2311 10691

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 31

GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES

arXiv:2311.10691v1 [math.DG] 17 Nov 2023

ELEFTERIOS SOULTANIS

Abstract. We construct a Lorentzian length space with an orthogonal splitting on a product I ×X


of an interval and a metric space, and use this framework to consider the relationship between metric
and causal geometry, as well as synthetic time-like Ricci curvature bounds.
The generalized Lorentzian product naturally has a Lorentzian length structure but can fail the
push-up condition in general. We recover the push-up property under a log-Lipschitz condition on
the time variable and establish sufficient conditions for global hyperbolicity. Moreover we formulate
time-like Ricci curvature bounds without push-up and regularity assumptions, and obtain a partial
rigidity of the splitting under a strong energy condition.

1. Introduction
1.1. Background and outline of results. A (smooth) globally hyperbolic metric g on a space-
time M admits an orthogonal splitting
(1.1) g = −h2 dt2 + gt , M = (a, b) × S
where gt is a “time dependent” Riemannian metric on the spacelike hypersurface {t} × S ⊂ M and
h : M → R is a positive (smooth) function [6]. Galloway’s splitting theorem [15] further improves
(1.1) to the isometric identification (M, g) ≃ (R × S, −dt2 + gS ) (with gS constant in t) assuming
the strong energy condition Ric(X, X) ≥ 0 for time-like vectors X, and that M contains a time-like
geodesic line. Galloway’s result can be interpreted as rigidity imposed on (1.1) by the strong energy
condition and the existence of complete time-like lines.
For continuous globally hyperbolic metrics only a topological splitting M ≈ (a, b)× S is available,
as a result of the stability of global hyperbolicity [34], see also [17]. The continuous case is an
important step towards a theory of non-smooth Lorentzian geometry, motivated e.g. by the study
of singularity theorems and by Penrose’s cosmic censorship conjecture. In this direction, the so
called Lorentzian pre-length spaces, introduced by Kunzinger–Sämann [29], have become a popular
framework. They consist of a causal set Y = (Y, ≤, ≪) satisfying the push-up property, and a
time separation function τ : Y × Y → [0, ∞] satisfying a reverse triangle inequality (cf. Definition
2.2). Lorentzian pre-length spaces provide an axiomatization of the causal geometry of smooth
spacetimes in the spirit of metric geometry.
In this setting, synthetic curvature conditions have been proposed in analogy with the metric
theories of sectional (Alexandrov) and Ricci (RCD) curvature bounds. For sectional curvature
bounds [1, 4, 29], an analogue of Galloway’s splitting theorem has been obtained in [4]. For Ricci
curvature bounds, introduced in [12, 32] using optimal transport, general singularity theorems were
proven in [13] but splitting theorems are not yet available. This is due in part to the lack of a
notion of infinitesimal Hilbertianity. In the metric theory this notion arises from the “analytic”
1
2 ELEFTERIOS SOULTANIS

description of synthetic Ricci bounds via the Bochner inequality [2] and is known to be a necessary
condition for the validity of splitting theorems [18, 19]. Analytic non-smooth Lorentzian geometry
is based on lower gradients (to appear in the forthcoming work [3]), but its connection to optimal
transport is yet to be explored.
In general, despite the keen interest in Lorentzian length spaces and metric techniques in recent
years [5, 9, 23, 30, 33], a number of questions remain open, e.g. (i) examples of Lorentzian length
spaces, (ii) the relationship between metric and causal geometry, and (iii) splitting theorems under
Ricci bounds. Indeed, not many examples of Lorentzian length spaces are known beyond those
arising from smooth (or C 1,1 ) spacetimes, while the connection of metric products and causal
geometry has only been considered in the work of Alexander et. al. [1] on Lorentzian warped
products – a special case of (1.1) – for sectional curvature bounds.

In this article we construct a generalization of (1.1) on product spaces where the hypersurface
S is a metric space. Indeed, given a metric space X, an interval I ⊂ R, a continuous function
h : I × X → (0, ∞), and a family F = {ds }s∈I of compatible metrics on X, we construct the
generalized Lorentzian product, denoted hI ×F X. For it, we establish (1) the push-up principle
(Theorem 1.3) under a log-Lipschitz condition, (2) sufficient conditions for global hyperbolicity
(Theorem 1.5), and (3) partial rigidity of splittings under strong energy conditions (Corollary 1.7).
The construction of the generalized Lorentzian product gives new examples of Lorentzian length
spaces, cf. Corollary 1.4 and Theorem 1.6. It moreover generalizes the Lorentzian warped product
considered in [1] and is compatible with (1.1) in the setting of manifolds: if X is a smooth manifold
and F is given by continuous Riemannian metrics, then hI ×F X coincides with the Lorentzian
length structure induced by (1.1), cf. Corollary 1.1.
In contrast to the warped products of [1] (which correspond to h = 1 and ds = f (s)d for a
continuous function f : I → (0, ∞) and a fixed metric d on X), pathological causal phenomena
may occur in the generality of our setting (see e.g. [14,16]), violating the push-up property required
in the definition of Lorentzian length spaces. In Lorentzian manifolds, causal theory is known to
break down for metrics below Lipschitz regularity [14, 21]. We identify a log-Lipschitz condition
(1.4) on F – inspired by the work of Kopfer–Sturm [26, 27, 28] – ensuring that the generalized
Lorentzian product hI ×F X is a Lorentzian length space (cf. Corollary 1.4). Moreover we show
the regularity of hI ×F X for conformal metrics in Theorem 1.6, generalizing the corresponding
fact for Lipschitz continuous Lorentzian metrics [20, 31] in the special case of conformal metrics.
Applying synthetic time-like curvature bounds to the generalized Lorentzian product, we obtain
the following rigidity statement (Corollary 1.7): if R ×F X equipped with a measure m = ms ⊗ ds
satisfies the strong energy condition, then the sliced measures ms are independent of s.

1.2. Generalized Lorentzian product. Throughout this paper X is a locally compact topo-
logical space and I ⊂ R an (open) interval. Suppose h : I × X → (0, ∞) is continuous, and
F = {ds }s∈I is a family of metrics on X whose induced length metrics generate the topology of X.1
We make the following local continuity assumption: for each compact J ⊂ I and K ⊂ X, there
exists r = rJ,K > 0 and a modulus of continuity ω = ωJ,K : [0, ∞) → [0, ∞) such that

1This is the case for example when each metric in F is a length metric or a quasiconvex metric.
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 3

ds (x, y)
(1.2) log ≤ ω(|s − s′ |), for s, s′ ∈ J, x, y ∈ K with inf ds (x, y) < r.
ds′ (x, y) s∈J

In particular all the metrics in F are locally bi-Lipschitz equivalent, and the absolute continuity
of a curve β : [a, b] → X is independent of the chosen metric in F. The generalized metric speed of
an absolutely continuous curve β : [a, b] → X is a map vβ : I × [a, b] → [0, ∞] with the following
properties (cf. Lemma 3.2):
• for each t ∈ [a, b], s 7→ vβ (s, t) is continuous in I, and
• for each s ∈ I we have that vβ (s, t) = |βt′ |ds a.e. t.2
The generalized length element of an absolutely continuous curve γ = (α, β) : [a, b] → I × X is
(1.3) |τγ′ |2 (t) := h(γt )2 α′ (t)2 − vβ (αt , t)2 ,
and we say that γ is (a) causal, if |τγ′ |2 (t) ≥ 0 a.e. t ∈ [a, b], and (b) timelike, if |τγ′ |2 (t) > 0 a.e.
t ∈ [a, b], cf. Definition 4.1. A standard construction detailed in Section 4.1 gives rise to a causal
set (I × X, ≤, ≪) and time separation function τ on I × X. In the terminology of [1, Appendix
A], the collection (I × X, ≤, ≪, τ ) is a Lorentzian length structure, which we denote by hI ×F X
and call the generalized Lorentzian product. In Corollary 4.15 we show that when X is a smooth
manifold and F given by continuous Riemannian metrics, the length elements given by (1.1) and
(1.3) coincide. As a direct consequence we obtain the compatibility of hI ×F X and the Lorentzian
length structure induced by the Lorentzian product metric.
Corollary 1.1. Let X be a smooth manifold, h : I × X → (0, ∞) continuous, and suppose {gs }s∈I
is a family of Riemannian metrics on X so that the local coordinate representations (gs )ij (p) are
continuous in (s, p). Let F = {dgs }s∈I , where dgs is the length metric associated to gs for each
s ∈ I. Then hI ×F X is the Lorentzian length structure induced by g := −h2 ds2 + gs .
1.3. Lorentzian length spaces and global hyperbolicity. Let h and F be as in Section 1.2.
Under these hypotheses, the generalized Lorentzian product hI ×F X might have causal bubbles
(see [16]) and thus fail to be a Lorentzian length space, as a result of failing the push-up property.
Recall that a causal set (Y, ≤, ≪) satisfies the push-up property if x ≪ z whenever x ≤ y and
y ≪ z or x ≪ y and y ≤ z.
To prevent this failure, we impose a log-Lipschitz condition in the spirit of [26, 27, 28].
Definition 1.2. We say that F satisfies a local log-Lipschitz condition if, for every compact J ⊂ I,
K ⊂ X, there exists C = CJ,K , r = rJ,K > 0 such that
ds (x, y)
(1.4) log ≤ C|s − s′ |, for s, s′ ∈ J, x, y ∈ K with inf ds (x, y) < r.
ds′ (x, y) s∈J

In other words, we require that the moduli of continuity ω in (1.2) are linear, corresponding to a
Lipschitz-type control on the time-variable. We remark here that (1.4) in Definition 1.2 is inspired
by the work of Kopfer–Sturm on Super Ricci flows on metric spaces [26, 27, 28], where a similar
condition appears. Whereas they focus on the evolution and gradient flows of “variable metric
spaces,” we instead consider the Lorentzian geometry of the product.
2Here |β ′ | denotes the metric speed of β with respect to the metric d .
t ds s
4 ELEFTERIOS SOULTANIS

Theorem 1.3. Suppose F satisfies (1.4) and, for every compact J ⊂ I, K ⊂ X there exists C ′ > 0
such that
(1.5) |h(s, x) − h(s′ , x)| ≤ C ′ |s − s′ |, s, s′ ∈ J, x ∈ K.
Then hI ×F X satisfies the push-up property.
The push-up property further implies the lower semicontinuity of the associated time separation
function τ on hI ×F X. These facts together ensure the generalized Lorentzian product is a
Lorentzian length space.
Corollary 1.4. If F and h satisfy (1.4) and (1.5), respectively, then hI ×F X is a Lorentzian
length space.
The Lorentzian length space hI ×F X is moreover strongly local, see Proposition 5.4. However,
local conditions such as (1.4) do not guarantee global hyperbolicity. While it is easy to see that
hI ×F X is globally hyperbolic if h is bounded and the metrics in F are globally bi-Lipschitz
to a proper metric (with uniformly bounded bi-Lipschitz constants), examples in [35] show that
properness of the slices does not imply global hyperbolicity, nor does global hyperbolicity imply
that the slices are proper. For smooth spacetimes, global hyperbolicity is nevertheless equivalent
the existence of a proper metric on the underlying space: in [10] this is shown by considering the
null-distance associated to a suitable time function. However in our context the natural choice of
time function (t, x) 7→ t does not necessarily yield a proper null-distance, cf. [35, § 6.1].
In Theorem 1.5 below we give a sufficient condition for the global hyperbolicity in terms of F and
h. The criterion (1.6) essentially characterizes the properness of a natural length metric on I × X
associated with F and h (see Section 3.2) and is phrased in terms of curves tending to infinity. A
locally absolutely continuous curve β : [0, ∞) → X is said to tend to infinity if, for every compact
K ⊂ X, there exists T > 0 such that K ∩ β([T, ∞)) = ∅.
Theorem 1.5. Suppose F and h satisfy (1.4) and (1.5), and moreover
Z ∞
vβ (αt , t)
(1.6) dt = ∞
0 h(αt , βt )
whenever β : [0, ∞) → X tends to infinity and α : [0, ∞) → I is locally absolutely continuous. Then
hI ×F X is globally hyperbolic.
Observe that (1.6) does not imply the properness of the metrics in F but instead of the confor-
mally weighted metrics (ds )h(s,·)−1 ; however it is a stronger condition than the propeness of each
(ds )h(s,·)−1 , s ∈ I.
1.4. Regularity. The regularity of hI ×F X seems more difficult to establish. Roughly speaking,
a Lorentzian length space is called regular, if maximizing curves are either null or time-like (have
a definite causal character). These are curves maximizing the Lorentzian length of curves joining
given points, and they replace geodesics in the current non-smooth framework where no geodesic
equation is available. Unlike geodesics, maximizing curves in Lorentzian length spaces need not have
a definite causal character; they can have null-segments while having positive Lorentzian length.
Regularity is known to hold for spacetimes with a Lorentzian metric of Lipschitz regularity, due
to a weak formulation of the geodesic equation [20, 31]. However, even in the setting of Lorentzian
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 5

metrics, (1.4) and (1.5) are weaker than Lipschitz regularity, and we do not know whether they
imply regularity of the spacetime in the absence of a suitable Lipschitz condition on the space
variable. However we establish regularity of hI ×F X in the current framework in the special case
F = {dρ(s,·) }s∈I , where ρ : I × X → (0, ∞) is a (locally Lipschitz) is a given conformal weight.
Here, given a length metric d on X and a continuous g : X → (0, ∞), the conformally weighted
metric dg is a length metric on X such that
|βt′ |dg = g(βt )|βt′ | a.e. t
for curves β in X. The precise definition is given in Section 2.1, compare (2.2).
Theorem 1.6. Let d be a locally compact length metric on X. Suppose that h, ρ : I × X → (0, ∞)
are locally Lipschitz, and set F = {dρ(s,·) }. Then F satisfies (1.4), and hI ×F X is a regular
Lorentzian length space.
1.5. Remarks on synthetic time-like Ricci curvature bounds. The metric techniques used
in [18, 19] to prove the splitting theorem of RCD-spaces are not available in the Lorentzian setting
as of yet. In the setting of generalized Lorentzian products hI ×F X, where the splitting is a priori
available, we may instead ask what restrictions do synthetic time-like Ricci curvature bounds place
on h and F, particularly concerning rigidity: if R ×F X equipped with a measure m = ms ⊗ ds
has non-negative synthetic time-like Ricci curvature, must s 7→ ms and s 7→ ds be constant?
Here, M = {ms }s∈I is a family of measures on X such that (a) s 7→ ms (E) is Borel whenever
m (B (x,2r))
E ⊂ X is Borel, (b) lim supr→0 mss (Bdds (x,r)) < ∞ ms -a.e. x ∈ X for each s ∈ I (the infinitesimal
s
doubling property), and (c) for each compact J ⊂ I and K ⊂ X, there exists C = C(J, K) > 0
with
(1.7) ms′ |K ≤ Cms |K , s, s′ ∈ J.
In Section 2.4 we formulate the wTCDp (K, N )-condition on Lorentzian length structures; the
ensuing notion (Definition 2.9) agrees with the [7, Definition 2.20] for (K-)globally hyperbolic
regular Lorentzian length spaces. Although the tools to obtain splitting and its rigidity are not
available, utilizing the product structure and techniques akin to the needle decomposition in [11],
we show that, if hI × FX satisfies the wTCDp (K, N )-condition, then the space slices hI × {x} are
RCD(K, N )-spaces for a.e. x ∈ X. We refer to Corollary 6.3 for the detailed statement. We remark
moreover that, in contrast to [11], no non-branching assumption on hI ×F X is needed due to the
product structure available in our setting. As a corollary of Theorem 6.2 we obtain the following
partial rigidity of time-like non-negatively curved generalized Lorentzian products.
Corollary 1.7. Suppose F and M satisfy (1.2) and (1.7), respectively. Let m = ms ⊗ ds and
suppose (R ×F X, m) satisfies the wTCDp (0, N )-condition for p ∈ (0, 1) and N ≥ 1. Then there is
a measure m on X such that ms = m for a.e. s ∈ I.
Organization of the paper. We give some background on in Section 2: metric spaces (Section
2.1), Lorentzian length spaces and structures (Sections 2.2 and 2.3), and time-like Ricci curvature
bounds (Section 2.4). We then define the generalized metric speed and establish its basic properties
in Section 3.
6 ELEFTERIOS SOULTANIS

In Section 4 we define the generalized Lorentzian product, based on generalized metric speed.
We discuss its causal structure in Section 4.1, the basic properties of causal curves in Section 4.2,
and prove Corollary 1.1 in Section 4.3.
In Section 5 we prove Theorem 1.3, Corollary 1.4. The proofs boil down to using Caratheodory
theory of ODE’s (Appendix A), which the log-Lipschitz condition (1.4) makes available. Theorem
1.5 is proved in Section 5.3 using a properness criterion given by (1.6). Theorem 1.6, proved in
Section 5.4, follows by reducing the problem to Lipschitz regular Lorentzian metrics.
Finally, Section 6 contains the discussion of synthetic time-like Ricci curvature bounds on gen-
eralized Lorentzian products, and the proof of Corollary 1.7.

Acknowledgements. The research in this manuscript was done during my stay at Radboud Uni-
versity and the University of Warwick, as well as the Thematic Program on Nonsmooth Riemannian
and Lorentzian Geometry at the Fields Institute. I gratefully acknowledge the financial support
of the Radboud Excellence Initiative, the Fields Institute and the ERC (grant no. 948021) of
Prof. Bate (Warwick). I would also like to thank Leonardo Garcia-Heveling, Annegret Burtscher,
Mathias Braun and Nicola Gigli for many interesting discussions on this manuscript and beyond.

2. Background
2.1. Metric spaces. Let (X, d) be a metric space. A curve β : [a, b] → X is called absolutely
continuous if there exists g ∈ L1 (a, b) such that
Z s
(2.1) d(βs , βt ) ≤ g(r)dr, a ≤ t ≤ s ≤ b.
t
d(βt+h , βt )
In this case, the limit |βt′ | := lim exists a.e. t ∈ [a, b], and t 7→ |βt′ | is the pointwise a.e.
h→0 |h|
Z b
smallest function for which (2.1) holds. We define the length of β as ℓ(β) = |βt′ |dt. Moreover,
a
there exists a non-decreasing function sβ : [0, ℓ(β)] → [a, b] and a parametrization β̄ : [0, ℓ(β)] → X,
called the arc-length parametrization of β, such that β ◦ sβ = β̄ and |β̄t′ | = 1 for every t ∈ [0, ℓ(β)].
See [8, 22]. We say that d is a length metric if
d(x, y) = inf{ℓ(β)| β : x y y}, x, y ∈ X.

To define metrics with a given conformal factor, we first define the line integral of a Borel function
g : X → [0, ∞] along an absolutely continuous curve β : [a, b] → X by
Z Z b
g ds := g(βt )|βt′ |dt.
β a

Note that the line-integral is independent of the parametrization of β, cf. [22, Theorem 3.12].
Suppose now d is a locally compact length metric on X, and let ρ : X → (0, ∞) be a continuous
function. We define the weighted length
Z
Lρ (β) := ρ ds
β
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 7

for any absolutely continuous curve β in X. Then Lρ is lower semicontinuous, and the arising
length structure (Lρ , AC(X)) on X is compatible with the metric topology of (X, d); the length
metric
nZ b
dρ (x, y) = inf ρ(βt )|βt′ |dt β : x y y}, x, y ∈ X,
a
induces the same topology as d. Moreover by [8, Theorem 2.4.3] we have that
(2.2) |βt′ |dρ = ρ(βt )|βt′ | a.e. t ∈ [a, b]
for every absolutely continuous curve β : [a, b] → (X, d).

Z ∞ 2.1. Assume (X, d) is a locally compact length space, and ρ : X → (0, ∞) continuous.
Lemma
If ρ(βt )|βt′ |dt = ∞ whenever β : [0, ∞) → X is locally absolutely continuous and leaves every
0
compact subset of X, then (X, dρ ) is proper and geodesic.
Proof. By the Hopf–Rinow theorem it suffices to show that (X, dρ ) is complete. Suppose (xj ) ⊂ X is
a Cauchy sequence for dρ . The claim follows if we show that (xj ) contains a subsequence converging
to a point x ∈ X with respect to dρ . Pass to a subsequence (labeled with the same indices) such that
dρ (xj , xj+1 ) < 2−j−1 , and let βj : [j, j + 1] → X be an absolutely continuous curve connecting xj to
xj+1 constant speed parametrized (with respect to d) such that Lρ (βj ) < dρ (xj , xj+1 )+2−j−1 < 2−j .
Consider the concatenation β : [0, ∞) → X defined such that β|[0,N ] = βN −1 · · · β1 β0 . We have that
β is locally absolutely continuous and
Z ∞ ∞
X

ρ(βt )|βt |d dt = Lρ (βj ) < 2,
0 j=0
hence– by our assumption – it cannot leave every compact set. Consequently there exists a compact
set K ⊂ X such that (xj ) ⊂ K. We conclude that a subsequence of (xj ) converges to x ∈ X in d,
and thus in dρ .

2.2. Lorentzian length spaces. The definitions below are taken from [29], which the interested
reader should consult for more details. A causal set is a space Y equipped with a transitive and
reflexive relation ≤, together with a transitive relation ≪ contained in ≤. We say that a causal set
(Y, ≤, ≪) satisfies the push-up property, if x ≪ z whenever x ≤ y and y ≪ z or x ≪ y and y ≤ z.
Definition 2.2. A Lorentzian pre-length space Y = (Y, ≤, ≪, τ ) consists of a causal set (Y, ≤, ≪)
together with a metric d on Y , and a lower semicontinuous time-separation function τ : Y × Y →
[0, ∞] such that τ (x, y) > 0 if and only if x ≪ y, satisfying the reverse triangle inequality:
τ (x, y) + τ (y, z) ≤ τ (x, z) if x ≤ y ≤ z.
A Lorentzian pre-length space satisfies the push-up property, and the past I + (x) = {y : x ≪ y}
and future I − (x) = {y : y ≪ x} of a point x ∈ Y are open. The corresponding sets with ≪ replaced
by ≤ are denoted J ± (x). A Lorentzian pre-length space Y is called
• locally causally closed, if every point in Y has a neighbourhood U such that ≤ is closed in
Ū × Ū ;
8 ELEFTERIOS SOULTANIS

• strongly causal, if the metric topology of Y is generated by the chronological diamonds


I(x, y) = I + (x) ∩ I − (y), x, y ∈ Y ;
• totally non-imprisoning, if for any compact K ⊂ Y there exists C such that Lτ (γ) ≤ C for
any causal curve γ in K;
• Globally hyperbolic, if it is totally non-imprisoning and causal diamonds J(x, y) := J + (x) ∩
J − (y) (x, y ∈ Y ) are compact.
A curve γ : [a, b] → Y is future-directed causal (resp. time-like) if γt ≤ γs (resp. γt ≪ γs )
whenever a ≤ t < s ≤ b. The Lorentzian length or τ -length of a causal curve γ is
N
nX o
(2.3) Lτ (γ) = inf τ (γti−1 , γti ) : a = t0 < · · · < tN = b ,
i=1

and γ is called maximal, if Lτ (γ) = τ (γa , γb ). We call a Lorentzian pre-length space Y


(i) causally path connected, if there exists a causal curve γ : x y y whenever x ≤ y, and a
time-like curve γ : x y y whenever x ≪ y;
(ii) geodesic, if there exists a maximal curve γ : x y y whenever x ≤ y;
(iii) localizable, if every point in Y has a neighbourhood U such that
LU
τ (p, q) := sup{Lτ (γ) : γ : p y q causal, and γ ⊂ U }, p, q ∈ U
is bounded from above and LU
τ ≤ τ |U ×U , and moreover YU := (U, ≤U ×U , ≪U ×U , LU
τ ) is a
geodesic Lorentzian pre-length space;
(iv) locally regularizable, if the neighbourhood U above can be chosen so that maximizing curves
in YU with positive τ -length are time-like.
A neighbourhood U as in (iii) and/or (iv) is called a localizing neighbourhood. Locally regulariz-
ability ensures that maximizing curves have a definite causal character, i.e. they are either time-like
or null (zero τ -length), cf. [29, Theorem 3.18].
Definition 2.3. (Definition 2.8 and 3.22 in [29]) A localizable, locally causally closed, and causally
path connected Lorentzian pre-length space (Y, ≤, ≪, τ ) is called a
(1) Lorentzian length space, if τ (x, y) = sup{Lτ (γ) : γ : x y y causal} whenever x ≤ y;
(2) regular Lorentzian length space, if it is a locally regularizable Lorentzian length space;
(3) Lorentzian geodesic space, if it is a geodesic Lorentzian length space.
2.3. Lorentzian length structures. We sketch the idea of Lorentzian lengths structures here
and refer to [1, Appendix A] for the precise definitions. Suppose I + ⊂ C + ⊂ AC(Y, d) are families
of curves in a metric space (Y, d) closed under concatenation, restriction, and reparametrization.
← ← ←
If γ denotes the curve γ with reversed orientation and I − = { γ : γ ∈ I + }, C − = { γ : γ ∈ C + },
then (I ± , C ± ) is admissible in the sense of [1, Definition A.1]. For the following definition set
I := I − ∪ I + , C := C − ∪ C + .
Definition 2.4. A Lorentzian length structure on Y consists of such a tuple (I ± , C ± ) together with
a Lorentzian length functional L : C → [0, ∞] that is additive, invariant under reparametrization,
strictly positive on I, and compatible with the topology: if γ : [a, b] → Y is in C, then t 7→ L(γ|[a,t] )
is continuous.
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 9

We call curves in I ± (resp. C ± ) past/future time-like (resp. causal). A Lorentzian length


structure (Y, I ± , C ± , L) induces a causal set (Y, ≤, ≪), where x ≤ y if there exists γ : x y y in C + ,
and x ≪ y if there exists γ : x y y in I + , cf. [1, Lemma A.4]. Moreover,
τL (x, y) = sup{L(γ) : γ : x y y is in C + }
(with the convention sup ∅ = 0) defines a time-separation function which is positive on ≪ and
satisfies the reverse triangle inequality, cf. [1, Lemma A.6 and A.7]. The induced tuple (Y, ≤, ≪, τL )
is causally path connected, and satisfies the length condition in Definition 2.3(1). However τ is not
lower semicontinuous and the push-up property may not hold in general.
Remark 2.5. If (Y, ≤, ≪, τ ) is a Lorentzian length space and I + , C + are the families of future
directed time-like and causal curves, respectively, and if L is given by (2.3), then (I ± , C ± , L) is a
Lorentzian length structure whose induced causal set coincides with (Y, ≤, ≪), and τ = τL .
2.4. Synthetic time-like Ricci curvature bounds. We define the (weak) TCDp (K, N )-condition
for Lorentzian length structures Y for p ∈ (0, 1), K ∈ R and N ∈ [1, ∞) following the ideas in [7,12].
Given two measures µ, ν ∈ P(Y ), a coupling of (µ, ν) is a measure π ∈ P(Y × Y ) such that
(p1∗ π, p2∗ π) = (µ, ν). A coupling is causal if it is concentrated on ≤, and their collection is
denoted Π≤ (µ, ν). Causal couplings are closely related with causal plans, that is measures η ∈
P(C([0, 1], Y )) concentrated on C + : if η is a causal plan, then (e0 , e1 )∗ η is a causal coupling of
(e0∗ η, e1∗ η). Here et : C([0, 1], Y ) → Y is the evaluation map γ 7→ γt for t ∈ [0, 1].
Definition 2.6. A functional S : P(Y ) → [−∞, ∞] is (K, N )-convex along a causal plan η, if the
function t 7→ u(t) := exp(−S(et∗ η)/N ) satisfies
Z 1/2
1−t t 2
u(t) ≥ σK/N (Tη )u(0) + σK/N (Tη )u(1), where Tη = L(γ) dη(γ) .

K
Equivalently, u is semiconvex and satisfies u′′ (t) ≤ − Tη u(t) a.e. t ∈ [0, 1].
N
In the definition above we use the auxiliary functions
 sκ (tθ)  sin(√κθ)
 sκ (θ) κθ 2 < 0 or κθ 2 ∈ (0, π 2 ) 
 √
κ
κ>0
σκt (θ) := t κθ 2 =0 , sκ (θ) := θ √ κ=0

∞ κθ 2 ≥ π 2  sinh(√ −κθ)

κ < 0.
−κ

Next we define p-optimal time-like plans. Given p ∈ (0, ∞), define the Lorentz–Wasserstein
p-distance
n Z 1/p o
ℓp (µ, ν) = sup τL (x, y)dπ(x, y) : π ∈ Π≤ (µ, ν) ,

with the convention sup ∅ = 0. Moreover, let OpT(Y ) denote the family of curves γ ∈ I + such
that
τ (γs , γt ) = (s − t)L(γ), a ≤ t ≤ s ≤ b.
Note that, since OpT(Y ) ⊂ I ⊂ AC(Y, d), its elements are automatically continuous.
10 ELEFTERIOS SOULTANIS

Definition 2.7. A p-optimal time-like plan is a causal plan concentrated on OpT(Y ) such that
Z
ℓp (e0∗ η, e1∗ η) = L(γ)p dη(γ).
p

If e0∗ η = µ and e1∗ η = ν, η is called a p-optimal plan for (µ, ν), and their collection is denoted
OpTp (µ, ν).
Remark 2.8. Let p ∈ (0, 1). If Y is a (K-)globally hyperbolic regular Lorentzian length space, then
every strongly time-like p-dualizable pair (µ, ν) of probability measures on Y admits a p-optimal
time-like plan for (µ, ν), see [7, Lemma 2.13].
Definition 2.9. Let (Y, I ± , C ± , L) be a Lorentzian length structure and m be a Radon measure on
Y . Let p ∈ (0, 1), K ∈ R and N ∈ [1, ∞). We say that (Y, m) satisfies the wTCDp (K, N )-condition
if, for any µ, ν ∈ Pcac (Y, m) with OpTp (µ, ν) 6= ∅, there exists η ∈ OpTp (µ, ν) such that Entm is
(K, N )-convex along η.
Here, the entropy functional Entm : P(Y ) → [−∞, ∞] is given by
 Z
ρ log ρdm, µ = ρm with (ρ log ρ)+ ∈ L1 (m)

Entm (µ) =
 +∞, otherwise

3. A generalized product metric


3.1. Generalized metric speed. Let X be a topological space, I ⊂ R an interval, and F =
{ds }s∈I a family of metrics on X satisfying (1.2).
Definition 3.1. Let β : [a, b] → X be an absolutely continuous curve. We define vβ : I × [a, b] → R
by
1 t+h ′
Z
vβ (s, t) = lim sup |βτ |ds dτ.
h→0 h t
Lemma 3.2. Suppose J ⊂ I and K ⊂ X are compact. Let β : [a, b] → K be an absolutely
continuous curve. There exists C = C(J, K) such that
(1) for every s ∈ I, vβ (s, t) = |βt′ |ds a.e. t ∈ [a, b];
(2) for every compact subinterval J ⊂ I we have that
vβ (s, t)
≤ 1 + Cω(|s′ − s|), t ∈ [a, b], s, s′ ∈ J,
vβ (s′ , t)
where ω = ωJ,K is the modulus of continuity in (1.2).
In the proof we use the observation eA ≤ 1 + C(D)A for A ∈ [0, D], which implies that (1.2) can
be equivalently expressed as
ds (x, y)
(3.1) ≤ 1 + CωJ,K (|s − s′ |) for s, s′ ∈ J, x, y ∈ K with inf ds (x, y) < r,
ds′ (x, y) s∈J

for some constants C = CJ,K , r = rJ,K > 0 for each compact J ⊂ I and K ⊂ X. If F satisfies
(1.4), the right hand side of (3.1) can be replaced by 1 + C|s − s′ |. In light of this, Lemma 3.2 has
the following immediate corollary.
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 11

Corollary 3.3. Suppose F satisfies (1.4). For every compact K ⊂ X and J ⊂ I, there exists
C = CK,J such that
vβ (s′ , t)
≤ 1 + C|s′ − s|, t ∈ [a, b], s, s′ ∈ J,
vβ (s, t)
for any absolutely continuous curve β : [a, b] → K.
Proof. The first claim follows from Lebesgue’s differentiation theorem since, for fixed s ∈ I, the
map t 7→ |βt′ |ds is L1 -integrable by the absolute continuity of β with respect to ds .
By (3.1) we have that
ℓds (β|[t,t+h] ) ℓd ′ (β|[t,t+h] )
≤ s (1 + Cω(|s − s′ |))
h h
and, upon taking limsup as h → 0,
vβ (s, t) ≤ vβ (s′ , t)(1 + Cω(|s − s′ |))
for s, s′ ∈ J. This proves (2). 

The following change of variables formula for vβ will be very useful.


Lemma 3.4. Let β : [a, b] → X be an AC-curve, and ϕ : [c, d] → [a, b] an absolutely continuous
bijection. Then
vβ◦ϕ (s, t) = |ϕ′ (t)|vβ (s, ϕ(t))
for all s ∈ I and t ∈ [a, b] for which ϕ′ (t) exists.
In particular if α : [a, b] → I is an absolutely continuous curve we have that
vβ◦ϕ (α(t), t) = |ϕ′ (t)|vβ (α(t), ϕ(t)) a.e. t.

Proof. We may assume that ϕ is increasing. Let s ∈ I and observe that


ℓs (β|[ϕ(t),ϕ(t)+δ(h)] )
vβ◦ϕ (s, t) = lim sup ,
h→0 h
where δ(h) := ϕ(t + h) − ϕ(t). If ϕ′ (t) exists we thus have that
δ(h) ℓs (β|[ϕ(t),ϕ(t)+δ(h)] ) ℓs (β|[ϕ(t),ϕ(t)+δ] )
vβ◦ϕ (s, t) = lim sup = ϕ′ (t) lim sup = ϕ′ (t)vβ (s, ϕ(t)).
h→0 h δ(h) δ→0 δ

Lemma 3.5. Suppose a sequence γj = (αj , βj ) : [a, b] → I × X of AC-curves converges uniformly
to an AC-curve γ = (α, β). Then
Z d Z d
vβ (α(t), t)dt ≤ lim inf vβj (αj (t), t)dt
c j→∞ c
whenever a ≤ c ≤ d ≤ b.
12 ELEFTERIOS SOULTANIS

Proof. By considering the restriction to [c, d] we may assume that c = a and d = b. Fix ε > 0. By
uniform convergence (αj ) is equicontinuous and thus there exists δ > 0 such that
supj |αj (t) − αj (s)| ≤ ε whenever |t − s| < δ. Let a = a0 < · · · < aM = b be a partition of
[a, b] such that ai − ai−1 < δ, i = 1, . . . , M . For each i we have by Lemma 3.2(2) that
Z ai Z ai Z ai
vβ (α(t), t)dt ≤ (1 + ω(ε)) vβ (α(ai ), t)dt ≤ (1 + ω(ε)) lim inf vβj (α(ai ), t)dt
ai−1 ai−1 j→∞ ai−1
Z ai
≤ (1 + ω(ε))2 lim inf vβj (αj (t), t)dt
j→∞ ai−1
Summing over i and taking the limit ε → 0 we have the estimate
Z b Z b
vβ (α(t), t)dt ≤ lim inf vβj (αj (t), t)dt,
a j→∞ a
as claimed. 
3.2. Generalized product metric. We define the generalized product metric dF ,h on I × X as
the length metric arising from the length functional
Z bq
LF ,h (γ) = h(γt )α′ (t)2 + vβ (αt , t)2 dt, γ = (α, β) : [a, b] → I × X,
a
see Section 2.1. We note that dF ,h generates the product topology on I × X, and (I × X, dF ,h )
is therefore a locally compact length space. Note that, if J ⊂ I is a subinterval (we will mostly
consider compact subintervals) and FJ = {ds }s∈J , then (J × X, dFJ ,h ) is homeomorphic with
(J × X, dF ,h ) but the metrics need not coincide.
Another convenient choice of background metric on I × X in what follows will be the conformally
weighted metric (dF ,h )h−1 , see Section 2.1. We have that (dF ,h )h−1 generates the product topology,
and a curve γ = (α, β) : [a, b] → I × X is in AC(X, dF ,h ) if and only if it is in AC(X, (dF ,h )h−1 ),
and in this case
s

|γt′ |dF ,h vβ (αt , s)2
|γt |(dF ,h )h−1 = = α′ (t)2 + a.e. t ∈ [a, b].
h(γt ) h(γt )2
If not explicitly stated otherwise, we tacitly equip I × X with the metric (dF ,h )h−1 .

4. The Lorentzian length structure on hI ×F X


4.1. Causal structure. Let X and F be as in Section 3.1 – we assume in particular that F
satisfies (1.2) – and let h : I × X → (0, ∞) be continuous. Throughout this subsection we denote
Y = I × X.
Definitions. An absolutely continuous curve γ = (α, β) : [a, b] → Y is said to be future directed if
α′ > 0 a.e. and past directed if α′ < 0 a.e.
Definition 4.1. Let γ = (α, β) : [a, b] → Y be a future or past directed curve. We say that γ is
(a) causal if vβ (αt , t)2 ≤ h(γt )2 α′ (t)2 a.e. t ∈ [a, b];
(b) timelike if vβ (αt , t)2 < h(γt )2 α′ (t)2 a.e. t ∈ [a, b];
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 13

We also say that γ is a null curve if vβ (αt , t)2 = h(γt )α′ (t)2 a.e. t ∈ [a, b]. Note that by
Lemma 3.4 the notions above are invariant under absolutely continuous orientation preserving
reparametrizations. Given a causal curve γ = (α, β) : [a, b] → Y recall its Lorentzian length
given by
Z bq
(4.1) τ (γ) = h(γ(t))2 α′ (t)2 − vβ (α(t), t)2 dt.
a

Lorentzian length is independent of parametrization (cf. Lemma 3.4) and additive under concate-
nation of (causal) curves.
Remark 4.2. If γ = (α, β) : [a, b] → hI ×F X is a (future or past directed) causal curve, then

ℓ(dF ,h )h−1 (γ) ≤ 2|αb − αa |.
Indeed,
Z bs √ b √
vβ (αt , t)2
Z
ℓ(dF ,h )h−1 (γ) = α′ (t)2 + 2
dt ≤ 2 |α′ (t)|dt = 2|αb − αa |.
a h(γt ) a

We now define the causal structure and time separation function on Y .


Definition 4.3. Let p, q ∈ Y . We say that p and q are
(a) causally related, denoted p ≤ q, if there exists a future directed causal curve γ : p y q;
(b) timelike related, denoted p ≪ q, if there exists a future directed timelike curve γ : p y q;
A standard argument yields that ≤ is a preorder and ≪ is a transitive relation contained in ≤,
so that (Y, ≪, ≤) is a causal set, cf. [29, Definition 2.1]. The causal and timelike future and past of
a point p ∈ Y are denoted
I + (p) = {q ∈ Y : p ≪ q}, J + (p) = {q ∈ Y : p ≤ q}
I − (p) = {q ∈ Y : q ≪ p}, J − (p) = {q ∈ Y : q ≤ p},
and the chronological and causal diamonds I ± (p, q) := I + (p) ∩ I − (q), J ± (p, q) := J + (p) ∩ J − (q).
The time separation function is defined in a standard way.
Definition 4.4. The time separation function τ : Y × Y → R is defined as
τ (p, q) = sup τ (γ),
γ:pyq

where the supremum is taken over all causal curves joining p and q (with the convention sup ∅ = 0).
Remark 4.5. The time separation function satisfies the following two standard properties:
(1) If p ≪ q then τ (p, q) > 0;
(2) If p ≤ z ≤ q then τ (p, z) + τ (z, q) ≤ τ (p, q).
However we do not know at this stage whether τ is lower semicontinuous. We will establish this in
the next section assuming (1.4).
14 ELEFTERIOS SOULTANIS

4.2. Basic properties of causal curves. We first establish the stability of causal curves and the
upper semicontinuity of Lorentzian length.
Proposition 4.6. Suppose γj = (αj , βj ) : [a, b] → Y are causal curves converging uniformly to an
AC-curve γ = (α, β) : [a, b] → Y . If α′ > 0 (or < 0) then γ is a causal curve.
Proof. Without loss of generality we assume that α′ > 0 a.e. We will prove the following equivalent
form of the causality condition:
Z b′ Z b′
(4.2) vβ (α(τ ), τ )dτ ≤ h(γ(τ ))α′ (τ )dτ, a ≤ a′ < b′ ≤ b.
a′ a′
By Lemma 3.5 we have that
Z b′ Z b′
vβ (α(τ ), τ )dτ ≤ lim inf vβj (αj (τ ), τ )dτ.
a′ j→∞ a′
Fix ε > 0. The causality of γj yields
Z b′ Z b′ Z b′
vβj (αj (τ ), τ )dτ ≤ h(γj (τ ))α′j (τ )dτ ≤ (h(γ(τ )) + ε)α′j (τ )dτ
a′ a′ a′
for large enough j. Since α′j (t)dt → α′ (t)dt
weakly as measures we obtain that
Z b′ Z b′

(h(γ(τ )) + ε)αj (τ )dτ → (h(γ(τ )) + ε)α′ (τ )dτ.
a′ a′
Letting ε → 0 we obtain (4.2). 
Proposition 4.7. Lorentzian length is upper semicontinuous, i.e. if γj is a sequence of causal
curves converging uniformly to a causal curve γ then
lim sup τ (γj ) ≤ τ (γ).
j→∞

Proof. By passing to a subsequence we may assume that lim supj→∞ τ (γj ) is a limit. By the
invarience of reparametrization we may assume that γj (t) = (cj t, βj (t)), t ∈ [0, 1], and that cj → c,
βj → β (uniformly) and
Z 1q
τ (γ) = c2 h(ct, β(t))2 − vβ (ct, t)2 dt.
0
Since vβj (cj t, t) ≤ cj h(cj t, βj (t)) a.e. we may pass to a (further) subsequence such that gj :=
q
vβj (cj ·, ·) and fj := c2j h(cj ·, βj (·))2 − gj2 converge weakly in L2 ([0, 1]) to g and f , respectively.
The lower semicontinuity of the L2 -norm with respect to weak convergence implies that, for every
0 ≤ a < b ≤ 1 we have
Z b Z b Z b Z b
2 2 2
g dt ≤ lim inf gj dt, f dt ≤ lim inf fj2 dt
a j→∞ a a j→∞ a
and thus Z b Z b Z b
2 2
(g + f )dt ≤ lim inf (gj2 + fj2 )dt =c 2
h(ct, β(t))2 dt,
a j→∞ a a
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 15

By Lemma 3.5 we have that


Z b Z b Z b Z b
vβ (ct, t)dt ≤ lim inf vβj (cj t, t)dt = lim inf gj dt = gdt.
a j→∞ a j→∞ a a
Since a, b are arbitrary it follows that
vβ (ct, t)2 + h(t)2 ≤ g(t)2 + f (t)2 ≤ c2 h(ct, β(t))2 a.e. t ∈ [0, 1].
From this we obtain that
Z 1q Z 1 Z 1
τ (γ) = 2 2 2
c h(ct, β(t)) − vβ (ct, t) dt ≥ hdt = lim hj dt = lim τ (γj ).
0 0 j→∞ 0 j→∞


We single out the following corollary of Propositions 4.6 and 4.7.
Corollary 4.8. (a) Assume p, q ∈ I × X and J(p, q) is contained in a compact set. Then
J(p, q) is compact and, for every p′ , q ′ ∈ J(p, q), there exists a maximizing causal curve
γ : p′ y q ′ .
(b) Let p0 ∈ I × X and h > 0. If B (dF ,h )h−1 (p0 , 4h) is compact, then J(p, q) is compact for
every p, q ∈ B (dF ,h )h−1 (p0 , h).
Proof. It follows from Remark 4.2 that

(dF ,h )h−1 (p, q ′ ) + (dF ,h )h−1 (q ′ , q) ≤ 2(b − a)
for any q ′ ∈ J(p, q), whenever p = (a, x) and q = (b, y) are causally related. Assume J(p, q) is
contained in a compact set. To show that J(p, q) is compact it suffices to show it is closed. If
(qj′ ) ⊂ J(p, q) converges to q ′ , let γjp : p y qj′ and γjq : qj′ y q be causal curves parametrized
with affine first component. We have that the curves γjp and γjq lie inside a compact set and have
uniformly bounded length. Thus the sequences converge up to a subsequence to causal curves
γ p : p y q ′ and γq : q ′ y q, respectively. This shows that q ′ ∈ J(p, q), and thus J(p, q) is closed.
Moreover, if p′ , q ′ ∈ J(p, q) and γj : p′ y q ′ is a sequence of causal curves with τ (γj ) → τ (p′ , q ′ )
for p′ , q ′ ∈ J(p, q), then up to reparametrization a subsequence of (γj ) converges to a causal curve
γ : p′ y q ′ satisfying τ (p′ , q ′ ) = limj→∞ τ (γj ) ≤ τ (γ). Thus γ is a maximizing curve. We have
proved (a).
To prove (b) note that, if p, q ∈ B (dF ,h )h−1 (p0 , h), then |b − a| ≤ 2h, and thus

(dF ,h )h−1 (p0 , q ′ ) ≤ (dF ,h )h−1 (p0 , p) + (dF ,h )h−1 (p, q ′ ) ≤ (1 + 2 2)h < 4h.
This shows that J(p, q) ⊂ B(dF ,h )h−1 (p0 , 4h) for all p, q ∈ B (dF ,h )h−1 (p0 , h). By (a) J(p, q) is
compact. 

Variational length of causal curves. Recall the Lorentzian length (2.3) associated to the time sep-
aration function τ .
Our next proposition states that it agrees with (4.1).
Proposition 4.9. Let γ : [a, b] → Y be a causal curve. Then Lτ (γ) = τ (γ).
16 ELEFTERIOS SOULTANIS

Proof. The inequality τ (γ) ≤ Lτ (γ) is a straightforward consequence of the definitions. We will
prove the opposite inequality.
By reparametrization we may assume that γ = (id, β) : [a, b] → Y . For each t ∈ [a, b] and h > 0
such that t + h ∈ [a, b], let γt,h = (id, βt,h ) be a maximizing geodesic γt y γt+h . Using Hölder’s
inequality
Z t+h
2
τ (γt , γt+h ) ≤ h [h(s, βt,h (s)) − vβt,h (s, s)2 ]ds, or equivalently
t
)2
τ (γt , γt+h 1 t+h 1 t+h
Z Z
2
+ vβt,h (s, s) ds ≤ h(s, βt,h (s))2 ds.
h2 h t h t
For fixed t0 ∈ [a, b], the estimate
2 2
1 t+h
 Z t+h  Z t+h
1 1 1
Z
2
vβt,h (s, s) ds ≥ vβt,h (s, s)ds ≥ vβt,h (t0 , s)ds
h t h t (1 + ω(t − t0 ))2 h t
dt0 (βt , βt+h ) 2
 
1

(1 + ω(t − t0 ))2 h
implies (using a similar argument in case h < 0) that
τ (γt , γt+h )2 vβ (t0 , t)2
lim sup + ≤ h(t, β(t))2 a.e. t.
h→0 h2 (1 + ω(t − t0 ))2
Choosing a countable dense set of t0 ∈ [a, b] this easily implies that
τ (γt , γt+h ) 2
 
lim sup ≤ h(t, β(t))2 − vβ (t, t)2 a.e. t
h→0 h
Since Lτ (γ|[a,t+h] ) − Lτ (γ|[a,t] ) ≤ τ (γt , γt+h ), it follows that the function t 7→ Lτ (γ[a,t] ) is absolutely
continuous and its derivative squared is bounded a.e. from above by h(t, β(t))2 − vβ (t, t)2 . This
implies that Lτ (γ) ≤ τ (γ) and completes the proof. 
The following result will be useful in proving that future and past lightcones are open under the
assumption (1.4).
Proposition 4.10. Suppose γ0 : q0 y p0 is a future directed Lipschitz time-like curve with
|τγ′ 0 | ≥ c0 > 0 a.e. Then there exists c > 0 and open neighbourhoods U, V of q0 and p0 , respectively
such that for any q ∈ U and p ∈ V there exists a future directed time-like curve γ : q y p with
|τγ′ | ≥ c.
In the proof we need a technical lemma.
Lemma 4.11. Let J0 ⊂ I and K ⊂ X be compact. Suppose γ = (α, β) : J → J0 × K ⊂ hI ×F X
is a future directed curve such that
τγ′ (t)2 = h(αt , βt )α′ (t)2 − vβ (αt , t)2 ≥ c20 > 0
for a.e. t ∈ J, and
max{ sup h, |α(J)|−1 , kα′ kL∞ (J) , kvβ kL∞ (J0 ×J) } ≤ L.
J0 ×K
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 17

Then there exists δ0 = δ0 (L, J0 , K) with the following property. Whenever J ′ ⊂ J0 is a compact
subinterval with |J ′ △α(J)| < δ0 , and A : α(J) → J ′ is an increasing affine bijection, the perturbed
curve γA := (A ◦ α, β) satisfies τγ′ A (t) ≥ c0 /2 a.e. t ∈ J.
Proof. Let α(J) = [a, b] and J ′ = [a′ , b′ ]. Then
b′ − a′
A(s) = a + (s − a), s ∈ [a, b]
b−a
and δ := |α(J)△J ′ | = |a − a′ | + |b − b′ |. Observe that

sup |A(s) − s| ≤ 2δ and |(A′ )2 − 1| ≤
s∈α(J) |α(J)|
as long as δ ≤ |α(J)|. By (1.2) and Lemma 3.2 have that
h(A(αt ), βt )2 ≥ h(αt , βt )2 − ωh2 (2δ)
vβ (A(αt ), t)2 ≤ [vβ (αt , t) + kvβ kL∞ (J0 ×J) ω(2δ)]2 ≤ vβ (αt , t)2 + 3kvβ k2L∞ (J0 ×J) ω(2δ),
where ωh2 is the modulus of continuity of h2 |J0 ×K . Denoting ω̃ = ωh2 + ω we have the estimate

τγ′ A (t)2 =h(A(αt ), βt )2 (A′ )2 α′ (t)2 − vβ (A(αt ), t)2


≥(A′ )2 [h(αt , βt )2 − ωh2 (2δ)]α′ (t)2 − vβ (αt , t)2 − 3kvβ k2L∞ (J0 ×J) ω(2δ)

≥h(αt , βt )2 α′ (t)2 − kα′ k2L∞ (J) ( sup h)2 − 4kα′ k2L∞ (J) ω̃(2δ)
|α(J)| J0 ×K

− vβ (αt , t)2 − 3kvβ k2L∞ (J0 ×J) ω̃(2δ)


≥c20 − 3L5 δ − 7L2 ω̃(2δ)
The proof is completed by choosing δ0 small enough such that 3L5 δ − 7L2 ω̃(2δ) < c20 /2. 
Proof of Proposition 4.10. Let J0 ⊂ I and K ⊂ X be compact sets so that J0 × K contains a
neughbourhood of im(β). Write γ0 = (α0 , β0 ) : J → hI ×F X and let L be such that the metrics
{ds }s∈J0 are pairwise L-bi-Lipschitz on K, and
(4.3) ( inf h)−2 + sup h + |α(J)|−1 + kα′ kL∞ (J) + kvβ kL∞ (J0 ×J) ≤ L.
J0 ×K J0 ×K
Define
U = (a0 − δ, a0 + δ) × Bda (x0 , δ), V = (b0 − δ, b0 + δ) × Bdb (y0 , δ),
δ0
where δ ≤ is small enough so that U ∪ V ⊂ J0 × K; here δ0 is the constant given by Lemma
2(L2 +1)
4.11. Given q = (a, x) ∈ U and p = (b, y) ∈ V , let
βx : [a, a + Lda (x, x0 )] → X, βy : [b − Ldb (y, y0 ), b] → X
be constant speed parametrized geodesics x y x0 and y0 y y, with respect to da and db , respec-
tively. It follows that
1 1
(4.4) vβx (t, t) ≤ L|(βx )′t |da = , vβy (t, t) ≤ L|(βy )′t |db ≤
L L
18 ELEFTERIOS SOULTANIS

a.e. t. By (4.3) and (4.4) we have that


 2
1 1 1 1
h(t, βx (t)) − vβx (t, t) ≥ − L2 ·
2 2
= − 2 , a.e. t ∈ [a, a + L2 da (x0 , x)],
L L L L
 2
1 1 1 1
h(t, βy (t))2 − vβy (t, t)2 ≥ − L2 · = − 2 a.e. t ∈ [b − L2 db (y0 , y), y].
L L L L
Thus the curves γx = (id, βx ) and γy = (id, βy ) satisfy
r r
′ 1 1 ′ 1 1
τγx ≥ − 2 , τγy ≥ − 2 a.e.
L L L L
Oberve that γx : q y q̃0 and γy : p̃0 y p, where
q0 = (a + L2 da (x, x0 ), x0 ), p0 = (b − L2 db (y0 , y), y0 ).
Choosing J ′ = [a + L2 da (x, x0 ), b − L2 db (y0 , y)], we note that
|[a0 , b0 ]△J ′ | = |a + L2 da (x0 , x) − a0 | + |b − L2 db (y0 , y) − b0 | < 2δ + 2L2 δ = δ0 .
Thus we may apply Lemma 4.11 to obtain that γA := (A ◦ α0 , β0 ), where A : [a0 , b0 ] → J ′ is the
increasing affine bijection, satisfies τγ′ A ≥ c0 /2. Since τA : q̃0 y p̃0 . The proof is now complete
because the concatenation γ := γy γA γx : p y q is Lipschitz and satisfies
( r )
′ c0 1 1
τγ ≥ min , − a.e.
2 L L2

Remark 4.12. Define a strict time-like relation by setting q ≪s p if there exists a causal Lipschitz
curve γ : q y p such that τγ′ ≥ c for some c > 0. Then Proposition 4.10 shows that the corre-
sponding future (and past) cones Is± (p) are open assuming only (1.2) (typically (1.4) is needed to
guarantee openness of time-like future cones). In general Lorentzian length spaces ≪s need not
coincide with ≪, and we do not know whether they can differ in our setting. In Section 5 we will
show that ≪=≪s in hI ×F X assuming (1.4).
4.3. Lorentzian manifolds. Assume that X = M is a smooth manifold, h : I × M → (0, ∞) is
continuous, and {gs }s∈I is a collection of continuous Riemannian metrics on M such that the local
coordinate representations (s, p) 7→ (gs )ij (p) are continuous in (s, p).
Remark 4.13. It is standard that the continuity assumption above is equivalent to the following:
(s, p) 7→ gs (Xp , Yp ) is continuous for any smooth vector fields X, Y ∈ C ∞ (T M ).
It follows that
(4.5) g := −h2 ds2 + gs
is a continuous Lorentzian metric on I × M . Thus (I × M, g) is a continuous space-time and we
denote by (I × M, ≤g , ≪g , τg ) the induced Lorentzian length structure.
Now consider the family of length metrics F = {dgs }s∈I on M associated to {gs }s∈I .
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 19

Lemma 4.14. Under the hypotheses above, F satisfies (1.2). Moreover, for any absolutely contin-
uous curve β : [a, b] → X there exists a null set N ⊂ [a, b] such that
(4.6) vβ (s, t)2 = gs (βt′ , βt′ ), (s, t) ∈ I × ([a, b] \ N ).

Proof. Let J ⊂ I and K ⊂ M be compact, and let g̃ = ds2 + gs be the Riemannian metric on
I × M . Set d˜s = dg̃ |{s}×X . Choose r > 0 so that B dg̃ (J × K, r) is compact. By the continuity
p
of the metric, there exists a modulus of continuity ω : [0, ∞) → [0, ∞) such that gs (v, v) ≤
gs′ (v, v) + ω(|s − s′ |) gs′ (v, v) for all s, s′ ∈ Jr := prI (B dg̃ (J × K, r)) and v ∈ Tp M with
p p

p ∈ Kr := prX (B dg̃ (J × K, r)). In particular if x, y ∈ K and inf s∈J dg̃ ((s, x), (s, y)) < r, then
ds (x, y) ≤ (1 + ω(|Jr |)))d˜s′ (x, y). Thus the metrics ds , d˜s′ are locally bi-Lipschitz equivalent for
s, s′ ∈ J with uniform bi-Lipschitz constant C = C(J, K). For x, y ∈ K with inf s∈J ds (x, y) < r/C,
we have that any ds -geodesic β : x y y lies in K for any s ∈ J, and thus
Z 1q Z 1 q
ds′ (x, y) ≤ℓds′ (β) = ′ ′
gs′ (βt , βt )dt ≤ (1 + ω(|s − s′ |)) gs (βt′ , βt′ )dt
0 0

≤(1 + ω(|s − s |))ds (x, y).
This proves (1.2).
To prove the second claim no that, since β is absolutely continuous, βt′ exists for a.e. t. Let
N ⊂ [a, b] be a null-set such that every t ∈ [a, b] \ N is a Lebesgue point of t 7→ βt′ . For fixed s ∈ I
we have that |βt′ |2dgs = gs (βt′ , βt′ ) a.e. t. In particular for (s, t) ∈ I × ([a, b] \ N ) we have that

1 t+h p 1 t+h ′
q Z Z
′ ′
gs (βt , βt ) = lim ′ ′
gs (βτ , βτ )dτ = lim sup |βτ |dgs dτ = vβ (s, t)
h→0 h t h→0 h t


In particular, the length element of g agrees with (1.3) a.e. on absolutely continuous curves.
Corollary 4.15. If γ = (α, β) is an absolutely continuous curve, we have that
g(γt′ , γt′ ) = −|τγ′ |2 (t) = −h(γt )α′ (t)2 + vβ (αt , t)2 a.e. t.
We now give the proof of Corollary 1.1.
Proof of Corollary 1.1. By Corollary 4.15 we have that a curve γ in I × M is causal with respect
to (I × M, ≤g , ≪g , τg ) if and only if it is causal in hI ×F M . Moreover, in this case τg (γ) = τ (γ)
and thus τg = τ . This completes the proof. 
Next we verify (1.4) for Lipschitz metrics.
Corollary 4.16. Suppose that g in (4.5) is Lipschitz (i.e. h and the local coordinate representations
(s, p) 7→ (gs )ij (p)) are locally Lipschitz). Then F = {dgs }s∈I satisfies (1.4).
Proof. Observe that, assuming (s, p) 7→ gij (p) is Lipschitz instead of merely continuous, we may
choose ω(a) = Ca for some C = C(J, K) in the proof of (1.2) in Lemma 4.14. This directly implies
(1.4). 
20 ELEFTERIOS SOULTANIS

5. Lorentzian length space structure of hI ×F X


In this section we prove Theorems 1.3 and 1.5 as well as Corollary 1.4.
5.1. Modifying causal curves via Caratheodory theory of ODE’s. Throughout this subsec-
tion we assume (1.4). We prove Theorem 1.3 by reducing it to the two dimensional case and using
Caratheodory theory of ODE’s via the following proposition. We refer the reader to Appendix A
for the results of Caratheodory theory of ODE’s needed below.
Proposition 5.1. Suppose γ = (α, β) : q y p is a causal curve with τ (γ) > 0 defined on an interval
J. Then there exists an absolutely continuous increasing function y : J → I with y(J) = α(J) such
that
h(yt , βt )2 y ′ (t)2 − vβ (yt , t)2 = c ∈ (0, τ (q, p)]
for all t ∈ J.
Remark 5.2. (a) In particular (y, β) is a time-like curve with the same start and end-points
as γ, and
kvβ kL∞ (α(J)×J) + τ (γ)
(b) the Lipschitz constant of y is at most (which can be +∞ if β is not
inf f
α(J)×β(J)
Lipschitz).
Proof. Without loss of generality we may assume that J = [0, 1]. Denote γ0 = (a, x) and γ1 = (b, y).
Given ε ∈ [0, τ (γ)] we define Φε : I × [0, 1] → R by
(vβ (s, t)2 + ε2 )1/2
Φε (s, t) = .
h(s, βt )
By (1.4), (1.5) and Lemma 3.2 there exists L ∈ L1 (0, 1) such that
|Φε (s, t) − Φε (s′ , t)| ≤ L(t)|s − s′ |, s, s′ ∈ im(α),
i.e. Φε is Caratheodory-Lipschitz (uniformly in ε ∈ [0, τ (γ)]). Theorem A.1 implies that the ODE
 ′
y = Φε (y, t)
y(0) = a
admits a unique absolutely continuous solution yε which varies continuously with ε. For each ε the
curve γε = (yε , β) satisfies

h(yε (t), βt )2 yε′ (t)2 − vβ (yε (t), t)2 = ε2


and τ (γε ) = ε. We show that
(5.1) y0 (1) < b and yτ (q,p)(1) ≥ b.
Since
vβ (y0 (t), t)
y0′ − = y0′ − Φ0 (y0 , t) = 0 ≤ α′t − Φ0 (α, t),
h(y0 (t), βt )
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 21

the comparison principle for ODE’s (Theorem A.2) there exists c ∈ [0, 1] such that y0 (t) = α(t)
whenever t ∈ [0, c] and y0 (t) < α(t) for t ∈ (c, 1]. If c = 1, then y0 = α which implies γ = γ0 ,
contradicting the fact that τ (γ) > 0. Thus we have that y0 (1) < α(1) = b. To finish the proof
of (5.1) suppose that yτ (q,p) (1) < b. Then the concatenation γ̃ of γτ (q,p) and the vertical segment
[yτ (q,p) (1), b]× {y} is a time-like curve connecting q and p with Lorentzian length τ (γ̃) = τ (γτ (q,p) )+
(b − yτ (q,p) ) > τ (q, p), which is a contradiction.
We conclude the proof of the proposition using (5.1). Since yε (1) varies continuously in ε, (5.1)
implies that there exists εb ∈ (0, τ (q, p)] for which yεb (1) = b. Thus y = yεb is the required
function. 
5.2. hI ×F X is a Lorentzian length space.
Proof of Theorem 1.3. Suppose q ≤ p ≤ r. By the reverse triangle inequality (cf. Remark 4.5(2))
τ (q, r) ≥ τ (q, p) + τ (p, r). If q ≪ p or p ≪ r, one of the summands is positive and thus τ (q, r) > 0.
It follows that there exists a causal curve γ : q y r with τ (γ) > 0. Proposition 5.1 yields the
existence of a time-like curve p y r, showing that p ≪ r. This completes the proof. 
Next we focus on Corollary 1.4. To establish it we need to show the lower semicontinuity of τ .
This, in turn will be an easy consequence of the openness of the relation ≪.
Proposition 5.3. Suppose q0 , p0 ∈ hI ×F X, q0 ≪ p0 . Then there exist neighbourhoods U and V
of q0 and p0 , respectively, such that q ≪ p whenever q ∈ U and p ∈ V .
In particular, I ± (p) is open for any p ∈ hI ×F X.
Proof. The claim follows from Proposition 4.10 once we establish the existence of a Lipschtiz curve
γ0 : q0 y p0 with τγ′ 0 ≥ c for some c > 0. To this end let γ = (α, β) : q0 y p0 be a causal curve
with τ (γ) > 0 parametrized so that α and β are Lipschitz. Then γ0 = (y, β), where y is given by
Proposition 5.1, satisfies the required properties, and thus the proof is complete. 
Proof of Corollary 1.4. We have that (hI ×F X, ≤, ≪) is a causal set which satisfies the push-up
property (Theorem 1.3). Moreover τ satisfies the reverse triangle inequality (Remark 4.5).
We use the argument in the proof of [1, Lemma 3.25] to show that τ is lower semicontinuous.
Let p, q ∈ hI ×F X, ε > 0 be arbitrary and assume without loss of generality that τ (p, q) > 0.
Let γ : [a, b] → hI ×F X be a causal curve p y q with τ (γ) > 0 such that τ (γ) > τ (q, p) − ε/3.
We may find a < t1 ≤ t2 < b such that 0 < τ (γ|[a,t1 ] ), τ (γ[t2 ,b] ) < ε/3. Setting U := I − (γ(t1 ))
and V := I + (γ(t2 )) we have that U and V are open and (p, q) ∈ U × V , cf. Proposition 5.3. For
(p̃, q̃) ∈ U × V there are timelike curves p̃ y γ(t1 ) and γ(t2 ) y q̃ and their concatenation with
γ|[t1 ,t2 ] is a causal curve p̃ y q̃. Thus
τ (p̃, q̃) ≥ τ (γ|[t1 ,t2 ] ) > τ (γ) − ε/2 − ε/2 = τ (p, q) − ε.
This proves the claimed lower semicontinuity. Thus we have that hI ×F X is a causally path
connected Lorentzian pre-length space. Proposition 4.6 and Corollary 4.8 implies that hI ×F X is
locally causally closed. For any p = (a, x) ∈ hI ×F X there exists h > 0 such that [a − h, a + h] ⊂ I,
and the chronological diamond Ip = ((a − h, x), (a + h, x)) is a localizing neighbourhood of p,
cf. Corollary 4.8. Finally, Proposition 4.9 implies that τ (q, p) = supγ:qyp Lτ (γ), so that it is a
Lorentzian length space. 
22 ELEFTERIOS SOULTANIS

Before we consider the global hyperbolicity of the generalized Lorentzian product, we record here
the fact that it is always strongly causal.
Proposition 5.4. hI ×F X strongly causal: chronological diamonds I(p, q) form a basis of the
topology of I × X.
Proof. Every neighbourhood U of a point p = (a, x) contains the chronological diamond I((a −
h, x), (a + h, x)) ∋ p for small enough h > 0 by Corollary 4.8 and Remark 4.2. Since chronological
diamonds are open, it follows that they form a basis of the topology of hIF X. 

5.3. Global hyperbolicity of hI ×F X. Recall the metric (dF ,h )h−1 on I × X from Section 3.2.
Global hyperbolicity is connected to the properness of this metric.
Lemma 5.5. Let J ⊂ I be compact and denote FJ = {ds }s∈J . If (J × X, (dFJ ,h )h−1 ) is proper,
then every causal diamond J(p, q) ⊂ hI ×F X with p, q ∈ J × X is compact.
Proof. Let p = (a, x) and q = (b, y) ∈ J × X. Since (J × X, (dFJ ,h )h−1 ) is proper there exists a
compact
√ set K ⊂ X such that any absolutely continuous curve starting at p with length at most
2(b − a) lies in ′ ′
√ J × K. By Remark 4.2 every causal curve γ : p y q with q ∈ J(p, q) has
length at most 2(b − a), and thus it follows that J(p, q) ⊂ J × K. By Corollary 4.8(a) J(p, q) is
compact. 
Since dFJ ,h is a locally compact length metric on J × X, Lemma 2.1 immediately yields the
following condition for properness.
Corollary 5.6. Suppose h : I × X → (0, ∞) is continuous and F satisfies (1.2). If J ⊂ I is
compact and
Z ∞
vβ (αt , t)
(5.2) dt = ∞
0 h(γt )
whenever γ = (α, β) : [0, ∞) → J × X is a locally absolutely continuous curve tending to infinity,
then (J × X, (dFJ ,h )h−1 ) is a proper metric space.
Proof of Theorem 1.5. It follows from Remark 4.2 that hI ×F X is totally non-imprisoning. By
(1.6) and Corollary 5.6 we have that (J × X, (dFJ ,h )h−1 ) is proper for each compact subinterval
J ⊂ I. By Lemma 5.5 causal diamonds in hI ×F X are compact, and thus hI ×F X is globally
hyperbolic. 

5.4. Regularity of conformal generalized Lorentzian products. Throughout this subsection


we assume that d is a given locally compact length metric on X, and that h, ρ : I × X → (0, ∞)
are locally Lipschitz functions. For each s ∈ I, consider the conformally weighted length metric
ds := dρ(s,·) on X, satisfying
|βt′ |ds = ρ(s, βt )|βt′ |d a.e. t
for every β ∈ AC(X, d). Set F = {ds }s∈I . It follows from (2.2) that
(5.3) vβ (s, t) = ρ(s, βt )vβ (t), (s, t) ∈ I × [a, b]
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 23

for every absolutely continuous β : [a, b] → X. Here


1 t+h ′
Z
vβ (t) = lim sup |βu |d du, t ∈ [a, b].
h→0 h t

Lemma 5.7. The family F defined above satisfies (1.4).


Proof. Let J ⊂ I and K ⊂ X be compact. Let r0 = r0 (K) > 0 be such that K ′ = B d (K, r0 ) is
compact, and set A = A(J, K) = inf J×K ′ ρ. Observe that, if s ∈ J and β is a curve in X such that
ℓds (β) < Ar0 , then β lies in K ′ .
Now suppose x, y ∈ K and inf ds (x, y) < Ar0 . Then there is s0 ∈ J such that ds0 (x, y) < Ar0 . By
s∈J
the observation above any sequence (βj ) of constant speed curves x y y with ℓds0 (βj ) → ds0 (x, y)
eventually lies in K ′ and thus subconverges to a ds -geodesic β : x y y lying in K ′ . It follows that
Z Z
ds (x, y) ≤ ρ(s, βt )|βt |dt ≤ ds0 (x, y) + |ρ(s, βt ) − ρ(s0 , βt )||βt′ |dt

 LIP(ρ|J×K ′ ) 
≤ds0 (x, y) 1 + |s − s0 |
A
LIP(ρ|J ×K ′ )
for any s ∈ J. Set C = A and r = Ar0 /(1 + C|J|). In particular if inf ds (x, y) < r0 , then
s∈J
ds (x, y) < Ar0 for all s ∈ J, and the argument above yields
ds′ (x, y)
≤ 1 + C|s − s′ |, x, y ∈ K, s, s′ ∈ J,
ds (x, y)
proving (1.4) 
We prove Theorem 1.6 as a corollary of the following result.
Theorem 5.8. Let γ : [a, b] → hI ×F X be a maximizing causal curve with Lτ (γ) > 0. Then γ is
time-like, i.e. τγ′ (t) > 0 a.e. t ∈ [a, b].
Since reparametrization does not change causal character, we may assume that γ = (α, β)
parametrized such that |βt′ |d = c a.e. for some constant c > 0, and α is Lipschitz. Consider
the space Q = α([a, b]) × [a, b] ⊂ R2 equipped with with the Lorentzian metric
gQ = −h(s, βt )2 ds2 + vβ (s, t)2 dt2 .
Note that vβ (s, t) = cρ(s, βt ), cf. (5.3).
Lemma 5.9. The Lorentzian metric gQ is Lipschitz.
Proof. Denote F (s, t) = h(s, β(t))2 . Since β and h2 |α([a,b])×β([a,b]) are Lipschitz, it follows that −F
is Lipschitz. Similarly (s, t) 7→ vβ (s, t)2 = c2 ρ(s, βt )2 is Lipschitz. 
Lemma 5.10. A future directed curve y = (y0 , y1 ) in Q is causal with respect to gQ if and only if
γy := (y0 , β ◦ y1 ) is causal in hI ×F X. In this case we have
LgQ (y) = τ (γy ).
24 ELEFTERIOS SOULTANIS

Proof. By Lemma 3.4 we have that


gQ (yt′ , yt′ ) = − h(y0 (t), β(y1 (t)))2 y0′ (t)2 + vβ (y0 (t), y1 (t))2 y1′ (t)2
= − h(y0 (t), β ◦ y1 (t))2 y0′ (t)2 + vβ◦y1 (y0 (t), t)2 = −τγ′ y (t)2
a.e. t from which the first claim immediately follows. It also follows that, if y is causal, then
Z q Z
LgQ (y) = −gQ (yt , yt )dt = τγ′ y (t)dt = τ (γy ).
′ ′


Corollary 5.11. For any t, t′ ∈ [a, b] with t ≤ t′ we have that x : [t, t′ ] → Q, x(λ) = (αλ , λ) is a
maximizing curve, and
τQ (x(t), x(t′ )) = τ (γ|[t,t′ ] ).
Proof. We use the notation of Lemma 5.10. Note that γx = γ|[t,t′ ] and thus LgQ (x) = τ (γ|[t,t′ ] ). If
x were not maximal, there would exist a causal curve y : [t, t′ ] → Q with the same end points and
τ (γy ) = LgQ (y) > LgQ (x) = τ (γ|[t,t′ ] ). Since γy is a causal curve with the same endpoints as γ|[t,t′ ] ,
this would contradict the maximality of γ. 

Proof of Theorem 5.8. Consider the Lorentzian manifold (Q, gQ ) and note that x = (α, id) : [a, b] →
Q is a maximizing curve (αa , a) y (αb , b), cf. Corollary 5.11. Recall that γx = γ (using the notation
of Lemma 5.10). Since LgQ (x) = τ (γ) > 0 and gQ is Lipschitz, it follows that (αa , a) ≪Q (αb , b)
and that x is time-like, cf. [Lorentz-meets-Lipschitz, Prop 1.2]. In particular
0 < −gQ (x′t , x′t ) = h(αt , βt )2 α′ (t)2 − vβ (αt , t)2 a.e. t,
showing that γ = (α, β) is time-like. 

Proof of Theorem 1.6. Since hI ×F X is a Lorentzian length space the only remaining claim to
establish is that every point has a localizing neighbourhood where maximizing curves with positive
τ -length are time-like. Indeed, given p = (a, x) ∈ I × X, the localizing neighbourhood Up =
I((a − h, x), (a + h, x)) (for small enough h > 0) satisfies the locally regularizable condition (iv) in
Section 2.2 by Theorem 5.8. This completes the proof. 

6. Synthetic curvature bounds


We apply the synthetic time-like Ricci curvature bounds defined in Section 2.4 to the Lorentzian
length structure hI ×F X, where F satisfies (1.2), and h : I → (0, ∞) is a continuous function of
one variable.
Let M = {ms }s∈I be a family of measures as in Section 1.5 satisfying (1.7), and consider the
measure m = ms ⊗ hds defined on the generalized Lorentzian product hI ×F X by
Z Z Z
F (s, x)dm(s, x) = h(s) F (s, x)dms (x)ds
I X
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 25

for Borel functions


Z F : I × X → [0, +∞]. Let G : I × X → (0, ∞), with G and G−1 locally bounded,
be such that Gdm = 1, and define m = mG ∈ P(X) by
Z
m(E) = Gdm, E ⊂ X.
I×E

Let g be the Radon–Nikodym derivative of m with respect to hL1 ⊗ m, i.e. m = g · hL1 ⊗ m. Using
(1.7) we see that, for each compact J ⊂ I and K ⊂ X, there exists a constant C = C(J, K) > 0
with
(6.1) C −1 ≤ g ≤ C m − a.e. on J × K.
In what follows we choose a Borel representative of g such that this holds pointwise everywhere for
all J, K.
Remark 6.1. For each compact J ⊂ I and K ⊂ X there exists C = C(J, K) > 0 such that
C −1 ms |K ≤ m|K ≤ Cms |K for every s ∈ J.
In the next theorem we denote by | · |h the metric on I given by
Z t
|t − s|h = h(τ )dτ, s ≤ t.
s

Theorem 6.2. Suppose F and M satisfy (1.2) and (1.7), respectively, h : I → (0, ∞) is continuous,
and assume that (hI ×F X, m) satisfies the wTCDp (K, N )-condition. For any compact interval
1 -representative g on J that
J ⊂ I we have that, for m-a.e. x ∈ X, g(·, x) has a (continuous) H|·| h
x
satisfies
(6.2) gx (at )1/N ≥ σK/N
1−t
(|b − a|h )gx (a)1/N + σK/N
t
(|b − a|h )gx (b)1/N , a, b ∈ J, t ∈ [0, 1],
where at : [0, 1] → J is the | · |h -geodesic a y b.
1 = hL1 . Moreover, denoting L = khk
Observe that H|·| 1
h L1 (J) = H|·|h (J), the function
Z
H : J → [0, L], H(t) := hdτ,
J∩(−∞,t]
1 ) = g ◦ H −1 L1 . Since in
defines an isometric bijection (J, | · |h ) → ([0, L], | · |) with H∗ (gH|·| h
the 1-dimensional situation ([0, L], | · |, gL1 ), the RCD(K, N )-condition is characterized by (6.2)
(see [25, Theorem 1.1]), Theorem 6.2 has the following immediate corollary.
Corollary 6.3. Under the hypotheses of Theorem 6.2, for m-a.e. x ∈ X, we have that
1 ) is an RCD(K, N )-space for every compact interval J ⊂ I.
Jx = (J, | · |h , g(·, x)H|·| h

We set up notation for the proof of Proposition 6.2. Let ν0 , ν1 ∈ P(J) and let t 7→ νt be the
(unique) geodesic with respect to the Wasserstein distance
( Z 1/2 )
1
(6.3) Wh (ν, σ) = inf |x − y|2h dπ(x, y) : π ∈ Π(ν, σ) .
2 J×J
26 ELEFTERIOS SOULTANIS

Recall that t 7→ νt is represented by an optimal dynamical plan η ∈ P(Geo(J, | · |h )), i.e νt = et∗ η,
and that the optimal pairing in (6.3) is given by π = (e0 , e1 )∗ η. By [36, Theorem 2.9] π (and
thus η) is optimal for any cost c(x, y) = f (|y − x|h ) with f : R → R strictly convex. Therefore
C := sptπ is c-cyclically monotone for any such c. Choosing f (t) = Lp − |t|p for p ∈ (0, 1) we
obtain the following reverse inequality in cyclic monotonicity.
k k
|ti − si |ph ≥ |ti − sσ(i) |ph ,
X X
(6.4) p ∈ (0, 1)
i i

for all (t1 , s1 ), . . . , (tk , sk ) ∈ C and every permutation σ of {1, . . . , k}.


Remark 6.4. By the same argument we have that
Z Z
|α1 − α0 | dη ≥ |α1 − α0 |p dη̃
p

for any p ∈ (0, 1) and any η̃ ∈ P(C([0, 1], J)) with ei∗ η̃ = νi , i = 0, 1, with equality if and only if
η̃ = η.
Lemma 6.5. Let ν0 , ν1 ∈ P ac (J, H|·|
1 ) be such that η ∈ OptGeo(ν , ν ) is concentrated on increas-
h
0 1
ing curves, and let B ⊂ X be a bounded Borel set. Define
m|B
µB
t := νt × .
m(B)
Then (µB B ac B
0 , µ1 ) ∈ P (hI ×F X, m) is time-like p-dualizable and t 7→ µt is the unique ℓp -geodesic
B B
between µ0 and µ1 . Moreover
Z 1/p
B B p
(6.5) ℓp (µ0 , µ1 ) = |α1 − α0 |h dη(α) .

In particular (µB B
0 , µ1 ) is strongly time-like p-dualizable.

Proof. The curve t 7→ µB B


t is represented by η ∈ P(OpT(hI ×F X)), defined by
Z Z
F (γ)dη B (γ) := − F ((α, x))dη(α)dm(x), F : C([0, 1]; I × X) → [0, ∞].
B

Consequently πB := (e0 , e1 )∗ηB


∈ Π≪ (µ0 , µ1 ). We show that C := sptπ B is ℓp -cyclically monotone.
Indeed, if (pj , qj ) ∈ C for j = 1, . . . , N then pj = (tj , xj ) and qj = (sj , xj ) (so that pj and qj have
the same x-component) and we have the estimate
N N N N
|tj − sσ(j) |ph ≤ |tj − sj |ph =
X X X X
ℓp (pj , qσ(j) ) ≤ ℓp (pj , qj )
j j j j

for any permutation σ by (6.4). This shows that πB


is optimal, and a direct calculation yields
(6.5).
To show uniqueness suppose π ′ ∈ Π≪ (µB B B B
0 , µ1 ) is an optimal coupling between µ0 and µ1 , and
′ ′
let η ∈ P(OpT(hI ×F X)) represent π . By Remark 6.4 we have that
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 27

Z Z Z Z
ℓp (µB B p
0 , µ1 ) = Lτ (γ)p dη ′ ≤ |p1 (γ1 ) − p1 (γ0 )|ph dη ′ = |α1 − α0 |ph dp1∗ η ′ ≤ |α1 − α0 |ph dη.

Since by (6.5) the last inequality is an equality, Remark 6.4 implies that p1∗ η ′ = η. In particular
sptπ ′ = sptπ B . Since for every (p, q) ∈ sptπ B there is a unique time-like geodesic p y q (the
vertical line with constant X-component), it follows that η ′ = η. 
Remark 6.6. In particular OpTp (µB B B
0 , µ1 ) = {η }.

In the notation of Lemma 6.5, define


eB (t) := Entm (µB
t ), ex (t) := Entλx (νt ), t ∈ [0, 1],
1 .
where λx = g(·, x)H|·| h

Lemma 6.7. For m-a.e. x ∈ X we have that


Z
(6.6) eB (t) = − ex (t)dm − log m(B).
B
1 and set
Proof. Let {ρ̃t }t∈[0,1] be Borel functions on J such that νt = ρ̃t H|·| h

ρ̃t (s)
ρt (s, x) := ,
g(s, x)
so that νt = ρt (·, x)λx for m-a.e. x ∈ X. Since m = gH|·| 1 ⊗ m, we have that
h
B χJ×B
µt = ρt m. Thus
m(B)
ρt ρt
Z
eB (t) = log dm
J×B m(B) mJ (B)
Z Z Z Z
=− ρt log ρt dλx dm(x) − − ρt log m(B)dλx dm(x)
B J B J
Z
=− ex (t)dm(x) − log m(B),
B
completing the proof. 

Assume now that hI ×F X satisfies the wTCDp (K, N )-condition for some K ∈ R, N ≥ 1 and
p ∈ (0, 1). It follows from the discussion above that eB satisfies
(6.7) e−eB (t)/N ≥ σK/N
1−t
(D)e−eB (0)/N + σK/N
t
(D)e−eB (1)/N , t ∈ [0, 1],
where
Z Z 1/2
2
D := kτ kL2 (πB ) = − |α1 − α0 |h dη(α)dm(x) = Wh (ν0 , ν2 ).
B
28 ELEFTERIOS SOULTANIS

Lemma 6.8. Let ν0 , ν1 ∈ P ac (J, H|·|


1 ) be as in Lemma 6.5. For m-a.e. x ∈ X, the function
f
ex : [0, 1] → R satisfies
1−t
(6.8) e−ex (t)/N ≥ σK/N (D)e−ex (0)/N + σK/N
t
(D)e−ex (1)/N , t ∈ [0, 1].
Proof. Let E ⊂ X be a m-null set such that whenever x ∈ X \ E we have that
Z
ex (t) = lim − ey (t)dmJ (y)
r→0 B(x,r)

for a countable dense set of t’s. (Recall that the measures ms are infinitesimally doubling, and
therefore m is a Vitali measure, cf. [24, Theorem 3.4.3]). Estimate (6.7) yields
R 1−t R t R
1 −1− e (t)dm
σK/N (D) − 1 − e (t)dm
σK/N (D) − 1 − e (t)dm
e N B(x,r) y ≥ e N B(x,r) y + e N B(x,r) y
m(B(x, r)) m(B(x, r)) m(B(x, r))
for all t ∈ [0, 1] and r > 0, cf. (6.6). Multiplying by m(B(x, r)) and letting r → 0 we obtain
the claimed inequality for a countable dense set of t’s. Since, for every x ∈ X \ E, ex (as well as
s
s 7→ σK/N (D)) is continuous, the inequality follows for all t ∈ [0, 1], completing the proof of the
claim. 
Proof of Theorem 6.2. Let a, b ∈ intJ and ãt := H(a) + t(H(b) − H(a)) for t ∈ [0, 1]. For small
δ > 0, the curve in t 7→ ν̃tδ in P([0, L]) given by
χ[ãt −δ,ãt +δ] 1
ν̃tδ := L |[0,L]

is a W2 -geodesic, and consequently νtδ := h−1 δ
∗ ν̃t is a geodesic in P(J) with respect to Wf . We have
δ δ δ δ
Wf (ν0 , ν1 ) = W2 (ν̃0 , ν̃1 ) = |H(b) − H(a)| = |b − a|h . Observe that
χB|·| (at ,δ) χB|·| (at ,δ)
νtδ = h 1
H|·| = h
λx ,
2δ h 2δg(·, x)
where at := H −1 (ãt ) is the | · |h -geodesic a y b. Thus
1 1 1
Z Z
δ δ 1
ex (t) := Entλx (νt ) = log dλx = −− log g(·, x)dH|·| + log .
B|·| (at ,δ) 2δg(·, x) 2δg(·, x) 2δ
h
h
B|·| h
(at ,δ)

By Lemma 6.8 we may find a m-null set E ⊂ X such that, if x ∈ X \ E, eδx satisfies (6.8) for all
a, b ∈ intJ ∩ Q with a < b and all δ ∈ Q+ .
For δ ∈ Q+ and x ∈ X \ E, define
!
1
Z
(6.9) uδx (a) := exp − 1
log g(·, x)dH|·| , a ∈ Jδ := (inf J + δ, sup J − δ).
N B|·| (a,δ) h
h
δ
For each a, b ∈ Jδ ∩ Q with a < b we have that uδx (at ) := 2δe−ex (t)/N satisfies (6.8) and by continuity
the same holds true for all a, b ∈ Jδ with a < b. It follows that, for δ ∈ Q+ and x ∈ X \ E, uδx
satisfies
uδx (at ) ≥ σK/N
1−t
(|b − a|h )uδx (a) + σK/N
t
(|b − a|h )uδx (b)
whenever a, b ∈ Jδ , t ∈ [0, 1].
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 29

Now on the one hand, for m-a.e. x ∈ X \ E, we have


lim uδx (a) = gJ (a, x)1/N a.e. a ∈ J,
δ→0+

while on the other hand, up to a subsequence, uδx converges locally uniformly as δ → 0+ to a


continuous function ux : J → R satisfying
1−t t
ux (at ) ≥ σK/N (|b − a|h )ux (a) + σK/N (|b − a|h )ux (b), a, b ∈ J, t ∈ [0, 1],

whenever x ∈ X \ E. Thus for m-a.e. x ∈ X we have that uN 1


x is a H|·|h -representative of g(·, x)
and satisfies (6.2). 
Proof of Corollary 1.7. Suppose that I = R, h ≡ 1, and K = 0. Let Jk = [−k, k]. By Theorem
1 ) is an RCD(0, N )-space and g 1/N is
6.2 there is a set E ⊂ X with m(E) = 0, and (Jk , | · |, gx H|·| h
x
1/N
concave in Jk for all k, whenever x ∈
/ E. Thus gx is a non-negative concave function on R, and
therefore it must be constant, which we denote by c(x). It follows that
Z Z Z
ms (E)ds = m(J × E) = g(s, x)dsdm(x) = |J| c dm
J J×E E
for any compact J ⊂ I and Borel E ⊂ X. The claim readily follows from this. 

Appendix A. Caratheodory theory of ODE’s


Let D ⊂ R2 be open. We say that Φ : D → R satisfies the Carathéodory condition if Φ(·, t)
is continuous for fixed t and f (s, ·) is measurable for fixed s. We say that Φ satisfies a (local)
Carathéodory-Lipschitz condition if in addition, for every compact I × J ⊂ D with I, J closed
intervals, there exists L ∈ L1 (J) such that
(A.1) |Φ(s, t) − Φ(s′ , t)| ≤ L(t)|s − s′ |, s, s′ ∈ I.
We have the following existence and uniqueness theorem.
Theorem A.1 (Theorem XVIII and XX, Ch. III §10 in [37]). If Φ is locally Carathéodory-Lipschitz
and (y0 , a) ∈ D, then the ODE
(A.2) y ′ = Φ(y, t)
with initial value y(a) = y0 has a unique solution which can be extended to the left and right (of a)
up to the boundary of D. In particular the unique solution varies continuously with y0 .
The following is a differential comparison theorem.
Theorem A.2 (Theorem XXI, Ch. III §10 in [37]). Suppose Φ is locally Carathéodory-Lipschitz
continuous, and that ϕ, ψ ∈ AC(J), J = [ξ, ξ + a] satisfy
(a) ϕ(ξ) ≤ ψ(ξ),
(b) P ϕ ≤ P ψ a.e. in J, where P φ = φ′ − Φ(φ, x).
Then either ϕ < ψ in J or there exists c ∈ [ξ, ξ + a] s.t. ϕ = ψ on [ξ, c] and ϕ < ψ in (c, ξ + a]. A
corresponding statement holds for the interval J ′ = [ξ − a, ξ] with the inequality reversed in (b).
30 ELEFTERIOS SOULTANIS

References
[1] S. Alexander, M. Graf, M. Kunzinger, and C. Sämann. Generalized cones as Lorentzian length spaces: Causality,
curvature, and singularity theorems. Preprint (arXiv:1909.09575), 2021.
[2] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré. Calculus and heat flow in metric measure spaces and appli-
cations to spaces with ricci bounds from below. Inventiones mathematicae, 195:289–391, 06 2014.
[3] Tobias Beran, Mathias Braun, Matteo Calisti, Nicola Gigli, Robert McCann, Argam Ohanyan, Felix Rott, and
Clemens Sämann. Untitled. In preparation, 2023.
[4] Tobias Beran, Argam Ohanyan, Felix Rott, and Didier A. Solis. The splitting theorem for globally hyperbolic
Lorentzian length spaces with non-negative timelike curvature. Lett. Math. Phys., 113(2):Paper No. 48, 47, 2023.
[5] Tobias Beran and Clemens Sämann. Hyperbolic angles in Lorentzian length spaces and timelike curvature bounds.
J. Lond. Math. Soc. (2), 107(5):1823–1880, 2023.
[6] Antonio N. Bernal and Miguel Sánchez. Smoothness of time functions and the metric splitting of globally
hyperbolic spacetimes. Comm. Math. Phys., 257(1):43–50, 2005.
[7] Mathias Braun. Good geodesics satisfying the timelike curvature-dimension condition. Nonlinear Anal., 229:Pa-
per No. 113205, 30, 2023.
[8] Dmitri Burago, Yuri Burago, and Sergei Ivanov. A course in metric geometry, volume 33 of Graduate Studies in
Mathematics. American Mathematical Society, Providence, RI, 2001.
[9] Annegret Burtscher and Leonardo Garcia-Heveling. Time functions on lorentzian length spaces. Preprint
arXiv:2108.02693, 2022.
[10] Annegret Burtscher and Leonardo Garcia-Heveling. Global hyperbolicity through the eyes of the null distance.
Preprint arXiv:2209.15610, 2023.
[11] Fabio Cavalletti and Andrea Mondino. Sharp and rigid isoperimetric inequalities in metric-measure spaces with
lower Ricci curvature bounds. Invent. Math., 208(3):803–849, 2017.
[12] Fabio Cavalletti and Andrea Mondino. A review of Lorentzian synthetic theory of timelike Ricci curvature
bounds. Gen. Relativity Gravitation, 54(11):Paper No. 137, 39, 2022.
[13] Fabio Cavalletti and Andrea Mondino. Optimal transport in lorentzian synthetic spaces, synthetic timelike ricci
curvature lower bounds and applications. Preprint arXiv:2004.08934, 2023.
[14] Piotr T. Chruściel and James D. E. Grant. On Lorentzian causality with continuous metrics. Classical Quantum
Gravity, 29(14):145001, 32, 2012.
[15] Gregory J. Galloway. The Lorentzian splitting theorem without the completeness assumption. J. Differential
Geom., 29(2):373–387, 1989.
[16] Leonardo Garcı́a-Heveling and Elefterios Soultanis. Causal bubbles in globally hyperbolic spacetimes. Gen. Rel-
ativity Gravitation, 54(12):Paper No. 155, 7, 2022.
[17] Robert Geroch. Domain of dependence. J. Mathematical Phys., 11:437–449, 1970.
[18] N. Gigli. The splitting theorem in non-smooth context. Preprint arXiv:1302.5555, 2013.
[19] Nicola Gigli. An overview of the proof of the splitting theorem in spaces with non-negative Ricci curvature. Anal.
Geom. Metr. Spaces, 2(1):169–213, 2014.
[20] Melanie Graf and Eric Ling. Maximizers in Lipschitz spacetimes are either timelike or null. Classical Quantum
Gravity, 35(8):087001, 6, 2018.
[21] James D. E. Grant, Michael Kunzinger, Clemens Sämann, and Roland Steinbauer. The future is not always
open. Lett. Math. Phys., 110(1):83–103, 2020.
[22] P. Hajlasz. Sobolev spaces on metric-measure spaces. In Heat kernels and analysis on manifolds, graphs, and
metric spaces (Paris, 2002), volume 338 of Contemp. Math., pages 173–218. Amer. Math. Soc., Providence, RI,
2003.
[23] Luis Ake Hau, Saul Burgos, and Didier A. Solis. Causal completions as Lorentzian pre-length spaces. Gen.
Relativity Gravitation, 54(9):Paper No. 108, 20, 2022.
[24] J. Heinonen, P. Koskela, N. Shanmugalingam, and J. Tyson. Sobolev spaces on metric measure spaces: an ap-
proach based on upper gradients. New Mathematical Monographs. Cambridge University Press, United Kingdom,
first edition, 2015.
GENERALIZED PRODUCTS AND LORENTZIAN LENGTH SPACES 31

[25] Yu Kitabeppu and Sajjad Lakzian. Characterization of low dimensional RCD∗ (K, N ) spaces. Anal. Geom. Metr.
Spaces, 4(1):187–215, 2016.
[26] Eva Kopfer. Super-Ricci flows and improved gradient and transport estimates. Probab. Theory Related Fields,
175(3-4):897–936, 2019.
[27] Eva Kopfer and Karl-Theodor Sturm. Heat flow on time-dependent metric measure spaces and super-Ricci flows.
Comm. Pure Appl. Math., 71(12):2500–2608, 2018.
[28] Eva Kopfer and Karl-Theodor Sturm. Functional inequalities for the heat flow on time-dependent metric measure
spaces. J. Lond. Math. Soc. (2), 104(2):926–955, 2021.
[29] Michael Kunzinger and Clemens Sämann. Lorentzian length spaces. Ann. Global Anal. Geom., 54(3):399–447,
2018.
[30] Michael Kunzinger and Roland Steinbauer. Null distance and convergence of Lorentzian length spaces. Ann.
Henri Poincaré, 23(12):4319–4342, 2022.
[31] Christian Lange, Alexander Lytchak, and Clemens Sämann. Lorentz meets Lipschitz. Adv. Theor. Math. Phys.,
25(8):2141–2170, 2021.
[32] Robert J. McCann. Displacement convexity of Boltzmann’s entropy characterizes the strong energy condition
from general relativity. Camb. J. Math., 8(3):609–681, 2020.
[33] Robert J. McCann and Clemens Sämann. A Lorentzian analog for Hausdorff dimension and measure. Pure Appl.
Anal., 4(2):367–400, 2022.
[34] Clemens Sämann. Global hyperbolicity for spacetimes with continuous metrics. Ann. Henri Poincaré, 17(6):1429–
1455, 2016.
[35] Miguel Sánchez. Globally hyperbolic spacetimes: slicings, boundaries and counterexamples. Gen. Relativity
Gravitation, 54(10):Paper No. 124, 52, 2022.
[36] Filippo Santambrogio. Optimal transport for applied mathematicians, volume 87 of Progress in Nonlinear Dif-
ferential Equations and their Applications. Birkhäuser/Springer, Cham, 2015. Calculus of variations, PDEs, and
modeling.
[37] Wolfgang Walter. Ordinary differential equations, volume 182 of Graduate Texts in Mathematics. Springer-
Verlag, New York, 1998. Translated from the sixth German (1996) edition by Russell Thompson, Readings in
Mathematics.

Department of Mathematics and Statistics, University of Jyväskylä, P.O. Box 35, FI-40014 Univer-
sity of Jyväskylä, Finland
Email address: elefterios.soultanis@gmail.com

You might also like