizing the r.h.s., F (u) := ∆u − ∇ · (u∇v), of (18.4) on R(|x|) and denoting
L := −dF (R), we find

Lξ = −∆ξ − ∇ · ξ∇∆−1 R − R∇∆−1 ξ .



(18.19)

R operator L is the hessian of F(u) at R defined on the space of the linearized constraint
The
{ ξ = 0} and therefore it is self-adjoint in the inner product hv, wiJ := hv, J −1 wiL2 ,
where, recall, J := −∇ · u∇ > 0. By a standard technique (see e.g. [6]), one has
σess (L) = [0, ∞).
Similarly, the rescaled equation (18.16), linearized around the rescaled stationary so-
lution R(|y|), leads to the operator La := −dFa (R), where Fa (u) := ∆u − ∇ · (u∇v) −
a∇ · (yu). Explicitly,
La ξ = Lξ + a∇ · (yξ), (18.20)
The operator R La is the hessian of Fa (u) at R defined on the space of the linearized
constraint { ξ = 0} and therefore it is self-adjoint in the inner product hv, wiJ :=
2
hv, J −1 wiL2 . Since the operator −∆ − a∇ · y on the space L2 (R2 , e−a|y| /2 dy) is unitary
equivalent to −∆ + 14 a2 |y|2 − a on the space L2 (R2 , dy), its spectrum is purely discrete
with gaps of the order O(a).
176 Lectures on Applied PDEs, November 3, 2017

∗ ∗ ∗ The information about the total mass M (of the initial condition), i.e. whether
M < 8π or M > 8π, will enter through the choice of the modalities of the behaviour of
a: for M < 8π, we expect a < 0 and converges to a nonzero limit and for M > 8π, we
expect a > 0 and, for M close to 8π, a → 0, as τ → ∞.
For a < 0 (more precisely, a = 1), the operator La was studied in detail in a number
of papers, see e.g. [?, ?].
We address the case a > 0 and very small. In the case, we can think of La and a small
perturbation of L. So, first, we examine the operator L.
Since R(|x|) breaks the translational and scaling symmetries of (18.4), the operator L
has the the translational and scaling zero modes
32xj 34 32
∂j R = and ∇ · (xR) = − .
(1 + |x|2 )3 2
(1 + |x| )2 (1 + |x|2 )3

Since L commutes with rotations, it can be decomposed into spherical harmonics (the
Fourier series in the polar angular variable θ) as

L = ⊕m≥0 Lm ,

where the operator Lm acts on the space L2 ([0, ∞), rdr) of radial functions and can be
written out explicitly. Then ϕ0 := ∇ · (xR) becomes the zero eigenfunction of L0 and
32|x|
∂j R give the zero eigenfunction, ϕ1 := (1+|x| 2 )3 , of L1 . Though ϕ0 and ϕ1 are positive,

since the operators Lm are non-local, we cannot use the Perron - Frobenius argument, to
conclude that 0 is the lowest and simple eigenvalue of L0 and L1 .
Returning to the operator La , we see that it has 3 zero or almost zero eigenvalues
originating from the triple degenerate eigenvalue 0 of L. Note that, since L∂j R = 0, we
have
h∂j R, La ∂j Ri = ah∂j R, ∇ · (y∂j R)i = ak∂j Rk2 .
Since La commutes with rotations, it can be decomposed into spherical harmonics (the
Fourier series in the polar angular variable θ) as

La = ⊕m≥0 Lam ,

where the operator Lam acts on the space L2 ([0, ∞), rdr) of radial functions and is related
to the operator Lm above as Lam = Lm + a(r∂r + 2).
For a > 0 and sufficiently small, the radially symmetric part, La0 , was investigated in
[?], where it was shown that it has the following spectrum in the interval . a:
• one negative eigenvalue −2a + lna1 + O a ln−2 a1 (corresponding to the scaling

a
mode—for a fixed parabolic scaling it is connected to possible variation of the blowup
time)

• one near zero eigenvalue (due to the shape instability),


Lectures on Applied PDEs, November 3, 2017 177

2a
+ O a ln−2 1

• the third eigenvalue, 2a + ln a1 a
, is positive, but vanishing as a → 0.

The paper [?] also isolated the correct perturbation (adiabatic) parameter— ln11 . Finally,
a
we record the more precise information about the eigenvalues obtained in [?]:
µ+a −3 1
( 
1
ln a +K+γ
+ O a ln a
n=0
λn = 2a −3 1
 (18.21)
2na + ln 1 +K+γ−H − µ+a + O a ln a n ≥ 1,
a n−1 2an

where K :=P ln 2−1−2γ (here γ = −Ψ(1) = 0.577216 . . . is the Euler-Mascheroni constant)


and Hn := nk=1 1/k.
We proceed with the radially symmetric case and make comments about the general
case later.

Modification of the leading term. Non-positive EVs La ⇒ Mstat := { λ12 R(x/λ) | λ >
0} is unstable and we have to construct a one-parameter deformation of it. For technical
reasons it is convenient to use a two-parameter family, Rbc (|y|)
8b
Rbc (|y|) := , (18.22)
(c + |y|2 )2
with b > 1 and both parameters b and c are close to 1, with an extra relation between the
parameters a, b and c. The family Rbc (|y|) gives a two-parameter family of approximate
solutions to (18.16) (see (18.7)) and forms the deformation (or almost center-unstable)
manifold M := {Rbc (|y|/λ) | λ > 0, p}.
Introduce the deformation (or almost center-unstable) manifold
Mstat deform := {(1/λ2 )Rbc (|x|/λ) | λ > 0, b, c}. (18.23)
It a key point that the tangent vectors ∂b Rb(τ )c(τ ) and ∂c Rb(τ )c(τ ) , spanning the tangent
space to Mstat deform at Rbc are the approximate eigenfunctions of the operator La0 corre-
sponding to the negative and almost zero eigenvalues displayed above.
This manifold absorbs all unstable/neutral degrees of freedom. The previous result
gives the linear stability of Mstat deform .

Splitting the solution. We expect that the solution to the rescaled KS approaches
this manifold as τ → ∞. Hence we decompose the solution u(y, τ ) to the rescaled KS as
the leading term, Rb(τ )c(τ ) (y), and the fluctuation, φ(y, τ ),
u(y, τ ) = Rb(τ )c(τ ) (y) + φ(y, τ ) , (18.24)
| {z } | {z }
leading term, finite dim fluctuation, infinite dim

and require that the fluctuation φ(y, τ ) is orthogonal to the tangent space of Mstat deform
at Rb(τ )c(τ ) (y),
h∂b,c Rb(τ )c(τ ) (·), φ(·, τ )i = 0.
178 Lectures on Applied PDEs, November 3, 2017

The leading term, Rb(τ )c(τ ) (y), and the fluctuation, φ(y, τ ), evolve on a different spatial
scales, as Rbc can rewritten as Rbc (y) = R b4 ,1 ( √yc ).
c

Collapse dynamics. In parametrizing solutions as above, we split the dynamics of


(18.4) into a finite-dimensional part describing motion over the manifold, M, and an
infinite-dimensional fluctuation (the error between the solution and the manifold approx-
imation) which is supposed to stay small.
Substituting the splitting u = Rbc +φ into the rescaled KS and projecting the resulting
equation onto Mstat deform and the orthogonal complement (to the tangent space) (this is
the Lyapunov-Schmidt decomposition we used extensively in the previous lectures), we
arrive at the equations for parameters b and c

 cτ = 2a − 4(b−1) + R(φ, a, b, c),
ln( a1 )
(18.25)
 baτ = − 2(b−1) + R(φ, a, b, c),
ln( 1 )a

and for the fluctuation φ.


Using the equation for the fluctuation φ and differential inequalities for Lyapunov func-
tionals, we estimate the remainders in (18.25) in the linear approximation as |R(φ, a, b, c)| .
1
a
ln2 ( a1 ) ln( a1 )
[(b − 1)kφkL2 + k(1 + |y|2 )−1 φk2L2 ]. As a result, these equations give the differ-
ential equation for a
2a2
aτ = − , (18.26)
ln( a1 )
2
with an error term O( ln2a( 1 ) ), given an a priori estimate of φ.
a
We solve the above equation in the leading order. First, we have a1 ln a1 + O(1) = 2τ


which results in ln a(τ1 ) = ln 2τ −ln ln 2τ + lnlnln2τ2τ +O ln12τ . Now, recalling that λ(t)λ̇(t) =


a(τ ) and using that λ(t)−1 λ̇(t) = λ(τ (t))−1 ∂τ λ(τ (t)), we obtain, after some derivations,
the law
1 1 1
λ(t) = (T − t)− 2 e| 2 ln(T −t)| 2 (c1 + o(1)). (18.27)
∗∗∗∗∗∗∗∗∗
∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗∗

18.4 Discussion of static solutions


The static solutions of (18.4)
R are critical points of the energy functional F, given in (18.6),
0
under the constraint that ρ = M = const.. Thus, they satisfy E (ρ) = C, where C is a
constant. Explicitly E 0 (ρ) = C reads
1
log(ρ) + ρ = C ⇔ ∆ log(ρ) + ρ = 0 ⇔ ∆u + eu = 0, (18.28)

Lectures on Applied PDEs, November 3, 2017 179

where u = log(ρ). Solutions to (18.28) can be written in the form of ’Gibbs states’
c
ρ = M Reec , with the concentration c considered as a negative potential (remember that
∆c = −ρ). In two dimensions, this equation has the solutionRfor M = 8π, given by (18.7).
This solution is a minimizer of E under the constraint that ρ = 8π.
The equations (18.4) have the family of homogeneous static solutions, (ρ∗ , c∗ ) : ρ∗ =
0, c∗ =constant. These solutions have no organisms present (zero total mass) and constant
concentration of the chemoattractant. A simple calculation shows that the linearized
operator has the spectrum filling in the positive semi-axis. This indicates that these
static solutions might be stable. However, any initial condition ρ0 ≥ 0, ρ0 is not identical
0 will have positive mass and therefore, since the total mass is conserved, could converge
to (ρ∗ , c∗ ) only in a weak sense with the loss of the mass.
∗ ∗ ∗ ∗ ∗ ∗ ∗∗
Consider the RKS with the degradation term for the chemical allowing for the ho-
mogeneous static solutions; investigate the stability of the latter and connection to the
Turing instability.

18.5 Blowup criterion


R
We note that R2 R dx = 8π. This mass turns out to be the threshold for the blowup of
solutions of (18.4). Indeed, we have the following result:

Theorem 5 (See [?]). Take ρ0 ≥ 0. If the dimension d = 2 and the total mass satisfies
Z
M := ρ0 dx > 8π, (18.29)
R2

then the solution to (18.4) breaks down in finite time.

Proof. The time derivative of the second moment of mass


Z
W := x2 ρ(x) dx
R2

is, using that ρ is a solution to (18.4),


Z
x2 ∆ρ + ∇ ρ∇∆−1 ρ dx.

∂t W =
R2

Integrating by parts and using that R2 x2 ∆ρdx = 2d R2 ρdx = 2d R2 ρ0 dx = 2dM , by


R R R

the conservation of total mass, we obtain

∂t W = 2(dM − J). (18.30)


180 Lectures on Applied PDEs, November 3, 2017

−1
ρ) · x dx. Using the integral representation of ∆−1 ρ (in two dimen-
R
where J := ρ (∇∆
1
sions ∆−1 ρ = 2π
R
R2
ln |x − y| ρ(y) dy), and symmetrizing the integral we compute

1 x−y
Z
J= x· ρ(y)ρ(x) dy dx
2π R2d |x − y|d
1 x−y 1 y−x
Z Z
= x· ρ(y)ρ(x) dy dx + y· ρ(y)ρ(x) dy dx
4π R2d |x − y|d 4π R2d |x − y|d

This gives
1 1
Z
J= ρ(y)ρ(x) dy dx.
4π R2d |x − y|(d−2)
1
For d = 2, this expression gives that J = 2π M 2 . Substituting this into (18.30), we obtain
that
1
∂t W = 4M (1 − M ).

If M > 8π, then the right hand side is constant and negative, and hence, W becomes
negative in finite time. When this happens we have a contradiction since W is by definition
always positive (recall that if ρ0 ≥ 0 then ρ(t) ≥ 0 by the maximum principle). Thus, if
M > 8π, then the solution ρ exists only for a finite time (t < t∗ , where t∗ is the point of
time when W vanishes).

18.6 Appendix: Gradient formulation


The Keller-Segel models (18.1) and (18.4) are gradient systems. We begin by formulating
a normalized version of (18.1),

∂t ρ = ∆ρ − ∇ · (f (ρ)∇c)
(18.31)
ε∂t c = ∆c + ρ − γc,

as a gradient system. This system is obtained from (18.1) by setting unimportant con-
stants to 1.
Define the energy (or Lyapunov) functional

1 γ
Z
Ef (ρ, c) := |∇c|2 − ρc + c2 + G(ρ) dx, (18.32)
Ω 2 2
Rρ Rρ 1
where G(ρ) := g(s) ds and g(ρ) := f (s)
ds. The L2 -gradient of Ef (ρ, c) is
 
−c + g(ρ)
gradL2 Ef (ρ, c) = ,
−∆c − ρ + γc
Lectures on Applied PDEs, November 3, 2017 181

and hence, if we define U = (ρ, c), then (18.31) can be written in the form ∂t U = IEf0 (U ),
where  
∇ · f (ρ)∇ 0
I= .
0 − 1ε
The operator I is non-positive and may be degenerate, however, assuming it is invertible,
the operator I defines the metric hv, wiI := − hv, I −1 wiL2 ⊕L2 . In this metric, grad F(U ) =
−IF 0 (U ) and hence
∂t U = −grad Ef (U ).
This shows that (18.31) has the structure of a gradient system. A consequence of this is
that the energy decreases on solutions of the KS system. Indeed, if f > 0, then
1
2 1
∂t Ef (ρ, c) = − f (ρ) ∇ (c − g(ρ) ) − k∆c + ρ − ck2L2 .
2

L 2 ε
The free energy F in (18.6) is obtained from (18.32) by dropping the quadratic term
1 2
2
c,replacing c with −∆−1 ρ in the remaining terms and using that f (ρ) = ρ.

18.7 Appendix 2: Criterion of break-down in the dimension


d≥3
For d ≥ 3, we have the following result
x2 ρ dx
R
0
Theorem 18.1. Take ρ0 ≥ 0. If d ≥ 3 and RRd d ρ0 dx is sufficiently small (this means that
R
ρ0 is concentrated at x = 0), then the solution to (18.4) blows up in finite time.
Proof. We now prove the case when d ≥ 3. We proceed as above and derive an upper
bound on ∂t W which is negative for certain initial conditions. To this end we derive a
lower bound on J. Let γ = d − 2. We estimate M 2 from above using J by first rewriting
M 2:
Z  α
u(x)u(y) 
Z 1−α
αγ
2
M = u(x)u(y) dx dy = u(x)u(y)|x − y| 1−α dx dy.
R2d R2d |x − y|γ
Hölder’s inequality then gives
Z α Z 1−α
2 u(x)u(y) αγ
M ≤ dx dy u(x)u(y)|x − y| 1−α dx dy .
R2d |x − y|
γ
R2d
αγ 2
The first integral is J α . We choose 1−α
= 2 or α = γ+2
so that
Z 2 Z
 γ+2 γ
 γ+2
2 u(x)u(y) 2
M ≤ dx dy u(x)u(y)|x − y| dx dy
R4 |x − y|
γ
R2d
Z γ
 γ+2
2
2
=J γ+2 u(x)u(y)|x − y| dx dy .
R2d
182 Lectures on Applied PDEs, November 3, 2017

Expanding the square we obtain that


Z Z
2
u(x)u(y) x2 + y 2 − 2x · y dx dy

u(x)u(y)|x − y| dx dy =
R2d ZR
2d

u(x)u(y) x2 + y 2 + 2|x||y|

=
R2d
Z 2
= 2W M + 2 |x|u(x) dx .
Rd

By Hölder’s inequality again


Z 2 Z Z
1 1
2
|x|u 2 u 2 dx ≤ x u dx u dx = W M.
Rd Rd Rd

Combining the estimates, one obtains that


2 γ
M 2 ≤ J γ+2 (4W M ) γ+2
or γ+4 γ
J ≥ CM 2 W−2 .
Substituting into (18.30) gives that
γ+4 −γ
∂t W ≤ 2(dM − CM 2 W 2 ).
Thus, if initially M/W (0)  1, then ∂t W is negative and W will decrease. Since it
decreases, the time derivative ∂t W becomes more negative and hence W continues to
decrease and will becomes negative in a finite time (since the time derivative will always
be less that the initial value of ∂t W ). This is again a contradiction since W is by definition
always positive.

19 Minimization: direct methods


19.1 General result
The problem we address is the following: given a functional E on a space M , find a
function u0 ∈ M (if such exists) that minimizes E:
E(u0 ) = inf E(u).
u∈M

Such a function u0 is called a minimizer for E. Thus to begin with, we want to assume
that E is bounded below, i.e.
E0 := inf E(u) > −∞.
u∈M
Lectures on Applied PDEs, November 3, 2017 183

We also assume that M is a closed subset of a Banach space, X. Let us first analyze the
finite dimensional situation: X = RN . How would we minimize a functional on M ? We
do this in three steps (that will be suitable to be generalized to the infinite dimensional
case):

• Step 1. Since E is bounded on M from below, E0 := inf u∈M E(u) > −∞, we can
pick a sequence {un } ⊂ M s.t. E(un ) → E0 , as n → ∞. Such a sequence is called a
minimizing sequence. Clearly we can take the sequence {un } s.t. every element un
satisfies
E(un ) ≤ E0 + 1 (19.1)
(just throw out those un for which E(un ) > E0 + 1).

• Step 2. We hope that either such a sequence converges or at least contains a


convergent subsequence. The limit of such a subsequence clearly is a candidate
for a minimizer if the latter exists. How do we show that {un } has a convergent
subsequence? Assume that

E(u) → ∞ as kukX → ∞. (19.2)

A functional E on M satisfying (19.2) is called coercive. Due to (19.1) and (19.2), we


have that kun kX ≤ C, ∀n, and for some C > 0. Hence by the Bolzano–Weierstrass
theorem, {un } has a convergent subsequence, which for notational convenience we
denote again by {un }.

• Step 3. Let u0 := limn→∞ un . If E is continuous, then

lim E(un ) = E(u0 ).


n→∞

Since on the other hand we have limn→∞ E(un ) = E0 , we conclude that E(u0 ) = E0 ,
i.e. u0 is a minimizer of E.

Let us look closer at the last two steps. Recall that a set K s.t. every infinite sequence
of elements of K contains a convergent subsequence is called (sequentially) compact. The
Bolzano–Weierstrass theorem states that a closed ball in RN is compact. This property
does not hold in general in the infinite dimensional case. For instance, closed balls in L2 (Ω)
are not compact. As a concrete example, take for instance L2 (Rn ) 3 un (x) := u(x − n),
for some u ∈ L2 (Rn ), kuk2 = 1. Clearly, this sequence does not have a convergent
subsequence.
We have however, the following weaker result (this result follows from the Banach-
Alaoglu theorem): If X is a reflexive, separable Banach space, then every uniformly
bounded sequence {un } in X (i.e. kun kX ≤ C) has a weakly convergent subsequence
{unk }. (In the previous example, un (x) := u(x − n) converges to 0 weakly. In the
184 Lectures on Applied PDEs, November 3, 2017

finite dimensional spaces, weak convergence is equivalent to strong convergence, and the
Banach-Alaoglu theorem reduces to the Bolzano-Weierstrass theorem.) 2
Thus we have to assume that M is weakly closed: We say that M is weakly closed in
w
X if and only if un −→u0 in X and un ∈ M , imply u0 ∈ M .
Next, the continuity w.r.to the weak convergence is a property hard to come by for
functionals on infinite dimensional spaces (e.g. u → kuk2X is not weakly continuous in X).
But there is a weaker property which often holds: the weak lower semicontinuity, which is
w
defined as follows. E is called weakly lower semicontinuous (w.l.s.c.) if and only if un −→u0
in M implies lim inf n→∞ E(un ) ≥ E(u0 ) (recall that lim inf un := limk→∞ (inf n≥k un )) (e.g.
u → kuk2X is w.l.s.c. in X).
w
We now check that in fact w.l.s.c. suffices to carry out step 3 above. If un −→u0 ,
then lim inf n→∞ E(un ) ≥ E(u0 ). On the other hand, by the definition of {un }, we have
lim inf n→∞ E(un ) = E0 = inf u∈M E(u). Therefore E(u0 ) = E0 , and hence u0 is a minimizer
of E.
Thus, we have essentially proven the following

Theorem 6 (Key Theorem). Assume that

(α) M is weakly closed in X,

(β) E is w.l.s.c.,

(γ) E is coercive.

Then E is bounded below and attains its minimum in M (i.e. there is a minimizer of E
on M ).

Proof. Let E0 := inf u∈M E(u), and let un be a minimizing sequence for E, i.e.

lim E(un ) = E0 . (19.3)


n→∞

Clearly, we can assume that E(un ) ≤ E0 + 1 (we get rid of those un ’s in the minimizing
sequence for which E(un ) > E0 + 1). Then by the coercivity of E, there is a constant
K s.t. kun k ≤ K, ∀n. Hence, by the Banach-Alaoglu theorem, {un } contains a weakly
w
convergent subsequence, {un0 }, un0 −→u0 ∈ X. The element u0 is a candidate for a
minimizer. Since M is weakly closed, u0 ∈ M . W.l.s.c. of E gives lim inf n0 →∞ E(un0 ) ≥
E(u0 ). This together with equation (19.3) implies that E0 ≥ E(u0 ). On the other hand,
E0 = inf u∈M E(u) ≤ E(u0 ), and therefore E(u0 ) = E0 . This shows that u0 is a minimizer
and that inf u∈M E(u) > −∞.
2
Here we recall some definitions. A Banach space is said to be reflexive, iff its second dual (the dual
of the dual) is isometrically isomorphic to the space itself. Every Hilbert space is reflexive, see e.g., [12],
Section 6.3, Theorem 2 and its Corollary. Furthermore, we say a sequence {un } in X is weakly convergent
w w
iff ∃u0 s.t. `(unk ) → `(u0 ), ∀` ∈ X 0 . The weak convergence is denoted by −→, as in un −→u0 .
Lectures on Applied PDEs, November 3, 2017 185

19.2 Applications
This is a simple but powerful result. It says that in order to show that E has a minimizer,
we have to check three conditions (α) - (γ). We begin with the discussion of these
conditions. We begin with (β).
Lemma 2. Let X be a Hilbert space. Then the functional u → kuk2X is w.l.s.c. on X.
Proof. Dropping the subindex X, we have kun k2 = hun , un i = hun − u + u, un − u + ui =
hu, ui + 2 Rehun − u, ui + hun − u, un − ui ≥ kuk2 + 2 Rehun − u, ui. Since hun − u, ui → 0,
this gives lim inf kun k2 = kuk2 .
In particular, this lemma implies
1
|∇u|2 is w.l.s.c. on H 1 (Ω).3
R
Corollary 7. The functional E(u) := 2 Ω
R
Lemma 3. Let G(x, u) ≥ 0, be l.s.c. in u and s.t. Ω G(x, u) is well defined for u ∈
H 1 (Ω). Then the functional E(u) := Ω G(x, u) is w.l.s.c. on H 1 (Ω).
R

w
Proof. Indeed, let un −→u in H 1 (Ω) and let Q be a bounded subset of Ω (Q = Ω if Ω is
bounded). Then by the Rellich-Kondrashov theorem, un → u in L2 (Ω) (up to picking a
subsequence) and therefore un → u a.e. in Q. Since Q is arbitrary bounded R subset of Ω, we
4
have that u n → u a.e.. Therefore, by Fatou’s lemma , we get lim inf n→∞ Ω G(x, un (x)) ≥
1
R R

lim inf n→∞ G(x, un (x)) ≥ Ω G(x, u(x)), and E is w.l.s.c. on H (Ω).

Lemmas 2 and 3 imply


Corollary 8. If G(x, u) ≥ 0, l.s.c. in u and s.t. Ω G(x, u) is well defined for u ∈ H 1 (Ω),
R

then the functional


1
Z
E(u) = ( |∇u(x)|2 + G(x, u(x)))dx. (19.4)
Ω 2
is w.l.s.c. on H 1 (Ω).
3
We show this by the direct computation:
1 1
Z Z
2
|∇un | = |∇(u + un − u)|2
2 Ω 2 Ω
1 1
Z Z Z
= |∇u|2 + Re ∇u · ∇(un − u) + |∇(un − u)|2
2 Ω 2
ZΩ Ω
1
Z
2
≥ |∇u| + Re ∇u · ∇(un − u).
2 Ω Ω
w
If un −→u in H 1 (Ω), then Ω ∇u · ∇(un − u) → 0, and therefore
R

1 1
Z Z
lim inf |∇un |2 ≥ |∇u|2 .
n→∞ 2 Ω 2 Ω

4
R
R For any sequesnce {fn } of non-negative and measurable functions, lim inf n→∞ Ω
fn (x) ≥

lim inf n→∞ fn (x).
186 Lectures on Applied PDEs, November 3, 2017

Now, we discuss (γ), the coercivity.


(c) The functional E(u) = Ω ( 12 |∇u|2 + c|u|2 ), c > 0, is obviously coercive. Indeed,
R

if c > 0, then E(u) ≥ δkuk2(1) , where δ = min(1/2, c) > 0, and, recall, kuk(1) = kukH 1 (Ω) ,
and therefore E is coercive on H 1 (Ω), i.e. E(u) → ∞ whenever kuk(1) → ∞.
A more delicate situation is with the Dirichlet functional. We have
1
(d) The functional E(u) := 2 Ω |∇u|2 is is coercive on H 1 (Ω), provided Ω is bounded
R
in
oneRdirection. Indeed, by the Poincaré inequality (see (19.12) below), we have Ω |u|2 ≤
R

D2 Ω |∇u|2 , for any u ∈ H01 (Ω), where D is the smallest diameter of Ω. So we get
E(u) ≥ 41 min(1, D−2 )||u||2(1) for every u ∈ H01 (Ω). Therefore E is coercive on H01 (Ω).
Collecting above statements we obtain
Proposition 1. Let G(x, u) ≥ c|u|2 , for some c ≥ 0. If c > 0, then the functional (19.4)
is coercive on H 1 (Ω); if c = 0 and Ω is bounded in one direction, then (19.4) is coercive
on H01 (Ω).
Now we give examples of weakly closed sets.
(f ) M = X or M = g + X (for a fixed g), where X is a reflexive Banach space (in
particular, a Hilbert space). Hence Hg1 (Ω) is weakly closed in H 1 (Ω), if g has an
extension, g̃, from ∂Ω to Ω. Indeed, in the latter case can write Hg1 = g̃ + H01 (Ω).
Since H01 (Ω) is a Hilbert space, the result follows.
(g) A convex, closed subset of a reflexive Banach space (by Mazur’s theorem).
(h) M0 = {u ∈ H01 (Ω) : Ω |u|p dn x = 1}, where Ω is a bounded, smooth domain in Rn
R
2n
and p < n−2 if n > 2 and p < ∞ if n ≤ 2.
Proposition 2. The set M0 defined above is weakly closed in H01 (Ω).

Proof. By the Rellich-Kondrashov theorem, H01 (Ω) is compactly embedded into


Lp (Ω), for p as in the proposition, and Ω bounded. This means that any weakly
w
convergent sequence un −→u0 in H01 (Ω) contains a subsequence {un0 } s.t. un0 → u0
in Lp (Ω). Hence ku0 kp = limn→∞ kun kp = 1, and therefore u0 ∈ M0 .

Let us draw some conclusions from what we have just shown:


Theorem 9. Assume G(x, u) ≥ 0, l.s.c. in u, is integrable over Ω and satisfies G(x, u) ≥
c|u|2 . Then the functional (19.4) has a minimizer in H01 (Ω) and in M0 , if c ≥ 0 and Ω is
bounded in one direction, and in H 1 (Ω), if c > 0.
Exercise 2. (Nonlinear Dirichlet problem) Use results above to show that the Dirichlet
problem
∆u − |u|2 u = 0 in Ω,
(19.5)
u=g on ∂Ω,
has a (weak) solution if the domain Ω is bounded in one direction and g ∈ L2 (∂Ω, R).
Lectures on Applied PDEs, November 3, 2017 187

Thie result above deals with the functional (19.4) with G(x, u) ≥ 0. What happens
when G(x, u) ≤ 0? To see possible pitfalls, consider the functional
Z  
1 2 1 p λ 2 n
F (u) = |∇u| − |u| − |u| d x. (19.6)
Ω 2 p 2
However, for p > 2, this functional is not bounded from below. Indeed, take uµ = µu
with a fixed function u, and some µ > 0. Then

21 µp
Z Z
2 2
 n
F (uµ ) = µ |∇u| − λ|u| d x − |u|p → −∞, as µ → ∞.
2 Ω p Ω

Consequently, this functional does not have a minimizer. Taking u(µ) := µα u(µx) with a
fixed function u and some µ, we find

2+2α−n 1 αp−n 1 2α−n λ


Z Z Z
(µ) 2
F (u ) = µ |∇u| − µ p
|u| − µ |u|2 .
2 Ω p Ω 2 Ω
2
Take now α so that 2 + 2α − n > 0, and 2 + 2α − n > αp − n, i.e. p−2 > α > n2 − 1.
2
Since p−2 > n−2
2
(because p2 < n−2
2 n
+ 1 = n−2 ), this is possible. Then we get F (u(µ) ) → ∞
as µ → ∞, which shows that F is not bounded from above, and consequently, it has no
maximizer either.
We discuss below how to go around this problem.

The nonlinear eigenvalue problem. Now we show how to use the Key Theorem in
order to prove existence of solutions of differential equations. Let Ω be a smooth bounded
domain in Rn . For λ ∈ R and p > 2, we consider the problem
−∆u − |u|p−2 u = λu in Ω,
(19.7)
u = 0 on ∂Ω.
We want to prove existence of solutions of this boundary value problem. Denote λ1 the
lowest eigenvalue of −∆ on Ω with Dirichlet boundary conditions. We have the following
2n
Theorem 10. Let 2 < p < n−2 if n > 2 and 2 < p < ∞ if n ≤ 2. Then for any λ < λ1 ,
there is a positive (weak) solution to the problem (19.7).
Exercise 3. Prove the theorem above. (Hint: Minimize the functional
1
Z
|∇u|2 − λ|u|2 dn x,

E(u) = (19.8)
2 Ω
subject to the constraint J(u) = 1, where
1
Z
J(u) := |u|p dn x. (19.9)
p Ω
188 Lectures on Applied PDEs, November 3, 2017

Show that the solution to the corresponding Euler-Lagrange equation can be rescaled to
satisfy (19.7). Specifically, show that the Euler-Lagrange equation for the problem (19.8)
- (19.9) is
− ∆v − λv − µ|v|p−2 v = 0, (19.10)
and
1
Z
|v|p dn x = 1, (19.11)
p Ω

for some µ ∈ R. But now we have the undesirable coefficient µ. (The parameter µ is a
function of λ determined through equation (19.11)). To get rid of this coefficient, we first
show that µ > 0. Indeed, multiplying (19.10) by v, integrating the result over Ω, and
then integrating by parts, we obtain
Z Z
p n
|∇v|2 − λ|v|2 dn x.

µ |v| d x =
Ω Ω

The r.h.s. is Ω (v(−∆)v−λ|v 2 |) ≥ (λ1 −λ) Ω |v|2 , where, recall, λ1 is the lowest eigenvalue
R R

of the negative Laplacian, −∆, with Dirichlet boundary conditions. Since λ < λ1 , the
r.h.s is positive and so µ > 0. Now we rescale v as
1
u(x) := µ p−2 v(x),

then clearly u satisfies (19.7).)

Discussion. Differential equation (19.7) is the Euler–Lagrange equation for the func-
tional (19.6), which as was shown above, for p > 2, is unbounded below and above and,
consequently, does not have a minimizer or maximizer. To get out of this dilemma, we
consider instead the constraint problem (19.8) - (19.9).

Discussion. One can show that µ ↓ 0 as λ ↑ λ1 . This shows that a branch of non-trivial
solutions of (19.7) bifurcates from the trivial solution u ≡ 0 at λ = λ1 .
We give the following general result about w.l.s.c..

Theorem 11. If E is convex, then E is w.l.s.c..

Proof. By Proposition 10.6, we have E(un ) ≥ E(u) + dE(u)(un − u). By the weak conver-
gence we have dE(u)(un − u) → 0, as n → ∞, which, together with the previous relation,
implies that lim inf n→∞ E(un ) ≥ E(u).

V (x)|u|2
R
This result implies, in particular, that if C ≥ V (x) ≥ 0, then the functional Ω
is w.l.s.c. in L2 (Ω).
Lectures on Applied PDEs, November 3, 2017 189

w
Exercise 4. 1) Let V (x) ≥ 0. Show that if un −→u in L2 (Ω), then
Z Z
2
lim inf V (x)|un | ≥ V (x)|u|2 .
n→∞ Ω Ω

2) Let Ω be aR bounded domain in R , and let f ∈ L2 (Ω). Show that the functional
n
1
|∇u|2 dn x − Ω f u is coercive and w.l.s.c. on Hg1 (Ω).
R
2 Ω
3) Let Ω be a bounded domain in Rn , and let g(u) = (gij (u)) be a family of m × m
positive definite matrices satisfying g(u) ≥ δ1l, for some δ > 0. Show that the functional
1
Z X
E(u) = gij (u)∇ui · ∇uj dn x
2 Ω i,j

is coercive and w.l.s.c. on Hg1 (Ω, Rm ). Here, u = (u1 , . . . , um ) : Ω → Rm .


We give more examples of weakly closed sets.
(h) Let Ω ⊂ Rn be bounded and
Z Z
1 2
M := {u ∈ L (Ω) | |∇u| < ∞ and u = c}
Ω Ω
for some c ∈ R.
Proposition 3. M ⊂ H 1 (Ω) and is weakly closed there.

Proof. By passing from u(x) to u(x) − c we reduce our problem to the case of
2 2
R R
c = 0. By the Poincaré inequality, Theorem 13, Ω |u| ≤ C Ω |∇u| , we have that
M ⊂ H 1 (Ω).
Now, if un → u weakly in H 1 (Ω), then by the Kondrashov-Rellich theorem (see e.g.
[12] and Section A.6), un → u strongly in L2 (Ω) and therefore
R in L1 (Ω)R (remember
that Ω is a bounded domain). Hence, if un ∈ M , i.e., Ω un = 0, then u = 0, i.e.,
u ∈ M . Thus M is weakly closed in H 1 (Ω).

(i) Let Ω ⊂ Rn be a bounded domain, and let M ⊂ H 1 (Ω) be given by M = {u ∈


H 1 (Ω) : u(x0 ) = 0} for some x0 ∈ Ω (say x0 = 0).
Proposition 4. M is weakly closed in H 1 (Ω) for n = 1 and is not weakly closed
for n > 1.

Proof. Let first n = 1. Since Ω is bounded, we have by the Rellich–Kondrashov


w
theorem that if un −→u0 in H 1 (Ω), then there is a subsequence {un0 } s.t. un0 → u0
in C(Ω). If un ∈ M , then un (0) = 0, and therefore u0 (0) = limn0 →0 un0 (0) = 0.
Now let n ≥ 2. Assume for simplicity that Ω is a ball and take ϕn ∈ M , ϕn (x) =
f (n|x|) with f (r) = r for r ≤ 1 and f (r) = 1 for f ≥ 1. Then (see in the exercise
below) ϕn → 1 weakly in H 1 (Ω) as n → ∞.
190 Lectures on Applied PDEs, November 3, 2017

Exercise 5. Show the statement above.

The key theorem implies that the following functionals have minimizers on the specified
sets:
1. Ω ( 12 |∇u|2 + G(x, u)) on M , if G(x, u) ≥ 0, where M is either Hg1 with Ω bounded
R

in one direction or
Z
{u ∈ H01 : |u|p dn x = 1}


2n m if m > 0
with Ω bounded and 2 ≤ p < , m+ = .
(n−2)+ 0 if m ≤ 0
1
gij (u)∇ui · ∇uj on Hg1 (Ω, Rm ), provided Ω is bounded in one direction, and
R P
2. 2 Ω ij
g(u) = (gij (u)) ≥ δ1l, for some δ > 0.
The Euler-Lagrange equations for the variational problems in the examples 1 and 2
above are
−∆u + G0 (x, u) + λ|u|p−2 u = 0 and
X
gij0 (u)∇ui · ∇uj + div(aij (u)∇u) = 0,

ij

respectively, where λ is a Langrange multiplier, G0 (x, u) = ∂u G(x, u) and gij0 (u) =


∂u gij (u). Existence of minimizers yields existence of weak solutions of the above equations
in a bounded Ω ⊂ Rn .??
We have the following special cases for example 1:
i) If G has a minimum at u0 which is independent of x, then u0 (x) ≡ u0 is a minimizer:
E(u0 ) = 0,

ii) G(x, u) = V (x)|u|2 for some V (x) ≥ 0, i.e. G is quadratic (remember that in
this case the equation for the critical points of E is linear !). We can write E(u) on
H0 (Ω) as E(u) = Ω u(− 12 ∆+V (x))u, i.e. E is the quadratic form of the Schrödinger
2
R

operator − 21 ∆ + V (x).
Exercise 6. Prove existence of (weak) solutions of the following boundary value problems
(below, Ω is a bounded domain in Rn ):
(a) the nonlinear eigenvalue problem:

∆u + a(x)|u|p−1 u = µu in Ω,
u = 0 on ∂Ω,
2n
where a(x) is a smooth and positive function on Ω, n ≥ 3, 2 < p < n−2
and µ ≥ 0;
Lectures on Applied PDEs, November 3, 2017 191

(b) the nonlinear Dirichlet problem:


∇ |∇u|2 ∇u = f

in Ω,
u = 0 on ∂Ω,
for any f ∈ L4/3 (Ω). (Hint: reduce this to a minimization problem on the Sobolev
space Z
4,1 4
W0 (Ω) = {u ∈ L (Ω) : |∇u|4 < ∞ and u|∂Ω = 0}).

How to gain smoothness: elliptic regularity. Assume we show that the following
equation has a (weak) solution in H 1 (Ω):
∆u = a(x)u4 in Ω,
and u = 0 on ∂Ω (Dirichlet boundary conditions). Here, a is smooth and Ω ⊂ R3 is
bounded. This is not so good since u4 does not below to a good space and ∆u ∈ H −1 (Ω).
But it turns out that in fact u is smooth!
We can show this in the following way, which we just sketch. By the Sobolev embedding
2n
theorem (i.e. H 1 (Ω) ⊂ Lα (Ω) with α < n−2 = 6 for n = 3), we have that u ∈ Lα (Ω) with
α < 6. Hence u4 ∈ Lβ (Ω) with β < 3/2, so ∆u ∈ Lβ (Ω) since a is smooth. Then by the
Sobolev embedding theorem, u ∈ Lα (Ω) for any α < ∞. Iterating this we obtain that u
is bounded and differentiable.

Poincaré inequality
Theorem 12 (Poincaré inequality). Let Ω have a diameter d < ∞ in some direction (i.e.
it is possible to place Ω between two parallel hyperplanes at a distance d from each other).
Then for any u ∈ H01 (Ω), we have
Z Z
2 2
|u| ≤ (2d) |∇u|2 . (19.12)
Ω Ω

Proof. We can assume that this hyperplanes are {x1 = 0} and {x1 = d}. Assume u is
real and estimate

∂ ∂u ∂u
Z Z Z Z
2 2 2
kuk2 = 1 · |u| = − x1 |u| = −2Re x1 u∗ ≤ 2d |u| .
Ω Ω ∂x1 Ω ∂x1 Ω ∂x1
Applying now the Schwartz inequality to the integral on the r.h.s., we obtain

2
∂u
kuk2 ≤ 2dkuk2 ∂x1 ≤ 2dkuk2 k∇uk2 ,

2

where k∇uk22 = Ω |∇u|2 = Ω n1 | ∂x∂u 2


R R P
n
| . The latter inequality implies kuk2 ≤ 2dk∇uk2 .
192 Lectures on Applied PDEs, November 3, 2017

We mention without proof, the following variant of the result above.


1
R
Theorem 13. Let Ω be bounded and ū := |Ω| Ω
u. Then
Z Z
2
|u − ū| ≤ C |∇u|2 .
Ω Ω

19.3 Existence of ground state of nonlinear Schrödinger equa-


tion (without and with potential)
*(under construction)* (see [20], Introduction and Section II, and also [?], Sections 1
( in particular, Theorem 1 i) - ii)) and 3 and Appendix)

20 Interfaces, vortices, vortex lattices and harmonic


maps
20.1 Allen-Cahn energy functional and interfaces
The difference in free energy of two phases of the same substance or of two substances in
a domain Ω ⊂ Rd can be often described by the following simple functional:
Z  
1 2
E(u) = |∇u| + λG(u) dd x, (20.1)
Ω 2

where u : Ω → R, λ > 0, and G(u) ≥ 0, has two strict, non-degenerate minima at u = 1


and u = −1, and G(u) → ∞ as |u| → ∞. In other words, G is of the form of a double-well
potential:

The typical and most important example is G(u) = 41 (|u|2 − 1)2 . This functional is called
G

u
−1 1

Figure 20: Double well potential.

the Allen - Cahn (or Ginzburg-Landau) energy functional. It plays an important role in
many areas of sciences and engineering. Ω is a large domain, say a box with sides of size
2L or a ball of radius L or Rd .
Lectures on Applied PDEs, November 3, 2017 193

Note that E(ϕ) ≥ 0 and has two absolute minimizers, −1 and +1, so that E(−1) =
E(+1) = 0. These minimizers correspond to two homogeneous substance or phases, which
we will call the −1 and +1 phase. We are interested in minimizers for which these phases
co-exist. To obtain such minimizers we to impose constraints and/or boundary conditions.
For instance, if the total amount of the −1 phase is fixed and Ω = Rd , then imposing
Z
(1 − u(x))dd x = α, (20.2)

would guarantee that the total amount of the −1 phase is finite and determined by α.
Before addressing these questions, we mention that the Euler-Lagrange equation for
critical points of E(u) is
− ∆u + λG0 (u) = µ, (20.3)
where µ 6= 0 if condition (20.2) is imposed. In the latter case µ is the corresponding
Lagrange multiplier and is chosen so that (20.2) holds. For µ 6= 0, this is the static
Cahn-Hilliard equation and for µ = 0, the static Allen - Cahn equation (also known as
the Ginzburg-Landau equation).

Exercise 7. Find the Euler-Lagrange equation for the functional (20.1) with side condi-
tion (20.2).

Now we apply the direct method of variational calculus in order to find minimizers
of the Allen - Cahn energy functional (20.1), with the function G(u) described in the
paragraph after (20.1).

Planar interface. We want to consider the situation where the minimizers describe
the planar interface. Assume this interface is the plane {x1 = 0} (by a translation and
a rotation, we can always reduce to this case). Then our unknown u depends on one
variable only - on x = x1 - and the problem becomes the one dimensional one and the
functional (20.1) becomes
Z  
1 2
E(u) = |∇u| + λG(u) dx, (20.4)
R 2

and can interpreted as the energy per unit area of the interface {x1 = 0}.
We have four distinct boundary conditions (BC): u(x1 ) → ±1 as x1 → ∞ and u(x1 ) →
±1, as x1 → −∞. Consequently, we are led to consider E on the following four spaces:

M±,± = {u ∈ H 1 (Ω) : one of the above BC holds}.

Clearly, E attains its strict minimum on M+,+ and M−,− at u+ (x1 ) ≡ 1 and u− (x1 ) ≡ −1,
respectively. In the phase separation model described by the Allen - Cahn (Ginzburg–
Landau) functional, these minimizers describe homogeneous phases. Next, a minimizer
194 Lectures on Applied PDEs, November 3, 2017

on M+,− is obtained from a minimizer on M−,+ by reflection u(x1 ) → u(−x1 ). Thus it


suffices to consider only minimizers on M−,+ . Observe that since E has the reflection
symmetry, we can simplify our task by looking for odd minimizers, i.e. we pass from
M−,+ to
odd
M−,+ := {u ∈ M−,+ : u(−x1 ) = −u(x1 )}.
To fix ideas, we let G(u) = 41 (u2 − 1)2 . Our result here is the following
odd
Theorem 14. The Ginzburg-Landau energy functional (20.4), defined on M−,+ , has a
minimizer.
Proof. We identify M−,+ with the space M−,+ = χ + H01 (R), where χ is a smooth (and
odd) function satisfying χ(x) = 1 for x ≥ δ and χ(x) = −1 for x ≤ −δ. Then, since
H01 (R) is weakly closed, M−,+
odd
is weakly closed in H 1 (R).
Now, we would like to show that E on Hg1 (R) is w.l.s.c. and coercive, and therefore
odd
on M−,+ . Since G ≥ 0, the w.l.s.c. follows from Corollary 8.
(g)
The coercivity of E(u) on Mα is a more subtle question. Define v := χ − u ∈ H01 (R).
We write    2
1 2 2 2 1 2 1
G(u) = (1 − χ ) − 2(1 − χ )v χ − v + v χ − v .
4 2 2
Assume |v|R≤ 1 for |x| ≥ Rδ. Then using Rthat χ(x) = ±1 for |x| ≥ δ, we find G(u) ≥ 14 v 2 .
This gives G(u)dx ≥ 41 |x|≥δ v 2 dx ≥ 41 R v 2 dx − Cδ. Therefore
1 1
Z
E(1 − v) ≥ (|∇v|2 + v 2 )dx − C ≡ kvk2H 1 − C.
4 R 4
This implies the coercivity of E on Hg1 (R).
Now, by the properties established above and the key theorem on minimization, E(u)
has a minimizer u∗ on the set
odd odd
M̃−,+ := {u ∈ M−,+ | |χ − u| ≤ 1}.
odd
Is this minimizer also a minimizer on M−,+ ? Observe that
E(u) ≥ E(max(min(u, 1), −1)).

+1

-1

u maxminu
Lectures on Applied PDEs, November 3, 2017 195

Indeed, one can see that replacing the function u by the function max(min(u, 1), −1)
decreases both gradient and potential term. Hence, if {un } is a minimizing sequence, then
so is {wn := max(min(un , 1), −1)} and the minimizer, w∗ , satisfies |w∗ | ≤ 1. Similarly, we
can show that
E(u) ≥ E(w), w := |u|χx≥0 − |u|χx<0 .
So we have w∗ ≥ 0 for x ≥ 0 and w∗ ≤ 0 for x ≤ 0. The last two properties imply
odd odd
|χ − w∗ | ≤ 1. Hence a minimizer on M̃−,+ belongs to and therefore M−,+ is also a
odd
minimizer on M−,+ .
The minimizers w∗ (x) are called kinks and w∗ (−x), anti-kinks. In what follows, we
denote them by χ(x) are and χ(−x). They describe planar interfaces. Of course, by
shifting χ(x) to χ(x − h), we obtain a one–parameter family of minimizers, the kinks
centred at different points of R. The kinks are solutions the Euler - Lagrange equation
for (20.4), i.e. (20.3), with µ = 0, the static Allen - Cahn equation,

− ∆χ + λG0 (χ) = 0. (20.5)

Lamellar phase. In this situation, layers of oil and water coexist in a periodic array.
To get a solution to (20.3), glue together a kink at z1 and an antikink at z2 . There is
no exact solution of this form: the kink and the antikink interact at any distance. They
attract each other and as a result, they move toward each other and collapse. This means
that if ϕ is a function consisting of a kink and an antikink glued together at a distance R,
then E(ϕ) is monotonically increasing as a function of R. Here, ϕ is a function consisting
of a kink and an antikink glued together at a distance R. However, presumably one can
construct a periodic solution corresponding to an array of kinks and antikinks.

C  C  C 
C  C  C 
C  C  C 
C  C  C 
C  C  C 
C  C  C 

Idea of proving the existence of the lamellar solutions. Consider the variational prob-
lem of minimizing the functional
Z c 
1 2
E(u) = |∇u| + λG(u) dx, (20.6)
−c 2
196 Lectures on Applied PDEs, November 3, 2017

Rc
under constraint that −c udx = α, i.e. on the space {u ∈ H 1 ([−c, c]) : u(±c) =
Rc
1, −c udx = α, for α ∈ (−2c, 2c). For α = 2c, the minimizer is trivial, u∗ = 1, and
for α = −2c, there is no minimizer. The parameters a and b satisfy a + b = c and their
ration is determined by α. Then the minimizer u∗ is extended periodically to the entire
real line.

Spherical drops and cylinder solutions. Here Ω is either the ball, BR , of radius R,
centered at the origin, or the cylinder, CR of radius R or Rd , d = 2, 3. We minimize the
energy functional E, as was described above, on the set
Z
(g) 1
Mα := {u ∈ Hg (Ω) | (1 − u) = α}

where g ≡ 1 on ∂B, for some α > 0.

Theorem 15. Let G(u) = 14 (u2 − 1)2 and Ω be bounded. The Ginzburg-Landau energy
(g)
functional E(u) defined on Mα , α > 0 and g ≡ 1, has a minimizer.
The proof of this theorem is similar to the one of Theorem 14.
(g)
Exercise 8. Check that the set Mα is weakly closed and that the functional E(u) defined
(g)
on Mα is w.l.s.c.
One can show that this minimizer, u∗ , is smooth. What is the shape of u∗ ? Is it
spherically symmetric? One can show that the minimizer must have exactly one zero, i.e.,
it is of the second type as shown on the figure below.

+1 +1
 C  C 
 C  C 
 C  C or 
 C  C \
 C  C  \
-1  C  C -1  R

The parameter R - the radius of the drop - is found from the condition (20.2).
One can also show that u∗ is spherically symmetric. If one does not want to work
hard, then one can look directly for spherically symmetric minimizers, i.e. minimizers of
E(u) of the form
u(x) = ψ(|x|), (20.7)
subject to the side condition (20.2). But then one would have only minimizer among
spherically symmetric functions.
Lectures on Applied PDEs, November 3, 2017 197

20.2 Vortices in Superfluids


Macroscopically, equilibrium states of superfluids and Bose-Einstein condensates are de-
scribed by a function ψ : Ω → C, Ω ⊂ Rd called the order parameter, which satisfies the
nonlinear differential equation

− ∆ψ + (|ψ|2 − 1)ψ = 0 in Ω (20.8)

called the Gross-Pitaevskii or Ginzburg-Landau equation, with the boundary condition

|ψ| → 1 as |x| → ∂Ω. (20.9)

In 1958, V.L Ginzburg and L. Pitaevskii conjectured that for d = 2 and Ω = R2 (the
common ‘cylindrical’ geomemtry) this equation has solutions of the form

ψn (x) = fn (r)einθ (20.10)

where (r, θ) are polar coordinates of x ∈ R2 and n is an integer. It was conjectured


that these solutions describe vortices observed in superfluids. (In 1947, L. Onsager has
conjectured that these vortices are analogous to normal fluid vortices, except for the fact
that the vortices of the superfluid were quantized while normal vortices were arbitrary.)

The rigorous result about existence and uniqueness of vortices came much later and
is stated in the following

Theorem 16. Let d = 2 and Ω be either BR or R2 . Then, for all n, there exists a solution
of form (20.10) unique up to symmetry transformation. The function fn (r) can be taken
to be positive and it vanishes at r = 0 as

fn (r) = arn

for some a > 0 and is monotonically increasing to 1.

fn
psin

r x

Portraits of vortices:
198 Lectures on Applied PDEs, November 3, 2017

s1 s2 nt s1

We prove this theorem by reformulating the b.v problem (20.8) - (20.9) as a variational
problem of finding a minimizer of an appropriate functional. For a bounded domain Ω,
(20.8) is the Euler-Lagrange equation for the functional

1 1
Z
E(ψ) = |∇ψ|2 + (|ψ|2 − 1)2 ,
Ω 2 4

called the Gross-Pitaevskii or Ginzburg-Landau energy functional, on the space


Z
1
M = {ψ ∈ Hg (Ω) | (|ψ|2 − 1)2 < ∞}

for some function satisfying |g| = 1. For Ω = BR , we choose

g(x) = einθ

Renormalized Ginzburg-Landau functional. For d = 2 and Ω unbounded, say,


Ω = R2 , E(ψ) = ∞ for functions ψ of interest. We explain this more carefully for the
most important case of Ω = R2 .
For C 1 complex functions (vector-fields) ψ satisfying the boundary condition |ψ| → 1
as |x| → ∞, we define the degree by

1
Z
deg ψ = d(arg ψ)
2π |x|=R

for R sufficiently large. Here arg ψ = 1i ln(ψ/|ψ|).

Exercise 9. Show that if f (r) → 1 as r → ∞, then deg(f (r)einθ ) = n.

It is shown in [?] that if ψ : R2 → C is a C 1 vector-field such that |ψ(x)| → 1 as


|x| → ∞ and degψ 6= 0, then E(ψ) = ∞. Thus the Ginzburg-Landau energy functional
E(ψ) is not defined in infinite domains in the most interesting cases (i.e., in the presence
of vortices). To go around this problem, [?] have introduced the renormalized Ginzburg-
Landau energy functional defined as follows
Lectures on Applied PDEs, November 3, 2017 199

1 (degψ)2 1
Z
Eren (ψ) = |∇ψ|2 − 2
χ + (|ψ|2 − 1)2
2 r 2
where χ is a cut-off function with the properties: χ ∈ C ∞ (R2 ) and

1 for |x| ≥ 2,
χ(x) =
0 for |x| ≤ 1.
It is shown in [?] that Eren (ψ) is defined on a large class of functions which include n–
vortices and their combinations and that

dEren (ψ) = −∆ψ + (|ψ|2 − 1)ψ,

i.e. Eren (ψ) produce the correct Euler-Lagrange equations (20.8). However, as should be,
the functional Eren (ψ) is not bounded below for |deg ψ| ≥ 2. Indeed, consider a function
ψ describing n := deg ψ ≥ 2 single vortices. Then moving these vortices apart to infinity
would decrease Eren (ψ) indefinitely. The energy functional Eren (ψ) is suitable and natural
for the variational approach to the vortex problem in an unbounded domain.
Now, we can state the following variational problem: Minimize Eren (ψ) among func-
tions ψ with a fixed degree deg ψ = n. This the variational problem with the topological
constraint.
For simplicity, we prove the theorem only in the case Ω = BR . A proof in the case
Ω = R2 uses the renormalized Ginzburg-Landau energy and can be found in [?]. We prove
a somewhat stronger statement.

Proof of Theorem 16 for Ω = BR . We minimize the functional E(ψ) on the functions of


the form ψ = f (x)einθ for some fixed n, where f (x) is a real valued function satisfying
the boundary condition f (x) = 1 for |x| = R. In other words, we minimize the functional

En (f ) := E(f einθ )

on a space of functions f (r) satisfying f (x) = 1 for |x| = R.


In what follows, we abbreviate BR = B. Write out this functional explicitly:

1 n2 2 1 2
Z
En (f ) = ( |∇f | + 2 f + (f − 1)2
2
B 2 2r 4

where we have used that


∇(f einθ ) = (∇f + in∇θ)einθ ,
and so for f real
|∇(f einθ )|2 = |∇f |2 + n2 |∇θ|2 f 2 ,
and that |∇θ| = 1/rn .
200 Lectures on Applied PDEs, November 3, 2017

Lemma 4. For all n, en (f ) has a minimizer on the set

M = {f real and En (f ) < ∞}.

and all minimizers are radially symmetric.

Proof. Let ξ = 1 − f and define the functional F (ξ) = en (1 − ξ). Explicitly this functional
is given as
1
Z
F (ξ) = |∇ξ|2 + G(ξ)
B 2
where
n2 1
G(ξ) =2
(1 − ξ)2 + ξ 2 (1 − ξ)2 .
r 2
Observe that ξ → 0 as |x| → R. We consider F (ξ) on the set

M = {ξ ∈ H01 (B) | ξ real, G(ξ) < ∞}.

Since G(ξ) ≥ 0, we have that F (ξ) is w.l.s.c. on H01 (B). The functional F (ξ) is also
coercive by the difficult part of Theorem ?? One can avoid using this difficult part (which
follows if B = R2 ) proceeding as follows: we have

1
G(ξ) ≥ ξ 2 if ξ ≤ 1
4
and therefore F (ξ) is coercive under the additional condition that ξ ≤ 1. Now we can
show that we can make this condition automatically satisfied for minimizing sequences.
Indeed, let ξn be a minimizing sequence for the functional F (ξ). Define a new sequence
{ξn0 (x) := min(ξn (x), 1)}. Clearly,
xn xnp

|∇ξn | ≥ |∇ξn0 | and G(ξn ) ≥ G(ξn0 )

Exercise 10. Check these statements.


Lectures on Applied PDEs, November 3, 2017 201

Hence F (ξn ) ≥ F (ξn0 ) so that {ξn0 } is also a minimizing sequence. Of course, ξn0 ≤ 1
so F (ξ) is coercive on this sequence and the proof of the Key Theorem can be modified
so as to show that the functional F (ξ) has a minimizer on the set M .
Let now f0 be a minimizer of the functional En (f ). We show that f0 must be radially
symmetric. Introduce the function
1
u = (f02 ) 2 ,
R 2π
where g(r) := 0 g(r, θ) 2π dθ
. Then f02 = u2 and

f04 ≥ u4 and |∇r f0 |2 ≥ |∇r u|2 . (20.11)

Indeed, the first of these inequalities follows from the Cauchy-Schwartz inequality and the
second is obtained as follows: for r such that u(r) 6= 0,
2
Z 1/2 2 R 2π f ∇ f dθ R 2π 2 dθ R 2π

dθ 0 0 r 0 2π f |∇r f0 |2 2π

∇ r 2
f0 (r, θ) = ≤ 0 0 2πR 2π0 2 dθ

R 2π 2 dθ
0 2π f0 f0
0 2π 0 2π

by the Cauchy-Schwartz inequality again. Inequality (20.11) implies that


Z Z ∞
2 2
|∇f0 | d x ≥ |∇r u|2 rdr
0

with the equality taking place only if f0 is independent of θ. Hence

En (f0 ) ≥ En (u)

and the equality holds only if f0 is radially symmetric. Since f0 is a minimizer, we can
conclude that it is radially symmetric. We omit the proof of monotonicity of f0 and refer
the reader to [?] for this proof.

Since f0 is a minimizer of En (f ) and is radially symmetric, it satisfies the Euler-


Lagrange equation
n2
−∆r f + 2 f + (f 2 − 1)f = 0 in Ω,
r
where ∆r is radial Laplacian in R2 : ∆r = 1r ∂r (r ∂r

). Since f0 is radially symmetric, we
have that ∇f0 · ∇θ = 0. This together with the relations ∆θ = 0 and |∇θ| = 1/r implies

n2
∆(f0 einθ ) = (∆r f0 − f0 )einθ .
r2
Therefore, the function ψn (x) = f0 (r)einθ satisfies Ginzburg-Landau equations (20.8) and
(20.9).
202 Lectures on Applied PDEs, November 3, 2017

20.3 The Ginzburg-Landau Equations


(See also Section 13, we repeat some definition from that section.) The Ginzburg-
Landau theory gives a macroscopic description of superconducting materials and serves
as an integral and paradigmatic part of particle physics. It is formulated in terms of a
pair (Ψ, A) : Rd → C × Rd , d = 1, 2, 3, satisfying the system of nonlinear PDE called the
Ginzburg-Landau equations:
−∆A Ψ = κ2 (1 − |Ψ|2 )Ψ
(20.12)
curl2 A = Im(Ψ̄∇A Ψ)
where ∇A = ∇ − iA, and ∆A = ∇2A , the covariant derivative and covariant Laplacian,
respectively, and κ > 0 is a parameter, coupling constant. For d = 2, curl A := ∂1 A2 −∂2 A1
is a scalar, and for scalar B(x) ∈ R, curl B = (∂2 B, −∂1 B) is a vector.
Superconductivity. The complex-valued function Ψ(x) is called an order parameter, |Ψ(x)|2
gives the local density of (Cooper pairs of) superconducting electrons, and the vector field
A(x)is the magnetic potential, so that B(x) := curl A(x) is the magnetic field. The param-
eter κ > 0 is called the Ginzburg-Landau parameter, it depends on the material properties
of the superconductor. The vector quantity J(x) := Im(Ψ̄∇A Ψ) is the superconducting
current. (See eg. [19, 18]).
Particle physics. In the Abelian-Higgs model, ψ and A are the Higgs and U (1) gauge
(electro-magnetic) fields, respectively. Geometrically, one can think of A as a connection
on the principal U (1)-bundle Rd × U (1), d = 2, 3.
Cylindrical geometry. In the commonly considered idealized situation of a superconductor
occupying all space and homogeneous in one direction, one is led to a problem on R2 and
so may consider Ψ : R2 → C and A : R2 → R2 . This is the case we deal with in this
contribution.

Symmetries of the equations The Ginzburg-Landau equations (20.12) admit several


symmetries, that is, transformations which map solutions to solutions.
Gauge symmetry: for any sufficiently regular function γ : R2 → R,
Tγgauge : (Ψ(x), A(x)) 7→ (eiγ(x) Ψ(x), A(x) + ∇γ(x)); (20.13)
Translation symmetry: for any h ∈ R2 ,
Thtrans : (Ψ(x), A(x)) 7→ (Ψ(x + h), A(x + h)); (20.14)
Rotation symmetry: for any ρ ∈ SO(2),
Tρrot : (Ψ(x), A(x)) 7→ (Ψ(ρ−1 x), ρ−1 A((ρ−1 )T x)), (20.15)
One of the analytically interesting aspects of the Ginzburg-Landau theory is the fact
that, because of the gauge transformations, the symmetry group is infinite-dimensional.
Lectures on Applied PDEs, November 3, 2017 203

Abrikosov lattices In 1957, A. Abrikosov discovered a class of solutions, (Ψ, A), to


(13.1), presently known as Abrikosov lattice vortex states (or just Abrikosov lattices),
whose physical characteristics, density of Cooper pairs, |Ψ|2 , the magnetic field, curl A,
and the supercurrent, JS = Im(Ψ̄∇A Ψ), are double-periodic w.r. to a lattice L. By a
lattice we understand here the Bravais lattice, i.e.
L = {mν1 + nν2 : m, n ∈ Z},
where (ν1 , ν2 ) is a basis of L.
Denote by Ω a fundamental cell of the lattice L, say {xν1 + yν2 : x, y ∈ [0, 1)},
R where
(ν1 , ν2 ) is a basis of L. For Abrikosov states, for (Ψ, A), the magnetic flux, Ω curl A,
through a lattice cell, Ω, is quantized,
Lemma 20.1.
1
Z
curl A = deg Ψ = n, (20.16)
2π Ω
for some integer n.
Proof. The periodicity of ns = |Ψ|2 and J = Im(Ψ̄∇A Ψ) imply that ∇ϕ − A, where

R = |Ψ|e , His periodic,
Ψ H provided Ψ 6= 0 on ∂Ω. This, together with Stokes’s
R theorem,

curl A = ∂Ω
A = ∂Ω
∇ϕ and the single-valuedness of Ψ, imply that Ω
curl A = 2πn
for some integer n.
Using the reflection symmetry of the problem, one can easily check that we can always
assume n ≥ 0.

Abrikosov lattices as gauge-equivariant states. We say a state (Ψ, A) is gauge


- equivariant (with respect to a lattice L, or L−equivariant) iff there exists (possibly
multivalued) function ηs : R2 → R, s ∈ L, such that
Tstrans (Ψ, A) = Tηgauge
s
(Ψ, A). (20.17)
A key point in proving both theorems is to realize that a state (Ψ, A) is an Abrikosov
lattice if and only if (Ψ, A) is gauge - equivariant.
Lemma 20.2. A state (Ψ, A) is an Abrikosov lattice if and only if (Ψ, A) is gauge-
equivariant.
Proof. If state (Ψ, A) satisfies (20.17), then all associated physical quantities are L−periodic,
i.e. (Ψ, A) is an Abrikosov lattice. In the opposite direction, if (Ψ, A) is an Abrikosov
lattice, then curl A(x) is periodic w.r.to L, and therefore A(x + s) = A(x) + ∇ηs (x),
for some functions ηs (x). Next, we write Ψ(x) = |Ψ(x)|eiφ(x) . Since |Ψ(x)| and J(x) =
|Ψ(x)|2 (∇φ(x) − A(x)) are periodic w.r.to L, we have that so is ∇φ(x) − A(x), which, to-
gether with the relation A(x+s) = A(x)+∇ηs (x), gives that ∇φ(x+s) = ∇φ(x)+∇ηs (x),
which implies that φ(x + s) = φ(x) + ηs (x) + cs , for some constants cs .
204 Lectures on Applied PDEs, November 3, 2017

Since Tstrans is a commutative group, we see that the family of functions gs has the
important cocycle property
ηs+t (x) − ηs (x + t) − ηt (x) ∈ 2πZ. (20.18)
This can be seen by evaluating the effect of translation by s + t in two different ways. We
call gs (x) the gauge exponent.

Ginzburg-Landau energy. The Ginzburg-Landau equations (20.12) are the Euler-


Lagrange equations for critical points of the Ginzburg-Landau energy functional (written
here for a domain Q ∈ R2 )
κ2
Z  
2 2 2 2
EQ (Ψ, A) := |∇A Ψ| + (curl A) + (|Ψ| − 1) . (20.19)
Q 2

Superconductivity: In the case of superconductors, the functional E(ψ, A) gives the dif-
ference in (Helmholtz) free energy (per unit length in the third direction) between the
superconducting and normal states, near the transition temperature.
Particle physics: In the particle physics case, the functional E(Ψ, A) gives the energy of
a static configuration in the U (1) Yang-Mills-Higgs classical gauge theory.

Existence of Abrikosov lattices


Theorem 20.3. For any lattice L, there exists a smooth Abrikosov lattice solution u∗ =
(Ψ∗ , A∗ ) (i.e. satisfying (20.17)) for this L.
Proof. We want to solve the Ginzburg-Landau equations (20.12) for functions satisfying
the condition (20.17). Notice that these equations are the Euler-Lagrange equations for
the energy functional EQ (Ψ, A), see (20.19), on the space H 1 (Ω), satisfying (20.17).
We say two states u0 and u are gauge-equivalent, iff there is differentiable real function
χ s.t. u0 = Tχgauge u. We begin with the following statement whose proof can be found in
[?]
Lemma 20.4. Any L-equivariant state, (Ψ0 , A0 ), is gauge-equivalent to a L-equivariant
state, (Ψ, A), satisfying div A = 0.
Now, we identify the quotient R2 /L with a fundamental cell, Ω, of the lattice L and
introduce the space
1 1
Hequiv (Ω) := {u ∈ Hloc (R2 ) : u satisfies (20.17) and div A = 0},
1
equipped with the norm kukHequiv
1 (Ω) := kukH 1 (Ω) . The space Hequiv (Ω) is weakly closed in
H 1 (Ω). Indeed, if un → u∗ weakly in H 1 (Ω), then
un → u∗ pointwise on Ω. (20.20)
Lectures on Applied PDEs, November 3, 2017 205

Using that R2 = ∪s∈L (Ω + s), we define u∗ on R2 by


u∗ (x + s) := es (x)u∗ (x), ∀x ∈ Ω, s ∈ L, (20.21)
where es (x) := Tggauge
s
. By (20.17), es (x) obeys et (x + s)es (x) = es+t (x). Since
u∗ (x + s + t) = es+t (x)u∗ (x) = et (x + s)es (x)u∗ (x) = et (x + s)u∗ (x + s),
which can be rewritten as u∗ (y + t) = et (y)u∗ (y), ∀y ∈ Ω + s, s, t ∈ L, we see that u∗
satisfies (??). The equations (20.20) and (20.21) imply un (x) → u∗ (x) pointwise on R2 .
Moreover, since un → u∗ weakly in H 1 (Ω) and div An = 0, we have that div A∗ = 0. Thus
1
u∗ ∈ Hequiv (Ω).
Now, we analyze EΩ (Ψ, A). Since div A = 0, (|Ψ|2 − 1)2 = |Ψ|4 − 2|Ψ|2 + 1 ≥ 21 |Ψ|4 − 1
and Ω has a finite area, |Ω| < ∞, the energy EΩ (Ψ, A) is obviously coercive,
κ2 4 κ2
Z  
2 2
EΩ (Ψ, A) ≥ |∇A Ψ| + (curl A) + |Ψ| − |Ω|
Ω 4 2
2
≥ ck(Ψ, A)kH 1 (Ω) − C|Ω|, (20.22)
2
for come constants c, C > 0. Finally, since G = κ2 (|Ψ|2 − 1)2 ≥ 0, the w.l.s.c. follows from
1
Corollary 8. Hence EΩ (Ψ, A) has a minimizer on the space Hequiv (Ω) and this minimizer
satisfies the Ginzburg-Landau equations (20.12).

20.3.1 Appendix I. Relation to line bundles


The equivariant pairs u = (Ψ, A) are related to sections and connections on the line
bundle, E, defined by the automorphy exponents gs (x) as
E := R2 × C/L, (20.23)
where L acts on R2 × C and on R2 as s : (x, Ψ) → (x + s, eigs (x) Ψ), and A’s, as connection
one-forms on E, with the connections ∇A , defined on [Ψ(x)]’s as ∇A : X → (∇A )X , where
X is a vector field on the base manifold R2 /L, (∇A )X [Ψ] := [(dX(x) − iA(X)(x))Ψ(x)],
and the equivalence class [∇A Ψ] is defined by the relation ∇A Ψ ∼ ∇A0 Ψ0 ⇔ ∃χ : R2 →
G : (Ψ0 , A0 ) = Tχgauge (Ψ, A). (We will use the notation ∇A [Ψ] := [∇A Ψ] = [(d − iA)Ψ].)
Line bundles have a topological invariant - the degree, which is defined as the Chern
number, c(e) = c(η) of the automorphy exponents gs (x),
1  
c(η) := ην2 (x + ν1 ) + ην1 (x) − ην1 (x + ν2 ) − ην2 (x) , (20.24)
2πi
where {ν1 , ν2 } is a basis of L. (The degree of line bundles can be also defined through
devisors or zero of sections; the expression (20.24) is convenient for our purposes.) The
next result shows that c(g) are topological invariants and are related to the quantization
of the magnetic flux.
206 Lectures on Applied PDEs, November 3, 2017

Proposition 20.5. • (characteristic class) The quantity (20.24) is independent of


x and of the choice of the basis {ν1 , ν2 } and is an integer. Thus (20.24) is the
topological invariant classifying Abrikosov lattice states.
1
R
• (magnetic flux quantization) The quantity 2π Ω
curl A (magnetic flux) is an integer
and is equal to c(η) (magnetic flux quantization),

1
Z
curl A = c(η), (20.25)
2π Ω

where A is the corresponding connection and Ω is a fundamental lattice cell.

Proof. By the relation (20.18), ην2 (x + ν1 ) + ην1 (x) − ην1 +ν2 (x) ∈ 2πZ and ην1 (x + ν2 ) +
ην2 (x) − ην1 +ν2 (x) ∈ 2πZ. Subtracting the second relation from the first shows that c(η)
is independent of x and is an integer.
To prove the second statement,
R we note
R that, by Stokes’ theorem, the magnetic flux
through a lattice cell Ω is Ω curl A = ∂Ω A. Now, using the definition of the gauge
transformation, (20.13), and the condition (20.17), we obtain
Z Z 1  
A= ν1 · (A(aν1 + ν2 ) − A(aν1 )) − ν2 · (A(aν2 + ν1 ) − A(aν2 )) da (20.26)
∂Ω 0
Z 1  
= (ν1 · ∇ην2 (aν1 )) − (ν2 · ∇ην1 (aν2 )) da. (20.27)
0
R1 R1
Next, the relation x · ∇η(ax) = ∂a η(ax), gives 0 ν · ∇η(aν)da = 0 ∂a η(aν))da = η(ν) −
η(0), which yields
Z
A = ην2 (ν1 ) − ην2 (0) − ην1 (ν2 ) + ην1 (0),
∂Ω

which, by (20.24), gives (20.25).

20.4 Harmonic maps


The harmonic map or the stationary sigma model (see (3.1)) is a map from a d−dimensional
euclidean space-time, Rd , with the euclidean metric η = (δij ), to a Riemannian manifold,
N , with a metric (gab ), which is a critical point of the energy functional, given by

1
Z
E(Φ) := h∂j Φ, ∂ j Φi. (20.28)
2 Rd

Here h∂j Φ, ∂ j Φi is the Riemann scalar product in N , which, in local coordinates, is


gab ∂j Φa ∂ j Φb , as usual, ∂j = ∂x∂ j and ∂ i := η ij ∂j , and we assume the summation over
Lectures on Applied PDEs, November 3, 2017 207

repeated indices i, j = 1, . . . , d, a, b, c = 1, . . . , dim(N ), i, j = 1, . . . , d. With this nota-


tion, we can write energy the energy functional as,
1
Z
E(Φ) := gab ∂i Φa ∂ i Φb .
2 Rd
The Euler-Lagrange equation for critical points of E(Φ), i.e. for harmonic maps is obtained
from (16.18) by restricting the indices to i = 1, . . . d, which gives
∆Φa + Γabc (Φ)∂i Φb ∂ i Φc = 0, (20.29)
where ∆ = ∂j ∂ j and Γabc (Φ) is the Christoffel symbols on N .
For a map, Φ, to have finite energy, it should converge to a constant at infinity. In
this case for each moment of time, t, Φ can be extended to a continuous map from S d to
N taking the point at infinity to the limit of Φ(x) at the spatial infinity. (In other words,
we pass to the one-point compactification of Rd .) Then one can define the degree, degΦ,
as the homotopy class of Φ as a map from S d to N , i.e. a member of the homotopy group
πd (N ). This degree is conserved under the dynamics generated by the Euler-Lagrange
equations above.
In the most important case N = G/H, where G is a compact Lie group and H is its
subgroup, specifically, G = SO(n + 1) and H = SO(n), so that N = S n . In the particle
physics one takes d = 2, 3 and n = 2, and in the condensed matter physics one takes
d = n = 2, i.e. Φ : Rd → S 2 , d = 2, 3. In both cases, the degree of Φ is an integer (the
degree for maps from S d to S 2 ), i.e. πd (S 2 ) = Z, for d = 2, 3.
Consider N = S n and let S n be embedded in Rn+1 in the standard way, so that Φ can
be thought of as a map from Md+1 to Rn+1 , satisfying |Φ| = 1. Then the energy can be
written as
1
Z
E(Φ) = ∂i Φ · ∂i Φdd x.
2 Rd
(a · b denotes the dot product in Rn+1 .)
Let d = n = 2. Then the degree is given by
1 1 1
Z Z Z
2 ∗
degΦ = Φ · (∂1 Φ ∧ ∂2 Φ)d x (= Φ (dS)) = d2 xj 0 ,
8π R2 4π S 2 2π R2
where
1
j 0 :=
Φ · (∂1 Φ ∧ ∂2 Φ)dx1 ∧ dx2 . (20.30)

A key point here is that one has the Bogomolnyi-type inequality
E(Φ) ≥ 4π|degΦ|.
Indeed, we have the Bogomolnyi-type identity
1
Z
E(Φ) = ±4πdegΦ + |∂i Φ ± ij Φ ∧ ∂j Φ|2 d2 x,
4 R2
208 Lectures on Applied PDEs, November 3, 2017

where ij is the Levi-Cevita antisymmetric symbol with 12 = −21 = 1, 11 = 22 =
0, which implies the inequality above. Moreover, the last relation yields that in every
homotopy class solutions of the self-dual/ anti-dual equations,
∂i Φ ± ij Φ ∧ ∂j Φ = 0
are the minimizers of E(Φ).
For any degΦ = k, these equations have explicit solutions (harmonic or anti-harmonic
maps), Φstat stat
k , given by Φk (ρ, φ) = (Uk (ρ), kφ), where (ρ, φ) are the polar coordinates in
R and (ϕ, θ) are the spherical coordinates in S 2 and Uk (ρ) = 2arctan ρk .
2

The solutions above can be found by going to the stereographic projection, Φ → W ,


of S 2 to the complex plane C, as

2 Re W −2ImW 1 − |W |2 
Φ= , , .
1 + |W |2 1 + |W |2 1 + |W |2
Here we identified R2 with the complex plane C, by z := x1 + ix2 . In the new variables
the energy is given by Z ¯ |2
|∂W |2 + |∂W
E(W ) := dz,
(1 + |W |2 )2
where we used the complex derivatives ∂ := ∂1 − i∂2 and ∂¯ := ∂1 + i∂2 , while the degree
is given by
1 ∂W ∂¯W̄ − ∂W
¯ ∂ W̄
Z
deg Φ ≡ deg W = dz.
4π (1 + |W |2 )2
The Euler - Lagrange equation, dW̄ E(W ) = 0, is given by
2W ¯ |2 ).
∆W = (|∂W |2 + |∂W
1 + |W |2
¯ = 0 or ∂W = 0 (the Cauchy -
Clearly, E(W ) is minimized by W satisfying either ∂W
Riemann equations), i.e. W is either a holomorphic or anti-holomorphic function. They
are mapped into each other by the complex conjugation. We consider holomorphic solu-
tions. One can show that they are of the form W (x) = Q(z)
P (z)
, where P (z) and Q(z) are
polynomials with no common factors. The degree of denominator = deg Φ (the harmonic
maps of degree n.).

21 PDEs of quantum mechanics and statistics


21.1 Hartree, Hartree - Fock and Gross-Pitaevski equations
Even for a few particles the Schrödinger equation is prohibitively difficult to solve. Hence
it is important to have approximations which work in various regimes. One such approx-
imation, which has a nice unifying theme and connects to a large areas of physics and
Lectures on Applied PDEs, November 3, 2017 209

mathematics, is the one approximating solutions of n-particle Schrödinger equations by


products of n one-particle functions (i.e. functions of 3 variables). This results in a single
nonlinear equation in 3 variables, or several coupled such equations. The trade-off here is
the number of dimensions for the nonlinearity. This method, which goes under different
names, e.g. the mean-field or self-consistent approximation, is especially effective when
the number of particles, n, is sufficiently large.
For simplicity we consider a system of n identical, spinless bosons. It is straightforward
to include spin. To extend our treatment to fermions requires a simple additional step
(see discussion below). The Hamiltonian of the system of n identical bosons of mass m,
interacting with each other, is
n
X ~2 1X
Hn := (− ∆xi ) + gv(xi − xj ), (21.1)
j=1
2m 2 i6=j

acting on the state space sn1 L2 (Rd ), d = 1, 2, 3. Here ∆x is the Laplacian acting on
the variable x, g > 0 is a parameter called the coupling constant, v is the interaction
potential, and s is the symmetric tensor product. As we know, the quantum evolution
is given by the Schrödinger equation
∂Ψ
i~ = Hn Ψ.
∂t
This is an equation in 3n + 1 variables, x1 , ..., xn and t, and it is not a simple matter to
understand properties of its solutions.
The Hartree equation is the Euler-Lagrange equation for stationary points of the action
functional
~
Z  
1
S(Ψ) := − ImhΨ, ∂t Ψi − hΨ, Hn Ψi dt,
2 2
considered on the set of functions
{Ψ := ⊗n1 ψ|ψ ∈ H 1 (R3 )}.
Here (⊗n1 ψ) is the function of 3n variables defined by (⊗n1 ψ)(x1 , ..., xn ) := ψ(x1 )...ψ(xn ).
The Euler-Lagrange equation for S(Ψ) on the set above is the following nonlinear equation
∂ψ
i~ = (h + v ∗ |ψ|2 )ψ. (21.2)
∂t
This nonlinear evolution equation is called the Hartree equation (HE).
If the interparticle interaction, v, is significant only at very short distances (one says
that v is very short range, which technically can be quantified by assuming that the
“particle scattering length” a is small), we can replace v(x) → 4πaδ(x), and Equation (??)
becomes
∂ψ
i~ = hψ + 4πa|ψ|2 ψ (21.3)
∂t
210 Lectures on Applied PDEs, November 3, 2017

(with the normalization (A.9)). This equation is called the Gross-Pitaevski equation
(GPE) or nonlinear Schrödinger equation. It is a mean-field approximation to the original
quantum problem for a system of n bosons. The Gross-Pitaevski equation is widely used
in the theory of superfluidity, and in the theory of Bose-Einstein condensation.
For (spinless) fermions, we consider the action S(Ψ) on the following function space

{Ψ := det[ψi (xj )]|ψi ∈ H 1 (R3 ) ∀i = 1, ..., n}

where [ψi (xj )] stands for the n × n matrix with the entries indicated. Then the Euler-
Lagrange equation for S(Ψ) on the latter set gives a system of nonlinear, coupled evolution
equations
∂ψj X X
i~ = (h + v ∗ |ψi |2 )ψj − (v ∗ ψi ψ̄j )ψi , (21.4)
∂t i i

for the unknowns ψ1 , ..., ψn . This systems plays the same role for fermions as the Hartree
equation does for bosons. It is called the Hartree-Fock equations (HFE).
Reconstruction of solutions to the n-particle Schrödinger equation. How do solutions of
(HE) or (GPE) relate to solutions of the original many-body Schrödinger equation? One
can show rigorously (see a review in [?]) that the solution of the Schrödinger equation

∂Ψ
i~ = Hn Ψ, Ψ|t=0 = ⊗n1 ψ0
∂t
satisfies, in some weak sense and and in the mean-field regime of n → ∞ and g → 0, with
ng fixed
Ψ − ⊗n1 ψ → 0
where ψ satisfies (HE) with initial condition ψ0 .
(HE), (HFE) and (GPE) are invariant under the time shifts and the gauge transfor-
mations,
ψ(x) → eiα ψ(x), α ∈ R.
Consequently, the energy, E(ψ), and the number of particles, N (ψ), (see below) are
conserved quantities.
To fix ideas, we will hereafter discuss mainly (GPE). For (HE) and (HFE) the results
should be appropriately modified. For (GPE) the energy functional is

~
Z  2 
2 2 4
E(ψ) := |∇ψ| + V |ψ| + 2πa|ψ| dx.
R3 2m

The number of particles for (GPE) and (HE) is given by


Z
N (ψ) := |ψ|2 dx
R3
Lectures on Applied PDEs, November 3, 2017 211

while for (HFE), by N (ψ) := i R3 |ψi |2 dx. Note that the energy and number of parti-
P R
cle conservation laws are related to the time-translational and gauge symmetries of the
equations, respectively.
Moreover, (HFE) is invariant under time-independent unitary transformations of {ψ1 , ..., ψn }.
As a result, (HFE) conserves the inner products, hψi , ψj i, ∀i, j.
The last item shows that the natural object for (HFE) is the subspace spanned by
{ψi },P
and the equation can be rewritten as an equation for the corresponding projection
γ := i |ψi ihψi |.
The above notions of the energy and number of particles are related to corresponding
notions in the original microscopic system. Indeed, let Ψ := √1n ⊗n1 ψ. Then

1
hΨ, Hn Ψi = E(ψ) + O
n
where E(ψ) is the energy for (HE) and
Z Z
2
n | Ψ(x1 , ..., xn ) | dx1 ...dxn = | ψ(x) |2 dx.

The notion of bound state can be extended to the nonlinear setting as follows. The
bound states are stationary solutions of (HE) or (GPE) of the form

ψ(x, t) = φµ (x)eiµt

where the profile φµ (x) is in H 2 (R3 ). Note that the profile φµ (x) satisfies the stationary
Gross-Pitaevski equation:
hφ + 4πa|φ|2 φ = −~µφ (21.5)
(we consider here (GPE) only). Thus we can think of the parameter −µ as a nonlinear
eigenvalue.
A ground state is a bound state such that the profile φµ (x) minimizes the energy for
a fixed number of particles:

φµ minimizes E(ψ) under N (ψ) = n

(see Chapter ?? which deals with variational, and in particular minimization problems).
Thus the nonlinear eigenvalue µ arises as a Lagrange multiplier from this constrained
minimization problem. In Statistical Mechanics µ is called the chemical potential (the
energy needed to add one more particle/atom, see Section ??).
Remark 1. 1. Mathematically, the ground state can be also defined as a stationary
solution with a positive (up to a constant phase factor) profile, ψ(x, t) = φµ (x)eiµt
with φµ (x) > 0. Let δ(µ) := kφµ k2 . Then we have (see [?])

δ 0 (µ) > 0 =⇒ φµ minimizes E(ψ) under N (ψ) = n.


212 Lectures on Applied PDEs, November 3, 2017

2. The Lagrange multiplier theorem in Section 10.6 implies that the ground state profile
φµ is a critical point of the functional

Eµ (ψ) := E(ψ) + ~µN (ψ).

In fact, φµ is a minimizer of this functional under the condition N (ψ) = n.

If φµ is the ground state of (GPE), then ⊗n1 φµ is close to the ground state of the n−body
Hamiltonian describing the Bose-Einstein condensate (see [?] for a review, and [?, ?] and
the Appendix below for rigorous results).
It is known that for natural classes of nonlinearities and potentials V (x) there is a
ground state. Three cases of special interest are
~ 2 2m
1. h := − 2m ∆ + V (x) has a ground state, and ~2
n|a| 1
2m
2. V has a minimum, ~2
n|a|  1, and a < 0

3. V (x) → ∞ as |x| → ∞ (i.e. V (x) is confining) and a > 0.

(The first and third cases are straightforward and the second case requires some work
[?, ?, ?].)
Stability. We discuss now the important issue of stability of stationary solutions under
small perturbations. Namely, we want to know how solutions of our equation with initial
conditions close to a stationary state (i.e. small perturbations of φµ (x)) behave. Are
these solutions stay close to the stationary state in question, do they converge to it, or do
they depart from it? This is obviously a central question. This issue appeared implicitly
in Section ?? (and in a stronger formulation in Chapter ??) but has not been explicitly
articulated yet. This is because the situation in the linear case that we have dealt with so
far is rather straightforward. On the other hand, in the nonlinear case, stability questions
are subtle and difficult, and play a central role.
We say that a stationary solution, φµ (x)eiµt , is orbitally (respectively, asymptotically)
stable if for all initial conditions sufficiently close to φµ (x)eiα (for some constant α ∈
R), the solutions of the evolution equation under consideration stay close (respectively,
converge in an appropriate norm) to a nearby stationary solution (times a phase factor),
0
φµ0 (x)ei(µ t+β(t)) . Here µ0 is usually close to µ, and the phase β depends on time, t. The
phase factors come from the fact that our equations have gauge symmetry: if ψ(x, t) is a
solution, then so is eiα ψ(x, t) for any constant α ∈ R. One should modify the statement
above if other symmetries are present. The notion of orbital stability generalizes the
classical notion of Lyapunov stability, well-known in the theory of dynamical systems, to
systems with symmetries.
For the linear Schrödinger equation, all bound states, as well as stationary states corre-
sponding to embedded eigenvalues, are orbitally stable. But they are not asymptotically
Lectures on Applied PDEs, November 3, 2017 213

stable in general. For most nonlinear evolution equations in unbounded domains, the
majority of states are not even Lyapunov/orbitally stable.
For (GPE), if V → ∞ as |x| → ∞ (i.e. V is confining), the ground states are orbitally
stable, but not asymptotically stable. If V → 0 as |x| → ∞, the ground states can be
proved to be asymptotically stable in some cases (see [?, ?, ?, ?, ?, ?] and references
therein).

21.2 Appendix II: Heuristic derivation of the Hartree equation


We give a heuristic derivation of the mean-field approximation for this equation. A
rigorous derivation is scetched in [6]. First, we observe that the potential experienced by
the i-th particle is X
W (xi ) := V (xi ) + v(xi − xj ).
j6=i

Assuming v(0) is finite, it can be re-written, modulo the constant term v(0), which we
neglect, as W (xi ) = V (xi ) + (v ∗ ρmicro )(xi ). Here, recall, f ∗ g denotes the convolution of
the functions f and g, and ρmicro stands for the (operator of) microscopic density of the
n particles, defined by X
ρmicro (x, t) := δ(x − xj ).
j

Note that the average quantum-mechanical (QM) density in the state Ψ is


hΨ, ρmicro (x, t)Ψi = ρQM (x, t)
where ρQM (x, t) := n | Ψ(x, x2 , ..., xn ) |2 dx2 ...dxn , the one-particle density in the
R

quantum state Ψ.
In the mean-field theory, we replace ρmicro (x, t) with a continuous function, ρM F (x, t),
which is supposed to be close to the average quantum-mechanical density, ρQM (x, t), and
which is to be determined later. Consequently, it is assumed that the potential experienced
by the i-th particle is
W M F (xi ) := V (xi ) + (v ∗ ρM F )(xi ).
Thus, in this approximation, the state ψ(x, t) of the i-th particle is a solution of the
following one-particle Schrödinger equation i~ ∂ψ ∂t
= (h + v ∗ ρM F )ψ where, recall, h =
~ 2
− 2m ∆x +RV (x). Of course, the integral of ρmicro (x, t) is equal to the total number of
particles,
R R3
ρmicro (x, t)dx = n. We require that the same should be true for ρM F (x, t):
R3
ρM F (x, t)dx = n. We normalize the one-particle state, ψ(x, t), in the same way
Z
|ψ(x, t)|2 dx = n. (21.6)
R3
Consider a situation in which we expect all the particles to be in the same state ψ. Then
it is natural to take ρM F (x, t) = |ψ(x, t)|2 . In this case ψ solves the Hartree equation
(21.2).
214 Lectures on Applied PDEs, November 3, 2017

21.3 Quantum statistics


We formalize the theory of quantum statistics by making the following postulates (for
details see [6]):

• States: positive trace-class operators on a Hilbert space H (as usual, up to


normalization);

• Evolution equation : i~ ∂ρ
∂t
= [h, ρ], where h is a self-adjoint operator on H;

• Observables : self-adjoint operators on H;

• Averages : hAiρ := Tr(Aρ).

We call the theory described above quantum statistics. Assuming H = L2 (Rn ), the last
two items lead to the following expression for the probability density for the coordinates:

• ρ(x; x) - probability density for coordinate x;

and similarly for the momenta. In particular, if ρ = Pψ , then

ρ(x; x) = |ψ(x)|2 ,

as should be the case according to our interpretation.


Note that the state space here is not a linear space but a positive cone in a linear
space. It can be identified with the space of all positive (normalized) linear functionals
A → ω(A) := Tr(Aρ) on the space of bounded observables. Denote the spaces of bounded
observables and of trace class operators on H as L∞ (H) and L1 (H), respectively. There
is a duality between density matrices and observables

hρ, Ai = Tr(Aρ)

for A ∈ L∞ (H) and ρ ∈ L1 (H).


Quantum mechanics is a special case of this theory, and is obtained by restricting the
density operators to be rank-one orthogonal projections.
Another special case of quantum statistics is probability.

21.4 Self-consistent approximation


Consider a system of n particles in an external potential V (x), interacting via a pair
potential v(x). Assuming the particles are identical and in the P same state, in the
R self-
consistent approximation, one replace the many P body potential, j v(x − xj ) = v(x −
y)nexact (y)dy = (v∗nexact )(x), where nexact (y) := j δ(x−xj ), created by n−1 particles at
P R
x and affecting the remaining one, by the ‘mean-field’ potential j v(x − y)nγ (y)dy =
Lectures on Applied PDEs, November 3, 2017 215

(v ∗ nγ )(x). This leads to the time-dependent generalized Hartree or the Hartree-von


Neumann equation,
∂γ
i = −[hγ , γ] (21.7)
∂t
with
hγ = h + v ∗ nγ and nγ (x) = γ(x, x). (21.8)
Here h = −∆+V acting H = L2 (Rn ). We describe some basic properties of this equation.
If the external potential V is zero (or independent of time), then the equation (21.7)
is invariant, under spatial (or time) translations

Thtrans : γ 7→ Uh γUh−1 , (21.9)

for any h ∈ Rd (space translations) or h ∈ R (time translations), and rotations

Tρrot : γ 7→ Uρ γUρ−1 , (21.10)

for any ρ ∈ SO(d). Here Uhtransl and Uρrot are the standard translation and rotation
transforms Uhtransl : φ(x) 7→ φ(x + h) and Uρrot : φ(x) 7→ φ(ρ−1 x).
Since the equation (21.7) is a hamiltonian system (see an appendix to this section),
these symmetries lead to the conservation laws. In particular, we have

• Time translation invariance → conservation of energy,


1 
E(γ) := Tr (h + v ∗ nγ )γ (21.11)
2Z
1
= Tr(hγ) + nγ v ∗ nγ dx. (21.12)
2

(The half compensates for the differentiating the quadratic term.)

• Unitarity of the Schrödinger evolution (gauge invariance) and the and cyclicity of
the trace → conservation of number of particles (total charge),

Tr γ = const. (21.13)

• Unitarity of the Schrödinger evolution → conservation of positivity: if γ0 ≥ 0, then


for all times γ ≥ 0,

γ0 ≥ 0 ⇒ γ ≥ 0. (21.14)

• Unitarity of the Schrödinger evolution → conservation of the eigenvalues γj ’s of γ,

∂t γj = 0. (21.15)
216 Lectures on Applied PDEs, November 3, 2017

Indeed, for (21.13), we have



−i∂t Tr γ = Tr [hg , γ] = 0. (21.16)
To prove (21.14), we let U (γ) be the evolution generated by the self-adjoint, time-
dependent operator hγ . We can rewrite the equation (21.7) as
γ = U (γ)∗ γ0 U (γ). (21.17)
Then the positivity of γ0 implies the positivity of γ. ((21.17) implies also (21.13).)
Eq. (21.15) follows from the isospectral properties of (21.7): γ can be written in the
form (21.17). (A different proof is given as follows. Using that γj = hφj , γφj i, we find
∂t γj = h∂t φj , γφj i + hφj , γ∂t φj i + hφj , (∂t γ)φj i. The first two terms give γj h∂t φj , φj i +
γj hφj , ∂t φj i = γj ∂t hφj , φj i = 0, while the third terms gives hφj , i[hγ , γ]φj i = 0. This
proves (21.15).)
Proposition 21.1. γ satisfies the equation (21.7) iff the eigenfunctions φj of γ satisfy
the equations
i∂t φj = hγ φj , (21.18)
where hγ is given in (21.8) and can be expressed in terms of φj ’s using
X
nγ (x) = γ(x, x) = |φj (x)|2 . (21.19)
j

Proof. To prove (21.18), we note that, since the spectrum of γ is discrete, we can write it
in terms of the projections on the eigenfunctions its φj (associated with the eigenvalues
γj ):
X
γ= γj Pφj . (21.20)
j

We plug (21.20) into i ∂γ


∂t
+ [hγ , γ] and use (21.15), to obtain
∂γ X
i + [hγ , γ] = γj (|(i∂t − hγ )φj ihφj | − |φj ih(i∂t − hγ )φj |. (21.21)
∂t j

Hence, if (21.18) are satisfied then so is (21.7). On the other hand, if (21.7) is satisfied,
then multiplying (21.21) scalarly by φk we obtain
X
γk (i∂t − hγ )φk = γj φj h(i∂t − hγ )φj , φk i
j
X
= γj φj hφj , (i∂t − hγ )φk i = γ(i∂t − hγ )φk .
j

Assuming the eigenvalues γk are non-degenerate, this implies that there are real numbers
µk s.t. (i∂t − hγ )φk = µk φk and therefore (i∂t − hγ )(eiµk t φk ) = 0.
Lectures on Applied PDEs, November 3, 2017 217

Remark. Solutions to the equations (21.18) are parametrized by the numbers γj ’s.

21.5 Equilibrium states and entropy


Clearly, if, for some function f , the operator γ satisfies the equation γ = f (hγ ), then
it also satisfies (21.7), namely it is a static solution to (21.7). However, not all such
solutions are of physical interest. We select those which are using the entropy principle,
namely, requiring that they minimize the energy for the fixed entropy (and the number
of particles).
We have already defined the energy functional E(γ) on trace class operators γ, see
(21.11). Now, we take the entropy of γ to be of the form

S(γ) = Tr g(γ), (21.22)

where
1
g(x) := − (x ln x + (1 − x) ln(1 − x)). (21.23)
2
We are interested in minimizing the internal energy E(γ) on the set S∗ := {0 ≤ γ ≤
1} ∩ {the entropy, S(γ), and the number of particles, N (γ) := Tr γ, are fixed}. As usual,
we define the free energy on the convex set S := {0 ≤ γ ≤ 1} as

FT µ (γ) = E(γ) − T S(γ) − µN (γ), (21.24)

fixing the chemical potential, µ, rather than the number of particles, N (γ). We define
FT µ (γ) on the Sobolev space space H 1,1 (H) defined as follows. Let b := (c1 + h)1/2 , where
c > 0 is such that c1 + h ≥ 1. Define the Sobolev space H s,1 (H) as

H s,1 (H) := {γ ∈ L1 (H) : bs/2 γbs/2 ∈ L1 (H)}. (21.25)

Define hγµ := hγ − µ. We begin with the following lemma proven in Appendix 21.8,

Lemma 21.2. Minimizers, γ, of the internal energy E(γ) on the convex set S, with
S(γ) and N (γ) fixed, are critical points of FT µ (γ), i.e. they satisfy the Euler-Lagrange
equations

dγ FT µ (γ) = 0, (21.26)

for some T and µ (the latter are determined by fixing S(γ) and Tr(γ)).

For S given in (21.22), the equation (21.26) becomes

1
γ = g # ( hγµ ), (21.27)
T
218 Lectures on Applied PDEs, November 3, 2017

where g # (h) := (g 0 )−1 (h). Indeed, the Gâteaux derivatives, ∂γ FT (γ), is given by (see
Appendix 21.8)

∂γ FT µ (γ) = hγµ − T g 0 (γ), (21.28)

where hγµ = hγ − µ. This together with (21.26) gives

hγµ = T g 0 (γ), (21.29)

which upon inverting the function g 0 gives (21.27).


Note that for g(x) := − 21 (x ln x + (1 − x) ln(1 − x)), we have

1 x
g 0 (x) = − ln and g # (h) = (1 + e2h )−1 . (21.30)
2 1−x

21.6 Local and global existence


Let Y := H 1,1 (H), where H s,1 (H) is the Sobolev space space defined in (21.25). Recall,
that we think of γ as a path γ : t ∈ I → u(t) ∈ Y in Y and we are looking for weak
solutions to (21.7), i.e. for solutions of this equation in the integral form, in the space
C([0, T ], Y ), for some T > 0. Define the ball BR,T := {γ ∈ C([0, T ], Y ) : kγkC([0,T ],Y ) ≤
R}. We have the following result (cf. J. M. Chadam, J. M. Chadam and R. T. Glassey, A.
Bove, G. Da Prato and G. Fano ([?], [?], [?]). (needs checking and filling in details)

Theorem 17. (i) Assume v ∈ L2 ∩ L1 . There are functions LR and MR , R > 0, s.t. for
γ0 ∈ Y , R > 2Kkγ0 kY and T < (KLR )−1 , R/2KMR , the equation (21.7) has a unique
weak solution γ ∈ BR,T . The solution γ depends continuously on the initial condition γ0 .
Furthermore, either the solution is global in time or blows up in Y in a finite time (i.e.
either kγ(t)kY < ∞, ∀t, or kγ(t)kY < ∞ for t < t∗ and kγ(t)kY → ∞ as t → t∗ for some
t∗ < ∞).
(ii) If in addition v∗ is positive definite, then the solutions are global.

Proof. We will use Theorem D.3. Note that the equation (21.7) is in the form (??) with
the linear operator A acting on density operators as A(ρ) := i[h, ρ] and f (γ) := [v ∗ nγ , γ].
We have to check that the conditions (D.12) - (D.14) are satisfied.
The operator A is an operator of adjoint representation and it generates the one
parameter group eAt (γ) = eiht γe−iht , which, by the cyclicity of the trace, unitarity of e−iht
and the fact that b commutes with e−iht , satisfies (D.12), for K = 1.
Next, it is straightforward to show that f (γ) satisfies (D.13) - (D.14) for some MR , LR <
∞. Then Theorem D.3 implies the first statement of the theorem.
To prove that the solutions are global one uses the conservation of the energy and
number of particles.
Lectures on Applied PDEs, November 3, 2017 219

21.7 Existence of ground states and Gibbs states


(needs checking and filling in details)

Theorem 18. Assume v ∈?? and g is strictly convex and satisfies ?? and g∗ ≥ 0 (g∗ is
the Legandre transform of g). Then the equation (21.7) has equilibrium states satisfying
(21.27) (see Lemma 21.2).

Proof. We can minimize the free energy functional (21.31) on the convex set S := {γ ∈
H 1,1 (H) : 0 ≤ γ ≤ 1}, where H s,1 (H) is the Sobolev space space defined in (21.25),
directly. However, we prefer to follow, under additional condition that

v∗ = v̌(−i∇) = (−∆)−1 ,

an elegant proof of Markowich, Rein and Wolansky ([?]) In this proof we pass from the
functional FT µ (γ) to the dual one

ΦT µ (V ) = − inf FT µ (γ, V ), (21.31)


γ

where
1
Z
V
|∇V |2 ,

FT µ (γ, V ) := Tr h γ − T Tr g(γ) − µ Tr γ − (21.32)
2

with hV := h + V . Notice that F (γ, vγ ) = FT µ (γ), where vγ := v ∗ nγ . Using that


dγ FT µ (γ, V ) = hV −T g 0 (γ)−µ and Tr (hV −µ)γ −T Tr g(γ) hV −µ−T g0 (γ)=0 = T g∗ ( T1 (hV −
µ)), we compute

1 1 V
Z
|∇V |2 − T Tr g∗

ΦT µ (V ) = (h − µ) . (21.33)
2 T

Now, one can show easily that

• If V∗ is a critical point of ΦT µ (V ), then γ∗ := g∗0 T1 (hV∗ − µ) is a critical point




of FT µ (γ). (Indeed, since dV ΦT µ (V ) = −∆V − nγ , where γ := g∗0 T1 (hV − µ) , a




critical point of ΦT µ (V ) satisfies V = (−∆)−1 nγ =: Vγ , where γ := g∗0 T1 (hV − µ) ,


or γ := g∗0 T1 (hV − µ) satisfies γ := g∗0 T1 (hVγ − µ) , which is (21.27).)

• The dual functional ΦT µ (V ) is coercive and weakly lower semi-continuous and there
fore has a minimizer. Moreover, it is strictly convex and so the minimizer is unique.

The last two statements imply the desired result.


220 Lectures on Applied PDEs, November 3, 2017

21.8 Appendix: Proof of Lemma 21.2


Proof of Lemma 21.2. To prove (21.27), we observe that if γ ∈ S is a minimizer, then for
any γ 0 , we have E((1 − s)γ + sγ 0 ) − E(γ) ≥ 0, provided S((1 − s)γ + sγ 0 ) − S(γ) = 0
and Tr((1 − s)γ + sγ 0 ) − Tr(γ) = 0. Dividing this by s > 0 and taking s → 0, we obtain
E 0 (γ)(γ 0 − γ) ≥ 0 for any γ 0 , satisfying S 0 (γ)(γ 0 − γ) = 0 and Tr(γ 0 − γ) = 0. Hence γ 0 = γ
minimizes E 0 (γ)(γ 0 − γ) on the set S, provided S 0 (γ)(γ 0 − γ) = 0 and Tr(γ 0 − γ) = 0.
Since E 0 (γ)(γ 0 − γ) is a linear function of γ 0 , this is possible iff γ satisfies
E 0 (γ) − T S 0 (γ) − µN 0 (γ) = 0, ?? (21.34)
for some T and µ. On the other hand, by the definition (21.31), we have
FT0 (γ, a)Φ = [E 0 (γ, a) − T S 0 (γ) − µN 0 (γ)]Φ, (21.35)
which together with (21.34) gives (21.26).
To prove (21.26), we first compute S 0 (γ)Φ = ∂s S(γ + sΦ)|s=0 . To Rthis end, we recall
1
that S(γ) = Tr g(γ) and write g(γ) as the Cauchy integral g(γ) = 2π dzg(z)(γ − z)−1 .
It is straightforward to compute
1
Z
∂s g(γ + sΦ)|s=0 = − dzg(z)(γ − z)−1 Φ(γ − z)−1 , (21.36)

which gives
1
Z
∂s S(γ + sΦ)|s=0 =− dzg(z) Tr((γ − z)−1 Φ(γ − z)−1 )

1
Z
=− dzf (z) Tr((γ − z)−2 Φ),

which implies
S 0 (γ)Φ = Tr(g 0 (γ)Φ). (21.37)
Next, the relation (21.11) implies that
E 0 (γ)Φ = Tr((hγµ |µ=0 )Φ), (21.38)
where hγµ was defined above. Eq. (21.35), together with (21.37) and (22.9) and the
computation N 0 (γ) = 1, gives (21.28).

21.9 Appendix: Hamiltonian formulation


Following [?], as canonical variables, we take κ and κ∗ , where κκ∗ = γ. Then the Hamil-
tonian and Poisson bracket are defined as
1
H (κ, κ∗ ) := T r(κ∗ (h + v ∗ nκ )κ), (21.39)
2
Lectures on Applied PDEs, November 3, 2017 221

|κ(x, y)|2 dy, and


R
with nκ (x) := nκ∗ κ (x) :=

{A, B}(κ, κ∗ ) = − − i Tr (∂κ A∂κ∗ B − ∂κ B∂κ∗ A) (κ, κ∗ ). (21.40)

Another way to introduce a hamiltonian structure is as follows. For a functional A(ρ)


we define the gradient operator, ∂ρ A(ρ), in the trace metric by the equation T r(∂ρ A(ρ)ξ) =
∂s A(ρ + sξ)|s=0 . On the space of classical field observables we define the Poisson bracket
by
{A(ρ), B(ρ)} = −i Tr (∂ρ A(ρ)ρ∂ρ B(ρ) − ∂ρ B(ρ)ρ∂ρ A(ρ)) . (21.41)
The Jacobi identity for (21.41) is proven in Appendix ??. We observe that
Z
{A(ρ), B(ρ)}|ρ=Pψ = i (∂ψ(x) A∂ψ(x) B − ∂ψ(x) A∂ψ(x) B)(ψ, ψ)dx. (21.42)

The r.h.s. is the standard Poisson bracket, {A, B}(ψ, ψ), for the Hartree equation (see
[?]). Indeed, ∂ρ A(ρ) and ∂ρ B(ρ) are operators on L2 (R3 ) (1−particle observables) and
therefore
T r((∂ρ B)(Pψ )Pψ (∂ρ A)(Pψ )) = h(∂ρ A)(Pψ )∗ ψ, (∂ρ B)(Pψ )ψi.
Since, as it is easy to see, (∂ρ B)(Pψ )ψ = ∂ψ(x) B(ψ, ψ) and (∂ρ A)(Pψ )∗ ψ = ∂ψ(x) A(ψ, ψ),
R
this gives T r((∂ρ B)(Pψ )Pψ (∂ρ A)(Pψ )) = ∂ψ(x) A∂ψ(x) B, which implies the desired rela-
tion.
We define the Hamiltonian functional on L1 (H) as

1
Z
H (ρ) := Tr(hρ ρ) + nρ v ∗ nρ dx, (21.43)
2

The resulting Hamilton equation,

∂t ρ = {H(ρ), ρ}, (21.44)

is exactly the Hartree-von Neumann equation considered above. Indeed, this fact follows
from the equation
{H(ρ), ρ} = −i[hρ , ρ], (21.45)
where hρ := −∆ + V (x) + (v ∗ nρ ). To show the latter equation we use the definition of
the Poisson bracket and the relation T r(∂ρ ρξ) = ξ, which follows from the definition of
∂ρ A(ρ) above, to obtain {H, ρ} = −i (∂ρ H(ρ)ρ − ρ∂ρ H(ρ)) . Next, computing ∂ρ H(ρ), we
conclude that {H, ρ} is equal to the r.h.s. of (21.45).

Remark 21.3. Equation (21.44) should be understood in the weak sense: for all a ∈ A1 ,

∂t Tr(aρ) = Tr(a {H(ρ), ρ}). (21.46)


222 Lectures on Applied PDEs, November 3, 2017

Using the linearity of the Poisson brackets in the second factor, we obtain

T r(a{H, ρ}) = {H(ρ), Tr(aρ)}. (21.47)

Thus equation (21.44) is equivalent to the equation

∂t Tr(aρ) = {H(ρ), Tr(aρ)}. (21.48)

The Hamiltonian (21.43) and the Poisson brackets (21.41) generate the Hartree-von
Neumann flow ϕt on one-particle density matrices. This flow induces the flow on gener-
alized classical observables:
A(ρ) → A(ϕt (ρ)). (21.49)

21.10 Appendix: Hilbert Space Approach


Quantum statistical dynamics can be put into a Hilbert space framework as follows.
Consider the space HHS of Hilbert-Schmidt operators acting on the Hilbert space H.
These are the bounded operators, K, such that K ∗ K is trace-class (see Section ??).
There is an inner-product on HHS , defined by

hF, Ki := Tr(F ∗ K). (21.50)

Exercise 1. Show that (21.50) defines an inner-product.

This inner-product makes HHS into a Hilbert space (see [?, 15]). On the space HHS ,
we define an operator L via
1
LK = [H, K],
~
where H is the Schrödinger operator of interest. The operator L is symmetric. Indeed,

~hF, LKi = Tr(F ∗ [H, K]).

Using the cyclicity of the trace, the right hand side can be written as

Tr(F ∗ HK − F ∗ KH) = Tr(F ∗ HK − HF ∗ K) = Tr([F ∗ , H]K)


= Tr([H, F ]∗ K) = ~hLF, Ki

and so hF, LKi = hLF, Ki as claimed. In fact, for self-adjoint Schrödinger operators, H,
of interest, L is also self-adjoint.
Now consider the Landau-von Neumann equation
∂k
i = Lk (21.51)
∂t
Lectures on Applied PDEs, November 3, 2017 223

where k = k(t) ∈ HHS . Since k(t) is a family of Hilbert-Schmidt operators, the operators
ρ(t) = k ∗ (t)k(t) are trace-class, positive operators. Because k(t) satisfies (21.51), the
operators ρ(t) obey the equation
∂ρ 1
i = Lρ = [H, ρ]. (21.52)
∂t ~
If ρ is normalized – i.e., Tr ρ = 1 – then ρ is a density matrix satisfying the Landau-von
Neumann equation (21.52). The stationary solutions to (21.51) are just eigenvectors of
the operator L with eigenvalue zero.
To conclude, we have shown that instead of density matrices, we can consider Hilbert-
Schmidt operators, which belong to a Hilbert space, and dynamical equations which are
of the same form as for density matrices. Moreover, these equations can be written in
the Schrödinger-type form (21.51), with self-adjoint operator L, sometimes called the
Liouville operator.

22 Existence of bubbles and Lyapunov - Schmidt de-


composition
For a field equation describing dynamics of interfaces, by a bubble we mean a smooth
spherically symmetric static solution, u(x) = v(|x|), with a single interface (or more
precisely a boundary layer) across which the solution changes from approximately constant
value to another.
We will consider two field equations describing interfaces, the Allen-Cahn equation,
(15.23), and the Cahn-Hilliard equation, introduced below. In this case, a bubble is a
smooth spherically symmetric static solution, u(x) = v(|x|), with the property that
v(r) → +1, r → 0, v(r) → −1, r → ∞. (22.1)
In additions, we make a technical assumption that v 0 (r) → −1, r → ∞.
First, we consider the Allen-Cahn equation, (9.3), which describes the phase separation
phenomena. We consider the Allen-Cahn equation in the entire space. We reproduce it
here
∂u
= 2 ∆u − g(u), (22.2)
∂t
where u : Rd × R+ → R and  is a small parameter, with the initial condition u|t=0 =
u0 (x), x ∈ Rd . Here g : R → R is the derivative, g = G0 , of a double-well potential:
G(u) ≥ 0 and has two non-degenerate global minima with the minimum value 0 (see
Figure 2). Specifically, we take g(u) = u3 − u and G(u) = 21 (u2 − 1)2 .
We consider static solutions of this equation in the entire space, Rd . They satisfy the
static Allen-Cahn equation,
− 2 ∆u + g(u) = 0. (22.3)
224 Lectures on Applied PDEs, November 3, 2017

Theorem 19. Let dimensions d ≥ 2. The Allen-Cahn equation (22.3) has no bubble
solutions.
Proof. We use the well-known approach which goes under names of virial relation, Der-
rick’s theorem or Pohozaev identity.
By rescaling, we reduce to the case of  = 1. Let r = |x| and u = v(r) be a spherically
symmetric, C 1 stationary solution to the Allen-Cahn equation:
d−1
− ∂r2 v − ∂r v + g(v) = 0, (22.4)
r
with v|r=0 = 1, and v|r=∞ = −1. We multiply (22.4) by ∂r v = x̂ · ∇v and integrate in r
to obtain Z ∞  2
1 2 ∞ ∂v 1
− (∂r v) |0 − (d − 1) dr + G(v)|∞0 = 0.
2 0 ∂r r
Since the last term is zero, this equation implies
Z ∞  2
1 2 ∂v 1
(∂r v) |r=0 = (d − 1) dr.
2 0 ∂r r
R∞ 2 1
If ∂v |
∂r r=0,∞
= 0,, then 0 ∂v ∂r r
dr = 0 and therefore v must be a constant; this however
contradicts the boundary conditions on v. If ∂v |
∂r r=0
6= 0 (by the continuity, this condition
must hold in a neighborhood of the origin), then the above relation leads to a contradiction
since the right side is infinite and the left side is finite. This completes the proof of the
first part of the theorem.
The reason the Allen-Cahn equation has no bubbles, or that bubble solutions collapse
to a point, is that the mass is not conserved and the surface tension shrinks the interface
to a point. In a variant of the Allen-Cahn equation with the mass conservation - the
Cahn-Hilliard equation - the bubbles do exist as shown below.
Now, we consider the Cahn-Hilliard equation, which play a central role in material
science. In a sense, this is a version of the Allen-Cahn equation with mass conservation.
As the Allen-Cahn equation, (22.2), it presents a key model with many generalizations
and extensions. In Rd , it is of the form
ut = −∆(2 ∆u − g(u)), (22.5)
where g is the same as for the Allen-Cahn equation. In particular, we can take g(u) =
u3 − u. The equation (22.5) is derived from the conservation law of mass
∂u
= − div J, (22.6)
∂t
where J flux of the material and Fick’s law,
J = −D∇µ, (22.7)
Lectures on Applied PDEs, November 3, 2017 225

connecting J to the chemical potential µ coming from thermodynamic consideration and


the expression of the latter in terms of the free energy, E,

µ = dE(u). (22.8)

(Recall, that dE(u) is the Gâteaux derivative of E(u).) If we take the standard expression
1 2
Z
E(u) := ( |∇u|2 + G(u)), (22.9)
Rd 2
with G0 = g, for the free energy E and D constant, say D = 1, then the above implies the
Cahn-Hilliard equation, (22.5).
It follows from Gauss’ theorem that
Z
u dx = constant
Rd

along solutions to (22.5), in agreement with the conservation of the average mass fraction
of the components of the alloy.
We consider static solutions of the Cahn-Hilliard equation in Rd . They satisfy the
static Cahn-Hilliard equation, which, after rescaling, is given by

− ∆u + g(u) = µ, (22.10)

It has a (homogeneous) solution u ≡ ū for any constant ū ∈ g −1 (µ). This gives a static
solution of (22.5).
Theorem 22.1. [Alikakos and Fusco] Let dimensions d ≥ 2. The Cahn-Hilliard equation,
(22.5), has a one-parameter family of static bubble solutions parametrized by their radii.
First, we present our main strategy. To solve the equation (22.10) equation for u we
(i) Construct a family, (φR (r), µR ) of approximate solutions to (22.10) parametrized by
R > 0 (radii of the bubbles). (Here we reparameterized our problem from µ to R.)
(ii) Solve the equation (22.10) near (φR (r), µR ) by using the Lyapunov - Schmidt de-
composition.
Ideas and sketch of the proof of Theorem 22.1. Recall that the static equation (22.10) with
µ = 0, i.e. the Allen-Cahn equation, has the kink solutions:

χx0 e (x) = χ (x − x0 ) · e , where x0 , e ∈ Rn ,



(22.11)

where χ(s) → ±1 as s → ±∞ (see Section 2.1 and Figure 2). Recall that for the key
g(u) = u3 − u, the function χ is given explicitly as
s
χ(s) = tahn( √ ). (22.12)
2
226 Lectures on Applied PDEs, November 3, 2017

Denote F (u, µ) := −∆u + g(u) − µ. Then the equation (22.10) can be rewritten as

F (u, µ) = 0. (22.13)

Let r = |x|. As an approximate solution, we try the shifted kink

χR (r) := χ(r − R) (22.14)

We run into two problems. First, using that χR is a spherically symmetric function and
therefore
d−1
∆χR = ∂r2 χR + ∂r χR ,
r
we obtain F (χR , µ) = −∂r2 χR − d−1
r
∂r χ R + g(χR ) − µ. Using that χR satisfies

− ∂r2 χR + g(χR ) = 0, (22.15)

we find furthermore
d−1
F (χR , µ) = − ∂r χR − µ.
r
The r.h.s. is small (if µ is small and R is large) but is not L2 : as r → ∞, we have
∂r χR → 0 and therefore F (χR , µ) → −µ. Hence we have to modify χR (r) at infinity.
We ignore this for now and look for a solution u(x) of (22.13) in the form u(x) =
χR (r) + α(r), expecting α to be small (the perturbation theory). We rewrite the equation
(22.13) as an equation for α, by expanding F (u, µ) in (22.13) as

F (χR + α, µ) = F (χR , µ) + LR α + N (α, µ), (22.16)

where LR = du F (u, µ)|u=χR and N (α, µ) is defined by this equation, explicitly, N (α) :=
g(χR + α) − g(χR ) − g 0 (χR )α, so that the equation F (χR + α, µ) = 0 becomes

LR α = −(F (χR , µ) + N (α, µ)).

We try to solve this equation for the fluctuation term α, by inverting the operator LR
and reducing it to the fixed point problem α = H(α), where H(α) := −L−1 R (F (χR , µ) +
N (α, µ)).
To show that H is a contraction, we need that the linearized map LR = du F (u, µ)|u=χR
is invertible and its inverse is not too large. The second problem is that, as we indicate
below,

• LR has an eigenvalue of the size O R1 and therefore it is either non-invertible or




invertible with the inverse of the size O (R).


Lectures on Applied PDEs, November 3, 2017 227

Indeed, using that the shifted kink χR satisfies the equation −∂r2 χR + g(χR ) = 0 and
differentiating this equation w.r.to R to obtain

(−∂r2 + g 0 (χR ))χ0R = 0, (22.17)

where χ0R ≡ ∂R χR . Thus we conclude that χ0R is the zero eigenfunction of the operator
L0 := −∂r2 + g 0 (χR ). Using this and LR = L0 − d−1
r
∂r , we compute

d − 1 00
LR (−χ0R ) = χR . (22.18)
r
Now we use that χ00R is concentrated near r = R, more precisely, that
(k)
|χR (r)| ≤ Ce−2|r−R| (22.19)

to obtain the estimate  


1
LR (−χ0R ) =O χ00R . (22.20)
R
(It is easy to check (22.19) for χ for the specific nonlinearity given in (22.12), but it holds
for the general nonlinearities described in (22.11), though with a different exponent.) By
the spectral theory (see Appendix E (to be done) and [6, 8]), this implies this implies
the statement above.
Thus we have to devise a way to overcome these problems. To tackle the first one,
we modify the shifted kink χR (r) := χ(r − R) to find a better approximate solution. To
handle the second problem, we use the Lyapunov - Schmidt reduction and contraction
mapping techniques. We ignore for the moment the first problem, treating χR as a good
approximate solution, and concentrate on the second one.
To begin with, the Lyapunov-Schmidt reduction consists of the following steps:
(a) Parameterization of solutions u(x). We parameterize our solution u(x) by a pair
(R, ξ), where R > 0 and the function α(r), so that u(x) can be represented uniquely as

u(x) = χR (r) + α(r), α ⊥ χ0R . (22.21)

The condition α⊥χ0R ≡ ∂R χR (in L2 (Rd )) determines R. Indeed, applying the implicit
function theorem to the equation f (R, v) := hv − χR , χ0R i = 0, we obtain the unique
solution R = R(v).
To clarify the geometric meaning of th decomposition (22.21), we define the kink
manifold Mkink := {χR (r)| R > 0} and and observe that (22.21) is equivalent to projecting
u(x) to this manifold, which gives the point χR (r) on Mkink and the function α(r) which
is orthogonal to this manifold, α ⊥ TχR Mkink . Since the tangent space to Mkink at χR (r)
is spanned by χ0R (r), this leads to (22.35). The function α(r) is called the fluctuation of
u(x) around χR (r).
228 Lectures on Applied PDEs, November 3, 2017

(b) Decomposition of the equation. Plug the decomposition (22.21) into (22.13) and
project the resulting equation onto the tangent space TχR Mkink and its orthogonal com-
plement, (TχR Mkink )⊥ to obtain the equations for two unknowns R and α:

PR F (χR + α, µ) = 0, (22.22)
PR⊥ F (χR + α, µ) = 0, (22.23)

where PR be the orthogonal projector onto χ0R , namely, PR f := χ0R χ0R f /norm, and
R

PR⊥ = 1l − PR . These are two equations for two unknowns ξ and h.


(c) Solution of the second equation (22.23) for α. We plug the expansion (22.16) into
(22.23) and use the relation α = PR⊥ α and the notation L⊥ ⊥ ⊥
R := PR LR PR , restricted to
Ran PR⊥ , to obtain the equation

PR⊥ F (χR , µ) + L⊥ ⊥
R α + PR N (α, µ) = 0. (22.24)

Now, we use the key fact (see Corollary 22.3, below) that the operator LR , restricted
to the orthogonal complement, Ran PR⊥ = (TχR Mkink )⊥ , of χ0R , is invertible, with the
uniformly bounded inverse. To prove this one needs a fair amount of spectral theory, see
[6, 8]. (For the relevant definitions and facts, see Appendix E (under construction).)
As a result, we can rewrite (22.24) as

α = Φ(α, µ), where Φ(α, µ) := −(L⊥ −1 ⊥


R ) PR [F (χR , µ) + N (α, µ)]. (22.25)

Now, it not hard to show that it is a contraction for R sufficiently large, which implies
that
• For R−1 and µ sufficiently small, the equation (22.23) has a unique solution for
α = α(R, µ) and this solution satisfies the estimate
d−1
α(R, µ) = O(R 2 (R−2 + µ)). (22.26)

We prove a similar result (for the correct approximate solution) later.


(d) Derivation of the reduced equation. Plug the solution α = α(R, µ) of (22.23) into
(22.22) to obtain the new equation for R,

f (R, µ) := hχ0R , F (χR + α(R, µ), µ)i = 0. (22.27)

This is a scalar equation for a single unknown R (the reduced equation). Solve the
reduced equation for R. Then χR + α(R, µ), where R is a solution to (22.27), is a solution
to the stationary Cahn-Hilliard.
(e) Solution of the reduced equation for R. Since PR f := χ0R χ0R f /norm, we can
R

rewrite (22.27) as hχ0R , F (χR + α, µ)i = 0. Remembering the expansion (22.16), this gives

hχ0R , F (χR , µ)i + hχ0R , LR α(R, µ)i + hχ0R , N (α(R, µ)i = 0. (22.28)
Lectures on Applied PDEs, November 3, 2017 229

To be specific, in what follows we take g(u) = u3 − u. Then the function χ is given


explicitly by (??) and one can easily compute the terms hχ0R , F (χR , µ)i, hχ0R , LR αi and
χ0R , N (α)i (see Appendix 22.1, below) and use the estimate (22.26), to derive from (22.28)
the following equation for R
d − 1 d−1 −1
− |S |R + O(R−4 + R−2 µ + µ2 ) = 0. (22.29)
2
This equation has the solution of the form R = O(µ−2 ).
(f) Summing up, we expect that if R is large enough, there exists a unique spherically
symmetric solution to the stationary Cahn-Hilliard equation of the form χR + α(R, µ),
d−1
where χR is defined in (22.14) and α(R, µ) = O(R 2 (R−2 + µ)).

Finally, we explain how to construct the family, φR (r), R > 0, of approximate solu-
tions. We assume R is sufficiently large and µ > 0, sufficiently small. To construct the
family, φR (r), R > 0, of approximate solutions to (22.13), we look for a spherically sym-
metric solution to (22.10), satisfying (approximately) the boundary conditions v(0) = −1
and v(∞) = 1.
As was discussed above, the shifted kink χR (r) := χ(r − R), where r = |x|, is not a
suitable approximate solution and we have to modify it. We write the family, φR (r), R >
0, of approximate solutions to (22.13) as follows

φR (r) = χR (r) + ηR (r), (22.30)

where ηR (r) := η(r − R). To be specific, we take g(u) = u3 − u. We show that ηR (r) can
be choosen so that

F (φR , µ) → 0, as r → ∞, (22.31)
F (φR , µ) is bounded, as r → 0. (22.32)
3 2
Namely, we take η(∞) = ηR (∞) = δ∞ , with the number δ∞ satisfying δ∞ −3δ∞ +2δ∞ = µ,
0 0 0 0
which gives δ∞ = O(µ), and η (−R) = ηR (0) = χR (0) = χ (−R). Indeed, using that χR
satisfies the equation (2.5), we compute

d−1 d−1
F (φR , µ) = − ∂r χR − ∂r2 ηR − ∂r ηR − ηR + 3ηR χ2R + 3ηR2 χR + η 3 − µ. (22.33)
r r
Since ∂r χR → 0 and χR (r) → −1, as r → ∞, the first relation requires that

−ηR + 3ηR − 3ηR2 + ηR3 − µ → 0,

as r → ∞. Hence, it suffices to choose η(∞) = ηR (∞) = δ∞ , with the number δ∞


3 2
satisfying δ∞ − 3δ∞ + 2δ∞ = µ. This gives δ∞ = O(µ) and F (φR , µ) → 0, as r → ∞.
230 Lectures on Applied PDEs, November 3, 2017

For the second relation, (22.32), we require that ∂r χR + ∂r ηR = O(r), as r → 0, and


therefore choose
η 0 (−R) = ηR0 (0) = χ0R (0) = χ0 (−R),
which gives (22.32). Finally, for simplicity, we take ηR (r) = 0, for R/2 ≤ r ≤ 3R/2.
Using the above expression, the estimate (22.19) and a convenient choice of ηR , we
obtain the simple estimates (this requires a fair amount of work not displayed here)
|F (φR , µ)| . e−δR , for r ≤ R/2 or r ≥ 3R/2, and |F (φR , µ)| . (R−1 + µ), for R/2 ≤ r ≤
3R/2. Using these estimates, we find
d−1
kF (φR , µ)kH r . R 2 (R−1 + µ). (22.34)
d−1
(The term R 2 comes from the volume of the shell of radius R and thickness 1.)
This completes the construction of the approximate solution φR . The Lyapunov-
Schmidt reduction based on this solution follows the outline above for the decomposition
(22.21), i.e. for the rough approximate solution χR and is done in Appendix to this
section.

22.1 Appendix. Details of the Lyapunov-Schmidt reduction


Our goal is to solve the equation (22.10) near φR (r), for some R depending on µ to be
chosen later. We look for a solution u(x) of (22.13) in the form u(x) = φR (r) + ξ(r), and
try to solve for the fluctuation term ξ. We follow the approach sketched above for the
decomposition (22.21), i.e. for the rough approximate solution χR . We begin with the
following proposition

Proposition 22.2. The operator LR = −∆r + g 0 (χR ) has the following properties:

1. The smallest point, inf σ(LR ), of the spectrum of LR is a non-degenerate, isolated


eigenvalue of size O R1 and with an approximate eigenfunction χ0R (r);

2. For large enough R, the gap between the smallest eigenvalue and the rest of the
spectrum is O(1).

Proving this proposition, as well as the corollary below, requires a fair amount of the
spectral theory, see e.g. [6, 8] and Appendix E (to be expanded), and the proof is given
below. The property 2 shows that

Corollary 22.3. The operator LR is invertible on the orthogonal complement of the sub-
space spanned by the vector χ0R (r) and its inverse on this orthogonal complement is uni-
formly bounded in R.
Lectures on Applied PDEs, November 3, 2017 231

Now, we follow closely the analysis in the main text. It consists of the following steps:
(a) Parameterization of solutions u(x). We parameterize our solution u(x) by a pair
(R, ξ), where R > 0 and the function ξ(r), so that u(x) can be represented uniquely as

u(x) = φR (r) + ξ(r), ξ ⊥ φ0R . (22.35)

where φ0R ≡ ∂R φR . The condition ξ⊥φ0R (in L2 (Rd )) determines R. Indeed, observing
that ∂R φR ≈ −φ0R and applying the implicit function theorem to the equation f (R, v) :=
hv − φR , φ0R i = 0, we obtain the unique solution R = R(v).
To clarify the geometric meaning of th decomposition (22.35), we define the kink
manifold M := {φR (r)| R > 0} and and observe that (22.35) is equivalent to projecting
u(x) to this manifold, which gives the point φR (r) on M and the function ξ(r) which
is orthogonal to this manifold, ξ ⊥ TφR M. Since the tangent space to M at φR (r) is
spanned by φ0R (r), this leads to (22.35). The function ξ(r) is called the fluctuation of u(x)
around φR (r).
(b) Decomposition of the equation. It is convenient to replace M := {φR (r)| R > 0}
by its approximate Mkink := {χR (r)| R > 0}. Plug the decomposition (22.35) into
(22.13) and project the resulting equation onto the tangent space TχR M and its orthogonal
complement, (TχR M)⊥ to obtain the equations for two unknowns R and ξ:

PR F (φR + ξ, µ) = 0, (22.36)
PR⊥ F (φR + ξ, µ) = 0, (22.37)

where PR be the orthogonal projector onto χ0R , namely, PR f := χ0R χ0R f /norm, and
R

PR⊥ = 1l − PR . These are two equations for two unknowns ξ and h.


(c) Solution of the second equation (22.37) for ξ. (Here the choice of the family
φR (r), R > 0, plays a crucial role.)

Lemma 22.4. For R−1 and µ sufficiently small, the equation (22.37) has a unique solu-
d−1
tion for ξ = ξ(R, µ) and this solution satisfies the estimates ξ(R, µ) = O(R 2 (R−2 + µ)).

We prove this lemma later. Meantime we explain the idea of the proof. Expand
F (φR + ξ, µ) in ξ to obtain

F (φR + ξ, µ) = F (φR , µ) + du F (φR , µ) ξ + N (ξ, µ), (22.38)


| {z }
LR,µ

where N (ξ, µ) is defined by this equation, explicitly,

N (ξ) := g(φR + ξ) − g(φR ) − g 0 (φR )ξ.


232 Lectures on Applied PDEs, November 3, 2017

Plug this into (22.37) and use the notation L⊥ ≡ L⊥ ⊥ ⊥


R,µ := PR du F (φR , µ)PR , restricted
to Ran PR⊥ , to obtain the equation
PR⊥ F (φR , µ) + L⊥ ξ + PR⊥ N (ξ, µ) = 0. (22.39)
By Corollary 22.3, the operator L⊥
R,µ is invertible, with the uniformly bounded inverse.
Hence we can rewrite (22.39) as
ξ = Φ(ξ, µ), where Φ(ξ, µ) := −(L⊥ )−1 PR⊥ [F (φR , µ) + N (ξ, µ)]. (22.40)
The estimation of the map Φ(ξ, µ) is done below and shows that it is a contraction for
R sufficiently large, which implies that that (22.40) has a unique solution, ξ = ξ(R, µ),
d−1
satisfying the estimate ξ(R, µ) = O(R 2 (R−2 + µ)).
(d) Derivation of the reduced equation. Plug the solution ξ = ξ(R, µ) of (22.37) into
(22.36) to obtain the new equation for R,
f (R, µ) := hχ0R , F (φR + ξ(R, µ), µ)i = 0. (22.41)
This is a scalar equation for a single unknown R (the reduced equation). Solve the
reduced equation for R. Then φR + ξ(R, µ), where R is a solution to (22.42), is a solution
to the stationary Cahn-Hilliard.
(e) Solution of the reduced equation for R. Since PR f := χ0R χ0R f /norm, we can
R

rewrite (22.41) as hχ0R , F (φR + ξ, µ)i = 0. Remembering the expansion (22.38), this gives
hχ0R , F (φR , µ)i + hχ0R , Lξ(R, µ)i + hχ0R , N (ξ(R, µ)i = 0. (22.42)
Using (22.33) and (??), we compute

d − 1 d−1 1 0 2 d−1
Z
hχ0R , F (φR , µ)i dr + O Rd−1 µ ,

≈− |S | χR r
2 0 r
 d−1 
hχ0R , Lξi = hLχ0R , ξ(R, µ)i = O R 2 −2 kξkL2 ,

hχ0R , N (ξ)i = O kξk2H 1 + kξk3H 1 .




For ξ = ξ(R, µ), the last two estimates give by Lemma 22.4, hχ0R , Lξi = O(Rd−3 (R−2 +µ))
and hχ0R , N (ξ)i = O(Rd−1 (R−4 + µ2 )). Inserting the estimates above into (22.42) and
dividing by Rd−1 , gives
d − 1 d−1 −1
− |S |R + O(R−4 + R−2 µ + µ2 ) = 0. (22.43)
2
This equation has a solution R = O(µ−2 ).
Summing up, we have shown that if R is large enough, there exists a unique spherically
symmetric solution to the stationary Cahn-Hilliard equation of the form φR + ξ(R, µ),
d−1
where φR is defined in (22.30) and ξ(R, µ) = O(R 2 (R−2 + µ)). Thus, the Cahn-Hilliard
equation has bubble solutions.
Lectures on Applied PDEs, November 3, 2017 233

Proof of Lemma 22.4. We estimate the map Φ(ξ, µ). To keep things specific we consider
the canonical nonlinearity g(u) = u3 − u and d = 3. Then N (ξ) = 3φR ξ 2 + ξ 3 . Using this
expression, the Sobolev embedding theorems and the simple inequality kξ m − η m kL2 ≤
m kξkm−1 m−1

H 1 + kηkH 1 kξ − ηkH 1 , we obtain

kN (ξ)kL2 . ρ2 , kN (ξ) − N (η)kL2 . ρkξ − ηkH 1 , (22.44)

provided kξkH 1 , kηkH 1 ≤ ρ ≤ 1. Next, by (22.34) and the boundedness of PR⊥ , we have
d−1
kPR⊥ F (φR , µ)kH r . R 2 (R−2 + µ). (22.45)

Using the last two relations, the uniform boundedness of L−1 and (22.40), we obtain
d−1
kΦ(ξ, µ)kH 1 . R 2 (R−2 + µ) + ρ2 ,
kΦ(ξ, µ) − Φ(η, µ)kH 1 . ρkξ − ηkH 1
d−1 d−1
Hence for ρ < 1 s.t. R 2 (R−2 + µ) + ρ2  ρ, which is easy to satisfy if R 2 (R−2 + µ)
sufficiently small (which means that R is large for d < 5), we have that Φ is a strict
contraction on the ball in H 1 of radius ρ, centered at the origin, and it has a unique fixed
point in this ball.

(Fill in details!)

Proof of Proposition 22.2. The proof below uses the definition and properties of the es-
sential spectrum (see [6, 8] and Appendix E (to be expanded)). The first statement
follows from the spectral fact that σess (L) = σ(L∞ ), where L∞ is the evaluation of L at
infinity: L∞ := −∆r + g 0 (χR (∞)) = −∆r + g 0 (1), and the simple computation, which uses
the definition of the spectrum (see [6, 8]) and the Fourier transform,

σ(L∞ ) = σ(−∆r + g 0 (1)) = [g 0 (1), ∞).

To prove the second statement, we observe that the equation (22.20) implies that
1
hχ0R , L 0

χ
R R i = O R
which shows that the operator LR has an eigenvalue of the order
1
O R below its essential spectrum. Using the fact that −χ0R is a positive function by the
Perron - Frobenius theory, we conclude that, for  R large enough, the smallest eigenvalue
of LR is non-degenerate and of the order O R1 , with the approximate eigenvector χ0R .
To prove the third statement, we perform a geometric analysis of L. Let (j0 , j1 ) be a
partition of unity, normalized as j02 + j12 = 1, and with supp j0 ⊂ {x ∈ Rd ||x| ≤ R/2} and
supp j1 ⊂ {x ∈ Rd ||x| ≥ R/3} and |∂ m ji | . R−m . We use the IMS formula
X X
L := ji Lji − |∇ji |2 , (22.46)
i i
234 Lectures on Applied PDEs, November 3, 2017

which holds for P1any 2operatorPL of the formPL = −∆ + PV1 (x).2 To P


prove it, we use the
relations L = i=0 ji L = − i=0 ji [ji , ∆] + i=0 ji V ji , i=0 ji ∆ = 1i=0 ji ∆ji + ji [ji , ∆]
1 1

and ji [ji , ∆] = −ji (2∇ji · ∇ + ∆ji ), to obtain


1
X 1
X
L=− ji ∆ji + ji (2∇ji · ∇ + ∆ji ) + ji V ji .
i=0 i=0

Similarly, we have
1
X 1
X
L=− ji ∆ji + (2∇ji · ∇ + ∆ji )ji + ji V ji .
i=0 i=0

Subtracting
P1 the second P1 relation from the first and dividing the result P by 2,2 we obtain
L = − i=0 ji ∆ji + i=0 ji V ji + [jP  i , ∇ji · ∇] = − i |∇ji | , this gives
i , ∇ji · ∇]. Since [j
(22.46). Since, by the choice of ji , i |∇ji |2 = O R12 , (22.46) implies that
1
X 1 
L := ji Lji + O . (22.47)
i=0
R2

We estimate each term ji Lji separately, using the properties of supp ji .


First we observe that χR ≈ 1 and therefore L ≈ −∆ + g 0 (1) ≥ g 0 (1) > 0 on {x ∈
R ||x| ≥ R/3}. Hence j1 Lj1 ≥ j1 g 0 (1)j1 = g 0 (1)j12 .
d

To show that the operator j0 Lj0 is bounded below by cj02 on the subspace (χ0R )⊥ , for
some c > 0 of the order O(1), we notice that the operator

d−1
L = −∂r2 − ∂r + g 0 (χR (r))
r
on L2 ([0, ∞), rd−1 dr) is unitarily equivalent to the operator

d−1 1
− ∂r2 + 2
+ g 0 (χR (r)) (22.48)
4 r
d−1 d−1
on L2 ([0, ∞), dr). Indeed, the transformation u 7→ r 2 u from L2 (rd−1 dr) to L2 (r 2 dr)
is unitary and maps L into (22.48) as follows from the computation

d−1 d−1 d−1 d2 − 1


(∂r2 + ∂r )(r− 2 v) = r− 2 (∂r − )v.
r 4r2
Next, under the unitary shift ξ(r) → ξ(r − R), the operator (22.48) is unitarily equivalent
to
d−1
L̃ := −∂r2 + 2
+ g 0 (χ(r))
4(r + R)
Lectures on Applied PDEs, November 3, 2017 235

on L2 ([−R, ∞], dr).


Now, under the unitary shift ξ(r) → ξ(r − R), supp j0 is mapped into a subset of
{r ≥ −2R/3} and therefore L̃ = L∗ + O(1/R) on {r ≥ −2R/3}, where L∗ := −∂r2 +
g 0 (χ(r)). The operator L∗ is independent of R, has a non-degenerate eigenvalue 0 with
the eigenfuction χ0R and the essential spectrum σess (L∗ ) = [g 0 (−1), ∞). Hence L∗ ≥ c
on the subspace (χ0R )⊥ , for some c > 0 of the order O(1), and therefore, by the unitary
equivalence, this implies that j0 Lj0 ≥ c0 j02 on the subspace (χ0R )⊥ .
The estimates j0 Lj0 ≥ c0 j02 on the subspace (χ0R )⊥ and j1 Lj1 ≥ g 0 (1)j12 , together with
the relation (22.47), imply that L ≥ c = min(c0 , g 0 (1)) on the subspace (χ0R )⊥ , for large
enough R. This gives the assertion four.
∗∗∗∗∗

A Spaces and Operators: Review


In this section, we introduce the simplest and most commonly used spaces, Banach and
Hilbert spaces, and describe their most important examples.

A.1 Vector Spaces


We begin with the key definitions. A vector space, X, is a collection of elements, denoted,
u, v, ..., for which the operations of addition, (u, v) → u + v and multiplication by a (real
or complex) number, (α, u) → αu, are defined in such a way that

(α + β)u = αu + βu
and
α(u + v) = αu + αv.
Recall that the operations of addition and multiplication have the following properties

u + v = v + u (commutativity)
and
α(βu) = (αβ)u (associativity).
Elements of a vector space are called vectors. As will be clear from the context most of
the vector spaces we consider in these lectures are defined for multiplication by complex
number (they are said to be vector spaces over complex numbers).
Let Ω ⊂ Rn be an open set of Rn . Main examples of vector spaces are:

(a) Rn = {x = (x1 , ..., xn )| − ∞ < xj < ∞ ∀ j}– the Euclidean space of dimension n;

(b) C k (Ω) – the space of k times continuously differentiable complex functions on Ω;


236 Lectures on Applied PDEs, November 3, 2017

(c) Lp (Ω)
R – the space of (measurable) complex functions, f on Ω s.t. |f |p is integrable,
i.e. Ω |f |p < ∞;

(e) S(Rn ) – space of C ∞ functions vanishing at ∞ together with all their derivatives
faster than |x|−n for all n.

The spaces Lp (Ω), 1 ≤ p ≤ ∞, are introduced and studied in Section A.3.


The sets above are vector spaces if we define addition and multiplication by real/complex
numbers in the point-wise way. Namely, for Rn we define

(x + y)j = xj + yj and (αx)j = αxj ∀ j


and, for the remaining spaces in (b) – (e), we define

(f + g)(x) := f (x) + g(x) and (αf )(x) := αf (x) ∀x ∈ Ω.

Exercise A.1. Show that (a)-(e) are vector spaces. (Hint: For (c), use the inequality
|f + g|p ≤ 2p (|f |p + |g|p ) (see (A.3) below).)

A.2 Banach Spaces


To measure the size of vectors, one uses the notion of norm. A norm is defined to be a
map

X 3 u → kuk ∈ [0, ∞)
which has the following properties:

(a) kuk = 0 ⇐⇒ u = 0;

(b) kαuk = |α|kuk;

(c) ku + vk ≤ kuk + kvk.

The last inequality is called the triangle inequality. (Observe an unusual notation for the
norm k · k, not f (·) or n(·) as one would denote other maps). A vector space equipped
with a norm is called a normed vector space.
Having defined a norm, we can define the notion of (norm–) convergence as follows.
Let {fn } ⊂ X be a sequence. We say that fn converges to f (∈ X), if and only if
kfn − f k → 0. We write fn → f .
A sequence {fn } ⊂ X is called a Cauchy sequence iff kfn − fm k → 0, as m, n → ∞.
A normed vector space X is called complete if and only if every Cauchy sequence
converges, i.e. if {fn } ⊂ X is a Cauchy sequence then {fn } converges (i.e. there is a
f ∈ X such that ||fn − f || → 0, as n → ∞). Remark that the converse is always true: any
Lectures on Applied PDEs, November 3, 2017 237

convergent sequence is necessarily a Cauchy sequence. A complete normed vector space


is called a Banach space.
Completeness is a very important property of a normed vector space, e.g. often one
solves an equation by successive approximations, and one wants to know that such ap-
proximate solutions converge to an actual solution.
Examples: The vector spaces Rn and Lp (Ω) defined above are Banach spaces under
the norms
Xn Z 1/p
2 1/2 p
|x| := ( xi ) , kf kp := |f | if 1 ≤ p < ∞.
i=1 Ω

Denote by Cbk (Ω) the subspace of C k (Ω) consisting of all functions in C k (Ω) which are
bounded together with all their derivatives up to the order k. We equip the space Cbk (Ω)
with the norm
X
kf kC k = sup |∂ α f (x)|,
x∈Ω
|α|=k
Qn α Pn
where α = (α1 , ..., αn ) with αj non-negative integers, ∂ α := j=1 ∂xjj and |α| = i=1 αj .

Exercise A.2. Show that

(a) Cbk (Ω) is a vector space;

(b) kf kC k is a norm on Cbk (Ω).

It is shown in [4], Proposition 4.13 and Exercise 5.7 that the spaces Cbk (Ω) are complete,
i.e., that they are Banach spaces.

A.3 Lp –spaces
Consider an open subset Ω of the Euclidean space Rn . In particular Ω can be a bounded
subset or the entire Rn . Let dx denote the Lebesgue measure on Rn . We define the
Lp –space for 1 ≤ p < ∞:
Z
L (Ω) := {f : Ω → C | f is measurable, and
p
|f |p dx < ∞}. (A.1)

In other words, f ∈ Lp (Ω) ⇔ |f |p ∈ L1 (Ω). We define the L∞ -space:

L∞ (Ω) := {f : Ω → C | f is measurable, and ess sup |f | < ∞}, (A.2)

where, recall that ess sup |f | := inf{sup |g| : g = f a.e.}. We often use the abbreviation
Lp for Lp (Ω).
238 Lectures on Applied PDEs, November 3, 2017

Strictly speaking, elements of Lp (Ω) are equivalence classes of measurable functions:


two functions define the same elements of Lp (Ω) if they differ only on a set of measure 0.
Lp , 1 ≤ p ≤ ∞ is a vector space. For p = ∞, this is obvious, and for 1 ≤ p < ∞, it
easily follows from the inequality
|f + g|p ≤ 2p (|f |p + |g|p ) . (A.3)
The latter inequality is obtained as follows: |f + g|p ≤ (2 max(|f |, |g|))p ≤ 2p (|f |p + |g|p ).
We define for every f ∈ Lp :
p 1/p
 R 
kf kp := |f | if 1 ≤ p < ∞, (A.4)
ess sup|f | if p = ∞.
Clearly, kf kp = 0 ⇔ f = 0 a.e., and kcf kp = |c| kf kp , ∀c ∈ C. We have also the triangle
inequality kf + gkp ≤ kf kp + kgkp , which we prove later.
From these properties, it follows that the map f 7→ kf kp is a norm. Hence Lp is a
normed vector space for every 1 ≤ p ≤ ∞. In fact, it is a Banach space. The proof is
given in [4], Theorem 6.6, Proposition 6.7 and Theorem 6.8.
There are three basic inequalities for Lp spaces:
1. Hölder’s inequality: Let 1 ≤ p, q, r ≤ ∞, p−1 + q −1 = r−1 , and f ∈ Lp , g ∈ Lq then
f g ∈ Lr and kf gkr ≤ kf kp kgkq .
2. Minkowski’s inequality: Let 1 ≤ p < ∞ and f, g ∈ Lp , then kf + gkp ≤ kf kp + kgkp .
We prove these inequalities in Appendix A.4. Here we prove Hölder’s inequality in easy,
but instructive, special cases, (p−1 , q −1 ) = (0, 1) and (p−1 , q −1 ) = (1/2, 1/2).
For (p−1 , q −1 ) = (0, 1), we have
Z Z
|f g| ≤ sup |f | |g| = kf k∞ kgk1 .

In the case (p−1 , q −1 ) = (1/2, 1/2), we define first the unit vectors fˆ = f /kf k2 and
ĝ = g/kgk2 (here we can assume f 6= 0 and g 6= 0 otherwise the result is trivial), so that
kfˆk2 = 1 and kĝk2 = 1 (check this!). Integrating the inequality

|fˆ|2 + |ĝ|2 − 2|fˆĝ| ≥ 0,


|fˆ|2 = 1 = |ĝ|2 , we obtain
R R
and using that
Z
1 ≥ |fˆĝ|. (A.5)

Since the integral on the right hand side is


Z
|f g|/kf k2 kgk2 ,
Lectures on Applied PDEs, November 3, 2017 239

we find after multiplying (A.5) by kf k2 kgk2 ,


Z
kf k2 kgk2 ≥ |f g|

as claimed. Note that this inequality, which is a special case of the Hölder inequality, has
its own name - the Schwarz inequality (with p = q = 2 and r = 1). One can prove the
general Hölder inequality by interpolating between the (0, 1) and (1/2, 1/2) cases. But it
is more elementary to prove it directly as we do in the next paragraph.
For more inequalities, see Appendix A.4.
We can further generalize the C and Lp spaces to spaces of continuous and Lp integrable
functions from an open set Ω ⊂ Rn to a Banach space X. We denote such spaces as
C(Ω, X) and L( Ω, X), respectively.

A.4 Supplement: Proofs of Hölder’s and Minkowski’s inequali-


ties
First, we prove Hölder’s inequality. Observe first that it suffices to proveR the Hölder
inequality for the case r = 1 (this follows easily from kf gkr = ( |f g| ) = ( |f |r |g|r )1/r ,
r 1/r
R

and e.g. calling f r = f1 and g r = g1 ).


We notice that the result is trivial if ||f ||p = 0 or ||g||p = 0 (for then f = 0 a.e. or
g = 0 a.e.), or if ||f ||p = ∞ or ||g||p = ∞. If neither of these cases hold, then we can
define
f (x) p g(x) q

1
a= , b= and λ = .
kf kp kgkq p
Below, we will show that for any a, b > 0, and 0 < λ < 1:

aλ b1−λ ≤ λa + (1 − λ)b. (A.6)

Applying this to our case, we get

|f (x)g(x)| |f (x)|p |g(x)|q


≤ R + R .
kf kp kgkq p |f |p q |g|q

Integrating this inequality, and using p−1 + q −1 = 1, we arrive at Hölder’s inequality.


We finish the proof by showing (A.6). We can assume b 6= 0, so we can divide (A.6)
by b on both sides. (A.6) is then equivalent to the inequality (where t = a/b)

tλ − λt − 1 + λ ≤ 0.

Since 0 < λ < 1, the maximum of the function on the l.h.s. is reached at t = 1, and is
equal to 0.
240 Lectures on Applied PDEs, November 3, 2017

Now we prove Minkowski’s inequality. If either p = ∞ or p = 1, then the result is


obvious. Now assume p > 1 and f + g 6= 0. Then we estimate

|f + g|p ≤ (|f | + |g|)|f + g|p−1


Integrating this inequality and applying Hölder inequality with exponents p and p0 (1/p +
1/p0 = 1) we obtain

Z Z Z
p p−1
|f + g| ≤ |f ||f + g| + |g||f + g|p−1

≤ kf kp k |f + g|p−1 kp0 + kgkp k |f + g|p−1 kp0

Taking into account that (p − 1)p0 = p, we find furthermore that


p
Z
p p0
|f + g| ≤ (kf kp + kgkp )kf + gkp

Now remember that |f + g|p = kf + gkpp and observe that p/p0 = p − 1. Hence dividing
R

both sides of the latter inequality by kf + gkp−1


p gives Minkowski’s inequality. 
Remark 2. Inequality (A.6) is a special case of the very useful Jensen’s inequality : let
ϕ be a convex function on [a, b] (i.e., ϕ satisfies ϕ(tx + (1 − t)y) ≤ tϕ(x)
Pn + (1 − t)ϕ(y)
for all t ∈ [0, 1] and x, y ∈ [a, b]), and pk positive numbers satisfying 1 pk = 1. We can
think about pk as probabilities. Then
n
X n
X
ϕ( pk tk ) ≤ pk ϕ(tk ), (A.7)
1 1

for all tk ∈ [a, b]. This inequality indeed implies (A.6) for ϕ(t) = et , n = 2, p1 = λ, t1 =
ln a and t2 = ln b.
Exercise A.3. Prove Jensen’s inequality (A.7).
Theorem 20 (Hausdorff-Young inequality). Let 1 ≤ p ≤ 2 and p−1 + q −1 = 1. Then
kfˆkq ≤ kf kp .

We sketch the proof of the Hausdorff-Young inequality. Clearly, kfˆk∞ ≤ kf k1 . More-


over, we have shown that kfˆk2 = kf k2 . For 1 < p < 2, the result follows from an
interpolation theorem (see e.g. [4]).
Note that the Hausdorff-Young inequality implies that F extends to a bounded oper-
ator from Lp to Lq . We omit the proof of this statement.
Theorem 21 (Riemann-Lebesgue Lemma). Suppose that g is s.t. ĝ ∈ L1 . Then
i) g is bounded and continuous,
Lectures on Applied PDEs, November 3, 2017 241

ii) g decays at infinity: lim|x|→∞ g(x) = 0.


Proof. i) The boundedness is easily seen: ∀x,
Z Z
ikx

|g(x)| = e ĝ(k) ≤ |ĝ(k)| = kĝkL1 .

Next, we show continuity. Since ĝ ∈ L1 , then


Z
lim (g(x + h) − g(x)) = lim eik·x eik·h − 1 ĝ(k)dk = 0,

h→0 h→0

by the dominated convergence theorem (|eik·h − 1||ĝ| ≤ 2|ĝ|). This shows that g is contin-
uous. Next, let us show ii). Since the Schwartz space S is dense in L1 , there is a sequence
ϕj ∈ S such that ||ϕj − ĝ||1 → 0 as j → 0. Thus
Z
||ϕ̌j − g||∞ ≤ |ϕj (k) − ĝ(k)| = ||ϕj − ĝ||1 → 0,

which shows that ϕ̌j → g uniformly on Rn . But ϕ̌j ∈ S, so ϕ̌j → 0 as |x| → ∞, and
therefore g → 0 as |x| → ∞.

A.5 Hilbert Spaces


To measure angles between vectors one introduces the inner product (sometimes called
the scalar product ).
The map (f, g) → hf, gi ∈ C is called an inner product if and only if
• hf, gi is linear in the second argument, i.e. hf, αg + βhi = α hf, gi + β hf, hi, for
any α, β ∈ C,
• hf, gi = hg, f i (this together with the above implies that hf, gi is anti-linear in the
first argument: hαf + βh, gi = α hf, gi + β hh, gi),
• hf, f i ≥ 0 with equality if and only if f = 0.
We remark that sometimes, the scalar product is taken to be linear in the first argument
and anti-linear in the second one.
The space L2 (Ω) is a Hilbert space if we define the inner product as
Z
hf, gi := f g. (A.8)

Exercise A.4. Show that (A.8) is an inner product.


An inner product defines a norm according to
p
kf k = hf, f i. (A.9)
242 Lectures on Applied PDEs, November 3, 2017

The Schwarz inequality


| hf, gi | ≤ kf k kgk.
It follows from the obvious relation

0 ≤ ku ± vk2 = hu ± v, u ± vi = kuk2 + kvk2 ± hu, vi ± hv, ui (A.10)

applied to u = ±f /kf k and v = g/kgk and to u = ±if /kf k and v = g/kgk.


Exercise A.5. Check that (A.9) defines a norm. (Hint: to prove the triangle inequality,
use the Schwarz inequality).
Thus a space with an inner product, or an inner product space, is also a normed space.
A complete inner product space is called a Hilbert space. Clearly a Hilbert space is also a
Banach space. An inner product, in addition to measuring sizes and distances, measures
relative directions, e.g.,
f ⊥ g ⇔ hf, gi = 0.
Given a norm kf k, how can we tell whether this norm comes from an inner product?
The answer to this question is that kf k comes from an inner product if and only if it
satisfies the parallelogram law

kf + gk2 + kf − gk = 2kf k2 + 2kgk2 .


In this case the inner product is given by (in the real case)
1
kf + gk2 − kf k2 − kgk2

hf, gi = (A.11)
2
(see (A.10)).
Exercise A.6. Check that (A.11) defines a (real) inner product.
The space Rn has an inner product defined by x · y = ni=1 xi yi .
P

A.6 Sobolev spaces


Now we want to introduce an additional structure on Lp –spaces which measures smooth-
ness, similar to the smoothness properties of functions in C k . We do so only for p = 2,
i.e. for the space L2 (Rn ). This is the most useful space among the Lp –spaces as it has an
inner product: Z
hf, gi := fg

and therefore it is a Hilbert space


p (i.e., recall, an inner product space which is complete with
respect to the norm kf k := hf, f i induced by the inner product). Another advantage
of the L2 –space is that the Fourier transform leaves it invariant (i.e. f ∈ L2 ⇒ fˆ ∈ L2 ).
Lectures on Applied PDEs, November 3, 2017 243

We now define for s integer, s ≥ 0, the new spaces

Hs (Rn ) = {f ∈ L2 (Rn ) : ∂ α f ∈ L2 (Rn ) ∀α s.t. |α| ≤ s}. (A.12)

This definition is very similar to the definition of the C s (Rn )–spaces: in fact, by
replacing L2 (Rn ) in (A.12) by C(Rn ), one obtains the definition of C s (Rn ). But there
is one crucial difference: in the C s (Rn )–case, the functions f are assumed to be s times
continuously differentiable, but in the Hs –case, they are not. Namely, the derivatives ∂ α f
in the above definition are understood in the distributional sense:

∂ α f (ϕ) := (−1)|α| f (∂ α ϕ),


R
where f (ϕ) = f ϕ, and ϕ ∈ S(Rn ). In other words ∂ α f is a linear functional defined
by the r.h.s. of the above equation and the
2 n α
R relation in (A.12) says that ∀α with |α| ≤ s,
there exists a gα ∈ L (R ) s.t. ∂ f (ϕ) = gα ϕ ∀ϕ.
The space Hs (Rn ) equipped with the inner product
X
hf, hi(s) = h∂ α f, ∂ α gi (A.13)
|α|≤s

is a Hilbert space.
There is another way of defining the spaces Hs (Rn ) using the Fourier transform defined
in the next appendix:

Hs (Rn ) = {f ∈ L2 (Rn ) : hkis fˆ(k) ∈ L2 (Rn )}, (A.14)

where hki = (1 + |k|2 )1/2 . The definition (A.14) has the advantage that it makes sense for
an arbitrary s ∈ R. Besides, it does not require extra explanations. Of course we have
to show that definitions (A.12) and (A.14) are equivalent for positive integers s. One can
show easily that the definitions (A.12) and (A.14) are equivalent for positive integers s.
The following result is often used in applications

Theorem A.1 (Sobolev embedding theorem). H 2 (Ω) ⊂ C α (Ω), with α < k + 2 − n2 .

A.7 Linear operators


Linear operators or simply operators are linear maps from one vector space Y into another
vector space X. We denote linear operators usually by capital roman letters, A, B, . . .
and use the notation
A:Y →X (A.15)
and Au to denote an operator A mapping Y into X and application of A to a vector
u ∈ Y , respectively. To define an operator A, we have to give a rule that prescribes to
244 Lectures on Applied PDEs, November 3, 2017

each element of Y an element of X (the image of u). We require this rule to be linear,
i.e. ∀u, v ∈ Y , and α, β ∈ C:

A(αu + βv) = αAu + βAv. (A.16)

To fix ideas here and in what follows, we consider vector spaces over the complex numbers
C, i.e. complex vector spaces. All the material of this section, except for spectral theory,
remains unchanged if we substitute R for C.
If the space Y in (A.15) is a subset of the space X then sometimes one calls Y the
domain of A (in X) and denotes it D(A) ≡ Y . In this case we say that A is defined in X
with domain D(A)(= Y ). The range (or image) of A is defined as

Ran (A) := {Au : u ∈ Y } ≡ AY.

Ran (A) is a vector space (show this). We may assume that D(A) is dense in X, i.e. for
any u ∈ X, there is a sequence {un } ⊂ D(A) s.t. un → u as n → ∞. Indeed, if D(A) is
not dense to begin with, we consider instead of the space X simply the space Y := D(A),
the closure of D(A), which is obtained by adding to Y limits of all possible sequences
{un } convergent in X.

Examples.
1) The identity operator 1l : Lp → Lp ;
2) The multiplication operator Mf : L→ Lp , u 7→ f u for a fixed f ∈ L∞ ;
3) The differentiation operator ∂x∂ j in L2 (Rn ) with the domain D( ∂x∂ j ) = H 1 (Rn );
P ∂2
4) The Laplacian ∆ := n1 ∂x 2
2 in L (Ω) with the domain D(∆) = H (Ω);
2
j
5) Integral operators, i.e., operators of the form
Z
(Ku)(x) = K(x, y)u(y)dy,

for some function K(x, y) (called the kernel or integral kernel). The domain and range of
the integral operator K depend on the properties of the kernel K(x, y).
6) The Fourier transform F : L1 (Rn ) → L∞ (Rn ),
Z
−n/2
F : u(x) → (2π) e−ik·x u(x)dx;

7) Convolution operator Cf : Lp → Lp , Cf : u → f ∗ u for fixed f ∈ L1 , where


Z
(f ∗ u)(x) = f (x − y)u(y)dy;

8) The wavelet transform


Lectures on Applied PDEs, November 3, 2017 245

Z
Wψ : f → ψab f dx,

|ψ̂(k)|2
where ψab = √1 ψ x−b
 R
a
for a fixed function ψ satisfying k
dk < ∞.
|a|

Note that the Fourier transform and convolution are integral operators with the inte-
gral kernels (2π)−n/2 eik·x and f (x − y), respectively.
In fact also in examples 1)-4), the operators can be represented as integral operators,
but with distributional kernels, e.g. K(x, y) = f (x)δ(x − y) for Mf , and K(x, y) =
0
−δ (xj − yj ) i6=j δ(xi − yi ) for ∂x∂ j .
Q

For an operator A : Y → X we define the norm

kAk ≡ kAkY →X = sup kAukX . (A.17)


kukY =1

If Y = X and kAk < ∞, then A is said to be bounded (in X). Observe that our definition
implies that

kAuk ≤ kAkkuk (A.18)


for all u ∈ Y . Now, if there is a constant C (independent of u) such that

kAuk ≤ Ckuk, (A.19)

for all u ∈ D(A), and if the space X is complete (i.e. a Banach space), then the operator A
can be extended by continuity to the whole space X as a bounded operator. The smallest
constant C satisfying (A.19) is the norm kAk of A, and so kAuk ≤ kAk kuk (so bounded
operators form a Banach algebra).
Examples of bounded operators are the identity operator, 1, of example 1) above (in
fact, k1k = 1), the multiplication operator, Mf , of example 2), the integral operator, K,
of example 6) with a kernel K(·, ·) ∈ L2 (Rn × Rn ), as an operator from L2 (Rn ) to L2 (Rn ).

Exercise A.7. Show that

(1) kABk ≤ kAkkBk;

(2) kwk = supkvk=1 | hw, vi |, and therefore

kAk = sup | hAu, vi |;


kuk,kvk=1

(3) kMf k = kf k∞ (Example 2 above);


246 Lectures on Applied PDEs, November 3, 2017

(4) for integral operators (Example 5 above)


Z 1/2
2
kKkL2 →L2 ≤ |K(x, y)| dxdy

provided the r.h.s. is finite.

Exercise A.8. Show that the differentiation operator in example 3) is not bounded in
L2 (Rn ) by finding a sequence fn of functions from D( ∂x∂ j ) such that kfn k ≤ 1, ∀n, and
k ∂x∂ j fn k2 → ∞, as n → ∞.

It is easy to see that F : L1 (Rn ) → L∞ (Rn ),


Z Z
−n/2 −n/2
ix·k

(2π)
e f (x)dx ≤ (2π)
|f (x)|dx.

It is considerably more difficult to show that F extends from L1 (Rn )∩L2 (Rn ) to a bounded
operator on L2 (Rn ). In fact F is an isometry in the sense that

kFf kL2 = kf kL2

(the Plancherel theorem, see Appendix B).


Moreover, by the Hausdorff-Young inequality (see Theorem 20 of Appendix A.4), F
extends to a bounded operator from Lp to Lq , for 1 ≤ p ≤ 2 and p−1 + q −1 = 1. More
precisely,
kFf kq ≤ kf kp .
We say an operator A : Y → X is invertible if and only if there is an operator
A : X → Y such that A−1 A = 1lY and AA−1 = 1lX , where 1lX and 1lY are the identity
−1

operators in X and Y respectively.


Exercise A.9. Show that:
1) A is invertible if and only if for every f ∈ X, the equation Au = f has a unique
solution u(= A−1 f ) ∈ Y , i.e. if and only if A is one–to–one (Au = 0 ⇒ u = 0) and onto
(Ran A = X).
2) if A is just one-to-one (i.e. not necessarily onto), then A is invertible as an operator
from Y to Ran A ⊂ X, i.e. the equation Au = f has a unique solution u = A−1 f for any
f ∈ Ran A.
3) if A and B are invertible then so is AB and (AB)−1 = B −1 A−1 (generalize to an
arbitrary number of factors).
Example A.1. We show invertibility of several important operators we encountered
above.

1. 1 is clearly invertible.
Lectures on Applied PDEs, November 3, 2017 247

2. Mf is invertible for essinf |f | > 0 and Mf−1 = M1/f .

d
3. dx
(see discussion below).

4. F is invertible with F −1 given in (B.3) (see Theorem B.1(f )).

5. −∆+1 is invertible with (−∆+1)−1 f = G∗f where G := F −1 ( |k|21+1 ) and ∗ denotes


convolution.
(x−y)2
6. curl is invertible with curl−1 w =
R
|x−y|2
w(y)dy in 2d (Biot-Savart formula).

Consider the operator A = dx d


: H 1 (R) → L2 (R). Then A is one-to-one:
Rx Au = 0 and
u ∈ H1 (R) imply u ≡ 0. Thus A has the right inverse B: f (x) → x0 f (y)dy for some
fixed x0 ∈ R: AB = 1l. However, B does not map L2 (R) into H 1 (R). In fact it is not
defined on the entire L2 (R) but only on L2 (R) ∩ L1 (R) and mapsRthis space into L∞ . Put

differently, the operator A is not onto: Ran A = {f ∈ L2 (R)| −∞ f dx = 0} = 6 L2 (R).
Hence A in not invertible.
The above illustrates the importance of finding inverses of operators: existence of an
inverse for A : Y → X is equivalent to the equation Au = f having a unique solution for
every f ∈ X. The following simple statement gives a powerful criterion for existence of
inverses.

Theorem A.2. Assume an operator A : X → X is invertible and an operator B : X → X


is bounded with the norm satisfying the inequality

kBk < kA−1 k−1

Then the operator A + B (defined on the domain of A) is invertible. Moreover, its inverse
is given by the absolutely convergent series

X
−1
(A + B) = A−1 (−BA−1 )n , (A.20)
n=0

called the Neumann series (for A + B).

Exercise A.10. Prove this theorem. Hint: Show that the series (A.20) is absolutely
convergent and gives the inverse to A + B and use that if Tn is a Cauchy sequence of
operators from X to X, then there is an operator T : X → X s.t. Tn → T (one has to use
the fact that the space of bounded operators equipped with the operator norm is complete,
i.e. is a Banach space).
248 Lectures on Applied PDEs, November 3, 2017

A.8 Special classes of operators


The adjoint. Consider an operator A on a Hilbert space X. With it we associate its
adjoint A∗ defined by the relation hA∗ u, vi = hu, Avi, for all v ∈ D(A), and for all u’s
such that supv∈D(A),kvk=1 |hu, Avi| < ∞ (those u’s form the domain of the operator A∗ ,
D(A∗ )).
If A is a bounded operator then supkvk=1 |hu, Avi| ≤ ||u||||A|| < ∞ and it suffices to
check only the relation hA∗ u, vi = hu, Avi.
Exercise A.11. Show that if operators A and B are bounded, then
(a) A∗ is a bounded operator and kAk = kA∗ k (Hint: use that kAk = sup||u||,||v||=1 |hu, Avi|,
see Exercise A.7(2)).
(b) (A + B)∗ = A∗ + B ∗ ,
(c) (αA)∗ = αA∗ ,
(d) (AB)∗ = B ∗ A∗ ,
(e) (A−1 )∗ = (A∗ )−1 .
An important class of operators on a Hilbert space is the class of self–adjoint operators.
By definition, an operator A is called self–adjoint if and only if A∗ = A. By definition,
every self–adjoint operator is symmetric, i.e. hAu, vi = hu, Avi, for all u, v ∈ D(A). Notice
that the converse is not true. If an operator A is symmetric then all we know is that
D(A) ⊂ D(A∗ ) (show this!). However, a symmetric operator A obeying D(A) = D(A∗ ) is
also self-adjoint. Thus every symmetric bounded operator is self–adjoint.
Consider the examples of the operators 1)–5) above. We have the following: Mf is
symmetric if and only if f is a real function; ∂x∂ j is anti–symmetric, so the differentiation
operator is not symmetric, but −i ∂x∂ j is symmetric; the identity operator is obviously
symmetric; ∆ is symmetric and so is −∆ + V (x) for V (x) real; the integral operator
is symmetric if K(x, y) = K(y, x) (cf with matrices!). In addition, if we know that
K(x, y) ∈ L2 (Rn × Rn ), then the operator K is self-adjoint. A point is that while the
symmetry property is easy to verify, the self-adjointness property is hard. See [?] or [8]
for a proof of self-adjointness of −i ∂x∂ j , ∆ and −∆ + V (x).

We observe that for any operator A on a Hilbert space X, we can write

X = Null A ⊕ Ran A∗ (A.21)

Here, the null space is defined as Null A := {u ∈ X : Au = 0}.


Exercise 11. Show that for a bounded operator A, Null A is a closed set, and show (A.21).

Projections. A bounded operator P on X is called a projection operator (or simply a


projection) if and only if it satisfies
P 2 = P.
Lectures on Applied PDEs, November 3, 2017 249

This relation implies kP k ≤ kP k2 , i.e. kP k ≥ 1. We have


v ∈ Ran P ⇐⇒ P v = v and v ∈ (Ran P )⊥ ⇐⇒ P ∗ v = 0. (A.22)
Indeed, if v ∈ Ran P , then there is a u ∈ X s.t. v = P u, so P v = P 2 u = P u = v; the
second statement is left as an
Exercise 12. Prove that (a) P ∗ v = 0 if and only if v ⊥ Ran P , (b) Ran P is closed and
(c) P ∗ is also a projection.
Examples. The following are projection operators:
1) on a space of functions u(x), χx∈E : u(x) 7→ χE (x)u(x), where χE (x) = 1 if x ∈ E
and χE (x) = 0 if x ∈/ E,
2) on a space of functions u(x), χp∈E = F −1 χx∈E F acting as u(x) 7→ (χE (k)û(k))ˇ(x),
3) on any Hilbert space, where ϕ, ψ are two fixed elements s.t. hϕ, ψi = 1: u 7→ hϕ, ui ψ
4) as in 3), but where now {ψi } is an orthonormal set (i.e. hψi , ψj i = δi,j ): u 7→
PN
1 hψi , ui ψi .

A projection P is called an orthogonal projection if and only if it is self-adjoint, i.e. if


and only if P = P ∗ . Let P be an orthogonal projection, then (A.22) implies that
v ⊥ Ran P ⇐⇒ P v = 0, i.e., Null P = (Ran P )⊥ . (A.23)
The projections in Examples 1), 2) and 4) above are orthogonal. The projection in
Example 3) is orthogonal if and only if ϕ = ψ.
Exercise 13. Let P be an orthogonal projection. Show that
(a) ||P || ≤ 1, and therefore ||P || = 1 (Hint: Use (A.22)),
(b) 1l − P is also an orthogonal projection and Ran (1l − P )⊥ Ran P and Null 1l − P =
Ran P ,
(c) X = Ran P ⊕ Null P .
Remark. Orthogonal projections on X are in one-to-one correspondence with closed
subspaces of a Hilbert space X. This correspondence is obtained as follows. Let V =
Ran P . Then V is a closed subspace of X. To show that V is closed, let {vn } ⊂ V ,
and vn → v ∈ X, and show that v ∈ V . Since P is a projection, we have vn = P vn , so
||v − P v|| = ||v − vn − P (v − vn )|| ≤ ||v − vn || + ||P || ||v − vn || → 0, as n → ∞. Therefore
v = P v, so v ∈ V , and V is closed.
Conversely, given a closed subspace V , define a projection operator P by
P u = v, where u = v + v ⊥ ∈ V ⊕ V ⊥ . (A.24)
Exercise 14. Show that P defined in (A.24) is an orthogonal projection with Ran P = V .
For any given V , show that there is only one orthogonal projection (the one given in
(A.24)) such that Ran P = V .
(*Galerkin approximation (A → P AP )*)
250 Lectures on Applied PDEs, November 3, 2017

The space of bounded linear operators L(X, Y ) We assume that X and Y are
normed vector spaces over C, and consider the set of all bounded linear operators from
X into Y , i.e. each such operator is defined on the entire space X, and its range lies in
Y . This set of operators is denoted by L(X, Y ).
For A, B ∈ L(X, Y ), we define a new operator, called A + B, by setting (A + B)u :=
Au + Bu, for all u ∈ X. Also, for λ ∈ C and A ∈ L(X, Y ), we define a new operator
λA as (λA)u = λAu, for all u ∈ X. If in addition to these two operations on operators,
we equip the set L(X, Y ) with the norm introduced in (A.17), then L(X, Y ) is a normed
vector space.

Exercise 15. Show that L(X, Y ) is a vector space.

An important question is: when is L(X, Y ) a Banach space? The answer is given in
the following theorem, which is not difficult to prove (see e.g. [4], Proposition 5.3):

Theorem 22. If Y is a Banach space, then L(X, Y ) is a Banach space.

The dual space. In the special case when Y = C, the space L(X, Y ) is called the dual
space of X (or simply the dual, or adjoint space or conjugate space of X), and it is denoted
as X 0 . Hence the elements of X 0 := L(X, C) are linear maps from X to C, and they are
called linear functionals. Remark also that since C is complete, then the last theorem
shows that X 0 is always a Banach space, whether X is complete or not.
The operator norm induces a norm on X 0 : if l ∈ X 0 , then

||l|| = sup |l(x)|.


||x||=1

If X is a space of functions, then X 0 can be identified with either a space of functions or


a space of distributions or a space of measures. Here are some examples of dual spaces:

1) (Lp )0 = Lq , where 1/p + 1/q = 1, if 1 ≤ p < ∞ (space of functions),

2) (L∞ )0 is a space of measures which is much larger than L1 ,

3) (Hs )0 = H−s (space of distributions if s > 0).

Note that (Lp )0 ⊃ LRq , for 1 ≤ p < ∞ follows from the Hölder inequality. In fact, given
f ∈ Lq , define lf (u) := f u. Since |lf (u)| ≤ ||f ||q ||u||p , we see that lf is a bounded linear
functional on Lp . It can be shown that in fact any bounded linear functional on Lp can
be represented by lf for some f ∈ Lq .
Lectures on Applied PDEs, November 3, 2017 251

B Fourier transform
In this section, we describe one of the most powerful tools in analysis – the Fourier
transform. This transform allows us to analyze a fine structure of functions and to solve
differential equations. The Fourier transform takes functions of time to functions of
frequencies, functions of coordinates to functions of momenta, and vice versa.
Initially, we define the Fourier transform on the Schwartz space S(Rn ) = S:

S = {f ∈ C ∞ (Rn ) : hxiN |∂ α f (x)| is bounded ∀N and ∀α}, (B.1)

where hxi = (1 + |x|2 )1/2 and α = (α1 , ..., αn ), with αj non-negative integers, ∂ α :=
Qn αj
|α|
Pn ˆ
j=1 ∂x j and = i=1 αj . On S, we define the Fourier transform F : f 7→ f by
Z
ˆ
f (k) := (2π)−n/2
f (x)e−ik·x dx. (B.2)

Define also the inverse Fourier transform of f (k) as


Z
fˇ(x) := (2π)−n/2
f (k)eix·k dk. (B.3)

Some key properties of the Fourier transform are collected in the following
Theorem B.1. Assume f, g ∈ S(Rn ). Then we have:
(a) (−i∂)α f 7→ k α fˆ, and xα f 7→ (−i∂)α fˆ,

(b) f g 7→ (2π)−n/2 fˆ ∗ ĝ, and f ∗ g 7→ (2π)n/2 fˆĝ,

(c) (fˆ)ˇ= f = (fˇ)ˆ,

(d) fˆ = fˇ ,

(e) fˆg = f ĝ,


R R

fˆĝ =
R R
(f ) f g.
Properties (a) - (f ) hold (possibly, with signs changed in (a)) also when ˆ is replaced by ˇ.
We give a formal proof. Integrating by parts, we compute
Z
−n/2
−i(∂xj f )ˆ(k) = (2π) (−i)∂xj f (x)e−ik·x dx
Z
−n/2
= (2π) f (x)i∂xj e−ik·x dx

= kj fˆ(k).
252 Lectures on Applied PDEs, November 3, 2017

Exercise 16. Prove the remaining relations in (a)


Now we prove the second relation in (b). Using e−ik·x = e−ik·(x−y) e−ik·y and changing
the variable of integration as x0 = x − y, we obtain
Z Z
f[∗ g (k) := (2π) −n/2
e−ik·x
( f (x − y)g(y)dy)dx
Z Z
−n/2
= (2π) ( e−ik·(x−y) f (x − y)dx) e−ik·y g(y)dy

= (2π)n/2 fˆ(k 0 )fˆ(k)ĝ(k).

Exercise 17. Prove the first relation in (b) from the second one and (c).
The proof of (c) is more subtle. We use an approximation of unity ϕt (x) = t−n ϕ(x/t)
and compute ϕt ∗ (fˆ)ˇ. Let us define ϕx (y) := ϕ(x − y). Using property (b), we find
Z Z Z
ˆ ˆ x ˆ
ϕt ∗ (f )ˇ= ϕt · (f )ˇdy = (ϕt )ˇf dy = ((ϕxt )ˇ)ˆf dy.
x

Exercise 18. Show (formally, without justification of the interchange of the order of
integration etc.) that

x−y
 
x x −n
((ϕt )ˇ)ˆ= ((ϕ̂t )ˇ) = t (ϕ̂)ˇ .
t
Thus we have
ϕt ∗ (fˆ)ˇ= ((ϕ̂)ˇ)t ∗ f (B.4)
We can choose ϕ such that (ϕ̂)ˇ ∈ L1 , and (ϕ̂)ˇ(x)dx = 1. Indeed, take e.g. ϕ(x) =
R
2 2 2
(4π)−n/2 e−|x| and use the fact that ((e−|x| )ˆ)ˇ = e−|x| . With this in mind, we take the
limit t → 0 in (B.4) and use the properties of the approximation of identity to get

ϕt ∗ (fˆ)ˇ→ (fˆ)ˇ and ((ϕ̂)ˇ)t ∗ f → f as t → 0

to obtain (fˆ)ˇ= f . Similarly one shows that (fˇ)ˆ= f .


Exercise 19. Prove the relations in (d) – (f ).
By definition of the Dirac δ–function, we obtain

F : δ(x − x0 ) → (2π)−n/2 e−ik·x0 .

Hence the property (c) implies that F −1 : (2π)−n/2 e−ik·x0 → δ(x − x0 ), and, by taking the
complex conjugate (remember that F(f ) = F ∗ (f ) = F −1 (f )), we arrive at

F : (2π)−n/2 e−ik0 ·x 7→ δ(k − k0 ). (B.5)


Lectures on Applied PDEs, November 3, 2017 253

Exercise 20. Using (B.5), prove formally that (fˆ)ˇ = f = (fˇ)ˆ, and that (f g)ˆ =
(2π)−n/2 fˆ ∗ ĝ.

Statement (f) is called the Plancherel Theorem. The adjoint F ∗ of the Fourier trans-
form is defined by hF ∗ u, vi = hu, Fvi for all u, v ∈ S(Rn ), where h·, ·i is the standard
inner product in L2 (Rn ). Then (d) and (e) show that F ∗ = F −1 . This together with (e)
implies that FF ∗ = id = F ∗ F on S, which is a restatement of the Plancherel theorem.

Corollary B.2. F extends to a unitary operator on L2 , i.e. to a bounded operator


satisfying F ∗ = F −1 .

The next theorem gives the important example of the Fourier transform - the Fourier
transform of a Gaussian :

Theorem B.3. Let A be a n × n matrix s.t. ReA := (A + A∗ )/2 is positive definite (i.e.
x · ReA x > 0 if x 6= 0). Then we have
−1 k/2
F : e−x·Ax/2 7→ (det A)−1/2 e−k·A (B.6)

Proof. We prove the theorem only for positive definite matrices. If A is positive definite
(i.e. if x · Ax > 0 for x 6= 0), then there is an orthogonal matrix U (i.e. U is real and
U U T = U T U = id) s.t. A := U T AU is diagonal, say A = diag(λ1 , . . . , λn ). Letting x = U y
and noticing that x · Ax = y · U T AU y, and that det U = 1, we get
Z Z n Z
−y·Ay/2 ik0 ·y 2 0
Y
e−x·Ax/2 −ik·x
e dx = e e dy = e−λj yj /2 eikj yj dyj ,
1

where k 0 = U T k, and we have used k · U y = U T k · x. It is left as an exercise to show that


for n = 1,
2 2
F : e−λx /2 7→ λ−1/2 e−k /2λ . (B.7)
The last two relations imply the desired statement.

Exercise 21. Show (B.7).

The function e−x·Ax is called a Gaussian. It is one of the most common functions
in applications. There is another important function whose Fourier transform can be
explicitly computed:
Cn,α |k|−n+α if α 6= n,

−α
F : |x| 7→ (B.8)
Cn,n ln |k| if α = n.
The coefficients are given for α = 2 by

((2 − n)σn−1 )−1



for n 6= 2,
Cn,2 = (B.9)
−(σn−1 )−1 = −(2π)−1 for n = 2,
254 Lectures on Applied PDEs, November 3, 2017

where σn is the volume of the n–dimensional unit sphere S n = {x ∈ Rn+1 : |x| = 1}.
One can easily deduce formula (B.8) modulo the constants (B.9). Indeed, since |x|−α is
rotationally invariant, then so is its Fourier transform (see Exercise B.2 below). Also,
since |x|−α is homogeneous of degree −α, then its Fourier transform is homogeneous of
degree −n+α (see Exercise B.2 below). Hence (B.8) follows. Though it is easy to compute
the Fourier transform of |x|−α , it is not easy to justify it. Indeed, the function |x|−α is
rather singular and definitely does not belong to S(Rn ).
Exercise B.1. For n = 1, compute the Fourier transform of the characteristic function
χ(−a,a) (x), using definition (B.2).
As an example we show the following relation

e−µ|x|
((|k|2 + µ2 )−1 )ˇ= , for n = 3, (B.10)
4π|x|
which appears often in applications. Indeed, let f (k) = (|k|2 + µ2 )−1 and n = 3. We have

eik·x
Z
fˇ(x) = (2π)−3/2 2 2
dk
R3 |k| + µ
Z R Z 1 ir|x|v 2
−1/2 e r
= lim (2π) 2 2
drdv
R→∞ 0 −1 r + µ
(2π)−1/2 R eir|x| r
Z
= lim 2 2
dr.
R→∞ i|x| −R r + µ

k·x
In the second equality, we change to spherical coordinates v = cos φ = |k||x| with r = |k|.
Now, the last integral can be computed by changing to a contour
√ integral over
√ a rectangle
in the upper half complex plane with top vertices at −R + i R and R + i R and taking
the limit as R → ∞. By the Residue theory we have (B.10) (show this).
Exercise B.2. Let fh (x) := f (x − h) for h ∈ Rn , f (λ) (x) := λn/2 f (λx) for λ ∈ R+ and
f (g) (x) := f (gx) for ga ∈ SO(n) be the translation, dilation and rotation of f (x). Show
that
F : fh (x) 7→ e−ik·h fˆ(k), (B.11)
F : f (λ) (x) 7→ fˆ(1/λ) (k). (B.12)
−1
F : f (g) (x) 7→ fˆ(g ) (k). (B.13)
Define the unitary operators Th : f (x) 7→ f (x − h) (the operator of translation by h)
and Sλ : f (x) 7→ λn/2 f (λx) (the operator of dilation by λ). Then (B.11) and (B.12) imply

F ◦ Th = Me−ik·h ◦ F,
Lectures on Applied PDEs, November 3, 2017 255

where we recall that Me−ik·h is the operator of multiplication by e−ik·h , and


F ◦ Sλ = S1/λ ◦ F.
Note the following important property of the Fourier transform, called the uncertainty
principle: let f ∈ S(Rn ) be a Schwartz function on Rn , s.t. kf kL2 (Rn ) = 1. Then
Z 1/2 Z 1/2
2 2 n 2 ˆ 2 n
|x| |f (x)| d x |k| |f (k)| d k ≥ 1/2. (B.14)

Exercise B.3. Prove inequality (B.14) (Hint: in the case n = 1, notice that
hf, {(−i∂x )x − x(−i∂x )}f i = −ikf k2 ,
where the scalar product and the norm are in L2 (R). On the other hand,
hf, {(−i∂x )x − x(−i∂x )}f i = 2iIm h(−i∂x )f, xf i .
Use these two observations to show that kf k2 ≤ 2k(−i∂x )f k kxf k, and finish the proof by
invoking Plancherel’s theorem).
Let w ∈ L2 (Rn ) be a given function. Then the integral
Z
−n/2
(2π) w(x − y)f (x)e−ik·x dn x (B.15)

is called the windowed Fourier transform of f (with the window function w). In signal
2
analysis, i.e. for n = 1, one often uses the Gaussian (2π)−1/2 e−x /(2α) for w. In this case,
the transform (B.15) is called the Gabor transform.
For more details see [10], Chapter 5.

C Linear evolution and semigroups


Let X be a Banach space and A be a linear (closed) operator on X with dense domain
Y = D(A). Our goal is to solve the initial value problem
∂u
= Au, u|t=0 = u0 ∈ Y
∂t
for u ∈ C(R, Y ) ∩ C 1 (R, X). We use the following terminology.
• The family, U (t), t ≥ 0, of operators on X is called a (strongly continuous) semigroup
if and only if
(a) U (t) are bounded ∀t ≥ 0.
(b) U (0) = 1 and U (t + s) = U (t)U (s).
(c) t → U (t)ϕ is continuous ∀ϕ ∈ X.
256 Lectures on Applied PDEs, November 3, 2017

• The family U (t) is called a contraction semigroup if and only if U (t) is a semigroup
and ||U (t)|| ≤ 1.

• The (closed) operator

1
Au := lim (U (s) − 1)u (C.1)
s→0 s

on X with the domain

D(A) := {u ∈ X | the limit on the r.h.s of (C.1) exists }


is called the generator of U (t). If A is the generator of the semigroup U (t), then we
write

U (t) = eAt .

Theorem C.1. If A is the generator of U (t), then U (t)D(A) ⊂ D(A) and ∀u0 ∈ D(A),
u := U (t)u0 solves the equation

∂u
= Au (C.2)
∂t
with the initial condition u|t=0 = u0 .

Proof. If u ∈ D(A), then

1 1
AU (t)u = lim (U (s) − 1)U (t)u = U (t) lim (U (s) − 1)u exists.
s→0 s s→0 s

Hence U (t)u ∈ D(A). Furthermore,

1 ∂
AU (t)u = lim (U (t + s)u − U (t)u) ≡ U (t)u.
s→0 s ∂t

Corollary C.2. If an operator A is the generator of a semigroup U (t), then the initial
value problem

∂u
= Au, u|t=0 = u0
∂t
has a solution for any u0 ∈ D(A) and this solution is given by the formula u = U (t)u0 .
Lectures on Applied PDEs, November 3, 2017 257

Thus the main question here is: when does an operator A generate a semigroup? First
of all bounded operators generate semigroups. Indeed, for a bounded operator B we define

tB
X (tB)n
U (t) ≡ e := . (C.3)
n=0
n!
The series on the r.h.s. converges absolutely since
∞ ∞
(tB)n X tn

X

n!
≤ ||B||n = et||B|| < ∞.
n=0
n!
n=0

Exercise C.1. Show that equation (D.4) defines a semigroup and that this semigroup is
generated by B.

For unbounded operators the answer to the question above is given by the following.

Theorem C.3 (Hille-Yosida). Let A be a closed operator such that (a) (0, ∞) ⊂ ρ(A)
and (b) ||(A − λ)−1 || ≤ 1/λ for any λ > 0. Then A generates a unique semigroup and
this semigroup is contractive.

Proof. The idea is very simple: we approximate the operator A by bounded operators Aλ
so that Aλ u → Au ∀u ∈ D(A) as λ → ∞; construct the semigroup, Uλ (t), for Aλ by the
formula

X 1
Uλ (t) = (tAλ )n (C.4)
n=0
n!
(see equation (D.4)); define the semigroup, U (t), for A as the limit

U (t)u = lim Uλ (t)u (C.5)


λ→∞

for any u ∈ D(A) and then, by continuity, extend U (t) to the entire space X.
We define Aλ as

Aλ := Aλ(λ − A)−1 .
(Note A−λ for λ > 0 is invertible by the condition that (0, ∞) ⊂ ρ(A).) Then ∀u ∈ D(A),
by (b)

λ(λ − A)−1 u − u = (λ − A)−1 Au → 0,


as λ → ∞, and

kλ(λ − A)−1 k ≤ 1.
258 Lectures on Applied PDEs, November 3, 2017

Hence, by an /3 argument,

λ(λ − A)−1 u → u ∀ u ∈ X.

This implies ∀u ∈ D(A),

Aλ u = λ(λ − A)−1 Au → Au.


Now we consider the semigroup Uλ (t) defined in (C.4). Due to (b) and the relation
Aλ = λ2 (λ − A)−1 − λ, we have


−λt
X 1
ke Aλ t
k ≤ e k(tλ2 (λ − A)−1 )n k
n=0
n!

−λt
X tn
≤ e λ2n k(λ − A)−1 kn
n=0
n!
∞ n
X t n
≤ e−λt λ . (C.6)
n=0
n!

Hence

||eAλ t || ≤ 1, (C.7)
i.e., eAλ t is the contractive semigroup.
Now we show that {eAλ t , λ > 0} is a Cauchy family in the sense that

||(eAλ0 t − eAλ t )u|| → 0 (C.8)


as λ, λ0 → ∞ for any u ∈ X. To prove this we represent the operator acting on u inside
the norm as an integral of a derivative
Z t
Aλ0 t Aλ t ∂ Aλ0 s Aλ (t−s)
e −e = e e ds.
0 ∂s
Using the equation
∂ Aλ s
e = Aλ eAλ s = eAλ s Aλ
∂s
gives
Z t
eAλ0 t
−e Aλ t
= eAλ0 s eAλ (t−s) (Aλ0 − Aλ )ds.
0

The last equation together with (C.7) yields


Lectures on Applied PDEs, November 3, 2017 259

Z t
||(e Aλ0 t
−e Aλ t
)u|| ≤ ||eAλ0 s eAλ (t−s) (Aλ0 − Aλ )u||ds
Z0 t
≤ ||(Aλ0 − Aλ )u||ds = t||(Aλ0 − Aλ )u||
0

and therefore (C.8) follows for any t ≥ 0 first ∀u ∈ D(A) and then by continuity ∀u ∈ X.
Now equation (C.8) implies that the limit on the r.h.s. of (C.5) exists ∀t ∈ [0, ∞) ∀u ∈
X and satisfies

||U (t)|| ≤ 1.
Equations (C.5) and (C.7) imply also that U (t) is the semigroup: U (t + s) = U (t)U (s)
and U (0) = 1. It remains to prove that U (t) is Rstrongly continuous and is generated by
s
the operator A. Using the relation Uλ (s) − 1 = 0 dtUλ (t)Aλ , we find for u ∈ D(A)

lim ||(U (s) − 1)u|| = lim lim ||(Uλ (s) − 1)u||


s→0 s→0 λ→∞
Z s
= lim lim || dtUλ (t)Aλ u||
s→0 λ→∞ 0
Z s
≤ lim lim dt||Aλ u||
s→0 λ→∞ 0
= lim s||Au|| = 0.
s→0

Hence U (s)u → u as s → 0 ∀u ∈ D(A). Since U (t) is bounded uniformly in t we


conclude
Rs that it is strongly continuous. Similarly, using the relation Uλ (s) − 1 − Aλ =
0
dt(Uλ (t) − 1)Aλ and the strong continuity of Uλ (t) shown above, we conclude that U (t)
is generated by A.

How do we check the conditions of the Hille-Yosida theorem? Usually this is a hard
business. However, there are several cases where this can be easily done. Below, X is a
Hilbert Space.
A) A = −A∗ (A is anti-self-adjoint). Then σ(A) ⊂ iR and k(A − λ)−1 k ≤ λ−1 for λ > 0.
B) A = A∗ ≤ 0 (A is non-positive). Then σ(A) ⊂ (−∞, 0] and k(A − λ)−1 k ≤ λ−1 for
λ > 0.
C) Perturbations of generators. For example, A = A0 +B where A0 generates a semigroup
and kB(A0 − λ)−1 k ≤ β with β < 1 ∀λ ≥ λ0 , for some λ0 > 0. Indeed, in this case we
have for λ > λ0

A − λ = [1 + Tλ ](A0 − λ) (C.9)
260 Lectures on Applied PDEs, November 3, 2017

where Tλ = B(A0 − λ)−1 . By the condition above ||Tλ || ≤ β ∀λ ≥ λ0 . Since β < 1, 1 + Tλ


is invertible and therefore so is the r.h.s. of (C.9) ∀λ ≥ λ0 . Therefore (λ0 , ∞) ⊂ ρ(A).
Moreover, (C.9) implies that

(A − λ)−1 = (A0 − λ)−1 (1 + Tλ )−1


and therefore

||(A − λ)−1 || ≤ ||(A0 − λ)−1 || ||(1 + Tλ )−1 ||.


Since A0 generates a semigroup, we have that ||(A0 − λ)−1 || ≤ 1/λ. Since ||Tλ || ≤ β < 1
we have that ||(1 + Tλ )−1 || ≤ (1 − β)−1 . Collecting the last three estimates we conclude
that ∀λ ≥ λ0

||(A − λ)−1 || ≤ (1 − β)−1 λ−1 .


This is not quite what we need (remember (b)). However, for the operator Aµ = A − µ
with µ = min(λ0 , (1 − β)−1 ), we obtain

||(Aµ − λ)−1 || = ||(A − (µ + λ))−1 || ≤ µ(µ + λ)−1 < λ−1 .


Thus the operator Aµ generates a contraction semigroup eAµ t , ||eAµ t || ≤ 1. Now, eAt :=
eAµ t eµt gives a semigroup for the operator A and this semigroup satisfies the estimate

||eAt || ≤ eµt .
Exercise C.2. Check the last statement.
Examples.

1) The Schödinger equation. In this case A = −iH where H is a self-adjoint operator


(e.g., a Schrödinger operator H := −∆ + V (x) on L2 (Rd )).

2) The heat equation. In this case A = A∗ ≤ 0 or more generally A = −A0 − A1 with


A0 = A∗0 > 0 and
−1/2 −1/2
kA0 A1 A0 k < 1.
P P
For example A = − ∂xi aij (x)∂xj + bi (x)∂xi + c(x) with the matrix (aij (x)) ≥ δ1,
( |bi (x)|2 )1/2 ≤ γ and c(x) ≥ γ 2 /δ.
P
3) The wave equation. In this case
 
0 1
A= with H = H ∗ ≥ 0. (C.10)
−H 0
Indeed, if v satisfies the wave equation
Lectures on Applied PDEs, November 3, 2017 261

∂ 2v
= −Hv with H ≥ 0,
∂t2
then the element u = (v, ∂v/∂t) satisfies the equation

∂u
= Au
∂t
with the operator A given in (C.10). On the Hilbert space X = D(H) ⊕ L2 with the inner
product

hu, wiX = hu1 , Hw1 i + hu2 , w2 i


where u = (u1 , u2 ) and w = (w1 , w2 ), the operator (C.10) is anti-self-adjoint, A = −A∗ ,
and therefore it generates a unique contraction semigroup.

Example. The acoustical wave equation.

∂ 2v 1
2
= c2 ρ∇ · ∇v.
∂t ρ
The operator H := −c2 ρ∇ · ρ1 ∇ is self-adjoint and, in fact, non-negative on the space
L2 (R3 , (c2 ρ)−1 dx).

Exercise C.3. Show that H is symmetric, i.e.,

hHu, vi = hu, Hvi ∀u, v ∈ D(H).

4) Maxwell equations. In a vacuum, Maxwell’s equations for the electric and magnetic
fields, E(x, t) and H(x, t), read

∂H
curl E = −µ , curl H =  ∂E
∂t
∂t
div(E) = 0, div(µH) = 0

where  and µ are dielectric constant and magnetic permeability, respectively. These
equations can be written as

∂t u = JAu,
 
0 1
where u = (E, H), J = and
−1 0
 −1 
µ curl 0
A=
0 −1 curl
262 Lectures on Applied PDEs, November 3, 2017

on the Hilbert space (we use that  and µ are independent of x)

H1 = {(E, H) | E, H ∈ H1 (R3 , R3 ), divE = 0, divH = 0}.


One can show that the operator A with the domain H1 ⊂ L2 (R3 , R3 ) ⊕ L2 (R3 , R3 ) is
self-adjoint (show that it is symmetric) and σess (H) = [0, ∞). Hence the operator JA
generates a contraction semigroup according to criterion D) above.

D Local Existence for Evolution Equations


We address the problem of the short time existence of solutions for evolution PDEs of the
general form
∂t u = F (u), u|t=0 = u0 , (D.1)
where F is a map defined from an open set, U , in a Banach space Y to some Banach
space Z. Without loss of generality we can assume that F (0) = 0.

Canonical form of evolution equations. We say the initial problem (D.1) is written
in the canonical form if the map G(u) is presented as G(u) = Au + f (u), where A is a
linear operator and f (u) satisfies f (u) = o(kuk). Then (D.1) can be rewritten as

∂t u = Au + f (u), u|t=0 = u0 . (D.2)

where A is a linear operator on Y and f is a nonlinear map, i.e. f (u) = o(kuk) in some
norm, or f (0) = f 0 (0) = 0. We will call f the nonlinearity.
We note that (4.9) is the initial value problem (D.1), ∂t u = F (u), u|t=0 = u0 , in the
canonical form, (4.1), with A = ∆, a linear operator acting on H 2 and f (u) = λ|u|p−1 u,
a nonlinear map.
In fact, we can write (D.1) in the form (D.2) for a large class of maps G. To begin with,
without loss of generality, we can assume G(0) = 0. Now, if G(u) is once continuously
(Gâteaux) differentiable, it can be always split into linear and nonlinear parts,

G(u) = Au + f (u),

with A and f as above. (This is a general result, valid for all differentiable maps. It
generalizes the Taylor theorem of Calculus.)

D.1 Linear problem


We begin with the simplest situation, namely, G(u) linear, G(u) = Au, or f (u) = 0. Then
we arrive at the linear initial value problem

∂t u = Au, u|t=0 = u0 , (D.3)


Lectures on Applied PDEs, November 3, 2017 263

Assume this equation has a unique solution for every u0 ∈ Y and denote this solution
by u0 (t) = etA u0 . The family of operators etA is called the evolution semigroup or a
propagator. Properties of evolution semigroups are discussed in Appendix C. Here we just
mention that the evolution semigroup etA exists if

• A is a bounded operator;

• A is either self-adjoint (A∗ = A) and bounded above, A ≤ C for some C < ∞


(hu, Aui ≤ Ckuk2 ) or is anti-self-adjoint (A∗ = −A);

• A is a ‘constant coefficient pseudo-differential’ operator; more precisely, A = a(−i∇x ),


where a is some decent function.

For a bounded operator A the flow always exists since eAt can be defined by

tA
X (tA)n
e := . (D.4)
n=0
n!

The series on the r.h.s. converges absolutely since


∞ ∞ ∞
X (tA)n X k(tA)n k X tn
≤ ≤ kAkn = etkAk < ∞.
n=0
n! n=0
n! n=0
n!

Exercise D.1. Show that, if A is bounded, then the family, U (t) = etA , t ≥ 0, satisfies

(a) U (t) are bounded ∀t ≥ 0.

(b) U (0) = 1 and U (t + s) = U (t)U (s).

(c) t → U (t)ϕ is continuous ∀ϕ ∈ X.

For the second case we refer to [6]. In the third case A and eAt are defined using
the Fourier transform as (Au)(k)
c At u)(k) = ea(k)t û(k), respectively (see
= a(k)û(k) and (ed
Appendix B and [6] for more on the Fourier transform). We have

Proposition D.1 (exponential of A = a(−i∇x )). Let a(k) be such that the inverse
Fourier transform, gt (x), of the function ea(k)t . Then, for A = a(−i∇x ), we have

etA u = gt ∗ u. (D.5)

Proof. Applying the inverse Fourier transform F −1 : f 7→ fˇ (so that (fˆ)ˇ = f = (fˇ)ˆ)
At u)(k) = ea(k)t û(k) and using that F −1 : gf 7→ (2π)−n/2 ǧ ∗ fˇ, we obtain (D.5).
to(ed

As an illustrative example, we consider


264 Lectures on Applied PDEs, November 3, 2017

Exercise D.2. (a) Show that

et∆ u = pt ∗ u, eit∆ u = pit ∗ u (D.6)


2
where pt (x) = (4πt)−n/2 e−|x| /4t .
(b) Show that et∆ u → u in L∞ , as t → 0.
The evolution operator et∆ , or sometimes its integral kernel, is called heat kernel and
it plays an important role in the theory of stochastic processes.

D.2 Reduction to a fixed point problem


Duhamel principle. Consider the inhomogeneous initial value problem

∂t u = Au + f, u|t=0 = u0 , (D.7)

where f = f (t) is a given function f : [0, T ) → Y .


Proposition D.2 (Duhamel principle). Assume the evolution semi-groups etA is well-
defined (see above). Then the solution, u(t), of (D.7), is given by
Z t
tA
u(t) = e u0 + e(t−s)A f (s) ds. (D.8)
0

In opposite direction, the family u(t) given by (D.8), which is differentiable in t and is in
the domain of the operator A, satisfies (D.7).
Exercise D.3. Prove this proposition under the additional assumption that v(t) := e−tA u(t)
is well-defined. (Hint: For the first part, consider the vector-function e−tA and show that
satisfies the equation
∂t v = f, v|t=0 = u0 .
Then use the Fundamental Theorem of calculus.)

Mild (weak) solutions. Consider the initial value problem in the canonical form (D.2).
We apply (D.8) to (D.2) to obtain
Z t
tA
u(t) = e u0 + e(t−s)A f (u(s)) ds. (D.9)
0

If u(t) solves (4.1), then it also solves the equation (D.9). Conversely, if u(t) solves (D.9)
and is differentiable in t and is in the domain of the operator A, then it solves the equation
(D.2).
If u(t) solves (D.9), but we do not know whether it is differentiable or not, we call u(t)
a mild (or weak) solution to (D.2).
Lectures on Applied PDEs, November 3, 2017 265

Fixed point problem. Eq (D.9) can be written as the equation u = H(u), where
Z t
H(u)(t) := e u0 + tA
e(t−s)A f (u)(s) ds, (D.10)
0

called the fixed point equation or the fixed point problem. A solution of such an equation
is called a fixed point. In our next step we learn how to solve fixed point equations.

D.3 Abstract result on local existence.


Before proceeding to other equations we prove an abstract result on local existence of
evolution equations in canonical form. We prove existence and uniqueness of weak
solutions for the initial value problem (D.1), ∂t u = F (u), u|t=0 = u0 , in the canonical
form, (D.2), which we reproduce here

∂t u = Au + f (u), (D.11)

with the initial condition u|t=0 = u0 . Recall, that here A is a linear operator on Y and f
is a nonlinear map, i.e. f (u) = o(kuk). For simplicity we assumed that f does not depend
on t explicitly (and only through u).
Recall, that we think of u as a path u : t ∈ I → u(t) ∈ Y in Y and we are looking for
weak solutions to (D.2). Specifically, we will consider solutions in the space C([0, T ], Y ),
for some T > 0. We have the following result:

Theorem D.3. Consider the abstract nonlinear evolution equation (D.2) and assume
that

• A generates a semi-group eAt , satisfying, for some constant K,


sup etA w Y ≤ KkwkY , (D.12)
t≥0

• the nonlinearity f is a locally Lipschitz map, f : Y → Y , obeying, for some


MR , LR < ∞,
sup kf (w)kY ≤ MR , (D.13)
kwkY ≤R

and
sup kf (w1 ) − f (w2 )kY / kw1 − w2 kY ≤ LR . (D.14)
kw1 kY ,kw2 kY ≤R

Let u0 ∈ Y and R > Kku0 kY . Then equation (D.2) has a unique weak solution u ∈
C([0, T ], Y ), for
T < (KLR )−1 , (R − Kku0 kY )/KMR ,
266 Lectures on Applied PDEs, November 3, 2017

in the ball kukC([0,T ],Y ) ≤ R. The solution u depends continuously on the initial condition
u0 . Furthermore, either the solution is global in time or blows up in Y in a finite time
(i.e. either ku(t)kY < ∞, ∀t, or ku(t)kY < ∞ for t < t∗ and ku(t)kY → ∞ as t → t∗ for
some t∗ < ∞).
Proof. Using Duhamel’s principle, Eq (D.2) can be written as the fixed point equation
u = H(u), where Z t
tA
H(u)(t) := e u0 + e(t−s)A f (u)(s) ds (D.15)
0
and we have written f (u)(s) for f (u(s)). Let X := C([0, T ], Y ), with T < (KLR )−1 , (R −
Kku0 kY )/KMR and R > Kku0 kY . The proof of existence and uniqueness will follow if
we can show that the map H has a unique fixed point in the ball
BR := {u ∈ X, kukX ≤ R}.
We prove this statement via the contraction mapping principle.
We begin by proving that H is a well-defined map from BR to BR . Using the assump-
tions (D.12) and (D.13), we obtain the estimates
tA
e u0 ≤ Kku0 kY (D.16)
X

and, if t < T and u ∈ BR ,


Z t Z t
(t−s)A

≤ sup K kf (u)(s)kY ds ≤ T KMR .
e f (u)(s) ds (D.17)
t≤T
0 X 0

Estimates (D.16) and (D.17) imply that H : BR → BR , provided Kku0 kY + KT MR ≤ R.


Now, we prove that H : BR → BR is a strict contraction. Recalling the definition of
kukX and using (D.12) and (D.14), we obtain, for u1 , u2 ∈ BR ,
Z t
kH(u1 ) − H(u2 )kX ≤ sup K kf (u1 )(s) − f (u2 )(s)kY ds
t≤T 0
≤ KT LR ku1 − u2 kX .
Therefore, if LR KT < 1, then H is a strict contraction in BR . We see that the in-
equalities Kku0 kY + KT MR ≤ R and LR KT < 1 are satisfied if R > Kku0 kY and
T < (KLR )−1 , (R − Kku0 kY )/KMR . This completes the proof of existence and unique-
ness of the solution u and the estimate on it.
Now, we prove that the solution to the initial value problem is continuous with respect
to changes in the initial condition u0 . Let u and v be the solutions with initial conditions
u0 and v0 . We estimate
ku − vkX ≤ ketA (u0 − v0 )kX
Z t
+k e(t−s)A (f (u)(s) − f (v)(s)) dskX .
0
Lectures on Applied PDEs, November 3, 2017 267

The estimate of these terms proceeds as above (take u1 = u and u2 = v) and if u, v ∈ BR ,


then
ku − vkX ≤ Kku0 − v0 kY + KT LR ku − vkX .
Thus, if T is as above, then ku − vkX ≤ K(1 − KT LR )−1 ku0 − v0 kY completing the proof
of continuity.
(expand) Finally, assume [0, t∗ ) is the maximal interval of existence of u and sup0≤t<t∗
ku(t)kY =: U < ∞. Take R = 2KU and let

τ := min((3KLR )−1 , U (3MR )−1 ).

Then taking u(t∗ − τ ) as a new initial condition, we see that the solution exists in the
interval [0, t∗ +τ ), a contradiction. This proves the dichotomy claimed in the theorem.
Discussion: Generalize (D.12) to

sup ρ(t) etA w Y ≤ KkwkY , (D.18)
t≥0

for some constant K and an appropriate positive function ρ(t).

E Elements of spectral theory


E.1 Definitions
Consider an operator A acting on a Banach space X with a domain D(A) (i.e., A :
D(A) → X). The spectrum, σ(A), of an operator A is the set in C defined by

σ(A) := {z ∈ C : A − z1l is not invertible }. (E.1)

For notational convenience, the operator “multiplication by z ∈ C” will be simply written


as z instead of z1l. Clearly, eigenvalues of A belong to σ(A) (in fact, if λ is an eigenvalue,
then Auλ = λuλ for some nonzero uλ ∈ X (uλ is called an eigenvector), so (A − λ)uλ = 0,
and A − λ is not invertible). In general, the spectrum can also contain continuous pieces
and it can take very peculiar forms.
Exercise E.1. Referring to the examples in Appendix A.7, show that
(a) the spectrum of the multiplication operator introduced in example 1) is σ(Mf ) =
Ran f ;
(b) the differentiation operator ∂x∂ j has spectrum σ( ∂x∂ j ) = iR;
(c) the identity operator (see
P 3)) has the spectrum consisting of one point σ(1l) = {1};
∂2
(d) the Laplacian ∆ := n1 ∂x 2 2
2 on L (Ω) with the domain D(∆) = H (Ω) has the
j
spectrum [0, ∞);
268 Lectures on Applied PDEs, November 3, 2017

(e) the Laplacian ∆ on [−a, a]n with (i) Dirichlet boundary conditions (i.e. u = 0
on the boundary) or (ii) the periodic boundary conditions has only eigenvalues of finite
multiplicities; find these eigenvalues.
The study of the spectra of operators is called spectral analysis.
Classification of the spectrum. Inverting A on (Null A)⊥ , perturbation theory (the
continuous dependence of eigenvalues and their multiplicities of the perturbation param-
eter).

∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗∗
The complement of the spectrum is called the resolvent set ρ(A):

ρ(A) := C\σ(A).

One can show (see [6]) that the set σ(A) is closed (and, consequently, the set ρ(A) is
open).
We have
Theorem E.1. The set σ(A) is closed (and, consequently, the set ρ(A) is open).
Proof. We prove the equivalent statement that the set ρ(A) is open. Let z0 ∈ ρ(A). Then
A − z0 is invertible. We write

A − z = A − z0 + z0 − z = (A − z0 )[1l + (z0 − z)(A − z0 )−1 ].


If |z0 − z| < k(A − z0 )−1 k−1 , then the operator in the square brackets on the right is
invertible by the Neumann theorem (see Theorem A.2). Therefore the operator A − z is
invertible for |z − z0 | < k(A − z0 )−1 k−1 as a product of two invertible operators.
For z ∈ ρ(A) the operator A − z has a bounded in X inverse. Denote this inverse as

RA (z) := (A − z)−1 .
It is called the resolvent of A at z ∈ ρ(A). It plays an important role in analysis of
operators. The proof of the theorem above shows that the resolvent is an analytic operator
valued function in z ∈ ρ(A) in the sense that for any z0 ∈ ρ(A) and for any z such that
|z − z0 | < kRA (z0 )k−1 , the resolvent RA (z) can be expanded in the series

X
RA (z) = RA (z0 ) ((z − z0 )RA (z0 ))n (E.2)
n=0

which is absolutely convergent is the sense that



X ∞
X
n
k(z − z0 )RA (z0 )k = |z − z0 |n kRA (z0 )kn < ∞.
n=0 n=0
Lectures on Applied PDEs, November 3, 2017 269

Indeed the series above is just the Neumann series for the inverse of the operator A − z =
(A − z0 )[1l + (z0 − z)(A − z0 )−1 ] (see Theorem A.2).
The resolvent satisfies two equations:

RA (z) − RA (w) = (z − w)RA (z)RA (w) (E.3)


and
RA (z) − RB (z) = RA (z)(B − A)RB (z), (E.4)
called the first and second resolvent equations. The first equation follows from the second
one with B = (z − w)1l and the second equation is equivalent to the identity A1 − B1 =
1
A
(B − A) B1 which can be easily verified.

Proposition E.2. If A is a bounded operator then σ(A) ⊂ {z ∈ C | |z| ≤ kAk}.

Proof. Expand formally by the Neumann series


∞  n
1 −1 1 −1 X A
= = .
A−z z 1 − A/z z n=0 z
This shows that z ∈ ρ(A) (i.e., (A − z)−1 is bounded) if and only if kA/zk < 1. Equiva-
lently, z ∈ ρ(A) if and only if |z| > kAk. Therefore, z ∈ σ(A) if and only if |z| ≤ kAk.

E.2 Location of the essential spectrum


(under construction)

E.3 Perron-Frobenius Theory


Consider a bounded operator T on the Hilbert space X = L2 (Ω).

Definition E.3. An operator T is called positivity preserving/improving if and only if


u ≥ 0, u 6= 0 =⇒ T u ≥ 0/T u > 0.

Note if T is positivity preserving, then T maps real functions into real functions.

Theorem E.4. Let T be a bounded positive and positivity improving operator and let λ
be an eigenvalue of T with an eigenvector ϕ. Then
a) λ = kT k ⇒ λ is simple and ϕ > 0 (modulo a constant factor).
b) ϕ > 0 and kT k is an eigenvalue of T ⇒ λ is simple and λ = ||T ||.

Proof. a) Let λ = ||T ||, T ψ = λψ and ψ be real. Then |ψ| ± ψ ≥ 0 and therefore
T (|ψ| ± ψ) > 0. The latter inequality implies that |T ψ| ≤ T |ψ| and therefore

h|ψ|, T |ψ|i ≥ h|ψ|, |T ψ|i ≥ hψ, T ψi = λ||ψ||2 .


270 Lectures on Applied PDEs, November 3, 2017

Since λ = ||T || = sup||ψ||=1 hψ, T ψi, we conclude using variational calculus (see e.g. [6] or
[?]) that

T |ψ| = λ|ψ| (E.5)


i.e., |ψ| is an eigenfunction of T with the eigenvalue λ. Indeed, since λ = kT k =
supkψk=1 hψ, T ψi, |ψ| is the maximizer for this problem. Hence |ψ| satisfies the Euler-
Lagrange equation T |ψ| = µ|ψ| for some µ. This implies that µ||ψ||2 = h|ψ|, T |ψ|i =
λ||ψ||2 and hence µ = λ. Equation (E.5) and the positivity improving property of T
imply that |ψ| > 0.
Now either ψ = ±|ψ| or |ψ| + ψ and |ψ| − ψ are nonzero. In the latter case they are
eigenfunctions of T corresponding to the eigenvalue λ : T (|ψ| ± ψ) = λ(|ψ| ± ψ). By the
positivity improving property of T this implies that |ψ| ± ψ > 0 which is impossible. Thus
ψ = ±|ψ|.
If ψ1 and ψ2 are two real eigenfunctions of T with the eigenvalue λ then so is aψ1 + bψ2
for any a, b ∈ R. By the above, either aψ1 + bψ2 > 0 or aψ1 + bψ2 < 0 ∀a, b ∈ R \ {0},
which is impossible. Thus T has a single real eigenfunction associated with λ.
Let now ψ be a complex eigenfunction of T with the eigenvalue λ and let ψ = ψ1 + iψ1
where ψ1 and ψ2 are real. Then the equation T ψ = λψ becomes

T ψ1 + iT ψ2 = λψ1 + iλψ2 .
Since T ψ2 and T ψ2 and λ are real (see above) we conclude that T ψi = λψ2 , i = 1, 2, and
therefore by the above ψ2 = cψ1 for some constant c. Hence ψ = (1 + ic)ψ1 is positive
and unique modulo a constant complex factor.

b) By a) and eigenfunction, ψ, corresponding to ν := ||T || can be chosen to be positive,


ψ > 0. But then

λhψ, ϕi = hψ, T ϕi = hT ψ, ϕi = νhψ, ϕi


and therefore λ = ν and ψ = cϕ.

*Question: Can the condition that ||T || is an eigenvalue of T (see b) be


removed?*

Now we consider the Schrödinger operator H = −∆ + V (x) with a real, bounded


potential V (x). The above result allows us to obtain the following important

Theorem E.5. Let H = −∆ + V (x) have an eigenvalue E0 with an eigenfunction ϕ0 (x)


and let inf σ(H) be an eigenvalue. Then

ϕ0 > 0 ⇒ E0 = inf{λ|λ ∈ σ(H)} and E0 is non-degenerate


Lectures on Applied PDEs, November 3, 2017 271

and, conversely,

E0 = inf{λ|λ ∈ σ(H)} ⇒ E0 is non-degenerate and ϕ0 > 0


(modulo multiplication by a constant factor).
Proof. To simplify the exposition we assume V (x) ≤ 0 and let W (x) = −V (x) ≥ 0. For
µ > sup W we have

X
−1 −1
(−∆ − W + µ) = (−∆ + µ) [W (−∆ + µ)−1 ]n (E.6)
n=0

where the series converges in norm as

kW (−∆ + µ)−1 k ≤ kW kk(−∆ + µ)−1 k ≤ kW kL∞ µ−1 < 1


by our assumption that µ > sup W = kW kL∞ . To be explicit we assume that d = 3.
Then the operator (−∆ + µ)−1 has the integral kernel

e− µ|x−y|
>0
4π|x − y|
while the operator W (−∆ + µ)−1 has the integral kernel

e− µ|x−y|
W (x) ≥ 0.
4π|x − y|
Consequently, the operator

1 e− µ|x−y|
Z
(−∆ + µ)−1 f (x) = f (y) dy
4π |x − y|
is positivity improving (f ≥ 0, f 6= 0 ⇒ (−∆ + µ)−1 f > 0) while the operator

1 e− µ|x−y|
Z
−1
W (−∆ + µ) f (x) = W (x) f (y) dy
4π |x − y|
is positivity preserving (f ≥ 0 ⇒ W (−∆ + µ)−1 f ≥ 0). The latter fact implies that the
operators [W (−∆+µ)−1 ]n , n ≥ 1, are positivity preserving (prove this!) and consequently
the operator

X
−1
(−∆ + µ) + [W (−∆ + µ)−1 ]n
n=1

is positivity improving (prove this!).


Thus we have shown that the operator (H + µ)−1 is positivity improving. Series (E.6)
shows also that (H + µ)−1 is bounded. Since
272 Lectures on Applied PDEs, November 3, 2017

hu, (H + µ)ui ≥ (− sup W + µ)kuk2 > 0,


we conclude that the operator (H + µ)−1 is positive (as an inverse of a positive operator).
Finally, k(H + µ)−1 k = sup σ((H + µ)−1 ) = (inf σ(H) + µ)−1 is an eigenvalue by the
condition of the theorem. Hence the previous theorem applies to it. Since Hϕ0 = E0 ϕ0 ⇔
(H + µ)−1 ϕ0 = (E0 + µ)−1 ϕ0 , the theorem under verification follows. *This paragraph
needs details!*.

References
[1] A. Ambrosetti and G. Prodi, A Primer of Nonlinear Analysis. Cambridge Univ.
Press, 1995.

[2] B. Dubrovin, A. Fomenko, S. Novikov, Modern Geometry, vols 1 and 2.

[3] L.C. Evans, Partial Differential Equations. AMS, 2002.

[4] G. B. Folland, Real Analysis. John Wiley & sons, 1984.

[5] I. M. Gelfand and S. V. Fomin, Calculus of Variations. (Dover Books on Mathe-


matics).

[6] S. J. Gustafson and I. M. Sigal. Mathematical Concepts of Quantum Mechanics.


2nd edition. Universitext. Springer-Verlag, Berlin, 2011.

[7] S. Gustafson, I.M. Sigal, T. Tzaneteas, Statics and Dynamics of Magnetic Vortices
and of Nielsen-Olesen (Nambu) Strings, Journal Mathematical Physics, 2010 (An
anniversary issue).

[8] P. Hislop and I.M. Sigal, Spectral Theory, Springer-Verlag, Berlin.

[9] J. Jost, Geometric Analysis, Springer

[10] E. Lieb and M. Loss, Analysis. 2nd edition, AMS, 2001.

[11] C. Mantegazza, Lecture Notes on Mean Curvature Flow. Birkhäuser, 2011.

[12] R. McOwen, Partial Differential Equations. Prentice Hall, 2003.

[13] J. Murray, Mathematical Biology: Part II: Spatial Models and Biomedical Appli-
cations, Springer-Verlag (2002).

[14] J. Ockedon, S. Howison, A. Lacey, A. Movchan, Applied Partial Differential Equa-


tions. Oxford Univ. Press, 1999.
Lectures on Applied PDEs, November 3, 2017 273

[15] M. Reed and B. Simon, Methods of Mathematical Physics, I. Academic Press.

[16] I.M. Sigal, Lectures on mean curvature flow.

[17] E.Stein and G. Weiss, Harmonic Analysis

[18] D. R. Tilley and J. Tilley, Superfluidity and Superconductivity. 3rd edition. Institute
of Physics Publishing, Bristol and Philadelphia (1990).

[19] M. Tinkham, Introduction to Superconductivity, McGraw-Hill Book Co., New York,


1996.

[20] M.I. Weinstein, Nonlinear Schrödinger equations and sharp interpolation estimates,
Comm. Math. Phys. 87, 567 – 576 (1983).

Das könnte Ihnen auch gefallen