The Null Energy Condition through the Lens of Optimal Transport: Smooth and Non-Smooth Spacetimes

The connection between Ricci curvature and optimal transport originates in Riemannian geometry. A series of seminal works by McCann [61], Otto–Villani [77], Cordero-Erausquin, McCann, and Schmuckenschläger [30], and von Renesse–Sturm [90] established that lower Ricci curvature bounds can be characterized through displacement convexity properties of suitable entropy functionals along Wasserstein geodesics. This discovery culminated in the celebrated Lott–Sturm–Villani theory [59, 85, 86], which introduced the curvature-dimension condition $\mathsf{CD}(K,N)$, providing a synthetic notion of a metric measure space with Ricci curvature bounded below by $K\in\mathbb{R}$ and dimension bounded above by $N\in[1,\infty]$.

One of the most remarkable features of the $\mathsf{CD}(K,N)$ theory is that curvature is encoded entirely through optimal transport and entropy convexity, without relying on any differentiable structure. As a consequence, meaningful notions of Ricci curvature extend far beyond the class of smooth Riemannian manifolds.

A major advance came with the work of Ambrosio, Gigli and Savaré [5, 6], who uncovered a deep connection between lower Ricci curvature bounds and gradient-flow structures. More precisely, they proved that in $\mathsf{CD}(K,\infty)$ spaces the heat flow can be characterized as the $W_{2}$-Wasserstein gradient flow of the Boltzmann–Shannon entropy. Even more significantly, they showed that the associated gradient flow satisfies the Evolution Variational Inequality (EVI) formulation if and only if the heat flow is linear (see also the subsequent generalizations [4, 3]). This observation led to the identification of the genuinely “Riemannian” synthetic structures within the larger class of $\mathsf{CD}$ spaces, which also includes Finsler geometries.

These ideas gave rise to the theory of $\mathsf{RCD}$ spaces, whose fundamental additional requirement is the infinitesimal Hilbertianity of the Sobolev space $W^{1,2}$. The analytic and geometric consequences of this condition were systematically developed by Gigli [39], leading to a rich first- and second-order differential calculus in the absence of smoothness. A fundamental property of the $\mathsf{RCD}$ condition is the stability under pointed measured Gromov-Hausdorff convergence, proved by Gigli, Mondino and Savaré [41].

A further milestone was the realization that the optimal transport formulation of lower Ricci bounds is equivalent to the Eulerian viewpoint based on Bochner-type inequalities, also known as the Bakry–Émery condition. This equivalence was first established in the infinite-dimensional setting by Ambrosio, Gigli and Savaré [7], and subsequently in finite dimensions by Erbar, Kuwada and Sturm [34] and, independently, by Ambrosio, Mondino and Savaré [8]. Another cornerstone of the theory is the local-to-global property of the $\mathsf{RCD}(K,N)$ condition, established by Cavalletti and Milman [24].

The resulting $\mathsf{RCD}(K,N)$ theory has undergone spectacular development over the last decade and now provides one of the most powerful frameworks for geometric analysis in non-smooth spaces. We refer to the surveys [89, 2, 87, 40] for comprehensive introductions and further references.

More recently, analogous ideas have been developed in Lorentzian geometry. In contrast to the Riemannian setting, the presence of a causal structure fundamentally alters the nature of the transport problem: admissible transport plans must respect causality, and the resulting geometry is inherently asymmetric in time. In the timelike setting, McCann [61] obtained an optimal transport characterization of the strong energy condition. Independently, and shortly thereafter, Mondino and Suhr [71] showed that lower bounds on timelike Ricci curvature—and, more generally, the Einstein equations themselves—can be reformulated through displacement convexity properties of suitable entropy functionals along causal transport geodesics. These works established a conceptual bridge between Lorentzian curvature and optimal transport, providing the smooth foundation for the synthetic theory of timelike Ricci curvature bounds developed by Cavalletti and Mondino [26] within the framework of Lorentzian (pre-)length spaces introduced by Kunzinger and Sämann [54].

One of the main achievements of the theory developed in [26] (see also the survey [25]) is the extension of Hawking’s singularity theorem to a fully synthetic setting. Remarkably, the resulting theorem requires neither a differentiable structure nor a smooth Lorentzian metric, thereby showing that singularity formation can be understood as a genuinely synthetic phenomenon. Closely related developments include inextendibility results for Lorentzian pre-length spaces satisfying synthetic lower bounds on timelike sectional curvature [44, 1, 10], as well as singularity theorems in the general setting of closed cone structures [64]. Together, these works provide strong evidence that several fundamental features of Lorentzian geometry persist far beyond the classical smooth framework.

More broadly, this line of research has led to synthetic counterparts of a number of classical comparison, rigidity, and singularity theorems from Lorentzian geometry; see, for instance, [12, 9, 14, 11, 15]. In addition, the optimal transport approach has recently yielded new Lorentzian isoperimetric and geometric inequalities under synthetic timelike Ricci curvature lower bounds [27], highlighting the richness of the emerging synthetic theory.

A notable feature of the synthetic framework is that it accommodates singular spacetimes lying beyond the scope of classical Lorentzian geometry. In particular, the class of Lorentzian pre-length spaces with synthetic timelike Ricci curvature bounded below includes physically significant examples such as Penrose’s impulsive gravitational waves [70] and spacetimes endowed with Lipschitz Lorentzian metrics whose timelike Ricci curvature is bounded below in the sense of distributions [13]. This demonstrates that synthetic timelike Ricci bounds are sufficiently robust to capture geometries with genuinely low regularity while retaining meaningful curvature information.

While timelike Ricci curvature bounds provide a synthetic counterpart of the strong energy condition, a natural and equally important question is whether curvature bounds in null directions admit a similar formulation. From the perspective of general relativity, this amounts to developing a synthetic theory of the null energy condition (NEC), one of the most fundamental assumptions in causal and singularity theory. Since this topic constitutes the main focus of the next sections, we conclude this overview by briefly summarizing the principal developments in the area.

A first synthetic characterization of the NEC for smooth spacetimes was obtained by McCann [62]. The key idea is to view the NEC as a limiting case of timelike Ricci curvature lower bounds and to exploit the corresponding optimal transport formulations developed in [61, 71]. This approach is closely related to the synthetic $\mathsf{TCD}(K,N)$ theory introduced in [26] for Lorentzian length spaces. A notable feature of McCann’s characterization is that it naturally extends to non-smooth Lorentzian length spaces. However, as already observed in [62], the resulting condition is not stable under suitable notions of convergence.

A different perspective was subsequently introduced by Ketterer [51]. Rather than approaching the NEC through a limiting procedure from timelike directions, Ketterer directly investigated optimal transport along null hypersurfaces. More precisely, he established a characterization of the NEC in terms of displacement convexity properties of entropy functionals for singular probability measures concentrated on codimension-two spacelike submanifolds contained in null hypersurfaces.

The authors [22] later obtained an optimal transport characterization of the NEC based on diffuse probability measures supported on null hypersurfaces. More precisely, they showed that the NEC is equivalent to suitable entropy convexity inequalities along null transport geodesics. Their work develops a systematic framework linking Ricci curvature in null directions, rigged measures on null hypersurfaces, and optimal transport of probability measures that are absolutely continuous with respect to such measures. A key advantage of the diffuse-measure approach over the singular-measure framework of [51] is its compatibility with stability and convergence properties. This makes it particularly suitable for the development of a synthetic theory and ultimately yields the stable synthetic formulation of the null energy condition obtained in [23]. The main achievements of [23], which form the basis of the subsequent sections, are fourfold:

• •

  the introduction of well-posed and geometrically natural synthetic notions of null Ricci curvature lower bounds and of the null energy condition;

• •

  the proof that these synthetic notions are fully compatible with the classical smooth theory;

• •

  the establishment of stability under a suitable notion of convergence for synthetic null hypersurfaces, thereby overcoming a major limitation of previous approaches;

• •

  the extension of fundamental results of Lorentzian geometry, including Hawking’s area theorem and Penrose’s singularity theorem, to highly non-smooth settings.

To formulate the relevant entropy inequalities, let $(X,\mathfrak{m})$ be a measure space. The Boltzmann–Shannon entropy of a probability measure $\mu\in\mathcal{P}(X)$ with respect to the reference measure $\mathfrak{m}$ is defined by

provided $\mu=\rho\,\mathfrak{m}$ and $(\rho\log\rho)_{+}\in L^{1}(\mathfrak{m})$; otherwise one sets

Originally introduced by Boltzmann in statistical mechanics and later rediscovered by Shannon in information theory, this functional has become a central object linking geometry, probability, analysis, and optimal transport.

We denote by

the effective domain of the entropy.

Following [34], it is convenient to consider the dimensional entropy-power functional. Given $N>0$, define

and set $\mathcal{U}_{N}(\mu|\mathfrak{m}):=0$ otherwise.

The functional $\mathcal{U}_{N}$ is closely related to the classical Shannon entropy power from information theory (see, e.g., [32]); for instance, in Euclidean space $\mathbb{R}^{N}$ the entropy power corresponds to $\mathcal{U}_{N/2}$. Its relevance in geometry stems from the fact that convexity properties of $\mathcal{U}_{N}$ along optimal transport geodesics provide a synthetic formulation of lower Ricci curvature bounds. As we shall see, an analogous phenomenon occurs in the Lorentzian setting when optimal transport is performed along null hypersurfaces.

We are now in a position to state the main result of [22], which provides an optimal transport characterization of the null energy condition.

Let $H\subset M$ be a smooth null hypersurface and let $\nu\in\mathrm{OptGeo}^{H}(\mu_{0},\mu_{1})$ be a null-geodesic dynamical transport plan connecting two probability measures $\mu_{0},\mu_{1}\in\mathcal{P}_{ac}(H)$. Denote by

the associated interpolation, and define

The appearance of the exponent $n-1$ is natural, since a null hypersurface is an $(n-1)$-dimensional object and the corresponding entropy power is computed relative to the rigged measure $\mathrm{Vol}_{L}$.

The following theorem shows that the null energy condition is precisely equivalent to the displacement concavity of the entropy power along null-geodesic transports.

In other words, the null energy condition is equivalent to the concavity of the Shannon entropy power along suitable optimal transports contained in null hypersurfaces. This result should be viewed as the null counterpart of the displacement convexity characterizations of Ricci curvature lower bounds in Riemannian geometry.

The theorem admits a weighted generalization. Let

be a smooth weighted measure on a null hypersurface, and let $N\geq n$. Consider the Bakry–Émery $N$-Ricci tensor

Then $\operatorname{Ric}^{g,\Phi,N}\geq 0$ in null directions if and only if the same displacement-concavity property holds with $\mathcal{U}_{N-1}$ in place of $\mathcal{U}_{n-1}$. Thus entropy power continues to encode weighted curvature bounds in exactly the same way as in the smooth Riemannian theory.

One of the main advantages of the optimal transport formulation is that it remains meaningful under very weak regularity assumptions. In particular, the displacement-concavity inequalities above still make sense even when the weight function $\Phi$ is merely continuous.

Motivated by this observation, [22] introduced the following synthetic condition.

Beyond providing a synthetic characterization of the null energy condition, the optimal transport approach developed in [22] yields several geometric and physical applications. Among the main consequences are:

• •

  stability of the $\mathsf{NC}^{e}(N)$ condition under natural approximation procedures, namely $C^{1}_{\rm loc}$-convergence of the Lorentzian metrics and $C^{0}_{\rm loc}$-convergence of the weights, see [22, Thm. 9.1]. This robustness is one of the key features of the optimal transport formulation and foreshadows the stability results under measured Gromov–Hausdorff-type convergence obtained in [23] and described in the next section.

• •

  a weighted version of the classical light-cone theorem, together with a corresponding rigidity statement, extending earlier results of Choquet-Bruhat, Chruściel and Martín-García [28] and Grant [43], see [22, Sect. 10.1];

• •

  a weighted Hawking area theorem and an associated rigidity theorem.

We focus here on the weighted Hawking area theorem, which illustrates particularly well the geometric content of the $\mathsf{NC}^{e}(N)$ condition.

Recall that a null hypersurface $H\subset M$ is said to be future geodesically complete if every future-directed null geodesic contained in $H$ can be extended indefinitely towards the future. More precisely, whenever

is a future-directed null geodesic, there exists an extension

which is again a future-directed null geodesic.

Let $S\subset H$ be a spacelike cross-section of $H$, i.e. $H$ is spanned by null geodesics that pass through $S$ exactly once. The quantity

may be interpreted as the weighted $(n-2)$-dimensional area of $S$ with respect to the density $e^{\Phi}$. The following theorem asserts that such weighted areas are non-decreasing towards the future under the synthetic null energy condition.

When $\Phi\equiv 0$, the theorem reduces to the classical monotonicity of the area of spacelike sections of a future-complete null hypersurface under the null energy condition discovered by Hawking [48]. Thus the $\mathsf{NC}^{e}(N)$ condition recovers one of the most important consequences of the NEC in black-hole geometry. Moreover, the theorem continues to hold for merely continuous weights, a level of regularity for which the Bakry–Émery Ricci tensor is not even defined in a distributional sense.

The proof illustrates a recurring theme in synthetic geometry: rather than studying the expansion of null geodesic congruences through differential equations such as the Raychaudhuri equation, one derives the desired monotonicity directly from entropy-concavity inequalities along null-geodesic transport plans. In this way, a classical geometric consequence of the null energy condition emerges from a purely optimal transport argument.

We refer to [22] for the corresponding rigidity statements, the proofs, and for the weighted light-cone theorem, which provide further evidence that the displacement-concavity approach captures the essential geometric content of the null energy condition.

We begin with some standard terminology. Let $(X,\leq,\ll,\mathfrak{T})$ be a topological causal space. A subset $S\subset X$ is said to be achronal if

Equivalently, no two points of $S$ are in chronological relation.
Given a subset $S\subset X$, its chronological future is defined by

In a smooth spacetime, a null hypersurface can be characterized by the degeneracy of the induced metric. Since such a description is unavailable in the absence of a differentiable structure, a different viewpoint is required.

The key observation is that null hypersurfaces enjoy a strong causal property: they are achronal (at least locally), yet the causal relation does not trivialize on them. This suggests replacing smooth null hypersurfaces by suitable achronal subsets. An especially important class is provided by achronal boundaries.

One can easily prove that achronal boundaries are indeed achronal (see [23, Lem. 3.2]).

The relevance of this notion stems from the smooth theory. Indeed, in smooth spacetimes, achronal boundaries are classical objects in causality theory (see for instance [80, 49]). Under suitable assumptions they are ruled by null geodesics, and every maximal achronal set arises as an achronal boundary (see for instance [65, Thm. 2.97]). Consequently, achronal boundaries provide a natural synthetic replacement for null hypersurfaces.

Although achronality is a fundamental property of null hypersurfaces, it is not sufficient to characterize them synthetically. Indeed, spacelike hypersurfaces are also achronal, but they do not possess the causal generators that are the hallmark of null geometry. To capture this additional structure, one must require that the causal relation remains non-trivial on the set. The next sections, culminating in Definition 4.7, make this idea precise.

A second ingredient of the smooth theory is the rigged measure $\mathrm{Vol}_{L}$, which serves as the reference measure in the definition of entropy.

Recall that, although the rigged measure depends on the choice of a null geodesic vector field $L$, the resulting $\mathsf{NC}^{e}(N)$ condition is independent of this choice. This suggests that the measure itself should be regarded as auxiliary data rather than as part of the intrinsic geometry.

In the synthetic setting, one therefore starts with a non-negative Radon measure

defined on the achronal boundary $H$. The measure $\mathfrak{m}$ plays the role of the rigged measure and provides the reference measure appearing in the entropy functionals.

As in the smooth setting, the resulting synthetic curvature condition will ultimately be shown to be invariant under a suitable equivalence relation on reference measures. Roughly speaking, this equivalence corresponds to multiplying $\mathfrak{m}$ by functions that remain constant along the null generators of $H$, mirroring the change of rigging vector field in the smooth theory.

A further ingredient is required before one can formulate a synthetic version of the $\mathsf{NC}^{e}(N)$ condition. Recall that the latter is expressed through concavity properties of entropy along interpolating measures $(\mu_{t})_{t\in[0,1]}$ generated by null-geodesic transport plans. The parameter $t\in[0,1]$ plays a crucial role, since concavity is a statement about the evolution of the entropy along the interpolation. Consequently, one needs a distinguished way of parametrizing the null geodesics supporting the transport.

In Riemannian geometry, every minimizing geodesic $\gamma:[0,1]\to X$ admits a distinguished parametrization by constant speed, characterized purely in metric terms by the identity

This relation simultaneously encodes both the minimizing property of the curve and the fact that it is parametrized proportionally to arc length. Equivalently,

An analogous phenomenon occurs in Lorentzian geometry: timelike geodesics admit a canonical parametrization by proper time, which can be characterized in terms of the time-separation function $\ell$. Null geodesics are fundamentally different. For smooth spacetimes $(M,g)$, if $\gamma:[0,1]\to M$ is a null geodesic, any reparametrization still satisfies

It follows that the class of maximizing null curves is invariant under arbitrary reparametrizations. As a consequence, the Lorentzian length functional does not determine a distinguished parametrization of null curves. This stands in sharp contrast with the Riemannian and timelike Lorentzian settings, where the length functional naturally selects constant-speed and proper-time parametrizations, respectively.

This becomes problematic because the $\mathsf{NC}^{e}(N)$ condition is formulated through convexity properties of entropy along geodesics, and convexity is not preserved under arbitrary reparametrizations. In the smooth setting, one resolves this ambiguity by requiring null geodesics to satisfy the geodesic equation

which fixes the parametrization up to affine transformations. Since convexity is invariant under affine reparametrizations, this is exactly the amount of structure needed.

The synthetic analogue of this affine parametrization is encoded by the notion of a gauge function. Intuitively, a gauge function plays the role of an affine parameter along null generators. In the smooth setting, one may think of $G$ as a function satisfying

along suitably normalized null geodesics.

Assumption (1) implies that, for every causal curve $\gamma$ contained in ${H}^{\circ}$, the composition $G\circ\gamma$ is continuous and non-decreasing. Indeed, if $\gamma$ is not injective, one can first reparametrize it as an injective causal curve and then apply Assumption (1).

Assumption (2) prevents the gauge from blowing up along finite segments of causal curves and ensures that the affine parameter encoded by $G$ remains finite whenever the underlying causal curve has finite parameter length.

The notion of properness plays a particularly important role in the synthetic theory. As we shall see in Section 4.5, properness of the gauge provides a synthetic counterpart of future null geodesic completeness and constitutes a key ingredient in the formulation of the synthetic Penrose’s singularity theorem.

A gauge function allows one to recover, in a purely synthetic manner, the affine structure carried by the null generators of a smooth null hypersurface. This leads to the following synthetic analogue of an affinely parametrized null geodesic.

The third condition should be viewed as the synthetic counterpart of the geodesic equation in the smooth setting. Recall that, for smooth null hypersurfaces, the geodesic equation singles out affine parametrizations of the null generators. In the absence of a differentiable structure, this role is played by the gauge function $G$: requiring $G\circ\gamma$ to be affine provides a synthetic notion of affine parameter.

More precisely, if $H$ is a smooth null hypersurface and $G$ is induced by an affine parameter along its null generators, then the $G$-causal curves are precisely the affinely parametrized null geodesics contained in $H$. Thus the notion of $G$-causal curve extends the classical concept of affinely parametrized null geodesic to the synthetic setting.

Achronality is a fundamental feature of null hypersurfaces. Yet it is far from sufficient to grasp the distinctive features of null geometry, in particular curvature. Indeed, even spacelike hypersurfaces are achronal, yet they lack the causal generators that constitute the hallmark of null geometry. Any synthetic notion of null hypersurface must therefore retain not only achronality, but also a non-trivial causal structure.

The second and third conditions are designed precisely to capture this additional geometric information. The gauge function $G$ is defined on the non-endpoint part ${H}^{\circ}$ of $H$ and encodes the affine structure of its causal generators. At the same time, the requirement $\mathfrak{m}(\mathfrak{s}(H))=0$ ensures that the reference measure is concentrated on the dynamically relevant part of the hypersurface, namely the points through which non-trivial causal generators pass.

Taken together, the data $(H,G,\mathfrak{m})$ provide a synthetic analogue of a smooth null hypersurface endowed with an affine parameter along its null generators and a rigged measure. In particular, they furnish exactly the structures needed to develop a theory of entropy, optimal transport, and synthetic curvature bounds in null directions.

We now discuss several important classes of examples showing that the notion of synthetic null hypersurface naturally encompasses the classical objects of smooth Lorentzian geometry.

As reference measure on $H$, one takes the rigged measure $\mathrm{Vol}_{L}$ associated with the null geodesic vector field $L$. The resulting triple $(H,G_{L,S},\mathrm{Vol}_{L})$ is a synthetic null hypersurface. For more details see [23, Rem. 3.10].

Although achronal boundaries are generally far from smooth, they retain a remarkable amount of null geometry. Classical results in causality theory show that $H$ is ruled by null generators and enjoys many of the structural properties of smooth null hypersurfaces; see, for instance, [80, Lem. 3.17] and [49, Prop. 6.3.1]. This residual geometric structure is sufficient to construct, in a natural way, both a gauge function and a reference measure on $H$. Thus $H$ naturally carries the structure of a synthetic null hypersurface (see [23, Rem. 3.12] for more details).

This example is particularly important for applications. Indeed, achronal boundaries naturally appear throughout Lorentzian geometry and general relativity, including in the study of light cones, event horizons, Cauchy horizons, trapped surfaces, and singularity theorems. Consequently, the synthetic framework applies not only to smooth null hypersurfaces but also to many geometrically significant hypersurfaces that arise naturally in low-regularity settings.

The first fundamental requirement is that the synthetic theory recovers the classical one whenever a smooth spacetime structure is available. This is indeed the case, see [22, Thms. 3.18, 3.19]: for smooth null hypersurfaces $H$ endowed with their natural gauge function and rigged measure, the synthetic $\mathsf{NC}^{e}(N)$ is equivalent to the non-negativity of Ricci tensor in null directions along $H$. Thus the synthetic theory genuinely extends, rather than replaces, the classical smooth framework.

The definition of a synthetic null hypersurface involves two pieces of auxiliary data: a gauge function $G$, encoding the affine parametrization of the causal generators, and a reference measure $\mathfrak{m}$, playing the role of a rigged measure. A priori, it is therefore not clear whether the resulting $\mathsf{NC}^{e}(N)$ condition depends on these choices.

To formulate the appropriate notion of invariance, one first identifies the transformations that preserve the underlying null geometry.

Let $H$ be a closed achronal set. A function $f:{H}^{\circ}\to\mathbb{R}$ is called transverse if it is constant along the causal generators of $H$, that is, if $f\circ\gamma$ is constant for every causal curve $\gamma$ contained in ${H}^{\circ}$.

Using transverse functions, one introduces an equivalence relation on reference measures. Given $\mathfrak{m}_{1},\mathfrak{m}_{2}\in\mathcal{M}^{+}(H),$ we write $\mathfrak{m}_{1}\sim\mathfrak{m}_{2}$ if

and

for some positive transverse function $f$.

Similarly, given two gauges $G_{1},G_{2}:{H}^{\circ}\to\mathbb{R},$ we write $G_{1}\sim G_{2}$ whenever they determine the same family of distinguished causal curves, namely

These equivalence relations have a clear interpretation in the smooth setting. If $H$ is a smooth null hypersurface, changing the null geodesic vector field $L$ modifies both the associated rigged measure and the affine parametrization of the null generators. The relations above are precisely the synthetic counterparts of these classical gauge freedoms.

The following result shows that the $\mathsf{NC}^{e}(N)$ condition is intrinsic. For the notion of a tight gauge see  [23, Def. 3.16]; roughly, it means that the set of $G$-causal curves satisfy a pre-compactness condition á-la Arzerlá-Ascoli.

Consequently, the synthetic null energy condition depends only on the underlying null geometry and not on the particular gauge function or reference measure used to describe it. In this sense, the condition is geometrically well posed.

The third fundamental property is stability. One of the principal motivations for introducing synthetic curvature conditions is their applicability to singular limits of smooth geometries. Accordingly, one expects the $\mathsf{NC}^{e}(N)$ condition to be preserved under suitable notions of convergence.

This is indeed the case. In [23, Sect. 9], the authors introduce Lorentzian analogues of pointed measured Gromov–Hausdorff convergence and Sturm’s $\mathbb{D}$-convergence adapted to the setting of synthetic null hypersurfaces. Roughly speaking, the role traditionally played by the distance function in metric geometry is replaced by the causal and affine structures encoded by the gauge function $G$. The actual construction is considerably more subtle, however, as it is designed to be invariant under the natural equivalence relation $G_{1}\sim G_{2}$ discussed above.

One of the main results of [23] states that the $\mathsf{NC}^{e}(N)$ condition is stable under these notions of convergence. In particular, limits of synthetic null hypersurfaces satisfying $\mathsf{NC}^{e}(N)$ continue to satisfy the same condition.

This stability property is one of the major strengths of the optimal-transport approach. It allows one to pass to geometric limits and to formulate meaningful lower bounds on null Ricci curvature for singular spacetime models arising as limits of smooth ones. More generally, it provides a robust framework for studying low-regularity and non-smooth Lorentzian geometries beyond the reach of classical differential-geometric methods.

We conclude by mentioning that, motivated by similar questions, several notions of convergence for smooth and non-smooth spacetimes have recently been proposed in the literature; see, for instance, [26, 67, 73, 19, 82, 72]. Understanding the relationship between these convergence theories and the one introduced in [23, Sect. 9] remains an interesting and largely open problem.

A key observation is that, in the smooth setting, null geodesic completeness imposes a strong constraint on the associated gauge function. More precisely, if every null generator can be extended indefinitely towards the future, then the corresponding gauge is proper.

This suggests taking properness of the gauge as the synthetic analogue of null geodesic completeness. Consequently, a singularity theorem should assert that, under suitable geometric assumptions, the gauge cannot be proper. This characterization of null completeness is discussed in details in [23, Sect. 8.1].

The second issue concerns the trapped-surface assumption.

Classically, the relevant hypothesis is that a compact spacelike surface be future converging, namely that its mean curvature vector points sufficiently towards the future. Since mean curvature depends on the Christoffel symbols, this notion is unavailable for continuous metrics.

The key idea in [23, Sect. 7 and 8] is to replace the classical mean-curvature condition by a second-order variational property involving the reference measure. The motivation comes from the smooth theory: mean curvature arises as the first variation of area, while area itself may be viewed as the first variation of volume. It is therefore natural to seek a synthetic analogue of future convergence in terms of a second variation of volume.

In the synthetic setting, the role of volume is played by the reference measure $\mathfrak{m}$. The corresponding notion of area is given by the Minkowski content $\mathfrak{m}_{G}^{+}$, defined with respect to the gauge function $G$; see [23, Def. 7.1]. This leads to a purely measure-theoretic formulation of future convergence, which remains meaningful even in the absence of differentiable structures and curvature quantities.

Let $(H,G,\mathfrak{m})$ be a synthetic null hypersurface satisfying $\mathsf{NC}^{e}(N)$, and suppose that

Although somewhat technical at first sight, this condition admits a natural geometric interpretation: it expresses a strictly negative second-order variation of volume along the future null generators of $H$, and therefore serves as a synthetic analogue of negative null expansion.

With these notions in place, one obtains the following synthetic version of Penrose’s singularity theorem. The assumption of null non-branching [23, Def. 5.6] is the synthetic counterpart of the uniqueness of null generators in the smooth setting. Roughly speaking, it requires that null generators do not branch, so that two generators that agree on a non-trivial segment must coincide. This condition is automatically verified for achronal sets in $C^{2}$-Lorentzian manifolds; see [23, Cor. 5.9].