UNIVERSIDADE FEDERAL DE SANTA CATARINA

CENTRO DE CIÊNCIAS FÍSICAS E MATEMÁTICAS

PROGRAMA DE PÓS-GRADUAÇÃO EM MATEMÁTICA PURA E APLICADA

Víctor Luis Espinoza

On the genericity of singularities in Lorentzian geometry

Florianópolis
2024
 Victor Luis Espinoza

On the genericity of singularities in Lorentzian geometry

Tese submetida ao Programa de Matemática Pura e

Aplicada da Universidade Federal de Santa Catarina
para a obtenção do título de Doutor em Geometria e
Topologia.
Orientador: Prof. Ivan Pontual Costa e Silva, Dr.

Florianópolis
2024
Ficha catalográfica gerada por meio de sistema automatizado gerenciado pela BU/UFSC.
Dados inseridos pelo próprio autor.

Espinoza, Victor Luis

On the genericity of singularities in Lorentzian
geometry / Victor Luis Espinoza ; orientador, Ivan Pontual
Costa e Silva, 2024.
81 p.

Tese (doutorado) - Universidade Federal de Santa

Catarina, Centro de Ciências Físicas e Matemáticas,
Programa de Pós-Graduação em Matemática Pura e Aplicada,
Florianópolis, 2024.

Inclui referências.

1. Matemática Pura e Aplicada. 2. geometria

Lorentziana. 3. incompletude geodésica. 4. genericidade.
I. Costa e Silva, Ivan Pontual. II. Universidade Federal
de Santa Catarina. Programa de Pós-Graduação em Matemática
Pura e Aplicada. III. Título.
 Victor Luis Espinoza
On the genericity of singularities in Lorentzian geometry

O presente trabalho em nível de doutorado foi avaliado e aprovado por banca examinadora
composta pelos seguintes membros:

Prof. Ivan Pontual Costa e Silva, Dr.

Universidade Federal de Santa Catarina

Prof. Jónatan Herrera Fernandez, Dr.

Universidad de Córdoba

Prof. José Luis Flores Dorado, Dr.

Universidad de Málaga

Prof. Francisco Carlos Caramello Junior, Dr.

Universidade Federal de Santa Catarina

Profa. Marianna Ravara Vago, Dra. [suplente]

Universidade Federal de Santa Catarina

Certificamos que esta é a versão original e final do trabalho de conclusão que foi julgado
adequado para obtenção do título de doutor em geometria e topologia.

Coordenação do Programa de Pós-Graduação

Prof. Ivan Pontual Costa e Silva, Dr.

Orientador

Florianópolis, 2024.
Agradecimentos
Esta tese é resultado de um trabalho que efetivamente durou quatro anos, passando por
um longo ano de pandemia onde diversos obstáculos tiveram que ser contornados.
Primeiramente, agradeço aos meus pais, Fernando Luis Espinoza e Maristela Cardoso,
pelo apoio durante todo meu percurso acadêmico.
Agradeço muito a meu orientador, Prof. Dr. Ivan Pontual Costa e Silva, que me guiou
durante todo este trajeto, com muitas ideias e insights, e que depois muitas reuniões, cálculos e
demonstrações, eventualmente culminaram na presente tese. Seu vasto conhecimento tanto em
matemática quanto em física são uma fonte de inspiração, e espero continuar colaborando com
ele no futuro.
Agradeço também os demais colegas dos departamentos de matemática e física por todas
as boas lembranças que carrego destes onze em que passei nestes departamentos. Em especial,
agradeço meu amigo e colega de graduação Paulo Henrique Dos Santos por todas as conversas
iluminadoras nos mais diversos e aleatórios assuntos, indo de política internacional, passando
por medicina, até animes, agradeço minha querida amiga Maria de Lourdes Deglmann por todo o
apoio, conversas e fofocas em nossas diversos encontros presenciais e vituais ao longo destes mais
de dez anos de amizade, e por fim também agradeço a meu amigo Gabriel Michels, por diversos
papos sobre matemática, geometria, mercado financeiro, e por recentemente me acompanhar
ganhando e perdendo (mais frequentemente perdendo) dinheiro no mercado de criptomoedas.
O presente trabalho foi realizado com apoio da Coordenação de Aperfeiçoamento de
Pessoal de Nível Superior – Brasil (CAPES) – Código de Financiamento 001.
Resumo
O objetivo desta tese é obter condições em espaços-tempo para que o fenômeno de
incompletude geodésica causal é genérico no sentido topológico (ou seja, é válido em um
conjunto residual).
Em nosso primeiro conjunto de novos resultados, baseamos nossas técnicas topológicas
naquelas desenvolvidas por Lerner (1973) e obtemos um teorema de genericidade da incompletude
geodésica em uma classe de métricas de espaço-tempo na topologia de Whitney forte (sob
condições abertas adequadas) para uma variedade não compacta fixada contendo subvariedades
fracamente presas de codimensão dois. Com algumas restrições extras para a curvatura em tais
classes de espaços-tempos, obtemos um resultado análogo no caso de codimensão maior que
dois.
Para nosso último conjunto de novos resultados, exploramos uma situação semelhante,
agora para conjuntos de dados iniciais contendo MOTS, sob condições adequadas. Para este
caso, estruturas de variedades com dimensão infinita já estabelecidas na literatura podem ser
adotadas. Com métodos de análise funcional baseados no trabalho de Biliotti, Javaloyes, and
Piccione (2009), obtemos a genericidade da incompletude luminosa para este caso.

Palavras chave: geometria Lorentziana, incompletude geodésica, genericidade.

Abstract
The objective of this thesis is to obtain conditions on spacetimes for when the phenomena
of nonspacelike geodesic incompleteness is generic in the topological sense (i.e. is valid in a
residual set).
In our first set of new results, we basis our topological techniques from the ones developed
by Lerner (1973) and obtain a genericity of geodesic incompleteness theorem in a class of
spacetime metrics with strong Whitney topology (under suitable open conditions) for a fixed
noncompact manifold containing codimension two weakly trapped submanifolds. With some
extra restrictions for the curvature on such classes of spacetimes we can give an analogous higher
codimensional result.
For our last set of new results we explore a similar situation, now for initial data sets
containing MOTS, under suitable conditions. For this case known infinite dimensional manifold
structures can adopted. With functional-analytical methods based on the work of Biliotti,
Javaloyes, and Piccione (2009) we obtain genericity of null incompleteness for this case as well.

Keywords: Lorentzian geometry, geodesic incompleteness, genericity.

Resumo Expandido

Introdução
A importância dos teoremas de singularidade (Beem, Ehrlich, and Easley (1999), Hawking
(1966), Hawking and Penrose (1970) and Penrose (1965)) em teorias geométricas da gravidade
não podem ser subestimadas. Como é bem conhecido, todos os teoremas deste tipo estabelecem
a existência de geodésicas causais inextensíveis incompletas (as chamadas “singularidades”)
em espaços-tempos sob suposições geométricas fisicamente motivadas. Quase tão importante
é garantir que as conclusões e/ou suposições nesses teoremas são estáveis sob “pequenas
perturbações” da métrica no espaço-tempo se tais devem ser fisicamente relevantes, uma vez que
atingir a precisão absoluta para medição de campos físicos tais como o gravitacional é impossível
em princípio.
O trabalho seminal de Lerner (1973) introduziu um método natural para discutir questões
de estabilidade de forma rigorosa, utilizando topologias fortes de Whitney 𝐶 𝑠 no espaço
das métricas Lorentzianas em uma dada variedade. Lerner também apresentou a forma de
adequar estas topologias na relatividade matemática, analisando a estabilidade de uma série de
propriedades causais e de curvatura usadas nos teoremas de singularidade (veja também Beem,
Ehrlich, and Easley (1999), cap. 7, para uma discussão detalhada e mais resultados e referências
sobre o assunto).
Embora o interior dos buracos negros seja o lugar principal onde se espera que as
singularidades ocorram, é bem conhecido que as noções matemáticas de buracos negros e
singularidades são logicamente independentes. Roger Penrose propôs a chamada conjectura da
censura cósmica para preencher essa lacuna, afirmando aproximadamente que os buracos negros
deveriam genericamente (em um sentido adequado) surgir quando existem singularidades (veja,
por exemplo, Wald (1984), pgs. 299-308, para uma discussão didática e referências originais).
Embora a conjectura não tenha sido provada ainda, uma maneira de abordá-la é considerando a
existência de singularidades na presença de superfícies marginalmente exteriormente aprisionadas
(MOTS), que são especialmente úteis para modelar horizontes de buracos negros em um conjunto
de dados iniciais. Pertinente aos nossos propósitos aqui, um “teorema de singularidade genérico”
pode ser dado neste caso (Chruściel and Galloway (2014), prop. 1.1; veja também Silva (2012)
para resultados relacionados). Esses resultados dependem de uma variante da condição genérica
no tensor de curvatura (cf. Beem, Ehrlich, and Easley (1999), seção 2.5), uma suposição já
usada no teorema de singularidade clássico de Penrose-Hawking (Hawking and Penrose (1970)).
Embora a condição genérica parecesse ser uma restrição de curvatura um tanto forçada, seu
caráter “verdadeiramente genérico” foi analisado em espaços tangentes por Beem and Harris
(1993), e globalmente - também usando topologias de Whitney - na dissertação de mestrado mais
recente de Larsson (2014).
Em qualquer caso, uma perspectiva conceitual mais clara dos teoremas de singularidade
em Chruściel and Galloway (2014) and Silva (2012) é se eles são vistos como manifestações
da densidade/genericidade de uma classe inteira de espaços-tempos singulares próximos (em
relação a uma geometria/topologia adequada) de um espaço-tempo contendo uma subvariedade
fechada fracamenta aprisionada tal como uma MOTS. Para um exemplo de tal caso, o trabalho de
Chruściel and Galloway (2014) (ver teo. 1.2) mostra que, além de certos “casos excepcionais”,
conjuntos de dados iniciais que satisfazem a condição de energia dominante (DEC) e contêm
uma MOTS Σ podem ser arbitrariamente aproximados (na topologia 𝐶 ∞ ) por dados iniciais que
também satisfazem DEC e para os quais Σ se torna uma superfície aprisionada externa, cujo no
desenvolvimento de Cauchy a existência de uma geodésica causal inextensível incompleta pode
ser diretamente provada (se a variedade subjacente for adicionalmente não compacta).

Objetivos
A Proposta desta tese é estudar condições para topológicas e geométricas em espaços-
tempo para estudar a genericidade (no sentido topológico, ou seja, válido em um conjunto
residual) da incompletude geodésica. Abordamos estas questões em contextoa distintos mas que
se complementam: primeiro no conjunto de métricas, e em segundo no formalismo de dados
iniciais. Em cada situação precisamos lidar com condições topologicas e geometricas adequeadas
para que técnicas já bem estabelecidas na literatura possam ser adaptadas em tais contextos.

Metodologia
A metologia utilizada neste trabalho de matemática pura é a pesquisa bibliográfica por
meio de artigos, teses e livros relevantes e ja estabelicidos na literatura, e reuinões do autor com
orientador e outros pesquisadores para discutir, propor e checar a validade dos resultados obtidos.

Resultados e Discussão
O primeiro conjunto de problemas abordados são inspirados no trabalho de Lerner (1973),
onde, induzindo topologias de Whtiney 𝐶 𝑠 forte no espaço de métricas Lorentzianas sobre uma
variedade não compacta, obtem-se propriedades de estabilidade (i.e. propriedades abertas) de
estruturas relevantes em tais topologias.
Introduzimos noção auxiliar para genericidade topológica que chamamos de prevalência,
que irá captar a noção de o que é ou não “grande” dentro de conjuntos fechados em sua topologia
de subespaço, uma vez que tratar de genericidade diretamente em conjuntos que não são abertos
traz alguns fenomenos indesejáveis (exemplo 2.1.5).
Para obter resultados relevantes, nos restringimos ao conjunto de métricas temporalmente
orientadas em uma direção pré-fixada 𝑋, que são causalmente estaveis, e para as quais uma
subvariedade Σ compacta fixada é espacial (talé aberto na topologia de interesse). A partir
disso, consideramos o conjunto de métricas para as quais Σ é fracamente futuro-aprisionada,
e F A seu fecho (métricas para as quais Σ é fracamente futuro-aprisionada). Obtemos um
resultado de prevalência da incompletude causal no caso de Σ ter codimensão 2 (teorema 3.1.7),
e com considerações extras no tensor de Riemann das métricas, obtemos o análogo para Σ de
codimensão maior (teorema 3.2.6).
Em uma segunda abordagem, exploramos o problema por meio de dados iniciais de
vácuo. Com algumas restrições geométricas, tal espaço possui uma estrutura de variedade de
dimensão infinita (Chruściel and Delay (2004), Bartnik (2005)), possibilitando uma aplicação
do clássico teorema de Sard-Smale. A técnica de obter um conjunto genérico é baseada numa
adaptação das técnicas em Biliotti, Javaloyes, and Piccione (2009), para a situação do funcional
de expansão escalar 𝜃 + , onde a pré-imagem por zero é exatamente o conjunto de dados iniciais
possuindo MOTS. Com uma sequência de aplicações de tais métodos e com algumas restrições
técnicas, e um teorema de singularidade em presença de MOTS desenvolvido aqui (teorema
1.6.3) obtemos um resultado de genericidade na presença de MOTS para dados iniciais (teoremas
4.3.1 e 4.3.4).

Considerações Finais
Neste trabalho conseguimos obter essencialmente dois resultados para genericidade de
singularidades em situanções distintas mas complementares. Este problema é de interesse tanto
téorico quando aplicado para a relatividade geral, e com respostas pouco satisftórias na literatura
até o momento. Apesar de respostas positivas, algumas restrições fortes precisaram ser assumidas
para utilizar resultados técnicos disponíveis na literatura, mas não está claro se tais restrições
são realmente necessárias, e isto o que abre possibilidade para novos estudos com o objetivo
de enfraquecer tais hipóteses. Também, as técnicas envolvendo topologias de Whitney são um
tanto gerais e aparentam ter outras aplicações na geometria Lorentziana, que motivam projetos
de investigação futuros.

Palavras chave: geometria Lorentziana, incompletude geodésica, genericidade.

Contents

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

I Preliminaries 15
1 Review of Lorentzian Geometry . . . . . . . . . . . . . . . . . . . . 16
1.1 Semi-Riemmanian and Lorentzian Manifolds . . . . . . . . . . . . . . . 16
1.1.1 Connection and Curvature . . . . . . . . . . . . . . . . . . . . . . . 17
1.1.2 Connections Over Maps . . . . . . . . . . . . . . . . . . . . . . . . 18
1.1.3 Geometry of Immersions . . . . . . . . . . . . . . . . . . . . . . . 19
1.2 Spacetimes and Causality . . . . . . . . . . . . . . . . . . . . . . . 21
1.3 The Generic Property for Spacetimes . . . . . . . . . . . . . . . . . . . 23
1.4 Trapped Surfaces and Submanifolds . . . . . . . . . . . . . . . . . . . 24
1.5 Einstein Field Equations and the Initial Data Formulation . . . . . . . . . . . 25
1.5.1 Constraint Equations and Initial Data . . . . . . . . . . . . . . . . . . . 26
1.5.2 MOTS and Initial Data . . . . . . . . . . . . . . . . . . . . . . . . 27
1.5.3 MOTS Stability Operator . . . . . . . . . . . . . . . . . . . . . . . 28
1.6 Singularity Theorems . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.6.1 A Singularity Theorem in the Presence of MOTS . . . . . . . . . . . . . . 30
2 Genericity and Prevalent Properties for Spacetimes - Basic Notions . . . . . 32
2.1 The Notion of Topological Genericity . . . . . . . . . . . . . . . . . . 32
2.2 Topological and Geometrical Settings . . . . . . . . . . . . . . . . . . 34
2.2.1 How Generic is the Generic Condition for Spacetimes . . . . . . . . . . . . 35
2.3 Further Generic and Stability Properties. . . . . . . . . . . . . . . . . . 36

II Prevalence of Singularities - Whitney Topologies 38

3 Genericity with Weakly Trapped Surfaces . . . . . . . . . . . . . . . . 39
3.1 Main Results I: Codimension Two . . . . . . . . . . . . . . . . . . . . 39
3.1.1 Singularities in Codimension Two . . . . . . . . . . . . . . . . . . . . 44
3.2 Main results II: Higher Codimensions . . . . . . . . . . . . . . . . . . 45
3.2.1 Singularities in Higher Codimensions . . . . . . . . . . . . . . . . . . 49

III Genericity of Singularities - Initial Data 51

4 Genericity of Singularities in Initial Data sets with MOTS . . . . . . . . . 52
4.1 Abstract Banach Manifold Genericity Result . . . . . . . . . . . . . . . . 52
4.2 Manifold Structures. . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2.1 Manifold of Initial Data Sets . . . . . . . . . . . . . . . . . . . . . . 57
4.2.2 Manifold of Embeddings. . . . . . . . . . . . . . . . . . . . . . . . 58
4.2.3 Null Expansion Scalar Function 𝜃 + . . . . . . . . . . . . . . . . . . . . 59
4.2.4 Linearization of 𝜃 + . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3 Geometric Interpretation. . . . . . . . . . . . . . . . . . . . . . . . 62
Appendices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
A Whitney Topologies . . . . . . . . . . . . . . . . . . . . . . . . . 65
A.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
A.2 Space of Jets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
A.3 Whitney Topologies . . . . . . . . . . . . . . . . . . . . . . . . . 68
A.3.1 Weak and Strong Topologies in the Space of Continuous Functions . . . . . . . 68
A.3.2 𝐶 𝑟 Whitney Topologies . . . . . . . . . . . . . . . . . . . . . . . . 69
A.3.3 Convergence of Sequences . . . . . . . . . . . . . . . . . . . . . . . 71
A.3.4 Baire Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
A.3.5 Thom’s Tranversality . . . . . . . . . . . . . . . . . . . . . . . . . 72
B MOTS Stability Operator . . . . . . . . . . . . . . . . . . . . . . 74
B.1 MOTS Stability Operator . . . . . . . . . . . . . . . . . . . . . . . 74
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
12

Introduction

The importance of the singularity theorems (Beem, Ehrlich, and Easley (1999), Hawking
(1966), Hawking and Penrose (1970) and Penrose (1965)) in geometric theories of gravity cannot
be overstated. As is well known, all theorems of this kind establish the existence of incomplete
inextendible causal geodesics (the so called “singularities”) in spacetimes under physically
motivated geometric assumptions. Almost as important is to ensure that the conclusions and/or
assumptions in these theorems are stable under “small perturbations” of the spacetime metric if
they are to be truly physically relevant, since attaining infinite precision for the values of physical
fields such as the gravitational one is untenable on principle.
The seminal work by Lerner (1973) introduced a natural framework to discuss stability
issues rigorously and in full nonlinear generality, under strong Whitney 𝐶 𝑠 topologies on the
space of Lorentzian metrics on a given manifold. Lerner also presented a cogent case for the
special suitability of these topologies in mathematical relativity, analyzing and establishing the
stability of a number of causal and curvature properties used in the original singularity theorems
(see also Beem, Ehrlich, and Easley (1999), Ch. 7, for a detailed discussion and further results
and references on the subject).
Although the interior of black holes are prime places where singularities are expected
to occur, it is well-known that the mathematical notions of black holes and singularities are
logically independent. This was already recognized by R. Penrose himself, who proposed
the so-called cosmic censorship conjecture to bridge this gap, roughly stating that black holes
should generically (in a suitable sense) arise when singularities exist (see, e.g., Wald (1984), pgs.
299-308, for a didactic discussion and original references). Although the full conjecture remains
elusive, one way to approach it is by considering the existence of singularities in the presence of
marginally outer trapped surfaces (MOTS), which are especially useful to model black holes
horizons in initial data sets. Pertinent to our purposes here, a “generic singularity theorem” can
be given in this case (Chruściel and Galloway (2014), prop. 1.1; see also Silva (2012) for related
results). These results rely on a variant of the generic condition on the curvature tensor (cf. Beem,
Ehrlich, and Easley (1999), section 2.5), an assumption already used in the Penrose-Hawking
classic singularity theorem (Hawking and Penrose (1970)). Although the generic condition at
first seemed to be a somewhat contrived curvature constraint, its “truly generic character” has
been analyzed at tangent spaces by Beem and Harris (1993), and globally - also using Whitney
topologies - in the more recent master’s thesis of Larsson (2014).
In any case, an arguably more transparent conceptual perspective of the singularity
theorems in Chruściel and Galloway (2014) and Silva (2012) is if they are viewed as manifestations
of the density/genericity of a whole class of singular spacetimes near (with respect to a suitable
geometry/topology) a spacetime containing a closed weakly trapped submanifold such as a
MOTS. For an example of such case, the work of Chruściel and Galloway (2014) (see thm. 1.2)
 13

shows that apart from certain “exceptional cases”, initial data sets satisfying the dominant energy
condition (DEC) and containing a MOTS Σ can be arbitrarily approximated (in the 𝐶 ∞ -topology)
by initial data sets also satisfying DEC for which Σ becomes an outer trapped surface, in whose
Cauchy development the existence of an incomplete inextendible causal geodesic can be directly
proven (if the underlying manifold is in addition noncompact).
The objective of this thesis is, in the spirit of Chruściel and Galloway (2014), to develop
techniques as to obtain conditions where geodesic incompleteness is generic for suitable topologies
on the set metrics over a given manifold with certain topological and geometrical properties that
will allow us to use some well established (and variations of) topological and functional-analytical
methods.
The original part of the work has been divided into two main sets. The first set of
genericity results presented here have been heavily influenced by the ideas in Lerner (1973) and in
particular we work with the strong (or fine) Whitney topologies throughout, using comparatively
more well-known topological notions, and applying these to a broader problem: that of analyzing
the genericity of causal geodesic incompleteness in spacetimes containing the so-called weakly
trapped submanifolds, a class which includes MOTS. The second group of genericity results
discussed here uses rather different techniques, and are now directly influenced by the arguments
in Chruściel and Galloway (2014). Our goal here is obtain, in similar fashion, a “genericity
of incompleteness near MOTS” type of result. Dealing with with the so-called initial data
sets rather than working direct with spacetimes has the advantage that in this context there are
infinite dimensional (Banach/Hilbert) manifold structures for the set of initial data (under some
reasonable restrictions, cf. Bartnik (2005) and Chruściel and Delay (2003)), and here the basic
technical tool to obtain genericity is the Sard-Smale theorem (Smale (1965)).
Of course, there are advantages and disadvantages of each approach. Among the perks
of the first approach we might count: (i) the proofs are less technical since they rely on
already standard topological techniques, (ii) our curvature assumption in the codimension 2 case
(strong energy condition) is strictly weaker than the dominant energy condition, (iii) we handle
weakly trapped submanifolds, a larger class than just MOTS, (iv) we include a result for higher
codimension with little extra cost, and last but not least (v) we weaken the causality requirements
on spacetime. The latter point is relevant especially in physical applications, because unless
strong cosmic censorship applies, any incomplete causal geodesic one predicts in the (globally
hyperbolic) maximal Cauchy development of a given initial data set might still be complete in an
isometric extension of lower causality, as the case of data induced in a suitable smooth partial
Cauchy hypersurfaces in anti-de Sitter spacetime (which is in particular stably causal but not
globally hyperbolic) illustrates.
On the flip side, however, the initial data approach, while technically harder, is more
convenient if one wishes to focus one’s attention only on MOTS, which are natural models, in
this context, for black hole horizons. Initial data sets also have broader applicability in PDE
analysis of the Einstein fields equations of general relativity, not least in numerical methods (cf.
14

Cook (2000)). From a physical perspective, again, one might argue that one can hardly expect to
glean information of spacetimes as a whole. Rather, all one can expect is to make predictions
from current data, and initial data sets are a natural model for such a situation. While we can
only guarantee genericity of causal incompleteness for the Cauchy development of initial data
sets, and in particular only for globally hyperbolic spacetimes, such already cover a vast amount
of interesting and relevant cases.
This thesis is organized as follows. In part I we recall some basic notions of geometry
and topology forms the basic language of the work, mostly to establish notation and terminology.
In chapter 1 we review semi-Riemannian and spacetime geometry. Most of the contents in this
part is quite standard, but we also introduce a more specialized singularity theorem (due to Silva
(2012)) on the presence of MOTS (theorem 1.6.3).
Chapter 2 is dedicated to review what one understands by genericity in the context of this
work, as well as to introduce a weaker notion of “topological largeness” which we call prevalence,
and also review stability and genericity properties for the space of Lorentzian metrics with strong
Whitney topologies (the main results here are due to Lerner (1973)).
With the basics established, we go to the first part of our main results. In chapter 3 we in-
troduce a suitable context were our causal structures are well defined. Our first prevalence/generic
on codimension 2 is theorem 3.1.7, and under some extra curvature restrictions, we give a higher
codimensional analogous (theorem 3.2.6).
Finally, chapter 4 is dedicated to the initial data/MOTS case. We start with an abstract,
Banach manifold genericity method (which has been especially adapted from Biliotti, Javaloyes,
and Piccione (2009)). "This abstract approach has the enormous advantage of flexibility: we can
choose among a number of variants of Banach/Hilbert manifold structures on initial data sets and
set of embeddings extant in the literature, subject only to relatively mild technical restrictions. It
also bypasses many of the tremendously technical details behind each of these structures. With
a functional-analytical method and a concrete separable Hilbert manifold structure, we obtain
the genericity result (theorem 4.3.1 and theorem 4.3.4) as a straightforward consequence of the
abstract machinery.
 Part I

Preliminaries
16

1 Review of Lorentzian Geometry

In this chapter we give a e review some of the main aspects of Lorentzian geometry that
will be relevant and recurrent throughout this thesis, mainly to introduce notation and terminology,
and refer the reader to standard textbooks for semi-Riemannian and Lorentzian geometry, eg.
O’Neill (1983), Beem, Ehrlich, and Easley (1999) and John M. Lee (2018) for other topics that
won’t be mentioned in this review (see also Costa e Silva (n.d.) and Espinoza (2020), also for a
discussion on initial data and MOTS see Hafemann (2023) for a comprehensive review on the
subject).

1.1 Semi-Riemmanian and Lorentzian Manifolds

Let 𝑀 be an 𝑛-dimensional smooth manifold, which will usually be denoted by 𝑀 𝑛 .
A symmetric smooth 1 (0, 2)-tensor is said to be a semi-Riemannian metric of index 𝜈 ∈ N
(0 ≤ 𝜈 ≤ 𝑛) if at each point 𝑝 ∈ 𝑀, the symmetric bilinear form 𝑔 𝑝 : 𝑇𝑝 𝑀 × 𝑇𝑝 𝑀 → R is
a non-degenerate bilinear form of index 𝜈. The pair (𝑀, 𝑔) is said to be a semi-Riemannian
manifold (of class 𝐶 𝑟 ). The relevant cases for this thesis are when 𝜈 = 0, called a Riemannian
manifold, and 𝜈 = 1 and 𝑛 ≥ 2, called a Lorentzian manifold.

Example 1.1.1. The simplest example of a semi-Riemannian manifold is given on 𝑀 = R𝑛 ,

Consider the usual Cartesian coordinates (𝑥 1 , . . . , 𝑥 𝑛 ), define a metric 𝜂 𝜈 by the line element
𝜈
∑︂ 𝑛
∑︂
𝑑𝑠2𝜈 = − (𝑑𝑥 𝑖 ) 2 + (𝑑𝑥 𝑗 ) 2 .
𝑖=1 𝑗=𝜈+1

This metric is called the semi-Euclidean metric of index 𝜈. R𝑛 with this metric is denoted by R𝑛𝜈 .
For 𝜈 = 1, we refer to R1𝑛 as the Minkowski spacetime. ◀

For (𝑀, 𝑔) a semi-Riemannian manifold with index 0 < 𝜈 < 𝑛, we can partition the
tangent vectors in mutually disjoint classes. We say 𝑣 ∈ 𝑇𝑝 𝑀 is
(i) timelike if 𝑔 𝑝 (𝑣, 𝑣) < 0,

(ii) lightlike (ou null) if 𝑔 𝑝 (𝑣, 𝑣) = 0 and 𝑣 ≠ 0,

(iii) spacelike if either 𝑔 𝑝 (𝑣, 𝑣) > 0 or 𝑣 = 0.

This is the causal character of tangent vectors. We also say that 𝑣 ∈ 𝑇𝑝 𝑀 is a causal vector if 𝑣
is either timelike or null. The set of timelike (resp. causal) vectors is called the time cone (resp.
causal cone).
1While standard textbooks on the subject describe semi-Riemannian theory for smooth objects, here finite
differentiability will be relevant at some points.
 17

Figure 1.1: Cones in Minkowski spacetime.

timelike
lightlike

spacelike spacelike

timelike

The causal character is also meaningful for other objects. A vector field 𝑋 is timelike
(resp. lightlike, spacelike or causal) if for each 𝑝 ∈ 𝑀 we have 𝑋 𝑝 timelike (resp. lightlike,
spacelike or causal). Similarly, a differentiable curve 𝛾 : 𝐼 → 𝑀 has a causal character if 𝛾 ′ (𝑡)
has the same causal character for all 𝑡 ∈ 𝐼.

1.1.1 Connection and Curvature

One of the main properties of semi-Riemannian manifolds is the existence of a canonical
connection associated with its metric. We denote by 𝔛(𝑀) the set of smooth vector fields over
𝑀 (smooth sections of 𝑇 𝑀 ), and define an affine connection over the tangent bundle 𝑇 𝑀 to be
an application ∇ : 𝔛(𝑀) × 𝔛(𝑀) → 𝔛(𝑀) satisfying
(i) ∇ 𝑋1 + 𝑓 𝑋2 𝑌 = ∇ 𝑋1 𝑌 + 𝑓 ∇ 𝑋2 𝑌 , for 𝑋1 , 𝑋2 ∈ 𝔛(𝑀), 𝑌 ∈ 𝔛(𝑀) and 𝑓 ∈ 𝐶 ∞ (𝑀).

(ii) ∇ 𝑋 (𝑌1 + 𝑐𝑌2 ) = ∇ 𝑋 𝑌1 + 𝑐∇ 𝑋 𝑌2 , for 𝑋 ∈ 𝔛(𝑀), 𝑌1 , 𝑌2 ∈ 𝔛(𝑀) and 𝑐 ∈ R.

(iii) ∇ 𝑋 ( 𝑓 𝑌 ) = (𝑋 𝑓 )𝑌 + 𝑓 ∇ 𝑋 𝑌 , for 𝑋 ∈ 𝔛(𝑀), 𝑌 ∈ 𝔛(𝑀) and 𝑓 ∈ 𝐶 ∞ (𝑀).

Such a connection is said to be symmetric or torsion free if ∇ 𝑋 𝑌 − ∇𝑌 𝑋 = [𝑋, 𝑌 ] (the Lie bracket
between 𝑋 and 𝑌 ) for all 𝑋, 𝑌 ∈ 𝔛(𝑀).
The fundamental theorem of semi-Riemmanian geometry is the following:

Theorem 1.1.2 (Costa e Silva (n.d.), thm. 3.2.1). For semi-Riemannian manifold (𝑀, 𝑔) there
exists a unique affine connection ∇𝑔 ≡ ∇ that is torsion-free and compatible with the metric 𝑔,
meaning that given 𝑋, 𝑌 , 𝑍 ∈ 𝔛(𝑀), ∇ satisfies

𝑍 (𝑔(𝑋, 𝑌 )) = 𝑔(∇𝑍 𝑋, 𝑌 ) + 𝑔(𝑋, ∇𝑍 𝑌 ).

Such connection is called the Levi-Civita connection of (𝑀, 𝑔) and can be characterized by the
Koszul formula:

2𝑔(∇ 𝑋 𝑌 , 𝑍) = 𝑋 (𝑔(𝑌 , 𝑍)) + 𝑌 (𝑔(𝑋, 𝑍)) − 𝑍 (𝑔(𝑋, 𝑌 ))

− 𝑔(𝑋, [𝑌 , 𝑍]) + 𝑔(𝑌 , [𝑍, 𝑋]) + 𝑔(𝑍, [𝑋, 𝑌 ]),
18

for all 𝑋, 𝑌 , 𝑍 ∈ 𝔛(𝑀). ■

Therefore, computations involving a connection ∇ on a semi-Riemannian manifold (𝑀, 𝑔)

will be implicitly understood to be with respect to the Levi-Civita connection.
The connection acting on local coordinate vector field defines the Christoffel symbols Γ𝑖𝑘𝑗 :

∇𝜕𝑖 𝜕 𝑗 = Γ𝑖𝑘𝑗 𝜕𝑘 .

For a symmetric connection, the Christoffel symbols are symmetric on the covariant indices
(Γ𝑖𝑘𝑗 = Γ 𝑘𝑗𝑖 ).
A connection gives rise to the curvature tensor 𝑅 : 𝔛(𝑀) × 𝔛(𝑀) × 𝔛(𝑀) → 𝔛(𝑀)
defined by
𝑅(𝑋, 𝑌 )𝑍 = ∇ 𝑋 ∇𝑌 𝑍 − ∇𝑌 ∇ 𝑋 𝑍 − ∇ [𝑋,𝑌 ] 𝑍, 𝑋, 𝑌 , 𝑍 ∈ 𝔛(𝑀).

In coordinates, 𝑅𝑖𝑗 𝑘𝑙 𝜕𝑖 = 𝑅 (𝜕𝑘 , 𝜕𝑙 ) 𝜕 𝑗 , where

𝑅𝑖𝑗 𝑘𝑙 = 𝜕𝑘 Γ𝑖𝑗𝑙 − 𝜕𝑙 Γ𝑖𝑗 𝑘 + Γ𝑚𝑗𝑙 Γ𝑚𝑘

𝑖
− Γ𝑚𝑗𝑘 Γ𝑚𝑙
𝑖
.

The Riemann tensor on the semi-Riemannian manifold (𝑀, 𝑔) is the (0, 4) tensor
metrically equivalent to the curvature tensor, here denoted by 𝑅𝑖𝑒𝑚. Its coordinate symbol is
𝑅𝑖 𝑗 𝑘𝑙 = 𝑔𝑖𝑟 𝑅𝑟𝑗 𝑘𝑙 .
The Ricci curvature tensor on (𝑀, 𝑔) is a symmetric (0, 2)-tensor Ric given by the
∑︁
contraction of the curvature tensor. In coordinates it will be denoted by 𝑅𝑖 𝑗 = 𝑛𝑘=1 𝑅𝑖𝑘
𝑘 .
𝑗
∞
Lastly, the scalar curvature 𝑆𝑐𝑎𝑙 ∈ 𝐶 (𝑀) is the trace of the Ricci curvature, 𝑆𝑐𝑎𝑙 =
𝑖 𝑗
𝑔 𝑅𝑖 𝑗 .

Remark 1.1.3. Since we will be interested in varying the metric over a manifold, these curvature
tensors will be a function of the metric, and we will sometimes use the descriptive notation 𝑅𝑔 or
𝑅(𝑔), 𝑅𝑖𝑐 𝑔 or 𝑅𝑖𝑐(𝑔) and so on to emphasize this dependence on the metric of such objects.

1.1.2 Connections Over Maps

Let 𝑀 and 𝑁 be smooth manifolds and let 𝐹 : 𝑁 → 𝑀 be a smooth map. A vector field
over 𝐹 is a map 𝑉 : 𝑁 → 𝑇 𝑀 for which 𝐹 = 𝜋 𝑀 ◦ 𝑉 holds, where 𝜋 𝑀 : 𝑇 𝑀 → 𝑀 is the standard
projection. For any smooth map 𝐹 : 𝑁 → 𝑀, we define 𝔛(𝐹) to be the set of smooth vector
fields over 𝐹, which is a 𝐶 ∞ (𝑁)-module with respect to pointwise operations. For instance, if 𝐹
is a smooth map, then given 𝑋 ∈ 𝔛(𝑀) and 𝑌 ∈ 𝔛(𝑁), the maps 𝑋 ◦ 𝐹 and 𝑑𝐹 ◦ 𝑌 are smooth
vector fields over 𝐹 where here (and hereafter) 𝑑𝐹 denotes the derivative of 𝐹.
At this point, the definition of connection can be extended to vector fields over maps as
follows: Let 𝐹 : 𝑁 → 𝑀 be a smooth map. A connection on 𝐹 is a map

𝐷 : (𝑋, 𝑉) ∈ 𝔛(𝑁) × 𝔛(𝐹) ↦→ 𝐷 𝑋 𝑉 ∈ 𝔛(𝐹),

 19

such that
1. 𝐷 is R-bilinear;

2. 𝐷 𝑓 𝑋 𝑌 = 𝑓 𝐷 𝑋 𝑌 , for 𝑋 ∈ 𝔛(𝑁), 𝑉 ∈ 𝔛(𝐹), 𝑓 ∈ 𝐶 ∞ (𝑁);

3. 𝐷 𝑋 ( 𝑓 𝑌 ) = (𝑋 𝑓 )𝑌 + 𝑓 𝐷 𝑋 𝑌 for 𝑋 ∈ 𝔛(𝑁), 𝑉 ∈ 𝔛(𝐹), 𝑓 ∈ 𝐶 ∞ (𝑁).

We can also define a curvature tensor for the connection 𝐷. The curvature tensor of a
connection 𝐷 on the map 𝐹 : 𝑁 → 𝑀 is given by

𝑅 𝐷 (𝑋, 𝑌 )𝑉 := 𝐷 𝑋 𝐷𝑌 𝑉 − 𝐷𝑌 𝐷 𝑋 𝑉 − 𝐷 [𝑋,𝑌 ] 𝑉

for all 𝑋, 𝑌 ∈ 𝔛(𝑁) and 𝑉 ∈ 𝔛(𝐹). This curvature tensor is 𝐶 ∞ (𝑁)-trilinear. The following
results regarding the so-called induced connection will be constantly employed in our calculations.
Theorem 1.1.4 (Costa e Silva (n.d.)). Let ∇ be a connection on the manifold 𝑀, let 𝑁 be any
smooth manifold and 𝐹 : 𝑁 → 𝑀 be a smooth map. Then, there exists a unique connection 𝐷 ∇
on 𝐹 such that

𝐷 ∇𝑋 (𝑉 ◦ 𝐹)( 𝑝) = ∇𝑑𝐹𝑝 ( 𝑋 𝑝 ) 𝑉 (𝐹 ( 𝑝)), ∀𝑝 ∈ 𝑁, ∀𝑋 ∈ 𝔛(𝑁), ∀𝑉 ∈ 𝔛(𝑀).

𝐷 ∇ is called the induced connection on 𝐹.

Proposition 1.1.5 (Costa e Silva (n.d.)). Let ∇ be a connection on the manifold M, let 𝑁 be any
manifold and let 𝐹 : 𝑁 → 𝑀 be a smooth map. Finally, let 𝐷 = 𝐷 ∇ be the induced connection
on 𝐹. Then
∇
𝑅𝐷𝑝 (𝑥, 𝑦)𝑣 = 𝑅 𝐹 ( 𝑝) (𝑑𝐹𝑝 (𝑥), 𝑑𝐹𝑝 (𝑦))𝑣, (1.1)

for any 𝑝 ∈ 𝑁 and ∀𝑥, 𝑦 ∈ 𝑇𝑝 𝑁, ∀𝑣 ∈ 𝑇𝐹 ( 𝑝) 𝑀. Moreover, if ∇ is symmetric, then

𝐷 𝑋 (𝑑𝐹 ◦ 𝑌 ) − 𝐷𝑌 (𝑑𝐹 ◦ 𝑋) = 𝑑𝐹 ◦ [𝑋, 𝑌 ].

When computations involving connections over maps appear in the context of semi-
Riemannian manifolds, we will always assume that 𝐷 is the induced connection from the
Levi-Civita connection.

1.1.3 Geometry of Immersions

It will be important to introduce not only some terminology and notation for the geometry
of semi-Riemannian submanifolds, but also more generally for semi-Riemannian immersions.
Let (𝑀, 𝑔) be a semi-Riemannian manifold. Given a smooth map 𝐹 : 𝑁 → 𝑀, the
pullback 𝐹 ∗ 𝑔 defines a symmetric (0, 2)-type smooth tensor field on 𝑁, but this is not necessarily a
semi-Riemannian metric. A necessary condition to ensure it is a metric is that 𝐹 is an immersion:
20

Proposition 1.1.6 (Costa e Silva (n.d.), prop. 4.1.1). Let 𝐹 : 𝑁 → 𝑀 be a smooth map into
the semi-Riemannian manifold (𝑀, 𝑔). If 𝐹 ∗ 𝑔 is a semi-Riemannian metric on 𝑁, then 𝐹 is an
immersion. ■

Given a semi-Riemannian manifold (𝑀, 𝑔) we then define a semi-Riemannian immersion

to be a smooth immersion 𝐹 : 𝑁 → 𝑀 for which (𝑁, 𝐹 ∗ 𝑔) is also a semi-Riemannian manifold.
A semi-Riemannian embedding is an embedding that is also a semi-Riemannian immersion. In
this case, 𝐹 ∗ 𝑔 is called the induced metric. When 𝑁 ⊆ 𝑀 is a submanifold, we will say that
𝑁 itself is a semi-Riemannian submanifold if the inclusion 𝑖 : 𝑁 ↩→ 𝑀 is a semi-Riemannian
embedding.
Moreover, any semi-Riemannian immersion 𝐹 : 𝑁 → 𝑀 for a Lorentzian (𝑀, 𝑔) induces
a metric 𝐹 ∗ 𝑔 which is either

1) Riemannian, in which case we say that 𝐹 is spacelike;

2) Lorentzian, in which case we say that 𝐹 is timelike;

If 𝑁 ⊆ 𝑀 is a semi-Riemannian submanifold of the Lorentzian manifold (𝑀, 𝑔), then it is said

to be spacelike [resp. timelike] if the inclusion 𝑖 : 𝑁 ↩→ 𝑀 is spacelike [resp. timelike].
The following notation for the scalar product over semi-Riemmanian immersions will be
frequent: for 𝐹 : 𝑁 → 𝑀 a semi-Riemannian immersion, given any 𝑉, 𝑊 ∈ 𝔛(𝐹), we define at
each 𝑝 ∈ 𝑁
⟨⟨𝑉, 𝑊⟩⟩ 𝑝 = 𝑔 𝐹 ( 𝑝) (𝑉𝑝 , 𝑊 𝑝 ). (1.2)

The is a natural orthogonal decomposition of vectors over a semi-Riemannian immersion:

given any 𝑝 ∈ 𝑁, and any 𝑣 ∈ 𝑇𝐹 ( 𝑝) 𝑀, there exist unique 𝑣 ⊤ ∈ 𝑑𝐹𝑝 (𝑇𝑝 𝑁) and 𝑣 ⊥ ∈ 𝑑𝐹𝑝 (𝑇𝑝 𝑁) ⊥ ,
called the tangent and normal parts of 𝑣, respectively, such that

𝑣 = 𝑣⊤ + 𝑣⊥. (1.3)

More generally, given a smooth vector field 𝑉 ∈ 𝔛(𝐹) on 𝐹, we have a pointwise decomposition
in tangent and normal parts. A smooth vector field 𝑉 ∈ 𝔛(𝐹) is said to be tangent [resp, normal]
if 𝑉𝑝 ∈ 𝑑𝐹𝑝 (𝑇𝑝 𝑁) [resp. 𝑉𝑝 ∈ 𝑑𝐹𝑝 (𝑇𝑝 𝑁) ⊥ ] for all 𝑝 ∈ 𝑁. We denote the subset of smooth
tangent [resp. normal] vector fields over 𝐹 by

𝔛 ⊤ (𝐹) [𝑟𝑒𝑠𝑝. 𝔛 ⊥ (𝐹)].

The normal-tangent decomposition of the connection over 𝐹 gives rise to a very important
geometric object.

Theorem 1.1.7 (Costa e Silva (n.d.), thm. 4.1.8). For any 𝑋, 𝑌 ∈ 𝔛(𝑁)

𝐷 𝑋 (𝑑𝐹 ◦ 𝑌 ) = 𝑑𝐹 (∇ 𝑁𝑋 𝑌 ) + 𝐼 𝐼 (𝑋, 𝑌 ), (1.4)

 21

where ∇ 𝑁 denotes the Levi-Civita connection of 𝐹 ∗ 𝑔 in 𝑁, and

𝐼 𝐼 (𝑋, 𝑌 ) := (𝐷 𝑋 (𝑑𝐹 ◦ 𝑌 )) ⊥ .

Moreover, the map 𝐼 𝐼 : 𝔛(𝑁) × 𝔛(𝑁) → 𝔛 ⊥ (𝐹) thus defined is 𝐶 ∞ (𝑀)-bilinear and symmetric,
and is called the second fundamental form tensor or shape tensor (of 𝐹). ■

From the shape tensor we define 𝐻 ∈ 𝔛 ⊥ (𝐹) the mean curvature vector of the semi-
Riemann immersion 𝐹 as the trace of the shape tensor with respect to the metric 𝐹 ∗ 𝑔. At each
𝑝 ∈ 𝑁 this vector can be computed as

dim
∑︂𝑁
𝐻𝑝 = 𝜀𝑖 𝐼 𝐼 (𝑒𝑖 , 𝑒𝑖 ),
𝑖=1

where {𝑒 1 , . . . , 𝑒 dim 𝑁 } is a 𝐹 ∗ 𝑔-orthonormal basis at 𝑇𝑝 𝑁. One can then easily check that the
definition is actually independent of the choice of the orthonormal basis.

Remark 1.1.8. As in remark 1.1.3, the mean curvature vector is also dependent on the metric
and 𝐹, so when an emphasis on the metric and the map is needed the notations 𝐻𝑔𝐹 or 𝐻 𝐹 (𝑔)
will be used.

1.2 Spacetimes and Causality

We shall restrict our analysis almost exclusively to a class of Lorentzian manifolds where
a notion of “past and future” can be given in a precise manner.
Let (𝑀, 𝑔) be a Lorentzian manifold. At each tangent space we choose a time cone.
Similarly to the usual definition of orientability of manifolds, the notion of time orientatility is
related to a continuous choice of such time cones along 𝑀. More precisely, let 𝜏 be a function
such that at each point 𝑝 ∈ 𝑀 chooses a time cone 𝜏( 𝑝) in 𝑇𝑝 𝑀.

Definition 1.2.1. We say that 𝜏 is smooth if at each 𝑝 ∈ 𝑀 there exists a smooth vector field 𝑋
defined on a open neighborhood of 𝑈 of 𝑝 such that 𝑋𝑞 ∈ 𝜏(𝑞) for all 𝑞 ∈ 𝑈 (figure 1.2). Such a
smooth function 𝜏 is called a time-orientation on (𝑀, 𝑔). If (𝑚, 𝑔) admits a time orientation, we
say that it is time orientable. The act of choosing a specific time orientation then time orients
(𝑀, 𝑔).

Proposition 1.2.2. If (𝑀, 𝑔) is a connected, time orientable Lorentzian manifold, there exists
exactly two time orientations for (𝑀, 𝑔). ■

In practice, the following lemma gives a more useful characterization of time orientability
that we shall adopt.
22

Figure 1.2: Time orientability.

𝜏𝑞
𝑈 𝑋𝑞
𝑝 𝑞

Lemma 1.2.3. A Lorentzian manifold (𝑀, 𝑔) is time orientable if and only if there exists a
globally defined smooth timelike vector field 𝑋 ∈ 𝔛(𝑀). ■

If there is such vector field 𝑋 as in lemma 1.2.3, the time orientation defined by 𝑋 is
called the time orientation induced by 𝑋.
A time orientation now gives us the notion of past and future directions. In a spacetime
(𝑀, 𝑔) with 𝑋 ∈ 𝔛(𝑀) inducing its time orientation, a causal vector 𝑣 ∈ 𝑇 𝑀 is said to be future-
directed if 𝑔(𝑣, 𝑋) < 0, and past-directed if 𝑔(𝑣, 𝑋) > 0. Similarly a vector field 𝑉 ∈ 𝔛(𝑀)
is future-directed if 𝑉𝑝 is future-directed for all 𝑝 ∈ 𝑀, and a differentiable curve 𝛾 : 𝐼 → 𝑀
is future-directed if 𝛾 ′ (𝑡) is future-directed for all 𝑡 ∈ 𝐼, with the notion of past-directed being
analogous.

Remark 1.2.4. The spacetime condition is not as restrictive as it might seem, since for any
connected Lorentzian manifold (𝑀, 𝑔) it is possible to find a double covering 𝜋 : 𝑀 ˜︁ → 𝑀 that is
also a connected Lorentzian manifold and is time orientable, so it is a spacetime locally isometric
to (𝑀, 𝑔) (see the discussion in Beem, Ehrlich, and Easley (1999) after definition 3.1).

With spacetimes having a precise notion of future and past directions, we can define the
causal relations: given 𝑝, 𝑞 points in a spacetime (𝑀, 𝑔),

• 𝑞 is in the chronological future of 𝑝 if there is a timelike future-directed curve starting at 𝑝

and ending at 𝑞.

• 𝑞 is in the causal future of 𝑝 if there is a causal future-directed curve starting at 𝑝 and

ending at 𝑞.

The notion of chronological and causal past points are defined analogously.
The chronological [resp. causal] future set of a point 𝑝 is the set of all points in the
chronological [resp. causal] future of 𝑝, denoted by 𝐼 + ( 𝑝) [resp. 𝐽 + ( 𝑝)]. By time duality, we the
denote 𝐼 − ( 𝑝), 𝐽 − ( 𝑝) for the past case, and we also have a analogous notion for past and future of
a set 𝐴 ⊆ 𝑀. Concretely, 𝐼 + ( 𝐴) = ∪ 𝑝∈𝐴 𝐼 + ( 𝑝), and analogously for 𝐽 + ( 𝐴).
 23

We define the so-called causality conditions on a spacetime that will be relevant in this
work2 .

Definition 1.2.5. Let (𝑀, 𝑔) be a spacetime. We say that (𝑀, 𝑔) is

(1) chronological if no point in 𝑀 is an element of its own chronological future (there are no
timelike loops).

(2) causal if no point in 𝑀 is an element of its own causal future (there are no causal loops).
Lastly, a subset 𝑆 ⊆ 𝑀 is said to be a Cauchy hypersurface if every timelike inextendible curve
intersects 𝑆 exactly once, and a spacetime that has a Cauchy hypersurface is said to be globally
hyperbolic3 .

1.3 The Generic Property for Spacetimes

The generic propriety is a restriction for causal vectors that is needed as a hypothesis for
some singularity theorems. Seemingly technical and algebraic in nature, it has a more physical
intepretation that we will point out shortly. But the generic property has also another, precise
topological meaning that will be discussed in section 2.2.1 and that will be its most appropriate
connotation for us here.

Definition 1.3.1. On a Lorentzian manifold (𝑀, 𝑔), let 𝑣 ∈ 𝑇 𝑀 be any vector, and denote 𝑅𝑖 𝑗 𝑘𝑙
the components of the (0, 4) Riemann curvature tensor 𝑅𝑖𝑒𝑚 with respect to some local basis.
We say that 𝑣 satisfies the generic condition if

𝑣 𝑘 𝑣 𝑙 𝑣 [𝑖 𝑅 𝑗] 𝑘𝑙 [𝑚 𝑣 𝑟] ≠ 0, (1.5)

for some combination of free indices. Such a vector 𝑣 satisfying the generic condition is said to
be a generic vector.
Similarly, we say that the generic condition holds for an inextendible geodesic 𝛾 : 𝐼 → 𝑀
if for some time 𝑡 0 ∈ 𝐼 we have that 𝛾 ′ (𝑡 0 ) is a generic vector in the sense of (1.5). In this sense,
we say that generic condition holds for the Lorentzian manifold (𝑀, 𝑔) if the generic condition
holds on each inextendible causal geodesic.

The condition (1.5) for a vector 𝑣 can be expressed in the following invariant way: 𝑣 is
generic if and only if
(𝑣 ♭ ⊗ 𝑣 ♭ ) ∧⃝ 𝑅𝑖𝑒𝑚( · , 𝑣, · , 𝑣) ≠ 0,

2There are other causality conditions that we shall have no use for here and are therefore omitted.
3 It
can then be shown (cf. O’Neill (1983), ch. 14) that a Cauchy hypersurface 𝑆 is a closed subset of 𝑀, that it is
indeed a 𝐶 0 hypersurface, and that 𝑆 is achronal, i.e., there are no two of its points that can be connected via a
timelike curve segment. Also, in a globally hyoerbolic space it is also possible to choose a smooth and spacelike
Cauchy hypersurface 𝑆 (Bernal and Sánchez (2003)).
24

where 𝑣 ♭ is the 1-form metrically equivalent to 𝑣 and ∧⃝ denotes the Kulkarni-Nomizu product of
two (0,2)-tensors (John M. Lee (2018), p. 213).
The generic condition for timelike vectors can be interpreted more geometrically to be
equivalent to the condition that the tidal force operator 𝑅𝑣 = 𝑅( · , 𝑣)𝑣 for 𝑣 ∈ 𝑇 𝑀 timelike
(Beem, Ehrlich, and Easley (1999), prop. 2.7) is not the zero operator. A similar, slightly more
technical characterization exists for lightlike vectors (Beem, Ehrlich, and Easley (1999), sect.
2.5).
Geodesics of interest for the singularity theorems are causal. Physically speaking, a
timelike geodesic fails to be generic if a corresponding observer along the geodesic fails to
measure a tidal force effect. A similar analysis holds for lightlike vectors and geodesics, with
the usual vector space quotient analysis to deal with the degeneracy of certain associated vector
spaces (Beem, Ehrlich, and Easley (1999), prop. 2.11).

1.4 Trapped Surfaces and Submanifolds

We now discuss the notion of trapped and marginally outer trapped submanifolds and
surfaces; these special surfaces play many roles in Lorentzian geometry and general relativity.

Definition 1.4.1 (Trapped Submanifolds). Let (𝑀 𝑚+𝑘 , 𝑔) be a spacetime, and let Σ be an

𝑚-dimensional manifold. A map 𝜓 : Σ → 𝑀 is said to be a future trapped immersion if 𝜓 is a
spacelike immersion and the mean curvature vector field 𝐻 of 𝜓 is past-directed and timelike.
An immersed submanifold Σ ⊂ 𝑀 is a future-trapped submanifold if the inclusion Σ ↩→ 𝑀 is a
trapped immersion.

A particular case of interest is when 𝑘 = 2 and the spacelike immersion 𝜓 : Σ → 𝑀 has

trivial normal bundle in the spacetime (𝑀, 𝑔) (when Σ ⊆ 𝑀 is a codimension 2 submanifold we
say that Σ is a surface). In this case, one can see that there exists two linearly independent and
globally defined null future-directed vector fields ℓ+ , ℓ− ∈ 𝔛 ⊥ (𝜓), that can be normalized such
that ⟨⟨ℓ+ , ℓ− ⟩⟩ = −1. In this situation we can define (0, 2)-tensors over Σ called the null second
fundamental forms X± associated with ℓ± as

X± (𝑋, 𝑌 ) = ⟨⟨𝐷 𝑋 ℓ± , 𝑑𝜓 ◦ 𝑌 ⟩⟩, for 𝑋, 𝑌 ∈ 𝔛(Σ).

The null mean curvatures (or null expansion scalars) 𝜃 ± ∈ 𝐶 ∞ (Σ) are defined as

𝜃 ± = tr𝜓 ∗ 𝑔 X± = div𝜓 ∗ 𝑔 ℓ±

where 𝜓 ∗ 𝑔 is the induced Riemannian metric in Σ. With some computations we obtain a more
friendly formula for the null expansion scalars:

𝜃 ± = −⟨⟨𝐻, ℓ± ⟩⟩.
 25

These quantities are all dependent on the choice of the null vector fields ℓ± , but their
signs remain the same when multiplying by a positive smooth function, and the signs of the
null expansion scalars are the meaningful quantitiy here. For the case of a codimension two
immersion 𝜓, we see that the mean curvature vector 𝐻 of 𝜓 has the form

𝐻 = 𝜃 − ℓ+ + 𝜃 + ℓ− .

Also, ⟨⟨𝐻, 𝐻⟩⟩ = −2𝜃 + 𝜃 − . If 𝜓 is future trapped, then 𝐻 is past-directed and timelike, and
this implies both 𝜃 ± < 0. Conversely, if both 𝜃 ± are negative over Σ then clearly 𝐻 is timelike
past-directed, so 𝜓 is future-trapped, establishing the equivalence between the negative sign of
the null expansion scalars and the future-trapped characteristic of the immersion.
Other signs for 𝜃 ± are also meaningful,and here a very important case is when 𝜃 + = 0.
A surface (immersion) satisfying this condition is called a marginally outer trapped surface
(immersion) (for surfaces this is abbreviated to MOTS). For this thesis MOTS will be more
relevant to the initial data formulation, discussed in the next section.

Remark 1.4.2. Although we have developed the submanifold geometry in the broader class of
immersions, for simplicity we shall focus on embeddings from now on. However, most results
can be adapted for immersions as well (but not all, especially those in chapter 4).

1.5 Einstein Field Equations and the Initial Data Formulation

The theory of general relativity first envisioned by Einstein (1915), provides a geometric
description of the phenomenon of gravity. From the the general relativistic point of view, a
region of the universe of interest can be described via a 4-dimensional spacetime (𝑀, 𝑔), and
gravitational phenomena are a manifestation of its underlying geometry, rather than by forces as
in classical mechanics.
The dynamics of general relativity is then described by the Einstein Field Equations
(abbreviated to EFE). In coordinates, these are a system of second-order nonlinear partial
differential equations that associate the geometry of spacetime with the distribution of matter and
energy in the universe. These equations are significantly hard to solve for a number of reasons,
not only purely technical, but also conceptual. Among the latter is the fact that a distribution of
matter only makes sense in the context of a background spacetime and therefore the distribution
of matter and the geometry of spacetime must be solved simultaneously.
In the language of Lorentzian geometry, the universe is modeled as spacetime (𝑀, 𝑔),
where 𝑔 represents the gravitational field. The distribution of matter and energy is represented by
a symmetric (0, 2)-tensor field 𝑇, known as the stress-energy tensor, rather than a mass density
function in classical mechanics. Additionally, the model also incorporates a constant called
cosmological constant Λ ∈ R that represents a form of energy that is inherent to space itself. The
relationship between the spacetime (𝑀, 𝑔), the tensor 𝑇 and the constant Λ is established via the
26

EFE:

Definition 1.5.1 (Einstein field equations). Let (𝑀, 𝑔) be a spacetime and 𝑇 be a symmetric
(0, 2)-tensor field on 𝑀. We say that the spacetime (𝑀, 𝑔) is a solution to the Einstein field
equations with stress-energy tensor 𝑇 and cosmological constant Λ if 𝑔 satisfies

1
𝑅𝑖𝑐 − 𝑆𝑐𝑎𝑙 · 𝑔 + Λ · 𝑔 = 𝑇, (1.6)
2

where 𝑅𝑖𝑐, 𝑆𝑐𝑎𝑙 are the Ricci and scalar curvature associated to 𝑔. Alternatively, the equations
can be written as 𝐺 = 𝑇 − Λ𝑔, where 𝐺 is the Einstein tensor of 𝑔 defined by

1
𝐺 := 𝑅𝑖𝑐 − 𝑆𝑐𝑎𝑙 · 𝑔.
2

The simplest version of the EFE is the vacuum spacetime case, which is obtained when 𝑇
is identically zero and Λ = 0. In this scenario, the Minkowski space satisfies the EFE and it is
considered as a fundamental model of empty spacetime. The vacuum case also includes other
non-trivial solutions of interest, such as the Schwarzschild spacetime. In particular, a spacetime
(𝑀, 𝑔) is a vacuum solution of the EFE if and only if it is Ricci-flat, i.e, 𝑅𝑖𝑐 = 0. Whenever
we refer to the EFE in this thesis, we shall almost always consider only the vacuum case. In
particular, we set Λ = 0 for the rest of the discussion.

1.5.1 Constraint Equations and Initial Data

It is natural from the PDE point of view to formulate an initial value problem for the EFE,
specifying an initial metric and stress-energy tensor at a given starting time to then evolve the
EFE forward on time. The notion of “time” in general relativity is rather delicate, and unlike
usual PDE theory, there isn’t a fixed notion of time with respect to which one could define
evolution equations in the usual sense, as the idea of time for a given observer only has meaning
after we know the spacetime in question, which is actually a variable to be obtained only after we
solve the EFE. All these questions gave rise to the initial data formulation of general relativity,
which we briefly discuss now.
For (𝑀 𝑛+1 , 𝑔) a spacetime with 𝐺 its associated Einstein tensor, given 𝑢 ∈ 𝑇 𝑀 a future-
directed unit timelike vector, the energy density associated with 𝑢 is defined by 𝜌𝑢 = 𝐺 (𝑢, 𝑢),
and we also define the energy momentum current as the one-form over the vector space {𝑢}⊥
given by 𝐽𝑢 (·) = −𝐺 (𝑢, ·).
Consider now S ⊆ 𝑀 an embbeded spacelike hypersuface in 𝑀 with induced Riemannian
metric ℎ, and let K denote its scalar second fundamental form with respect to 𝑈 ∈ 𝔛 ⊥ (S) which
is the unique future-directed unit timelike vector field normal to S. The following equations,
called the constraint equations, hold on S (Wald (1984), ch. 10):
 27

Scalℎ − |K | 2ℎ + (trℎ K) 2 = 2𝜌𝑈 , (Hamiltonian Constraint Equation)

divℎ K − 𝑑 (trℎ K) = 𝐽𝑈 . (Momentum Constraint Equation)

Although the previous equations are induced on a spacelike hypersurface in a previously

defined solution of the EFE, and thus give a necessary condition on such a solution, a surprising
fact is that these conditions are actually sufficient to characterize the underlying spacetime in a
precise sense we now discuss. This allows one to start with abstractly defined initial data and
obtain a solution of the EFE from that: this summarizes the idea behind the so-called initial data
formulation of general relativity.

Definition 1.5.2 (Initial Data Set). An initial data set is a triple (S 𝑛 , ℎ, K) where (S 𝑛 , ℎ) is
a Riemannian manifold and K is a symmetric (0, 2)-tensor field on S. Given an initial data
(S, ℎ, K), we define a function 𝜌 ∈ 𝐶 ∞ (S) and a one-form 𝐽 ∈ Ω1 (S), called respectively the
energy density and energy-momentum current associated with the data by

1 [︁
Scalℎ − |K | 2ℎ + (trℎ K) 2 ,
]︁
𝜌=
2 (1.7)
𝐽 = divℎ K − 𝑑 (trℎ K).

The seminal work of Choquet-Bruhat and Geroch (1969) shows that for vacuum initial
data, i.e., those for which 𝜌 = 0 and 𝐽 = 0, and (S, ℎ, K) is one such initial data, the manifold
S can be viewed as suitably embedded spacelike hypersurfaces in a uniquely defined maximal
globally hyperbolic vacuum spacetime. We summarize this result in the following theorem:

Theorem 1.5.3. Let (S 𝑛 , ℎ) be a Riemannian manifold and let K be a smooth symmetric (0, 2)-
tensor field on S. Suppose that eqs. (1.7) are satisfied for vacuum conditions with 𝜌 = 0 and
𝐽 = 0. Then there exists an (𝑛 + 1)-dimensional vacuum spacetime (𝑀 𝑛+1 , 𝑔) such that (S 𝑛 , ℎ)
isometrically embeds into (𝑀, 𝑔) as a Cauchy hypersurface with second fundamental form K.
Furthermore, there is a unique (up to isometry) maximal such solution in the sense that any other
solution satisfying these conditions can be isometrically embedded therein. ■

1.5.2 MOTS and Initial Data

For our applications we are interested in MOTS and its null expansion scalar from an
initial data set point of view. Let Σ𝑛−1 ⊆ S 𝑛 be an embedded submanifold4 , and recall that Σ is
two-sided if it has a trivial normal bundle in S.

Definition 1.5.4 (Null expansion and MOTS - Initial data version). Let (S, ℎ, K) be an ini-
tial data set and Σ ⊆ S a two-sided embedded submanifold, with one of the two normal vectors
4 Forsimplicity we work with an embbeded submanifold for the last of the chapter. Slight modifications to the
formulas exhibited here are needed for general immersions, and we return to them on appendix B.
28

ν ∈ 𝔛 ⊥ (Σ) fixed, and referred to as the outward-pointing unit normal vector field on Σ. The
outward null expansion 𝜃 + [resp. inward null expansion 𝜃 − ] of Σ is defined as

𝜃 ± = trΣ K ± 𝐻ν , (1.8)

where 𝐻ν is the mean curvature scalar of Σ with respect to the normal ν and the partial trace is
in respect to the induced metric. For the sign of 𝜃 + we then define Σ to be

• outer trapped if 𝜃 + < 0,

• weakly outer trapped if 𝜃 + ≤ 0,

• marginally outer trapped if 𝜃 + = 0.

MOTS are important in general relativity because they provide a “quasi-local” analgue of
the fully global notion of (black hole) event horizon, and as such can be adapted to numerical
studies, for example Cook (2000). In purely mathematical terms, MOTS are important because
they provide spacetime/initial data analogues of minimal surfaces (cf. Dan A. Lee (2019), section
7.5).

1.5.3 MOTS Stability Operator

As observed above, MOTS can be viewed as spacetime analogues of minimal surfaces.
More concretely, since 𝜃 + = 0, if K is traceless (for example if K = 0, called a symmetric
data set) then by eq. (1.8) Σ has vanishing mean curvature, so the MOTS is a minimal surface.
However, while minimal surfaces can be described via a variational formulation, there is no
known analogue for a MOTS.
A powerful tool for studying minimal surfaces is the notion of stability, which comes
from the sign of the second variation of the volume measure; this concept of stability can be
generalized to the setting of MOTS through the linearization of the null expansion 𝜃 + . Such
notion was introduced by Andersson, Mars, and Simon (2008). We give a brief account of more
technical aspects of the MOTS stability operator in appendix B. For now, the relevant definition
is the following:

Definition 1.5.5 (MOTS Stability Operator - Initial Data). Let Σ𝑛−1 be a closed MOTS (com-
pact without boundary) within an initial data (S 𝑛 , ℎ, K). We define the MOTS stability operator
𝐿 : 𝐶 ∞ (Σ) → 𝐶 ∞ (Σ) to be

𝐿(𝜓) = −Δ𝜓 + 2⟨𝑋, grad 𝜓⟩ + (𝑄 + div 𝑋 − ∥ 𝑋 ∥ 2 )𝜓, ∀𝜓 ∈ 𝐶 ∞ (Σ), (1.9)

where
1 1
𝑄 = ScalΣ − [𝐽 (ν) + 𝜌] − |Kν + K | 2 . (1.10)
2 2
 29

All the geometric objects her are defined on Σ, ν is the outward pointing unit normal vector field
on Σ, Kν is the scalar second fundamental form of Σ with respect to the induced metric from
(S, ℎ) on the direction ν, 𝑋 is the vector field dual to the one-form K (ν, ·) on Σ. Finally, 𝜌 and
𝐽 are defined as in definition 1.5.2.

In the case of time-symmetric initial data (K = 0), the operator 𝐿 reduces to the self-
adjoint, classic stability (or Jacobi) operator of the minimal surface theory, which consists of
the second variation of the volume. Although the operator 𝐿 is not self-adjoint in general, the
operator possesses crucial spectral properties as stated below.

Proposition 1.5.6 (Andersson, Mars, and Simon (2008), Galloway (2018)). Let Σ be a closed
MOTS (compact without boundary) within an initial data set (S 𝑛 , ℎ, K). The following statements
hold for the MOTS stability operator 𝐿.

1. There is a real eigenvalue 𝜆 1 = 𝜆 1 (𝐿), called the principal eigenvalue of 𝐿, such that for
any other eigenvalue 𝜇, 𝑅𝑒(𝜇) ≥ 𝜆 1 . The associated eigenfunction 𝜙 ∈ 𝐶 ∞ (Σ), 𝐿𝜙 = 𝜆 1 𝜙,
is unique up to a multiplicative constant, and can be chosen to be strictly positive.

2. 𝜆 1 ≥ 0 (resp., 𝜆 1 > 0) if only and if there exist some 𝜓 ∈ 𝐶 ∞ (Σ), 𝜓 > 0, such that
𝐿 (𝜓) ≥ 0 (resp., 𝐿 (𝜓) > 0). ■

1.6 Singularity Theorems

For later reference we end this chapter by recalling the statements of the classic singularity
theorem by Hawking and Penrose, and also present a less frequent singularity theorem with the
presence of MOTS. . Both types of singularity theorems will be relevant in chapters 3 and 4.
The following is the celebrated theorem of Hawking and Penrose (1970).

Theorem 1.6.1 (Hawking-Penrose). Let (𝑀, 𝑔) be a spacetime with dimension 𝑛 ≥ 3 such that
(i) (𝑀, 𝑔) is chronological;

(ii) every inextendible causal geodesic has a pair of conjugated points;

(iii) there is a future (or past) trapped set 𝐴 ⊆ 𝑀.

Then (𝑀, 𝑔)has at least one inextendible, incomplete causal geodesic. ■

The following corollary for the Hawking and Penrose theorem is the statement needed in
chapter 3.

Corollary 1.6.2 (Beem, Ehrlich, and Easley (1999), thm. 12.47). Let (𝑀, 𝑔) be a chronolog-
ical spacetime of dimension 𝑛 ≥ 3 that satisfies the generic condition (definition 1.3.1) and the
timelike convergence condition (i.e. 𝑅𝑖𝑐(𝑣, 𝑣) ≥ 0 for all 𝑣 ∈ 𝑇 𝑀 timelike). Then (𝑀, 𝑔) is
nonspacelike incomplete if any of the following conditions are satisfied:
30

(i) (𝑀, 𝑔) has a closed trapped surface (definition 1.4.1);

(ii) (𝑀, 𝑔) has a point 𝑝 such that any null geodesic starting at 𝑝 reconverges5 at some point
in the past or future of 𝑝;

(iii) (𝑀, 𝑔) has a compact spacelike hypersurface. ■

1.6.1 A Singularity Theorem in the Presence of MOTS

Let Σ𝑛−1 ⊆ S 𝑛 be connected manifolds. We say that Σ separates S if S \ Σ is not
connected. In this case, we write S \ Σ = S+ ⊔ S− , where S± are open submanifolds of S. We
now give a singularity theorem for a spacetime in the presence of a MOTS that will be of crucial
importance in the last part of this thesis.

Theorem 1.6.3. Let (𝑀 𝑛+1 , 𝑔) be a globally hyperbolic spacetime satisfying the lightlike con-
vergence condition, let S 𝑛 ⊆ 𝑀 be a spacelike Cauchy hypersurface, and Σ𝑛−1 ⊆ S a connected,
compact MOTS without boundary that separates S, with S \ Σ = 𝑆+ ⊔ 𝑆− , where 𝑆± are
open disjoint sets in 𝑆 and 𝑆+ is not compact. Assume that the principal eigenvalue for the
MOTS stability operator of Σ is non zero. Under these hypotheses, (𝑀, 𝑔) is null geodesically
incomplete.

Remark 1.6.4. Regarding the assumptions of this theorem, the separation of S by Σ is not
essential,because one can show (the proof is analogous to that of prop. 14.48 of O’Neill (1983))
that there exists a smooth covering manifold of 𝑀 for which there exists a spacelike Cauchy
hypersurface S̃ covering S and containing a isometric copy of Σ separating S̃. Now, insofar as
we are interested in geodesic incompleteness, one may as well work in covering manifolds.

Proof (sketch). Our hypotheses are a restriction of a more general study for singularities in the
presence of MOTS done by Silva (2012), so we just give here an outline of the argument, referring
to that reference for further details.
We first argue that there is a future-directed null Σ-ray 𝜂 : [0, 𝑎) → 𝑀 with 𝜂′ (0)
being parallel to ℓ+ at 𝜂(0) (called an outward-pointing ray, with a inward-pointing ray being
defined analogously for 𝜂′ (0) parallel to ℓ− ). Since the spacetime (𝑀, 𝑔) is globally hyperbolic,
𝐸 + (Σ) = 𝜕𝐼 + (Σ). Given 𝑝 ∈ 𝐸 + (Σ) \ Σ, consider 𝜂 : [0, 1] → 𝑀 a segment of a future-directed
null generator of 𝜕𝐼 + (Σ) that starts at a point at Σ and ends in 𝑝, that is in particular normal do
Σ, so 𝜂′ (0) is either parallel to ℓ+ or ℓ− . Denote by H+ the set of points in 𝐸 + (Σ) \ Σ for which
the case 𝜂(0) parallel to ℓ+ occurs, and H− for the ℓ− case. We then have 𝐸 + (Σ) \ Σ = H+ ∪ H− ,
One sees that H+ and H− are disjoint sets: assuming they are not, taking a point 𝑝 in
their intersection we have 𝜂+ , 𝜂− two future-directed outward- and inward-pointing, respectively,
null Σ-rays starting at Σ and ending in 𝑝. We concatenate this two curves to a curve 𝛾 starting
5The more technically precise meaning of “reconvergence” in this context is the appearance of a conjugate point to
𝑝 along any null geodesic.
 31

at Σ and going to 𝑝 following 𝜂− , then going back to Σ following 𝜂+ backwards. Then using a
standard continuous function 𝜌 𝑋 : S → 𝑀 that fixes S, where 𝑋 is a timelike vector field over
𝑀 such that all maximally extended integral curves intersect S exactly once (since S is a Cauchy
hypersurface such 𝑋 exists), and 𝜌 𝑋 is defined by “sliding points of 𝑀 along the integral curves”
to their intersection with S (see e.g. O’Neill (1983), prop. 14.31, or Espinoza (2020), thm.
3.8.12). One sees that the curve 𝜌 𝑋 ◦ 𝛾 intersects Σ in a point 𝑞 = 𝜌 𝑋 (𝛾(𝑡0 )) distinct from its
endpoints, so 𝑞 ∈ Σ ∩ 𝐼 − ( 𝑝), implying 𝐼 + (Σ) ∩ 𝜕𝐼 + (Σ) ≠ ∅, a contradiction. One also sees that
𝜌 𝑋 (H+ ∪ Σ) = S+ , so by our hypothesis H+ ∪ Σ is not compact (see Silva (2012) prop. 2.4 for
the details of these arguments). The construction of the desired Σ-ray follows an argument with
the limit curve lemma laid out in Silva (2012), prop. 2.1, replacing 𝐸 + (Σ) there with H+ ∪ Σ.
It is worth pointing out that Σ need not be a MOTS in order to establish the existence of
such a null Σ-ray, as this is relevant for the next step. Since we assume the stability operator of
Σ has a non-zero principal eigenvalue 𝜆 0 ≠ 0 by proposition 1.5.6 its eigenfunction 𝜙0 can be
chosen to be strictly positive. For the normal variation with vector field 𝑉 = 𝜙0 ν, where ν is the
outward pointing unit normal vector field of Σ in S, one can show that
|︁
𝜕𝜃 + |︁|︁
= 𝐿 (𝜙0 ) = 𝜆 0 𝜙0 ,
𝜕𝑡 |︁𝑡=0

therefore under this variation we can deform Σ to be outer-trapped (i.e. 𝜃 + < 0) by moving
Σ along the variation either outwards or inwards according to the sign of 𝜆 0 . Still denoting
this deformation as Σ, we then have an outward-pointing null Σ-ray 𝜂 : [0, 𝑎) → 𝑀 normal to
this outer-trapped Σ as argued above, and this geodesic if future-incomplete because if it were
complete, then by prop. 10.43 in O’Neill (1983), it would have a focal point of Σ, which is a
contradiction since 𝜂 is a Σ-ray. ■

Example 1.6.5. In the four dimensional Minkowski space R41 with usual coordinates (𝑡, 𝑥, 𝑦, 𝑧),
consider S = {𝑡 = 0} a spacelike Cauchy hypersurface and the two dimensional plane Σ = {𝑡 =
𝑥 = 0}. By identifying 𝑦 ∼ 𝑦 +1 and 𝑧 ∼ 𝑧 +1 in R41 , the resulting spacetime (𝑀, 𝑔) is geodesically
complete, vacuum globally hyperbolic spacetime. The plane Σ under this identification wraps
Σ, therefore is a MOTS that separates S,
itself around to a totally geodesic 2-torus ˜︁ ˜︁ but this does
not contradict theorem 1.6.3. The MOTS operator of ˜︁ Σ is just the Laplacian operator over the flat
2-torus, then it has zero as a eigenvalue (Colbois (2010)). ◀
32

2 Genericity and Prevalent Properties for

Spacetimes - Basic Notions

This chapter is dedicated to introduce terminology and standard notation for the main
results in chapter 3. In particular, some key notions presented here seek to clarify what it means for
a subset to be “small”, or "negligible" in a topological space, especially in the absence of a natural
measure-theoretic analogue of smallness (“measure zero”) available on (finite-dimensional)
manifolds. We will deal with a standard notion of topological genericity usually introduced in
the context of Baire spaces, and also a less standard notion of prevalence to be introduced here,
that is simple but will be frequent in chapter 3. We also briefly review the topological setting in
the space of Lorentzian metrics and some key stability results introduced by Lerner (1973) that
will be instrumental in the first part of this thesis.

2.1 The Notion of Topological Genericity

The goal of this work is to establish, under suitable conditions, that the appearance of
singularities in a spacetime is generic, in the following precise way:

Definition 2.1.1. Let 𝑋 be topological space. A subset 𝐴 ⊂ 𝑋 is said to be residual if it contains

a countable union of open dense sets. A property of elements of 𝑋 is said to be generic if the set
of all those elements that possess it is residual.

The complementary notion of a residual set is called a meager set, that is, a set contained
in a countable union of nowhere dense sets (sets such that the interior of their closure is empty).
If every residual set is dense, we say that the topological space is a Baire space. The Baire
category theorem states that completely metrizable spaces or locally compact Hausdorff spaces
are Baire spaces (John M. Lee (2011), Thm. 4.68).
While genericity implies density on Baire spaces, the converse is certainly false: the set
of rational numbers is dense in the real line, but being countable it is also a meager set (being a
countable union of singletons).
The idea of searching for generic properties is that its complementary meager set is
topologically negligible: a nowhere dense set is so “topologically small” that the interior of its
closure is empty, and a meager set is contained in a union of such “small” sets. While such notion
of smallness doesn’t seem to be as good as a measure-theoretical idea of a null set, it is widely
adopted because there are no infinite-dimensional analogues of the Lebesgue measure (see Hunt,
Sauer, and Yorke (1992) for a discussion on this topic and for an alternative notion of smallness
in infinite dimensional linear spaces that is closer to measure theory, but that is delicate to adapt
to infinite dimensional manifolds and won’t be pursued here, and also Oxtoby (1980) for a survey
 33

of the similarities and shortcomings of Baire category in relation to measure theory). In any
case, the notion residual sets and genericity and its dual notion of topological smallness have
proven to be useful in many geometrical situations, giving rise to a number of important results
as we shall see later.
We also introduce the following alternative notion of topological smallness that will be
relevant for our first “topological smallness” theorem.
Definition 2.1.2. Let 𝑋 be a topological space and 𝐶 ⊆ 𝑋 a closed set. We shall say that a set
𝐴 ⊆ 𝐶 is prevalent1 in C if 𝐶 \ 𝐴 is a meager set in 𝑋.
It is clear from this definition, for example, that the set of irrational numbers is prevalent
in the real line, whereas the set of rationals is not.
The following straightforward topological lemma is relevant in this context, as it in
particular gives a condition for prevalence.
Lemma 2.1.3. Let 𝐶 be a closed subset of a topological space 𝑋, and 𝑈 ⊆ 𝐶 an open subset of
𝑋 such that 𝑖𝑛𝑡 (𝐶) ⊆ 𝑈. Then 𝐶 \ 𝑈 is nowhere dense in 𝑋 (and 𝑈 is prevalent in 𝐶).
(︂ )︂
Proof. Since 𝐶 \ 𝑈 is closed, 𝑖𝑛𝑡 𝐶 \ 𝑈 = 𝑖𝑛𝑡 (𝐶 \ 𝑈). Now, 𝑖𝑛𝑡 (𝐶 \ 𝑈) ⊆ 𝑖𝑛𝑡 (𝐶) \ 𝑈, and
(︂ )︂
from our hypothesis, 𝑖𝑛𝑡 (𝐶) \ 𝑈 = ∅, implying 𝑖𝑛𝑡 𝐶 \ 𝑈 = ∅, therefore 𝐶 \ 𝑈 is nowhere
dense in 𝑋. ■
We see that every residual set is prevalent.
Proposition 2.1.4. Let 𝐶 ⊆ 𝑋 be a closed subset that is a Baire space in its subspace topology
(this happens e.g. 𝑋 is completely metrizable). Then every residual set of 𝐶 (residual in the
subspace topology) is prevalent in 𝐶 (in the sense of definition 2.1.2).
⋂︁
Proof. Let 𝐴 ⊆ 𝐶 be residual in 𝐶. Then 𝐴 contains 𝑛 O𝑛 a countable collection of open dense
⋃︁
subsets of 𝐶. We have 𝐶 \ 𝐴 ⊆ 𝑛 𝐶 \ O𝑛 , with each 𝐶 \ O𝑛 being closed in 𝐶, therefore closed
in 𝑋. Now 𝑖𝑛𝑡 𝑋 (𝐶 \ O𝑛 ) is an open subset of 𝐶 contained in 𝐶 \ O𝑛 that doesn’t intersect O𝑛 ,
however O𝑛 is dense in 𝐶, therefore 𝑖𝑛𝑡 𝑋 (𝐶 \ O𝑛 ) must an empty set, so 𝐶 \ 𝐴 is a meager set of
𝑋. ■
As the next example shows, the converse to proposition 2.1.4 is false.
Example 2.1.5. Let 𝑋 = R2 with the usual topology, 𝐶 = {𝑦 ≤ 0} ∪ {𝑥 = 0, 𝑦 ≥ 0}. For
𝑈 = {𝑦 < 0}. Since (fig. 2.1) 𝐶 \ 𝑈 = {𝑥 = 0, 𝑦 > 0}, this closed line has empty interior, so 𝑈 is
prevalent in 𝐶.
Consider now the interior operation with respect to the subspace topology of 𝐶, that will
be denoted by 𝑖𝑛𝑡𝐶 . The set 𝐶 \ 𝑈 = {𝑥 = 0, 𝑦 > 0} now has interior with respect to the subspace
topology (this half line is open in 𝐶) so it cannot be meager in 𝐶, since 𝐶 is a Baire space and any
meager set therein must have empty interior. Therefore 𝑈 cannot be residual in 𝐶 (thus neither
can 𝑈) with the subspace topology. ◀
1This nomenclature is inspired by the work of Hunt, Sauer, and Yorke (1992), but it is unrelated to the measure
theoretic notion of prevalence introduced therein.
34

Figure 2.1: Sets 𝐶 and 𝑈 in example 2.1.5

𝐶 𝑈

𝐶 \𝑈

2.2 Topological and Geometrical Settings

For the remaining of this chapter and we shall fix 𝑀 a connected and non-compact
smooth (i.e. 𝐶 ∞ ) real manifold without boundary of dimension 𝑛 = 𝑚 + 𝑘 ≥ 3. Let 𝑆𝑦𝑚 2 (𝑀)
be the vector bundle of (0, 2)-type symmetric tensors on 𝑀. We denote by 𝐿 ⊂ 𝑆𝑦𝑚 2 (𝑀)
the smooth subbundle whose sections are Lorentzian metric tensors (or Lorentzian metrics for
short) on 𝑀, which is open in 𝑆𝑦𝑚 2 (𝑀) when the latter is endowed with its standard manifold
topology. In other words, a section of 𝐿 is a map 𝑔 which associates with each 𝑝 ∈ 𝑀 a
symmetric nondegenerate bilinear form 𝑔 𝑝 : 𝑇𝑝 𝑀 × 𝑇𝑝 𝑀 → R of index 1. Let Γ𝑟 (𝐿) be the
set of 𝑟-differentiable sections with 0 ≤ 𝑟 ≤ ∞, that is, the set of Lorentzian metrics 𝑔 over 𝑀
whose components 𝑔𝑖 𝑗 in local coordinates have continuous partial derivatives up to order 𝑟 for
𝑟 ≥ 1, or that are simply continuous when 𝑟 = 02 . By a metric we always mean here a Lorentzian
metric, unless stated otherwise, but we specify its degree of differentiability as needed.
Let us briefly describe the relevant topologies we shall adopt on Γ𝑟 (𝐿) here (in appendix
A we give brief description of the more technical aspects of such topologies). In general, given
any other smooth manifold 𝑁, we denote by 𝐶 𝑟 (𝑀, 𝑁) the set of all 𝑟-differentiable functions
𝑓 : 𝑀 → 𝑁. Over this set we shall always adopt the strong 𝐶 𝑠 Whitney topologies, 0 ≤ 𝑠 ≤ 𝑟. A
basis for the (strong) 𝐶 𝑠 Whitney topology over 𝐶 𝑟 (𝑀, 𝑁) (𝑠 finite) can be most readily defined
via jet bundles as follows. Given the bundle 𝐽 𝑠 (𝑀, 𝑁) of 𝑠-jets of 𝑟-differentiable maps, for each
open set O ⊆ 𝐽 𝑠 (𝑀, 𝑁) we set

𝐵 𝑠 (O) = { 𝑓 ∈ 𝐶 𝑟 (𝑀, 𝑁) : 𝑗 𝑠 𝑓 (𝑀) ⊆ O},

so the collection of all sets of this form is the desired basis. A basis for the so-called Whitney
𝐶 ∞ topology on 𝐶 ∞ (𝑀, 𝑁) is defined by taking as a basis the collection of all 𝐶 𝑠 -open sets, for
2We do not refer to these as “(of class) 𝐶 𝑟 ” to avoid notational confusion with the Whitney topologies discussed
ahead.
 35

all 0 ≤ 𝑠 < ∞.
In particular, consider a smooth fiber bundle 𝐸 over 𝑀. Then the set Γ𝑟 (𝐸) of its 𝑟-
differentiable sections will also said to be given the 𝐶 𝑠 topology since we take Γ𝑟 (𝐸) ⊆ 𝐶 𝑟 (𝑀, 𝐸)
and endow Γ𝑟 (𝐸) with the induced 𝐶 𝑠 topology.
Γ𝑟 (𝐿) will always be assumed to be endowed with the (induced) 𝐶 𝑠 topology as described. In
most of our main results we shall take 𝑠 = 2 ≤ 𝑟 ≤ ∞. In case we fix some 𝑠 ∈ Z+ in connection
with the Whitney topology 𝐶 𝑠 in a statement, that statement is meant to hold separately on Γ𝑟 (𝐿)
for each 𝑠 ≤ 𝑟 ≤ ∞.
An alternative, perhaps more concrete way of defining Whitney topologies is by using
local charts. Following Mukherjee (2015), for a given function 𝑓 ∈ 𝐶 𝑠 (𝑀, 𝑁) with 𝑠 finite,
consider the following families:
(i) Φ = {(𝑈𝑖 , 𝜑𝑖 )}𝑖∈𝐼 a locally finite family of coordinate charts of 𝑀;

(ii) Ψ = {(𝑉𝑖 , 𝜓𝑖 )}𝑖∈𝐼 a family of coordinate charts of 𝑁;

(iii) K = {𝐾𝑖 }𝑖∈𝐼 a family of compact sets in 𝑀 such that 𝐾𝑖 ⊆ 𝑈𝑖 and 𝑓 (𝐾𝑖 ) ⊆ 𝑉𝑖 for all 𝑖 ∈ 𝐼.

(iv) E = {𝜀𝑖 }𝑖∈𝐼 a family of positive real numbers.

Denote by N 𝑠 ( 𝑓 , Φ, Ψ, K, E) the set of functions 𝑔 ∈ 𝐶 𝑠 (𝑀, 𝑁) such that 𝑔(𝐾𝑖 ) ⊆ 𝑉𝑖 for all
𝑖 ∈ 𝐼 and for all multi-index 𝛼 with |𝛼| ≤ 𝑠,
∥︁ ∥︁
∥︁ 𝜕 𝛼 (𝜓 ◦ 𝑓 ◦ 𝜑−1 ) 𝜕 𝛼 (𝜓 ◦ 𝑔 ◦ 𝜑−1 ) ∥︁
𝑖 𝑖 𝑖 𝑖
(𝑎) − (𝑎)
∥︁ ∥︁
∥︁ ∥︁ < 𝜀𝑖 ,
∥︁ 𝜕𝑥 𝛼 𝜕𝑥 𝛼 ∥︁

for all 𝑎 ∈ 𝜑𝑖 (𝐾𝑖 ). Then the collection of all sets N 𝑠 ( 𝑓 , Φ, Ψ, K, E) ⊆ 𝐶 𝑠 (𝑀, 𝑁) form a basis
for the 𝐶 𝑠 topology (Mukherjee (2015), prop. 8.2.9).
In the seminal work of Lerner (1973), a number of key results using the strong Whitney
topologies on Γ𝑟 (𝐿) were obtained - some of which are briefly reviewed in the next section - and
these have ever since been widely accepted as the natural topologies to adopt in the particular
geometric setting of interest here. In any case, we work exclusively with the strong Whitney
topologies on Γ𝑟 (𝐿) for the main results in and from now on by 𝐶 𝑠 topology we always mean the
strong 𝐶 𝑠 Whitney topology.
Strong Whitney topologies are not metrizable, not even first countable if the domain
manifold is not compact (Mukherjee (2015), pgs. 239-240) so we will rely on net convergence
arguments when needed.

2.2.1 How Generic is the Generic Condition for Spacetimes

In section 1.3 we looked at an algebraic condition for vectors and inextendible geodesics
on spacetimes called the “generic condition” that is relevant for singularity theorems, this
condition been seemingly unrelated to the notion of topological genericity.
36

The relation between generic condition and topological genericity has been expected to
be positive since, heuristically speaking, physically reasonable space-times usually satisfy the
generic condition (Hawking and Ellis (1973), p. 101). The work of Beem and Harris (1993)
showed that the generic condition on each tangent space where the Riemann curvature is not
zero is in fact generic in the sense of definition 2.1.1 (more specifically, the generic vectors at
𝑝 ∈ 𝑀 form an open dense set in 𝑇𝑝 𝑀). More recently the masters thesis work of Larsson (2014)
proved using differential topology methods, specifically the Thom transversality theorem, that
the generic condition is indeed generic (Larsson (2014), Thm. 2.6.3): on the function space of
Lorentzian metrics for a given manifold 𝑀 of dimension 𝑛 ≥ 4 with strong Whitney topology
𝐶 𝑟 , 4 ≤ 𝑟 ≤ ∞, the set of metrics satisfying the generic condition forms a generic set (and is
in particular dense, since spaces of suitable tensor fields endowed with Whitney topologies are
Baire spaces).

2.3 Further Generic and Stability Properties

To keep the presentation reasonably self-contained as well as to establish further notation,
we reproduce here, for later reference, some results established in Lerner (1973) that will be
relevant for us later on.
Fix some 0 ≤ 𝑟 which is either an integer or else 𝑟 = ∞. By a 𝐶 𝑠 stable property in
𝐶 𝑟 (𝑀, 𝑁), [resp. Γ𝑟 (𝐸)] we mean a property that is valid for all functions in a 𝐶 𝑠 open subset
of 𝐶 𝑟 (𝑀, 𝑁) [resp. Γ𝑟 (𝐸)]. For each 𝑡 > 𝑠 the 𝐶 𝑡 topology is finer than the 𝐶 𝑠 topology, so a
stable property in 𝐶 𝑠 is also stable in 𝐶 𝑡 .
Denote by S T 𝑟 ⊂ Γ𝑟 (𝐿) the set of 𝑟-differentiable metrics 𝑔 such that (𝑀, 𝑔) is
time-orientable. The first relevant result is the 𝐶 0 -stability of time-orientability.

Proposition 2.3.1 (Lerner (1973), prop. 4.7). Let 𝑋 ∈ 𝔛(𝑀) be an everywhere nonzero vector
field. Then the set
S T 𝑟 (𝑋) := {𝑔 ∈ Γ𝑟 (𝐿) : 𝑋 is 𝑔-timelike}

is 𝐶 0 -open in Γ𝑟 (𝐿). In particular, S T 𝑟 is 𝐶 0 - open (and therefore 𝐶 𝑠 -open for each 0 ≤ 𝑠 ≤ 𝑟)

in Γ𝑟 (𝐿). ■

A convenient aspect of working with S T 𝑟 (𝑋) is that we can simultaneously choose the
time-orientation in all of its elements so that 𝑋 is future-directed, and we shall implicitly assume
this choice from now on. Fix a codimension 𝑘 submanifold Σ ⊆ 𝑀, and denote by SΣ𝑟 the set of
metrics in Γ𝑟 (𝐿) for which Σ is spacelike.

Proposition 2.3.2 (Lerner (1973), prop. 4.2). SΣ𝑟 is 𝐶 0 -open in Γ𝑟 (𝐿). ■

Consider also the set SC 𝑟 ⊂ S T 𝑟 of (time-orientable) metrics that are stably causal,
meaning that 𝑔 ∈ SC 𝑟 if there is a 𝐶 0 -neighborhood U ∋ 𝑔 in Γ𝑟 (𝐿) such that every 𝑔′ ∈ U
 37

is a time-orientable causal metric. The set SC 𝑟 is nonempty because 𝑀 is noncompact Lerner

(1973, p. 27, item (a)). It is 𝐶 0 -open by definition, and if we denote as CH 𝑟 ⊂ S T 𝑟 the set
of chronological, time-orientable metrics, then this set is 𝐶 0 -closed in S T 𝑟 , with SC = CH
(closure in the 𝐶 0 topology, Lerner (1973, p. 27, item (b)).
Thus, stably causal metrics is 𝐶 0 -generic in the set of chronological metrics, in the sense
that they form an open dense subset of the latter, or equivalently, that chronological but not stably
causal metrics form a nowhere dense subset of the set of all chronological metrics.
We emphasize, however, that unlike stability, this does not imply 𝐶 𝑠 -genericity for 𝑠 > 0.
This is so, of course, because while SC 𝑟 would still be 𝐶 𝑠 -open in CH 𝑟 , it might no longer be
dense in this finer topology. A consequence for us here is that in order to obtain our higher-order
genericity results we shall need to work with stably causal metrics even if the needed singularity
theorems only require chronology.
Assume now 𝑟 ≥ 2 and denote by SE 𝑟 the set of 𝑟-differentiable metrics 𝑔 ∈ Γ𝑟 (𝐿) for
which the respective Ricci tensor, denoted by 𝑅𝑖𝑐(𝑔), satisfies

𝑅𝑖𝑐(𝑔)(𝑣, 𝑣) > 0, 𝑣 ∈ 𝑇 𝑀 𝑔-causal.

This set is 𝐶 2 -open in Γ𝑟 (𝐿) (Lerner (1973, Prop. 4.3). Importantly for us here, for each 𝑔 ∈ SE 𝑟
all 𝑔-causal vectors are generic in (𝑀, 𝑔) in the sense of definition 1.3.1.
Similarly, consider E 𝑟 the set of 𝑟-differentiable metrics 𝑔 ∈ Γ𝑟 (𝐿) satisfying

𝑅𝑖𝑐(𝑔)(𝑣, 𝑣) ≥ 0, 𝑣 ∈ 𝑇 𝑀 𝑔-causal.

This set is 𝐶 2 -closed, with SE 𝑟 ⊆ E 𝑟 , where now the overbar indicates 𝐶 2 -closure ( Lerner
(1973, p. 28, item 4.4)). The main result (Lerner (1973, Prop. 4.5)) we will need in our later
arguments is the following relation between E 𝑟 and SE 𝑟 :
Theorem 2.3.3. In the 𝐶 2 strong Whitney topology, 𝑖𝑛𝑡 (E 𝑟 ) = SE 𝑟 , for all 2 ≤ 𝑟 ≤ ∞. In
particular, E 𝑟 \ SE 𝑟 is 𝐶 2 -nowhere dense in Γ𝑟 (𝐿). ■
Theorem 2.3.3 is the statement as proved by Lerner, but in the sense of definition 2.1.2,
we see that the set SE 𝑟 is actually prevalent in E 𝑟 .
Remark 2.3.4. The set E 𝑟 consists precisely of those metrics 𝑔 satisfying the so-called timelike
convergence condition: 𝑅𝑖𝑐(𝑔)(𝑣, 𝑣) ≥ 0, ∀𝑣 ∈ 𝑇 𝑀 timelike (the null vectors case being obtained
via limits). This is often referred to as the strong energy condition in the physics literature,
because it arises via the Einstein field equation in the context of general relativity, by coupling the
spacetime metric with physically relevant classical matter fields, most of which satisfy it. Hence,
it is a very common assumption in singularity theorems. Theorem 2.3.3 has thus a simple but
very suggestive meaning: only a “negligibly small” subset of the metrics satisfying the timelike
convergence condition do not admit of an arbitrarily close approximation by a metric in SE 𝑟 , a
condition which in turn will often lead to causal geodesic incompleteness.
 Part II

Prevalence of Singularities - Whitney

Topologies
 39

3 Genericity with Weakly Trapped Sur-

faces

Having understood what is meant to mean “topologically small” in the context of this
work, and with the established stability and genericity results from Lerner (1973) all laid out,
in this chapter we prove a prevalence result for future-trapped submanifolds inside the set of
weakly trapped submanifolds, and use this condition, together with theorem 2.3.3 to establish
a genericity theorem for metrics that have singularities,under the existence of future trapped
surfaces (co-dimension 2) . With some extra control over the Riemann curvature as well, we can
obtain a similar theorem for situations where higher codimensional submanifolds appear.

3.1 Main Results I: Codimension Two

We fix in this section a smooth embedded codimension 𝑘 ≥ 2 submanifold Σ 𝑚 ⊆ 𝑀 𝑚+𝑘 .
In some of our main results we shall need that Σ be compact and without boundary, and then
we will simply say that Σ is closed, not to be confused standard topological closure. In order to
simplify the notation, unless stated otherwise we work in Γ2 (𝐿) endowed with the 𝐶 2 Whitney
topology, with the understanding that everything remains valid for 𝑟-differentiable metrics with
𝑟 ≥ 2, and thus omit any 𝑟 superscripts in what follows.
Let 𝑔 ∈ Γ2 (𝐿), and let ∇ = ∇𝑔 denote its Levi-Civita connection. Recall if Σ is spacelike
in (𝑀, 𝑔), then we can define the second fundamental form tensor (or shape tensor for short) of
Σ by
𝐼 𝐼 Σ (𝑔)(𝑉, 𝑊) := (∇𝑉 𝑊) ⊥ , ∀𝑉, 𝑊 ∈ 𝔛(Σ).

where ⊥ denotes the normal part with respect to 𝑔. Given any local 𝑔-orthonormal frame
{𝐸 1 , . . . , 𝐸 𝑚 } ⊂ 𝔛(Σ) on Σ, the associated mean curvature vector of Σ is
𝑚
∑︂
Σ Σ
𝐻 (𝑔) := 𝑡𝑟 Σ 𝐼 𝐼 (𝑔) = 𝐼 𝐼 Σ (𝑔)(𝐸𝑖 , 𝐸𝑖 ).
𝑖=1

When the metric is unambiguously understood, we shall denote the associated mean
curvature vector simply by 𝐻 Σ . Straightforward coordinate computations and net convergence
arguments similar to the ones given in the proof of Lerner (1973, Prop. 4.7(b)) show that the
mapping
𝑔 ∈ Γ𝑟 (𝐿) ↦→ 𝐻 Σ (𝑔) ∈ Γ0 (𝑇 𝑀 | Σ ) (3.1)

is continuous in the 𝐶 𝑟 topology on Γ𝑟 (𝐿) for all 1 ≤ 𝑟 ≤ ∞, where 𝑇 𝑀 | Σ is the restriction of

the tangent bundle 𝑇 𝑀 to Σ.
40

Fix a vector field 𝑋 ∈ 𝔛(𝑀). Our ambient topological space for the next definitions and
results is the set S T (𝑋) ∩ SΣ ∩ SC of 𝐶 2 stably-causal, time-oriented metrics in which 𝑋 is
future-directed timelike, and for which Σ is spacelike (which is 𝐶 2 -open in Γ(𝐿) - conf. Props.
2.3.1 and 2.3.2). In that set, consider the subset A of metrics for which Σ is a future-trapped
submanifold. Recall that Σ is future-trapped if and only if 𝐻 Σ𝑝 is past-directed timelike for each
𝑝 ∈ Σ (O’Neill calls such a set future-converging O’Neill (1983, p. 435)).
More precisely, we define A as

A = {𝑔 ∈ S T (𝑋) ∩ SΣ ∩ SC : 𝑔(𝐻 Σ , 𝐻 Σ ) < 0, 𝑔(𝐻 Σ , 𝑋) > 0}.

Since 𝐻 Σ only depends on the metric coefficients and their first derivatives in suitable coordinates
(conf. e.g. O’Neill (1983, Ex. 1, p. 123)) the definition of A readily implies (again by arguments
entirely analogous to those in the proof of Lerner (1973, Prop. 4.7(b))) that this set is 𝐶 1 -open in
Γ2 (𝐿), and hence 𝐶 2 -open therein as well. Consider also the set

F A = {𝑔 ∈ S T (𝑋) ∩ SΣ ∩ SC : 𝑔(𝐻 Σ , 𝐻 Σ ) ≤ 0, 𝑔(𝐻 Σ , 𝑋) ≥ 0}.

We say that Σ is weakly future-trapped for the metric 𝑔 if 𝑔 ∈ F A. In the context of a

fixed Σ as we have here, we informally refer to metrics in A as “future-trapped metrics,” and
analogously, we say the metrics in F A are “weakly future-trapped”.
Observe that for 𝑔 ∈ F A and a point 𝑝 ∈ Σ, either 𝐻 Σ𝑝 (𝑔) is past-directed and causal, or
else it is zero. In particular, F A includes metrics 𝑔 for which Σ is an extremal submanifold, i.e.,
𝐻 Σ (𝑔) ≡ 0 identically, which occurs for example if Σ is totally geodesic with respect to 𝑔.

Remark 3.1.1. For another key class of examples of elements of F A, assume for the moment
that Σ is closed and with codimension 𝑘 = 2, and that the normal bundle of Σ is trivial. (This
can be ensured independently of the choice of the ambient metric provided suitable orientability
assumptions on Σ are made.) Then, given 𝑔 ∈ S T (𝑋) ∩ SΣ ∩ SC we may choose two normal
future-directed null vector fields ℓ± ∈ 𝔛 ⊥ (Σ) globally defined on Σ and spanning its normal
bundle. The null expansion scalars 𝜃 ± ∈ 𝐶 1 (Σ) associated with this choice are defined by

𝜃 ± := −𝑔(𝐻 Σ (𝑔), 𝐾± ).

Observe that in this case 𝑔 ∈ A, i.e., it is future-trapped, if and only if 𝜃 ± < 0. If by convention
we say that ℓ+ is outward-pointing, then Σ is a marginally outer trapped surface (MOTS) if
𝜃 + ≡ 0 on Σ. Then, for each 𝑝 ∈ Σ, since since 𝐻 Σ𝑝 , 𝐾+ ( 𝑝) ∈ (𝑇𝑝 Σ) ⊥ and the latter vector space
is a two-dimensional Lorentz space, they can be orthogonal if and only if either 𝐻 Σ𝑝 is zero or
null (parallel to 𝐾+ ), and hence 𝑔 ∈ F A. In other words, 𝐶 2 metrics for which Σ is a MOTS are
all in F A, i.e., they are weakly future-trapped.

Returning now to our main discussion, we evidently have A ⊆ F A. We will be

 41

able to prove much more regarding A and F A, the main point being that A is prevalent in
F A. Informally, weakly future-trapped spacetime metrics can “almost always” be arbitrarily
𝐶 2 -approximated by future-trapped ones.
A key tool for these proofs will be conformal perturbations of the metric, so let’s briefly
how the mean curvature vector of Σ transforms under a conformal change of metric. Let
𝑔 ∈ Γ2 (𝐿) and consider the usual conformal transformation ˆ︁𝑔 = 𝑒 2 𝑓 𝑔, where 𝑓 : 𝑀 → R is any
smooth function. The Levi-Civita connection transforms as (conf. John M. Lee (2018), p. 217)

∇ 𝑋 𝑌 = ∇ 𝑋 𝑌 + (𝑋 𝑓 )𝑌 + (𝑌 𝑓 ) 𝑋 − 𝑔(𝑋, 𝑌 )𝑔𝑟𝑎𝑑𝑔 𝑓 ,
ˆ︁

for 𝑋, 𝑌 smooth vector fields over 𝑀. Then, the shape tensor of Σ associated with the metric ˆ︁
𝑔 is

∇𝑉 𝑊) ⊥ = 𝐼 𝐼 (𝑉, 𝑊) − 𝑔(𝑉, 𝑊)(𝑔𝑟𝑎𝑑𝑔 𝑓 ) ⊥ ,

𝐼ˆ︁𝐼 (𝑉, 𝑊) = ( ˆ︁

for 𝑉, 𝑊 smooth vector fields tangent to Σ. It now readily follows that

ˆ︁Σ = 𝑒 −2 𝑓 𝐻 Σ − 𝑚𝑒 −2 𝑓 (𝑔𝑟𝑎𝑑𝑔 𝑓 ) ⊥ ,
𝐻 (3.2)

and the scalar product of 𝐻 Σ transforms as

𝑔( 𝐻
ˆ︁ ˆ︁Σ ) = 𝑒 −2 𝑓 𝑔(𝐻 Σ , 𝐻 Σ ) − 2𝑒 −2 𝑓 𝑚𝑔(𝐻 Σ , 𝑔𝑟𝑎𝑑𝑔 𝑓 )
ˆ︁Σ , 𝐻
(3.3)
+ 𝑒 −2 𝑓 𝑚 2 𝑔((𝑔𝑟𝑎𝑑𝑔 𝑓 ) ⊥ , (𝑔𝑟𝑎𝑑𝑔 𝑓 ) ⊥ ).

We are now ready to prove the announced result.

Theorem 3.1.2. F A is 𝐶 2 -closed in S T (𝑋) ∩ SΣ ∩ SC, with F A = A, and also we have

𝑖𝑛𝑡 (F A) = A. In particular, A is prevalent in F A.

Proof. Firt lets show that F A is 𝐶 2 -closed, implying A ⊆ F A. Consider {𝑔𝜆 }𝜆∈Λ a net of
metrics in F A converging to a metric 𝑔 ∈ S T (𝑋) ∩ SΣ ∩ SC in the 𝐶 2 topology. Denoting
by 𝐻𝜆Σ the mean curvature vector of Σ associated with the metric 𝑔𝜆 , by the continuity in
(3.1) we have 𝐻𝜆Σ → 𝐻 Σ in the 𝐶 0 topology on Γ(𝑇 𝑀 | Σ ), and since 𝐶 2 convergence implies
𝐶 0 convergence, we have a pointwise convergence 𝑔𝜆 (𝐻𝜆Σ , 𝐻𝜆Σ ) → 𝑔(𝐻 Σ , 𝐻 Σ ), which implies
𝑔(𝐻 Σ , 𝐻 Σ ) ≤ 0. Similarly, we have a pointwise convergence 𝑔𝜆 (𝑋, 𝐻𝜆Σ ) → 𝑔(𝑋, 𝐻 Σ ), thus
showing that 𝑔(𝑋, 𝐻 Σ ) ≥ 0, i.e., 𝑔 ∈ F A.
For F A ⊆ A, let 𝑔 ∈ F A, and if we can find a sequence 𝑔𝑛 ∈ A of metrics converging
to 𝑔 in the 𝐶 2 topology to 𝑔 such that

𝑔𝑛 (𝐻𝑛 , 𝐻𝑛 ) < 0 and 𝑔𝑛 (𝐻𝑛 , 𝑋) > 0, (3.4)

then we will have 𝑔 ∈ A.

We find such a sequence as follows: since 𝑔 is stably causal, there exists a 𝐶 2 𝑔-time
42

cosmic function 𝑡 : 𝑀 → R (Hawking and Ellis (1973), prop. 6.4.9), with 𝑔𝑟𝑎𝑑𝑔 𝑡 being
future-directed with respect to the fixed vector field 𝑋. With Σ being compact, we can find
pre-compact open neighborhoods 𝑉 and 𝑈 of Σ such that 𝑉 ⊆ 𝑈. Consider a smooth bump
function 𝜙 that is constant to 1 in 𝑉 and with support in 𝑈. By defining 𝑓 = 𝜙𝑡, then 𝑓 is
zero outside of a compact set, and has the same 𝑔-gradient as 𝑡 for points over Σ, so we define
𝑓𝑛 = 𝑓 /𝑛 and 𝑔𝑛 = 𝑒 2 𝑓𝑛 𝑔. Since 𝑓𝑛 → 0 in the 𝐶 2 topology by proposition A.3.8, we see that
𝑒 2 𝑓𝑛 converges to 1 in the 𝐶 2 topology, and then also 𝑔𝑛 → 𝑔 in the 𝐶 2 topology.
To see 𝑔𝑛 satisfies (3.4), we make some elementary computations. By eq. 3.3,

2𝑚𝑒 −2 𝑓𝑛
𝑔𝑛 (𝐻𝑛 , 𝐻𝑛 ) = 𝑒 −2 𝑓𝑛 𝑔(𝐻 Σ (𝑔), 𝐻 Σ (𝑔)) − 𝑔(𝐻 Σ (𝑔), 𝑔𝑟𝑎𝑑𝑔 𝑡)
𝑛
𝑚 2 𝑒 −2 𝑓𝑛
+ 𝑔((𝑔𝑟𝑎𝑑𝑔 𝑡) ⊥ , (𝑔𝑟𝑎𝑑𝑔 𝑡) ⊥ ).
𝑛2

With 𝑔𝑟𝑎𝑑𝑔 𝑡 future-directed and 𝐻 Σ (𝑔) being either zero or past directed (with respect to 𝑔 and
𝑋), 𝑔(𝐻 Σ (𝑔), 𝑔𝑟𝑎𝑑𝑔 𝑡) ≥ 0, and also 𝑔(𝐻 Σ (𝑔), 𝐻 Σ (𝑔)) ≤ 0, but 𝑔((𝑔𝑟𝑎𝑑𝑔 𝑡) ⊥ , (𝑔𝑟𝑎𝑑𝑔 𝑡) ⊥ ) < 0
strictly, and 𝑔𝑛 (𝐻𝑛 , 𝐻𝑛 ) < 0. Similarly, by eq. (3.2),

𝑚𝑒 −2 𝑓𝑛
𝑔𝑛 (𝐻𝑛 , 𝑋) = 𝑒 −2 𝑓𝑛 𝑔(𝐻 Σ (𝑔), 𝑋) − 𝑔((𝑔𝑟𝑎𝑑𝑔 𝑡) ⊥ , 𝑋),
𝑛

and since 𝑔(𝐻 Σ (𝑔), 𝑋) ≥ 0 and 𝑔((𝑔𝑟𝑎𝑑𝑔 𝑡) ⊥ , 𝑋) < 0 because 𝑔𝑟𝑎𝑑𝑔 𝑡 is also future directed,
we obtained the desired sequence.
A ⊆ F A is 𝐶 2 -open in Γ2 (𝐿), so certainly A ⊆ 𝑖𝑛𝑡 (F A). Consider now 𝑔 ∈ F A \ A.
We will exhibit a sequence 𝑔𝑛 ∉ F A with 𝑔𝑛 → 𝑔 in the 𝐶 2 topology, so that every 𝐶 2 -open
neighborhood of 𝑔 will have a metric not in F A, meaning 𝑔 ∉ 𝑖𝑛𝑡 (F A).
Since 𝑔 is not a metric in A, there exists a point 𝑝 ∈ 𝑀 for which either 𝑔(𝐻 Σ𝑝 , 𝐻 Σ𝑝 ) ≥ 0,
or else 𝑔(𝐻 Σ𝑝 , 𝑋 𝑝 ) ≤ 0. But 𝑔 ∈ F A, so either 𝐻 Σ𝑝 (𝑔) is lightlike, or otherwise 𝐻 Σ𝑝 (𝑔) = 0.
For the lightlike case, choose a past-directed lightlike vector 𝑣 ∈ 𝑇𝑝 Σ⊥ not collinear
to 𝐻 Σ𝑝 (which exists because 𝑘 = 𝑐𝑜𝑑𝑖𝑚 Σ ≥ 2). We easily find on some relatively compact
neighborhood 𝑈 of 𝑝 a smooth function 𝜙 ∈ 𝐶 ∞ (𝑈) such that 𝑔𝑟𝑎𝑑𝑔 𝜙| 𝑝 = 𝑣. We then extend
𝜙 globally to smooth real-valued function 𝑀 with support in 𝑈 by a standard bump function
argument, and define the sequence 𝜙𝑛 = 𝜙/𝑛. This then converges to the zero function in the 𝐶 2
topology, by proposition A.3.8, implying that 𝑒 2𝜙𝑛 converges to the constant function 1 in the
𝐶 2 topology, and thus that 𝑔𝑛 = 𝑒 2𝜙𝑛 𝑔 converges to 𝑔 in the 𝐶 2 topology on Γ2 (𝐿). Denoting
by 𝐻 𝑛 the mean curvature vector of Σ associated with the metric 𝑔𝑛 , we can now employ (3.3).
Since 𝐻 Σ𝑝 (𝑔) is (past-directed) 𝑔-lightlike, and 𝑔𝑟𝑎𝑑𝑔 𝜙| 𝑝 = 𝑣 is also past-directed lightlike and
𝑔-normal to Σ, we obtain

2𝑚 2𝜙𝑛 ( 𝑝)
𝑔𝑛 (𝐻 𝑛𝑝 , 𝐻 𝑛𝑝 ) = − 𝑒 𝑔(𝐻 Σ𝑝 (𝑔), 𝑣) > 0.
𝑛
 43

The case 𝐻 Σ𝑝 = 0 is very similar: we just choose 𝑣 ≠ 0 𝑔-spacelike in 𝑇𝑝 Σ⊥ , and the rest
of the argument proceeds analogously, but now applying (3.3) we get

𝑚 2 −2𝜙𝑛 ( 𝑝)
𝑔𝑛 (𝐻 𝑛𝑝 , 𝐻 𝑛𝑝 ) = 𝑒 𝑔(𝑣, 𝑣) > 0.
𝑛

In any case, we obtain 𝑔𝑛 ∉ F A, as desired. ■

Remark 3.1.3. Some important remarks are appropriate here.

(1) By time-duality, we obviously have analogous results for past-trapped submanifolds, with
an analogous notion of weakly past-trapped submanifolds.

(2) Consider now metrics in Γ∞ (𝐿). Then A is still 𝐶 ∞ -open, F A is 𝐶 ∞ -closed, and
A ⊆ F A (𝐶 ∞ closure). In the proof of theorem 3.1.2, one sees (because 𝜙 had compact
support) that actually the sequence 𝜙𝑛 of smooth functions converges in any 𝐶 𝑟 topology
(𝑟 finite) to 0, and thus 𝑔𝑛 converges to 𝑔 in any 𝐶 𝑟 topology. Therefore 𝑔𝑛 → 𝑔 in the 𝐶 ∞
topology, and theorem 3.1.2 can be restated for the 𝐶 ∞ topology. The exact same kind of
proof inspection in Lerner (1973) gives an analogous “𝐶 ∞ statement” for theorem 2.3.3
provided one considers only smooth metrics.

(3) Since a pointwise converging net {𝑔𝜆 } of metrics in F A has its pointwise limit in F A
this set is weakly closed therefore a Baire space (theorem A.3.11).

In Lerner (1973) it was questioned that perhaps SE = E, but left the issue open in that
reference, and this question doesn’t seem to have been addressed in the literature yet. However,
using later results on 𝐶 1 -stability of causal geodesic completeness in globally hyperbolic
spacetimes (Beem, Ehrlich, and Easley (1999, cor. 7.37)), we can now give a very simple
counterexample that shows one may indeed have E \ SE ≠ ∅.

Example 3.1.4. Concretely, in the standard Minkowski spacetime R1𝑚+1 consider the quotient
manifold 𝑀 = R𝑚+1 /Z𝑚 with induced metric 𝑔, where the Z𝑚 isometric action defined by
(𝑡, 𝑥 1 , . . . , 𝑥 𝑚 ) ∼ (𝑡, 𝑥 1 + 𝑛1 , . . . , 𝑥 𝑚 + 𝑛𝑚 ). Thus, (𝑀, 𝑔) is a flat (𝑔 ∈ E), globally hyperbolic
geodesically complete spacetime. The spacelike Cauchy hypersurface Π = {𝑡 = 0} in R𝑚+1
quotients to an 𝑚-torus and thus is compact. Since global hyperbolicity, causal geodesic
completeness and the spacelike character of Π are all 𝐶 1 -stable via the cited result, there is a
open 𝐶 2 neighborhood U of 𝑔 such that all metrics in U satisfy these properties. Now, if we had
𝑔 ∈ SE, then there would exist some ℎ ∈ SE ∩ U. However, in that case the spacetime (𝑀, ℎ)
would satisfy all the conditions in the well-known Hawking-Penrose singularity theorem (conf.
Beem, Ehrlich, and Easley (1999, Thm. 12.47)), while being causally geodesically complete, a
contradiction. Therefore 𝑔 ∈ E \ SE. ◀

While the statement of theorem 3.1.2 is an analogous version of theorem 2.3.3 in the
context of trapped and weakly trapped submanifolds, and the idea of the proof is also an adaptation
44

of conformal perturbation techniques used by Lerner (1973), in example 3.1.4 we have constructed
a situation where E 𝑟 ≠ SE 𝑟 . But in the setting of theorem 3.1.2, since we are dealing with stably
causal metrics, we were actually able to prove F A = A with a similar sequence of conformal
perturbations argument.

3.1.1 Singularities in Codimension Two

We can now combine these results with the “energy” conditions for the Ricci tensor
(theorem 2.3.3). As pointed out in remark 3.1.3(ii), the results in theorem 2.3.3 and theorem 3.1.2
can be stated for the 𝐶 ∞ topology. Because the latter assumption is the most common one in
geometry, the following discussion will be carried out for smooth metrics, while easily restated
for 𝐶 𝑟 metrics with the 𝐶 𝑟 topology with 𝑟 ≥ 2.
Recall that the set A of 𝐶 ∞ metrics for which Σ is future trapped is 𝐶 ∞ -open in
Γ∞ (𝐿), and so is SE. Therefore, M := A ∩ SE is also 𝐶 ∞ -open in Γ∞ (𝐿), and contained in
S T (𝑋) ∩ SΣ ∩ SC.
Let F M := F A ∩E be the set of smooth spacetime metrics which are both weakly trapped
and satisfy the timelike convergence conditions. This set is 𝐶 ∞ -closed in S T (𝑋) ∩ SΣ ∩ SC;
thus M ⊆ F M, and also

𝑖𝑛𝑡 (F M) = 𝑖𝑛𝑡 (F A) ∩ 𝑖𝑛𝑡 (E) = M,

therefore M is prevalent in F M.

Remark 3.1.5. Just as in the discussion showing that E \ SE is not necessarily empty given in
example example 3.1.4, the same spacetime given as counterexample therein works to show that
F M \ M may also not be empty: just consider in addition the spacelike surface originating from
the quotient of 𝑆 = {𝑡 = 𝑥 1 = 0}. Therefore 𝑔 ∈ F M, but the possibility of 𝑔 ∈ M leads to the
same contradiction.

This counterexample can now be used to display a similar phenomenon as that exhibited
in example 2.1.5.

Example 3.1.6. Denote by 𝑁 the underlying manifold in remark 3.1.5. We always have M
prevalent in F M, in particular F M \ M is nowhere dense in S T (𝑋) ∩ SΣ ∩ SC. However, we
can show that F M \ M is not nowhere dense in the subspace topology of F M for this example.
In fact, since there is some 𝑔 ∈ F M \ M, consider an open neighborhood 𝑈 of 𝑔 such
that 𝑈 ∩ M = ∅. The set 𝑉 = 𝑈 ∩ F M is now an open neighborhood of 𝑔 in F M, and clearly
𝑉 ⊆ F M \ M; thus, it cannot be nowhere dense in F M, as claimed. ◀

Now, consider the following key observation. Assume that (𝑖) 𝑔 ∈ M, (𝑖𝑖) Σ is a closed
submanifold of codimension 𝑘 = 2 (following the physics usage we simply say that Σ is a closed
surface in this case), and (𝑖𝑖𝑖) (𝑀, 𝑔) is a chronological spacetime. Then, since in particular we
 45

will have 𝑅𝑖𝑐(𝑔)(𝑣, 𝑣) > 0 for any 𝑔-causal vector, the generic condition (definition 1.3.1), will
be trivially satisfied (conf. Beem, Ehrlich, and Easley (1999, Prop. 2.12, p. 39)). But then all
the hypotheses in the Hawking-Penrose singularity theorem (corollary 1.6.2) hold and therefore
(𝑀, 𝑔) must possess at least one incomplete causal geodesic1 .
Now, since our metrics are restricted to the set S T (𝑋) ∩ SΣ ∩ SC (so in particular
chronological), for every 𝑔 ∈ M, the spacetime (𝑀, 𝑔) is causally geodesically incomplete. Since
M is prevalent in F M, we conclude that “nearly all” metrics in F M are arbitrarily 𝐶 ∞ -close to
causally incomplete metrics. More precisely, we summarize the discussion of this subsection in
the following theorem.

Theorem 3.1.7. Let (𝑀, 𝑔) be a spacetime of dimension ≥ 3 and smooth metric containing a
spacelike closed surface Σ. Assume that the following conditions hold.

1. Σ is weakly future-trapped;

2. 𝑅𝑖𝑐(𝑔)(𝑣, 𝑣) ≥ 0 for all 𝑔-timelike 𝑣 ∈ 𝑇 𝑀;

3. (𝑀, 𝑔) is stably causal.

There exists a prevalent set 𝜒 in F M such that if 𝑔 ∈ 𝜒, then there exists a net (𝑔𝜆 )𝜆∈Λ of smooth
metrics on 𝑀 such that 𝑔𝜆 → 𝑔 in the 𝐶 ∞ topology and such that for each 𝜆 ∈ Λ, the spacetime
(𝑀, 𝑔𝜆 ) satisfies the following conditions

1. Σ is future-trapped;

2. 𝑅𝑖𝑐(𝑔𝜆 )(𝑣, 𝑣) > 0 for all 𝑔𝜆 -causal 𝑣 ∈ 𝑇 𝑀;

3. (𝑀, 𝑔𝜆 ) is stably causal.

In particular, (𝑀, 𝑔𝜆 ) has at least one incomplete causal geodesic. ■

3.2 Main results II: Higher Codimensions

Our goal in this final section for the chapter is to establish a version of Theorem 3.1.7
that is valid for all higher codimensions 𝑘 ≥ 2 of Σ 𝑚 ⊂ 𝑀. Since singularity theorems in the
presence of higher co-dimensional submanifolds are more delicate, we must introduce some
further notation and results as to have a better control of the Riemannian curvature as well.
For a Lorentz metric 𝑔 ∈ Γ2 (𝐿) on 𝑀 denote by 𝑅𝑖𝑒𝑚(𝑔) the covariant Riemann curvature
(0, 4)-tensor associated with 𝑔. Applying a jet bundle argument completely analogous to the
one used to prove continuity of the Ricci tensor 𝑅𝑖𝑐 as a function of 𝑔 (conf. Lerner (1973,
p. 23–24)), one readily sees that the function

𝑅𝑖𝑒𝑚 : 𝑔 ∈ Γ𝑟 (𝐿) ↦→ 𝑅𝑖𝑒𝑚(𝑔) ∈ Γ0 (𝑇 (0,4) 𝑀)

1 For this particular result, if (𝑀, 𝑔) is chronological it is enough that 𝑔 ∈ F A ∩ SE (Silva (2012)).
46

is continuous with respect to the 𝐶 𝑟 -topology on Γ𝑟 (𝐿) for each 2 ≤ 𝑟 ≤ ∞, where now 𝑇 (0,4) 𝑀
denotes the smooth vector bundle of (0, 4)-tensors over 𝑀. Consider in Γ2 (𝐿) the subset

P = {𝑔 ∈ Γ2 (𝐿) : 𝑅𝑖𝑒𝑚(𝑔)(𝑤, 𝑣, 𝑣, 𝑤) > 0 ∀𝑣 ∈ 𝑇 𝑀 𝑔-causal and all 𝑤 non-collinear to 𝑣}.

Proposition 3.2.1. The set P is 𝐶 2 -open in Γ2 (𝐿).

Proof. Again, our arguments adapt some ideas in the proof of Lerner (1973, Prop. 4.3), but
with a number of modifications of detail. Denote by 𝐶𝑢𝑟𝑣(𝑀) the vector subbundle of 𝑇 (0,4) 𝑀
of all curvature-like tensors over 𝑀, i.e., if 𝐹 ∈ 𝐶𝑢𝑟𝑣(𝑀), then for all 𝑝 ∈ 𝑀 and all vectors
𝑤, 𝑥, 𝑦, 𝑧 ∈ 𝑇𝑝 𝑀 we have

CL1) 𝐹 (𝑤, 𝑧, 𝑥, 𝑦) = −𝐹 (𝑤, 𝑧, 𝑦, 𝑥);

CL2) 𝐹 (𝑤, 𝑧, 𝑥, 𝑦) = −𝐹 (𝑧, 𝑤, 𝑥, 𝑦);

CL3) 𝐹 (𝑤, 𝑧, 𝑥, 𝑦) + 𝐹 (𝑤, 𝑥, 𝑦, 𝑧) + 𝐹 (𝑤, 𝑦, 𝑧, 𝑥) = 0;

CL4) 𝐹 (𝑥, 𝑦, 𝑤, 𝑧) = 𝐹 (𝑤, 𝑧, 𝑥, 𝑦).

Fix 𝑔 ∈ P. Denote by 𝑇ˆ︃ 𝑀 the tangent bundle minus all zero vectors, and consider, for
0
𝐶 ∈ Γ (𝐶𝑢𝑟𝑣(𝑀)), the set

𝑉𝐶 = {𝑣 ∈ 𝑇ˆ︃
𝑀 : 𝐶 (𝑤, 𝑣, 𝑣, 𝑤) > 0 for all 𝑤 non-collinear to 𝑣}.

We argue that 𝑉𝐶 is open in 𝑇ˆ︃ 𝑀. Indeed, suppose not. Then there is a sequence 𝑣 𝑛 ∈ 𝑇ˆ︃ 𝑀 \ 𝑉𝐶
converging to some vector 𝑣 ∈ 𝑉𝐶 . For each 𝑣 𝑛 , not being an element of 𝑉𝐶 implies that there exists
𝑤 𝑛 not collinear to 𝑣 𝑛 and based at the same points respectively such that 𝐶 (𝑤 𝑛 , 𝑣 𝑛 , 𝑣 𝑛 , 𝑤 𝑛 ) ≤ 0.
Now, fix some background Riemannian metric ℎ for 𝑀. By taking into account the symmetries
for 𝐶 we can assume without loss of generality that all vectors are ℎ-unitary, and each 𝑤 𝑛 is
ℎ-normal to 𝑣 𝑛 . By the compactness of the ℎ-sphere bundle over compact subsets of 𝑀 (since 𝑣 𝑛 is
convergent), passing to a subsequence if necessary we can assume that 𝑤 𝑛 converges to a nonzero
𝑤 with same base point in 𝑀 as 𝑣, and such that ℎ(𝑣, 𝑤) = 0, implying non-collinearity. Thus,
on the one hand, 𝑣 ∈ 𝑉𝐶 , and hence 𝐶 (𝑤, 𝑣, 𝑣, 𝑤) > 0, and on the other hand the convergence
implies 𝐶 (𝑤, 𝑣, 𝑣, 𝑤) ≤ 0, a contradiction.
Denote by C𝑔 ⊆ 𝑇ˆ︃ 𝑀 the set of all 𝑔-causal vectors. Now, since 𝑔 ∈ P, we have
C𝑔 ⊆ 𝑉𝑅𝑖𝑒𝑚(𝑔) . However, C𝑔 is evidently closed in 𝑇ˆ︃ 𝑀, therefore, since the latter is a manifold
and hence a normal topological space, there exists an open set 𝑈 ⊆ 𝑇ˆ︃ 𝑀 satisfying C𝑔 ⊆ 𝑈 ⊆
𝑈 ⊆ 𝑉𝑅𝑖𝑒𝑚(𝑔) . In addition, we can assume without loss of generality that 𝑈 can be chosen to
satisfy 𝑣 ∈ 𝑈 =⇒ 𝛼𝑣 ∈ 𝑈, for nonzero 𝛼 since both sets C𝑔 and 𝑉𝑅𝑖𝑒𝑚(𝑔) have this property.
Denoting as 𝜋 : 𝐶𝑢𝑟𝑣(𝑀) → 𝑀 the bundle standard projection, we now prove the
following statement for this chosen 𝑈: for each point 𝑝 ∈ 𝑀, there exists an open set 𝑊 ⊆
𝐶𝑢𝑟𝑣(𝑀) containing 𝑅𝑖𝑒𝑚(𝑔)( 𝑝) with the property that, if for 𝐶 ∈ Γ0 (𝐶𝑢𝑟𝑣) and all 𝑞 ∈ 𝜋(𝑊)
 47

we have 𝐶 (𝑞) ∈ 𝑊, then 𝑈 ∩ 𝑇𝑞 𝑀 ⊆ 𝑉𝐶 ∩ 𝑇𝑞 𝑀. Suppose by way of contradiction that this is

false. Thus we have a nested sequence of sets 𝑊𝑛 ⊇ 𝑊𝑛+1 , all neighborhoods of 𝑅𝑖𝑒𝑚(𝑔)( 𝑝)
in 𝐶𝑢𝑟𝑣(𝑀) with 𝑊𝑛 = {𝑅𝑖𝑒𝑚(𝑔)( 𝑝)}, a sequence 𝐶𝑛 ∈ Γ0 (𝐶𝑢𝑟𝑣) and points 𝑞 𝑛 ∈ 𝜋(𝑊𝑛 )
⋂︁

such that there exists some 𝑣 𝑛 ∈ (𝑈 \ 𝑉𝐶𝑛 ) ∩ 𝑇𝑞 𝑛 𝑀, which means that there is some 𝑤 𝑛 ∈ 𝑇𝑞 𝑛 𝑀
not collinear to 𝑣 𝑛 satisfying 𝐶𝑛 (𝑞 𝑛 )(𝑤 𝑛 , 𝑣 𝑛 , 𝑣 𝑛 , 𝑤 𝑛 ) ≤ 0. Since 𝐶𝑛 (𝑞 𝑛 ) → 𝑅𝑖𝑒𝑚(𝑔)( 𝑝) in
𝐶𝑢𝑟𝑣(𝑀) and 𝑞 𝑛 → 𝑝 in 𝑀, an argument using a background Riemannian metric similar as
the one in the first part of this proof shows that 𝑣 𝑛 → 𝑣 and 𝑤 𝑛 → 𝑤 up to a subsequence,
for some nonzero 𝑣 ∈ 𝑈 ∩ 𝑇𝑝 𝑀, and 𝑤 ∈ 𝑇𝑝 𝑀 not collinear to 𝑣. Thus, on the one hand,
taking limits we get 𝑅𝑖𝑒𝑚(𝑔)( 𝑝)(𝑤, 𝑣, 𝑣, 𝑤) ≤ 0, and on the other hand 𝑣 ∈ 𝑈 ⊂ 𝑉𝑅𝑖𝑒𝑚(𝑔) , so
𝑅𝑖𝑒𝑚(𝑔) ( 𝑝)(𝑤, 𝑣, 𝑣, 𝑤) > 0, which is the desired contradiction.
The rest of the proof is quite similar to the final part that of Lerner (1973, Prop. 4.3, p.
28). Namely, using the local triviality of the bundle 𝐶𝑢𝑟𝑣(𝑀) and the neighborhoods 𝑊 above,
we can obtain a 𝐶 0 neighborhood W = W (𝑅𝑖𝑒𝑚(𝑔)) ⊂ Γ0 (𝐶𝑢𝑟𝑣(𝑀)) of 𝑅𝑖𝑒𝑚(𝑔) with the
property that for any 𝐶 ∈ W we have 𝑈 ⊂ 𝑉𝐶 . Now, by the continuity of the map 𝑅𝑖𝑒𝑚 we have
that Z := 𝑅𝑖𝑒𝑚 −1 (W) is a 𝐶 2 -open set containing 𝑔 and by construction Z ⊂ P as desired. ■

Now, let us define

F P = {𝑔 ∈ Γ2 (𝐿) : 𝑅𝑖𝑒𝑚(𝑔)(𝑤, 𝑣, 𝑣, 𝑤) ≥ 0 ∀𝑣 ∈ 𝑇 𝑀 𝑔-causal and all 𝑤 non-collinear to 𝑣}.

If we have a net (𝑔𝜆 )𝜆∈Λ in F P converging in 𝐶 2 to 𝑔, we have that 𝑅𝑖𝑒𝑚(𝑔𝜆 ) → 𝑅𝑖𝑒𝑚(𝑔) in

the 𝐶 0 topology on Γ0 (𝑇 (0,4) 𝑀), whence we conclude that 𝑔 ∈ F P; that is, F P is 𝐶 2 -closed.
We now have:

Theorem 3.2.2. In the 𝐶 2 topology on Γ2 (𝐿) we have 𝑖𝑛𝑡 (F P) = P, so P is prevalent in F P.

Proof. Clearly, P ⊂ 𝑖𝑛𝑡 (F ). To prove the other inclusion, we adapt the arguments in the proof
of theorem 3.1.2 above. Specifically, we fix 𝑔 ∈ F P \ P, and build via conformal rescalings
a sequence 𝑔𝑛 ∉ F P of metrics converging in 𝐶 2 to 𝑔. We will use the standard formulas for
how global conformal rescalings of the metric change the Riemann curvature tensor. (Conf. e.g.
John M. Lee (2018, Eq. 7.44), which although presented for Riemannian metric, does remain
valid in this Lorentzian context.)
Since 𝑔 ∉ P, we can pick a 𝑔-causal vector 𝑣 ∈ 𝑇𝑝 𝑀 for some 𝑝 ∈ 𝑀, and some 𝑤 ≠ 0
at the same base point not collinear to 𝑣 with 𝑅𝑖𝑒𝑚(𝑔)(𝑤, 𝑣, 𝑣, 𝑤) = 0. Assume first that 𝑣 is
𝑔-timelike, so 𝑤 can be assumed to be spacelike and normal to 𝑣, and both can also be assumed to
be 𝑔-unit vectors. Choose a 𝑔-normal neighborhood system (𝑥 1 , . . . , 𝑥 𝑑𝑖𝑚 𝑀 ) centered at 𝑝, which
can be chosen such that 𝑣 has components (1, 0, . . . , 0), and also such that the first component of
1
𝑤 is zero. Using such coordinates, consider the function 𝜉 (𝑥 1 , . . . , 𝑥 𝑑𝑖𝑚 𝑀 ) := 𝑒 𝑥 . We globally
extend 𝜉 with usual bump function arguments, (denoting the extension the same way) so that 𝜉 is,
1
in coordinates, 𝑒 𝑥 around 𝑝, and zero outside some compact subset of 𝑀 containing 𝑝. Define,
for each 𝑛 ∈ N, 𝜉𝑛 = 𝜉/𝑛, 𝑔𝑛 = 𝑒 2𝜉𝑛 𝑔, so that again 𝑔𝑛 → 𝑔 in the 𝐶 2 topology. With our choice
48

of coordinates, a straightforward computation gives

𝑒 2/𝑛
𝑅𝑖𝑒𝑚(𝑔𝑛 ) 𝑝 (𝑤, 𝑣, 𝑣, 𝑤) = − < 0,
𝑛

so 𝑔𝑛 ∉ F P as desired. The case when 𝑣 is 𝑔-lightlike is slightly more involved. In this

case we pick normal coordinates (𝑥 1 , . . . , 𝑥 𝑑𝑖𝑚 𝑀 ) centered at 𝑝 such that 𝑒𝑖 := 𝜕/𝜕𝑥 𝑖 ( 𝑝)
(𝑖 = 1, . . . , 𝑑𝑖𝑚 𝑀) form a 𝑔-orthonormal basis with 𝑒 1 timelike and 𝑣 = 𝑒 1 + 𝑒 2 . Let ℓ := 𝑒 1 − 𝑒 2 .
ℓ is also null, and not collinear with 𝑣. Now, write

𝑑𝑖𝑚
∑︂𝑀
𝑤 =𝑎·𝑣+𝑏·ℓ+ 𝑤 𝑖 · 𝑒𝑖 .
𝑖=3

Since 𝑤 is not collinear with 𝑣, at least one of the 𝑏, 𝑤 𝑖 are nonzero, and since the part parallel to
𝑣 gives no contribution to 𝑅𝑖𝑒𝑚(𝑔)(𝑤, 𝑣, 𝑣, 𝑤) due to the curvature symmetries we can assume
∑︁𝑑𝑖𝑚 𝑀 𝑖
without loss of generality that 𝑎 = 0. Observe that the vector 𝑖=3 𝑤 · 𝑒𝑖 is normal to both
𝑣 and ℓ, and we have the possibility that either 𝑤 is lightlike (when each 𝑤 𝑖 ≡ 0) or spacelike.
Consider first the spacelike case. Define 𝜉 (𝑥 1 , 𝑥 2 , . . . , 𝑥 𝑛 ) = (𝑥 1 + 𝑥 2 ) 2 . Proceeding thenceforth
just as in the timelike case, we now obtain

4
𝑅𝑖𝑒𝑚(𝑔𝑛 )(𝑤, 𝑣, 𝑣, 𝑤) = − 𝑔(𝑤, 𝑤) < 0.
𝑛

Suppose now 𝑤 is lightlike. Rescaling, we can assume 𝑤 = ℓ. Thus, define 𝜉 (𝑥 1 , 𝑥 2 , . . . , 𝑥 𝑛 ) =

(𝑥 1 ) 2 . Ckmputing as before we obtain

8
𝑅𝑖𝑒𝑚(𝑔𝑛 )(𝑤, 𝑣, 𝑣, 𝑤) = − < 0.
𝑛

Therefore in each case we conclude that 𝑔𝑛 ∉ F P as desired, thus completing the proof. ■

Remark 3.2.3. Again, by arguments analogous to those in remark 3.1.3(ii), the theorem 3.2.2
remains valid for the 𝐶 ∞ topology, provided we work with smooth metrics.

The following set is actually more relevant: let O ⊂ Γ2 (𝐿) be the the of 2-differentiable
metrics for which the associated tidal force operators (see Beem, Ehrlich, and Easley (1999, pp.
35 and 38) for definitions and notation) along causal directions are positive semidefinite, that is,
for 𝑔 ∈ Γ2 (𝐿) and 𝑣 ∈ 𝑇 𝑀 𝑔-causal, the linear operators

⎨ 𝑅𝑣𝑔 : 𝑣 ⊥ → 𝑣 ⊥ , 𝑣 for 𝑔-timelike,

⎧
⎪
⎪
⎪ 𝑅 𝑔 𝑣 : 𝑣 ⊥ → 𝑣 ⊥ , 𝑣 for 𝑔-lightlike,
⎪
⎩
 49

are both positive semidefinite. It is straightforward to see that

P ⊆ SE; (3.5)
P ⊆ O ⊆ FP ⊂ E (3.6)

Therefore, P ⊆ O ⊆ F P, and since 𝑖𝑛𝑡 (F P) = P, then 𝑖𝑛𝑡 (F P) ⊆ O, and the following

corollary of theorem 3.2.2 is immediate.

Corollary 3.2.4. In the 𝐶 2 topology, O is prevalent in F P. ■

Remark 3.2.5. It is straightforward to check that 𝑔 ∈ F P if and only if

𝑅𝑖𝑒𝑚(𝑔)(𝑤, 𝑣, 𝑣, 𝑤) ≥ 0, ∀𝑣 ∈ 𝑇 𝑀 𝑔-timelike and all 𝑤 ∈ 𝑣 ⊥𝑔 .

3.2.1 Singularities in Higher Codimensions

We consider a similar analysis as the one carried out in section 3.1.1 when the codimension
𝑘 of the submanifold Σ ⊂ 𝑀 is higher than 2. However, we cannot use the Hawking-Penrose
singularity theorem in this context, but rather use some analogous singularity theorem in the
presence of future-trapped submanifolds of codimension higher than 2. Just such a result has
been obtained by Galloway and Senovilla (2010). Also, similar to what we did in section 3.1.1
(cf. remark 3.2.3), the following discussion uses the 𝐶 ∞ topology on the set of smooth metrics,
and is easily adapted to 𝐶 2 metrics.
We need only to check that the positive-semidefiniteness of the tidal force operators
along causal directions codified in the set O does implies the required conditions. Specifically,
let 𝑔 ∈ A ∩ O and assume that Σ is closed. Following Galloway and Senovilla (2010), let
𝛾 : [0, 𝑏) → 𝑀 be some future-directed causal geodesic with 𝛾(0) ∈ Σ and 𝛾 ′ (0) normal to
Σ. Consider 𝑒 1 , . . . , 𝑒 𝑚 some coordinate basis of tangent vectors for 𝑇𝛾(0) Σ, and let 𝐸 1 , . . . , 𝐸 𝑚
denote their parallel-transport vector fields along 𝛾. Write 𝑔𝑎𝑏 = 𝑔(𝐸 𝑎 , 𝐸 𝑏 ) (which is constant
along 𝛾). Along this geodesic, since we have assumed that 𝑅𝛾 ′ (or 𝑅𝛾 ′ if 𝛾 is null) is positive
semidefinite in the subspaces spanned by the vectors 𝐸 𝑎 , the trace

𝑔 𝑎𝑏 𝑅𝑖𝑒𝑚(𝛾 ′, 𝐸 𝑎 , 𝐸 𝑏 , 𝛾′) ≥ 0.

But this is precisely the condition 3.1 in Galloway and Senovilla (2010). Observe also that
the inclusions (3.6) imply that the positive semidefiniteness of the tidal forces also entails the
timelike convergence condition. Therefore, if in addition we assume that (𝑖) (𝑀, 𝑔) is also
chronological and (𝑖𝑖) satisfies the generic condition for causal vectors, then (𝑀, 𝑔) is causally
geodesically incomplete by Galloway and Senovilla (2010, Thm. 3). Therefore, the arguments in
section 3.1.1 by taking proposition 3.2.1, theorem 3.2.2, remark 3.2.5 and the inclusions (3.5),
(3.6) into account can be now adapted as follows.
50

Theorem 3.2.6. Let (𝑀, 𝑔) be a spacetime of dimension ≥ 3 and smooth metric containing a
spacelike closed submanifold Σ of codimension 𝑘 ≥ 2. Assume that the following conditions
hold.

1. Σ is weakly future-trapped;

2. 𝑅𝑖𝑒𝑚(𝑔)(𝑤, 𝑣, 𝑣, 𝑤) ≥ 0 for all 𝑔-timelike 𝑣 ∈ 𝑇 𝑀 and all 𝑤 at the same base point
𝑔-orthogonal to 𝑣;

3. (𝑀, 𝑔) is stably causal.

There exists a prevalent subset 𝜒 in F P such that if𝑔 ∈ 𝜒, then there exists a net (𝑔𝜆 )𝜆∈Λ of
smooth metrics on 𝑀 such that 𝑔𝜆 → 𝑔 in the 𝐶 ∞ topology and such that for each 𝜆 ∈ Λ, the
spacetime (𝑀, 𝑔𝜆 ) satisfies the following conditions

1. Σ is future-trapped;

2. 𝑅𝑖𝑐(𝑔𝜆 )(𝑣, 𝑣) > 0 for all 𝑔𝜆 -causal 𝑣 ∈ 𝑇 𝑀, and thus the causal genericity condition
holds in (𝑀, 𝑔𝜆 );

3. the tidal force operators along 𝑔𝜆 -causal directions are all positive semidefin0ite;

4. (𝑀, 𝑔𝜆 ) is stably causal.

In particular, each (𝑀, 𝑔𝜆 ) has at least one incomplete causal geodesic. ■

 Part III

Genericity of Singularities - Initial Data

52

4 Genericity of Singularities in Initial Data

sets with MOTS

The objective of this chapter is to obtain a genericity result of singular spacetimes under
suitable conditions in similar fashion to the ones in chapter 3, but using the approach of initial
data containing a MOTS. Under concrete situations to be discussed, we will have a separable
Hilbert manifold structure for the set of triples (ℎ, K, 𝜓), where (ℎ, K). is an initial data on
a fixed connected 𝑛-manifold S and 𝜓 : Σ → S a MOTS embedding, where again Σ is a
fixed compact and connected manifold (without boundary). The genericity statement we seek
will follow from an abstract functional-analytic approach we will develop, together with the
well-known Sard-Smale theorem.

4.1 Abstract Banach Manifold Genericity Result

In order to obtain our main genericity results, we shal give a suitable Banach manifold
structure both on the set of initial data on a fixed connected 𝑛-manifold S, and on the set of
embeddings of a fixed connected compact (𝑛 − 1)-manifold Σ into S. In order to do that however,
we shall first establish in this section a key but purely abstract result on Banach manifolds,
adapting the main analysis in Biliotti, Javaloyes, and Piccione (2009). Our first lemma is an
adaptation of lemma 2.2 in the latter reference, where we substitute the kernel condition therein
for a suitable alternative on the orthogonal complement of images. For completeness we give
proofs of other results from Biliotti, Javaloyes, and Piccione (2009), clearly indicating the source
when we do so, as needed.
Here, for 𝑊 a closed subspace in a Hilbert space 𝐻, we denote by 𝑃𝑊 : 𝐻 → 𝑊 the
orthogonal projection onto 𝑊 .
We also recall that a bounded linear operator 𝑇 : 𝐸 → 𝐹 between Banach spaces is
said to be Fredholm if ker 𝑇 and coker 𝑇 have finite dimensions1 , where coker 𝑇 = 𝐸/Im 𝑇 (the
dimension of the cokernel is also called the codimension of Im 𝑇). The index of a Fredholm
operator 𝑇 is defined as ind(𝑇) = dim ker 𝑇 − dim coker 𝑇 . "For later reference, we also recall
that if 𝑀, 𝑁 are (perhaps infinite dimensional) Banach manifolds and 𝑎 is an integer number, then
a 𝐶 1 map 𝑓 : 𝑀 → 𝑁 is said to a Fredholm map of index 𝑎 if its derivative 𝑑𝑓 𝑝 : 𝑇𝑝 𝑀 → 𝑇 𝑓 ( 𝑝) 𝑁
is a Fredholm bounded linear operator of fixed index 𝑎 for every 𝑝 ∈ 𝑀

Lemma 4.1.1. Let 𝐸, 𝐹 be Banach spaces, and let 𝐻 be a Hilbert space with inner product ⟨ , ⟩.
Consider 𝑇 : 𝐸 → 𝐻, 𝑆 : 𝐹 → 𝐻 bounded linear operators, with Im 𝑆 closed and such that

1 Someaauthors explicitly require the range of 𝑇 to be closed; however, this condition follows easily from the finite
dimensionality of the cokernel.
 53

𝑃Im 𝑆⊥ (Im 𝑇) is closed in Im 𝑆 ⊥ (this happens for example if 𝑆 is a Fredholm operator). Then the
sum operator 𝑇 ⊕ 𝑆 : 𝐸 × 𝐹 → 𝐻 (given by 𝑇 ⊕ 𝑆(𝑥, 𝑦) = 𝑇 (𝑥) + 𝑆(𝑦)) is surjective if and only
if Im 𝑇 ⊥ ∩ Im 𝑆 ⊥ = {0}.

Proof. Assuming the sum operator to be surjective, given 𝑧 ∈ Im 𝑇 ⊥ ∩ Im 𝑆 ⊥ , then there is a pair
(𝑥, 𝑦) with 𝑧 = 𝑇 (𝑥) + 𝑆(𝑦), but also by perpendicularity

∥𝑧∥ 2 = ⟨𝑧, 𝑧⟩ = ⟨𝑧, 𝑇 (𝑥) + 𝑆(𝑦)⟩ = 0.

Assuming now Im 𝑇 ⊥ ∩ Im 𝑆 ⊥ = {0}, take some 𝑧 ∈ 𝐻 \ {0}. With Im 𝑆 closed,

𝐻 = Im 𝑆 ⊕ Im 𝑆 ⊥ , so 𝑧 = 𝑆(𝑥) + 𝑢, 𝑢 ∈ Im 𝑆 ⊥ . We can also split Im 𝑆 ⊥ (since it is a Hilbert
space under the restriction of the inner product and 𝑃Im 𝑆⊥ (Im 𝑇) is closed in Im 𝑆 ⊥ ) as follows:

Im 𝑆 ⊥ = [𝑃Im 𝑆⊥ (Im 𝑇)] ⊕ [(𝑃Im 𝑆⊥ (Im 𝑇)) ⊥ ∩ Im 𝑆 ⊥ ].

Now, 𝑢 = 𝑃Im 𝑆⊥ (𝑇 (𝑥)) + 𝑣, where 𝑣 ∈ (𝑃Im 𝑆⊥ (Im 𝑇)) ⊥ ∩ Im 𝑆 ⊥ . Also, 𝑇 (𝑥) = 𝑆(𝑦ˆ ) +
𝑃Im 𝑆⊥ (𝑇 (𝑥)), and we have 𝑧 = 𝑇 (𝑥) + 𝑆(𝑦) − 𝑆(𝑦ˆ ) + 𝑣. If 𝑣 = 0 we are done. Now, 𝑣 ∈ Im 𝑆 ⊥ ,
and for any 𝑇 (𝑒), 𝑒 ∈ 𝐸, again 𝑇 (𝑒) = 𝑆( 𝑓 ) + 𝑃Im 𝑆⊥ (𝑇 (𝑒)), and we obtain,

⟨𝑣, 𝑇 (𝑒)⟩ = ⟨𝑣, 𝑆( 𝑓 )⟩ + ⟨𝑣, 𝑃Im 𝑆⊥ (𝑇 (𝑒))⟩ = 0,

meaning 𝑣 ∈ Im 𝑇 ⊥ , implying 𝑣 = 0. ■

Remark 4.1.2. For the Banach adjoint 𝑇 ∗ : 𝐻 ∗ → 𝐸 ∗ , from standard functional analysis
(Bachman and Narici (2000), p. 285) we have Im 𝑇 ⊥ = ker 𝑇 ∗ , so the trivial intersection of
images in lemma 4.1.1 can be substituted with trivial intersection of the adjoint kernels.

Recall also that in a Banach space 𝐸, a closed subspace 𝑉 ⊆ 𝐸 is said to be complemented

if there is another another closed subspace 𝑊 ⊆ 𝐸, called the complement of 𝑉, satisfying
𝐸 = 𝑉 ⊕ 𝑊.

Proposition 4.1.3 (Biliotti, Javaloyes, and Piccione (2009), lemma 2.4). Let 𝐸, 𝐹, 𝐺 be Ba-
nach spaces, and 𝑇 : 𝐸 → 𝐺, 𝑆 : 𝐹 → 𝐺 bounded linear operators, with ker 𝑆 complemented in
𝐹 and Im 𝑆 finite codimensional in 𝐺 (this happens for example if 𝑆 is a Fredholm operator).
Then the kernel of 𝑇 ⊕ 𝑆 is complemented in 𝐸 × 𝐹.

Proof. One sees that Im 𝑇 ∩ Im 𝑆 ⊆ 𝐺 has finite codimension in Im 𝑇 : for the quotient projection
𝜋 : 𝐺 → 𝐺/Im 𝑆, the restriction 𝜋| Im 𝑇 : Im 𝑇 → 𝐺/Im 𝑆 has ker 𝜋| Im 𝑇 = Im 𝑇 ∩ Im 𝑆, thus
inducing an injective linear map

Im 𝑇 𝐺
[𝜋| Im 𝑇 ] : → ,
Im 𝑇 ∩ Im 𝑆 Im 𝑆
54

and 𝐺/Im 𝑆 is finite dimensional, therefore Im 𝑇 ∩ Im 𝑆 has finite codimension as a subspace of

Im 𝑇 .
Now, Im 𝑇 ∩ Im 𝑆 has a finite dimensional (therefore closed) algebraic complement 𝑍
in Im 𝑇, that is, (Im 𝑇 ∩ Im 𝑆) ⊕ 𝑍 = Im 𝑇 . Clearly ker 𝑇 has finite codimension in 𝑇 −1 (𝑍),
since the induced quotient of 𝑇 mapping [𝑇] : 𝑇 −1 (𝑍)/ker 𝑇 → 𝑍 is surjective. we then set
𝑇 −1 (𝑍) = ker 𝑇 ⊕ 𝑉, 𝑉 ⊆ 𝐸 finite dimensional. Lastly, ker 𝑆 has a closed complement 𝑊 in 𝐹
from our hypothesis.
We claim 𝑉 × 𝑊 to be the desired complement of ker (𝑇 ⊕ 𝑆). Take (𝑥, 𝑦) ∈ (𝑉 × 𝑊) ∩
ker(𝑇 ⊕ 𝑆), then 𝑇 (𝑥) + 𝑆(𝑦) = 0. With 𝑉 ⊆ 𝑇 −1 (𝑍), 𝑇 (𝑥) ∈ 𝑍, but also 𝑇 (𝑥) = −𝑆(𝑦) ∈ Im 𝑆,
and since (Im 𝑇 ∩ Im 𝑆) ∩ 𝑍 = {0}, it then follows 𝑇 (𝑥) = 𝑆(𝑦) =, so 𝑥 ∈ ker 𝑇 ∩ 𝑉 = {0} and
𝑦 ∈ ker 𝑆 ∩ 𝑊 = {0}, implying (𝑥, 𝑦) = 0 and the intersection (𝑉 × 𝑊) ∩ ker(𝑇 ⊕ 𝑆) is trivial.
To prove the sum (𝑉 × 𝑊) + ker(𝑇 ⊕ 𝑆) is all 𝐸 × 𝐹, for (𝑥, 𝑦) ∈ 𝐸 × 𝐹 arbitrary, write
𝑇 (𝑥) = 𝑇 (𝑒) + 𝑧, where 𝑒 ∈ 𝐸, 𝑇 (𝑒) ∈ Im 𝑆 and 𝑧 ∈ 𝑍. Since 𝑍 ⊆ Im 𝑇, by the decomposition
of 𝑇 −1 (𝑍) we see that 𝑧 = 𝑇 (𝑣) for some 𝑣 ∈ 𝑉, and then 𝑥 = 𝑒 + 𝑣 + 𝑎, for some 𝑎 ∈ ker 𝑇 .
Also, 𝑇 (𝑒) = 𝑆(𝑤) for some 𝑤 ∈ 𝑊 . Setting 𝑦 = 𝑏 + 𝑤 ′, where 𝑏 ∈ ker 𝑆 and 𝑤 ′ ∈ 𝑊 . Then, the
vectors (𝑒 + 𝑎, 𝑏 − 𝑤) ∈ ker(𝑇 ⊕ 𝑆) and (𝑣, 𝑤′ + 𝑤) ∈ 𝑉 × 𝑊 are such that

(𝑒 + 𝑎, 𝑏 − 𝑤) + (𝑣, 𝑤′ + 𝑤) = (𝑒 + 𝑣 + 𝑎, 𝑏 + 𝑤 ′) = (𝑥, 𝑦),

thus proving our claim. ■

Proposition 4.1.4 (Biliotti, Javaloyes, and Piccione (2009), lemma 2.3). Let 𝑈, 𝑉 be vector
spaces, 𝐿 : 𝑈 → 𝑉 a linear map, and 𝑆 ⊆ 𝑉 a finite codimensional subspace. Then
𝐿 −1 (𝑆) has finite codimension in 𝑈, with

codim𝑈 (𝐿 −1 (𝑆)) = codim𝑉 (𝑆) − codim𝑉 (𝑆 + Im 𝐿). (4.1)

Proof. Considering the quotient projection 𝜋 : 𝑉 → 𝑉/𝑆, the composition 𝜋 ◦ 𝐿 : 𝑈 → 𝑉/𝑆 has
kernel 𝐿 −1 (𝑆), so the induced quotient map [𝜋 ◦ 𝐿] : 𝑈/𝐿 −1 (𝑆) → 𝑉/𝑆 is injective, with proves
𝐿 −1 (𝑆) has finite codimension in 𝑈.
For (4.1), using standard dimensional analysis from linear algebra applied to the linear
map [𝜋 ◦ 𝐿] : 𝑈/𝐿 −1 (𝑆) → 𝑉/𝑆 we have

codim𝑉 𝑆 = dim (𝑉/𝑆) = dim (Im[𝜋 ◦ 𝐿]) + codim𝑉/𝑆 (Im[𝜋 ◦ 𝐿])

= dim (𝑈/𝐿 −1 (𝑆)) + codim𝑉 (Im (𝐿) + 𝑆). ■

For the abstract Banach manifold result we’ll need,we have announced, we fix a separable
Banach manifold 𝑋, a Hilbert manifold 𝑌 , and a Hilbert space 𝑉 . We apply the general functional-
analytic results above to prove the following proposition, an adaptation of Biliotti, Javaloyes, and
Piccione (2009), prop. 3.1 with more suitable hypotheses for our concrete case of interest.
 55

Proposition 4.1.5. Consider 𝐴 ⊆ 𝑋 × 𝑌 an open set, and let 𝑓 : 𝐴 → 𝑉 be a 𝐶 𝑘 function,

𝑘 ≥ 1. For every (𝑥 0 , 𝑦 0 ) ∈ 𝐴 such that 𝑓 (𝑥 0 , 𝑦 0 ) = 0, assume that the partial derivative
𝜕𝑓
(𝑥 0 , 𝑦 0 ) : 𝑇𝑦0 𝑌 → 𝑉 is a Fredholm operator. Then 0 is a regular value of 𝑓 if and only if
𝜕𝑦
(︃ )︃ ⊥ (︃ )︃ ⊥
𝜕𝑓 𝜕𝑓
Im (𝑥 0 , 𝑦 0 ) ∩ Im (𝑥 0 , 𝑦 0 ) = {0}. (4.2)
𝜕𝑥 𝜕𝑦

Proof. 0 is a regular value of 𝑓 if and only if for all (𝑥0 , 𝑦 0 ) ∈ 𝑓 −1 (0), 𝑑𝑓 (𝑥0 ,𝑦0 ) : 𝑇𝑥0 𝑋 ×𝑇𝑦0 𝑌 → 𝑉
is surjective with complemented kernel. Now 𝑑𝑓 (𝑥0 ,𝑦0 ) can be written as the sum operator

𝜕𝑓 𝜕𝑓
𝑑𝑓 (𝑥0 ,𝑦0 ) = (𝑥 0 , 𝑦 0 ) ⊕ (𝑥 0 , 𝑦 0 ).
𝜕𝑥 𝜕𝑦

𝜕𝑓
Applying lemma 4.1.1 since (𝑥 0 , 𝑦 0 ) is Fredholm, it follows that if 𝑑𝑓 (𝑥0 ,𝑦0 ) is surjective
𝜕𝑦
then (4.2) holds. Conversely if (4.2) holds, by lemma 4.1.1 we obtain surjectivity, and also by
proposition 4.1.3 the kernel of 𝑑𝑓 (𝑥0 ,𝑦0 ) is complemented. ■

Remark 4.1.6. For the sake of brevity, we will call the trivial intersection condition (4.2) as the
T condition. As stated in remark 4.1.2, we can substitute this condition for

𝜕𝑓∗
(︃ )︃ (︃ ∗ )︃
𝜕𝑓
ker (𝑥 0 , 𝑦 0 ) ∩ ker (𝑥0 , 𝑦 0 ) = {0}
𝜕𝑥 𝜕𝑦

at points (𝑥0 , 𝑦 0 ) ∈ 𝑓 −1 (0). This might be helpful in concrete situations as it directly states the
condition in the form of system of equations for the adjoints.

Under the hypothesis of proposition 4.1.5, assuming (4.2) to be valid, then the level set
𝑀= 𝑓 −1 {0}
⊆ 𝐴 is an embedded submanifold or 𝑋 ×𝑌 , with tangent space at some (𝑥0 , 𝑦 0 ) ∈ 𝑀
given by
{︃ }︃
𝜕𝑓 𝜕𝑓
𝑇(𝑥0 ,𝑦0 ) 𝑀 = (𝑣, 𝑤) ∈ 𝑇𝑥0 𝑋 × 𝑇𝑦0 𝑌 : (𝑥 0 , 𝑦 0 )(𝑣) + (𝑥0 , 𝑦 0 )(𝑤) = 0 .
𝜕𝑥 𝜕𝑦

The main abstract genericity result we’ll need is the following (adapted from Biliotti, Javaloyes,
and Piccione (2009, corollary 3.4)).
𝜕𝑓
Theorem 4.1.7. Under the hypothesis of proposition 4.1.5, assume also that is a Fredholm
𝜕𝑦
operator of index zero for points in 𝑀. Denoting by Π : 𝑋 × 𝑌 → 𝑋 the canonical projection,
then,

(i) Π| 𝑀 : 𝑀 → 𝑋 is Fredholm map of index zero.

56

(ii) The critical points of Π| 𝑀 are the points (𝑥0 , 𝑦 0 ) ∈ 𝑀 such that 𝑦 0 is a critical point of the
functional
𝑦 ∈ 𝐴𝑥0 ↦→ 𝑓 (𝑥 0 , 𝑦) ∈ 𝑉,

where 𝐴𝑥 = {𝑦 ∈ 𝑌 : (𝑥, 𝑦) ∈ 𝐴}.

(iii) Assuming 𝑋, 𝑌 and 𝑉 to be separable, the set of points 𝑥 ∈ 𝑋 such that the functional
𝑦 ∈ 𝐴𝑥 ↦→ 𝑓 (𝑥, 𝑦) ∈ 𝑉 has no critical points is generic in Π( 𝐴).

Proof. For (i), given (𝑥 0 , 𝑦 0 ) ∈ 𝑀, we have ker 𝑑Π|𝑇( 𝑥0 ,𝑦0 )𝑀 = 𝑇(𝑥0 ,𝑦0 ) 𝑀 ∩ ({0} × 𝑇𝑦0 𝑌 ), and this
(︂ )︂
vector space is isomorphic to ker 𝜕𝜕𝑦𝑓 (𝑥 0 , 𝑦 0 ) so it is finite dimensional. Also, one readily sees
that (︃ )︃ −1 (︃ )︃
𝜕𝑓 𝜕𝑓
Im 𝑑Π|𝑇( 𝑥0 ,𝑦0 ) 𝑀 = (𝑥 0 , 𝑦 0 ) Im (𝑥 0 , 𝑦 0 ) ,
𝜕𝑥 𝜕𝑦
so from proposition 4.1.4, Im 𝑑Π|𝑇( 𝑥0 ,𝑦0 ) 𝑀 is finite codimensional, and since 𝑑𝑓 (𝑥0 ,𝑦0 ) is surjective,
by eq. (4.1) (︃ )︃
𝜕𝑓
codim(Im 𝑑Π|𝑇( 𝑥0 ,𝑦0 ) 𝑀 ) = codim (𝑥 0 , 𝑦 0 ) .
𝜕𝑦

Since we are assuming 𝜕𝜕𝑦𝑓 (𝑥0 , 𝑦 0 ) to be Fredholm of index zero, it follows that 𝑑Π|𝑇( 𝑥0 ,𝑦0 )𝑀 is
also Fredholm of index zero.
For (ii), it is easy to see that (𝑥0 , 𝑦 0 ) ∈ 𝑀 is a regular point of Π| 𝑀 if and only if
(︃ )︃ (︃ )︃
𝜕𝑓 𝜕𝑓
Im (𝑥0 , 𝑦 0 ) ⊆ Im (𝑥0 , 𝑦 0 ) .
𝜕𝑥 𝜕𝑦

Taking orthogonal complements, this is equivalent to

(︃ )︃ ⊥ (︃ )︃ ⊥
𝜕𝑓 𝜕𝑓
Im (𝑥 0 , 𝑦 0 ) ⊆ Im (𝑥 0 , 𝑦 0 ) ,
𝜕𝑦 𝜕𝑥
(︂ )︂
𝜕𝑓
and by condition (4.2), equivalent to Im 𝜕𝑦 (𝑥 0 , 𝑦 0 ) be a zero codimensional subspace, in turn
(︂ )︂
equivalent to the triviality of ker 𝜕𝜕𝑦𝑓 (𝑥 0 , 𝑦 0 ) .
(︂ )︂
To summarize, (𝑥0 , 𝑦 0 ) ∈ 𝑀 is a regular point of Π| 𝑀 if and only if ker 𝜕𝜕𝑦𝑓 (𝑥 0 , 𝑦 0 ) = {0},
that is, if and only if 𝑦 0 is a regular point of the functional 𝑦 ∈ 𝐴𝑥0 ↦→ 𝑓 (𝑥 0 , 𝑦) ∈ 𝑉 . Item (iii)
now follows, because the set of points 𝑥 ∈ 𝑋 such that 𝑦 ∈ 𝐴𝑥 ↦→ 𝑓 (𝑥, 𝑦) ∈ 𝑉 has no critical
points coincides with the set of regular values for Π| 𝑀 , and genericity follows by the Sard-Smale
theorem (Smale (1965)). ■

Remark 4.1.8. We make here an important comment on the plausibility of the hypotheses in
theorem 4.1.7.
The condition 𝑇 together with the Fredholm condition is equivalent to 0 being a regular
value of a 𝐶 𝑘 function 𝑓 : 𝐴 ⊆ 𝑋 × 𝑌 → 𝑉 . Now, suppose 𝑋, 𝑌 , 𝑉 were all finite-dimensional.
 57

Then, the simplest version of elementary transversality theorem (corollary A.3.15) applied to the
submanifold 𝑊 = {0} ⊂ 𝑉 would mean the the set of 𝐶 𝑘 functions 𝑓 : 𝐴 → 𝑉 for which 0 is a
regular value is residual with respect to a Whitney topology. This suggests that the condition 𝑇
might not be restrictive.
While we expect to have an analogous situation for our infinite dimensional manifolds,
tranversality theory is far harder for the infinite dimensional case, and we were unable to find an
analogous Thom transversality theorem in the literature applied to our case. Therefore the 𝑇
condition will remain a technical condition for us.

4.2 Manifold Structures

The function we will concretely analyze in order to apply theorem 4.1.7 is the expansion
scalar 𝜃 + defined on a suitable open set (when manifold structures are defined) around a triple
(ℎ0 , K0 , 𝜓0 ), with (ℎ0 , K0 ) an initial data set for which 𝜓0 : Σ → S is a MOTS embedding.
With a domain well defined, the expansion scalar is a function of the form

𝜓
(ℎ, K, 𝜓) ↦→ 𝜃 + (ℎ, K, 𝜓) = trℎ K ◦ 𝜓 + 𝐻 ℎ,ν .

First,we describe possible infinite dimensional manifold structures on the set of vacuum
initial data, then discuss a Hilbert manifold structure for the set of embeddings Σ → S.
These infinite dimensional structures are very technical in nature, and describing them
precisely is out of the scope of this work. Fortunately, we shall need only very general feature of
these here, so we shall content ourselves to briefly summarize the ideas behind them. With this
structures well established, we then define the 𝜃 + on this “suitable open set”.

4.2.1 Manifold of Initial Data Sets

Denote by 𝐼 𝐷 (S) the set of vacuum initial data (ℎ, K). We set to establish a reasonable
structure for the initial data set 𝐼 𝐷 (S).
A technical condition that arises frequently in the study of initial data sets is the existence
of the so called Killing initial data (abbreviated to KID). For a initial data set (S, ℎ, K), given an
open set Ω ⊆ S, consider the following operator2 for 𝑌 a vector field and 𝑁 a function over Ω,
(︁ )︁
2 ∇ (𝑖𝑌 𝑗) − ∇𝑙 𝑌𝑙 ℎ𝑖 𝑗 − K𝑖 𝑗 𝑁 + tr K 𝑁 ℎ𝑖 𝑗
𝑃∗ (𝑌 , 𝑁) = ⎜⎜
⎛ ⎞
∇𝑙 𝑌𝑙 K𝑖 𝑗 − 2K 𝑙 (𝑖 ∇ 𝑗)𝑌𝑙 + K 𝑞 𝑙 ∇𝑞𝑌 𝑙 ℎ𝑖 𝑗 − Δ𝑁 ℎ𝑖 𝑗 + ∇𝑖 ∇ 𝑗 𝑁 ⎟.
⎟
(︁ )︁
⎝ + ∇ 𝑝 K𝑙 𝑝 ℎ𝑖 𝑗 − ∇𝑙 K𝑖 𝑗 𝑌 𝑙 − 𝑁 Ric(ℎ)𝑖 𝑗 + 2𝑁K𝑖𝑙 𝐾 𝑗𝑙 − 2𝑁 (tr K)K𝑖 𝑗 ⎠

The equations 𝑃∗ (𝑌 , 𝑁) = 0 are the called the vacuum KID equations on Ω. A KID on Ω is

2This
equation derived as the formal adjoint of the linearization for the constraint equation (cf. Chruściel and Delay
(2003), section 2).
58

then defined as the pair (𝑌 , 𝑁) on Ω that is a nontrivial solution of 𝑃∗ (𝑌 , 𝑁) = 0. If there are no

nontrivial solutions of 𝑃∗ (𝑌 , 𝑁) = 0, the initial data set (S, ℎ, K) satisfies the no initial Killing
data (no KIDs condition) on Ω. As shown by Moncrief (1975), solutions to the vacuum KID
equations are in one-to-one correspondence with Killing vector fields in the maximal Cauchy
development of the initial data.
Since we want conditions on S as to apply theorem 1.6.3, S cannot be compact. For
a concrete manifold structure under such conditions we will settle for S to be a non-compact,
oriented asymptotically flat manifold without boundary. Asymptotically flatness here means that
S has a compact set 𝐾 such that S \ 𝐾 is the union of a finite number of regions S1 , . . . , Sℓ
called the ends of S, each diffeomorphic to R𝑛 \ 𝐵, where 𝐵 denotes the closed ball of radius 1
centered at zero3 .
For the sake of simplicity we here assume S has only one end. Then, under these
conditions we find a few options in literature, but for our practical purposes we choose the smooth
Hilbert manifold structure laid out by Bartnik (2005). This structure4 is locally modeled on
Sobolev spaces 𝐻 2 × 𝐻 1 , therefore is separable.

4.2.2 Manifold of Embeddings

Infinite dimensional structures for functions spaces Σ → S with Σ compact is a well-
established idea in the literature. Here we follow Alias and Piccione (2011), section 3.
The idea of the process is to choose a “well suitable” class of regularity R less regular
than 𝐶 regularity, that densely contains 𝐶 ∞ and can be continuously embedded into those
∞

of 𝐶 2,𝛼 regularity (Alias and Piccione (2011), section 2), then model R (Σ, S) as a Banach or
Hilbert manifold of such regularity, and finally argue that the set embeddings 𝐸𝑚𝑏(Σ, S) is open
in R (Σ, S).
Here we choose Sobolev 𝐻 𝑘+2 regularity (a Hilbert space), with 𝑘, 𝑟 ∈ N and 𝛼 ∈ (0, 1)
(we take 𝑟 ≥ 2 since at least two derivatives are computed in curvature tensors) satisfying
(𝑘 + 2 − 𝑟 − 𝛼)/𝑛 > 1/2 so that the Sobolev space 𝐻 𝑘+2 is continuously embedded into the Hölder
space 𝐶 𝑟,𝛼 (Aubin (1998), thms. 2.10 and 2.20), so even if it seems we are losing too much
regularity by locally modeling the embeddings over Sobolev spaces, these can be identified as to
have enough regularity. More precisely, Alias and Piccione (2011) show that 𝐸𝑚𝑏(Σ, S) in this
regularity can be viewed as a Hilbert manifold locally modeled5 on the Hilbert space 𝐻 𝑘+2 (Σ).

3The precise definition of asymptotic flatness also involves detailed falloff conditions for the metric and its derivatives

up to second order, that will need not concern us here.

4 Originally we wanted to model using a Banach manifold structure following Chruściel and Delay (2003), since
their work is not limited to the asymptotically flat case. However their Banach structure is of the form 𝐻 𝑙 ∩ 𝐶 𝑟 , 𝛼
(both Sobolev and Hölder) using a Sobolev-Hölder sum norm, and it is unclear to us if the Banach manifold locally
modeled on this space is separable, as this condition is crucial to apply the Sard-Smale theorem.
5 Integrals on Σ are computed with respect to some background Riemannian metric ℎ, but the resulting Sobolev

space does not depend on this choice. See Aubin (1998) for details.
 59

4.2.3 Null Expansion Scalar Function 𝜃 +

To make technical sense of the expansion scalar as a smooth function on the set of initial
data sets and two-sided, framed embeedings, it requires a chosen outward normal direction a
priori, therefore we can only work on a neighborhood of a fiducial (ℎ0 , K0 , 𝜓0 ).
To be more precise, consider (ℎ0 , K0 ) ∈ 𝐼 𝐷 (S) an asymptotically flat initial data set
satisfying the no KIDs condition, and 𝜓0 ∈ 𝐸𝑚𝑏(Σ, S) a smooth embedding that is a MOTS
with respect to initial data set (ℎ0 , K0 ), with ν0 ∈ 𝔛 ⊥ (𝜓0 ) the chosen outward pointing unit
normal vector field of the embedding 𝜓0 under ℎ0 . Then

𝜓
𝜃 + (ℎ0 , K0 , 𝜓0 ) = trℎ0 K0 ◦ 𝜓0 + 𝐻 ℎ00,ν0 = 0.

The process is now to “propagate” the normal vector ν0 over a small neighborhood around
𝜓0 (Σ). Since 𝜓0 (Σ) is an smooth compact submanifold of S, there is a smooth vector field
V ∈ 𝔛(S) such that V ◦ 𝜓0 = ν0 . Restricting to a small enough neighborhood around 𝜓0 (Σ),
we can assume V𝑝 ∉ (𝑑𝜓0 ) 𝑝 (𝑇𝑝 Σ) at all points in this neighborhood. We can then choose a
open neighborhood around our embedding 𝜓0 in 𝐸𝑚𝑏(Σ, S) such that all embeddings 𝜓 in this
neighborhood have V ◦ 𝜓 nowhere tangent. (These embeddings are called framed embeddings,
and we see via this argument that this is a stable property). With initial data (ℎ, K) close enough
to (ℎ0 , K0 ), we consider the ℎ-normal part of V ◦ 𝜓 :

(V ◦ 𝜓) ⊥,ℎ
ν(ℎ, 𝜓) = .
∥(V ◦ 𝜓) ⊥,ℎ ∥ ℎ

Denote by A0 the open set in 𝐼 𝐷 (S) × 𝐸𝑚𝑏(Σ, S) around (ℎ0 , K0 , 𝜓0 ) as describe above. Then
we can define
𝜓
𝜃 + : (ℎ, K, 𝜓) ∈ A0 ↦→ trℎ K ◦ 𝜓 + 𝐻 ℎ,ν(ℎ,𝜓) , (4.3)

4.2.4 Linearization of 𝜃 +
With 𝜃 + suitably defined, the formal partial linearization of 𝜃 + with respect to the
embedding is the MOTS stability operator (definition 1.5.5): if 𝜓 : Σ → S is a MOTS
embedding for the initial data (ℎ, K),
|︁
𝜕𝜃 + |︁|︁
( 𝑓 ) = 𝐿 ( 𝑓 ) = −Δℎ 𝑓 + 2⟨𝑋, gradℎ 𝑓 ⟩ℎ + (𝑄 + divℎ 𝑋 − ∥ 𝑋 ∥ 2ℎ ) 𝑓 ,
𝜕𝜓 |︁ (ℎ,K,𝜓)

where
1 1
𝑄 = ScalΣ − [𝐽 (ν) + 𝜌] − |Kν + K | 2 ,
2 2
and 𝑋 the vector field ℎ-dual to the one form K (ν, ·)
An issue that needs addressing in other to apply the analytical machinery developed is
60

the Fredholm property for the MOTS stability operator. However, since 𝐿 is a second order
linear eliptic operator with a specific form, we have the following general result for the Fredholm
propriety of such operators.
Proposition 4.2.1 (Dan A. Lee (2019), coro. A.9). Let (Σ, ℎ) be a compact Riemannian mani-
fold without boundary, and consider 𝐿 a second order linear elliptic operator on (Σ, ℎ) of the
form
𝐿( 𝑓 ) = −Δℎ 𝑓 + ⟨𝑉, 𝑔𝑟𝑎𝑑 ℎ 𝑓 ⟩ℎ + 𝑞 𝑓 ,

with 𝑉 ∈ 𝔛(𝑀), 𝑞 ∈ 𝐶 ∞ (Σ). Then

𝐿 : 𝑊 𝑘+2,𝑝 (Σ) → 𝑊 𝑘,𝑝 (Σ),

𝐿 : 𝐶 𝑘+2,𝛼 (Σ) → 𝐶 𝑘,𝛼 (Σ)

are Fredholm operators of index zero for each 𝑘 ∈ N. ■

For completeness, we can also compute the formal partial linearization of 𝜃 + with respect
to the variable of initial data: given (𝑏, Q) two (0, 2)-symmetric tensors over S, for simplicity
we calculate this assuming Σ is embedded in S (𝑖 : Σ ↩→ S is the embedding; for the general
case we will have to carry objects over the map 𝜓 as in section 1.1.2).
For a perturbation ℎ + 𝛿𝑏 of the metric ℎ, the first order expansion of the inverse is

(ℎ + 𝛿𝑏) 𝑖 𝑗 = ℎ𝑖 𝑗 − 𝛿ℎ𝑖𝑘 𝑏 𝑘𝑙 ℎ𝑙 𝑗 + 𝑜(𝛿2 ).

We first analyze the trace term of 𝜃 + in (4.3). A perturbation (ℎ + 𝛿𝑏, K + 𝛿Q) has first order
expansion
trℎ+𝛿𝑏 (K + 𝛿Q) = trℎ+𝛿𝑏 (K) + 𝛿 trℎ+𝛿𝑏 (Q)
= trℎ K − 𝛿ℎ𝑖𝑘 𝑏 𝑘𝑙 ℎ𝑙 𝑗 K𝑖 𝑗 + 𝛿 trℎ Q + 𝑜(𝛿2 ),
subtracting trℎ K, dividing by 𝛿, and taking the limit 𝛿 → 0,

trℎ+𝛿𝑏 (K + 𝛿Q) − trℎ K

lim = trℎ Q − ℎ𝑖𝑘 𝑏 𝑘𝑙 ℎ𝑙 𝑗 K𝑖 𝑗 , (4.4)
𝛿→0 𝛿

and ℎ𝑖𝑘 𝑏 𝑘𝑙 ℎ𝑙 𝑗 K𝑖 𝑗 is the ℎ-inner product of the two tensors 𝑏 and K, which we denote by ⟨𝑏, K⟩ℎ .
The second term in 𝜃 + is a bit trickier. Since Σ has a normal unitary vector field ν,
that is dependent on the metric ℎ. Denote by ν(𝛿) = ν(ℎ + 𝛿𝑏), with a first order expansion
ν(𝛿) = ν + 𝛿𝑉 + 𝑜(𝛿2 ). The scalar mean curvature on the direction ν(𝛿) has the form (here we
denote ∇𝑖𝛿𝑗 = ∇𝜕ℎ+𝛿𝑏𝑖
𝜕𝑗 , )

𝐻 ℎ+𝛿𝑏,ν(𝛿) = (ℎ + 𝛿𝑏) 𝑖 𝑗 (ℎ + 𝛿𝑏)(∇𝑖𝛿𝑗 , ν(𝛿))

= ℎ𝑖 𝑗 ℎ(∇𝑖𝛿𝑗 , ν(𝛿)) + 𝛿ℎ𝑖 𝑗 𝑏(∇𝑖𝛿𝑗 , ν(𝛿)) − 𝛿ℎ𝑖𝑘 𝑏 𝑘𝑙 ℎ𝑙 𝑗 ℎ(∇𝑖𝛿𝑗 , ν(𝛿)) + 𝑜(𝛿2 )

First we expand the Christoffel symbols (Γ𝑖𝑘𝑗 (𝛿) denotes the perturbed symbol, Γ𝑖𝑘𝑗 corresponds
 61

to ℎ):
𝛿 (𝑏)
Γ𝑖𝑘𝑗 (𝛿) = Γ𝑖𝑘𝑗 + ℎ 𝑘𝑚 𝛾𝑚𝑖 + 𝑜(𝛿2 ),
2 𝑗

(𝑏) (𝑏)
where 𝛾𝑚𝑖 𝑗 = 𝑏 𝑚𝑖; 𝑗 + 𝑏 𝑚 𝑗;𝑖 − 𝑏 𝑖 𝑗;𝑚 (covariant derivative with respect to ℎ), and we denote by 𝑇
(𝑏)
the (1,2)-tensor over S with components ℎ 𝑘𝑚 𝛾𝑚𝑖 𝑗 . The expansion of ∇𝑖 𝑗 is then
𝛿

𝛿
∇𝑖𝛿𝑗 = ∇𝑖 𝑗 + 𝑇𝑖(𝑏) + 𝑜(𝛿2 ),
2 𝑗

and the perturbed scalar mean curvature is

𝛿
𝐻 ℎ+𝛿𝑏,ν(𝛿) = ℎ𝑖 𝑗 ℎ(∇𝑖 𝑗 , ν(𝛿)) + ℎ𝑖 𝑗 ℎ(𝑇𝑖(𝑏)
𝑗 , ν(𝛿))
2
+ 𝛿ℎ𝑖 𝑗 𝑏(∇𝑖 𝑗 , ν(𝛿)) − 𝛿ℎ𝑖𝑘 𝑏 𝑘𝑙 ℎ𝑙 𝑗 ℎ(∇𝑖 𝑗 , ν(𝛿)) + 𝑜(𝛿2 ),

Expanding ν(𝛿) as well

𝛿
𝐻 ℎ+𝛿𝑏,ν(𝛿) = 𝐻 ℎ,ν + 𝛿ℎ𝑖 𝑗 ℎ(∇𝑖 𝑗 , 𝑉) + ℎ(trℎ 𝑇 (𝑏) , ν)
2
+ 𝛿ℎ 𝑏(∇𝑖 𝑗 , ν) − 𝛿ℎ 𝑏 𝑘𝑙 ℎ𝑙 𝑗 ℎ(∇𝑖 𝑗 , ν) + 𝑜(𝛿2 ).
𝑖𝑗 𝑖𝑘

The term ℎ𝑖𝑘 𝑏 𝑘𝑙 ℎ𝑙 𝑗 ℎ(∇𝑖 𝑗 , ν) is ⟨K𝜈ℎ , 𝑏⟩ℎ , where K𝜈ℎ is the scalar second fundamental form of
Σ in the direction ν. Observe also that, since (ℎ + 𝛿𝑏)(ν(𝛿), ν(𝛿)) = 1, we have ℎ(ν, 𝑉) =
−1/2𝑏(ν, ν) + 𝑜(𝛿). Applying this, and also observing that ℎ𝑖 𝑗 ∇𝑖 𝑗 = 𝐻 ℎ,ν ν + ℎ𝑖 𝑗 ∇𝑖⊤𝑗 (the last
term denoting the tangent part to Σ), and also decomposing 𝑉 = 𝑉 ⊤ + 𝑉 ⊥,ℎ , we finally obtain

𝛿 𝛿
𝐻 ℎ+𝛿𝑏,ν(𝛿) = 𝐻 ℎ,ν − 𝛿⟨K𝜈ℎ , 𝑏⟩ℎ + ℎ(trℎ 𝑇 (𝑏) , ν) + 𝐻 ℎ,ν 𝑏(ν, ν)
2 2
⊤ ⊤ ⊤
+ 𝛿ℎ ℎ(∇𝑖 𝑗 , 𝑉 ) + 𝛿ℎ 𝑏(∇𝑖 𝑗 , 𝜈) + 𝑜(𝛿2 )
𝑖𝑗 𝑖𝑗

so the linearization is
𝐻 ℎ+𝛿𝑏,ν(𝛿) − 𝐻 ℎ,ν 1 1
lim = − ⟨K𝜈ℎ , 𝑏⟩ℎ + ℎ(trℎ 𝑇 (𝑏) , ν) + 𝐻 ℎ,ν 𝑏(ν, ν)
𝛿→0 𝛿 2 2 (4.5)
𝑖𝑗 ⊤ ⊤ 𝑖𝑗 ⊤
+ ℎ ℎ(∇𝑖 𝑗 , 𝑉 ) + ℎ 𝑏(∇𝑖 𝑗 , 𝜈).

Adding (4.4) and (4.5),

|︁
𝜕𝜃 + |︁|︁ ℎ 1 (𝑏) 1
(𝑏, Q) = tr ℎ Q − ⟨𝑏, K⟩ ℎ − ⟨K 𝜈 , 𝑏⟩ ℎ + ℎ(tr ℎ 𝑇 , ν) + 𝐻 ℎ,ν 𝑏(ν, ν)
𝜕 (ℎ, K) |︁ (ℎ,K,𝑖) 2 2
+ ℎ𝑖 𝑗 ℎ(∇𝑖⊤𝑗 , 𝑉 ⊤ ) + ℎ𝑖 𝑗 𝑏(∇𝑖⊤𝑗 , 𝜈).
62

4.3 Geometric Interpretation

With the abstract analytical machinery in place, a suitable infinite dimensional manifold
structure and 𝜃 + all well established, we now give our main results.
To simplify our language, let us introduce some terminology: say that a triple (ℎ, K, 𝜓)
is a MOTS triple if 𝜃 + (ℎ, K, 𝜓) = 0 (i.e. the embedding 𝜓 is a MOTS in the initial data (ℎ, K)),
and call a MOTS triple (ℎ, K, 𝜓) non-degenerate if the associated MOTS stability operator has a
non-zero principal eigenvalue (in such case the stability operator is said to be non-degenerate).

Theorem 4.3.1. Let (ℎ0 , K0 ) be initial data set satisfying the no KIDs condition, 𝜓0 : Σ → S
an embedding that is a MOTS for the initial data (ℎ0 , K0 ) on S, and consider the neighborhood
A0 where 𝜃 + is defined. Assume 𝜃 + satisfies the T condition (remark 4.1.6). Then there exists an
open neighborhood O ⊆ 𝐼 𝐷 (S) around (ℎ0 , K0 ) where the set

G = {(ℎ, K) ∈ O : every MOTS in (ℎ, K) are non-degenerate} (4.6)

is a residual set in O.

Proof. Since the MOTS stability operator is a Fredholm operator of index 0, we are in conditions
to apply theorem 4.1.7: here, 𝑀 = 𝜃 +−1 (0) is the submanifold in A0 of MOTS triples. By item
(iii), it is a generic property in Π(A0 ) = O the following: initial data (ℎ, K) such that the
functional
𝜓 ∈ ( 𝐴0 ) (ℎ,K) ↦→ 𝜃 + (ℎ, K, 𝜓)

has no critical points. Having no critical points, then any MOTS embedding 𝜓 for such (ℎ, K) is
such that the associated stabability operator is non-degenerate. ■

Remark 4.3.2. What theorem 4.3.1 says about the residual set G ⊆ O is that if there is a MOTS
in the initial data (ℎ, K) ∈ G, then this MOTS is non-degenerate. However our method cannot
guarantee that such a MOTS exists for initial data in G. In order to obtain this last condition,
we have to make extra assumptions about the fiducial MOTS triple (ℎ0 , K0 , 𝜓0 ). We cite two
distinct situations where this is possible:

(1) Assume that (ℎ0 , K0 , 𝜓0 ) is itself non-degenerate. Then by condition 𝑇 the partial derivative
with respect to the embeddings variable at (ℎ0 , K0 , 𝜓0 ) is an isomorphism (this can be more
clearly seen from the abstract case in theorem 4.1.7). Then by a straightforward application
of the implicit function theorem in Banach spaces (cf. Abraham, Marsden, and Ratiu
(2012), thm. 2.5.7) We can reduce A0 to a open set of the form U0 × V0 , (ℎ0 , K0 ) ∈ U0 ,
𝜓0 ∈ V0 , where now all initial data in O = 𝑈0 have a non-degenerate MOTS, so when the
fiducial MOTS triple is non-degenerate we have obtained that existence of non-degenerate
MOTS for initial data in U0 is stable. (Actually, this argument bypasses the generic step
(iii) in theorem 4.1.7 thanks to condition 𝑇 and the implicit function theorem.)
 63

(2) For this second case, we restrict (ℎ0 , K0 ) and the dimension and topology of S to the
following theorem due to Andersson, Eichmair, and Metzger (2011):

Theorem 4.3.3 (Andersson, Eichmair, and Metzger (2011), thm. 3.3). Let (ℎ, K) be
a initial data set for S, where 3 ≤ dim S ≤ 7. Assume that there is a connected bounded
open set Ω ⊆ S with smooth embedded boundary 𝜕Ω. Assume this boundary consists of two
non-empty closed hypersurfaces 𝜕+ Ω and 𝜕− Ω, possibly consisting of several components,
so that
𝐻𝜕+ Ω − tr𝜕+ Ω K > 0 𝑎𝑛𝑑 𝐻𝜕− Ω + tr𝜕− Ω K > 0, (4.7)

where the mean curvature scalar is computed as the tangential divergence of the unit
normal vector field that is pointing out of Ω. Then there exists a smooth closed embedded
MOTS6 Σ ⊆ Ω for the initial data (ℎ, K) that is homologous to 𝜕+ Ω. ■

By restricting the topology and dimension of S as in theorem 4.3.3, consider an initial data
(ℎ0 , K0 ) for S, and then a MOTS 𝜓0 for this initial data is obtained by the latter theorem.
This MOTS triple (ℎ0 , K0 , 𝜓0 ) is now our fiducial MOTS triple, that may or may not be
non-degenerated. The point is that conditions (4.7) are open, so there is an open set U0
around (ℎ0 , K0 ) where all initial data in U0 have a MOTS by theorem 4.3.3. Applying
theorem 4.3.1 under these conditions, the residual set G will have the property that all
initial data (ℎ, K) in G have a non-degenerate MOTS.

Now let us organize theorem 4.3.1 together with conditions laid out in remark 4.3.2 as to
obtain a generic condition for incompleteness.
Consider a MOTS triple (ℎ0 , K0 , 𝜓0 ) and its open neighborhood A0 where 𝜃 + is defined,
assume now that 𝜓0 (Σ) separates S as defined section 1.6.1. Reducing A0 to a smaller open
set we can assume that all embeddings separate S. By theorem 4.3.1 there is an open set O
around (ℎ0 , K0 ) where the property (4.6) is generic in O. If (ℎ, K) ∈ G has a MOTS, this
MOTS is non-degenerate, therefore we are in conditions to apply theorem 1.6.3, and MOTS for
(ℎ, K) can be deformed to be outer-trapped, then the maximal Cauchy development of (ℎ, K)
(cf. theorem 1.5.3) is null incomplete.
Following remark 4.3.2, by suitably reducing A0 , under condition (1) null imcompleteness
is now stable around (ℎ0 , K0 ), and under condition (2) it is generic in an open subset around
(ℎ0 , K0 ). We summarize this in the following theorem:

Theorem 4.3.4. Let (ℎ0 , K0 ) be initial data set satisfying the no KIDs condition, 𝜓0 : Σ → S
an embedding that is a MOTS for the initial data (ℎ0 , K0 ), where 𝜓0 (Σ) separates S. Consider
the open set A0 around (ℎ0 , K0 , 𝜓0 ) where 𝜃 + is defined and all embeddings separates S, and
also assume that 𝜃 + satisfies the 𝑇 condition. Under these assumptions,

6 Otherproperties of Σ are obtained in Andersson, Eichmair, and Metzger (2011), but we need not concerned about
them here.
64

(1) if (ℎ0 , K0 , 𝜓0 ) is itself non-degenerate, then there is an open set O around (ℎ0 , K0 ) where
all initial data (ℎ, K) ∈ O are such that their maximal Cauchy development of (ℎ, K) is
null incomplete (null incompleteness is stable);

(2) if (ℎ0 , K0 , 𝜓0 ) is a MOTS triple originating from theorem 4.3.3 (where the topology and
dimensions of S have to be further restricted) then there is an open set O around (ℎ0 , K0 )
and a residual set G ⊆ O where all initial data (ℎ, K) ∈ G are such that their maximal
Cauchy development of (ℎ, K) is null incomplete (null incompleteness is generic). ■
 65

A Whitney Topologies

In this appendix we review the elements of differential topology pertinent to this thesis
on jet bundles and Whitney topologies and the associated Baire proprieties of such topologies
(for more on this subject see, e.g., Hirsch (1976) or Mukherjee (2015) for a comprehensive
introduction to Whitney topologies, and Mather (1969) for more specific technical results).

A.1 Notation
We establish basic notation that recurs while discussing jets of functions. If 𝑓 : 𝑈 ⊆
R → R𝑛 is a function of class 𝐶 𝑟 , we denote by 𝐷 𝑘 𝑓 𝑝 , 𝑘 ≤ 𝑟 its 𝑘th order derivative at a point
𝑚

𝑝 ∈ 𝑈. This can be identified with a symmetric 𝑘-multilinear mapping. We denote the set of
symmetric 𝑘-multilinear mappings from R𝑛 to R𝑚 by 𝑆 𝑘 (𝑚, 𝑛). In particular, 𝑆 1 (𝑚, 𝑛) is the set
of linear mappings R𝑚 → R𝑛 .
With 𝑘-multilinear mappings we can define polynomials. Given 𝐴 ∈ 𝑆 𝑙 (𝑚, 𝑛), 𝑥 ∈ R𝑚 ,
we denote by 𝐴(𝑥) 𝑙 = 𝐴(𝑥, . . . , 𝑥) (the same vector 𝑥 repeated 𝑙 times as an argument of 𝐴). For
𝐴 𝑗 ∈ 𝑆 𝑗 (𝑚, 𝑛), 𝑗 ≤ 𝑘 and 𝑦 0 ∈ R𝑛 , a polynomial of several variables of degree 𝑘 is a mapping
𝑝 : R𝑚 → R𝑛 defined by

𝑝(𝑥) = 𝑦 0 + 𝐴1 (𝑥) + 𝐴2 (𝑥) 2 + · · · + 𝐴 𝑘 (𝑥) 𝑘 .

To define jets it is convenient to work with polynomials without the constant term (𝑦 0 = 0). We
denote the set of such 𝑘-degree polynomials by 𝑃 𝑘 (𝑚, 𝑛). There is a natural identification

𝑃 𝑘 (𝑚, 𝑛) ≅ 𝑆 1 (𝑚, 𝑛) × 𝑆 2 (𝑚, 𝑛) × · · · × 𝑆 𝑘 (𝑚, 𝑛).

To simplify the notation of polynomials in several variables and multiple partial derivatives,
we introduce the idea of 𝑚-multi-indices which is an 𝑚-tuple of positive integers 𝛼 = (𝛼1 , . . . , 𝛼𝑚 ).
We write |𝛼| = 𝛼1 + · · · + 𝛼𝑚 , and for any 𝑥 = (𝑥 1 , . . . , 𝑥 𝑚 ) ∈ R𝑚 we use the notation
𝛼𝑚
𝑥 𝛼 = 𝑥 1𝛼1 . . . 𝑥 𝑚 . A partial derivative of order |𝛼| ≤ 𝑟 for the function 𝑓 is denoted by

𝜕𝛼 𝑓 𝜕 |𝛼| 𝑓
= 𝛼𝑚 .
𝜕𝑥 𝛼 𝜕𝑥1𝛼1 · · · 𝜕𝑥 𝑚

A.2 Space of Jets

Here we present a brief summary of the construction of space of 𝑟-jets as a smooth fiber
bundle; the main reference here is Mukherjee (2015) (see chapter 8 for the details and omitted
computations).
66

Let 𝑀 and 𝑁 be smooth manifolds of dimensions 𝑚 and 𝑛, respectively, both with

empty boundary1 . Given 𝑝 ∈ 𝑀 and 𝑟 ∈ N, consider the following relation between functions
𝑝
𝑓 , 𝑔 ∈ 𝐶 ∞ (𝑀, 𝑁): 𝑓 ∼ 𝑔 if
𝑟

• 𝑓 ( 𝑝) = 𝑔( 𝑝);

• Given local charts (𝑈, 𝜑) and (𝑉, 𝜓) around 𝑝 and 𝑓 ( 𝑝), respectively, such that 𝑓 (𝑈), 𝑔(𝑈) ⊆
𝑉, and for any 𝑘 ∈ N with 𝑘 ≤ 𝑟, one has

𝐷 𝑘 (𝜓 ◦ 𝑓 ◦ 𝜑−1 ) 𝜑( 𝑝) = 𝐷 𝑘 (𝜓 ◦ 𝑔 ◦ 𝜑−1 ) 𝜑( 𝑝) .

Using standard multivariable calculus arguments, it can be shown that this relation does
not depend on the choice of local charts, and therefore it is well-defined. It is also immediately
observed that this relation is an equivalence relation.
𝑝
Remark A.2.1. The last item in the definition of ∼ is equivalent to the following statement: for
𝑟
all 𝑚 multi-indice 𝛼 with |𝛼| ≤ 𝑟,

𝜕 𝛼 (𝜓 ◦ 𝑓 ◦ 𝜑−1 ) 𝜕 𝛼 (𝜓 ◦ 𝑔 ◦ 𝜑−1 )
(𝜑( 𝑝)) = (𝜑( 𝑝)).
𝜕𝑥 𝛼 𝜕𝑥 𝛼
𝑝
The equivalence class of a function 𝑓 ∈ 𝐶 ∞ (𝑀, 𝑁) under the relation ∼ is denoted by
𝑟
𝑗 𝑟𝑝 𝑓 and called the 𝑟-jet of 𝑓 at the point 𝑝. The set of all 𝑟-jets at the point 𝑝 is denoted by
𝐽 𝑟𝑝 (𝑀, 𝑁). Given 𝑞 ∈ 𝑁, the set of all 𝑟-jets 𝑗 𝑟𝑝 𝑓 with 𝑓 ( 𝑝) = 𝑞 is denoted by 𝐽 𝑟𝑝 (𝑀, 𝑁) 𝑞 . The
space of 𝑟-jets,denoted by 𝐽 𝑟 (𝑀, 𝑁), is the disjoint union
⨆︂
𝐽 𝑟 (𝑀, 𝑁) = 𝐽 𝑟𝑝 (𝑀, 𝑁) 𝑞 .
( 𝑝,𝑞)∈𝑀×𝑁

We establish the following terminology in jet theory:

Given 𝑗 𝑟𝑝 𝑓 ∈ 𝐽 𝑟 (𝑀, 𝑁), the point 𝑝 ∈ 𝑀 is called the source of the jet, and the point
𝑞 = 𝑓 ( 𝑝) is called the target of the jet. The functions

𝜎 : 𝐽 𝑟 (𝑀, 𝑁) → 𝑀 𝜏 : 𝐽 𝑟 (𝑀, 𝑁) → 𝑁
𝑗 𝑟𝑝 𝑓 ↦→ 𝑝, 𝑗 𝑟𝑝 𝑓 ↦→ 𝑓 ( 𝑝),

are called, respectively, the source and target maps, and are clearly surjective. The 𝑟-jet
prolongation for 𝑓 ∈ 𝐶 ∞ (𝑀, 𝑁) is the map

𝑗 𝑟 𝑓 : 𝑀 → 𝐽 𝑟 (𝑀, 𝑁)
𝑝 ↦→ 𝑗 𝑟𝑝 𝑓 .
1The space of jets can be developed in a similar way for manifolds with boundary, but then the resulting manifold
will have corners, as developed in Michor (1980), chapter 2, or Margalef-Roig and Outerelo Domínguez (1992),
chapter 1. Since our primary interest are manifolds without boundary, we will avoid such complications.
 67

Jets as Fiber Bundles

The set 𝐽 𝑟 (𝑀, 𝑁) is endowed with the structure of a manifold and, moreover, of a
fiber bundle, in the following way: given non-empty open sets 𝑈 ⊆ 𝑀 and 𝑉 ⊆ 𝑁, denote
by 𝐽 𝑟 (𝑈, 𝑉) ⊆ 𝐽 𝑟 (𝑀, 𝑁) the subset of 𝑟-jets 𝑗 𝑟𝑝 𝑓 of smooth functions 𝑓 : 𝑈 → 𝑉 . Consider
charts (𝑈, 𝜑) and (𝑉, 𝜓) of 𝑀 and 𝑁, respectively. It can be verified that the function
ℎ𝑈,𝑉 : 𝐽 𝑟 (𝑈, 𝑉) → 𝜑(𝑈) × 𝜓(𝑉) × 𝑃𝑟 (𝑚, 𝑛) defined by

ℎ𝑈,𝑉 ( 𝑗 𝑟𝑝 𝑓 ) = (𝜑( 𝑝), 𝜓( 𝑓 ( 𝑝)), 𝐷 (𝜓 ◦ 𝑓 ◦ 𝜑−1 ) 𝜑( 𝑝) ), . . . , 𝐷 𝑟 (𝜓 ◦ 𝑓 ◦ 𝜑−1 ) 𝜑( 𝑝) ))

is a bijection (Mukherjee (2015), lemma 8.1.3). For a collection of charts {(𝑈𝑖 , 𝜑𝑖 )} and
{(𝑉𝛼 , 𝜓𝛼 )} of 𝑀 and 𝑁 that cover these manifolds, the collection of sets 𝐽 𝑟 (𝑈𝑖 , 𝑉𝛼 ) covers
𝐽 𝑟 (𝑀, 𝑁). A more extensive analysis (Mukherjee (2015), lemma 8.6.1 and theorem 8.6.2) shows
that the transition maps ℎ𝑈𝑖 ,𝑉𝛼 ◦ ℎ𝑈 −1 are smooth diffeomorphisms. Thus, by means of usual
𝑗 ,𝑉𝛽
results on the construction of smooth manifolds (for example, John M. Lee (2012), lemma
1.35), we obtain a smooth manifold structure for 𝐽 𝑟 (𝑀, 𝑁), the collection {(𝐽 𝑟 (𝑈𝑖 , 𝑉𝛼 ), ℎ𝑈𝑖 ,𝑉𝛼 )}
thus being a collection of smooth charts that covers 𝐽 𝑟 (𝑀, 𝑁). With this structure, a simple
computation in coordinates (Mukherjee (2015), theorem 8.6.2) shows that the source 𝜎 and target
𝜏 are smooth submersions, and also that the prolongation 𝑗 𝑟 𝑓 is a smooth embedding.
Such charts, together with the source and target maps, induce a fiber bundle structure on the
manifold 𝐽 𝑟 (𝑀, 𝑁) in the following way: consider the projection 𝜋 : 𝐽 𝑟 (𝑀, 𝑁) → 𝑀 × 𝑁 defined
by 𝜋( 𝑗 𝑟𝑝 𝑓 ) = (𝜎( 𝑗 𝑟𝑝 𝑓 ), 𝜏( 𝑗 𝑟𝑝 𝑓 )) = ( 𝑝, 𝑓 ( 𝑝)), which is certainly surjective. We observe that
𝜋 −1 (𝑈 × 𝑉) = 𝐽 𝑟 (𝑈, 𝑉). Considering (𝑈, 𝜑), (𝑉, 𝜓) charts of 𝑀 and 𝑁, respectively, with ℎ𝑈,𝑉
the chart in 𝐽 𝑟 (𝑀, 𝑁) associated, consider the application Φ𝑈,𝑉 : 𝜋 −1 (𝑈 ×𝑉) → 𝑈 ×𝑉 × 𝑃𝑟 (𝑚, 𝑛)
setting2 Φ𝑈,𝑉 = (𝜑−1 × 𝜓 −1 × 𝐼𝑑 𝑃𝑟 (𝑚,𝑛) ) ◦ ℎ𝑈,𝑉 . This mapping is a smooth diffeomorphism, being
a composition of diffeomorphisms. It is readily verified that, for 𝑝𝑟 : 𝑈 × 𝑉 × 𝑃𝑟 (𝑚, 𝑛) → 𝑈 × 𝑉
the coordinate projection, the diagram

Φ𝑈,𝑉
𝐽 𝑟 (𝑈, 𝑉) 𝑈 × 𝑉 × 𝑃𝑟 (𝑚, 𝑛)

𝜋 𝑝𝑟
𝑈 ×𝑉

commutes, showing Φ𝑈,𝑉 has the local trivialization property for the projection 𝜋, which allows
us to interpret 𝐽 𝑟 (𝑀, 𝑁) as a smooth fiber bundle with base space 𝑀 × 𝑁 and fiber given by the
polynomials 𝑃𝑟 (𝑚, 𝑛).

2The Cartesian product of functions is defined by ( 𝑓 × 𝑔) ( 𝑝, 𝑞) = ( 𝑓 ( 𝑝), 𝑔(𝑞)).

68

A.3 Whitney Topologies

The main interest in jet theory is due to its use in defining a class of topologies on the set
of smooth functions between two manifolds, the so-called Whitney topologies. The first contact
with such topologies is often made by using derivatives of maps with respect to concretely chosen
atlases as in the classical reference of differential topology from Hirsch (1976); indeed, it is often
more useful when computations are needed; but for the purposes of this work the approach via
jets is more advantageous because it gives us an invariant construction of these topologies, and
such a construction allows us to easily import topological properties of the known topologies in
the space of continuous functions, which we will describe shortly.
The equivalence of the coordinate version with the one via jets is not evident only from
the definitions; in fact, the proof is somewhat laborious and can be found in Mukherjee (2015),
propositions 8.2.8 and 8.2.9.

A.3.1 Weak and Strong Topologies in the Space of Continuous Functions

To construct Whitney topologies by means of jets, an intermediate step requires us to
understand the topologies on the set of continuous functions. We briefly describe the necessary
theory.
If 𝑋 and 𝑌 are topological spaces, we denote by 𝐶 0 (𝑋, 𝑌 ) the set of continuous functions
with domain in 𝑋 and codomain in 𝑌 . Following Munkres (2000), chapter 7, the first relevant
topologies on the set of continuous functions are given by the following two definitions.

Definition A.3.1. If 𝑋 and 𝑌 are topological spaces, the compact-open topology in 𝐶 0 (𝑋, 𝑌 ) is
generated by subbases of the form

𝑆(𝐾, 𝑉) = { 𝑓 ∈ 𝐶 0 (𝑋, 𝑌 ) : 𝑓 (𝐾) ⊆ 𝑉, 𝐾 ⊆ 𝑋 compact and 𝑉 ⊆ 𝑌 open}. (A.1)

Definition A.3.2. If 𝑋 and 𝑌 are topological spaces with 𝑌 metrizable with a metric 𝑑, the
compact convergence topology on 𝐶 0 (𝑋, 𝑌 ) is defined by the basis sets

𝐵( 𝑓 , 𝐾, 𝜀) = {𝑔 ∈ 𝐶 0 (𝑋, 𝑌 ) : sup 𝑑 ( 𝑓 (𝑥), 𝑔(𝑥)) < 𝜀, 𝐾 ⊆ 𝑋 compact}. (A.2)

𝑥∈𝐾

When 𝑌 is a metric space these two topologies for 𝐶 0 (𝑋, 𝑌 ) coincide (Munkres (2000),
theorem 46.8). As an immediate corollary, if 𝑌 is metrizable, then the compact convergence
topology does not depend on the choice of metric on 𝑌 .
When 𝑋, 𝑌 are manifolds, we refer to the compact-open topology (or equivalently, the
compact convergence topology) as the weak topology, denoted by 𝐶𝑊 0 (𝑋, 𝑌 ). Since manifolds
 69

are completely metrizable topological spaces3 , the following theorem (Hirsch (1976), theorem
4.1) will be relevant.
Theorem A.3.3. Let 𝑋 be a second countable topological space with each connected component
being locally compact, and 𝑌 a completely metrizable topological space. Then 𝐶𝑊 0 (𝑋, 𝑌 ) is

completely metrizable. ■
We also define another relevant topology on 𝐶 0 (𝑋, 𝑌 ). Here, Gr( 𝑓 ) denotes the graph set
of the function 𝑓 , and we consider the usual product topology on 𝑋 × 𝑌 .
Definition A.3.4. The graph topology or strong topology on 𝐶 0 (𝑋, 𝑌 ) is the topology generated
by basis

𝐵(𝑊) = { 𝑓 ∈ 𝐶 0 (𝑋, 𝑌 ) : Gr( 𝑓 ) ⊆ 𝑊, 𝑊 ⊆ 𝑋 × 𝑌 open}.

We will denote the strong topology by 𝐶𝑆0 (𝑋, 𝑌 ).

With 𝑋 Hausdorff and paracompact, 𝑌 metrizable, we have an alternative basis4 associated
with the metric 𝑑 on 𝑌 (Mukherjee (2015), proposition 8.2.5) given as follows: for 𝜖 : 𝑋 →
(0, +∞) a continuous function, the basis is formed by sets of the form

𝑆( 𝑓 , 𝜖) = {𝑔 ∈ 𝐶 0 (𝑋, 𝑌 ) : 𝑑 ( 𝑓 (𝑥), 𝑔(𝑥)) < 𝜖 (𝑥) for all 𝑥 ∈ 𝑋 }. (A.3)

More general topological questions about the strong topology can be found in Naimpally
(1966). In particular, we note that the strong topology is finer than the weak topology, and that if
𝑋 is compact and Hausdorff, the strong and the weak topologies coincide (Mukherjee (2015),
lemmas 8.2.2-8.2.4 or Naimpally (1966), theorem 4.2).
When the domain 𝑋 is not compact, we find an important difference: the strong topology
is not in general metrizable. Proposition 2 in Krikorian (1969) shows that in general 𝐶 0 (𝑋, 𝑌 ) is
not first countable (and the topological conditions of this result include manifold topologies).

A.3.2 𝐶 𝑟 Whitney Topologies

With the main results of the weak and strong topologies in the space of continuous
functions given, we define the Whitney topologies by inducing the topology of 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁))
on 𝐶 ∞ (𝑀, 𝑁) as follows:
Definition A.3.5. Let 𝑀 and 𝑁 be smooth manifolds, and 𝐽 𝑟 (𝑀, 𝑁) the corresponding space of
𝑟-jets. Consider the function

𝑗 𝑟 : 𝐶 ∞ (𝑀, 𝑁) → 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁))

𝑓 ↦→ 𝑗 𝑟 𝑓 ,
3A topological space is completely metrizable if it is metrizable with a complete metric 𝑑.
4 In
this metrizable case, the topology generated by such bases is also found in the literature as the fine topology, for
example Krikorian (1969).
70

which is known to be injective. The weak (Whitney) 𝐶 𝑟 topology on 𝐶 ∞ (𝑀, 𝑁) is the topology
induced on 𝐶 ∞ (𝑀, 𝑁) by the function 𝑗 𝑟 considering the weak topology on 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)).
∞ (𝑀, 𝑁).
We will denote the weak 𝐶 𝑟 topology by 𝐶𝑊,𝑟
Similarly, the strong (fine) 𝐶 𝑟 topology on 𝐶 ∞ (𝑀, 𝑁) is induced by the function 𝑗 𝑟
considering the strong topology on 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)). We will denote the strong 𝐶 𝑟 topology by
∞ (𝑀, 𝑁).
𝐶𝑆,𝑟

∞ (𝑀, 𝑁) → 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)) becomes a topolog-

Since 𝑗 𝑟 is injective, the map 𝑗 𝑟 : 𝐶𝑊,𝑟 𝑊
ical embedding (the analogue for the strong 𝐶 topology). The weak 𝐶 𝑟 topology is induced
𝑟

by 𝑗 𝑟 from the weak topology 𝐶𝑊 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)), and it is described by the subbases in (A.1);

therefore a subbase for 𝐶𝑊,𝑟 ∞ (𝑀, 𝑁) is

𝑆𝑟 (𝐾, 𝑉) = ( 𝑗 𝑟 ) −1 (𝑆(𝐾, 𝑉))

= { 𝑓 ∈ 𝐶 ∞ (𝑀, 𝑁) : 𝑗 𝑟 𝑓 (𝐾) ⊆ 𝑉, 𝐾 ⊆ 𝑀 compact and 𝑉 ⊆ 𝐽 𝑟 (𝑀, 𝑁) open}.

For the strong topology, with some more work (cf. Michor (1980), 4.4.1), one can verify
∞ (𝑀, 𝑁) can be given via sets of the form
that a basis for 𝐶𝑆,𝑟

𝐵𝑟 (𝑉) = { 𝑓 ∈ 𝐶 ∞ (𝑀, 𝑁) : 𝑗 𝑟 𝑓 (𝑀) ⊆ 𝑉, 𝑉 ⊆ 𝐽 𝑟 (𝑀, 𝑁) open}.

For a fixed metric 𝑑𝑟 on 𝐽 𝑟 (𝑀, 𝑁), the same reasoning as in the weak 𝐶 𝑟 topology, inducing with
𝑗 𝑟 the basis in (A.2) for the 𝐶 0 weak topology, we see that it has a basis set given by

𝐵𝑟 ( 𝑓 , 𝐾, 𝜀) = {𝑔 ∈ 𝐶 ∞ (𝑀, 𝑁) : sup 𝑝∈𝐾 {𝑑𝑟 ( 𝑗 𝑟 𝑓 ( 𝑝), 𝑗 𝑟 𝑔( 𝑝))} < 𝜀, 𝐾 ⊆ 𝑀 compact}, (A.4)

and similarly to (A.3), the strong 𝐶 𝑟 topology has a base

𝑆𝑟 ( 𝑓 , 𝜖) = {𝑔 ∈ 𝐶 ∞ (𝑀, 𝑁) : 𝑑𝑟 ( 𝑗 𝑟 𝑓 ( 𝑝), 𝑗 𝑟 𝑔( 𝑝)) < 𝜖 ( 𝑝), 𝜖 ∈ 𝐶 0 (𝑀, (0, +∞))}.

Remark A.3.6. Since we are inducing the 𝐶 𝑟 Whitney topologies by the function 𝑗 𝑟 via the
strong or the weak topology on 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)), if 𝑀 is compact then these two topologies
coincide for 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)), therefore 𝐶 𝑟 weak and strong topologies are equal if 𝑀 is
compact.

Definition A.3.7 (Whitney 𝑪 ∞ Topology). Denote T𝑟 𝑊 and T𝑟 𝑆 (𝑟 = 0, 1, 2 . . . ,) as the weak

and strong 𝐶 𝑟 topologies, respectively. The strong Whitney 𝐶 ∞ topology is defined by the base
⋃︁∞
B 𝑆 = 𝑟=0 T𝑟 𝑆 , and the weak Whitney 𝐶 ∞ topology is defined analogously.

Since T𝑠 𝑆 ⊆ T𝑟 𝑆 if 𝑠 ≤ 𝑟, (and the analogous statement for the weak topologies), we see
that these topologies are well-defined. The strong and weak 𝐶 ∞ topologies will be denoted,
respectively, by 𝐶𝑆∞ (𝑀, 𝑁) and 𝐶𝑊 ∞ (𝑀, 𝑁). Also, by remark A.3.6 we see that the 𝐶 ∞ topologies

coincide if 𝑀 is compact.
 71

A.3.3 Convergence of Sequences

Convergence of sequences in the weak 𝐶 𝑟 topology is straightforward: using the basis set
in (A.4), one easily sees that a sequence 𝑓𝑛 ∈ 𝐶 ∞ (𝑀, 𝑁) converges to 𝑓 in the weak 𝐶 𝑟 topology
if and only if, for some metric 𝑑𝑟 defined in 𝐽 𝑟 (𝑀, 𝑁), 𝑗 𝑟 𝑓𝑛 converges 𝑑𝑟 -uniformly to 𝑓 on each
compact set of 𝑀.
For the strong 𝐶 𝑟 topology, convergence of sequences is trickier. We need to control all
derivatives of the sequence (codified in the 𝑟-jet prolongation) “at infinity” (over all 𝑀), not
just on compact sets, like with the weak topology. When 𝑀 is compact, as already said, this is
not a problem (as a matter of fact strong and weak topologies coincide in this case), the issue
then lies with noncompact 𝑀. The following result gives a necessary and sufficient condition for
convergence in the strong 𝐶 𝑟 topology (see Golubitsky and Guillemin (1973), pgs. 43-44, for a
proof).

Proposition A.3.8. Let 𝑀 be noncompact. A sequence 𝑓𝑛 ∈ 𝐶 ∞ (𝑀, 𝑁) converges to 𝑓 ∈

𝐶 ∞ (𝑀, 𝑁) in the strong 𝐶 𝑟 topology if and only if there is a compact set 𝐾 ⊆ 𝑀 such that
𝑗 𝑟 𝑓𝑛 = 𝑗 𝑟 𝑓 for all outside of 𝐾 for all but a finite number of elements in the sequence, and 𝑗 𝑟 𝑓𝑛
converges to 𝑗 𝑟 𝑓 uniformly on 𝐾 with respect to some metric 𝑑𝑟 in 𝐽 𝑟 (𝑀, 𝑁). ■

A.3.4 Baire Property

Defining the Whitney topologies as we have done so far allows one to verify the Baire
property of the weak and strong topologies via the respective properties on 𝐶 0 (𝑋, 𝑌 ).

Definition A.3.9. A topological space 𝑋 is said to be a Baire space if for any countable collection
⋂︁
of dense open sets {𝐴𝑛 } of 𝑋 the intersection 𝑛 𝐴𝑛 is dense in 𝑋.

The Baire category theorem then states that completely metrizable spaces are Baire
spaces, so from theorem A.3.3:

Corollary A.3.10. If 𝑋 is second countable and locally compact, and 𝑌 is completely metrizable,
0 (𝑋, 𝑌 ) is a Baire space.
then 𝐶𝑊 ■

As observed earlier, the strong topology is not metrizable, but we still have the Baire
property. If 𝑌 is metrizable, then a subset 𝐹 ⊆ 𝐶 0 (𝑋, 𝑌 ) is said to be uniformly closed if every
sequence in 𝐹 that converges uniformly is convergent to a function in 𝐹. In particular, weakly
closed sets of 𝐶 0 (𝑋, 𝑌 ) (i.e. closed in the weak topology) are uniformly closed. We have the
following general result (Hirsch (1976), theorem 4.2):

Theorem A.3.11. Let 𝑋 be paracompact and 𝑌 be metrizable. If 𝐹 ⊆ 𝐶 0 (𝑋, 𝑌 ) is uniformly

closed, then 𝐹 is a Baire space in the strong topology (i.e. in the subspace topology of 𝐶𝑆0 (𝑋, 𝑌 )).
In particular 𝐶𝑆0 (𝑋, 𝑌 ) is a Baire space. ■
72

Connecting these results with the Whitney topologies goes through the following lemma
(Michor (1980), lemma 4.2 or Hirsch (1976), theorem 4.3):

Lemma A.3.12. The image of the function 𝑗 𝑟 : 𝐶 ∞ (𝑀, 𝑁) → 𝐶 0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)) is closed in the
weak topology. ■

0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)) a completely metrizable space, Im 𝑗 𝑟 is completely

With Im 𝑗 𝑟 closed in 𝐶𝑊
metrizable as a subspace in the weak topology, and therefore a Baire space by corollary A.3.10.
Also, Im 𝑗 𝑟 in the weak topology is homeomorphic to 𝐶𝑊,𝑟∞ (𝑀, 𝑁), therefore the weak 𝐶 𝑟 topology

is completely metrizable and a Baire space.

For the strong 𝐶 𝑟 topology, we have Im 𝑗 𝑟 weakly closed in 𝐶𝑆0 (𝑀, 𝐽 𝑟 (𝑀, 𝑁)), therefore
a Baire space as a subspace in the strong topology by theorem A.3.11, and it being again
homeomorphic to 𝐶𝑆,𝑟 ∞ (𝑀, 𝑁), the latter is a Baire space.

We can also verify the Baire property for 𝐶 ∞ topologies. For the weak topology, a
straightforward way is to work with projective limits for topological spaces (cf. Ribes and
Zalesskii (2010), chapter 1). Consider the collection of maps

∞ ∞
𝐼𝑟,𝑠 : 𝐶𝑊,𝑟 (𝑀, 𝑁) → 𝐶𝑊,𝑠 (𝑀, 𝑁), 𝑠 ≤ 𝑟,

where 𝐼𝑟,𝑠 is the identity map. It can be easily verified that the projective limit of the collection
∞ (𝑀, 𝑁), 𝐼 , N} is 𝐶 ∞ (𝑀, 𝑁) (with the same being true for the strong topology).
{𝐶𝑊,𝑟 𝑟,𝑠 𝑊
In the case of the weak topology, since all the 𝐶𝑊,𝑟∞ (𝑀, 𝑁) are completely metrizable,

the Cartesian product of these spaces is completely metrizable. Now the projective limit is
homeomorphic to a closed subset of the Cartesian product (Ribes and Zalesskii (2010, lemma
1.1.2)), and it follows that 𝐶𝑊∞ (𝑀, 𝑁) is completely metrizable, and in particular a Baire space.

Despite the strong 𝐶 ∞ topology not having the good properties of the weak topology, it is
also true that 𝐶𝑆∞ (𝑀, 𝑁) is a Baire space, the proof (see Gravesen (1983)) follows an argument
with similar strucure to the proof of the usual Baire category theorem for complete metric spaces.

A.3.5 Thom’s Tranversality

We finish with a brief note on transversality. We leave the general notions and basic
results of transversality to standard manifold theory literature (e.g. John M. Lee (2012), chapter
6.)

Definition A.3.13. Let 𝐹 : 𝑀 → 𝑁 be a smooth function and let 𝑆 ⊆ 𝑁 be an embedded

−
submanifold. We say that 𝐹 is transverse to 𝑆, denoted by 𝐹 ⋔ 𝑆, if for every 𝑝 ∈ 𝑀, either
𝐹 ( 𝑝) ∉ 𝑆, or else
𝑑𝐹𝑝 (𝑇𝑝 𝑀) + 𝑇𝐹 ( 𝑝) 𝑆 = 𝑇𝐹 ( 𝑝) 𝑁.
 73

Theorem A.3.14 (Thom Transversality Theorem). Let 𝑍 ⊆ 𝐽 𝑟 (𝑀, 𝑁) be an embedded sub-

manifold. Then the set
−
T (𝑍) = { 𝑓 ∈ 𝐶 ∞ (𝑀, 𝑁) : 𝑗 𝑟 𝑓 ⋔ 𝑍 }

is a residual set in 𝐶𝑆∞ (𝑀, 𝑁), and is in particular dense. If 𝑍 is closed, then T (𝑍) is an open
dense set ■

The proof of the Thom transversality theorem is fairly elaborated and can be seen in
Mukherjee (2015), section 8.7, or Golubitsky and Guillemin (1973), chapter 2-§5.

Corollary A.3.15 (Elementary Transversality Theorem). Let 𝑍 be a submanifold of 𝑁. Then

−
W (𝑍) = { 𝑓 ∈ 𝐶 ∞ (𝑀, 𝑁) : 𝑓 ⋔ 𝑍 }

is a dense set in 𝐶𝑆∞ (𝑀, 𝑁). If 𝑍 is closed, then W (𝑍) is an open dense set. ■
74

B MOTS Stability Operator

Here we give a brief survey of MOTS and the MOTS stability operator, and present some
of its main properties.

B.1 MOTS Stability Operator

We now return to the notion of the MOTS stability operator and its properties. This
section is an expansion on the discussion done in section 1.5.2, where the necessary theory for
the main text was given. A fairly detailed account with explicit calculations of the results cited
here can be found in Hafemann (2023) (especially appendix A, where an in depth computation of
the time derivative of 𝜃 + is given).
We recall that a MOTS on a spacetime (𝑀, 𝑔) is a surface Σ, or more generally a
codimension two immersion 𝜓 : Σ → 𝑀, with trivial normal bundle such that its null expansion
scalar 𝜃 + is zero on Σ (cf. section 1.4). To transition for the initial data version of MOTS, the
following result gives us a way to reinterpret 𝜃 + solely from the information of the immersion.
Proposition B.1.1. Let (𝑀 𝑛+1 , 𝑔) be a spacetime, 𝜙 : S 𝑛 → 𝑀 𝑛+1 be a spacelike immersion
with u ∈ 𝔛 ⊥ (𝜙) the unique unit future-directed timelike normal vector field and 𝜓 : Σ𝑛−1 → 𝑆 𝑛 a
two-sided immersion with ν ∈ 𝔛 ⊥ (𝜓) a unit normal. Defining the the future-directed null vector
√
fields as ℓ± := (u ◦ 𝜓 ± 𝑑𝜙 ◦ ν)/ 2 ∈ 𝔛 ⊥ (𝜙 ◦ 𝜓), it follows that 𝜙 ◦ 𝜓 is a spacelike immersion
of codimension two with trivial normal bundle and the null expansion given by

𝜓
𝜃 ± = trΣ K ◦ 𝜓 ± 𝐻ν , (B.1)

𝜓
where 𝐻ν is the mean curvature scalar of 𝜓 with respect to the normal ν and K is the second
fundamental form of 𝜙 with respect to the normal u and the partial trace is respect to the induced
metric. ■

With this for of the null expansions scalars, the definition of a MOTS in initial data sets is
well motivated.
Definition B.1.2 (Null expansion and MOTS - Initial data version). Let (S, ℎ, K) be an ini-
tial data set and 𝜓 : Σ → S be a two-sided immersion with ν ∈ 𝔛(𝜓) as the outward pointing
unit normal vector field of 𝜓. The outward null expansion 𝜃 + [resp. inward null expansion 𝜃 − ] of
Σ is defined as
𝜓
𝜃 ± = trΣ K ◦ 𝜓 ± 𝐻ν , (B.2)
𝜓
where 𝐻ν is the mean curvature scalar of 𝜓 with respect to the normal ν and K is the second
fundamental form of 𝜓 with respect to the normal ν and the partial trace is in respect to the
induced metric. For the sign of 𝜃 + we then define Σ to be
 75

• outer trapped if 𝜃 + < 0,

• weakly outer trapped if 𝜃 + ≤ 0,

• marginally outer trapped if 𝜃 + = 0.

Some special classes of MOTS are important in some contexts.

Definition B.1.3 (Homologous Surfaces). Let (S 𝑛 , ℎ, K) be an initial data. A pair of codi-

mension one surfaces Σ and Σ′ in S are said to be homologous if there exists a smooth map
Φ : (𝑎, 𝑏) × Σ → S satisfying

(i) [0, 1] ⊆ (𝑎, 𝑏),

(ii) for each 𝑡 ∈ (𝑎, 𝑏), the map 𝜙𝑡 : 𝑥 ∈ Σ ↦→ Φ(𝑡, 𝑥) ∈ S is an embedding,

(iii) 𝜙0 = 𝑖𝑑 Σ and 𝜙1 (Σ) = Σ′.

|︁
We say that Σ and Σ′ are outward homologous if the variation vector field 𝑉 = 𝜕Φ
𝜕𝑡 𝑡=0 is equal to
|︁
𝑉 = 𝑓 ν, where ν the outward pointing unit normal vector field of 𝜙0 , for some strictly positive
function 𝑓 ∈ 𝐶 ∞ (Σ).

Definition B.1.4 (Outermost MOTS). Let Σ be a MOTS in an initial data set (S 𝑛 , ℎ, K) with
ν an outward pointing unit normal vector field of Σ in S.

(i) We say that Σ is outermost MOTS in S if there are no outer trapped (𝜃 + < 0) or marginally
outer trapped (𝜃 + = 0) surfaces outward homologous to Σ.

(ii) We say that Σ is a weakly outermost MOTS in S provided there are no outer trapped
surfaces (𝜃 + < 0) outward homologous to Σ.

Given a spacetime (𝑀 𝑛+1 , 𝑔) let Σ𝑛−1 ⊆ 𝑀 be a smooth closed (compact without

boundary) codimension two spacelike submanifold with trivial normal bundle. For ℓ± the
future-directed normal null vector fields on Σ normalized with 𝑔(ℓ+ , ℓ− ) = −1, we shall assume
that Σ is a MOTS with respect to ℓ+ (𝜃 + = 0). For convenience, we also define on Σ the normal unit
√ √
timelike vector field u = (ℓ+ + ℓ− )/ 2 and normal unit spacelike vector field ν = (ℓ+ − ℓ− )/ 2.
Finally, let Φ : (−𝑡0 , 𝑡0 ) × Σ → 𝑀 be a smooth variation of Σ in 𝑀 with a normal variation
vector field 𝑉. As a result, the normal vector field 𝑉 can be decomposed into
|︁
𝜕Φ |︁|︁
𝑉= = 𝜙ℓ+ + 𝜓ν, 𝜙, 𝜓 ∈ 𝐶 ∞ (Σ).
𝜕𝑡 |︁𝑡=0

We shall in addition assume that a smooth choice was made on each Σ𝑡 = 𝜙𝑡 (Σ) of
future-directed null normal vector fields ℓ± (𝑡) so that ℓ± (0) = ℓ± , and ⟨⟨ℓ+ (𝑡), ℓ− (𝑡)⟩⟩ = −1.
Thus, denote by 𝜃 + (𝑡) the null expansion with respect to ℓ+ (𝑡) (that is, 𝜃 + (𝑡) = −𝑔(ℓ+ (𝑡), 𝐻 𝜙𝑡 )).
With this convention we obtain, after some hefty computations for the derivative of 𝜃 + , the key
form of 𝜃 ′+ (0).
76

Proposition B.1.5. Let Σ𝑛−1 be a MOTS within a spacetime (𝑀 𝑛+1 , 𝑔). Let Φ : (−𝑡0 , 𝑡0 )×Σ → 𝑀
be a variation with normal variation vector field 𝑉 = 𝜙ℓ+ + 𝜓ν. Then, the variation of the null
expansion scalar 𝜃 + (𝑡) on Σ in the direction of the variation vector field 𝑉 is given by

𝜃 ′+ (0) = −(| 𝜒+ | 2 + 𝑅𝑖𝑐 𝑔 (ℓ+ , ℓ+ )) · 𝜙 + 𝐿 (𝜓), (B.3)

where
𝐿(𝜓) = −Δ𝜓 + 2⟨𝑋, grad 𝜓⟩ + (𝑄 + div 𝑋 − ∥ 𝑋 ∥ 2 )𝜓, (B.4)

1 1
𝑄 = ScalΣ − [𝐽 (ν) + 𝜌] − | 𝜒+ | 2 . (B.5)
2 2
The differential operators and scalar curvature defined for the induced metric in Σ, 𝑋 is the
vector field on Σ metrically dual - also with the induced metric - to the one-form Ku (ν, ·)|𝑇 Σ and
where 𝜌 and 𝐽 are, respectively, the energy density and energy-momentum density associated
with the timelike vector field u. ■

Proposition B.1.5 allows us to define a linear second-order elliptic differential operator

𝐿, called the MOTS stability operator. Furthermore, although we have considered a MOTS in
a spacetime, we can give a purely initial-data description. This is motivated by the following
observation.
Suppose that, in addition to the conventions adopted for proposition B.1.5, we have
Σ𝑛−1 ⊆ S 𝑛 , where is a S spacelike hypersurface in the spacetime (𝑀 𝑛+1 , 𝑔), with the unique unit
normal future-directed timelike normal vector field 𝑈, induced metric ℎ and second fundamental
form K with respect to 𝑈. Assume also that an ℎ-unit normal vector field ν is chosen on Σ,
√
so that ℓ± = (𝑈| Σ ± ν)/ 2. Setting u = 𝑈| Σ and noting that X+ = Kν + Ku , where Kν is the
second-fundamental form of Σ associated with the normal vector field ν. This motivates the
following definition.

Definition B.1.6 (MOTS Stability Operator - Initial Data Version). Let Σ be a closed MOTS
(compact without boundary) within an initial data (S 𝑛 , ℎ, K). We define the MOTS stability
operator 𝐿 : 𝐶 ∞ (Σ) → 𝐶 ∞ (Σ) to be

𝐿(𝜓) = −Δ𝜓 + 2⟨𝑋, grad 𝜓⟩ + (𝑄 + div 𝑋 − ∥ 𝑋 ∥ 2 )𝜓, (B.6)

1 1
𝑄 = ScalΣ − [𝐽 (ν) + 𝜌] − |Kν + K | 2 , (B.7)
2 2
where the geometric quantities are defined on Σ, ν is the outward pointing unit normal vector
field on Σ, Kν is the scalar second fundamental form of Σ wrt. the induced metric from (S, ℎ)
on the direction ν, 𝑋 is the vector field dual to the one-form K (ν, ·) along Σ and where 𝜌 and 𝐽
are defined as in Definition 1.5.2.
 77

It is worth mentioning that the MOTS stability operator can be derived from normal
variations in the initial data. In the case of time-symmetric initial data (K = 0), the operator 𝐿
reduces to the self-adjoint, classic stability (or Jacobi) operator of the minimal surface theory,
which consists of the second variation of the volume. Although the operator 𝐿 is not self-adjoint
in general, the operator possesses crucial properties for its spectrum.

Proposition B.1.7 (Andersson, Mars, and Simon (2008), Galloway (2018)). Let Σ be a closed
MOTS (compact without boundary) within an initial data set (S 𝑛 , ℎ, K). The following statements
hold for the MOTS stability operator 𝐿.

(1) There is a real eigenvalue 𝜆 1 = 𝜆 1 (𝐿), called the principal eigenvalue of 𝐿, such that for
any other eigenvalue 𝜇, 𝑅𝑒(𝜇) ≥ 𝜆 1 . The associated eigenfunction 𝜙 ∈ 𝐶 ∞ (Σ), 𝐿𝜙 = 𝜆 1 𝜙,
is unique up to a multiplicative constant, and can be chosen to be strictly positive.

(2) 𝜆 1 ≥ 0 (resp., 𝜆 1 > 0) if only if there exist some 𝜓 ∈ 𝐶 ∞ (Σ), 𝜓 > 0, such that 𝐿(𝜓) ≥ 0
(resp., 𝐿(𝜓) > 0). ■
78

References

ABRAHAM, Ralph; MARSDEN, Jerrold E; RATIU, Tudor. Manifolds, tensor analysis, and
applications. [S.l.]: Springer Science & Business Media, 2012. v. 75.
ALIAS, Luis J; PICCIONE, Paolo. On the manifold structure of the set of unparameterized
embeddings with low regularity. Bulletin of the Brazilian Mathematical Society, New Series,
Springer, v. 42, p. 171–183, 2011.
ANDERSSON, Lars; EICHMAIR, Michael; METZGER, Jan. Jang’s equation and its
applications to marginally trapped surfaces. In: COMPLEX Analysis and Dynamical Systems IV:
General relativity, geometry, and PDE (Contemporary Mathematics, 554). [S.l.: s.n.], 2011.
P. 13–46.
ANDERSSON, Lars; MARS, Marc; SIMON, Walter. Stability of marginally outer trapped
surfaces and existence of marginally outer trapped tubes. Advances in Theoretical and
Mathematical Physics, International Press of Boston, v. 12, n. 4, p. 853–888, 2008.
AUBIN, Thierry. Some nonlinear problems in Riemannian geometry. Berlin ; New York:
Springer, 1998. P. 394. (Springer monographs in mathematics).
BACHMAN, G.; NARICI, L. Functional Analysis. [S.l.]: Dover Publications, 2000. (Academic
Press textbooks in mathematics).
BARTNIK, Robert A. Phase space for the Einstein equations. Communications in Analysis and
Geometry, International Press of Boston, Inc., v. 13, n. 5, p. 845–885, 2005.
BEEM, John K.; EHRLICH, Paul E.; EASLEY, Kevin L. Global Lorentzian Geometry. 2. ed.
New York: CRC Press, 1999. P. 656.
BEEM, John K.; HARRIS, Steven G. The generic condition is generic. General Relativity and
Gravitation, v. 25, n. 9, p. 939–962, Sept. 1993. DOI: 10.1007/BF00759194.
BERNAL, Antonio N; SÁNCHEZ, Miguel. On Smooth Cauchy Hypersurfaces and Geroch’s
Splitting Theorem. Communications in Mathematical Physics, v. 243, n. 3, p. 461–470, 2003.
BILIOTTI, Leonardo; JAVALOYES, Miquel Angel; PICCIONE, Paulo. Genericity of
nondegenerate critical points and Morse geodesic functionals. Indiana Univ. Math. J., v. 58,
p. 1797–1830, 4 2009.
CHOQUET-BRUHAT, Yvonne; GEROCH, Robert. Global aspects of the Cauchy problem in
general relativity. Communications in Mathematical Physics, v. 14, n. 4, p. 329–335, Dec.
1969. DOI: 10.1007/BF01645389.
CHRUŚCIEL, Piotr T; GALLOWAY, Gregory J. Outer trapped surfaces are dense near MOTSs.
Classical and Quantum Gravity, v. 31, n. 4, p. 11, 2014. DOI:
10.1088/0264-9381/31/4/045013.
 79

CHRUŚCIEL, Piotr T.; DELAY, Erwann. Manifold structures for sets of solutions of the general
relativistic constraint equations. Journal of Geometry and Physics, v. 51, n. 4, p. 442–472,
2004. DOI: https://doi.org/10.1016/j.geomphys.2003.12.002.
. On mapping properties of the general relativistic constraints operator in weighted
function spaces, with applications. en. Société mathématique de France, n. 94, 2003. DOI:
10.24033/msmf.407. Available from:
<http://www.numdam.org/item/MSMF_2003_2_94__1_0/>.
COLBOIS, Bruno. Laplacian on Riemannian manifolds. 2010. Available from:
<https://faculty.fiu.edu/~lhermi/dido/colbois-course.pdf>.
COOK, Gregory B. Initial Data for Numerical Relativity. Living Reviews in Relativity,
Springer Science and Business Media, v. 3, n. 1, 2000. DOI: 10.12942/lrr-2000-5.
COSTA E SILVA, Ivan Pontual. Lecture Notes on Semi-Riemannian Geometry. [S.l.: s.n.]. No
prelo 2020.
EINSTEIN, Albert. Die Feldgleichungen der Gravitation. Sitzungsberichte der Preussischen
Akademie der Wissenschaften zu Berlin, p. 844–847, 1915.
ESPINOZA, Victor Luis. Linhas e raios geodésicos causais em espaços-tempos com
aplicações à relatividade. 2020. Dissertação de Mestrado – Universidade Federal de Santa
Catarina, Florianópolis. Available from:
<https://repositorio.ufsc.br/handle/123456789/216502>.
GALLOWAY, Gregory J; SENOVILLA, José M M. Singularity theorems based on trapped
submanifolds of arbitrary co-dimension. Classical and Quantum Gravity, v. 27, n. 15,
p. 152002, 2010. DOI: 10.1088/0264-9381/27/15/152002.
GALLOWAY, Gregory J. Rigidity of outermost MOTS: the initial data version. General
Relativity and Gravitation, Springer Science and Business Media LLC, v. 50, n. 3, Feb. 2018.
GOLUBITSKY, Martin; GUILLEMIN, Victor. Stable mappings and their singularities.
Berlin Heidelberg New York, N.Y: Springer, 1973. (Graduate texts in mathematics, 14).
GRAVESEN, Jens. Whitney 𝐶 ∞ -Topologies and the Baire Property. Mathematica
Scandinavica, Mathematica Scandinavica, v. 52, n. 1, p. 58–60, 1983. Available from:
<http://www.jstor.org/stable/24491467>.
HAFEMANN, Eduardo. Geometry and topology of black hole horizons. 2023. Masters Thesis
– Universidade Federal de Santa Catarina. Available from:
<https://repositorio.ufsc.br/handle/123456789/251595>.
HAWKING, S. W. The Occurrence of Singularities in Cosmology. Proceedings of the Royal
Society of London. Series A, Mathematical and Physical Sciences, The Royal Society, v. 294,
n. 1439, p. 511–521, 1966.
80

HAWKING, S. W.; PENROSE, R. The Singularities of Gravitational Collapse and Cosmology.

Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences,
The Royal Society, v. 314, n. 1519, p. 529–548, 1970. DOI: 10.1098/rspa.1970.0021.
HAWKING, S.W; ELLIS, G.F.R. The Large Scale Structure of Space-Time. 1. ed. Cambridge:
Cambridge University Press, 1973. P. 391. (Cambridge Monographs on Mathematical Physics).
HIRSCH, Morris W. Differential Topology. New York, NY: Springer New York, 1976. v. 33.
(Graduate Texts in Mathematics).
HUNT, Brian R.; SAUER, Tim; YORKE, James A. Prevalence: a translation-invariant “almost
every” on infinite-dimensional spaces. Bulletin of the American Mathematical Society, v. 27,
p. 217–238, 1992. DOI: https://doi.org/10.1090/S0273-0979-1992-00328-2.
KRIKORIAN, Nishan. A note concerning the fine topology on function spaces. Compositio
Mathematica, Wolters-Noordhoff Publishing, v. 21, n. 4, p. 343–348, 1969. Available from:
<http://www.numdam.org/item/CM_1969__21_4_343_0/>.
LARSSON, Eric. Lorentzian Cobordisms, Compact Horizons and the Generic Condition.
2014. Master of Science Thesis – KTH Royal Institute of Technology, Stockholm. Available
from: <http://kth.diva-
portal.org/smash/record.jsf?pid=diva2%3A723418&dswid=6483>.
LEE, Dan A. Geometric relativity. Providence, Rhode Island: American Mathematical Society,
2019. P. 361. (Graduate studies in mathematics, volume 201). ISBN 9781470450816.
LEE, John M. Introduction to Riemannian Manifolds. 2. ed. [S.l.]: Springer, 2018. v. 176,
p. 437. (Graduate Texts in Mathematics).
. Introduction to Smooth Manifolds. 2. ed. New York: Springer, 2012. v. 218, p. 708.
(Graduate Texts in Mathematics).
. Introduction to Topological Manifolds. 2. ed. [S.l.]: Springer, 2011. v. 176, p. 433.
(Graduate Texts in Mathematics).
LERNER, David E. The space of Lorentz metrics. Communications in Mathematical Physics,
v. 32, n. 1, p. 19–38, 1973. DOI: 10.1007/BF01646426.
MARGALEF-ROIG, J.; OUTERELO DOMÍNGUEZ, E. Differential Topology. Amsterdam:
North-Holland, 1992.
MATHER, John N. Stability of 𝐶 ∞ Mappings: II. Infinitesimal Stability Implies Stability.
Annals of Mathematics, Annals of Mathematics, v. 89, n. 2, p. 254–291, 1969. Available from:
<http://www.jstor.org/stable/1970668>.
MICHOR, Peter W. Manifolds of differentiable mappings. Orpington: Shiva Pub, 1980. (Shiva
mathematics series ; 3).
 81

MONCRIEF, Vincent. Spacetime symmetries and linearization stability of the Einstein

equations. I. Journal of Mathematical Physics, American Institute of Physics, v. 16, n. 3,
p. 493–498, 1975.
MUKHERJEE, Amiya. Differential topology. 2. ed. Cham Heidelberg: Birkhäuser, 2015.
MUNKRES, James R. Topology. 2. ed. Upper Saddle River, NJ: Prentice Hall, Inc, 2000.
NAIMPALLY, Somashekhar Amrith. Graph topology for function spaces. Transactions of the
American Mathematical Society, v. 123, n. 1, p. 267–272, 1966. DOI:
10.1090/S0002-9947-1966-0192466-4.
O’NEILL, Barrett. Semi-Riemannian Geometry: With Applications to Relativity. 1. ed. San
Diego: Academic Press, 1983. v. 108. (Pure and Applied Mathematics).
OXTOBY, John C. Measure and category: a survey of the analogies between topological and
measure spaces. 2. ed. New York ; Heidelberg ; Berlin: Springer-Verlag, 1980. (Graduate Texts
in Mathematics, 2). OCLC: 1204344049.
PENROSE, Roger. Gravitational Collapse and Space-Time Singularities. Phys. Rev. Lett.,
American Physical Society, v. 14, p. 57–59, 1965.
RIBES, Luis; ZALESSKII, Pavel. Profinite groups. 2nd ed. Berlin ; New York: Springer, 2010.
(Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge, v. 40).
SILVA, I P Costa e. On the geodesic incompleteness of spacetimes containing marginally (outer)
trapped surfaces. Classical and Quantum Gravity, IOP Publishing, v. 29, n. 23, p. 15, 2012.
DOI: 10.1088/0264-9381/29/23/235008.
SMALE, S. An Infinite Dimensional Version of Sard’s Theorem. American Journal of
Mathematics, JSTOR, v. 87, n. 4, p. 861–866, 1965.
WALD, Robert M. General Relativity. 1. ed. Chicago: The University of Chicago Press, 1984.
P. 491.