Bipartite Graph QM

BigraphQM is a web-based implementation of a graph-theoretic approach to quantum mechanics.

In the current version, quantum systems can be modelled by manually building bipartite graphs. The method for calculating probabilities is based on finding maximal cliques by using a variant of the Bron-Kerbosch algorithm.

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

The app uses @hpcc-js/wasm for rendering graphs and KaTeX for displaying LaTeX notation. The model is a spin-off of my earlier project Hypergraph. For a philosophical take on these ideas see my blog post The Game of Algorithms.

Overview

We start with the assumption that at the lowest level of physical reality all the possible sequences of interactions get actualized.

We model these sequences as maximal paths in a directed bipartite graph. The model is computational so that each maximal path describes a sequence of operations and their operands.

For two sequences to interact, they must be local and mutually consistent (i.e. spacelike).

In our model, for two operands to interact, their maximal paths must compute the same function and be composable in parallel. Technically, the latter means that their lowest common ancestors must be operations, not operands.

Since consistency is a pairwise property, there can be situations in which some sequence $A$ is consistent with both $B$ and $C$, but $B$ and $C$ are not consistent with each other. We can think of these situations as second-order inconsistencies or superpositions. If we make a new graph in which sequences are nodes and edges connect consistent pairs, second-order inconsistencies show up as open triplets.

Interaction/observation is a process in which some internal part (observer) interacts with some other spacelike part (system) in a way that resolves these second-order inconsistencies. The result will be a set of mutually inconsistent outcomes in each of which every pair is connected. In graph theory, these kind of subgraphs are called maximal cliques.

There are typically many sequences (permutations, ordered sets) that lead to the same maximal clique (image, unordered set). From the observer's point of view, the proportion of all the sequences that lead to each maximal clique is the probability — or more precisely the weight — of that outcome.

NOTE: In our current model we consider all the maximal paths as local.

The Model

Basic concepts

Let $H$ be a directed bipartite graph with two parts: operations $V_1$ and operands $V_2$.

$\displaystyle\qquad H= (V_1 \cup V_2, E),$

$\displaystyle\qquad E\subseteq (V_1{\times}V_2)\cup (V_2{\times}V_1)$

At each new step, the lowest set of operands $L_2$ is connected to a new set of operations that compute the next generation of operands.

$\displaystyle\qquad L_2=\big\lbrace v \in V_2\mid\mathbf{deg^+} (v)=0 \big\rbrace$

Two operands are spacelike if and only if their lowest common ancestors (LCA) are operations. Let an undirected graph $G$ track all such pairs.

$\displaystyle\qquad G= (L_2, F),$

$\displaystyle\qquad F = \big\lbrace (a,b)\mid a,b\in L_2 \wedge\mathbf{LCA}_H(a,b)\subset V_1\big\rbrace$

NOTE: In our current model we consider all maximal paths as local in the sense that they compute the same function.

Probabilities

Let the sample space $\Omega$ be the neighbourhood graph $N_G$ relative to an observer $X$.

$\displaystyle\qquad\Omega= N_G(X)$

One way to find all the possible measurement outcomes $\mathcal{F}$ is to first enumerate all the sequences of interactions and then take their images. The problem is that the number of sequences grows factorially so that the algorithm has time complexity $O(n!)$.

Instead, we approach the problem by finding all the maximal cliques of $\Omega$.

Finding all maximal cliques is an NP-complete problem, but solvable within our sample space sizes. In the app we use a variant of the Bron-Kerbosch algorithm with the worst-case time complexity $O(3^{n\over 3})$ and for k-degenerate graphs $O(kn3^\frac{k}{3})$.

$\displaystyle\qquad\mathcal{F} = \mathbf{BK}(\varnothing,\Omega,\varnothing)$

ALGORITHM BK(R, P, X) IS
    IF P and X are both empty THEN
        report R as a maximal clique
    choose a pivot vertex u in P ⋃ X
    FOR each vertex v in P \ N(u) DO
        BK(R ⋃ {v}, P ⋂ N(v), X ⋂ N(v))
        P := P \ {v}
        X := X ⋃ {v}

Let $p_j$ be the proportion of all the consistent interactions leading to the clique $\mathcal{F_j}\in\mathcal{F}$. Informally $p_j$ is the probability of the outcome $\mathcal{F_j}$. However, if we buy into the philosophy that all the consistent interactions are realized, we should really think of it as a clique weight.

$\displaystyle\qquad p_j={{|\mathcal{F_j}|!}\over{\sum\limits_{k} |\mathcal{F_k}|!}},\quad j\in \lbrace 1,\dots,|\mathcal{F}| \rbrace$

Based on the weights, we can calculate the weight $\lambda_i$ for the token $v_i\in\Omega$ to be included in some (any) clique.

$\displaystyle\qquad \lambda_i=\sum_{j}\mathbf{1}_\mathcal{F_j}{(v_i)} p_j,\quad i\in\lbrace 1,\dots,|\Omega|\rbrace$

Energy & Entropy

Let $E_{free}$ be the system's free energy. One interaction is considered equivalent to one unit of energy.

$\displaystyle\qquad E_{free} = \sum_j |\mathcal{F_j}||\mathcal{F_j}|!$

The observer gets only one sequence at measurement, so the expected energy $\langle E\rangle$ can be calculated by using the clique weights and sizes.

$\displaystyle\qquad \langle E\rangle = \sum_j p_j|\mathcal{F_j}|$

Using the clique weights we can calculate the Shannon entropy $H_1$.

$\displaystyle\qquad H_1 = -\sum_j p_j \log p_j$

Past Hypothesis

Suppose we start with the past hypothesis so that all the tokens in $\Omega$ are mutually consistent. It means that the system has the lowest entropy and the highest free energy.

$\displaystyle\qquad H_{1,min} = 0,\quad E_{max} = |\Omega| |\Omega|!$

If the system interacts only with itself, it tends to become more inconsistent over time. This seems to be compatible with the second law of thermodynamics.

Furthermore, since each interaction decreases the average size of the cliques, and thus the probability for that specific branch, the principle of least action applies and determines the most likely path.

Eventually the system will reach its highest entropy state with the lowest free energy.

$\displaystyle\qquad H_{1,max} = - \log \frac{1}{|\Omega|},\quad E_{min} = |\Omega|$

Entanglement

Let $D_{A,B}$ be the overlap between two cliques.

$\displaystyle\qquad D_{A,B} = {{2|A\cap B|}\over{|A|+|B|}}$

We can extend this measure to all combinations.

$\displaystyle\qquad D = {{\sum 2|A\cap B|}\over{\sum(|A|+|B|)}},\quad {(A,B)}\in\binom{\mathcal{F}}{2}$

If $D=0$, none of the outcomes overlap. If $D=1$, all the outcomes are the same. The measure is not defined with only one outcome.

Density matrix

The pure quantum state of each clique can be presented as a linear combination of orthonormal vectors $|\phi_i\rangle$ representing the tokens $v_i\in\Omega$. The square root left to the sum is the normalizing constant.

$\displaystyle\qquad |\psi_j\rangle = \frac{1}{\sqrt{|\mathcal{F_j}|}}\sum_{i=1}^{n}\mathbf{1}_\mathcal{F_j}{(v_i)}|\phi_i\rangle$

The density matrix $\rho$ is the weighted sum of the outer products of the pure states.

$\displaystyle\qquad \rho=\sum_{j=1}^{m} p_j|\psi_j\rangle\langle\psi_j|$

NOTE: The density matrix, as described here, is a limited projection and unable to represent the detailed ancestral structure. The bipartite graph is the actual data structure of the model.

Phases

Based the overlap of two cliques $D_{A,B}$ we can calculate their relative phase $\theta_{A,B}$ in radians.

$\displaystyle\qquad \theta_{A,B} = (1 - D_{A,B})\frac{\pi}{2}$

Computational Basis

From the observer's point of view all the tokens are identical. It means that they can't distinguish between two local cliques with the same number of tokens. At measurement they might, however, distinguish the number of interactions (i.e. energy), which is proportional to the size of the clique.

So, in order to create a two-state system that the observer can use for computation, we can define a qubit so that $|0\rangle$ represents cliques of size one (ground state) and $|1\rangle$ cliques of size two (first excited state). This scheme generalises to d-state system i.e. qudits.

The Editor

The bipartite graph on the app describes one possible set of consistent sequences relative to the observer. It doesn't follow sequences that have become inconsistent with the observer, because the two parts can't interact with each other.

To ensure consistency, the editor only allows new edges between spacelike elements.

SYMBOL	DESCRIPTION
^TOKEN	Tokens are abstract elements. The value inside the symbol represent the weight $\lambda\in [0,1]$ of the token. Informally it is the probability that the token would end up being a part of some (any) classical state at (hypothetical) measurement.
^EVENT	Events use tokens as input and produce tokens as output.
^EDGE	Directed edges represent input/output relations between tokens and events. Only consistent (i.e. spacelike) connections are allowed.
^OBSERVER	Observer represents a classical system relative to which probabilities are calculated. The observer, too, is a bipartite subgraph, but instead of showing individual tokens and events, they are grouped together for simplicity.
^ENVIRONMENT	Environment is a classical system that can interfere with the sample space.
^CLIQUE	Maximal clique in probability space represents one possible classical outcome. The value inside the symbol is the probability for the observer to see that outcome at (hypothetical) measurement.

User Actions / Shortcuts:

Click on a token/event/clique to select it. Remove all the selections by clicking on the background.
Double click on a token/event to add a new branch.
Drag a token/event on top of another event/token to connect the two with an edge. Note: Only consistent (spacelike) connections are allowed.
Drag a token to an empty space before or after another token to reorder the tokens.
Click on an edge to select it.
Double click on an edge to delete it.

TOOL	DESCRIPTION
^UNDO	Undo/redo operations. The maximum size of the undo buffer is 50 operations. The app also uses localStorage to store the current graph so that if you refresh the page without any URL parameters, you can restore the previous working model.
^ZOOM	Zoom the graph. If you want to increase/decrease the font size, zoom the browser window first and scale your graph to fit the screen using these zoom buttons.
^TIME	Decrease/increase time steps. When a new time step is added, new branches are added to the existing spacelike tokens. By decreasing the time steps you can go back all the way to the initial state.
^ADD	Add a new branch to the selected token/event. The new branch is automatically extended to the bottom of the graph.
^DEL	Delete the selected token/event/edge together with all its parents and children that are dependent on it.
^MEASURE	Shortcut for making a random proto-measurement.
^DECOHERE	Shortcut for adding random decoherence to the system.
B C ^VIEW	Switch to an alternative view: B) Maximal cliques (classical states) and their probabilities, C) the final state with the induced graph G' and its clique graph K(G'). Note: While an alternative view is shown the editing mode is disabled.
^INFO	Show the sidebar with details. The details include information about the selected tokens, probabilities, direct URL link to the graph, and the DOT source describing the shown graph. Refer to The Model for an explanation of the presented mathematical concepts.
^URL	Copy URL to clipboard. Parameter `g` includes the Base64 encoded bipartite graph.
^SVG	Download the shown bipartite graph as a SVG file.
^GitHub	Link to the GitHub project.

Gallery

MODEL	DESCRIPTION
Wave function collapse Open in App	All events are local by definition. However, new edges can add shortcuts through the graph's ancestral structure and break some existing spacelike relations. These instantaneous and non-local effects are known as the wave function collapse.
Bell state (entanglement) Open in App	Bell state is an example of a maximally entangled state. Here we start with two sequences $A$ and $B$. They both split into two inconsistent (orthogonal) branches. Lets call them \|↑> and \|↓>. These branches get entangled so that the observer can only see either \|↑↓> or \|↓↑>, but never \|↑↑> or \|↓↓>.
Decoherence Open in App	Decoherence is a process in which the system interacts with its environment so that the overlap decreases and the states become separable.
Interference Open in App	Two spacelike parts of the sample space can interact and thus interfere with each other. This changes the probabilities over time and disturbs the probability wave. This can be seen at measurement as an interference pattern. - Note: The shown example is the result of random interactions.

Notes and Acknowledgements

The idea that every possible outcome is realised is based on Many-Worlds Interpretation.
The terminology of consistent and inconsistent sequences is taken from Consistent Histories.
The idea that all the facts/outcomes are relative comes from Carlo Rovelli's Relational Quantum Mechanics.
The graph-based approach is heavily influenced by the Wolfram Model and Max Piskunov's work on Local Multiway Systems (SetReplace).
Several parts of the code are from my earlier project Hypergraph.
Many of the ideas were originally discussed with Tuomas Sorakivi.

Appendix A: The Born Rule

The Born rule is often considered a key postulate of quantum mechanics. It says that the absolute value of the vectors inner product, or equivalently the cosine squared of the angle between the lines the vectors span, corresponds to the transition probability.

The reason why the cosines are squared in the Born rule is geometric. First, Kolmogorov's axioms tell us that the probabilities should add up to one. Second, as shown below, the sum of cosines squared equals to one in high dimensional vector spaces:

Let $r$ be the length of a hypervector in $n$ dimensional space.

$\displaystyle\qquad r =\sqrt{x_1^2+x_2^2+\dots+x_n^2} \qquad \Bigg| \cdot ()^2 \quad \Bigg| \cdot {1 \over {r^2}}$

$\displaystyle\qquad 1 = \left({{x_1} \over r}\right)^2+\left({{x_2} \over r}\right)^2 + \dots + \left({{x_n} \over r}\right)^2$

$\displaystyle\qquad 1 = cos^2 \theta_1 + cos^2 \theta_2 + \dots + cos^2 \theta_n \quad\square$

This works both ways. So, if we want to move from classical probabilities to quantum probabilities - that is, from sets to vector spaces - we need to take square roots to get the probability amplitudes.

met4citizen/BigraphQM