⚛️ quantum physics

Connectivity determines the capability of sparse neural network quantum states

This paper extends the Lottery Ticket Hypothesis to quantum many-body systems, demonstrating that sparse neural network quantum states can achieve accuracy comparable to dense models through structure-dependent performance and universal scaling laws that reveal sparsity-induced quantum phase transitions and enhanced physical interpretability.

Original authors: Brandon Barton, Juan Carrasquilla, Christopher Roth, Agnes Valenti

Published 2026-01-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Brandon Barton, Juan Carrasquilla, Christopher Roth, Agnes Valenti

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

Imagine you have a massive, incredibly complex library (a neural network) designed to solve a specific, difficult puzzle: finding the "ground state" of a quantum system. In physics, the ground state is the lowest energy, most stable configuration of a system—like finding the perfect, calm arrangement of a shaken box of marbles.

Usually, to solve this puzzle, you need the whole library, with millions of books (parameters). But this paper asks a fascinating question: Do we actually need all those books, or is there a tiny, hidden "winning ticket" inside the library that can solve the puzzle just as well if we only use that specific section?

This idea is based on the "Lottery Ticket Hypothesis," which suggests that within huge, overgrown neural networks, there are small, sparse sub-networks that are naturally good at the job.

Here is what the researchers found, explained simply:

1. The Structure Matters More Than the "Seed"

In the original "Lottery Ticket" idea, scientists thought you needed two things to win:

The right structure (the specific arrangement of books on the shelves).
The right initial values (the specific words written in the books when you first open them).

The Paper's Twist: When applying this to quantum physics, they found that only the structure matters.

Analogy: Imagine you have a specific blueprint for a house (the structure). In normal machine learning, you might think you also need the exact same type of bricks and mortar (the initial weights) to build it.
The Finding: In this quantum context, if you take the blueprint of the winning sub-network but build it with random bricks and mortar, it still works perfectly! The "winning ticket" is just the shape of the connections, not the specific numbers inside them. This is a huge surprise because it means the network's "skeleton" is what holds the physics, not the specific starting values.

2. The "Goldilocks" Zones (Scaling Laws)

As the researchers pruned (cut away) more and more of the network, they didn't just see a slow decline in performance. Instead, they found three distinct "zones," like different climate zones on a map:

The Lush Forest (Low Error): When the network is still quite big, it solves the puzzle perfectly. Cutting away a few branches doesn't hurt the tree.
The Power-Law Slope: As you cut more, the performance starts to drop in a predictable, mathematical way.
The Desert (High Error): Once you cut past a certain point, the network collapses. It can no longer solve the puzzle, no matter how you train it.

The Discovery: The boundaries between these zones depend on the physics of the problem, not just the size of the network. For example, a "ferromagnetic" state (where all spins align) is easier to represent with a tiny network than a "critical" state (where things are chaotic).

3. The "Sparsity-Induced" Phase Transition

This is the most magical part. The researchers found that as they cut the network down, the network didn't just get "worse"; it actually changed its nature.

Analogy: Imagine you are slowly removing the walls of a house. At a certain point, the house doesn't just become a smaller house; it suddenly transforms into a completely different type of structure, like a tent.
The Finding: At a specific point of "sparsity" (how many connections are left), the network undergoes a Phase Transition. It jumps from representing the correct, complex quantum physics to representing a much simpler, broken version of it.
They proved this by measuring "fidelity" (how similar the network's answer is to the previous step). At the transition point, the similarity dropped to near zero instantly, like a light switch flipping. This is a "first-order" phase transition, a concept usually reserved for things like water turning to ice, but here it happened just by cutting wires in a computer program.

4. Cracking the "Toric Code" Puzzle

The researchers tested this on a famous, difficult puzzle called the "Toric Code" (which involves topological order, a very complex type of quantum arrangement).

The Result: By pruning the network, they accidentally discovered a perfect, simplified solution that humans hadn't explicitly designed.
The Analogy: It's like trying to solve a maze by randomly removing walls. Eventually, you remove so many walls that the maze disappears, and you are left with a straight, open path that leads directly to the exit.
They found that the remaining sparse network used a specific pattern of connections (filters) that acted like a mathematical "amplifier," boosting the correct answers and crushing the wrong ones. This solution was so efficient it was "asymptotically exact," meaning it could be made perfect just by adjusting the strength of the remaining connections.

Summary

In short, this paper shows that:

Connectivity is King: For quantum physics problems, the shape of the neural network is the most important part, not the specific numbers inside it.
Physics Dictates Limits: The point at which a network fails isn't random; it's determined by the physical laws of the system it's trying to simulate.
Sudden Shifts: Cutting a network can cause it to suddenly snap from one "state of matter" to another, revealing deep connections between computer pruning and quantum physics.
Discovery Tool: Pruning isn't just about making models smaller; it's a way to discover new, efficient ways to solve complex physics problems that humans might not have thought of.

Technical Summary: Connectivity Determines the Capability of Sparse Neural Network Quantum States

Problem Statement
This work investigates the applicability of the Lottery Ticket Hypothesis (LTH) to the unsupervised task of approximating the ground states of quantum many-body Hamiltonians using neural networks. Unlike typical machine learning tasks (e.g., image recognition) where data is sparse, quantum many-body wavefunctions are generally fully dense representations with amplitudes spanning many orders of magnitude. The authors address the challenge of finding sparse subnetworks within overparameterized neural networks that can match the performance of the original dense model when trained in isolation. Specifically, they examine whether the "winning ticket" in this context relies on specific initial weight values (as in the original LTH) or solely on the connectivity structure of the sparse subnetwork.

Methodology
The study employs Iterative Magnitude Pruning with Weight Rewinding (IMP-WR) to generate sparse masks. The process involves:

Pre-training: Training a randomly initialized dense neural network.
Pruning: Iteratively removing the smallest magnitude weights and rewinding the remaining weights to their values at the start of the pruning iteration.
Isolated Training: Testing three distinct initialization strategies for the resulting sparse subnetworks ("tickets"):
- $T(\theta_{init}, m_{imp})$ : Original LTH test (original weights + learned mask).
- $T(\theta_{rand}, m_{imp})$ : Random weights + learned mask.
- $T(\theta_{init}, m_{rand})$ : Original weights + random mask.

The experiments utilize two physical models:

Transverse Field Ising Model (TFIM): A 2D system exhibiting a quantum phase transition between ferromagnetic and paramagnetic phases.
Toric Code (TC): A topologically ordered model with an exact ground state solution.

The neural network architectures tested include shallow Feed-Forward Neural Networks (FFNNs), shallow Convolutional Neural Networks (CNNs), and Residual CNNs (ResCNNs). Training is performed using Natural Gradient Descent (Stochastic Reconfiguration) to minimize the variational energy cost function.

Key Contributions

Reformulation of the LTH for Quantum States: The authors demonstrate that for neural-network quantum states, the "winning ticket" is defined by the sparse connectivity structure alone. The specific initialization of weights is negligible; sparse subnetworks trained from random initializations ( $T(\theta_{rand}, m_{imp})$ ) perform comparably to those trained from the original initialization ( $T(\theta_{init}, m_{imp})$ ). This contrasts with the original LTH, where specific initial parameter values are crucial.
Universal Scaling Behavior: The paper identifies universal scaling regions in the error vs. sparsity curve that persist across different network sizes and physical models. These regions include:
- A Low-Error Plateau (LEP) at high parameter densities.
- An Approximate Power-Law (APL) region.
- A High-Error Plateau (HEP) at low parameter densities.
  The boundaries of these regions are determined by the underlying physics of the Hamiltonian.
Sparsity-Induced Phase Transitions: By treating the number of parameters per spin ( $\rho$ ) as a control parameter analogous to a physical field, the authors identify a first-order quantum phase transition in shallow networks. This transition occurs at the boundary between the APL and HEP, characterized by a discontinuity in the derivative of the energy and a sharp drop in fidelity between adjacent pruned states.
Interpretability via Pruning: The pruning process reveals that the structure of sparse subnetworks directly encodes the physics of the Hamiltonian. For the Toric Code, pruning uncovers an asymptotically exact solution composed of "odd parity filters" that enforce the plaquette conditions of the ground state.

Results

TFIM Performance: Sparse networks (with 5–20% of the original parameters) achieve accuracies comparable to dense models. The transition to the high-error plateau occurs at different sparsity levels depending on the phase (ferromagnetic vs. paramagnetic), reflecting the intrinsic sparsity of the true ground state eigenvector in each phase.
Phase Transition Signatures: At the critical point of the TFIM, the transition from the APL to the HEP is marked by a sudden change in total magnetization ( $M_z$ and $M_x$ ) and a sharp drop in fidelity, confirming a first-order phase transition induced by sparsity.
Toric Code Solution: Pruning a single-layer FFNN on the Toric Code yields a solution where the error decreases exponentially with weight magnitude. The resulting sparse structure consists of filters that effectively implement the exact ground state constraints (superposition of configurations satisfying $\lambda_{Bp}(\sigma) = 1$ ).
Architecture Independence: The phenomenon of connectivity-determined performance holds across FFNNs, CNNs, and ResCNNs, though ResCNNs show high performance even with random masks due to their localized kernel structure providing a strong initial prior.

Significance
The paper claims that pruning serves as a powerful tool not only for model compression but also for interpreting the physics of quantum many-body systems. By linking the structure of sparse subnetworks to the underlying Hamiltonian, the authors show that:

Neural networks can discover efficient, asymptotically exact solutions to complex quantum problems (like the Toric Code) without prior architectural design.
The dynamics of pruning reveal fundamental properties of quantum phases, such as phase transitions driven by representational capacity (sparsity).
The independence of performance from weight initialization suggests a strong, intrinsic connection between the topology of the network and the physics it represents, potentially simplifying the training of quantum neural states by removing the need for precise initialization.

The authors conclude that these findings may facilitate the use of compressed models for applications such as time evolution, quantum circuit simulation, and fine-tuning, where parameter efficiency is critical.