Information in Many-body Eigenstates: A Question of Learnability

This paper introduces "learnability" as a machine learning-based metric to quantify how much information individual many-body eigenstates encode about their underlying Hamiltonian, demonstrating that spectral-edge eigenstates are significantly more learnable and require fewer samples for accurate Hamiltonian reconstruction than mid-spectrum eigenstates.

The Gist

The Big Question: Can a Single Page Tell You the Whole Story?

Imagine you have a massive, incredibly complex machine (a quantum system). This machine is governed by a hidden instruction manual called the Hamiltonian. This manual contains all the rules, settings, and dials that make the machine work.

Usually, to figure out what the manual says, you need to see the machine running in many different ways. But this paper asks a different question: If you only look at one specific "snapshot" of the machine's behavior (an eigenstate), can you reverse-engineer the manual?

The authors use a tool called Machine Learning (specifically a type of AI called an "autoencoder") to act as a detective. They feed the AI a snapshot of the machine and ask: "Based on this picture, what were the original settings?"

The Two Types of Snapshots

The paper discovers that the answer depends entirely on which snapshot you pick. The machine has a spectrum of possible behaviors, ranging from the "calmest" states to the "chaotic" states.

1. The "Low-Energy" Snapshots (The Calm, Ordered States)

• The Analogy: Imagine looking at a library where the books are perfectly organized by author, title, and color. The shelves are neat, and the pattern is obvious.
• The Reality: These are the states near the bottom of the energy spectrum. They are highly structured and follow clear rules (locality).
• The Result: The AI detective is excellent at these. Even with just one of these snapshots, the AI can accurately guess the settings of the manual. It's easy to learn because the "clues" are very clear and organized.

2. The "Mid-Spectrum" Snapshots (The Chaotic, Random States)

• The Analogy: Now imagine looking at a library where someone has thrown all the books into a giant pile, mixed them up, and shaken them. It looks like random noise. There is no obvious pattern to the arrangement.
• The Reality: These are the states in the middle of the energy spectrum. They are "entangled" and look almost like random static. They follow the rules of chaos (Random Matrix Theory).
• The Result: The AI detective fails here. Even if you give it many of these chaotic snapshots, it struggles to guess the manual's settings. The information about the original rules has been "scrambled" so thoroughly that it's nearly impossible to recover.

The Experiment: How They Tested It

The researchers set up a simulation of a chain of tiny magnets (spins). They created thousands of different versions of this chain by tweaking two dials (parameters $J_1$ and $J_2$).

1. The Encoder: They took a snapshot of the magnets (an eigenstate) and fed it into the AI.
2. The Guess: The AI tried to guess what the dial settings were.
3. The Check: They compared the AI's guess to the actual settings.

They tested this in two ways:

• Single State: They gave the AI just one snapshot from different parts of the spectrum.
• Multiple States: They gave the AI a small group of snapshots.

The Key Findings

• Location Matters: The ability to "learn" the machine's rules drops off sharply as you move from the calm, low-energy states to the chaotic, middle-energy states.
• It's Not a Computer Problem: The researchers tried making the AI "smarter" (giving it more brainpower/neurons). While this helped slightly with the easy (low-energy) cases, it did not help with the hard (mid-spectrum) cases. This proves that the problem isn't that the AI is too dumb; the problem is that the information simply isn't there to be found in the chaotic snapshots.
• The "Learnability" Metric: The authors propose a new way to measure information called Learnability. Instead of just asking "Is this state complex?", they ask "Can a machine learn the rules from this state?" If the answer is "No," the state has low learnability.

The Takeaway

This paper suggests that in the quantum world, information is not stored equally everywhere.

• In the calm, low-energy states, the "fingerprint" of the machine's rules is clear and easy to read.
• In the chaotic, high-energy states, the fingerprint is washed out by randomness.

The authors conclude that Machine Learning isn't just a tool for solving problems; it's a new way to measure physics. By seeing how well an AI can guess the rules, we can understand how much information is actually preserved in different parts of a quantum system.

In short: If you want to know how a quantum machine works, look at its quiet, orderly moments. If you look at its chaotic, frantic moments, the clues are likely gone.

Technical Summary

Technical Summary: Information in Many-body Eigenstates: A Question of Learnability

Problem Statement
The paper addresses a fundamental question in quantum many-body physics: to what extent do individual eigenstates encode information about their underlying Hamiltonian, and how does this information content depend on the eigenstate's spectral position? While it is established that low-energy eigenstates exhibit strong structure (e.g., locality, constrained entanglement) and mid-spectrum eigenstates are often pseudo-random (consistent with Random Matrix Theory and the Eigenstate Thermalization Hypothesis), a quantitative measure of the informational difference between these regimes has been lacking. The authors propose that the ability to reconstruct Hamiltonian parameters from eigenstates serves as a novel probe of this information content.

Methodology
The authors introduce a machine learning (ML) framework to quantify "learnability," defined as the precision with which Hamiltonian parameters can be reconstructed from a single or a subset of eigenstates using a fixed neural network architecture.

1. Model System: The study focuses on a family of local one-dimensional spin-1/2 chains ($L=6$) described by a Hamiltonian $\hat{H}_{J_1, J_2}$ with nearest-neighbor ($J_1$) and next-nearest-neighbor ($J_2$) couplings, subject to periodic boundary conditions and local magnetic fields. The parameters are reduced to a latent vector $\theta = (J_1, J_2)$, with $J_2$ fixed in the main analysis and $J_1$ varied.
2. Network Architecture: An unsupervised autoencoder is employed:
   • Encoder: A point-wise multi-layer perceptron (MLP) that takes a matrix of $M$ eigenstates ($\Psi_\theta \in \mathbb{R}^{D \times M}$) as input and maps them to a latent representation $\tilde{\theta}$ (the inferred parameters).
   • Decoder: A parameter-free module that constructs the Hamiltonian $\hat{H}_{\tilde{\theta}}$ from the inferred parameters $\tilde{\theta}$ using a fixed basis of operators.
3. Loss Function: To avoid the computational prohibitions of repeated diagonalization and the need for labeled parameters during training, the authors utilize a physics-inspired Rayleigh loss ($L_{Rayleigh}$). This loss measures the deviation of the reconstructed Hamiltonian $\hat{H}_{\tilde{\theta}}$ from diagonality in the basis of the input eigenstates $\Psi_\theta$. It penalizes off-diagonal elements and spectral mismatches without requiring access to the true parameters $\theta$ during optimization.
4. Protocols: The study evaluates learnability under three selection protocols:
   • Single-state sweep: Using one eigenstate ($M=1$) swept across the entire spectrum.
   • Low-energy protocol: Using the first $M$ eigenstates.
   • Mid-spectrum protocol: Using $M$ consecutive eigenstates centered around the mean energy.

Key Results
The study demonstrates a sharp distinction in learnability based on spectral position:

• Spectral Dependence: Eigenstates near the spectral edges (low-energy) allow for highly accurate reconstruction of the Hamiltonian parameters. In contrast, eigenstates in the middle of the spectrum yield poor reconstruction accuracy, even with large training sets.
• Single-State Reconstruction: When using a single eigenstate ($M=1$), the reconstruction error remains low for low-energy states but saturates quickly for mid-spectrum states, regardless of the hidden layer width ($w_H$) or the number of training samples ($N_{sam}$).
• Effect of Data Quantity: Increasing the number of input eigenstates ($M$) significantly improves reconstruction for low-energy states. However, for mid-spectrum states, reconstruction only becomes accurate when $M$ approaches half the spectrum ($M \to D/2$), indicating that local information is suppressed in the highly entangled bulk.
• Network Capacity: Increasing the model capacity (hidden layer width) improves convergence for edge states but fails to remedy the lack of information in mid-spectrum states. This confirms that the failure to learn from bulk states is due to intrinsic properties of the eigenstates (their pseudo-random nature) rather than limitations of the neural network's expressivity.
• Generalization: While the network can accurately reconstruct parameters within the training domain, it struggles to generalize to excluded parameter intervals, suggesting that learnability and generalization are distinct challenges.

Significance and Claims
The paper claims to establish learnability as a new, alternative information-theoretic diagnostic for quantum many-body systems. Its primary contributions are:

1. Quantitative Probe: It provides a systematic, model-agnostic method to measure how information is distributed across the energy spectrum, linking spectral position directly to the feasibility of Hamiltonian reconstruction.
2. Validation of Physical Intuition: The results quantitatively confirm the theoretical expectation that low-energy states preserve structural information (local couplings) while mid-spectrum states effectively erase this information, becoming indistinguishable from random vectors.
3. Shift in ML Role: The work reframes machine learning from a purely computational tool for solving inverse problems to a conceptual probe of physical structure. It demonstrates that the "learnability" of a system is dictated by the underlying physics of the states themselves, specifically their entanglement and correlation properties.

The authors conclude that the feasibility of extracting Hamiltonian structure is not uniform across the spectrum but is fundamentally limited by the transition from structured, low-entanglement states to thermal, pseudo-random states in the bulk.