Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach

This paper introduces a novel MERA-inspired tensor network autoencoder for collider jet anomaly detection, demonstrating that its locality-preserving hierarchical structure and disentangling layers provide a superior inductive bias for reconstructing jet data compared to dense and tree-based architectures, particularly under strong compression.

The Gist

Imagine you are a detective trying to find a counterfeit bill in a massive pile of real money. The real bills (the "background") all follow a specific pattern: they have a certain texture, ink distribution, and watermark. The fake bills (the "anomalies") might look real at first glance, but they have subtle flaws in how the ink is layered or how the fibers are arranged.

In the world of particle physics, scientists at the Large Hadron Collider (LHC) face a similar problem. They are bombarded with trillions of particle collisions. Most of these are just "background noise" (ordinary particles like quarks and gluons behaving as expected). Occasionally, a "signal" appears—a rare particle or a new physics phenomenon that looks different. The challenge is to spot the fake bill (the anomaly) without knowing exactly what it looks like beforehand.

This paper introduces a new, clever detective tool called a MERA-based Autoencoder. Here is how it works, broken down into simple concepts:

1. The Problem with Standard Detectors

Usually, scientists use a "Dense Autoencoder" to solve this. Think of this as a detective who looks at the whole pile of money at once, shuffling every single bill around randomly to find patterns.

• The Flaw: While this detective is smart, they treat the bills as a messy, unorganized heap. They have to learn from scratch that the ink on the top left corner usually matches the ink on the bottom right. It's inefficient and misses the natural structure of the data.

2. The Physics of a "Jet"

In particle physics, when a high-energy particle decays, it doesn't just explode randomly. It creates a "jet" of particles that branch out like a tree.

• The Analogy: Imagine a tree growing. The trunk splits into big branches, which split into smaller twigs, which split into leaves. The leaves near each other on a twig are closely related; they share a recent history.
• The Insight: The authors realized that if you want to detect a fake jet, you shouldn't look at the whole mess at once. You should look at the branches and how they connect, step-by-step, from the big trunk down to the tiny leaves.

3. The New Tool: The "Tree-Structured" Detective (MERA)

The authors built a new detector inspired by a concept from quantum physics called MERA (Multiscale Entanglement Renormalization Ansatz).

Think of MERA as a specialized sorting machine that respects the tree structure of the jet:

• Step 1: Organizing the Neighbors. First, the machine rearranges the particles so that neighbors in the list are actually neighbors in space (like sorting books by their shelf location rather than by height).
• Step 2: The "Unscrambler" (Disentanglers). Before compressing the data, the machine has a special step called a "disentangler." Imagine you have a tangled ball of yarn. Before you can shrink it, you have to untangle the knots. This step untangles the local relationships between nearby particles so they make sense.
• Step 3: The "Shrinking" (Coarse-Graining). Then, the machine groups pairs of particles together and summarizes them into a single, simpler "super-particle." It repeats this process, moving up the tree, until the whole jet is summarized into a tiny, compact code.

4. How It Detects Anomalies

The machine is trained only on the "real" background jets (the real money).

• The Test: When a new jet comes in, the machine tries to compress it and then rebuild it.
• The Result: If the jet is "real," the machine can easily untangle it, shrink it, and rebuild it perfectly.
• The Anomaly: If the jet is "fake" (an anomaly), the machine gets confused. The "unscrambler" can't untangle it because the patterns don't match the training. When it tries to rebuild it, the result is blurry or wrong. The bigger the error in the rebuild, the more likely it is a fake.

5. Why This Paper is a Big Deal

The authors didn't just build this tool; they tested it rigorously to prove why it works:

• It's Smarter and Cheaper: The new MERA detector found more anomalies than the standard "messy heap" detector, but it used one-third fewer computer resources (parameters). It's like solving a puzzle with fewer pieces because you know where they fit.
• Order Matters: They proved that sorting the particles by their physical location (neighbors next to neighbors) is crucial. If you sorted them randomly, the detector got confused. This confirms that the "tree structure" of the universe is real and useful.
• The "Unscrambler" is Key: They tested a version of the machine without the "unscrambler" step. It worked okay, but when the data was very hard to compress (the "tightest" squeeze), the unscrambler made a huge difference. It proved that untangling local relationships before shrinking is a necessary step for high-precision detection.

The Bottom Line

This paper shows that by mimicking the natural, hierarchical way the universe builds particle jets (like a tree growing), we can build much better AI detectors. Instead of forcing the AI to learn everything from scratch, we give it a "head start" by telling it, "Hey, these particles are neighbors; treat them like neighbors."

This approach allows scientists to spot the rare, weird events that might hide new physics, using a tool that is both more powerful and more efficient than what we had before.

Technical Summary

1. Problem Statement

The search for physics beyond the Standard Model (BSM) at the Large Hadron Collider (LHC) increasingly relies on anomaly detection to identify unexpected phenomena without relying on specific signal hypotheses. A common approach is reconstruction-based anomaly detection, where a model is trained exclusively on background data (Standard Model jets) to learn the typical structure of these events. Events that cannot be accurately reconstructed are flagged as anomalies.

However, standard deep learning approaches, such as Dense Autoencoders (AEs), treat jet constituents as unstructured vectors. They fail to explicitly encode the hierarchical, multiscale nature of jet formation. Jets are produced via a branching cascade (QCD shower) where hard, large-angle splittings determine coarse structures, and softer, collinear radiation refines these structures at smaller scales. The authors hypothesize that an architecture explicitly designed to handle locality-preserving hierarchical compression would be a more effective inductive bias for jet data than generic dense networks.

2. Methodology

The authors propose and evaluate a MERA-inspired Autoencoder (Multiscale Entanglement Renormalization Ansatz), a tensor network architecture originally developed for quantum many-body physics.

A. Architecture Design

• Input Representation: Jets are represented as ordered lists of $N=48$ constituents. Each constituent has three features: transverse momentum ($p_T$) and angular differences relative to the jet axis ($\Delta\eta, \Delta\phi$).
• Locality-Preserving Ordering: Unlike standard benchmarks that order constituents by decreasing $p_T$, the authors reorder constituents based on geometric proximity in the $(\Delta\eta, \Delta\phi)$ plane. This ensures that "neighboring" sites in the tensor network correspond to physically adjacent particles in the jet.
• MERA Encoder:
  • The encoder acts on the ordered chain of constituents.
  • It alternates between Disentanglers ($U$) and Isometries ($W$).
  • Disentanglers: Local unitary transformations that reorganize short-range correlations before compression.
  • Isometries: Projections that coarse-grain pairs of sites into a single effective site, reducing the dimensionality (e.g., $48 \to 24 \to 12 \to 6 \to 3$).
  • The process ends at a latent bottleneck vector $z$.
• MERA Decoder: An untied decoder mirrors the encoder structure, expanding the latent vector back to the original constituent space using learned expansion maps and mixing operations.
• Training Objective: The model is trained on background-only QCD jets to minimize the Mean Squared Error (MSE) reconstruction loss. Soft penalties are added to enforce the orthogonality of disentanglers and the isometric nature of the projections.
• Anomaly Score: The reconstruction error ($L_{rec}$) serves as the anomaly score. High error indicates the event does not fit the learned background manifold.

B. Baselines and Comparisons

To validate the approach, the MERA model is compared against:

1. Dense Autoencoder (AE): A standard fully connected network with the same latent dimensions ($B \in \{8, 16, 32\}$).
2. Tree Tensor Network (TTN) Limit: A MERA variant where disentanglers are fixed to the identity matrix ($U=I$), removing the ability to reorganize correlations.
3. Classical Baselines: Principal Component Analysis (PCA), Gaussian Mahalanobis distance, and Isolation Forest.

3. Key Contributions

1. First MERA Application in Collider Anomaly Detection: This is the first study to formulate a MERA-based autoencoder operating directly on ordered jet constituents for anomaly detection, moving beyond previous applications of tensor networks in high-energy physics (which focused on event reconstruction or matrix product states in latent spaces).
2. Validation of Locality-Preserving Ordering: The paper demonstrates that the specific ordering of jet constituents is not merely a preprocessing step but a critical architectural component. Ordering by geometric proximity significantly outperforms $p_T$-ordering or random permutations.
3. Empirical Proof of Disentangler Utility: Through ablation studies, the authors show that the disentangling layers provide a measurable performance boost over a simple tree hierarchy, specifically when the compression bottleneck is tight.
4. Parameter Efficiency: The MERA architecture achieves superior performance with significantly fewer trainable parameters compared to dense autoencoders.

4. Results

The study was conducted on the Top Quark Tagging Reference Dataset, using QCD jets as background and hadronically decaying top jets as the "anomaly" proxy.

• Nominal Benchmark Performance:
  • At all latent dimensions ($B=8, 16, 32$), the MERA model outperformed the Dense Autoencoder.
  • AUC Gains: MERA achieved AUCs of 0.7959, 0.7885, and 0.7963 for $B=8, 16, 32$, respectively, compared to 0.7616, 0.7487, and 0.7550 for the Dense AE.
  • Parameter Efficiency: MERA used only ~5,000 parameters, roughly one-third of the parameters required by the Dense AE (~17,000), yet achieved higher accuracy.
• Locality and Compressibility:
  • The locality-preserving ordering reduced the mean angular separation between adjacent sites to $\langle \Delta R \rangle = 0.131$ (vs. 0.361 for $p_T$ ordering).
  • This ordering resulted in higher retained variance in the first two compression layers (0.869 and 0.920), confirming that the input data is naturally compressible in a geometrically local manner.
  • Models trained on locality-ordered data consistently outperformed those trained on $p_T$-ordered or random data.
• Disentangler Ablation:
  • In the strongest compression regime ($B=8$), the full MERA (with trainable disentanglers) achieved a mean AUC of 0.8002, compared to 0.7922 for the TTN limit (identity disentanglers).
  • This confirms that the ability to rotate local bases before coarse-graining is crucial for capturing the complex correlations in jet substructure when information must be aggressively compressed.

5. Significance

This paper establishes that multiscale, locality-aware compression is a highly effective inductive bias for jet anomaly detection.

• Physics Alignment: The results validate the physical intuition that jet substructure is organized hierarchically. By mimicking the renormalization group flow (coarse-graining while preserving relevant correlations), MERA models the QCD background more efficiently than generic neural networks.
• Efficiency: The ability to achieve state-of-the-art anomaly detection with fewer parameters suggests that tensor networks could be a resource-efficient alternative to large dense networks, which is valuable for deployment in high-throughput collider environments.
• Methodological Insight: The study highlights that the success of tensor networks in this domain is not just due to the architecture itself, but critically depends on aligning the data representation (ordering) with the inductive bias of the model.

In conclusion, the authors demonstrate that a MERA-based autoencoder is not only viable but superior to standard dense autoencoders for collider anomaly detection, provided the input data is ordered to respect the geometric locality of the jet constituents.