⚛️ quantum physics

Simulating dynamics of the two-dimensional transverse-field Ising model: a comparative study of large-scale classical numerics

This paper presents a comprehensive comparative study of state-of-the-art classical numerical methods, including tensor networks and neural quantum states, to simulate the dynamics of the two-dimensional transverse-field Ising model under quantum annealing and quench protocols, thereby establishing benchmarks for future classical and quantum computing capabilities.

Original authors: Joseph Vovrosh, Sergi Julià-Farré, Wladislaw Krinitsin, Michael Kaicher, Fergus Hayes, Emmanuel Gottlob, Augustine Kshetrimayum, Kemal Bidzhiev, Simon B. Jäger, Markus Schmitt, Joseph Tindall, Constan

Published 2026-06-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Joseph Vovrosh, Sergi Julià-Farré, Wladislaw Krinitsin, Michael Kaicher, Fergus Hayes, Emmanuel Gottlob, Augustine Kshetrimayum, Kemal Bidzhiev, Simon B. Jäger, Markus Schmitt, Joseph Tindall, Constantin Dalyac, Tiago Mendes-Santos, Alexandre Dauphin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A Race Between Computers

Imagine you are trying to predict how a massive crowd of people will move through a city square. Some people are holding hands (interacting), and the wind is blowing in different directions (magnetic fields).

This paper is about a "race" to see who can predict the crowd's movement best:

The Classical Computers: These are the super-smart, traditional calculators we have today. They use clever math tricks to guess the crowd's path.
The Quantum Computers: These are new, experimental machines that actually act like the crowd, simulating the physics directly.

The authors of this paper didn't build a new quantum computer. Instead, they acted as referees. They took the best "classical" math tools available and used them to simulate a specific type of crowd movement called the 2D Transverse-Field Ising Model. They wanted to see:

How far can these classical computers go before they get confused?
Where do they start to make mistakes?
This helps scientists know exactly when they need a quantum computer to solve a problem because the classical ones have hit a wall.

The Two Scenarios: The Slow Walk vs. The Sudden Push

The researchers tested the classical computers in two different "games" or scenarios.

1. The Slow Walk (Quantum Annealing)

Imagine a crowd slowly walking from a chaotic, disorganized state into a perfectly organized line.

The Game: The researchers slowly changed the rules (the "wind") to guide the crowd into an ordered formation.
The Result: Most of the classical math tools did a great job here. They could predict the crowd's path accurately, even when the crowd was moving slowly through a "critical point" (a moment of high tension where the crowd is deciding how to organize).
The Catch: One tool (called 2DTN) started to stumble when the crowd got too large or the loops in the crowd became too tight, like a GPS getting lost in a maze with too many turns.

2. The Sudden Push (Post-Quench Dynamics)

Imagine the crowd is standing still, and suddenly, a giant drumbeat hits, causing everyone to jump and spin wildly.

The Game: The rules change instantly, and the crowd goes into a chaotic, energetic frenzy.
The Result: This was much harder for the classical computers.
- Strong Interactions: If the crowd was tightly knit, the math tools worked well.
- The Critical Zone: When the crowd was in a "tipping point" state (neither fully organized nor fully chaotic), the classical computers started to disagree with each other. Some said the crowd would calm down; others said they would keep spinning.
- The Limit: As time went on, the "entanglement" (the complex web of connections between people) grew so huge that the classical computers ran out of memory or accuracy. They couldn't keep track of the chaos anymore.

The Tools in the Toolbox

The authors used a "toolbox" of different mathematical strategies to solve these problems. Think of them as different ways to map the crowd:

MPS (Matrix Product States): Imagine trying to map the crowd by looking at them one row at a time, like reading a book line by line. It works great for simple lines, but if the crowd is a big 2D square, you have to twist the line into a snake shape. This gets messy and inaccurate when the crowd gets too complex.
TTN (Tree Tensor Networks): Imagine mapping the crowd using a family tree structure. It's better than the snake, but if the crowd forms a tight circle (a loop), the tree structure breaks down because trees don't have loops.
2DTN (2D Tensor Networks): This tool tries to map the crowd exactly as a 2D grid, respecting the square shape. It's very good at short distances but uses a shortcut (called "Belief Propagation") to save time. When the crowd gets too complex, the shortcut fails, and the map becomes wrong.
NQS (Neural Quantum States): This uses an Artificial Intelligence (a neural network) to learn the crowd's behavior. It's very flexible but sometimes gets "confused" by the math equations it has to solve, leading to errors that aren't necessarily about how complex the crowd is, but about the AI's internal math glitches.

The "Symmetry Check" (The New Rule)

One of the paper's clever ideas was a new way to check if the computers were lying.

Since the city square is perfectly symmetrical (it looks the same if you rotate it 90 degrees), the crowd's behavior should also look the same from every angle.

The Trick: The researchers checked if the math tools respected this symmetry. If a tool said "People on the left are calm, but people on the right are panicking" (when the rules were identical), the tool had failed.
The Finding: They found that even when the math tools claimed to be "converged" (finished calculating), they were sometimes breaking this symmetry. This new "Symmetry Error" check helped them spot exactly when the tools stopped being reliable.

The Conclusion: Where Do We Stand?

The paper concludes with a clear map of the landscape:

For Slow, Organized changes: Classical computers are still kings. They can handle these simulations very well.
For Sudden, Chaotic changes (near critical points): Classical computers are hitting a wall. They start to disagree with each other and lose accuracy quickly as the system gets bigger.
The Quantum Opportunity: This is where the new quantum computers (like the Rydberg atom arrays mentioned) might finally beat the classical ones. The paper suggests that for these specific "sudden push" scenarios, quantum computers could provide answers that classical computers simply cannot calculate accurately anymore.

In short: The authors built a benchmark to show us exactly where the "classical" limit is. They found that while classical computers are great for slow, steady problems, they struggle significantly with fast, chaotic, and highly connected quantum systems, opening the door for quantum computers to take the lead.

Technical Summary: Simulating Dynamics of the 2D Transverse-Field Ising Model

Problem Statement

The simulation of many-body quantum dynamics is a central challenge in quantum physics, particularly as programmable quantum processing units (QPUs) based on Rydberg atom arrays advance. While QPUs offer a potential route to simulate regimes inaccessible to classical computers, the boundaries of classical simulability remain poorly understood, especially regarding diverse dynamical phenomena like quantum annealing and post-quench dynamics. Recent claims of quantum advantage have been challenged by classical numerical methods, necessitating a rigorous, comprehensive benchmark to delineate the capabilities and limitations of state-of-the-art classical solvers. This work addresses the gap in understanding by performing an extensive cross-benchmark of classical numerical approaches for the two-dimensional transverse-field Ising model (2D TFIM), a model relevant to Rydberg atom experiments.

Methodology

The authors employ a "toolbox" of four distinct state-of-the-art classical numerical methods to simulate the dynamics of the 2D TFIM on square lattices with open boundary conditions. All methods were executed on identical hardware (NVIDIA A100 GPUs) to ensure a fair comparison.

Tensor Network (TN) Approaches:
- Matrix Product States (MPS): Implemented via the two-site Time-Dependent Variational Principle (TDVP) on a "snaking" path across the 2D lattice.
- Tree Tensor Networks (TTN): Implemented via the one-site TDVP algorithm on a binary tree structure.
- 2D Tensor Networks (2DTN): Utilized a 2D projected-entangled-pair-state (PEPS) ansatz. Time evolution was performed using a Trotterized gate approach with the Belief Propagation (BP) approximation for environment contraction. Expectation values were computed using a more accurate boundary MPS contraction scheme.
Neural Quantum States (NQS):
- Implemented via Time-Dependent Variational Monte Carlo (tVMC) using a Convolutional Neural Network (CNN) ansatz. Simulations were performed in a rotated local basis to avoid pathologies associated with product states in the computational basis.
Protocols and Observables:
- Quantum Annealing: Two protocols were studied: (I) a quasi-adiabatic sweep crossing a critical point, and (II) a non-adiabatic sweep exhibiting Kibble-Zurek (KZ) scaling.
- Post-Quench Dynamics: The system was evolved from a fully polarized state under a time-independent Hamiltonian for three transverse field strengths ( $h_x/J = 0.5, 2.0, 3.0$ ), covering perturbative, dynamical phase transition, and quantum critical regimes.
- Observables: Local magnetization ( $\langle \sigma^z_i \rangle$ ) and connected two-point correlations ( $C(\delta_i)$ ).
Convergence and Error Analysis:
- Cross-Benchmarking: Discrepancies between methods were quantified using error metrics $\epsilon_z$ (magnetization) and $\epsilon_{zz}$ (correlations).
- Projection Errors: Truncation errors for TN methods and accumulated TDVP errors for TTN/NQS were monitored.
- Symmetry Errors: A novel convergence criterion was proposed based on the $D_4$ lattice symmetries. Since MPS and TTN break spatial symmetries due to their 1D mapping or tree structure, deviations from symmetry-equivalent observables serve as a diagnostic for convergence failure.

Key Results

1. Quantum Annealing

Quasi-Adiabatic Regime (Annealing I): MPS, TTN, and NQS showed qualitative and quantitative agreement, successfully capturing the buildup of antiferromagnetic (AFM) order. 2DTN deviated at later times due to the breakdown of the BP approximation in the presence of large correlation lengths and small loops.
Non-Adiabatic Regime (Annealing II / KZ): MPS and 2DTN showed strong agreement, capturing the suppression of long-range order due to non-adiabatic effects. TTN and NQS struggled to converge with these results, suggesting limitations in handling the specific diabatic effects in this regime.

2. Post-Quench Dynamics

Perturbative Regime ( $h_x/J = 0.5$ ): All methods agreed well, consistent with low entanglement and perturbative dynamics.
Critical and Intermediate Regimes ( $h_x/J = 2.0, 3.0$ ):
- Short Times ( $tJ \lesssim 1.5$ ): MPS, TTN, and 2DTN showed reasonable agreement.
- Long Times ( $tJ \gtrsim 2.0$ ): All methods began to diverge significantly.
- MPS/TTN: Accuracy degraded rapidly with system size due to the exponential growth of entanglement, which exceeds the representational capacity of the fixed bond dimensions.
- NQS: Failed to capture long-time dynamics even at small system sizes, with convergence issues linked to numerical instabilities in the tVMC equations of motion rather than expressivity.
- 2DTN: Showed the most favorable scaling with system size for short-time dynamics but suffered from abrupt failure when the BP approximation broke down (e.g., in quasi-adiabatic regimes or at long times).

3. Convergence Diagnostics

Projection Errors: For TN methods, errors remained low until the bond dimension limit was reached, after which they increased exponentially. For NQS, accumulated errors plateaued or increased when the ansatz complexity was insufficient to capture the dynamics.
Symmetry Errors: The proposed symmetry-based metric ( $\epsilon_{rel}$ ) effectively identified the onset of unphysical anisotropies in MPS and TTN results. It revealed that while local magnetization might remain converged, correlation functions often lose accuracy earlier, particularly in TTN due to its tree structure.

Significance and Claims

The paper claims to provide a comprehensive benchmark that clarifies the boundary between classical and quantum computational capabilities for 2D many-body dynamics.

Complementarity: The study demonstrates that classical methods are complementary. MPS and TTN are effective for low-entanglement or specific annealing protocols but fail for high-entanglement post-quench dynamics. 2DTN offers better scaling for 2D geometries but is sensitive to the validity of the BP approximation. NQS is promising for quasi-adiabatic processes but currently faces numerical stability challenges in highly non-equilibrium regimes.
Limitations of Classical Simulability: The results indicate that while classical methods can simulate short-time post-quench dynamics and quasi-adiabatic annealing, they struggle significantly with long-time dynamics near quantum critical points. This suggests that Rydberg atom QPUs operating in these specific regimes (e.g., $h_x/J \approx 2.0 - 3.0$ for times $tJ > 2$) may offer a regime where classical simulation is unreliable, potentially enabling the observation of quantum advantage.
Methodological Contribution: The introduction of symmetry-based error metrics provides a robust, physically motivated convergence criterion that is more sensitive to the failure of local observables than global truncation errors alone.
Experimental Context: The work connects these numerical limits to experimental capabilities of Rydberg atom arrays, noting that current devices can access the challenging time scales ( $tJ \gtrsim 1$ ) where classical methods begin to diverge, provided noise levels are sufficiently controlled.

The authors conclude that while classical algorithms like MPS are mature, emerging methods (NQS, 2DTN) are evolving rapidly. The observed limitations in NQS are attributed to algorithmic instabilities rather than fundamental entanglement barriers, suggesting potential for future improvement. The study emphasizes that claiming quantum advantage requires careful characterization of both classical numerical limits and experimental noise.