mgm79 / Hypergraph

Hypergraph Rewriting System

Hypergraph Rewriting System

Run it: https://met4citizen.github.io/Hypergraph/

This is a Hypergraph Rewriting System that is able to visualize both single-way and multiway evolutions in 3D.

The app uses 3d Force-Directed Graph for representing graph structures, ThreeJS/WebGL for 3D rendering and d3-force-3d for the force engine.

Introduction

A hypergraph is a generalization of a regular graph so that whereas an edge typically connects only two nodes, a hyperedge can join any number of nodes. In a Hypergraph Rewriting System some initial hypergraph is transformed incrementally by making a series of updates that follow some abstract rewriting rule.

As an example, consider a rewriting rule (x,x,y)(y,z,u)->(x,v,u)(y,v,z)(v,v,u). Wherever/whenever a subhypergraph having the form of the left-hand side pattern (x,x,y)(y,z,u) is found in the hypergraph, it is replaced with a new subhypergraph having the form of the right-hand side pattern (x,v,u)(y,v,z)(v,v,u).

Sometimes matches overlap. For example, when using the previous rule with the initial state (1,1,2)(2,2,3)(3,3,4) there are two overlapping matches (x=1,y=2,z=2,u=3) and (x=2,y=3,z=3,u=4) (critical pair). One way to resolve this conflict is to pick just one of the two according to some ordering scheme and ignore the other (single-way evolution). Another approach is to rewrite both by allowing the system to branch (multiway evolution).

As the multiway system branches and diverges (quantum mechanics), the probability of ending up in some particular end state is related to the number of different evolutionary paths to that state (path counting). However, there can also be rules that make branches merge (critical pair completion). This branching and merging makes certain end states more/less likely (constructive/destructive interference). In the end we (observers) are likely to find ourselves in the part of the system in which the branches always merge (confluence) and the system converges (classical mechanics).

For more information about hypergraph rewriting systems and their potential to represent fundamental physics visit The Wolfram Physics Project website. According to their technical documents certain models reproduce key features of both relativity and quantum mechanics.

Rules

Click RULE to modify the rewriting rule, change its settings, and run the rewriting process.

The system supports several rules separated with a semicolon ; or written on separate lines. The two sides of any one-way rule must be separated with an arrow ->. The separator == can be used as a shortcut for a two-way setup in which both directions (the rule and its inverse) are possible.

Hyperedge patterns can be described by using numbers, characters or words. Several types of parentheses are also supported. For example, a rule [{x,y}{x,z}]->[{x,y}{x,w}{y,w}{z,w}] is considered valid and can be validated and converted to the default number format by clicking Scan.

The system also supports a filter/negation \. As an example, the rule (1)(1,2)\(2)->(1)(1,2)(2) is applied only if there is no unary edge (2). If the branchlike interactions are allowed, the check is made relative to all possible branches of history.

A rule without any right-hand side, such as (1,1,1)(1,1,1), is used as the initial graph. An alternative way to define an initial state is to use some predefined function:

Initial graph	Description
complete(n)	Complete graph with n vertices so that each vertex is connected to every other vertex.
grid(d1,d2,...)	Grid with sides of length d1, d2, and so on. The number of parameters defines the dimension of the grid. For example, grid(2,4,5) creates a 3-dimensional 2x4x5 grid.
line(n)	Line with n vertices.
points(n)	n unconnected unary edges.
prerun(n)	Pre-run the current rule for n events in one branch (singleway). The leaves of the result are used as an initial state.
random(n,d,nedges)	Random graph with n vertices so that each vertex is sprinkled randomly in d dimensional space and has at least nedges connections.
rule('rule',n)	Run rewriting rule rule for maximum n events in one branch (singleway). The leaves of the result are used as an initial state.
sphere(n)	Fibonacci sphere with n vertices.

It is also possible to define some specific branch, or a combination of branches, for the initial state. As an example, (1,1)(1,1)/7, would specify branches 1-3 (the sum of the first three bits 1+2+4). By default, the initial state is set for all the tracked branches.

If the initial state is not specified, the left-hand side pattern of the first rule is used, but with only a single node. For example, a rule (1,2)(1,3)->(1,2)(1,4)(2,4)(3,4) gives an initial state (1,1)(1,1).

The Evolution option defines which kind of evolution is to be simulated:

Evolution	Description
1,2,4	Single-way system with 1/2/4 branches. By default, random event order is used to resolve overlaps for each branch. If the WM option is set, Wolfram model's standard event order is used for branch 1 and its reverse for branch 2.
FULL	Full multiway system. All the matches are instantiated and four branches tracked.

The Interactions option defines the possible interactions between hyperedges. Any combination of the three possible interactions can be selected.

Interactions	Description
SPACE	Allow interactions between spacelike separated hyperedges. Two edges are spacelike separated if their lowest common ancestors are all updating events. In practice this should always be selected, because the nodes in a typical initial state are spacelike separated.
TIME	Allow interaction between timelike separated hyperedges. Two edges are timelike separated if either one of them is an ancestors of the other one, that is, inside the other's past causal cone.
BRANCH	Allow interaction between branchlike separated hyperedges. Two edges are branchlike separated if any of their lowest common ancestors is a hyperedge.

Other options:

Option	Description
WM	Wolfram Model. If set, the first branch uses Wolfram Model's standard event order (LeastRecentEdge + RuleOrdering + RuleIndex) and the second branch its reverse. By default the setting is off and all tracked branches use random event ordering.
RO	Rule order (index). Regardless of other settings, always try to apply the events in the order in which the rules have been specified. By default the setting is off and the individual rules are allowed to mix.
DD	De-duplicate. The overlapping new hyperedges on different branches are de-duplicated at the end of each step. This allows branches to merge. EXPERIMENTAL, FUNCTIONALITY LIKELY TO CHANGE.

Simulation/Observer

Simulation can be run in three different modes: Space, Time or Phase.

• In Space mode the system shows the evolution of the spatial hypergraph. According to the Wolfram Model, the spatial hypergraph represents a spacelike state of the universe with nodes as "atoms of space".
• In Time mode the system builds up the transitive reduction of the causal graph. In this view nodes represent updating events and directed edges their causal relations. According to the Wolfram Model, the flux of causal edges through spacelike and timelike hypersurfaces is related to energy and momentum respectively.
• In Phase mode hyper-dimensional multiway space is projected in 3D by first constructing a graph based on groups of hyperedges and their nearest neighbours (k-NN) using Hamming distances. See Appendix A: Coordinatization of Local Multiway System by using Hyper-dimensional Vectors. According to Wolfram Model, positions in this so-called "branchial" space are related to quantum phase.

Media buttons let you reset the mode, start/pause the simulation and skip to the end. Whenever the system has branches, the first four branches can be shown separately or in any combination. If Past is selected, the full history of the local multiway system is shown in space mode. By default only the leaf edges of the system are visible.

The two sliders change the visual appearance of the graph by tuning the parameters of the underlying force engine. Note: Changing the viewpoint or the forces do not in any way change the multiway system itself only how it is visualized on the screen.

Highlighting

Subhypergraphs can be highlighted by clicking RED/BLUE and using one or more of the following commands:

Command	Highlighted	Status Bar
curv(x,y)	Two n-dimensional balls of radius one and the shortest path between their centers.	Curvature based on Ollivier-Ricci (1-Wasserstein) distance.
dim([x],[radius])	N-dimensional ball with an origin x (random, if not specified) and radius r (automatically scaled if not specified).	The effective dimension d based on nearby n-ball volumes fitted to r^d.
geodesic(x,y,[dir],[rev],[all]) dir = directed edges rev = reverse direction all = all shortest paths	Shortest path(s) between two nodes.	Path distance as the number of edges.
lightcone(x,length)	Lightcone centered at node x with size length. TIME mode only.	Size of the cones as the number of edges.
phase(x,y)	Multiway distance between two nodes.	Multiway distance 0-10,240. Mid-point 5,120 corresponds to orthogonal vectors (phase difference PI).
nball(x,radius,[dir],[rev]) dir = directed edges rev = reverse direction	N-dimensional ball is a set of nodes and edges within a distance radius from a given node x.	Volume as the number of edges.
nsphere(x,radius,[dir],[rev]) dir = directed edges rev = reverse direction	N-dimensional sphere/hypersurface within a distance radius from a given node x.	Area as the number of nodes.
random(x,distance,[dir],[rev]) dir = directed edges rev = reverse direction	Random walk starting from a specific node x with maximum distance.	Path distance as the number of edges.
surface(x,y)	Space-like hypersurface based on a range of nodes.	Volume as the number of nodes.
worldline(x,...)	Time-like curve of space-like node/nodes. TIME mode only.	Distance as the number of edges.
(x,y)(y,z) (x,y)(y,z)\(z)->(x,y)	Subhypergraphs matching the given rule-based pattern. The right hand side pattern can be used to specify which part of the match is highlighted. SPACE mode only.	The number of rule-based matches.

The node parameters can be specified with identifiers you can find by hovering the mouse pointer over nodes. It is also possible to use variables x and y. To set variable x click on any node and to set y use right click.

Scalar Fields

Click GRAD to visualize predefined scalar fields. Relative intensity of the field is represented by different hues of colour from light blue (lowest) to green to yellow (mid) to orange to red (highest). Field values are calculated for each edge and the colours of the vertices represent the mean of their edges.

Scalar Field	Description
branch	Branch id. With two branches the main colours are blue and red. With four branches blue, green, orange and red. For shared edges the colour is in the middle of the spectrum.
created	Creation time from oldest to newest.
curvature	Ollivier-Ricci curvature. NOTE: Calculating curvature is a CPU intensive task. When used in real-time it will slow down the animation.
degree	The mean of incoming and outgoing edges.
energy	The mean of updated edges.
mass	The part of energy in which the right hand side edges connect pre-existing vertices.
momentum	The part of energy in which the right hand side edges have new vertices.
pathcnt	The number of paths leading to specific edge.
phase(x)	Multiway distance to a given token x. Multiway coordinates are calculated using 10,240-dimensional dense bipolar hypervectors. Distances close to 5,120, that is, 1/2 of the dimension, can be considered orthogonal. NOTE: Hyperdimensional computing is CPU intensive, so when used in real-time it will slow down the animation.
probability	Normalized path count for each edge in each step.
step	Rewriting step.

The value range can be limited by giving lower and higher limits as parameters. For example, branch(2,4) shows branches 2-4. A limit can also be given as a percentage. For example, energy(50%,100%) highlights only upper half of the full range.

Notes

The aim of this project has been to learn some basic concepts and ideas related to hypergraphs and hypergraph rewriting. Whereas the Wolfram physics project has been a great inspiration, this project is not directly associated with it, doesn't use any code from it, and doesn't claim to be compatible with the Wolfram Model.

As a historical note, the idea of "atoms of space" is not a new one. It already appears in the old Greek tradition of atomism (atomos, "uncuttable") started by Democritus (ca. 460–370 BCE). In the Hellenistic period the idea was revived by Epicurus (341–270 BCE) and put in a poetic form by Lucretius (ca. 99–55 BCE). Unfortunately, starting from the Early Middle Ages, atomism was mostly forgotten in the Western world until Lucretius' De Rerum Natura and other atomist teachings were rediscovered in the 14th century.

“The atoms come together in different order and position, like letters, which, though they are few, yet, by being placed together in different ways, produce innumerable words.” -- Epicurus (according to Lactantius)

Appendix A

Coordinatization of Local Multiway System using Hyperdimensional Vectors

The following idea/specification is based on Pentti Kanerva's work on Hyperdimensional computing [1]. It is based on the fact that in a hyperdimensional space (>10,000-D) a randomly chosen vector (hypervector) is quasi-orthogonal to all previously generated hypervectors.

• Let the coordinate of the initial event in a local multiway system be a randomly generated 10,240-dimensional dense bipolar {-1,1} vector.

• Let the coordinate of each new token/event be the sum of its parents' coordinates so that the 'sum' is the element-wise majority with random bipolar values for ties.

• Whenever the parents are timelike/branchlike separated, instead of summing up their own coordinates, use the coordinates of their lowest common ancestors.

• Whenever there is a new branch so that two or more events use an overlapping set of tokens, separate the events in orthogonal branches by adding random hypervectors to their previously calculated coordinates.

• The distance between the coordinates of any two tokens/events is the Hamming distance between their hypervectors. If the distance is close to 1/2 of the used dimension, the two hypervectors are quasi-orthogonal.

The local multiway coordinates can be projected into some lower dimension by first calculating the distance matrix and then doing multidimensional scaling. One possible way to do the projection is to build a graph in which each node represents a group of close-by tokens/events. These groups can then be connected to their 1-3 nearest neighbours by using k-NN algorithm.

[1] Kanerva, P. (2019). Computing with High-Dimensional Vectors. IEEE Design & Test, 36(3):7–14.