⚛️ quantum physics

Exploring the performance of superposition of product states: from 1D to 3D quantum spin systems

This paper investigates the superposition-of-product-states (SPS) ansatz as a geometry-independent, parallelizable variational framework for quantum spin systems, demonstrating its ability to achieve high accuracy in ground state searches across 1D and 3D tilted Ising models despite lower information compression compared to traditional tensor networks.

Original authors: Apimuk Sornsaeng, Itai Arad, Dario Poletti

Published 2026-04-08

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Apimuk Sornsaeng, Itai Arad, Dario Poletti

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 Problem: The "Unfathomable Library"

Imagine you are trying to find the perfect arrangement of furniture in a house. If the house has 3 rooms, it's easy. But if the house has 100 rooms, and every piece of furniture can be in any of 100 spots, the number of possible arrangements is so huge that it exceeds the number of atoms in the universe.

This is the problem physicists face when studying quantum systems (like atoms or electrons). As you add more particles, the number of possible states they can be in explodes exponentially. It's like trying to read every book in a library that grows faster than you can build shelves.

The Old Tools: "Tensor Networks"

For a long time, the best tool for this job has been called Tensor Networks.

The Analogy: Think of a Tensor Network like a folding map. It's brilliant at folding up a complex 2D map (a 1D line of atoms) into a small, manageable size.
The Flaw: However, if you try to use this folding map on a 3D object (like a cube of atoms) or a messy, random shape, the map gets wrinkled, torn, or impossible to fold. It struggles to capture the complexity of 3D or disordered systems without losing important details.

The New Tool: "SPS" (Superposition of Product States)

This paper introduces a new method called SPS.

The Analogy: Imagine instead of folding a map, you are building a collage.
- You take a bunch of simple, flat pictures (Product States).
- You stack them on top of each other, layering them with different weights (Superposition).
- The final image is the sum of all these layers.

The authors argue that while this "collage" method isn't as good at compressing data as the folding map (Tensor Networks), it has four superpowers:

It sees everything clearly: You can read the details of the collage without guessing.
It doesn't care about the shape: Whether the system is a line, a cube, or a messy random web, the collage works the same way.
It's a team sport: You can build different parts of the collage on different computers at the same time (parallelizable).
It has shortcuts: You can calculate things using math formulas instead of brute-force guessing.

The "Barren Plateau" Trap

In the world of AI and quantum computing, there's a famous problem called the Barren Plateau.

The Analogy: Imagine you are trying to find the lowest point in a valley (the ground state of energy) to park your car. In many complex methods, the landscape is so flat that you can't tell which way is "down." You wander aimlessly, and the computer gets stuck.
The SPS Solution: The authors tested their "collage" method and found that the landscape is not flat. There are clear slopes. Even in huge systems, the computer can always feel which way is "down" and slide toward the solution. This means the method is trainable and won't get stuck.

The Results: How Well Does It Work?

The team tested this new method on various "houses" (quantum systems):

Simple Lines (1D): It works well, though the old "folding map" (Tensor Networks) is still slightly faster here.
Cubes (3D): This is where the old method struggles. The "collage" (SPS) shines, finding the correct answers quickly and accurately.
Messy/Random Systems: Imagine a house where the rooms are connected randomly, like a spiderweb. The old methods get confused. The SPS method handles this chaos beautifully, finding the ground state even when the connections are random.

The Verdict

The paper concludes that the SPS method is a robust, flexible, and powerful new tool.

When to use it: If you are dealing with 3D systems, long-range interactions, or messy, disordered materials, the SPS "collage" is a fantastic choice.
The Trade-off: It requires a bit more memory than the old method for simple 1D lines, but it pays off by being able to tackle the complex, 3D, and chaotic problems that were previously too difficult to solve accurately.

In short: They built a new kind of "quantum telescope" that might not be the sharpest for looking at a single straight line, but it is the best tool we have for looking at the messy, 3D, and complex universe of quantum matter.

1. Problem Statement

The study of quantum many-body systems is hindered by the exponential growth of the Hilbert space with the number of particles. While Tensor Networks (TNs) are highly effective for one-dimensional (1D) systems, their application to generic geometries (e.g., 2D and 3D) faces two primary limitations:

Expressive Power: Standard TNs (like MPS) struggle to capture high entanglement in higher dimensions without exponentially increasing computational cost.
Information Extraction: Accurately extracting physical observables from TNs in higher dimensions often requires approximate contraction methods (e.g., CTMRG, Belief Propagation) or sampling, which introduce errors.

Furthermore, variational quantum algorithms often suffer from the barren plateau problem, where gradient variances decay exponentially with system size, rendering optimization infeasible. The authors aim to investigate an alternative variational framework that avoids these pitfalls: the Superposition of Product States (SPS) ansatz.

2. Methodology: The SPS Ansatz

The authors propose and analyze the Superposition of Product States (SPS) ansatz.

Mathematical Formulation: The SPS is defined as a linear combination of $M$ normalized product states:
$|\Psi\rangle = \frac{1}{\sqrt{Z}} \sum_{m=1}^M c_m \bigotimes_{l=1}^L |\theta_m^l\rangle$
where $|\theta_m^l\rangle$ are single-site spin-1/2 states parameterized by angles $\theta$ , $c_m$ are complex (or real in this work) coefficients, and $Z$ is the normalization factor.
Tensor Decomposition Perspective: The ansatz corresponds to a Rank- $M$ Canonical Polyadic (CP) decomposition (also known as CANDECOMP/PARAFAC) of the quantum state tensor.
Key Advantages:
- Geometry Independence: The structure does not depend on the lattice connectivity, making it naturally suitable for 1D, 3D, and random graphs.
- Analytical Tractability: Unlike many TNs requiring numerical contraction, expectation values and gradients for SPS can be computed analytically using trigonometric identities, avoiding sampling errors.
- Parallelizability: The computations are highly parallelizable, suitable for GPU acceleration.

3. Key Contributions and Analysis

A. Statistical Properties and Typicality

The authors analyzed the statistical behavior of randomly initialized SPS states:

Normalization: The norm $Z$ follows a Gaussian distribution for large $M$ , with fluctuations vanishing as $M$ increases.
Local Observables: The variance of local observables (e.g., $\langle \sigma_x \rangle$ ) scales as $1/M$ . Unlike Haar-random states (which show exponential suppression of fluctuations with system size $L$ ), SPS exhibits "restricted typicality" where fluctuations are suppressed polynomially in $M$ but remain independent of $L$ . This suggests stability in the thermodynamic limit.
Entanglement Entropy: The typical 2-Rényi entropy of random SPS states is bounded by $\ln M$ . While this is lower than the maximum possible entropy for a given bond dimension in Matrix Product States (MPS), it indicates that SPS can capture significant correlations, though it is inherently less expressive than high-bond-dimension TNs for highly entangled states.

B. Trainability and Barren Plateaus

A critical contribution is the demonstration that SPS avoids the barren plateau problem:

Gradient Analysis: The authors computed the variance of gradients for local observables with respect to variational parameters ( $c_m$ and $\theta_m^l$ ).
Scaling: Gradient variances scale as $1/M^2$ and saturate (do not vanish) as the system size $L$ increases.
Implication: This polynomial scaling ensures that optimization remains feasible even for large systems, as local gradients do not vanish exponentially.

C. Ground State Search Performance

The authors benchmarked the SPS ansatz against the Density Matrix Renormalization Group (DMRG) algorithm (using a "snake" configuration for 3D) on the Tilted Ising Model with various geometries:

1D and 3D Nearest-Neighbor: SPS achieved high accuracy in the ferromagnetic phase. In the paramagnetic phase and near critical points, accuracy improved with larger $M$ , though DMRG remained superior for 1D due to its area-law efficiency.
Long-Range Interactions: Tested on power-law decaying interactions ( $J \propto d^{-\alpha}$ ). SPS performed well, particularly for small $\alpha$ (long-range), where the ground state is less entangled.
Random Coupling: Tested on Erdős-Rényi random graphs. SPS maintained robust performance, achieving relative errors as low as $10^{-5}$ with moderate $M$ , demonstrating flexibility in disordered systems lacking translational symmetry.

4. Results and Performance Metrics

Accuracy: In the ferromagnetic regime, SPS with small $M$ (e.g., $M=1$ ) achieved relative errors down to $10^{-8}$ . In the paramagnetic regime, errors were slightly higher but remained small ( $\sim 10^{-3}$ to $10^{-5}$ ) with sufficient $M$ .
Computational Efficiency:
- Time Complexity: SPS scales as $O(LM^2)$ for energy evaluation and $O(LM^2T)$ for optimization (where $T$ is steps).
- Hardware: The authors utilized NVIDIA A100 GPUs for SPS and AMD EPYC CPUs for DMRG.
- Speedup: For 3D and random-coupling systems, SPS reached target relative errors faster than DMRG, especially when leveraging GPU parallelization. For 1D systems, DMRG remained faster, highlighting the specific niche of SPS for higher dimensions and complex topologies.
Table I Comparison: The paper provides a direct comparison showing that for 3D nearest-neighbor and random-coupling systems, SPS (with $M \approx 100-500$ ) can match DMRG accuracy with competitive or superior wall-clock times on GPUs.

5. Significance and Conclusion

The paper establishes the SPS ansatz as a powerful, geometry-agnostic tool for quantum many-body simulations.

Overcoming TN Limitations: It offers a viable alternative to Tensor Networks for 3D and disordered systems where TN contraction is expensive or approximate.
Optimization Stability: It provides a concrete example of a variational ansatz that is free from barren plateaus, making it suitable for large-scale variational quantum algorithms.
Practical Utility: The ability to compute gradients analytically and the high degree of parallelizability make SPS particularly attractive for modern GPU-accelerated computing environments.
Future Directions: The authors suggest that incorporating spatial modulation into the parameters could further enhance expressivity in the paramagnetic phase and that extending the framework to complex-valued states and time-evolution is a promising avenue for future research.

In summary, the SPS ansatz trades some expressive power (compared to high-bond-dimension TNs) for computational tractability, analytical exactness, and robust trainability, making it a strong candidate for exploring complex quantum phases in higher dimensions and disordered systems.