Initial tensor construction and dependence of the tensor renormalization group on initial tensors

This paper proposes a singular-value-decomposition-free method for constructing initial tensor networks and demonstrates that while Tensor Renormalization Group (TRG) results depend heavily on these initial tensors, this dependence can be eliminated and numerical robustness significantly improved by employing the boundary TRG technique.

The Gist

Imagine you are trying to solve a massive, incredibly complex jigsaw puzzle. This puzzle represents the universe of a physical system (like a magnet or a grid of tiny magnets). The goal is to understand the "big picture" behavior of the whole system without getting lost in the billions of tiny pieces.

This paper introduces a new, smarter way to build the puzzle and a better way to solve it, specifically for a method called Tensor Renormalization Group (TRG).

Here is the breakdown using simple analogies:

1. The Problem: The "Messy" Puzzle Box

In physics, to calculate things like temperature or energy, scientists turn the system into a giant network of numbers (tensors).

• The Old Way: To make this network solvable, previous methods required a lot of complicated math "pre-processing." It was like trying to fit a square peg in a round hole. You had to stretch, shrink, and expand the pieces (using things called "Taylor expansions" or "Singular Value Decompositions") to make them fit a specific shape.
• The Issue: If you changed how you pre-processed the pieces (the "initial tensor"), the final answer sometimes changed too. It was like if you built the puzzle base differently, the picture you ended up with looked slightly different, even though the puzzle pieces were the same. This made the results unreliable.

2. The New Solution: The "Identity" Shortcut

The authors propose a much simpler way to build the initial puzzle pieces.

• The Analogy: Instead of trying to reshape the pieces, they just take the raw pieces and insert a "transparent sheet" (an identity matrix) between them.
• Why it's cool: It's like taking a photo and just adding a clear frame around it. You haven't changed the photo; you've just organized it so the camera (the computer) can take a picture of it easily.
• The Benefit: This method works for almost any system (magnets, gauge theories, etc.) without needing complex, system-specific math. It's a "one-size-fits-all" recipe.

3. The Discovery: The "Symmetry" Trap

The authors tested their new puzzle pieces against different ways of solving the puzzle (different TRG algorithms). They found a surprising flaw in some popular methods:

• The Trap: Some algorithms were like biased judges. If the puzzle pieces were perfectly symmetrical (like a snowflake), the judge gave a perfect score. But if the pieces were slightly lopsided (asymmetrical), the judge got confused and gave a terrible score.
• The Reality: Mathematically, a lopsided piece and a symmetrical piece can represent the exact same thing. But the computer algorithm treated them differently, leading to errors.

4. The Fix: The "Squeezer" Technique

The paper shows how to fix these biased judges.

• The Analogy: Imagine you are trying to compress a suitcase.
  • Old Method (Isometry): You try to force the suitcase shut by pushing on one side. If the clothes inside are arranged unevenly, the suitcase won't close properly, and you lose some clothes (information).
  • New Method (Squeezer/Boundary TRG): You use a special compression tool that squeezes the suitcase evenly from all sides at once. It doesn't matter how the clothes are arranged inside; the suitcase closes perfectly every time.
• The Result: By using this "squeezer" technique, the algorithms stop caring about the shape of the initial pieces. They become robust. Whether you start with a symmetrical or lopsided puzzle, you get the same, accurate answer.

5. Real-World Testing: The 3D Gauge Theory

To prove it works, they tested this on a very complex 3D system (Z2 gauge theory), which is like trying to solve a 3D Rubik's cube that changes shape as you turn it.

• The Challenge: Usually, solving this requires "gauge-fixing" (a complicated rule to simplify the rules of the game).
• The Win: The authors solved it without those complicated rules. Their simple "identity sheet" method worked just as well as the complex, old-school methods. They calculated the "specific heat" (how much energy it takes to warm the system) and got results that matched the best supercomputer simulations.

Summary: Why Should You Care?

Think of this paper as an upgrade to the operating system of physics simulations.

1. Simpler Setup: You don't need to be a math wizard to set up the initial problem anymore.
2. More Reliable: The results don't depend on how you accidentally set up the starting numbers.
3. Versatile: It works on 1D, 2D, and 3D systems, and even on systems with long-range interactions (where pieces far apart affect each other).

In short, the authors found a way to make the "puzzle solver" less picky about the "puzzle pieces," making the whole process faster, easier, and much more accurate.

Technical Summary

1. Problem Statement

The Tensor Renormalization Group (TRG) is a powerful method for calculating partition functions and physical observables in statistical physics and lattice field theories without the sign problem inherent in Monte Carlo simulations. However, the accuracy of TRG algorithms depends heavily on two factors:

1. Initial Tensor Construction: The method used to convert a physical model (defined by a Hamiltonian or Lagrangian) into a locally connected tensor network. Common methods rely on Singular Value Decompositions (SVD) or series expansions (e.g., Taylor expansions), which introduce specific symmetries or approximations.
2. Algorithm Sensitivity: Different TRG variants (e.g., HOTRG, ATRG, MDTRG) exhibit varying degrees of sensitivity to the symmetry properties of these initial tensors. Specifically, algorithms using isometries to generate new coarse-grained indices often fail or lose accuracy when the initial tensor lacks specific symmetries.

The authors address the lack of a universal, simple, and robust method for constructing initial tensors and investigate how to eliminate the dependence of TRG accuracy on the choice of these tensors.

2. Methodology

A. Novel Initial Tensor Construction

The authors propose a generic method to construct a locally connected tensor network without using SVDs, series expansions, or gauge-fixing.

• Mechanism: Instead of expanding interaction terms, they introduce an identity matrix (or shifted delta functions) to decompose the original non-local interaction tensor.
• Procedure:
  1. Start with the Boltzmann weight of a system with periodic boundary conditions.
  2. Insert a delta function (identity matrix) to split a multi-site interaction into a product of local tensors.
  3. Shift the indices of the delta function to neighboring lattice sites.
  4. Combine the original spin indices with the new auxiliary indices to form a larger, locally connected tensor.
• Generality: This method applies to systems with long-range interactions (e.g., next-nearest neighbor) and higher dimensions. The bond dimension of the resulting tensor scales as $d^{2(n_{int}-1+n_s)}$, where $d$ is the local Hilbert space dimension, $n_{int}$ is the number of sites in the original interaction, and $n_s$ is the number of "Steiner points" needed to connect disconnected regions.

B. Investigation of TRG Variants

The authors benchmarked several TRG algorithms on the 2D Ising model and 3D $Z_2$ gauge theory using two types of initial tensors:

1. $K(\text{exp})$: Constructed via Taylor expansion (symmetric).
2. $K(\text{delta})$: Constructed via the proposed delta-function method (often asymmetric).

The algorithms tested included:

• Isometric Methods (iso): HOTRG, Iso-ATRG, Iso-MDTRG (use isometries to create new indices).
• Squeezer Methods (sqz): Standard TRG, Boundary TRG (b-HOTRG), ATRG (without isometries), and variants using "squeezers" derived from boundary TRG techniques.

C. Boundary TRG and Squeezers

To address the sensitivity of isometric methods, the authors implemented Boundary TRG techniques. Instead of choosing a single optimal isometry (which favors one direction and introduces bias), they construct squeezers ($P_1, P_2$) that combine information from both directions (left and right, or up and down). This effectively symmetrizes the truncation process.

3. Key Contributions

1. Universal Initial Tensor Construction: A simple, model-independent method to generate locally connected tensor networks using delta functions. It avoids complex expansions and gauge-fixing, making it applicable to a wide range of systems (including $Z_2$ gauge theory) with minimal code changes.
2. Identification of Symmetry Dependence: The study reveals that isometric TRG methods (like standard HOTRG) are highly dependent on the symmetry of the initial tensor. Asymmetric initial tensors lead to significant accuracy degradation in these methods.
3. Elimination of Dependence via Squeezers: The authors demonstrate that replacing isometries with squeezers (as used in Boundary TRG) eliminates the dependence on initial tensor symmetry. This makes algorithms like b-HOTRG, ATRG, and MDTRG robust against the choice of initial tensor.
4. Index Exchange Optimization: The paper analyzes the impact of index swapping (rotation vs. flip) between coarse-graining steps. It finds that "shifted" methods require specific rotation exchanges to avoid error accumulation, while "non-shifted" methods perform better with flips.

4. Results

• 2D Ising Model:
  • HOTRG: Showed a massive drop in accuracy when using the asymmetric $K(\text{delta})$ tensor compared to the symmetric $K(\text{exp})$.
  • b-HOTRG (Boundary): Produced identical high-accuracy results for both symmetric and asymmetric tensors, proving that the squeezer technique removes the symmetry bias.
  • ATRG/MDTRG: Standard isometric variants suffered from the same symmetry dependence. However, their "shifted" or "boundary" variants (using squeezers) showed negligible dependence on the initial tensor form.
• 3D $Z_2$ Gauge Theory:
  • The authors applied their delta-function construction to the 3D $Z_2$ gauge theory without gauge-fixing.
  • Using the shifted ATRG method, they calculated the free energy and specific heat.
  • The results matched previous high-precision calculations (which used Taylor expansions and gauge-fixing) and Monte Carlo simulations, confirming the validity of the gauge-unfixed construction.
  • The critical temperature was determined as $\beta_c = 0.6560(3)$, consistent with literature values.

5. Significance

• Robustness and Simplicity: The proposed initial tensor construction simplifies the setup of TRG simulations. Researchers no longer need to derive complex, model-specific expansions or worry about gauge-fixing ambiguities.
• Algorithmic Improvement: The paper provides a clear prescription for improving existing TRG codes: replace isometries with squeezers (or use Boundary TRG techniques). This small code modification significantly enhances the reliability of results, particularly for systems where constructing a perfectly symmetric initial tensor is difficult or impossible.
• Broad Applicability: The method is applicable to systems with long-range interactions, multi-site interactions, and higher dimensions, making it a versatile tool for studying frustrated systems, gauge theories, and quantum field theories where sign problems prevent Monte Carlo simulations.
• Theoretical Insight: The work clarifies the source of systematic errors in TRG, attributing them to the interplay between the asymmetry of the initial tensor and the directional bias introduced by isometric truncations.

In conclusion, the paper establishes that while the choice of initial tensor matters significantly for standard isometric TRG algorithms, this issue can be completely resolved by adopting Boundary TRG techniques (squeezers). Combined with their simple delta-function construction method, this offers a powerful, generic, and robust framework for tensor network calculations.