Distance learning from projective measurements as an information-geometric probe of many-body physics

This paper introduces an unsupervised "distance learning" framework that uses a neural discriminator to estimate statistical distances between quantum state snapshots, enabling the identification of correlation regimes, reconstruction of phase diagrams, and extraction of critical exponents across diverse many-body systems without relying on traditional representation learning.

The Gist

Imagine you are a detective trying to solve a mystery inside a giant, chaotic room filled with thousands of people (quantum particles). You can't talk to them, and you can't see their thoughts. All you have is a camera that takes a million "snapshots" of the room, freezing everyone in a single pose.

In the past, scientists tried to understand these snapshots by trying to compress the data into a simple summary (like a map or a chart). But this paper introduces a smarter, more direct way to solve the mystery called "Distance Learning."

Here is the breakdown of how it works, using simple analogies:

1. The Problem: Too Many Snapshots

Modern quantum computers (simulators) can take pictures of quantum systems. But these pictures are huge and messy.

• The Old Way: Scientists tried to take all these messy photos and force them into a low-dimensional map (like squishing a 3D globe into a 2D flat map). This often lost important details or required guessing which map was best.
• The New Way (Distance Learning): Instead of trying to draw a map, the scientists ask a simpler question: "How different are these two snapshots from each other?"

2. The Tool: The "Taste Tester" (The Discriminator)

To figure out how different two groups of snapshots are, the authors use a neural network (a type of AI) acting as a Taste Tester.

• The Setup: Imagine you have two jars of soup. Jar A is made by Chef 1, and Jar B is made by Chef 2. You have a bowl of soup from each jar, but you don't know which is which.
• The Training: You feed the AI thousands of soup samples, telling it, "This one is from Chef 1, this one is from Chef 2." The AI learns to recognize the subtle differences in flavor (the statistical patterns).
• The Result: Once trained, the AI doesn't just say "Chef 1" or "Chef 2." It calculates a score of how likely a new soup sample is to belong to Chef 1 versus Chef 2.
• The Magic: By comparing these scores, the AI can calculate the exact "statistical distance" between Chef 1's soup and Chef 2's soup. If the score is high, the soups are very different. If it's low, they are almost the same.

3. The Discovery: Finding the "Phases"

The scientists applied this to quantum physics. They took snapshots of a quantum system under different conditions (like changing the temperature or magnetic field).

• Clustering: They fed the "distance scores" into a clustering algorithm. Think of this like a party where guests are asked to stand next to people who taste similar.
  • Guests who taste like "Ferromagnetic Soup" (spins aligned) group together.
  • Guests who taste like "Paramagnetic Soup" (spins chaotic) group together.
• The Result: The AI successfully separated the quantum world into distinct "neighborhoods" (phases of matter) without ever being told what a "phase" was. It just knew that some snapshots felt very different from others.

4. The Bonus: Measuring the "Critical Point"

The most exciting part is that this method doesn't just find the neighborhoods; it finds the borderline.

• The Analogy: Imagine walking from a hot desert to a cold tundra. At some point, you hit a "transition zone" where the weather changes rapidly.
• The Metric: The AI measures how much the "taste" of the soup changes as you tweak the recipe slightly.
  • If you change the temperature a tiny bit and the soup tastes the same, you are far from the transition.
  • If you change the temperature a tiny bit and the soup tastes completely different, you are standing right on the critical point (the phase transition).
• The Power: By measuring this sensitivity, the scientists could calculate the "critical exponents" (mathematical numbers that describe how the universe behaves at the edge of change). This is usually very hard to do, but the AI did it just by looking at the snapshots.

5. Real-World Tests

The team tested this on three very different types of quantum systems:

1. The Classic Ising Model: Like a row of magnets flipping back and forth. The AI found the exact point where they stop flipping and line up.
2. The Toric Code: A system with "topological order" (like a knot that can't be untied). This is hard to see with normal tools, but the AI found the boundaries where the "knot" unravels.
3. The Triangular Lattice: A complex system where particles form strange "bound states" (like dance partners). The AI figured out exactly when these partners formed and when they broke apart, even when the number of particles changed.

Why This Matters

• No Preconceptions: The AI didn't need to know the laws of physics beforehand. It just looked at the data and said, "These groups are different."
• Robust: It works even if the data is noisy or if you look at the system from different angles (bases).
• Universal: It can find the "rules of the game" for any quantum system, whether it's a simple magnet or a complex superconductor.

In summary: Instead of trying to summarize a complex quantum world into a simple chart, this method asks the AI to simply measure how "far apart" different states are. By mapping these distances, the AI reveals the hidden structure of the quantum world, finding the boundaries between different states of matter and the critical points where the universe changes its rules.

Technical Summary

1. Problem Statement

Modern quantum simulators (both digital and analogue) generate large ensembles of single-shot projective measurements ("snapshots") of quantum many-body systems. While these snapshots contain rich information about the underlying quantum state, analyzing them poses significant challenges:

• Data Complexity: The data is high-dimensional, and traditional methods of extracting local observables or correlation functions may miss non-local or higher-order correlations.
• Limitations of Representation Learning: Previous unsupervised machine learning approaches typically rely on representation learning (e.g., PCA, diffusion maps) to compress snapshots into low-dimensional vectors before clustering. However, there is no systematic understanding of which representation best suits a specific physical system, and the choice of representation can introduce biases.
• Need for Direct Metric Estimation: There is a need for a method that can directly infer the statistical distance between quantum states from raw snapshots without intermediate dimensionality reduction, enabling robust phase diagram mapping and critical exponent extraction.

2. Methodology: Discriminative Distance Learning

The authors propose a framework called Distance Learning, which bypasses representation learning to directly estimate pairwise statistical distances between quantum states.

A. Core Concept

Instead of mapping snapshots to a latent space, the method treats the problem as estimating the Csiszár $f$-divergence between probability distributions (Born distributions) of different quantum states.

• Input: A dataset of snapshots $x$ collected from various points $\eta$ in the phase diagram.
• Discriminator: A single neural network (classifier) is trained in a self-supervised manner. Given a snapshot $x$, the network predicts the probability $P(i|x)$ that the snapshot originated from the quantum state at phase diagram point $\eta^{(i)}$.
• Density Ratio Estimation: Under the assumption of Bayesian optimality, the ratio of posterior probabilities $r(x) = P(q|x)/P(p|x)$ produced by the trained discriminator equals the density ratio $q(x)/p(x)$ of the underlying distributions.

B. Estimating $f$-Divergences

Using the estimated density ratio, the authors employ an importance sampling Monte Carlo estimator to compute the $f$-divergence $D_f(q||p)$:
$$D_f(q||p) \approx \frac{1}{2N_s} \sum_{x \sim p+q} w(x) \cdot f(r(x))$$
where $w(x)$ is an importance weight. The paper specifically utilizes the Squared Hellinger Divergence ($H^2$), as its square root satisfies the axioms of a true metric and relates directly to the overlap of real, non-negative wavefunctions.

C. Phase Clustering and Susceptibility

1. Clustering: The pairwise distances form a matrix $D_{ij}$. This matrix is fed into the HDBSCAN (Hierarchical Density-Based Spatial Clustering) algorithm. HDBSCAN is chosen because it requires no prior assumption about the number of clusters, can handle arbitrary cluster shapes, and identifies "noise" points (critical regions) that do not belong to distinct phases.
2. Critical Scaling: The infinitesimal limit of the inferred divergence relates to the Fisher Information Metric (or fidelity susceptibility for quantum ground states). By analyzing the finite-size scaling of this susceptibility, the method extracts critical exponents and identifies universality classes.
3. Multi-Basis Handling: The framework naturally incorporates measurements taken in different bases by treating the basis label as part of the input, summing the divergences weighted by the fraction of snapshots per basis.

3. Key Contributions

• Paradigm Shift: Moves from "representation learning" (compressing data) to "distance learning" (directly estimating statistical distances), avoiding the ambiguity of choosing optimal latent representations.
• Information-Geometric Foundation: Establishes a rigorous link between machine learning classifiers and information geometry, showing that the learned discriminator outputs provide access to the Fisher Information metric.
• Universality Class Identification: Demonstrates that critical exponents can be extracted directly from snapshot data via finite-size scaling of the learned susceptibility, without needing a known Hamiltonian or order parameter.
• Robustness: The method is shown to be robust to hyperparameter choices (network depth, width) and works with modest sampling budgets ($\sim 10^3$ snapshots per point).

4. Results and Applications

The method was applied to four distinct systems, successfully recovering known physics and revealing new insights:

A. 1D Transverse-Field Ising Model (TFIM)

• Task: Map the transition between ferromagnetic and paramagnetic phases.
• Result: The distance matrix clearly showed block structures corresponding to the two phases. HDBSCAN correctly identified the transition point.
• Critical Exponents: Finite-size scaling of the Hellinger susceptibility yielded critical exponents $\nu \approx 1.13$ and $\alpha_F \approx 0.8$, matching theoretical values ($\nu=1, \alpha_F=1$) within error margins.

B. 2D Classical Ising Model

• Task: Probe the thermal phase transition.
• Result: The learned susceptibility matched results from Ferrenberg-Swendsen histogram reweighting. The method successfully captured the logarithmic divergence of the specific heat (characteristic of the 2D Ising universality class, $\alpha=0$).

C. Extended Toric Code (Topological Order)

• Task: Map phases with non-local topological order (deconfined, confined, Higgs) where no local order parameter exists.
• Result:
  • Using snapshots from both $X$ and $Z$ bases, the method reconstructed the full phase diagram, identifying confinement and Higgs transitions.
  • Discovery: At high external fields, the method identified a distinct "supercritical-like" region adjacent to the first-order line that was not attributed to any single cluster. This region corresponds to a crossover where observables peak, suggesting a structure in the snapshot distribution distinct from the adjacent phases.

D. Fermionic $t-J$ Model on a Triangular Lattice

• Task: Analyze systems with higher-order correlations (Nagaoka polarons and kinetically induced bound states).
• Result:
  • Nagaoka Phase: The method clustered regions based on the position of the peak in the spin structure factor, identifying the Nagaoka ferromagnetic regime.
  • Higher-Order Correlations: By constructing surrogate distributions that matched only 1st and 2nd order marginals, the authors showed that the learned distance was sensitive to 3rd-order correlations (doublon-spin-spin), proving the method captures physics beyond standard correlation functions.
  • Cross-Sector Generalization: The authors demonstrated that a discriminator trained on one particle-number sector (e.g., 1 hole) could generalize to another (e.g., 2 holes), allowing for the comparison of non-overlapping quantum states and identifying bound state thresholds.

5. Significance and Outlook

• Experimental Relevance: The method is "Hamiltonian-agnostic," making it ideal for experimental quantum simulators where the exact Hamiltonian might be unknown or complex. It works directly on raw measurement data.
• Interpretability: While primarily an exploration tool, the paper shows how distance learning can be combined with surrogate data and cross-sector tests to interpret the nature of discovered clusters (e.g., identifying that a cluster is driven by higher-order correlations).
• Future Directions: The authors suggest extending this framework to non-equilibrium dynamics (using $f$-divergences as a proxy for Loschmidt echoes), exploring complex-valued distances for Berry phases, and combining with real-space renormalization flows for fully unsupervised discovery of operator content.

In summary, this work establishes distance learning as a powerful, information-geometric tool for analyzing quantum many-body systems, capable of mapping phase diagrams, extracting critical exponents, and identifying complex correlation structures directly from experimental snapshots.