Some Relativistic and Gravitational Properties of the Wolfram

Model

Jonathan Gorard1,2
arXiv:2004.14810v1 [cs.DM] 28 Apr 2020

1
University of Cambridge∗

2
Wolfram Research, Inc.†

May 1, 2020

Abstract

The Wolfram Model, which is a slight generalization of the model first introduced by Stephen Wolfram

in A New Kind of Science (NKS), is a discrete spacetime formalism in which space is represented by a

hypergraph whose dynamics are determined by abstract replacement operations on set systems, and in

which the conformal structure of spacetime is represented by a causal graph. The purpose of this article

is to present rigorous mathematical derivations of many key properties of such models in the continuum

limit, as first discussed in NKS, including the fact that large classes of them obey discrete forms of both

special and general relativity. First, we prove that causal invariance (namely, the requirement that all

causal graphs be isomorphic, irrespective of the choice of hypergraph updating order) is equivalent to a

discrete version of general covariance, with changes to the updating order corresponding to discrete gauge

transformations. This fact then allows one to deduce a discrete analog of Lorentz covariance, and the

resultant physical consequences of discrete Lorentz transformations. We also introduce discrete notions

of Riemann and Ricci curvature for hypergraphs, and prove that the correction factor for the volume of

a discrete spacetime cone in a causal graph corresponding to curved spacetime of fixed dimensionality is

proportional to a timelike projection of the discrete spacetime Ricci tensor, subsequently using this fact

(along with the assumption that the updating rules preserve the dimensionality of the causal graph in

limiting cases) to prove that the most general set of constraints on the discrete spacetime Ricci tensor
∗ [email protected]
† [email protected]

1
 corresponds to a discrete form of the Einstein field equations. Finally, we discuss a potential formalism

for general relativity in hypergraphs of varying local dimensionality, and the implications that this for-

malism may have for inflationary cosmology and the value of the cosmological constant. Connections to

many other fields of mathematics - including mathematical logic, abstract rewriting systems, λ-calculus,

category theory, automated theorem-proving, measure theory, fluid mechanics, algebraic topology and

geometric group theory - are also discussed.

1 Introduction

Perhaps the single most significant idea conveyed within Stephen Wolfram’s A New Kind of Science[1],

and the initial intellectual seedling from which the contents of the book subsequently grow, is the abstract

empirical discovery that the “computational universe” - that is, the space of all possible programs - is far

richer, more diverse and more vibrant than one might reasonably expect. The fact that such intricate and

complex behavior can be exhibited by computational rules as apparently elementary as the Rule 30[2][3] and

Rule 110[4][5] cellular automata, which are so straightforward to represent that they can easily be discovered

by systematic enumeration, is profoundly counterintuitive to many people.

However, once one has truly absorbed and internalized this realization, it leads to an exceedingly tan-

talizing possibility: that perhaps, lying somewhere out there in the computational universe, is the rule for

our physical universe[6]. If an entity as remarkable as Rule 30 could be found just by an exhaustive search,

then perhaps so too can a theory of fundamental physics. The idea that there could exist some elementary

computational rule that successfully reproduces the entirety of the physical universe at first seems somewhat

absurd, although there does not appear to be any fundamental reason (neither in physics, nor mathematics,

nor philosophy) to presume that such a rule could not exist. Moreover, if there is even a remote possibility

that such a rule could exist, then it’s slightly embarrassing for us not to be looking for it. The objective of

the Wolfram Physics Project is to enact this search.

Conducting such a search is far from straightforward, since (more or less by definition) any reasonable

candidate for our universe will inevitably be rife with computational irreducibility. Definite lower bounds on

the degree of computational irreducibility in the laws of physics were first conjectured by Wolfram[7], and

one such bound was proved recently by the author[8], with the corollary that there exists an infinite class

of classical physical experiments with undecidable observables. Furthermore, explicit complexity-theoretic

bounds have also been recently demonstrated by the author for the case of a discrete spacetime formalism,

in collaboration with Shah[9].

2
 If one wants to construct a minimal model for the physical universe, then a cellular automaton is clearly

an inappropriate formalism, as it already contains far too much inbuilt structure (such as a rigid array of

cells, an explicit separation of the notion of space from the notion of states of cells, etc.). In NKS, Wolfram

argued that a more suitable model for space would be a collection of discrete points, with connections defined

between them in the form of a trivalent graph; laws of physics can then be represented by transformation

rules defined on this graph (i.e. rules which replace subgraphs with other subgraphs, in such a way as

to preserve the total number of outgoing edges). More recently, we have discovered that a slightly more

general approach is to model space as a hypergraph, and hence to define transformation rules as replacement

operations on set systems; an example of such a transformation rule is shown in Figure 1, and an example

of its evolution is shown in Figures 2 and 3.

Figure 1: An example of a possible replacement operation on a set system, visualized as a transformation

rule between two hypergraphs (which, in this particular case, happen to be equivalent to ordinary graphs).
Adapted from S. Wolfram, A Class of Models with Potential to Represent Fundamental Physics[10].

The present article begins by outlining a basic mathematical definition of the Wolfram Model and its

causal structure (in terms of discrete causal graphs) in Section 2.1, before proceeding to formalize these

definitions using the theory of abstract rewriting systems from mathematical logic in Section 2.2, and dis-

cussing some connections to λ-calculus and category theory. Section 2.3 builds upon these formalizations

to develop a proof that causal invariance (namely, the requirement that all causal graphs be isomorphic,

irrespective of the choice of hypergraph updating order) is equivalent to a discrete form of general covari-

ance, with changes to the updating order corresponding to discrete gauge transformations. This fact allows

one to deduce a discrete analog of Lorentz covariance, and the resultant physical consequences of discrete

Lorentz transformations, as demonstrated in Section 2.4, using techniques from the theory of automated

theorem-proving.

In Section 3.1, we introduce a discrete analog of the Ricci scalar curvature (namely the Ollivier-Ricci

curvature for metric-measure spaces) for hypergraphs, and prove that the correction factor for the volume

of a discrete ball in a spatial hypergraph corresponding to curved space of fixed dimensionality is trivially

proportional to the Ollivier-Ricci curvature. We subsequently introduce discrete notions of parallel transport,

3
Figure 2: An example evolution of the above transformation rule, starting from an initial (multi)hypergraph
consisting of a single vertex with two self loops. Adapted from S. Wolfram, A Class of Models with Potential
to Represent Fundamental Physics.

holonomy, sectional curvature, the Riemann curvature tensor and the spacetime Ricci curvature tensor in

Section 3.2, and use these notions to prove the corresponding result for volumes of discrete spacetime

cones in causal graphs corresponding to curved spacetimes of fixed dimensionality, demonstrating that the

correction factors are now proportional to timelike projections of the discrete spacetime Ricci tensor. In

Section 3.3, we use this fact, along with the assumption that the updating rules preserve the dimensionality

of the causal graph in limiting cases, to prove that the most general set of constraints on the discrete

spacetime Ricci tensor correspond to a discrete form of the Einstein field equations, using an analogy to the

Chapman-Enskog derivation of the continuum hydrodynamics equations from discrete molecular dynamics.

Finally, in Section 3.4, we discuss a potential formalism for general relativity in hypergraphs of varying local

dimensionality, using techniques from algebraic topology and geometric group theory, and the implications

that this formalism may have for inflationary cosmology and the value of the cosmological constant.

4
Figure 3: The final state of the above Wolfram Model evolution. Adapted from S. Wolfram, A Class of
Models with Potential to Represent Fundamental Physics.

2 Space, Time and Causality in the Wolfram Model

2.1 Basic Formalism and Causal Graphs

A much more complete description of this model is given by Wolfram in [10]. The essential idea here

is to model space as a large collection of discrete points which, on a sufficiently large scale, resembles

continuous space, in much the same way as a large collection of discrete molecules resembles a continuous

fluid. A geometry can then be induced on this collection of points by introducing patterns of connections

between them, as defined, for instance, by a hypergraph or set system. Here, a hypergraph denotes a direct

generalization of an ordinary graph, in which (hyper)edges can join an arbitrary number of vertices[11]:

Definition 1 A “spatial hypergraph”, denoted H = (V, E), is a finite, undirected hypergraph, i.e:

E ⊆ P (V ) \ {∅}, (1)

where P (V ) denotes the power set of V .

We assume henceforth that all hypergraphs are actually multihypergraphs, in the sense that E is actually a

multiset, thus allowing for hyperedges of arbitrary multiplicity.

Note that, for many practical purposes, it suffices to consider a special case of the more general hypergraph

formulation, in which space is simply represented by a trivalent graph:

5
Definition 2 A “trivalent spatial graph”, denoted G = (V, E), is a finite, undirected, regular graph of degree

3.

This was the approach adopted in NKS, and it suffices as a minimal representation for all finite, undirected

graphs, because any vertex of degree greater than 3 could equivalently be replaced by a cycle of vertices of

degree exactly 3, without changing any of the large-scale combinatorial properties of the graph.

The update rules, or “laws of physics”, that effectively determine a particular candidate universe (up to

initial conditions) are then defined to be abstract rewrite operations acting on these spatial hypergraphs,

i.e. operations which take a subhypergraph with a particular canonical form, and replace it with a distinct

subhypergraph with a different canonical form:

Definition 3 An “update rule”, denoted R, for a spatial hypergraph H = (V, E) is an abstract rewrite rule

of the form H1 → H2 , in which a subhypergraph H1 is replaced by a distinct subhypergraph H2 with the same

number of outgoing hyperedges as H1 , and in such a way as to preserve any symmetries of H1 .

With the dynamics of the hypergraph thus defined, we are able to introduce something akin to the

causal structure of a spacetime/Lorentzian manifold by constructing a directed graph in which every vertex

corresponds to a spacetime “event” (i.e. an application of an update rule), and every edge specifies a causal

relation between events.

Definition 4 A “causal graph”, denoted Gcausal , is a directed, acyclic graph in which every vertex corre-

sponds to an application of an update rule (i.e. an “event”), and in which the edge A → B exists if and only

if the update rule designated by event B was only applicable as a result of the outcome of the update rule

designated by event A.

Pragmatically speaking, this implies that the causal relation A → B exists if and only if the input for

event B has a non-trivial overlap with the output of event A. If the Wolfram Model is to be a plausible

underlying formalism for fundamental physics, one must presumably assume that every edge of the causal

graph corresponds to a spacetime interval on the order of (at most) the Planck scale, i.e. approximately

10−35 metres, or 10−43 seconds, although dimensional analysis of the model indicates a potentially much

smaller spacetime scale[10].

A causal graph corresponding to the evolution of an elementary string substitution system is shown in

Figure 4.

In this respect, the Wolfram Model can be thought of as being an abstract generalization of the so-

called “Causal Dynamical Triangulation” approach to quantum gravity developed by Loll, Ambjørn and

6
Figure 4: A causal graph corresponding to the evolution of the string substitution system
{BB → A, AAB → BAAB}, starting from the initial condition ABAAB. Adapted from S. Wolfram, A
New Kind of Science, page 498.

Jurkiewicz[12][13][14][15], in which spacetime is topologically triangulated into a simplicial complex of 4-

simplices (also known as pentachora), which then evolve according to some deterministic dynamical law.

With subhypergraph replacement rules of the general form H1 → H2 , it is evident that the evolution of

a given spatial hypergraph will, in general, be nondeterministic, since there does not exist any canonical

order in which the replacement rules should be applied, and distinct orderings will generally give rise to

distinct spatial hypergraphs. In other words, the evolution history for an arbitrary candidate universe will,

in most cases, correspond to a directed acyclic graph (known as a “multiway system”), rather than a single

path. A simple example of such a multiway system, corresponding to the evolution of an elementary string

substitution system, is shown in Figure 5.

However, it turns out that there exist many situations in which one is able to mitigate this problem,

and effectively obtain deterministic evolution of the hypergraph, by considering only replacement rules that

exhibit a particular abstract property known as “confluence”, as described below.

2.2 Abstract Rewriting, the Church-Rosser Property and Causal Invariance

Firstly, one is able to make the notions of update rules and transformations between hypergraphs more

mathematically rigorous, by drawing upon the formalism of abstract rewriting systems in mathematical

logic[16][17].

7
Figure 5: A multiway system corresponding to the evolution of an elementary string substitution system
{AB → A, BA → B}, starting from the initial condition ABA. Adapted from S. Wolfram, A New Kind of
Science, page 205.

Definition 5 An “abstract rewriting system” is a set, denoted A (where each element of A is known as an

“object”), equipped with some binary relation, denoted →, known as the “rewrite relation”.

Concretely, a → b designates a replacement rule, indicating that object a can be replaced with (or rewritten

as) object b.

More generally, there may exist situations in which two objects, a and b, are connected not by a single

rewrite operation, but by some finite sequence of rewrite operations, i.e:

a → a0 → a00 → · · · → b0 → b, (2)

in which case we can use the notation a →∗ b to indicate the existence of such a rewrite sequence. More

formally:

Definition 6 →∗ is the reflexive transitive closure of →, i.e. it is the transitive closure of → ∪ =, where =

denotes the identity relation.

In other words, →∗ is the smallest preorder that contains →, or the smallest binary relation that contains

→ and also satisfies both reflexivity and transitivity:

a →∗ a, and a →∗ b, b →∗ c =⇒ a →∗ c. (3)

The concept of “confluence” allows us to formalize the idea that some objects in an abstract rewriting

system may be rewritten in multiple ways, so as to yield the same eventual result:

Definition 7 An object a ∈ A is “confluent”, if:

8
 ∀b, c ∈ A, such that a →∗ b and a →∗ c, ∃d ∈ A such that b →∗ d and c →∗ d. (4)

Definition 8 An abstract rewriting system A is (globally) “confluent” (or exhibits the “Church-Rosser prop-

erty”) if every object a ∈ A is confluent.

Thus, within a (globally) confluent rewriting system, every time there exists an ambiguity in the rewriting

order, such that distinct objects b and c can be obtained by different rewrite sequences from some common

object a, those objects can always be made to reconverge on some common object d after a finite number of

rewriting operations. (Global) confluence is demonstrated explicitly for four elementary string substitution

systems in Figure 6.

Figure 6: Four evolution histories for (globally) confluent elementary string substitution systems (the first two
corresponding to A → B, and the last two to {A → B, BB → B} and {AA → BA, AB → BA}, respectively)
demonstrating that, irrespective of the chosen rewriting order, the same eventual result is always obtained.
Adapted from S. Wolfram, A New Kind of Science, pages 507 and 1037.

The notion of confluence hence allows us to assuage the reader’s potential fears regarding the potentially

ambiguous ordering of update operations on spatial hypergraphs. If we consider only confluent hypergraph

replacement rules, then, for any apparent divergence in the spatial hypergraphs that one obtains by following

different paths of the multiway system (i.e. any divergence resulting from ambiguity in the ordering of the

update operations), there will always exist a new path in the multiway system (i.e. a new sequence of

update operations) which causes that ambiguity to disappear. We can formalize this notion by stating

that confluence is a necessary condition for such rules to exhibit an asymptotic property known as “causal

invariance”:

Definition 9 A multiway system is “causal invariant” if the causal graphs that it generates (i.e. the causal

graphs generated by following every possible path in the multiway graph, corresponding to every possible

updating order) are all, eventually, isomorphic as directed, acyclic graphs.

A pair of simple but nontrivial string substitution systems exhibiting causal invariance is shown in Figure 7.

9
Figure 7: A pair of nontrivial string substitution systems exhibiting causal invariance (A → AA and
{A → AB, B → A}, respectively), since in both cases the combinatorial structure of the causal graph is
independent of the order in which the string substitutions get applied. Adapted from S. Wolfram, A New
Kind of Science, page 500.

There are some subtleties here, since the definition of (global) confluence only guarantees that the rewrite

sequences b →∗ d and c →∗ d must exist for a given b, c, but it does not guarantee anything about their

potential length. As such, the paths that one must follow in order to obtain convergence may be arbitrarily

long, so although causal invariance necessitates that the causal graphs generated by following every path

through the multiway system must eventually become isomorphic, those causal graphs are not guaranteed

to be isomorphic after any finite number of update steps. As such, causal invariance is best interpreted as a

limiting statement about the global structure of the multiway system.

Outside of the theory of abstract rewriting systems, there are many equivalent ways to formalize these

notions of update rules, confluence, causal invariance, etc. For instance, one can formulate a category-

theoretic version of the same ideas, by first considering the rewrite relation → of the abstract rewriting

system (A, →) to be an indexed union of subrelations:

→1 ∪ →2 =→ . (5)

This system is mathematically equivalent to a labeled state transition system, (A, Λ, →), with Λ being the

set of indices (labels), and this system is itself simply a bijective function from A to the powerset of A

indexed by Λ, i.e. P (Λ × A), given by:

10
 p 7→ {(α, q) ∈ Λ × A : p →α q}. (6)

Now, using the fact that the power set construction on the category of sets is a covariant endofunctor, the

state transition system is an F-coalgebra for the functor P (Λ × (−)). Then, the more general case in which

→ is not an indexed union of subrelations, which corresponds to an unlabeled state transition system, is

simply the case in which Λ is a singleton set. Therefore, in general, if P is considered to be an endofunctor

on the category Set:

P : Set 7→ Set, (7)

then the system (A, →) is the object A equipped with a morphism of Set, denoted →:

→: A 7→ PA. (8)

Statements about confluence and convergences between states in the multiway system thus translate into

purely category-theoretic statements regarding limits of functors and the relationship between cones and

co-cones (which can be thought of as being, respectively, the category-theoretic analogs of the notions of

critical pair divergence and convergence in abstract rewriting theory)[18].

There exist certain weakenings of the property of (global) confluence, including “local confluence”, in

which only objects which diverge after a single rewrite operation are required to reconverge:

∀b, c ∈ A, such that a → b and a → c, ∃d ∈ A such that b →∗ d and c →∗ d, (9)

and “semi-confluence”, whereby one object is obtained by a single rewrite operation and the other is obtained

by an arbitrary rewrite sequence:

∀b, c ∈ A, such that a → b and a →∗ c, ∃d ∈ A such that b →∗ d and c →∗ d. (10)

There are also various strengthenings, such as the strong “diamond property”, in which objects that diverge

after a single rewrite operation are also required to reconverge after a single rewrite operation:

∀b, c ∈ A, such that a → b and a → c, ∃d ∈ A such that b → d and c → d, (11)

11
and “strong confluence”, whereby, for two objects which diverged after a single rewrite operation, one is

required to converge with a single rewrite operation (or no rewrite operations at all), whilst the other can

converge via an arbitrary rewrite sequence:

∀b, c ∈ A, such that a → b and a → c, ∃d ∈ A such that b →∗ d and (c → d or c = d) . (12)

These properties, and their relation to the critical pair lemma in mathematical logic, are believed to be of

foundational relevance to the derivation of quantum mechanics and quantum field theory within the Wolfram

Model, as outlined in the accompanying publication[19].

The property of global confluence is often referred to as the Church-Rosser property for largely historical

reasons; Alonzo Church and J. Barkley Rosser proved in 1936[20] that β-reduction of λ terms in the λ-

calculus is globally confluent. In other words, if one considers the operation of replacing bound variables in

the body of an abstraction as being a rewrite operation of the form:

((λx.M ) E) → (M [x := E]) . (13)

then two λ terms are equivalent if and only if they are joinable. Here, “equivalence” of terms x and y refers

to the binary operation x ↔∗ y, which is the reflexive transitive symmetric closure of →, or the smallest

equivalence relation containing → (otherwise described as the transitive closure of ↔ ∪ =, where ↔ denotes

the symmetric closure of →, i.e. → ∪ →−1 ). Moreover, “joinability” refers to the binary operation x ↓ y,

which designates the existence of a common term z to which x and y are both reducible:

∀x, y, x ↓ y ⇐⇒ ∃z such that x →∗ z ←∗ y. (14)

Having shown that (global) confluence is a necessary condition for causal invariance, we now proceed to

prove that causal invariance implies an appropriately discretized form of general covariance, from which one

is able to deduce both special and general relativity (in the form of Lorentz covariance and local Lorentz

covariance, respectively).

2.3 Causal Graph Hypersurfaces and Discrete Lorentz Covariance

The first essential step in the derivation of special relativity for causal-invariant Wolfram Model systems is

to make precise the formal correspondence between directed edges connecting updating events in a discrete

12
causal graph, and timelike-separation of events in a continuous Minkowski space (or, more generally, in a

Lorentzian manifold). In doing this, it is helpful first to introduce a canonical method for laying out a causal

graph in Euclidean space, which can be defined as an optimization problem of the following form[21][22]:

Definition 10 A “layered graph embedding” is an embedding of a directed, acyclic graph in the Euclidean

plane, in which edges are represented as monotonic downwards curves, and in which crossings between edges

are to be minimized.

The ideal case of a layered graph embedding, which is not guaranteed to exist in general, would be a so-called

“downward planar embedding”:

Definition 11 A “downward planar embedding” is an embedding of a directed, acyclic graph in the Euclidean

plane, in which edges are represented as monotonic downwards curves without crossings.

In the standard mathematical formalism for special relativity[23], one starts by considering the (n + 1)-

dimensional Minkowski space R1,n , in which every point p is a spacetime event (specified by one time

coordinate and n spatial coordinates). If one now considers performing a layered graph embedding of a

causal graph into the discrete “Minkowski lattice” Z1,n , then one can label updating events by:

p = (t, x) , (15)

where t ∈ Z is a discrete time coordinate, and:

x = (x1 , . . . , xn ) ∈ Zn , (16)

are discrete spatial coordinates. One can then induce a geometry on Z1,n by making an appropriate choice

of norm:

Definition 12 The “discrete Minkowski norm” is given by:

k(t, x)k = kxk2 − t2 , (17)

or, in more explicit form:

k(t, x)k = x21 + · · · + x2n − t2 .

(18)

13
Definition 13 Updating events p = (t, x) are classified as either “timelike”, “lightlike” or “spacelike” based

upon their discrete Minkowski norm:






timelike, if k(t, x)k < 0,


p∼ if k(t, x)k = 0, (19)
lightlike,




spacelike,
 if k(t, x)k > 0.

Definition 14 Pairs of updating events p = (t1 , x1 ), q = (t2 , x2 ) can be classified as either “timelike-separated”,

“lightlike-separated” or “spacelike-separated”, accordingly:



timelike-separated,


 if ((t1 , x1 ) − (t2 , x2 )) ∼ timelike,


(p, q) ∼ lightlike-separated, if ((t1 , x1 ) − (t2 , x2 )) ∼ lightlike, (20)





spacelike-separated,
 if ((t1 , x1 ) − (t2 , x2 )) ∼ spacelike.

One of the foundational features of conventional special relativity is that two events are causally related

if and only if they are timelike-separated. From our definition of the discrete Minkowski norm and the

properties of layered graph embedding, we can see that a pair of updating events are causally related (i.e.

connected by a directed edge in the causal graph) if and only if the corresponding vertices are timelike-

separated in the embedding of the causal graph into the discrete Minkowski lattice Z1,n , as required.

Different possible layerings of the causal graph, or, equivalently, different possible “slicings” through the

causal graph taken in the canonical layered graph embedding, will therefore correspond to different possible

permutations in the ordering of the updating events. Every possible update scheme hence corresponds to

a different possible discrete foliation of the causal graph into these “slices”, where each “slice” effectively

designates a possible spatial hypergraph that can be generated by some permutation of the updating events.

Each such slice thus constitutes a discrete spacelike hypersurface, i.e. a discrete hypersurface embedded in

Z1,n , in which every pair of updating events is spacelike-separated. A possible discrete hypersurface foliation

(and hence, a possible updating order) of the causal graph for a non-causal-invariant string substitution

system is shown in Figure 8.

Physically, each such slice through the causal graph is the discrete analog of a Cauchy surface in space-

time, which one can show by making an explicit correspondence with the so-called “ADM decomposition”,

14
Figure 8: A simple string substitution system (AA → BA, BBB → A, A → AB) that is not causal-invariant,
since the causal graph is not unique. As a result, different possible discrete foliations of a given causal graph
into discrete spacelike hypersurfaces (shown as dashed lines), corresponding to different possible choices of
updating order for the substitution system, are found to produce distinct eventual outcomes for the system.
Adapted from S. Wolfram, A New Kind of Science, page 516.

developed by Arnowitt, Deser and Misner in 1959[24][25], for Hamiltonian general relativity. In the ADM

formalism, one decomposes a 4-dimensional spacetime (consisting of a manifold and a Lorentzian metric) by

foliating it into a parameterized family of non-intersecting spacelike hypersurfaces; it is fairly straightforward

to extend this notion to the case of discrete causal graphs:

Definition 15 A “discrete spacetime” is any pair (M, g), where M is a discrete metric space (taken to be

the discrete analog of a pseudo-Riemannian manifold), and g is a discrete Lorentzian metric of signature

(− + ++).

(Note that the mathematical details of how such a discrete Lorentzian metric tensor can be explicitly obtained

are outlined in the next section.)

Definition 16 A “discrete Cauchy surface” is a discrete spacelike hypersurface, i.e a set of updating events:

Σ ⊂ M, (21)

with the property that every timelike or lightlike (i.e. null) path through the causal graph, without endpoints

(taken to be the discrete analog of a smooth timelike or lightlike curve), intersects an updating event in Σ

exactly once .

We proceed on the assumption that our spacetime is globally discretely hyperbolic, in the sense that

there exists some universal “time function”:

15
 ∃t : M → Z, (22)

with non-zero gradient everywhere:

∆t 6= 0 everywhere, (23)

such that our spacetime can be discretely foliated into non-intersecting level sets of this function, and that

the collection of such level sets successfully covers the entire spacetime, i.e. the level surfaces t = const. are

discrete hypersurfaces, with the property:

∀t1 , t2 ∈ Z, Σt1 = {p ∈ M : t(p) = t1 }, and Σt1 ∩ Σt2 = ∅ ⇐⇒ t1 6= t2 . (24)

Then, by correspondence with the standard ADM decomposition, the proper distances on each discrete

hypersurface, denoted ∆l, are determined by the induced 3-dimensional discrete spatial metric, denoted γij :

∆l2 = γij ∆xi ∆xj , (25)

where the spatial coordinates xi (t) label the points on the discrete hypersurface, for a given coordinate

time t. The normal direction to the discrete hypersurface is given locally by some vector, denoted n ∈ Z1,n ,

representing the discrete relativistic 4-velocity of a normal observer; by travelling along the normal direction,

the distance in proper time, denoted ∆τ , to the adjacent discrete hypersurface at time t + ∆t is given by:

∆τ = α∆t, (26)

where α is the “lapse function”, i.e. a gauge variable that determines how the spacetime is discretely foliated

in the timelike direction. Analogously, a set of three gauge variables, known collectively as the “shift vector”,

denoted β i , determines how the spacetime is discretely foliated in the spacelike direction. The shift vectors

effectively relabel the discrete spatial coordinates in accordance with the following scheme:

xi (t + ∆t) = xi (t) − β i ∆t, (27)

such that the overall spacetime line element in our discrete analog of ADM becomes:

16
 ∆s2 = −α2 + β i βi ∆t2 + 2βi ∆xi ∆t + γij ∆xi ∆xj .


(28)

Thus, we can see that, under this formal analogy, the gauge freedom of the ADM formalism (i.e. our

freedom to choose values of α and β i for each normal observer) corresponds directly to our freedom to choose

an updating order, and hence a discrete foliation, for a given causal graph. In particular, α designates the

effective number of updating events required to map between an input subhypergraph and the corresponding

output subhypergraph in two neighboring hypersurfaces, and β i designates the effective graph distance

between those corresponding subhypergraphs.

Note that, in the above, we have implicitly made a correspondence between the combinatorial structure

of causal graphs, and the causal structure of Lorentzian manifolds[26][27][28]. In particular, we can exploit

this correspondence to define notions of chronological and causal precedence for updating events, as follows:

Definition 17 An updating event x “chronologically precedes” updating event y, denoted x  y, if there

exists a future-directed (i.e. monotonic downwards) chronological (i.e. timelike) path through the causal

graph connecting x and y.

Definition 18 An updating event x “strictly causally precedes” updating event y, denoted x < y, if there

exists a future-directed (i.e. monotonic downwards) causal (i.e. non-spacelike) path through the causal graph

connecting x and y.

Definition 19 An updating event x “causally precedes” updating event y, denoted x ≺ y, if either x strictly

causally precedes y, or x = y.

The standard algebraic properties of these relations, such as transitivity:

x  y, y  z =⇒ x  z, x ≺ y, y ≺ z =⇒ x ≺ z, (29)

then follow by elementary combinatorics.

One can hence define the chronological and causal future and past for individual updating events:

Definition 20 The “chronological future” and “chronological past” of an updating event x, denoted I + (x)

and I − (x), are defined as the sets of updating events which x chronologically precedes, and which chronolog-

ically precede x, respectively:

I + (x) = {y ∈ M : x  y}, I − (x) = {y ∈ M : y  x}. (30)

17
Definition 21 The “causal future” and “causal past” of an updating event x, denoted J + (x) and J − (x),

are defined as the sets of updating events which x causally precedes, and which casually precede x, respectively:

J + (x) = {y ∈ M : x ≺ y}), J − (x) = {y ∈ M : y ≺ x}. (31)

For instance, in the discrete Minkowski space, Z1,n , considered above, I + (x) designates only the interior of

the future light cone of x, whereas J + (x) designates the entire future light cone (i.e. including the cone

itself). Chronological and causal future and past can accordingly be defined for sets of updating events,

denoted S ⊂ M, as follows:

[ [
I ± (S) = I ± (x) , J ± (S) = J ± (x) . (32)
x∈S x∈S

Once again, standard algebraic properties of the chronological and causal future and past, such as the fact

that the interiors of past and future light cones are always strict supersets of the light cones themselves:

I + (S) = I + I + (S) ⊂ J + (S) = J + J + (S) ,

(33)

and:

I − (S) = I − I − (S) ⊂ J − (S) = I − I − (S) ,

(34)

also hold combinatorially.

One rather welcome consequence of this new formalism is that it allows us to introduce a more rigor-

ous definition of a discrete Cauchy surface within a causal graph (and hence, to provide a more complete

mathematical description of a discrete causal graph foliation):

Definition 22 The “discrete future Cauchy development” of a set of updating events S ⊂ M, denoted

D+ (S), is the set of all updating events x for which every past-directed (i.e. monotonic upwards), inex-

tendible, causal (i.e. non-spacelike) path in the causal graph through x also intersects an updating event in

S at least once. Likewise for the “discrete past Cauchy development”, denoted D− (S).

Definition 23 The “discrete Cauchy development” of a set of updating events S ⊂ M, denoted D (S), is

the union of future and past Cauchy developments:

18
 D (S) = D+ (S) ∪ D− (S) . (35)

Definition 24 A set of updating events S ⊂ M is “achronal” if S is disjoint from its own chronological

future, i.e:

@q, r ∈ S, such that r ∈ I + (q) . (36)

Definition 25 A “discrete Cauchy surface” in M is an achronal set of updating events whose discrete

Cauchy development is M.

Another elegant byproduct of this formalism is that it makes manifest the connection between causal graphs

and conformal transformations; specifically, the causal graph represents the conformally-invariant structure

of a discrete Lorentzian manifold. Since the combinatorial structure of a causal graph is unchanged by its

embedding, one can conclude that the causal structure of a discrete Lorentzian manifold is invariant under

the conformal transformation:

ĝ = Ω2 g, (37)

for conformal factor Ω, since the timelike, null and spacelike qualities of tangent vectors, denoted X, remain

invariant under this map, e.g:

g (X, X) < 0, =⇒ ĝ (X, X) = Ω2 g (X, X) < 0. (38)

Therefore, the requirement of causal invariance for a Wolfram Model corresponds precisely to the claim

that the ordering of timelike-separated updating events is agreed upon by all observers, irrespective of

their particular choice of discrete spacelike hypersurface foliation (i.e. it is invariant under changes in the

updating order of the system), even though the ordering of spacelike-separated events is not in general. In

the most generic case, this can be interpreted as the claim that the eventual outcomes of updating events

are independent of the reference frame (i.e. discrete hypersurface) in which the observer is embedded, which

is a discretized version of the principle of general covariance, or diffeomorphism invariance, of the laws of

physics.

In the particular case in which the discrete hypersurfaces contained within the foliation are required to

be “flat” (a rigorous definition of curvature for a discrete hypersurface will be presented within the next

19
section), those hypersurfaces will thus correspond to inertial reference frames, and the statement of causal

invariance reduces to a statement of discrete Lorentz covariance. Since the other fundamental postulate

of special relativity - namely constancy of the speed of light - is enforced axiomatically by our definition

of the edge lengths in causal graphs, this completes the proof. More explicit details of how the physical

consequences of discrete Lorentz transformations may be derived are outlined below.

It is worth noting that the true distinction between inertial and non-inertial frames in the Wolfram

Model is ultimately a computability-theoretic one; one may think of an observer in relativity as being an

entity which, upon observing the evolution of a collection of clocks, attempts to construct a hypothetical

gravitational field configuration which is consistent with that evolution (i.e. to “synchronize” these clocks).

Therefore, the relative computational power of the observer and the clocks places constraints upon the types

of field configurations (i.e. reference frames) that the observer is able to construct within their internal model

of the world. For instance, an observer who is unable to decide membership of any set of integers defined

by a Π1 or Σ1 sentence in the arithmetical hierarchy must be unable to construct a “Malament-Hogarth”

spacetime, (M, g), of the form:

Z
∃γ1 ⊂ M and p ∈ M, such that dτ = ∞ and γ1 ⊂ I − (p) , (39)
γ1

since such spacetimes are known to permit the construction of hypercomputers, i.e. computers that are

able to solve recursively undecidable problems in finite time[29][30][31][32], thus contradicting the stated

computational power of the observer. Therefore, an inertial reference frame corresponds to the limiting case

of an arbitrarily computationally-bounded observer, and other constraints on the geometry of spacetime (and

the resultant energy conditions regarding the matter content of that spacetime) follow from the observer’s

own position in the arithmetical hierarchy. A formal statement of this argument, and a derivation of its

computability-theoretic and relativistic consequences, will be outlined in a subsequent publication.

Note also that causal disconnections in spacetime, such as event horizons and apparent horizons, are now

represented as literal disconnections in the combinatorial structure of the causal graph (for instance, the

problem of determining whether null rays can escape to future null infinity from a region of spacetime can

now be represented as a concrete reachability problem for vertices of the causal graph). This observation

is relevant both to our exploration of quantum mechanics in [19], and to the implications of the Wolfram

Model for such problems as the black hole information paradox and the weak cosmic censorship hypothesis,

which shall be explored in future work.

20
2.4 Discrete Lorentz Transformations

Consider now how an observer embedded within one inertial reference frame (i.e. flat discrete hypersurface),

denoted F , measures discrete coordinates relative an observer embedded within a different inertial reference

frame, denoted F 0 , which we take to be moving at some constant velocity, v ∈ Zn , relative to F . Let:

v = (tanh (ρ)) u, (40)

where u ∈ Zn is a vector representing the direction of motion, and:

kvk
tanh (ρ) = < 1, (41)
kuk

represents the magnitude, appropriately normalized. By convention, we shall denote the coordinates in F

by (t, x), and the coordinates in F 0 by (t0 , x0 ), with the two coordinate systems synchronized such that:

t = t0 = 0, (42)

when the inertial reference frames initially coincide.

Definition 26 The “discrete Lorentz transformation” expresses the F 0 coordinates in terms of the F coor-

dinates as:

t0 = (cosh (ρ)) t − (sinh (ρ)) x · u, (43)

where · denotes the standard (“dot”) inner product of vectors in Zn .

To more clearly illustrate the connection between discrete Lorentz transformations and different choices

of updating orders in a causal graph, consider the causal graph for the very simple (and trivially causal-

invariant) string substitution system shown in Figure 9.

The causal graph, with the coordinates for its default foliation into discrete spacelike hypersurfaces (as

determined by the standard layered graph embedding), along with the corresponding updating order of the

substitution system (which one can think of as being the ordering of updating events as seen by an observer

in the rest frame F ), are shown in Figure 10.

Now, applying a discrete Lorentz transformation allows us to consider a new inertial reference frame,

F 0 , presumed to be moving with a normalized velocity v = 5

13 relative to the rest frame F . The vertex

21
Figure 9: An elementary and trivially causal-invariant string substitution system, AB → BA, applied to the
periodic initial condition ABABABAB . . .. Adapted from S. Wolfram, A New Kind of Science, page 518.

Figure 10: On the left is the causal graph, with the coordinates for its standard layered graph embedding,
corresponding to the default choice of updating order/discrete spacelike hypersurface foliation. On the right
is the evolution history of the string substitution system for this default updating order (i.e. the updating
order as seen by an observer in the rest frame F ). Adapted from Ø. Tafjord, NKS and the Nature of Space
and Time, slide 11.

coordinates (i.e. the discrete spacetime coordinates of the updating events) corresponding to this new choice

of causal graph embedding are shown in Figure 11.

This discrete Lorentz transformation, leading to a new possible embedding of the causal graph, itself

defines a new possible foliation of the causal graph into discrete spacelike hypersurfaces, and hence a new

possible updating order for the substitution system. The new causal graph foliation and updating order are

shown in Figure 12.

As expected, we see that, following discrete Lorentz transformations corresponding to a variety of different

(effectively subliminal) velocities, the ordering of spacelike-separated events may indeed vary, but the inherent

causal invariance of the substitution rule guarantees that the ordering of timelike-separated events always

remains invariant, as shown in Figure 13.

To understand the precise details of how concepts such as relativistic mass emerge within this new

formalism, we must first introduce a notion of (elementary) particles in spatial hypergraphs. For the sake

of simplicity, let us consider the particular case of ordinary graphs (i.e. hypergraphs in which each edge

22
Figure 11: The new, discretely Lorentz-transformed coordinates for the causal graph vertices (updating
events) in the inertial frame F 0 , moving with normalized velocity v = 13
5
relative to the rest frame F .
Adapted from Ø. Tafjord, NKS and the Nature of Space and Time, slide 12.

Figure 12: On the left is the causal graph, with the new, discretely Lorentz-transformed updating or-
der/hypersurface foliation shown. On the right is the evolution history of the string substitution system for
this new choice of updating order (i.e. the updating order as seen by an observer in the inertial frame F 0 ).
Adapted from Ø. Tafjord, NKS and the Nature of Space and Time, slide 13.

connects exactly two vertices). We now exploit a fundamental result in graph theory, known as “Kuratowksi’s

theorem”, which states that a graph is planar (i.e. can be embedded in the Euclidean plane without any

crossings of edges) if and only if it does not contain a subgraph that is a subdivision of either K5 (the complete

graph on 5 vertices), or K3,3 (the “utility graph”, or bipartite complete graph on 3 + 3 vertices)[33][34]:

Definition 27 A “subdivision” of an undirected graph G = (V, E) is a new, undirected graph H = (W, F )

which results from the subdivision of edges in G.

Definition 28 A “subdivision” of an edge e ∈ E, where the endpoints of e are given by u, v ∈ V , is obtained

by introducing a new vertex w ∈ W , and replacing e by a new pair of edges f1 , f2 ∈ F , whose endpoints are

given by u, w ∈ W and w, v ∈ W , respectively.

This implies that any nonplanarity in a spatial graph must be associated with a finite set of isolable, nonplanar

23
Figure 13: On the left are the causal graphs, with discretely Lorentz-transformed hypersurface foliations
corresponding to normalized velocities of v = 0.05, v = 0.25, v = 0.5 and v = 0.8, respectively. On the right
are the evolution histories of the string substitution system, as seen by observers moving at normalized
velocities of v = 0.05, v = 0.15, v = 0.25, v = 0.35, v = 0.45, v = 0.55, v = 0.65, v = 0.75, v = 0.85 and
v = 0.95, respectively. Adapted from Ø. Tafjord, NKS and the Nature of Space and Time, slide 15.

“tangles”, each of which is a subdivision of either K5 or K3,3 , as shown in Figure 14.

Figure 14: On the left is a trivalent spatial graph containing an irreducible crossing of edges. On the right
is an illustration of how this irreducible crossing is ultimately due to the presence of a nonplanar “tangle”
that is a subdivision of K3,3 . Adapted from S. Wolfram, A New Kind of Science, page 527.

Thus, if the update rules for spatial graphs have the property that they always preserve planarity, then

this immediately implies that each such nonplanar tangle will effectively act like an independent persistent

structure that can only disappear through some kind of “annihilation” process with a different nonplanar

tangle. As first pointed out in NKS, these persistent nonplanar structures are therefore highly suggestive

of elementary particles in particle physics, with the purely graph-theoretic property of planarity playing the

role of some conserved physical quantity (such as electric charge). The mathematical details of how one

is able to generalize this correspondence, in order to yield a graph-theoretic analog of Noether’s theorem

(by virtue of Wagner’s theorem[35], the Robertson-Seymour theorem[36][37], and a correspondence between

conserved currents in physics and graph families that are closed under the operation of taking graph minors

24
in combinatorics) is outlined in our accompanying publication on quantum mechanics[19].

In this context, relativistic mass can be understood in terms of the so-called “de Bruijn indices”[38] of

mathematical logic. In the λ-calculus, abstractions of the form (λx.M ) cause the variable x to become bound

within the expression, and therefore one must perform α-conversion of the λ-terms, i.e. one must perform

renaming of the bound variables[39]:

(λx.M [x]) → (λy.M [y]) , (44)

in order to avoid name collisions. However, if one instead indexes the bound variables in such a way as to

ensure that λ-terms are invariant under α-conversion, i.e. such that checking for α-equivalence is the same

as checking for syntactic equivalence, then one can avoid having to perform α-conversion at all. The central

idea underlying the de Bruijn index notation is to index each occurrence of a variable within a λ-term by a

natural number, designating the number of binders (within scope) that lie between that variable occurrence

and its associated binder. For instance, the K and S combinators from combinatory logic, namely:

λx.λy.x, and λx.λy.λz.xz (yz) , (45)

respectively, can be represented using de Bruijn indices as:

λλ2, and λλλ3 1 (2 1) , (46)

respectively.

De Bruijn indices are correspondingly used within automated theorem-proving systems as a naming

convention for generated variables. In constructing an automated proof of an equational theorem, such as

x = y, a standard technique is to treat the problem as one of abstract term rewriting: one can think of the

axioms of the system as being abstract rewriting rules, and then the problem of finding a proof of x = y

is ultimately the problem of determining a rewrite sequence that transforms x to y and vice versa, i.e one

wishes to determine that x ↔∗ y. In constructing this proof, one may consider rewriting rules of the form

f (x, y) → g (x, y, z), in which the variable z is a generated variable (i.e. it appears on the right-hand-side,

but not the left-hand-side, of the rule), which automated theorem-proving systems generally tend to index

using de Bruijn index notation. These generated variables will eventually be eliminated, for instance by rules

of the form g (x, y, z) → h (x, y), but they may be necessary for certain intermediate steps of the proof.

The key point here, however, is that the longer the path that one traverses through the proof graph, the

25
more intermediate (generated) variables will be produced; as such, the generated variables can be thought

of as being consequences (artefacts) of particular choices of updating orders. Since all abstract rewriting

formalisms are mathematically equivalent (and therefore, in particular, hypergraph substitution systems

are ultimately isomorphic to automated theorem-proving systems), this statement corresponds to the claim

that, the longer the path that one traverses through a multiway system (i.e. the steeper the angle of

“slicing” in the foliation of the causal graph), the more intermediate/generated vertices will be produced. As

such, elementary particles will appear to be “heavier” (i.e. to have more vertices associated with them) on

steeper “slices” in the foliation of the causal graph, without changing the eventual outcomes of the updating

events. This is the discrete analog of relativistic mass increase due to a Lorentz transformation - a concrete

demonstration of this idea is given in [10].

3 Discrete Geometry, Curvature and Gravitation in the Wolfram

Model

3.1 A Discrete Ricci Scalar Curvature for (Directed) Hypergraphs

Intuitively, the notion of curvature in Riemannian geometry is some measure of the deviation of a manifold

from being locally Euclidean, or, more formally, a measure of the degree to which the metric connection fails

to be exact. For the case of an n-dimensional Riemannian manifold, (M, g), we can make this statement

mathematically precise, in the following way: the Ricci scalar curvature (i.e. the simplest curvature invariant

of the manifold) is related to the ratio of the volume of a ball of radius , denoted B ⊂ M, to the volume

of a ball of the same radius in flat (Euclidean) space[40]:

Vol (B (p) ⊂ M) R

2 + O 4 ,


n
=1− (47)
Vol (B (0) ⊂ R ) 6 (n + 2)

as  → 0, where R denotes the Ricci scalar curvature of (M, g) at the point p ∈ M. However, an equivalent

method of defining R, which extends more readily to the discrete case of a directed hypergraph, is related

to the ratio of the average distance between points on two balls to the distance between their centres, when

those balls are sufficiently small and close together. More specifically, if B (p) ⊂ M is a ball of radius ,

centred at point p ∈ M, and it is mapped via parallel transport to B (q) ⊂ M, a corresponding ball of

radius  centred at point q ∈ M, then the average distance between a point on B (p) and its image (i.e. its

corresponding point on B (q)) is given by:

26
 2
 
3 2


δ 1− R+O  + δ , (48)
2 (n + 2)

as , δ → 0, where δ = d (p, q) denotes the distance between the centres p and q.

This local characterization of curvature in terms of average distances between balls, crucially, allows one to

extend the notion of Ricci scalar curvature to more general metric spaces (including hypergraphs)[41][42][43].

If one now considers an arbitrary metric space (X, d), with distance function d, then the natural generalization

of the concept of a volume measure is a probability measure, with the natural generalization of an average

distance between volume measures being the so-called “Wasserstein distance” between probability measures:

Definition 29 For a Polish metric space (X, d), equipped with its Borel σ-algebra, a “random walk” on X,

denoted m, is a family of probability measures:

m = {mx : x ∈ X}, (49)

each satisfying the requirement that mx has a finite first moment, and that the map x → mx is measurable.

Definition 30 The “1-Wasserstein distance”, denoted W1 (mx , my ), between two probability measures mx

and my on a metric space X, is the optimal transportation distance between those measures, given by:

"Z #
W1 (mx , my ) = Q inf d (x, y) d (x, y) , (50)
∈ (mx ,my ) (x,y)∈X×X
Q
where (mx , my ) denotes the set of measures on X × X, i.e. the coupling between random walks projecting

to mx and those projecting to my .

Q
Instinctively, one may view (mx , my ) as designating the set of all possible “transportations” of the measure

mx to the measure my (where a “transportation” involves “disassembling” the measure, transporting it, and

“reassembling it” in the form of my ). Thus, the Wasserstein distance is essentially the minimal cost (in

terms of transportation distance) required to transport measure mx to measure my .

Definition 31 For a metric space (X, d), equipped with a random walk m, the “Ollivier-Ricci scalar curva-

ture” in the direction (p, q), for distinct points p, q ∈ X, is given by:

W1 (mp , mq )
κ (p, q) = 1 − . (51)
d (p, q)

27
In the particular case in which (X, d) is a Riemannian manifold, and m is the standard Riemannian volume

measure, the Ollivier-Ricci scalar curvature κ (p, q) reduces (up to some arbitrary scaling factor) to the

standard Riemannian Ricci scalar curvature R (p, q). As such, one can think of mp and mq as being the

appropriate generalization of the notion of volumes of balls centred at points p and q in a Riemannian

manifold.

Moreover, for the case in which X is a discrete metric space, one can give the following, more explicit,

form of the Wasserstein transportation metric:

Definition 32 The “discrete 1-Wasserstein distance”, denoted W1 (mx , my ), between two discrete proba-

bility measures mx and my on a discrete metric space X, is the multi-marginal (i.e. discrete) optimal

transportation distance between those measures, given by:

 
X
W1 (mx , my ) = inf
Q
 d (x0 , y 0 ) µx,y (x0 , y 0 ) , (52)
µx,y ∈ (mx ,my )
(x0 ,y 0 )∈X×X

Q
where (mx , my ) here denotes the set of all discrete probability measures, µx,y , satisfying:

X
µx,y (x0 , y 0 ) = mx (x0 ) , (53)
y 0 ∈X

and:

X
µx,y (x0 , y 0 ) = my (y 0 ) . (54)
x0 ∈X

Let us now demonstrate how this formalism works in practice, by considering the particular case of a

directed hypergraph H = (V, E)[44], in which every hyperedge e ∈ E designates a directional relation between

two sets of vertices, A and B, referred to as the “tail” and the “head” of the hyperedge, respectively. Then,

for a given vertex in the tail of a directed hyperedge, xi ∈ A, the number of incoming hyperedges to xi ,

denoted dxin
i
, designates the number of hyperedges that include xi as an element of their head set. Similarly,

for a given vertex in the head of a directed hyperedge, yi ∈ B, the number of outgoing hyperedges from yj ,

denoted dyjout , designates the number of hyperedges which include yj as an element of their tail set:

Definition 33 For a directed hypergraph H = (V, E), the “Ollivier-Ricci scalar curvature” of the directed

hyperedge e ∈ E, where:

A = {x1 , . . . , xn } →e B = {y1 , . . . , ym }, (55)

28
where n, m ≤ |V |, is given by:

κ (e) = 1 − W (µAin , µB out ) , (56)

where the probability measures µAin and µB out satisfy:

n
X m
X
µAin = µxi , and µB out = µyj , (57)
i=1 j=1

with:


0, if z = xi and dxin 6= 0,



 i




 n1 , if z = xi and dxin = 0,


i
∀1 ≤ i ≤ n, z ∈ V, µxi (z) = (58)
1
if z 6= xi and ∃e0 : z → xi ,
P
n×dxin ×|tail(e0 )| ,






 e0 :z→xi i


if z 6= xi and @e0 : z → xi ,

0,

and:


0, if z = yj and dyjout 6= 0,








1,

if z = yj and dyjout = 0,
 m
∀1 ≤ j ≤ m, z ∈ V, µyj (z) = (59)
1
if z 6= yj and ∃e0 : yj → z,
P
m×dyout ×|head(e0 )| ,




e0 :yj →z

 j


if z 6= yj and @e0 : yj → z.

0,

Definition 34 The “discrete 1-Wasserstein distance”, denoted W1 (µAin , µB out ), between two discrete prob-

ability measures µAin and µB out on a directed hypergraph H = (V, E), is the multi-marginal (i.e. discrete)

optimal transportation distance between those measures, given by:

" #
X X
W (µAin , µB out ) = min d (u, v)  (u, v) , (60)
u→A B→v

where d (u, v) represents the minimum number of directed hyperedges that must be traversed when travelling

from vertex u ∈ Ain (u → A) to vertex v ∈ B out (B → v),  (u, v) represents the coupling, i.e. the total mass

29
being moved from vertex u to vertex v, and one is minimizing over the set of all couplings, , between measures

µAin and µB out , satisfying:

X m
X
 (u, v) = µyj (v) , (61)
u→A j=1

and:

X n
X
 (u, v) = µxI (u) . (62)
B→v i=1

Therefore, using the combinatorial metric d (u, v) defined above, in which each hyperedge is assumed to

correspond effectively to a unit of spatial distance, one is able to determine the dimensionality of a given

spatial hypergraph by determining the number of vertices, denoted N (r), that lie within a distance r of a

chosen vertex. For a “flat” spatial hypergraph, N (r) will scale exactly like rn :

N (r) = arn , a ∈ Z, (63)

where n is the dimensionality of the spatial hypergraph, as shown in Figure 15.

Figure 15: The number of vertices that lie within a distance r of a chosen vertex, denoted N (r), plotted
for flat spatial hypergraphs in one, two and three dimensions, respectively. Adapted from Ø. Tafjord,
Fundamental Physics and NKS, slide 6.

For instance, a two-dimensional spatial hypergraph can be thought of as consisting entirely of hexagons,

in which case N (r) will scale asymptotically like the area of a circle, i.e. πr2 . However, a curved spatial

hypergraph in two dimensions will also contain pentagonal (corresponding to regions of positive Ollivier-Ricci

scalar curvature) and/or heptagonal (corresponding to regions of negative Ollivier-Ricci scalar curvature)

structures, wherein there will exist a correction factor to this dimensionality calculation, as shown in Figure

16.

Using the fact that the Ollivier-Ricci scalar curvature reduces to the standard Riemannian Ricci scalar

curvature in the case of an n-dimensional Riemannian manifold, (M, g), we infer that the ratio of the volume

of a discrete ball of radius r in a curved spatial hypergraph to the volume of a discrete ball of the same

radius in a flat spatial hypergraph is given by:

30
Figure 16: Two-dimensional spatial hypergraphs with different Ollivier-Ricci scalar curvatures, with the
presence of pentagons indicating regions of positive scalar curvature, and the presence of heptagons indicating
regions of negative scalar curvature. Adapted from Ø. Tafjord, Fundamental Physics and NKS, slide 7.

Vol (Br (p) ∈ M) R

r2 + O r4 ,


=1− (64)
Vol (Br (0) ∈ Zn ) 6 (n + 2)

for any sufficiently small value of r (with exactness when r = 1), thus allowing us to compute the correction

factor to N (r) directly as:

 
R
N (r) = arn 1 − r2 + O r4 .


(65)
6 (n + 2)

In other words, the correction factor to the dimensionality calculation is found to be proportional to the

Ollivier-Ricci scalar curvature, as required. However, in order to derive the complete mathematical formalism

of general relativity, one must consider curvature not only in space (as represented by spatial hypergraphs),

but in spacetime (as represented by causal graphs). This, in turn, requires introducing appropriately dis-

cretized notions of parallel transport, holonomy, the metric tensor, the Riemann curvature tensor, and the

spacetime Ricci curvature tensor, which we proceed to do below.

3.2 Parallel Transport, Discrete Holonomy, and the Discrete Spacetime Ricci
Tensor

In standard differential geometry, the Ricci scalar curvature, R, is often introduced as being the trace of the

Ricci curvature tensor, denoted Rab :

R = Tr (Rab ) = g ab Rab = Raa , (66)

for a metric tensor gab . In much the same way as the Ricci scalar curvature can be interpreted as a measure

of the ratio of the volume of a ball in an n-dimensional Riemannian manifold, (M, g) to the volume of a

ball of the same radius in flat (Euclidean) space, the Ricci curvature tensor also has the following direct

geometrical interpretation: R (ξ, ξ) is a measure of the ratio of the volume of a conical region in the direction

of the vector ξ, consisting of geodesic segments of length  emanating from a single point p ∈ M, to the

31
volume of the corresponding conical region in flat (Euclidean) space.

The Ricci curvature may, in turn, be defined as being the contraction of the Riemann curvature tensor,

denoted Rabcd , between the first and third indices:

c
Rab = Racb , (67)

where the Riemann curvature tensor effectively quantifies how the direction of an arbitrary vector, v, changes

as it gets parallel transported around a closed curve in (M, g):

δv α = Rβγδ
α
dxγ dxδ v β ; (68)

in other words, the Riemann curvature tensor is a measure of the degree to which a Riemannian manifold

fails to be “holonomic” (i.e. fails to preserve geometrical data, as that data gets transported around a

closed curve). A more algebraic way of stating the same thing is that the Riemann curvature tensor is the

commutator of the covariant derivative operator acting on an arbitrary vector:

∇c ∇d v a − ∇d ∇c v a = Rbcd
a
vb . (69)

When defining the Ricci curvature tensor, the reason for contracting Rabcd between the first and third indices

specifically is that, due to the intrinsic symmetries of the Riemann curvature tensor, it is not difficult to

prove that any other choice of contraction will either yield zero, or ±Rab .

In order to have some coherent notion of parallel transport on a standard Riemannian manifold (and

hence to introduce such concepts as holonomy and Riemann curvature), it is necessary first to define a

connection on that manifold. The standard choice in Riemannian geometry is the Levi-Civita connection

(otherwise known as the torsion-free or metric connection), whose components in a chosen basis are given

by the Christoffel symbols, denoted Γρµν , and which can be expressed in terms of the first derivatives of the

metric tensor as:

1 ρσ
Γρµν = g (∂µ gσν + ∂ν gµσ − ∂σ gµν ) . (70)
2

Intuitively, the Christoffel symbols designate how the basis vectors change as one moves around on the

manifold. In the Levi-Civita connection, the Riemann curvature tensor can be written explicitly in terms of

second derivatives the metric tensor as:

32
 a
Rbcd = ∂c Γabd − ∂d Γabc + Γaec Γebd − Γaed Γebc . (71)

In a locally flat, inertial reference frame (i.e. one in which the Christoffel symbols, but not their derivatives,

all vanish), these components can be expressed in lowered-index form as:

1
Rαβγδ = (∂β ∂γ gαδ − ∂β ∂δ gαγ + ∂α ∂δ gβγ − ∂α ∂γ gβδ ) . (72)
2

These concepts now allow us to make the geometrical interpretation of the Ricci curvature tensor more

precise: if one chooses to use geodesic normal coordinates around the point p ∈ M, i.e. coordinates in which

all geodesics passing through p correspond to straight lines passing through the origin, then the metric tensor

can be approximated by the Euclidean metric:

gij = δij + O kxk2 ,

(73)

where δij denotes the standard Kronecker delta function, indicating that the geodesic distance from p can

be approximated by the standard Euclidean distance, as one would expect. The correction factor can be

determined explicitly by performing a Taylor expansion of the metric tensor along a radial geodesic in

geodesic normal coordinates:

1
gij = δij − Rijkl xk xl + O kxk3 ,


(74)
3

where the Taylor expansion is performed in terms of Jacobi fields:

 
∂γ τ (t)
J (t) = , (75)
∂τ τ =0

i.e. in terms of the tangent space to a given geodesic, denoted γ 0 , in the space of all possible geodesics

(here, γ τ denotes a one-parameter family of geodesics). Expanding the square root of the determinant of

the metric tensor hence yields an expansion of the metric volume element of (M, g), denoted dµg , in terms

of the volume element of the Euclidean metric, denoted dµEuclidean :

 
1 j k 3


dµg = 1 − Rjk x x + O kxk dµEuclidean , (76)
6

as required.

33
 Therefore, in order to be able to proceed sensibly with the derivation, we must now introduce appro-

priately discretized notions of holonomy, Riemann curvature, etc., that maintain compatibility with the

definition of the Ollivier-Ricci scalar curvature outlined previously. Note that, in the formal definition of the

Ollivier-Ricci scalar curvature for (directed) hypergraphs given above, we have implicitly employed a discrete

notion of parallel transport when mapping points on one hypergraph ball onto corresponding points on a

nearby hypergraph ball. To make this definition more explicit, and hence to derive the discrete analog of the

full Riemann curvature tensor, it is helpful first to consider the notion of sectional curvature on Riemannian

manifolds:

hR (u, v) v, ui
K (u, v) = , (77)
hu, uihv, vi − hu, vi2

where R is the full Riemann curvature tensor, and u and v are linearly independent tangent vectors at some

point p ∈ M. Geometrically, the sectional curvature at point p ∈ M designates the Gaussian curvature of

the surface obtained from the set of all geodesics starting at p, and proceeding in the directions of the tangent

plane σp defined by the tangent vectors u and v. In particular, when u and v are orthonormal, one has:

K (u, v) = hR (u, v) v, ui. (78)

Consider now an arbitrary metric space (X, d). If γ denotes a unit-speed geodesic whose origin is the

point x ∈ X, and whose initial direction is v, then we denote the endpoint of that geodesic by expx (v),

so-called because of its relation to the exponential map from tangent spaces to manifolds in Riemannian

geometry. Now, a natural generalization of the Riemannian sectional curvature, K (u, v), to the metric

space (X, d) is as follows:

2
 
3 2


d (expx (wx ) , expy (wy )) = δ 1 − K (v, w) + O  +  δ , (79)
2

as , δ → 0. In the above, v and wx denote unit-length tangent vectors at some point x ∈ X, point y ∈ X

denotes the endpoint of the vector δv, wy is the vector obtained by the parallel transport of vector wx from

point x to point y, and δ = d (x, y) denotes the distance between the points x and y. In this more general

context, the sectional curvature is a measure of the discrepancy between the distance between points x and

y, and the distance between the points lying a distance  away from x and y along the geodesics starting

at wx and wy , respectively. The Ollivier-Ricci scalar curvature for an arbitrary metric space can then be

recovered by taking the average of the sectional curvature K (v, w) over all vectors w. More formally, if Sx

34
is the set of all tangent vectors of length  at point x ∈ X, and similarly for Sy , and if Sx is mapped to Sy

via parallel transport, then the average distance between a point in Sx and its image is given by:

2
 
3 2


δ 1− R+O  + δ , (80)
2 (n + 2)

as , δ → 0, where R is the Ollivier-Ricci scalar curvature, as required.

In the case of a (directed) hypergraph, one can set, as usual, all hyperedges to be of unit length, and

hence one can also take  = δ = 1 in all of the definitions given above. Geodesics in the hypergraph are then

given by solutions to the shortest paths problem. We also assume that all pairs of hyperedges which are not

parallel (i.e. incident to the same set of vertices) can be considered to be orthonormal. These assumptions

allow us to translate the generalized definition of sectional curvature given above to the case of (directed)

hypergraphs, in a manner that is fully consistent with the discrete definition of the Ollivier-Ricci scalar

curvature as discussed earlier. Rather gratifyingly, since the components of the sectional curvature tensor

completely determine the components of the Riemann curvature tensor:

K (u, v) = hR (u, v) v, ui, (81)

one immediately obtains discrete analogs of the Riemann curvature and Ricci curvature tensors for the case

of (directed) hypergraphs.

Finally, in addition to asking about how the number of vertices, N (r) contained within a ball of radius r

in a particular spatial hypergraph grows as a function of r, we now also have the requisite technical machinery

to ask about how the number of updating events, denoted C (t), contained within a cone of length t within

a particular causal graph grows as a function of t (where t can be thought of as corresponding to the

number of discrete spacelike hypersurfaces intersected by the cone, for a particular discrete foliation of the

casual graph). Henceforth, we assume that special relativity holds (i.e. that all update rules referenced are

causal-invariant).

By direct analogy to the spatial hypergraph case, for a causal graph corresponding to flat, n-dimensional

spacetime, the number of updating events reached will clearly scale like tn :

C (t) = atn , a ∈ Z, (82)

as shown in Figure 17.

However, in the presence of non-zero spacetime curvature, there exists (again, by analogy to the purely

35
Figure 17: A cone of updating events, embedded within a flat, two-dimensional causal graph. Adapted from
S. Wolfram, A New Kind of Science, page 518.

spatial case) a correction factor, in the form of a coefficient of tn+2 , which we now proceed to compute

explicitly, by exploiting the fact that the discrete spacetime Ollivier-Ricci curvature tensor, Rij , determines

the ratio of the volume of a conical region in the causal graph consisting of geodesic segments of unit length

emanating from a single vertex, to the volume of the corresponding conical region in a purely flat causal

graph. Since, in the Riemannian case, one has:

 
1 j k 3


dµg = 1 − Rjk x x + O kxk dµEuclidean , (83)
6

one now obtains, in the discrete case:

 
1
C (t) = atn 1 − Rjk tj tk + O ktk3 ,


(84)
6

i.e. the correction factor is indeed found to be a coefficient of tn+2 , and is proportional to the projection of the

discrete spacetime Ollivier-Ricci curvature tensor along the discrete time vector, t, defined by the particular

choice of foliation of the causal graph into discrete spacelike hypersurfaces. In terms of the discrete ADM

decomposition introduced earlier, this time vector has the generic form:

ta = αna + β a , (85)

for a given choice of lapse function α and shift vector β i .

All that remains is to deduce an appropriate set of constraints on the spacetime Ollivier-Ricci curvature

tensor, and to show (subject to certain assumptions about the limiting conditions of the causal graph) that

36
these constraints are equivalent to an appropriately discretized form of the Einstein field equations.

3.3 The Discrete Einstein Field Equations

In all that follows, we shall assume one further condition on the hypergraph update rules, beyond mere

causal invariance: namely, “asymptotic dimensionality preservation”. Loosely speaking, this means that the

dimensionality of the causal graph show converge to some fixed, finite value as the number of updating events

grows arbitrarily large. We can express this requirement more formally by stating that the growth rate of

the “global dimension anomaly”, i.e. the correction factor to the global dimensionality of the causal graph,

should converge to zero as the size of the causal graph increases (since converging to anything non-zero would

imply a global dimension anomaly that grows without bound, and would therefore correspond to a causal

graph with effectively unbounded dimensionality). Since, as established above, the local correction factor to

the dimensionality of the causal graph is proportional to a projection of the discrete spacetime Ollivier-Ricci

curvature tensor Rij , we begin by computing an average of this quantity across both all possible projection

directions and all possible vertices, in order to obtain a value for the global dimension anomaly; averaging

out over all timelike projection directions t corresponds to a standard tensor index contraction, thus reducing

Rij to the discrete spacetime Ollivier-Ricci scalar curvature R, and hence yielding a volume average (i.e.

global dimension anomaly) given by:

  X
S g ab = dµg R. (86)

Our condition that the rules be asymptotically dimensionality preserving therefore corresponds to the state-

ment that the change in the value of the global dimension anomaly, with respect to the discrete causal graph

metric tensor, should converge to zero in the limit of a large causal graph:

 
δS g ab
→ 0, (87)
δg ab

In the continuum limit of an arbitrarily large causal graph, this sum becomes (subject to a weak ergodicity

assumption on the dynamics of the causal graph) an integral, with the volume element now being given by

the determinant of the metric tensor, as demonstrated earlier:

√
Z
S g ab = d4 x −gR,
 
(88)

37
such that the constraint that the update rules must be asymptotically dimensionality preserving becomes

mathematically equivalent to the statement that the classical (vacuum) Einstein-Hilbert action[45], with the

standard general relativistic Lagrangian density:

√
LG = −gR, (89)

must be extremized. Therefore, we can adopt the standard approach of taking a functional derivative of

the Einstein-Hilbert action with respect to the inverse metric tensor, and enforcing the assumption of zero

surface terms, to obtain:

 
δS g ab √
 
1
= −g Rab − Rgab , (90)
δg ab 2

with minimization of the action hence yielding the vacuum Einstein field equations:

1
Rab − Rgab = 0. (91)
2

Thus, for any set of hypergraph updating rules that are both causal-invariant and asymptotically dimension-

ality preserving, then if a continuum limit exists in which the causal graph becomes a Riemannian manifold,

that limiting manifold must satisfy the vacuum Einstein field equations, as required. A more intuitive

statement of this derivation, in which the argument is made that any local appearance of curvature in the

causal graph must disappear globally in order to allow the curvature to be “distinguishable” from a global

dimension change, is given in [10].

Amongst many other consequences, this implies that the trajectory, denoted x (s), of a test particle (e.g.

a persistent nonplanar tangle embedded within a spatial hypergraph) through a causal graph, will, in the

continuum limit and in the absence of any non-gravitational forces, obey the relativistic geodesic equation:

d2 xµ α
µ dx dx
β
+ Γ αβ = 0. (92)
ds2 ds ds

This fact can be proved in the usual way, by considering, for instance, a geodesic between a pair of timelike-

separated events in the causal graph. One starts from the action for a timelike curve:

Z
S= ds, (93)

where ds designates the spacetime line element:

38
 p
ds = −gµν dxµ dxν , (94)

with the negative sign before the metric tensor indicating the timelike nature of the curve. Then, by

introducing a scalar variable, λ, by which to parameterize the action:

Z r
dxµ dxν
S= −gµν dλ, (95)
dλ dλ

we are now able to vary the action with respect to the curve xµ , thus yielding the geodesic equation via the

principle of least action. Analogous derivations exist for geodesics between lightlike- and spacelike-separated

events too. The qualitative consequences of the geodesic equation for spatial hypergraphs are illustrated in

Figure 18.

Figure 18: Geodesic trajectories of test particles (starting on initially parallel paths) through flat, positively-
curved and negatively-curved discrete spaces, respectively. Adapted from S. Wolfram, A New Kind of
Science, page 531.

To incorporate matter contributions within the field equations, one can use the first-principles definition

of the energy-momentum tensor, i.e. that T µν designates the flux of the relativistic 4-momentum P µ across

the hypersurface of constant xν . This definition readily extends to the discrete case, since (as demonstrated

in the special relativity derivation above) this flux of the relativistic 4-momentum is, in the simplest case,

the number of crossing edges, i.e. the number of edges bound within nonplanar tangles of some spatial

hypergraph, which pass through a given discrete hypersurface in the causal graph (a more formal statement

of this idea is presented towards the end of this subsection). This increase in the effective number of

hyperedges increases the number of reachable vertices, N (r), lying within a distance r of a given vertex,

and hence must be corrected for within the dimensionality calculation. We can accommodate this correction

term by effectively adding a matter contribution to the relativistic Lagrangian density:

L = LG + CM LM , (96)

where CM is an arbitrary constant (which we assume, by convention, to be equal to 16π in the case of

39
scalar field matter). Now, minimization of the combined action yields to the full (non-vacuum) Einstein field

equations:

1
Rab − Rgab = 8πTab , (97)
2

so long as the energy-momentum tensor is taken to be of the following general form in terms of the matter

Lagrangian density:

 
CM ∂LM g ab
Tab =− √ , (98)
8π −g ∂g ab

which can also be reformulated in terms of an effective matter action, as:

 
CM δSM g ab
Tab =− √ . (99)
8π −g δg ab

The qualitative consequences of the inclusion of matter fields are illustrated in Figure 19.

Figure 19: Geodesic trajectories of test particles (starting on initially parallel paths) through discrete spaces
containing a spherical mass distribution. Adapted from S. Wolfram, A New Kind of Science, page 531.

Note that this entire derivation defines the Einstein field equations only up to an integration constant;

more specifically, the constraint that the update rules should be asymptotically dimensionality preserving

could equally well have been enforced by using the sum over the dimensionality correction factor, plus some

arbitrary constant, which we (suggestively) denote 2Λ:

  X
S g ab = dµg (R + 2Λ) . (100)

Running through the same logic as above, we see that, in the continuum limit, the general relativistic

Lagrangian density will now become:

40
 √
LG = −g (R + 2Λ) , (101)

thus yielding the new field equations:

1
Rab − Rgab + Λgab = 8πTab . (102)
2

Therefore, we can conclude that the Wolfram Model is consistent with both zero and non-zero values of

the cosmological constant. This suggests a generalization of the definition of the discrete energy-momentum

tensor given above, as being instead a measure of the flux (i.e. intersection) of causal edges through a discrete

hypersurface in the causal graph, with causal edges associated with the evolution of elementary particles (i.e.

nonplanar tangles) corresponding to standard baryonic matter contributions, and causal edges associated

with the evolution of the background space corresponding to vacuum energy contributions. In the usual way,

pure energy and momentum may be extracted from T µν by considering fluxes through purely spacelike and

purely timelike hypersurfaces, respectively.

It is interesting to note that the nature of this derivation of the continuum Einstein field equations

from the underlying discrete geometry of the causal graph is formally analogous to the so-called “Chapman-

Enskog” hydrodynamic expansion[46][47], used in the derivation of the continuum Navier-Stokes equations

from underlying discrete molecular dynamics. Generalizations of Chapman-Enskog theory were previously

developed by Wolfram[48] in the derivation of continuum hydrodynamics equations from discrete cellular

automaton models. In standard Chapman-Enskog theory, one starts from the 1-particle distribution function,

f (r, v, t), and constructs its Boltzmann equation[49]:

∂f ∂f ∂f
+v· +F· = Ĉf, (103)
∂t ∂r ∂v

for some nonlinear integral operator, Ĉ, characterizing the collisions between particles and the resultant

effect on the evolution of the distribution function. One now proceeds to introduce a dummy parameter,

denoted , into the Boltzmann equation:

∂f ∂f ∂f 1
+v· +F· = Ĉf, (104)
∂t ∂r ∂v 

and then to expand f as a formal power series in :

41
 f = f (0) + f (1) + 2 f (2) + . . . . (105)

The zeroth-order solution, given by:

 23 "
2
#
m (v − v00 (r, t))

(0) 0 m
f = n (r, t) exp − , (106)
2πkB T 0 (r, t) 2kB T 0 (r, t)

yields the dissipation-free Euler equations (i.e. the hydrodynamic description of a gas in equilibrium, in the

absence of thermal conductivity and viscosity), for functions n0 (r, t), v00 (r, t) and T 0 (r, t) determined by the

moments of f (r, v, t). The second-order solution, given by:

"   12 #
(1) 1 2kB T 2
f = − A (v) · ∇ log (T ) − B (v) : ∇v0 f (0) , (107)
n m n

yields the dissipative full Navier-Stokes equations, where A (v) and B (v) are a vector and tensor, respectively,

determined by solutions to an integral equation, and · : · denotes the double dot product between tensors:

XX
T : T0 = 0
Tij Tji . (108)
i j

Higher-order terms in the expansion give rise to higher-order supersets of the Navier-Stokes equations, such

as the Burnett equations and the super-Burnett equations.

As such, one can now see that our derivation of the Einstein field equations in the continuum limit of

causal-invariant and asymptotically dimensionality preserving Wolfram Model systems is essentially equiva-

lent to a Chapman-Enskog hydrodynamic expansion, but with the function:

 
1
C (t) = atn 1 − Rjk tj tk + O ktk3 ,


(109)
6

playing the role of a “distribution function” for vertices in the causal graph. One then recovers the dy-

namics of the Ricci curvature tensor (standard general relativity) as the zeroth-order term in the expansion

- analogous to the Euler equations - and the dynamics of the Weyl curvature tensor (higher-order general

relativity) as the corresponding first-order term - analogous to the Navier-Stokes equations. Excitingly, this

approach leaves open the possibility of the existence of additional higher-order supersets of the Einstein field

equations, presumably analogous to the Burnett and super-Burnett equations, which have been conjectured

by Wolfram to have possible implications for dark matter[50], for instance by analogy to Milgrom’s “Modified

42
Newtonian Dynamics” (MOND) proposal[51].

We intend to compute and investigate these potential higher-order corrections to the discrete Einstein

field equations more fully within a future publication, but one possible way in which they may manifest

is in the form of higher-order contractions of the Riemann curvature tensor, as in the Lovelock theory of

higher-dimensional gravity[52][53]:

 
1 1
α − Rρσ Rρσ gµnu − ∇ν ∇µ R − 2Rρνµσ Rσρ + gµν RRµν + (110)
2 2
 
1 2
β R gµν − 2RRµν − 2∇ν ∇µ R + 2gµν R + (111)
2
 
1 −2 −2
γ − κ Rgµν + κ Rµν = 0, (112)
2

with free parameters α, β, γ, and where:

κ2 = 32πG. (113)

A directly geometrical intuition for a possible manifestation of higher-order corrections as follows: the

Einstein field equations provide constraints on the Ricci curvature tensor, and hence on the volumes of

geodesic bundles in spacetime, but do not provide any explicit constraints on the shapes of those bundles,

which are instead described by the Weyl curvature tensor[54][55]:

1
Ciklm = Riklm + (Rim gkl − Ril gkm + Rkl gim − Rkm gil ) (114)
n−2
1
+ R (gil gkm − gim gkl ) . (115)
(n − 1) (n − 2)

Consider now the hydrodynamics analogy. In fluid mechanics, the total viscous stress tensor (i.e. the tensor

describing the stresses within the fluid due purely to its strain rate) can be exactly decomposed into a sum of

a trace part (i.e. a scalar multiple of the identity tensor, manifesting as a bulk hydrostatic pressure), and a

trace-free part (i.e. the viscous shear stress tensor). The Euler equations place constraints on the trace part

(the pressure), and their higher-order corrections place constraints on the trace-free part (the shear stresses).

In exactly the same way, the Riemann curvature tensor in general relativity can be decomposed into a sum

of a trace part (i.e. the Ricci curvature tensor, a sum of Kronecker delta functions) and a trace-free part (i.e.

43
the Weyl curvature tensor). The Einstein field equations place constraints on the Ricci curvature tensor,

and their higher-order corrections may very well place explicit additional constraints on the Weyl curvature

tensor too.

3.4 Topology, Variable Dimensionality and Implications for Cosmology

One assumption that has been implicit in the reasoning presented above is that, when computing the

dimensionality of spatial hypergraphs:

 
n R 2 4


N (r) = ar 1− r +O r , (116)
6 (n + 2)

or causal graphs:

 
1
C (t) = atn 1 − Rjk tj tk + O ktk3 ,


(117)
6

we have assumed that the dimensionality, n, is constant, and that the curvature, R, varies from point to point

on the manifold. However, it is equally mathematically consistent to fix the value of R, and to allow n to

vary as a function of position, effectively treating curvature as a highly localized change in the dimensionality

of the manifold. Of course, this involves slightly generalizing the notion of a manifold, since for an ordinary

manifold M, dimensionality is a local invariant. This indicates that the function f : M → R, mapping every

point in M to the dimensionality of its neighborhood, is locally constant, in the sense that[56]:

∀x ∈ M, ∃ neighborhood U ⊂ M, such that f is constant on U, (118)

which, in turn, entails that every connected component of M must have a fixed dimensionality.

Developing an extension of general relativity that can accommodate spacetime geodesics with varying

fractal dimension has been investigated in the context of so-called “Scale Relativity”[57][58][59], in which the

principle of general covariance is extended to accommodate scale invariance, i.e. in which physical laws are

taken to be valid in all coordinate systems, irrespective of their absolute scale. Within general scale relativity,

this variation in fractal dimension is constrained by a logarithmic version of the Lorentz transformation, and

is treated as a form of “curvature in scale space”. However, a complete general theory of scale relativity,

including scale-invariant versions of the Einstein field equations, has yet to be constructed.

For our present purposes, we shall consider one possible method of formalizing the treatment of cur-

44
vature as a local change in dimensionality: a unification of the topological concepts of coordinate charts

and covering maps. In standard topology, if M is a topological manifold, then a coordinate chart is a

homeomorphism[60][61]:

ϕ : (U ⊂ M) → (ϕ (U ) ⊂ Rn ) , (119)

i.e. a homeomorphism from an open subset of M to an open subset of the flat (Euclidean) space of dimension

n. An atlas of M is then an indexed family of charts:

A = {(Uα , ϕα ) : α ∈ I}, (120)

which covers M:

[
Uα = M. (121)
α∈A

However, there also exists the related notion of a covering map in algebraic topology, which is a continuous

surjective map between topological spaces X and C[62][63]:

p : C → X, (122)

such that, for every point x ∈ X, there exists an open neighborhood of x, denoted U ⊂ X, such that the

pre-image of U under the map p, i.e. p−1 (U ), is a union of disjoint open sets in C, and such that p maps

each such open set homeomorphically onto U . Here, C is known as the total space (or covering space) of the

covering, and X is known as the base space.

If the covering space C is simply connected, then it is known as a universal covering space, since it covers

any connected cover of the base space X[64]. More precisely, if p : C → X is a universal cover of X, and if

q : D → X is an arbitrary cover of X, then there must exist a covering map f : C → D such that q ◦ f = p.

In particular, in the case where the base space X is a Riemannian manifold, then so too is its universal

cover C, and the map p becomes a local isometry. Therefore, one is able to generalize the concept of a

Riemannian manifold to allow for connected components with variable dimensionality in the following way:

one fixes a base space of a fixed dimensionality, and then allows for a universal covering space of variable

dimensionality, with the universal cover of the space playing the role of the coordinate chart of the manifold.

Thus, the “curvature” of such a generalized manifold becomes a measure of the discrepancy between the local

45
dimensionality of the universal covering space and the global dimensionality of the base space. In the discrete

case of (directed) hypergraphs, the base space can be considered to be a group-theoretical lattice, produced

by some “modding-out” procedure. Then, the relationship between the discrete generalized manifold and

the underlying base space is directly analogous to the relationship between a Bethe lattice[65][66] and its

underlying Cayley graph[67] in geometric group theory.

More specifically, if G is a group and S is its generating set, then the Cayley graph of G, denoted

Γ = Γ (G, S), is the colored directed graph whose vertex set is given by the elements of G[68][69]:

V (Γ) = {g : g ∈ G}, (123)

and whose edge set is given by:

E (Γ) = {(g, gs) : g ∈ G, s ∈ S}, (124)

where the element s ∈ S determines the color of the edge. On the other hand, the Bethe lattice for a given

coordination number, z ∈ N, is an infinite, connected, acyclic graph (i.e. a rooted tree), in which each vertex

has exactly z neighbors, such that the number of vertices contained within the kth shell surrounding the

root vertex is given by:

k−1
Nk = z (z − 1) , k > 0. (125)

In the context of geometric group theory, the Bethe lattice with coordination number z = 2n may be con-

sidered to be the Cayley graph of the free group on n generators. Since any presentation of a group G by a

set S of n generators can be thought of as a surjective map from the free group on n generators to the group

G, all group presentations are effectively maps from the Bethe lattice (with the root vertex corresponding

to the identity element) to the associated Cayley graph. In algebraic topology, this map is interpreted as

the universal cover of the Cayley graph (i.e. the base space), where the Bethe lattice (the universal covering

space) is guaranteed to be simply connected, even though the corresponding Cayley graph is not. More

complete mathematical details of these variable-dimensionality “pseudo-manifolds”, their connections to ge-

ometric group theory, and the possibility of formulating an analog of general relativity in which dimension

is treated as a dynamical variable, will be presented in a subsequent publication.

In contemporary physics, inflationary cosmology is a modification to the standard model (otherwise

known as the ΛCDM or “hot big bang” model) of cosmology, invoked in order to explain the observed large-

46
scale homogeneity and isotropy of the universe, and hence to solve the so-called “horizon” and “flatness”

problems that are endemic to ΛCDM [70][71][72][73]. Here, we sketch briefly how the Wolfram Model, when

allowing for both local and global variation in dimensionality, is compatible with a model of cosmology which

yields identical observational consequences to inflation, and hence also constitutes a valid solution to the

horizon and flatness problems.

We begin by assuming that the initial condition for the universe consists of a spatial hypergraph with an

abnormally high vertex connectivity, perhaps corresponding to a complete graph. As such, the universe starts

off with some arbitrarily large number of spatial dimensions (which we can treat as being effectively infinite),

but then the asymptotic dimensionality preserving property of the update rules causes the number of spatial

dimensions to converge to some finite, fixed value, such as three. Another way to view such a universe, from

the point of view of its causal structure, is that it starts with an arbitrary large value of the speed of light

(since the causal graph is arbitrarily densely connected, yielding an abnormally high effective maximum rate

of information propagation in the hypergraph), which then converges to a some much lower, fixed value. In

this way, we can make an explicit mathematical connection with so-called “Variable Speed of Light” (or VSL)

models of cosmology, as investigated by Petit, Moffat, Albrecht and Magueijo[74][75][76][77][78][79], many

of which are known to yield similar observational consequences to more conventional inflationary models.

An elementary mathematical model of VSL cosmology uses a scalar field, denoted χ, to describe the

dynamical behavior of the effective speed of light:

c (t) = ĉχ (t) , (126)

where ĉ is a constant with the dimensions of velocity. In the simplest case, we can consider a phase transition

in the effective speed of light at absolute time t ∼ tc , given by:

c (t) = c0 θ (tc − t) + cm θ (t − tc ) , (127)

where cm is the current (measured) value of the speed of light, c0  cm is the initial value of the speed of

light, and θ (t) is the Heaviside step function:



1,
 if t > 0,
θ (t) = (128)

0,
 if t ≤ 0.

47
As such, the standard FLRW (Friedmann-Lemaı̂tre-Robertson-Walker) metric for an expanding universe,

defined by the spacetime line element:

dr2
 
2 µ ν 2 2 2 2 2 2


ds = gµν dx dx = dt c (t) − a (t) + r dθ + sin (θ) dφ , (129)
1 − kr2

with scale factor a (t), now takes on a bimetric form:

gµν = g0µν + gmµν , (130)

where:

dr2
 
2
ds20 µ ν
dt2 c20 θ (tc 2 2 2 2


= g0µν dx dx = − t) − a (t) + r dθ + sin (θ) dφ , (131)
1 − kr2

and:

dr2
 
2
ds2m = gmµν dxµ dxν = dt2 c2m θ (t − tc ) − a2 (t) 2 2 2


+ r dθ + sin (θ) dφ , (132)
1 − kr2

thus producing two distinct light cones, ds20 = 0 and ds2m = 0, with the dimensionless ratio:

c0
γ= , (133)
cm

determining their relative sizes.

To see how this yields a valid solution to the horizon problem, one can compute the proper horizon scale

using:

t
c (t0 ) dt0
Z
dH (t) = a (t) , (134)
0 a (t0 )

with t0 defined by the diffeomorphism:

dt0 = a (t) dt. (135)

1
Since, in a radiation-dominated universe, a (t) ∼ t 2 , one obtains the standard result dH ∼ 2cm t, when t > tc

and c = cm , and dh ∼ 2c0 t when t ∼ tc and c = c0 . Therefore, when γ → ∞, the size of the proper horizon

becomes arbitrarily large, hence allowing all observers to be in causal contact. In particular, the forward

light cone from the initial singularity becomes much larger than the region from which cosmic microwave

48
background photons are currently being observed, as required for isotropy of the horizon.

On the other hand, the solution to the flatness problem can be seen by inspecting the Friedmann equation:

c2 k 8πGρ c2 Λ
H 2 (t) + = + , (136)
a2 (t) 3 3

where the Hubble parameter, denoted H, is given by:

ȧ (t)
H (t) = . (137)
a (t)

Assuming zero cosmological constant, Λ = 0, and defining:

c2 |k|
 (t) = |Ω (t) − 1| = , (138)
ȧ2 (t)

where:

8πGρ
Ω (t) = , (139)
3H 2 (t)

we obtain:

2c2 |k|ä (t)

  2 
ċ c |k|
˙ (t) = − +2 . (140)
ȧ3 (t) c ȧ2 (t)

In a radiation-dominated universe, ä (t) < 0, and when the speed of light decreases from c0 to cm , one has
ċ
c < 0, and therefore ˙ (t) < 0. This corresponds to a cosmological attractor solution, with  ∼ 0, i.e. an

approximately spatially-flat universe, as required.

This connection between higher-dimensional (i.e. more densely connected) initial spatial hypergraphs

and the exponential expansion rates associated with inflationary cosmology should not be entirely surprising,

since even within standard FLRW cosmology it is known that higher values for the dimensionality of space

yield higher initial expansion rates[80]. This can be seen explicitly by considering the FLRW spacetime line

element in n spatial dimensions[81]:

−1
kr2

2 2 2
ds = dt − a (t) 1 + δij dxi dxj , (141)
4

for i, j = 1, 2, 3, . . . , n, and:

49
 X 2
r2 = (xi ) . (142)
i

For the case of a perfect relativistic fluid, with energy density ρ and fluid pressure p, obeying the higher-

dimensional energy conservation law:

ȧ
ρ̇ + n (ρ + p) = 0, (143)
a

this yields a reduced form for the higher-dimensional Einstein field equations, with n-dimensional gravita-

tional constant Gn , namely:

n(n − 1) ȧ2
 
k
+ 2 = 8πGn ρ, (144)
2 a2 a

and:

(n − 1)ä (n − 1)(n − 2) ȧ2

 
k
+ + = −8πGn p. (145)
a 2 a2 a2

For a spatially flat universe k = 0, with a non-zero cosmological constant Λ, and a higher-dimensional

cosmological equation of state parameter ωn [82]:

p = ωn ρ, (146)

the expansion term in the FLRW equations therefore takes the following analytic form, as derived by Holanda

and Pereira[83]:

h  q  i 2
  n(ω 1 +1) exp (ω + 1) 2Λn t − 1 n(ωn +1)
n(n − 1) n n n−1
a(t) = q  , (147)
8Λ exp 2Λ
n(n−1) t

thus implying higher early-time expansion rates and lower late-time expansion rates for increasing values

of n, and consequently allowing one to reproduce the effects of initially exponential expansion rates from

cosmic inflation with arbitrarily high-dimensional initial spatial hypergraphs.

50
4 Concluding Remarks

The present article has demonstrated the Wolfram Model to be a novel, exciting and potentially highly

fruitful discrete model for spacetime geometry, exhibiting discrete analogs of many (and possibly all) of

the salient mathematical features of Lorentzian and pseudo-Riemannian manifolds in limiting cases. There

exist a variety of open problems arising from this work, ranging from the possibility of computing higher-

order corrections to the discrete Einstein field equations, to determining the computability-theoretic and

complexity-theoretic properties that distinguish inertial and non-inertial reference frames, to developing a

theory of general relativity that holds in manifolds with variable spacetime dimensions. A few of these

problems are discussed in greater depth in our accompanying publication on quantum mechanics[19], which

makes significant use of both the special relativistic and general relativistic formalisms that we develop in

this paper (especially the relationship between confluence, causal invariance and Lorentz covariance, and

the derivation of the discrete Einstein field equations), and we intend to investigate several more of these

questions in the course of future publications. The present work, however, has at least revealed the Wolfram

Model to be a plausible fundamental model for classical relativistic and gravitational physics, and we eagerly

await the implications that this will entail.

Acknowledgments

The author is greatly indebted to the other members of the Wolfram Physics Project, namely Stephen

Wolfram and Max Piskunov, for their many and various contributions and insights relating to the present

work. SW is responsible for originally proposing the formalism of the Wolfram Model, for determining

many of its salient empirical and conceptual properties, for contributing greatly to the development of the

intellectual framework of which this work is a part, and (most notably through the writing of A New Kind of

Science) for inventing and continuing to champion the very scientific paradigm within which this work was

conducted. MP is responsible for much of the software development work that has spearheaded the present

project, and in particular remains the principal developer of the SetReplace paclet[84], used (among other

things) to render the images in Section 1 of the present text.

The author would also like to thank Isabella Retter and Ed Pegg Jr. for useful conversations (regarding

discrete differential geometry and combinatorial enumeration, respectively), Øyvind Tafjord and Todd Row-

land for their ample supply of NKS-related source material, and Sushma Kini for keeping us all on speaking

terms.

51
References

[1] S. Wolfram (2002), “A New Kind of Science”. Wolfram Media, Inc. ISBN 1-57955-008-8. https://www.

wolframscience.com/nks/

[2] S. Wolfram (1983), “Statistical Mechanics of Cellular Automata”. Reviews of Modern Physics, 55 (3):

601-644. https://www.stephenwolfram.com/publications/cellular-automata-complexity/pdfs/

statistical-mechanics-cellular-automata.pdf

[3] S. Wolfram (1985), “Cryptography with Cellular Automata”. Proceedings of Advances in

Cryptology - CRYPTO ’85 : 429. https://www.stephenwolfram.com/publications/academic/

cryptography-cellular-automata.pdf

[4] M. Cook (2004), “Universality in Elementary Cellular Automata”, Complex Systems, 15: 1-40. https:

//wpmedia.wolfram.com/uploads/sites/13/2018/02/15-1-1.pdf

[5] M. Cook (2008), “A Concrete View of Rule 110 Computation”, The Complexity of Simple Programs.

Electronic Proceedings in Theoretical Computer Science, 1: 31-55. https://arxiv.org/pdf/0906.

3248.pdf

[6] S. Wolfram (2015), “What is Spacetime Really?”. https://writings.stephenwolfram.com/2015/12/

what-is-spacetime-really/

[7] S. Wolfram (1985), “Undecidability and Intractability in Theoretical Physics”, Physical Re-

view Letters, 54 (8): 735-738. https://www.stephenwolfram.com/publications/academic/

undecidability-intractability-theoretical-physics.pdf

[8] J. Gorard (2018), “The Slowdown Theorem: A Lower Bound for Computational Irreducibility in Phys-

ical Systems”, Complex Systems, 27 (2): 177-185. https://wpmedia.wolfram.com/uploads/sites/

13/2018/09/27-2-5.pdf

[9] R. Shah, J. Gorard (2019), “Quantum Cellular Automata, Black Hole Thermodynamics and the Laws of

Quantum Complexity”, Complex Systems, 28 (4): 393-409. https://content.wolfram.com/uploads/

sites/13/2019/12/28-4-1.pdf

[10] S. Wolfram (2020), “A Class of Models With Potential to Represent Fundamental Physics”. https:

//www.wolframphysics.org/technical-introduction/

52
[11] C. Berge (1989), “Hypergraphs: Combinatorics of Finite Sets”. North-Holland Publishing Company.

ISBN 978-0444548887.

[12] R. Loll (2001), “Discrete Lorentzian Quantum Gravity”, Nuclear Physics B - Proceedings Supplements,

94: 96-107. https://arxiv.org/pdf/hep-th/0011194.pdf

[13] J. Ambjørn, A. Dasgupta, J. Jurkiewicz, R. Loll (2002), “A Lorentzian Cure for Euclidean Trou-

bles”, Nuclear Physics B - Proceedings Supplements, 106: 977-979. https://arxiv.org/pdf/hep-th/

0201104.pdf

[14] R. Loll, J. Ambjørn, J. Jurkiewicz (2006), “The Universe from Scratch”, Contemporary Physics, 47:

103-116. https://arxiv.org/pdf/hep-th/0509010.pdf

[15] J. Ambjørn, J. Jurkiewicz, R. Loll (2005), “Reconstructing the Universe”, Physical Review D, 72

(064014). https://arxiv.org/pdf/hep-th/0505154.pdf

[16] F. Baader, T. Nipkow (1998), “Term Rewriting and All That”. Cambridge University Press. ISBN

978-0521779203.

[17] M. Bezem, J. Willem Klop, R. de Vrijer (2003), “Term Rewriting Systems (‘Terese’)”. Cambridge

University Press. ISBN 0-521-39115-6.

[18] K. Stokkermans (1995), “Towards a Categorical Calculus for Critical-Pair/Completion”, Automated

Practical Reasoning: Texts and Monographs in Symbolic Computation (A Series of the Research Institute

for Symbolic Computation), 91-124. Springer, Vienna, ISBN 978-3-211-82600-3.

[19] J. Gorard (2020), “Some Quantum Mechanical Properties of the Wol-

fram Model”. https://www.wolframcloud.com/obj/wolframphysics/Documents/

some-quantum-mechanical-properties-of-the-wolfram-model.pdf

[20] A. Church, J. Barkley Rosser (1936), “Some Properties of Conversion”, Transactions of the American

Mathematical Society, 39 (3): 472-482. https://www.cs.cmu.edu/~crary/819-f09/ChurchRosser36.

pdf

[21] G. di Battista, P. Eades, R. Tamassia, I. G. Tollis (1998), “Flow and Upward Planarity”, Graph Drawing:

Algorithms for the Visualization of Graphs: 171-213. Prentice Hall. ISBN 978-0-13-301615-4.

53
[22] G. di Battista, F. Frati (2012), “Drawing Trees, Outerplanar Graphs, Series-Parallel Graphs, and Planar

Graphs in Small Area”, Thirty Essays on Geometric Graph Theory, Algorithms and Combinatorics,

Springer, 29: 121-165.

[23] G. L. Naber (1992), “The Geometry of Minkowski Spacetime”, New York: Springer-Verlag. ISBN 978-

0-387-97848-2.

[24] R. Arnowitt, S. Deser, C. Misner (1959), “Dynamical Structure and Definition of Energy in General

Relativity”, Physical Review, 116 (5): 1322-1330. https://authors.library.caltech.edu/72877/1/

PhysRev.116.1322.pdf

[25] C. W. Misner, K. S. Thorne, J. A. Wheeler (1973), “Gravitation”. W. H. Freeman. ISBN 978-0-7167-

0344-0.

[26] R. Penrose (1972), “Techniques of Differential Topology in Relativity”. Society for Industrial and Ap-

plied Mathematics (SIAM). ISBN 0898710057.

[27] A. V. Levichev (1987), “Prescribing the Conformal Geometry of a Lorentz Manifold by Means of its

Causal Structure”, Soviet Mathematics Doklady, 35: 452-455. https://link.springer.com/content/

pdf/10.1007%2FBF00970346.pdf

[28] D. Malament (1977), “The Class of Continuous Timelike Curves Determines the Topology of Spacetime”,

Journal of Mathematical Physics, 18 (7): 1399-1404. https://aip.scitation.org/doi/pdf/10.1063/

1.523436?class=pdf

[29] P. D. Welch (2008), “The Extent of Computation in Malament-Hogarth Spacetimes”, The British

Journal for the Philosophy of Science, 59 (4): 659-674. https://people.maths.bris.ac.uk/~mapdw/

borel2.1.pdf

[30] M. Hogarth (1992), “Does general relativity allow an observer to view an eternity in a finite time?”,

Foundations of Physics Letters, 5 (2): 173-181.

[31] M. Hogarth (1994), “Non-Turing computers and non-Turing computability”, PSA: Proceedings of the

Biennial Meetings of the Philosophy of Science Association, Volume 1 : 126-138.

[32] M. Hogarth (2004), “Deciding arithmetic using SAD computers”, British Journal for the Philosophy of

Science, 55 (1): 681-991.

54
[33] K. Kuratowski (1930), “Sur le problème des courbes gauches en topologie”, Fundamenta Mathematicae,

15: 271-283. http://matwbn.icm.edu.pl/ksiazki/fm/fm15/fm15126.pdf

[34] W. T. Tutte (1963), “How to Draw a Graph”, Proceedings of the London Mathematical Society, Third

Series, 13: 743-767. http://www.cs.jhu.edu/~misha/Fall07/Papers/Tutte63.pdf

[35] K. Wagner (1937), “Über eine Eigenschaft der Ebenen Komplexe”, Mathematische Annalen, 114: 570-

590. https://link.springer.com/content/pdf/10.1007%2FBF01594196.pdf

[36] N. Robertson, P. Seymour (1983), “Graph Minors. I. Excluding a Forest”, Journal of Combinatorial

Theory, Series B, 35 (1): 39-61. https://core.ac.uk/download/pdf/82390181.pdf

[37] N. Robertson, P. Seymour (2004), “Graph Minors. XX. Wagner’s Conjecture”, Journal of Combinatorial

Theory, Series B, 92 (2): 325-357. https://web.math.princeton.edu/~pds/papers/GM20/GM20.ps

[38] N. G. de Bruijn (1972), “Lambda Calculus Notation with Nameless Dummies: A Tool for Modern

Formula Manipulation, with Application to the Church-rosser Theorem”, Indagationes Mathematicae,

34: 381-392. ISSN 0019-3577. https://www.win.tue.nl/automath/archive/pdf/aut029.pdf

[39] M. J. Gabbay, A. M. Pitts (1999), “A New Approach to Abstract Syntax Involving Binders”, 14th An-

nual IEEE Symposium on Logic in Computer Science: 214-224. https://www.cl.cam.ac.uk/~amp12/

papers/newaas/newaas.pdf

[40] G. Ricci (1904), “Direzioni e Invarianti Principali in una Varietà Qualunque”, Atti del Reale Istituto

veneto di Scienze, 63 (2): 1233-1239.

[41] Y. Ollivier (2007), “Ricci Curvature of Metric Spaces”, Comptes Rendus Mathématique de l’Académie

des Sciences, 345: 643-646. https://www.math.uchicago.edu/~shmuel/QuantCourse%20/Metric%

20Space/Ollivier,%20Ricci%20curvature%20of%20Metric%20Spaces.pdf

[42] Y. Ollivier (2009), “Ricci Curvature of Markov Chains on Metric Spaces”, Journal of Functional Anal-

ysis, 256 (3): 810-864. https://arxiv.org/pdf/math/0701886.pdf

[43] Y. Ollivier (2011), “A Visual Introduction to Riemannian Curvatures and Some Discrete Generaliza-

tions”, Analysis and Geometry of Metric Measure Spaces: Lecture Notes of the 50th Séminaire de

Mathématiques Supérieures (SMS), Montréal, 56: 197-219. http://www.yann-ollivier.org/rech/

publs/visualcurvature.pdf

55
[44] M. Eidi, J. Jost (2019), “Ollivier Ricci Curvature of Directed Hypergraphs”. https://arxiv.org/pdf/

1907.04727.pdf

[45] D. Hilbert (1924), “Die Grundlagen der Physik”, Mathematische Annalen, 92: 1-32. https://link.

springer.com/content/pdf/10.1007/BF01448427.pdf

[46] S. Chapman, T. G. Cowling (1970). “The Mathematical Theory of Non-Uniform Gases” (3rd Edition).

Cambridge University Press. ISBN 978-0521408448.

[47] R. Balescu (1975), “Equilibrium and Nonequilibrium Statistical Mechanics”. John Wiley & Sons. ISBN

978-0-471-04600-4.

[48] S. Wolfram (1986), “Cellular Automaton Fluids 1: Basic Theory”, Journal of Statisti-

cal Physics, 45 (3/4): 471-526. https://www.stephenwolfram.com/publications/academic/

cellular-automaton-fluids-theory.pdf

[49] C. Cercignani (1975), “Theory and Application of the Boltzmann Equation”. Elsevier. ISBN 978-0-444-

19450-3.

[50] S. Wolfram (2019), Personal Communication.

[51] M. Milgrom (1983), “A Modification of the Newtonian Dynamics as a Possible Alternative to the Hidden

Mass Hypothesis”, Astrophysical Journal, 270: 365-370. http://articles.adsabs.harvard.edu/pdf/

1983ApJ...270..365M

[52] D. Lovelock (1971), “The Einstein Tensor and its Generalizations”, Journal of Mathematical Physics,

12: 498-501. https://aip.scitation.org/doi/pdf/10.1063/1.1665613?class=pdf

[53] S. A. Woolliams (2013), “Higher Derivative Theories of Gravity”. https://www.imperial.

ac.uk/media/imperial-college/research-centres-and-groups/theoretical-physics/msc/

dissertations/2013/LI_INCOMPACT3D2014.pdf

[54] S. W. Hawking, G. F. R. Ellis (1973), “The Large Scale Structure of Space-Time”. Cambridge University

Press. ISBN 0-521-09906-4.

[55] A. Danehkar (2009), “On the Significance of the Weyl Curvature in a Relativistic Cosmological Model”,

Modern Physics Letters A, 24 (38): 3113-3127. https://arxiv.org/pdf/0707.2987.pdf

56
[56] J. Lee (2010), “Introduction to Topological Manifolds”. Springer Science & Business Media. ISBN 978-

1-4419-7940-7.

[57] L. Nottale (1989), “Fractals and the Quantum Theory of Spacetime”, International Journal of

Modern Physics A, 04 (19): 5047-5117. https://www.worldscientific.com/doi/pdf/10.1142/

S0217751X89002156

[58] G. N. Ord (1983), “Fractal Space-Time: A Geometric Analogue of Relativistic Quantum Mechanics”,

Journal of Physics A: Mathematical and General, 16 (9): 1869-1884. https://iopscience.iop.org/

article/10.1088/0305-4470/16/9/012/pdf

[59] L. Nottale, J. Schneider (1984), “Fractals and Nonstandard Analysis”, Journal of Mathematical Physics,

25 (5): 1296-1300. https://aip.scitation.org/doi/pdf/10.1063/1.526285

[60] J. Jost (2013), “Riemannian Geometry and Geometric Analysis”. Springer Science & Business Media.

ISBN 978-3-319-61859-3.

[61] J. M. Lee (2006), “Introduction to Smooth Manifolds”. Springer-Verlag New York. ISBN 978-0-387-

95448-6.

[62] E. H. Spanier (1994), “Algebraic Topology”. Springer-Verlag New York. ISBN 978-0-387-94426-5.

[63] A. V. Chernavskii (2001), “Covering”, Encyclopedia of Mathematics. Springer Science & Business Media.

ISBN 978-1-55608-010-4.

[64] J. R. Munkres (2000), “Topology” (2nd Edition). Upper Saddle River, New Jersey: Prentice Hall. ISBN

0131816292.

[65] H. A. Bethe (1935), “Statistical Theory of Superlattices”, Proceedings of the Royal Society of London

A, 150: 552-575. https://royalsocietypublishing.org/doi/pdf/10.1098/rspa.1935.0122

[66] R. J. Baxter (1982), “Exactly Solved Models in Statistical Mechanics”. Academic Press. ISBN 0-12-

083182-1.

[67] M. Ostilli (2012), “Cayley Trees and Bethe Lattices, a Concise Analysis for Mathematicians and Physi-

cists”, Physica A, 391: 3417-3424. https://arxiv.org/pdf/1109.6725.pdf

57
[68] A. Cayley (1878), “Desiderata and Suggestions: No. 2. The Theory of Groups: Graphical Representa-

tion”, American Journal of Mathematics, 1 (2): 174-176. https://www.jstor.org/stable/2369306?

seq=1#metadata_info_tab_contents

[69] W. Magnus, A. Karrass, D. Solitar (2004), “Combinatorial Group Theory: Presentations of Groups In

Terms of Generators and Relations”. Courier. ISBN 978-0-486-43830-6.

[70] J. A. Peacock (1998), “Cosmological Physics”. Cambridge University Press. ISBN 978-0-521-42270-3.

[71] R. H. Dicke (1970), “Gravitation and the Universe: Jayne Lectures for 1969”. American Philosophical

Society. ISBN 978-0871690784.

[72] A. Guth (1981), “Inflationary Universe: A Possible Solution to the Horizon and Flatness Prob-

lems”, Physical Review D, 23 (2): 347-356. http://tucana.astro.physik.uni-potsdam.de/~cfech/

lectures/galkos/guth1981.pdf

[73] D. J. Fixsen (2009), “The Temperature of the Cosmic Microwave Background”, The Astrophysical

Journal, 707 (2): 916-920. https://arxiv.org/pdf/0911.1955.pdf

[74] J. P. Petit (1988), “An Interpretation of Cosmological Model with Variable Light Velocity”, Modern

Physics Letters A, 3 (16): 1527-1532. https://www.jp-petit.org/science/f300/modern_physics_

letters_a1.pdf

[75] J. P. Petit (1988), “Cosmological Model with Variable Light Velocity: The Interpretation of Red Shifts”,

Modern Physics Letters A, 3 (18): 1733-1744. https://www.jp-petit.org/science/f300/modern_

physics_letters_a2.pdf

[76] J. P. Petit, M. Viton (1989), “Gauge Cosmological Model with Variable Light Velocity. Comparison with

QSO Observational Data”, Modern Physics Letters A, 4 (23): 2201-2210. https://www.jp-petit.org/

science/f300/modern_physics_letters_a3.pdf

[77] J. W. Moffat (1993), “Superluminary Universe: A Possible Solution to the Initial Value Problem in

Cosmology”, International Journal of Modern Physics D, 2 (3): 351-366. https://arxiv.org/pdf/

gr-qc/9211020.pdf

[78] J. W. Moffat (2002), “Variable Speed of Light Cosmology: An Alternative to Inflation”. https://

arxiv.org/pdf/hep-th/0208122.pdf

58
[79] A. Albrecht, J. Magueijo (1999), “A Time Varying Speed of Light as a Solution to Cosmological Puzzles”,

Physical Review D, 59 (4): 043516. https://arxiv.org/pdf/astro-ph/9811018.pdf

[80] J. A. S. Lima (2001), “Note on solving for the dynamics of the Universe”, American Journal of Physics,

59: 1245-1247. https://aapt.scitation.org/doi/10.1119/1.1405506

[81] F. R. Tangherlini (1963), “Schwarzschild field in n dimensions and the dimensionality

of space problem”, Il Nuovo Cimento, 27 (3): 636-651. https://www.osti.gov/biblio/

4749059-schwarzschild-field-inn-dimensions-dimensionality-space-problem

[82] M. S. Turner and M. White (1997), “CDM models with a smooth component”, Physics Review D 56

(8): R4439(R). https://journals.aps.org/prd/abstract/10.1103/PhysRevD.56.R4439

[83] R. F. L. Holanda and S. H. Pereira (2012), “On the Dynamics of the Universe in D Spatial Dimensions”,

Revista Mexicana de Astronomı́a y Astrofı́sica, 48 (2): 251-256. https://arxiv.org/abs/0707.3387

[84] M. Piskunov, SetReplace, https://github.com/maxitg/SetReplace.

59