Basis dependence of Neural Quantum States for the Transverse Field Ising Model

This paper investigates the basis dependence of Neural Quantum States for the Transverse Field Ising Model, revealing that their performance is governed by ground state properties like degeneracy and amplitude uniformity, which can be understood through the convergence of cluster or cumulant expansions of multi-spin operators.

The Gist

Imagine you are trying to teach a very smart, but slightly stubborn, robot to recognize a specific pattern in a room. This pattern is a complex quantum state (a specific arrangement of tiny magnets called "spins"). The robot uses a special type of brain called a Neural Quantum State (NQS) to learn this pattern.

The big question this paper asks is: Does it matter how you describe the room to the robot?

In the quantum world, you can describe the same physical reality from different angles (called "bases"). It's like looking at a sculpture: from the front, it looks like a face; from the side, it looks like a mountain. The object hasn't changed, but the picture you see has.

The authors of this paper discovered that yes, it matters a lot. If you describe the room from the "wrong" angle, the robot gets confused and fails to learn the pattern, even though the pattern itself hasn't changed.

Here is a breakdown of their findings using simple analogies:

1. The Robot's Dilemma: The "Simplest" Lie

The robot (a Restricted Boltzmann Machine, or RBM) is designed to find the "easiest" way to explain what it sees.

• The Problem: Sometimes, the true pattern the robot needs to learn is actually a mix of two different states that look almost identical (degenerate).
• The Analogy: Imagine you ask the robot to draw a picture of a "perfectly balanced seesaw." But in reality, the seesaw is slightly tipped to the left or right, and both positions are equally valid.
• The Result: Instead of learning the complex, tipped seesaw, the robot gets lazy. It draws a perfectly flat, horizontal line because that's the "simplest" superposition of the two options. It fails to capture the nuance of the real world because it's trying to find the path of least resistance.

2. The "Uniformity" Rule: Spreading the Butter

The second major factor is how "spread out" the pattern is.

• The Problem: Some patterns are like a thick layer of butter spread evenly over toast. Others are like a single, giant drop of butter stuck in one corner.
• The Analogy:
  • Uniform (Easy): If the quantum state is like the even butter, the robot can easily learn it. It's a smooth, predictable landscape.
  • Peaked (Hard): If the state is like a giant drop in one corner (a "peaked" wavefunction), the robot struggles. It's like trying to find a needle in a haystack where the needle is actually a mountain. The robot gets stuck trying to figure out why everything else is empty.
• The Finding: The robot performs best when the "butter" is spread out evenly. If you rotate your view of the room so the butter looks concentrated in one spot, the robot's performance crashes.

3. The "Sign" Confusion: The Secret Code

Quantum states have "signs" (positive or negative numbers) that act like a secret code.

• The Problem: Sometimes, the code is simple (all positive). Other times, it's a chaotic mix of pluses and minuses.
• The Analogy: Imagine the robot is trying to solve a maze.
  • Easy Mode: The maze has no traps; you just walk forward.
  • Hard Mode: The maze has invisible walls that flip your direction randomly.
• The Finding: If the "secret code" (the signs) is messy, the robot gets lost. However, the authors found that if the "butter" (amplitudes) is spread out evenly, the robot can handle a messy code better. But if the code is messy and the butter is clumped in a corner, the robot gives up.

4. The "Cumulant Expansion": The Recipe Book

To explain why the robot fails, the authors used a mathematical tool called a Cumulant Expansion.

• The Analogy: Think of the quantum state as a complex recipe.
  • Simple Recipe: "Mix flour and water." (This is easy to write down).
  • Complex Recipe: "Mix flour, water, salt, pepper, a pinch of cinnamon, a secret spice from the 14th century, and a whisper of wind." (This is hard to write down).
• The Discovery: The robot is like a chef who can only remember the first few ingredients in a recipe.
  • If the most important ingredients (the "big" terms in the recipe) are at the top of the list, the robot can cook a great meal.
  • If the important ingredients are buried deep in the list, or if the recipe requires every single ingredient to make sense, the robot fails.
• The Key Insight: The authors found that the robot's performance is directly linked to how quickly the "recipe" gets simpler. If you can describe the state using just a few main ingredients (a rapidly converging expansion), the robot wins. If you need thousands of ingredients, the robot loses.

The Takeaway: Choosing the Right Lens

The most important conclusion of the paper is a strategy for scientists: Don't just throw your problem at the robot and hope for the best.

Before you start training your AI, you should:

1. Check the "Recipe": Look at the quantum state and see if the "ingredients" (correlations) are simple or complex.
2. Rotate the View: Try describing the problem from different angles (bases).
3. Pick the Winner: Choose the angle where the "recipe" is shortest and the "butter" is most evenly spread.

In short: Neural networks are powerful, but they aren't magic. They are sensitive to how you present the data to them. By choosing the right "lens" to look at the quantum world, you can turn a impossible puzzle into an easy one.

Technical Summary

1. Problem Statement

Neural Quantum States (NQS), particularly those based on Restricted Boltzmann Machines (RBM), have proven effective for representing complex quantum many-body wavefunctions. However, a fundamental gap exists in understanding the limitations of these methods. Specifically, it is unclear why NQS performance varies drastically depending on the choice of the computational basis.

While physical properties (like entanglement entropy) are basis-independent, the wavefunction amplitudes and their sign structures are basis-dependent. The paper addresses the open question: What specific properties of a ground state wavefunction in a given basis determine whether an RBM can efficiently learn and represent it?

2. Methodology

The authors employ a controlled numerical approach to isolate learning and representability issues from sampling errors:

• Model System: The 1D Transverse-Field Ising Model (TFIM) with $L=10$ spins.
• Basis Rotation: Instead of rotating the state, they rotate the Hamiltonian $H(\theta)$ by an angle $\theta$. This generates a family of Hamiltonians with identical physical spectra but different ground state representations in the fixed computational basis (eigenbasis of $\sigma^z$).
  • $H(\theta) = -\sum \tilde{\sigma}^z_i \tilde{\sigma}^z_{i+1} - g \sum \tilde{\sigma}^x_i$, where Pauli matrices are rotated by $\theta$.
• Architecture: Restricted Boltzmann Machines (RBM) with complex-valued parameters.
• Training: Stochastic Reconfiguration (SR) is used for variational optimization.
• Exact Diagonalization: Due to the small system size ($L=10$), the authors perform exact summation over the entire Hilbert space ($2^{10}$ states) to compute expectation values and infidelities without Monte Carlo sampling noise.
• Analytical Tool: A truncated cumulant expansion (generalized Jastrow ansatz) is used to analyze the convergence properties of the exact ground state and compare it to the RBM's learned state.

3. Key Contributions & Findings

The paper identifies three critical factors governing RBM performance and establishes a theoretical framework linking them to convergence:

A. Ground State Uniqueness and Degeneracy

• Observation: If the true ground state is nearly degenerate (common in the symmetry-broken phase of TFIM), the RBM tends to converge to the simplest possible superposition of the degenerate states.
• Mechanism: The RBM favors states with uniform positive signs and uniform amplitudes in the computational basis.
• Result: If the true ground state has a complex sign structure or non-uniform amplitudes within the degenerate manifold, the RBM fails to learn it, instead converging to a "simpler" but physically incorrect superposition (e.g., $|\Psi_+\rangle$ instead of $|\Psi_-\rangle$).

B. Phase and Amplitude Uniformity

The authors decouple the effects of phase and amplitude uniformity:

• Phase Uniformity: It is generally easier to learn states with uniform phases (or simple sign structures like the Marshall sign rule). Complex sign structures increase optimization difficulty.
• Amplitude Uniformity (Crucial Finding): The authors highlight that amplitude uniformity is often overlooked but critical.
  • Peaked Distributions: When the probability distribution $|\Psi(s)|^2$ is highly peaked (non-uniform), RBM performance degrades significantly, even if the sign structure is simple.
  • Uniform Distributions: When $|\Psi(s)|^2$ is close to uniform (e.g., in the paramagnetic phase at $\theta=0$), the RBM learns the state efficiently.
  • Interplay: The optimization landscape becomes complex when both phases and amplitudes are non-uniform, leading to local minima and poor convergence.

C. Cumulant Expansion Convergence

The paper provides a rigorous link between NQS performance and the cumulant expansion of the wavefunction:
$$\Psi(s) = \exp\left( \sum c_A S_A(s) \right)$$

• Hypothesis: The performance of an RBM with $N_{var}$ parameters is equivalent to a truncated cumulant expansion that retains only the $N_{var}$ largest coefficients.
• Validation: By comparing the infidelity of the RBM against the truncated exact expansion, the authors show that:
  1. For small rotation angles $\theta$, the cumulant expansion converges rapidly (few large coefficients). The RBM captures these dominant terms effectively.
  2. For large $\theta$ (e.g., $\theta = \pi/2$), the expansion converges slowly or non-monotonically, requiring many high-order terms. The RBM fails to capture these, leading to high infidelity.
• Significance: This explains why basis choice matters: a "good" basis is one where the ground state's cumulant expansion converges rapidly with respect to the number of available variational parameters.

4. Results Summary

• Basis Dependence: The energy error and infidelity of the RBM vary by several orders of magnitude as the rotation angle $\theta$ changes, despite the physical Hamiltonian remaining equivalent.
• Degeneracy Trap: In the ferromagnetic phase ($g < 1$), the RBM struggles at angles where the ground state is degenerate and has complex signs/amplitudes, converging instead to a symmetric superposition.
• Amplitude Sensitivity: Even in stoquastic (sign-problem-free) regimes, the RBM fails if the amplitude distribution is too peaked (e.g., near the product state limit at $\theta = \pi/2$).
• Predictive Framework: The convergence rate of the cumulant expansion serves as a reliable predictor for RBM performance. If the expansion requires more terms than the RBM has parameters to represent the state accurately, the NQS will fail.

5. Significance and Implications

• Beyond Entanglement: The work challenges the prevailing view that NQS performance is solely determined by entanglement (which is basis-independent). It demonstrates that basis-dependent properties (sign structure, amplitude distribution) are equally, if not more, important for learnability.
• Practical Strategy: The authors propose a pre-computation strategy for applying NQS to new problems:
  1. Check for ground state degeneracies.
  2. Compute the cumulant expansion of the ground state (via exact diagonalization of small clusters).
  3. Identify the basis transformation that maximizes the convergence speed of this expansion.
• Generalizability: While tested on RBMs and TFIM, the authors argue these principles (degeneracy handling, amplitude/phase uniformity, and expansion convergence) likely apply to other neural network architectures and quantum models.

In conclusion, the paper provides a fundamental framework for understanding the "learnability" of quantum states by neural networks, shifting the focus from purely representational capacity to the interplay between the basis choice, the structure of the wavefunction, and the convergence of many-body expansions.