Finding and characterising physical states of Euclidean Abelianized loop quantum gravity using neural quantum states

This paper employs variational Monte Carlo with neural quantum states to characterize physical states of 4D Euclidean loop quantum gravity on a complete graph, revealing distinct solution families for the Hamiltonian constraint and its adjoint that correspond to the Ashtekar-Lewandowski and Dittrich-Geiller vacua, respectively, while also providing insights into their relationship with continuum solutions.

The Gist

The Big Picture: Building a Universe from Scratch

Imagine you are an architect trying to build a universe. In physics, the "blueprint" for our universe is General Relativity (Einstein's theory of gravity). But when you try to zoom in to the tiniest possible scale (the quantum level), the blueprint breaks down. The math gets messy, and the rules seem to contradict each other.

Loop Quantum Gravity (LQG) is a team of physicists trying to fix this blueprint. They propose that space isn't a smooth, continuous fabric, but is actually made of tiny, discrete "pixels" or "threads" woven together.

The problem? It's incredibly hard to find the specific patterns of these threads that actually look like our real, smooth universe. It's like trying to find a single, perfect melody in a room full of people shouting random notes.

The Experiment: A Digital Sandbox

In this paper, the authors (Hanno Sahlmann and Waleed Sherif) set up a digital sandbox to test this theory.

1. The Graph (The Skeleton): Instead of a full universe, they built a tiny, simplified skeleton of space. They used a shape called K5, which is like a 5-sided die where every corner is connected to every other corner. Think of it as a very small, rigid spiderweb.
2. The Rules (The Constraints): In their theory, there are strict rules (constraints) that any valid universe must follow. If the threads don't follow these rules, the universe collapses or doesn't make sense. The most important rule is the Hamiltonian Constraint, which dictates how space evolves and behaves.
3. The Problem: There are many ways to write these rules mathematically. It's like writing a recipe: you can say "add salt then pepper" or "add pepper then salt." In quantum physics, the order matters! The authors wanted to see if changing the order of the math changes the resulting universe.

The Tool: The Neural Network "AI Architect"

To find the right patterns of threads, they couldn't check every possibility (there are more combinations than atoms in the universe). Instead, they used Neural Network Quantum States (NQS).

Think of this as an AI Architect.

• The AI is given the rules (the constraints).
• It tries to guess a pattern of threads.
• It checks how well its guess follows the rules.
• If it fails, it tweaks the pattern and tries again.
• It does this millions of times until it finds a pattern that satisfies the rules almost perfectly.

The Discovery: Two Different Universes

Here is the surprising result. The authors tried two different orders for the math rules:

• Order A: "Do X, then Y."
• Order B: "Do Y, then X."

They expected the AI to find the same "perfect universe" regardless of the order. It didn't.

Instead, the AI found two completely different types of universes, depending on which order of rules it was given.

1. The "Flat, Spacious" Universe (Type A)

When the AI followed the first order, it built a universe that was:

• Flat: Like a calm, smooth ocean. The geometry was very regular.
• Spacious: It had a healthy amount of "volume." It felt like a real, 3D space.
• Chaotic but Normal: The tiny threads were all over the place, but they averaged out to look like a smooth, flat surface.
• The Metaphor: Imagine a Dittrich-Geiller vacuum. Think of this as a calm, flat lake. If you look at the water from far away, it's perfectly smooth. If you look close up, the water molecules are moving, but they don't clump together.

2. The "Crushed, Clumpy" Universe (Type B)

When the AI followed the second order, it built a universe that was:

• Curved: It wasn't flat; it had bumps and ripples.
• Crushed: The "volume" collapsed to almost zero. It was like a piece of paper crumpled into a tight ball.
• Clumpy: The threads were very specific and concentrated. They didn't spread out; they huddled together in a specific pattern.
• The Metaphor: Imagine an Ashtekar-Lewandowski vacuum. Think of this as a pile of sand where every grain is perfectly stacked in a specific, rigid tower. It's very structured, but it has no "room" inside; it's dense and collapsed.

The "Quasi-Solution": The Compromise

The authors then asked: "Can we get a universe that has the best of both worlds?"

They created a symmetric rule (a mix of Order A and Order B) and added some extra "penalties" to stop the AI from collapsing the universe or making it too flat.

The result was a Quasi-Solution. It was a hybrid:

• It had the "clumpy" correlations of the second universe (the threads were connected in interesting ways).
• But it kept the "spacious" volume of the first universe (it didn't collapse).
• It was a bit of a compromise, but it proved that by tweaking the math rules, you can steer the AI to find different kinds of physical realities.

Why Does This Matter?

1. Order Matters: In the quantum world, the order in which you do things changes the outcome. This paper proves that "Order A" and "Order B" aren't just different ways of writing the same thing; they lead to physically different universes. One is flat and spacious; the other is curved and collapsed.
2. AI is a Powerful Tool: This shows that using AI (Neural Networks) is a viable way to solve these incredibly complex physics problems. We can now "simulate" these universes and study their properties, like their shape and size, without needing a super-computer to calculate every single atom.
3. The Path Forward: This is a stepping stone. The authors used a simplified version of gravity (Abelian/flat). The next step is to use this same AI method to solve the real version of gravity (which is much more complex). If they can do that, we might finally understand how space and time emerge from the quantum foam.

Summary in One Sentence

By using an AI to solve the math of quantum gravity, the authors discovered that simply changing the order of the equations creates two totally different universes: one that is flat and spacious, and another that is curved and collapsed, proving that the "recipe" for reality is incredibly sensitive to how you mix the ingredients.

Technical Summary

1. Problem Statement

The central challenge in Canonical Loop Quantum Gravity (LQG) is the Hamiltonian constraint, which encodes the dynamics of the theory. Finding physical states (states in the kernel of the constraints) is technically difficult due to the infinite-dimensional nature of the kinematical Hilbert space and the complexity of the constraint operators.

This work addresses the problem by:

• Truncating the system: Working on a fixed, non-trivial graph (the complete graph $K_5$, the boundary of a 4-simplex) rather than the full continuum.
• Abelianization: Utilizing Smolin's weak coupling limit, reducing the $SU(2)$ gauge group to $U(1)^3$. This simplifies the algebra while retaining 4-dimensional kinematics and dynamics.
• Operator Ordering Ambiguity: Investigating whether different factor orderings of the Hamiltonian constraint (specifically $\hat{H}$, $\hat{H}^\dagger$, and their symmetric combination) lead to the same physical solution space or distinct sectors.

2. Methodology

Theoretical Framework

• Model: 4D Euclidean LQG in the $U(1)^3$ weak coupling limit.
• Graph: The complete graph $K_5$ (5 vertices, 10 edges). This graph is chosen for its high symmetry (reducing combinatorial artifacts) and non-planarity (requiring explicit embedding).
• Constraints: The authors construct the Thiemann Hamiltonian constraint $\hat{H}_v$ on the fixed graph. They define a quadratic "Master-like" constraint operator to minimize:
  • $\hat{Q}\hat{H} = \sum_v \hat{H}_v \hat{H}^\dagger_v$ (Standard ordering)
  • $\hat{Q}\hat{H}^\dagger = \sum_v \hat{H}^\dagger_v \hat{H}_v$ (Alternative ordering)
  • $\hat{Q}\hat{H}_S = \sum_v (\frac{1}{2}(\hat{H}_v + \hat{H}^\dagger_v))^2$ (Symmetric ordering)
• Cutoff: Representation labels (charges) on edges are truncated to a finite set $m \in [-m_{max}, m_{max}]$, effectively working with a quantum group $U_q(1)^3$.

Computational Approach

• Neural Quantum States (NQS): The wavefunction $\psi_\theta(\sigma)$ is parameterized by a neural network.
• Architecture: A Physics-Informed Neural Network (PINN) designed specifically for graph-based LQG.
  • It uses a Graph Neural Network (GNN) structure with edge-wise, vertex-wise, and global MLP towers.
  • It enforces gauge invariance and graph symmetries (permutation invariance of edges/vertices) by construction, significantly reducing the parameter space and improving sample efficiency.
  • The network outputs the complex log-amplitude (magnitude and phase) of the state.
• Optimization: Variational Monte Carlo (VMC) is used to minimize the expectation value of the quadratic constraints. The optimizer (Adam) searches for states where $\langle \hat{Q} \rangle \approx 0$.
• Software: Implemented using neuraLQX (built on NetKet, JAX, and Flax) with MPI-distributed execution.

3. Key Contributions

1. Scalable Numerical Framework: Demonstrated that NQS can handle LQG systems with Hilbert space dimensions far exceeding exact diagonalization capabilities (e.g., $m_{max}=5$ yields dimensions $\sim 10^{26}$), using a vanishingly small number of parameters ($< 10^{-20}$ compression rate).
2. Ordering-Induced Sector Selection: Proved that the factor ordering of the constraint is not a benign technical choice but a physical selector that drives the variational optimization into two distinct, robust families of solutions.
3. Comprehensive Characterization: Developed a multi-faceted diagnostic suite to characterize physical states beyond just the constraint value, including:
   • Normalizability analysis via stratified state-vector scanning.
   • Geometric observables (Area, Volume, Holonomies).
   • Long-range correlations (Chromaticity 1- and 2-point functions).
4. Interpolation Strategy: Successfully constructed "quasi-solutions" for the symmetric ordering that interpolate between the two distinct families by combining the symmetric constraint with penalty terms (no-vacuum projector and inverse-volume regularizer).

4. Key Results

The study identified two distinct solution families (Type-A and Type-B) and a third quasi-solution family:

Type-A Solutions (from $\hat{Q}\hat{H}$)

• Geometry: Non-degenerate volume. The expectation value of the volume operator $\langle \hat{V}_v \rangle > 0$ at all vertices.
• Flatness: Strictly flat. Minimal loop holonomies are trivial ($\langle \text{tr} \hat{h}_\alpha \rangle \approx 3$), indicating zero curvature.
• Correlations: Delocalized/Weak. Edge charge marginals are nearly uniform (maximally mixed), and connected 2-point correlation functions show no robust long-range structure (consistent with noise).
• Interpretation: These states resemble the Dittrich-Geiller (DG) vacuum (distributional, flat, non-normalizable in the kinematical inner product). They represent an extended, approximately isotropic 3-sphere geometry.

Type-B Solutions (from $\hat{Q}\hat{H}^\dagger$)

• Geometry: Volume Degenerate. The states lie in the kernel of the volume operator ($\langle \hat{V}_v \rangle \approx 0$).
• Flatness: Non-flat. Holonomies are non-trivial, indicating curvature excitations.
• Correlations: Localized/Structured. Charge marginals are sharply peaked at the zero-charge vector (Ashtekar-Lewandowski-like), with strong, stable correlations between adjacent edges (satisfying Gauss constraints locally).
• Interpretation: These states resemble the Ashtekar-Lewandowski (AL) vacuum (normalizable, zero volume, non-flat). They represent a "collapsed" geometry with sparse flux excitations.

Symmetric/Quasi-Solutions

• By minimizing the symmetric constraint $\hat{Q}\hat{H}_S$ with specific regularizers, the authors found states that:
  • Possess non-zero volume (avoiding Type-B collapse).
  • Exhibit Type-B-like charge correlations (strong adjacency).
  • Retain Type-A-like geometric flatness and isotropy.
• This demonstrates that the ordering dependence can be controlled to explore the space between these sectors.

Normalizability Analysis

• Type-A: As the cutoff $m_{max}$ increases, probability mass does not concentrate; it spreads to the outer shells. This indicates the states are non-normalizable in the kinematical inner product, consistent with distributional physical states.
• Type-B: Probability mass concentrates on the zero-charge sector as the cutoff increases, indicating normalizability.

5. Significance and Outlook

• Physical Insight: The work provides strong evidence that operator ordering in LQG is a physical parameter that selects different semiclassical sectors (flat/non-degenerate vs. curved/degenerate). This challenges the notion that different orderings merely represent different regularization schemes for the same physics.
• Methodological Advance: It establishes Variational Monte Carlo with NQS as a viable tool for solving the Hamiltonian constraint in LQG on non-trivial graphs, moving beyond symmetry-reduced models.
• Future Directions:
  • Extending the framework to the full non-Abelian $SU(2)$ theory, which introduces intertwiner degrees of freedom and more complex recoupling.
  • Using these numerical solutions to extract effective dynamics and define a notion of "quantum time" (Dirac quantization).
  • Investigating the continuum limit and the relationship between these discrete solutions and analytical results.

In summary, this paper successfully uses machine learning to navigate the complex landscape of LQG constraints, revealing that the choice of operator ordering fundamentally alters the geometric and topological nature of the resulting physical states.