Statistical mechanics in continuous space with tensor network methods

This paper extends tensor network methods to interacting particle systems in continuous space by formulating an effective lattice model through real-space discretization and coarse-graining, successfully applying the framework to the two-dimensional hard-disk problem to demonstrate its advantages over traditional Monte Carlo simulations.

The Gist

Imagine you are trying to understand how a crowd of people moves inside a room. In physics, this is similar to studying how particles (like atoms) behave in a gas or liquid. Usually, scientists use a method called "Monte Carlo simulation," which is like sending thousands of random scouts into the room to guess where people are standing. It's powerful, but it can be slow, and it sometimes struggles to give you the exact "cost" (free energy) of the whole system.

This paper introduces a new, more structured way to solve this problem using something called Tensor Networks (TN). Think of Tensor Networks not as random scouts, but as a highly organized, grid-based map that captures the rules of the room perfectly.

Here is a simple breakdown of what the authors did:

1. Turning a Continuous Room into a Grid

In the real world, particles can be anywhere in a continuous space (like a smooth floor). The authors realized that Tensor Networks work best on a grid (like a chessboard).

• The Trick: They didn't just chop the floor into tiny squares. Instead, they used a "cell-based" approach. Imagine grouping a small cluster of chessboard squares into one big "super-square" (a cell).
• The Rule: Inside each of these "super-squares," they applied a simple rule: either the whole cell is empty, or exactly one particle is in it. This is like saying, "In this small neighborhood, only one person can stand at a time."
• Why? This simplifies the math massively. It turns a messy, continuous problem into a neat, local puzzle that the Tensor Network can solve efficiently.

2. The "Infinite" Map vs. The "Box"

The authors tested their method in two ways:

• The Infinite Map: They used a technique to simulate an infinitely large room. This allows them to see what happens when the system gets huge, without having to build a bigger and bigger computer model. It's like looking at a pattern that repeats forever.
• The Box: They also simulated a specific, finite room with walls. This was crucial for watching a phase transition—specifically, when a liquid turns into a solid (like water freezing into ice). In their simulation, they could watch the particles spontaneously line up into a crystal structure as they got crowded, something that is hard to capture with standard random methods.

3. The Big Win: Calculating the "Price Tag"

The most significant claim in the paper is about Free Energy.

• The Problem: In standard simulations, calculating the "absolute free energy" (think of this as the total price tag or the fundamental cost of the system's state) is incredibly difficult. It's like trying to count every single grain of sand on a beach to find the total weight. The standard method (Wang-Landau algorithm) gets exponentially harder as the system gets bigger.
• The Solution: Because Tensor Networks represent the whole system as a connected map, calculating this "price tag" becomes much easier. The authors showed that as they made the system bigger, the time it took to calculate the energy only went up linearly (like adding one step at a time), whereas the old method went up exponentially (like doubling the effort every single time).

4. The Results

They tested this on a classic physics problem: Hard Disks. Imagine a floor covered in coins that cannot overlap.

• They calculated how dense the coins get and how they arrange themselves.
• Their results matched the standard "random scout" (Monte Carlo) methods perfectly, proving their new map is accurate.
• They successfully captured the moment the coins stopped flowing like a liquid and started locking into a solid crystal pattern.

Summary

The paper claims to have successfully taken a powerful mathematical tool (Tensor Networks), which was usually only used for grid-based problems, and adapted it to work for particles moving in continuous space. By creating a smart "cell" system, they proved this method is:

1. Accurate: It matches existing gold-standard simulations.
2. Efficient: It calculates the total energy of the system much faster as the system grows.
3. Versatile: It can handle both infinite systems and the tricky transition from liquid to solid.

In short, they built a better, more efficient map for navigating the complex world of interacting particles.

Technical Summary

1. Problem Statement

Traditional Tensor Network (TN) methods have been highly successful in classical statistical mechanics, but their application has been largely restricted to lattice models (e.g., Ising, XY models). Extending TN methods to interacting particle systems in continuous space has proven difficult. Existing approaches for continuous spaces often rely on a "particle" picture, where TNs capture inter-particle correlations, but this fails to exploit spatial locality, a key feature that makes TNs efficient for lattice models.

The authors aim to bridge this gap by formulating an effective lattice model for continuous interacting particle systems that preserves spatial locality, enabling the use of standard TN contraction techniques to compute partition functions and thermodynamic quantities without the sampling issues inherent in Monte Carlo (MC) methods.

2. Methodology

The authors propose a framework combining real-space discretization with a cell-based coarse-graining scheme to map continuous systems onto effective lattice models.

A. Discretization and Representation

• From Particle to Site Representation: The continuous grand canonical partition function $\Xi$ is first discretized onto a fine grid $G$. Instead of summing over particle coordinates (particle representation), the system is reformulated in terms of site occupation numbers $n_k \in \{0, 1\}$ (site representation).
• Cell-Based Coarse-Graining: To handle the short-range repulsion typical of interatomic potentials (like hard disks), the fine grid is grouped into cells.
  • Constraint: Within each cell, the maximum distance between sites is smaller than the interaction cutoff $r_c$. Consequently, the model enforces a constraint that a cell can be either empty or contain exactly one particle.
  • Benefit: This reduces the configuration space per cell from $2^N$ to $N+1$ (where $N$ is the number of fine grid points in the cell), significantly lowering the effective bond dimension required for the TN.

B. Tensor Network Construction

• Factor Graph to TN: The partition function is expressed as a factor graph with on-site terms ($g_k$) and pairwise interaction terms ($f_{kk'}$).
• Interaction Range: The authors restrict interactions to Nearest-Neighbor (NN) and Next-Nearest-Neighbor (NNN) cells.
• Lattice Transformation: Following a scheme for NNN interactions, the factor graph is transformed into a 2D TN on a square lattice rotated by $\pi/4$.
  • Tensors: Two types of tensors are defined:
    • $S$ (Site Tensor): Represents the on-site weight, contracted with identity tensors.
    • $T$ (Plaquette Tensor): Represents the pairwise interactions between four surrounding sites, constructed by taking the square root of interaction terms to distribute them symmetrically.
• Bond Dimension ($D$): The bond dimension corresponds to the number of allowed configurations per coarse-grained cell.

C. Contraction Strategy

The authors employ boundary contraction using Matrix Product States (MPS) to compute the partition function:

1. Infinite System (Thermodynamic Limit): Uses a uniform MPS with a $2 \times 2$ unit cell (due to the checkerboard arrangement of $S$ and $T$ tensors) to directly compute results in the thermodynamic limit.
2. Finite System (Open Boundary Conditions): Uses successive randomized compression algorithms to handle finite boxes, allowing for the study of phase transitions and boundary effects.

3. Key Contributions

• Extension to Continuous Space: Successfully mapped continuous interacting particle systems to effective lattice models suitable for TN methods while preserving spatial locality.
• Deterministic Free Energy: Demonstrated the ability to compute absolute free energies directly, a task notoriously difficult for Monte Carlo methods which typically rely on thermodynamic integration or flat-histogram techniques (like Wang-Landau) that scale poorly.
• Linear Scaling: Showed that the computational cost of TN contraction scales linearly with system size for a fixed bond dimension, contrasting sharply with the exponential scaling observed in Wang-Landau algorithms for similar problems.
• Phase Transition Capture: Applied the method to the 2D hard-disk system, successfully capturing the liquid-to-solid phase transition and symmetry breaking.

4. Results

The framework was tested on the 2D Hard-Disk (HD) potential ($V(r) = \infty$ for $r < \sigma$, $0$ otherwise).

• Density and Convergence:
  • At low to intermediate chemical potentials ($\mu$), the TN results for density converged rapidly with low bond dimensions ($\chi$).
  • Near the liquid-to-solid transition ($\mu > 10$), convergence slowed. The authors attribute this to the spontaneous breaking of translational invariance in the solid phase, which requires a larger bond dimension to represent the superposition of solid phases within a translationally invariant MPS.
• Pair Correlation Functions ($g(r)$):
  • TN results for $g(r)$ matched standard Markov Chain Monte Carlo (MCMC) results in the continuum limit.
  • While TN results showed high-frequency oscillations due to lattice discretization, they accurately captured the overall envelope and the increasing long-range order.
• Free Energy and Computational Cost:
  • TN: The free energy density converged with system size, and the required bond dimension for a target accuracy remained independent of system size.
  • Wang-Landau (WL) Comparison: The WL algorithm required a number of Monte Carlo steps that grew exponentially with system area to achieve comparable accuracy. In contrast, the TN method maintained linear scaling, offering a massive efficiency advantage for large systems.
• Phase Transition Visualization: Finite box simulations with Open Boundary Conditions (OBC) successfully visualized the formation of a crystal structure at high $\mu$, demonstrating the method's ability to handle symmetry breaking without enforcing translational invariance.

5. Significance and Outlook

This work establishes a robust pathway for applying Tensor Network methods to continuous space statistical mechanics.

• Efficiency: It offers a deterministic alternative to stochastic sampling, avoiding the "sign problem" and large-scale sampling issues of MC methods.
• Scalability: The linear scaling of computational cost with system size makes it feasible to study large systems and compute absolute free energies, which are critical for phase diagram construction.
• Future Directions: While currently demonstrated in 2D, the authors note the method is extensible to 3D using 3D TNs. Future work aims to extend the framework to longer-range potentials and non-equilibrium systems.

In summary, the paper provides a versatile tool for studying complex continuous systems in equilibrium, bridging the gap between the efficiency of lattice-based TN methods and the physical realism of continuous particle models.