Hypothesis Tests for Observing Quantum Entanglement in HWW at the LHC

This paper proposes a novel strategy for detecting quantum entanglement in Higgs boson decays to $W$ boson pairs at the LHC by combining a continuous CGLMP inequality formulation with conditional denoising diffusion probabilistic models for neutrino reconstruction, projecting that robust 5$\sigma$ evidence will be achievable with the high-luminosity dataset expected at the HL-LHC.

The Gist

Imagine the universe as a giant, high-speed dance floor where particles are the dancers. Usually, when two dancers meet and part ways, their moves are independent; what one does doesn't instantly dictate what the other does. But in the strange world of quantum mechanics, particles can become "entangled." This is like a pair of dancers who, even after being separated by miles, instantly mirror each other's moves. If one spins left, the other spins right, no matter the distance. This connection is so strong that it defies the rules of classical physics.

This paper presents a new, clever way to prove that this "quantum dance" is happening when a Higgs boson (a heavy particle discovered at the Large Hadron Collider, or LHC) decays into two W bosons.

Here is the story of how the researchers solved the puzzle, explained simply:

1. The Problem: The Invisible Partners

When the Higgs boson decays into two W bosons, those W bosons immediately turn into other particles, including neutrinos. Neutrinos are like ghosts; they pass through everything and leave no trace in the detectors.

• The Challenge: To prove the dancers were entangled, physicists need to know exactly how they were spinning. But because the neutrinos are invisible, the physicists can't see the full picture. It's like trying to figure out a dance routine by only watching the shadows of the dancers, while two of the dancers are invisible.
• The Old Way: Previous methods tried to guess where the invisible neutrinos went using math equations. But these equations often failed or gave messy, unreliable results, especially when there was "noise" from other particle collisions (background events).

2. The New Tool: The AI "Denoising" Machine

The authors introduced a new type of artificial intelligence called a Conditional Denoising Diffusion Probabilistic Model (cDDPM).

• The Analogy: Imagine you have a photo of a dance that has been heavily blurred and covered in static (noise). Traditional methods try to guess the original photo by solving a complex puzzle, often getting it wrong.
• The AI Approach: This new AI works like a master restorer. It starts with a completely blurry, noisy image and slowly "denoises" it, step-by-step, until the clear picture of the original dance emerges. It learns from millions of simulated examples what the "ghost" neutrinos should look like based on the visible particles.
• The Benefit: Unlike older methods that needed to know the "truth" beforehand to work, this AI can look at real data (including the messy background noise) and reconstruct the invisible parts without getting confused. It effectively "fills in the blanks" of the invisible neutrinos with high accuracy.

3. The Test: From "Average" to "Shape"

Once they reconstructed the dance, they needed to check if it was entangled.

• The Old Method (The Flawed Average): Previously, scientists would calculate a single "average score" (an expectation value) to see if entanglement existed. The problem is that if one weird, rare event happens (an outlier), it can skew the entire average, making the result unreliable. It's like judging a whole orchestra's performance based on the loudest single note; if that one note is off, you think the whole concert was bad.
• The New Method (The Shape Test): Instead of looking for a single average number, the authors looked at the entire shape of the data distribution. They asked: "Does the overall pattern of the dance moves look like an entangled dance, or does it look like two independent dancers?"
• The Analogy: Think of it like identifying a song. Instead of measuring the average volume of the music, you listen to the melody and rhythm. Even if there is some static (noise), you can still recognize the song by its unique shape. This method is much more robust against errors and outliers.

4. The Results: Seeing the Quantum Connection

By combining the AI reconstruction with this new "shape-based" test, the researchers simulated what would happen with real data from the LHC.

• The Prediction: They found that with enough data (specifically, about 555 units of "luminosity," which is a measure of how many collisions occur), they could see evidence of entanglement with a high degree of confidence (3-sigma, which is strong evidence).
• The Future: If they wait for the High-Luminosity LHC (which will run for several years and produce much more data, around 1600 units), they expect to reach a "5-sigma" result. In physics, 5-sigma is the gold standard for a discovery—it means there is less than a one-in-a-million chance that the result is a fluke.

Summary

In short, this paper proposes a new strategy to catch the "ghosts" (neutrinos) using a smart AI that cleans up the noise. Instead of relying on a fragile average number, they look at the overall shape of the data to prove that particles are dancing in perfect, mysterious unison. This method is robust, handles the messy reality of particle colliders well, and promises to confirm quantum entanglement in Higgs boson decays within the next few years of data collection.

Technical Summary

Problem Statement

The paper addresses the challenge of experimentally testing quantum entanglement in the decay of the Higgs boson to a pair of W bosons ($H \to WW^*$) at the Large Hadron Collider (LHC). While theoretical frameworks exist for testing Bell inequalities in these systems, current experimental approaches face two fundamental limitations:

1. Sensitivity to Outliers and Detector Effects: Traditional methods rely on computing expectation values of Bell operators (specifically the CGLMP inequality parameter $I_3$) derived from spin correlation coefficients. These expectation values are highly sensitive to events with extreme kinematic configurations caused by detector mismeasurements, background contamination, or selection cuts. The resulting distributions often exhibit heavy tails (resembling Cauchy distributions rather than Gaussians), making the assumption of a well-defined mean invalid and biasing the extracted Bell parameter.
2. Neutrino Reconstruction: The $H \to WW^* \to \ell\nu\ell\nu$ channel involves two invisible neutrinos, making the reconstruction of the W boson rest frames (essential for spin analysis) an underconstrained problem. Traditional analytical reconstruction methods fail for a significant fraction of events (approx. 35%) and, crucially, destroy per-event information. The paper demonstrates that classical unfolding of the reconstructed $I_3$ distribution yields zero coverage and fails to distinguish between entangled and separable hypotheses because the analytical reconstruction correlates poorly with truth-level values ($r \approx 0.08$).

Methodology

The authors propose a paradigm shift from discrete expectation value calculations to continuous hypothesis testing combined with generative unfolding for neutrino reconstruction.

1. Continuous Hypothesis Testing:
Instead of calculating a single expectation value $\langle I_3 \rangle$, the authors construct the full continuous probability distribution of the CGLMP observable. They perform a standard profile likelihood ratio test to distinguish between two hypotheses:

• Null Hypothesis ($H_0$): A fully separable (non-entangled) quantum state.
• Alternative Hypothesis ($H_1$): A fully entangled state (Standard Model prediction).
  The test utilizes a binned likelihood fit implemented in RooFit (using the HistFactory formalism), treating the separable configuration as the null hypothesis. This approach is robust against outliers and naturally incorporates background models through simultaneous fitting of signal and background components.

2. Neutrino Reconstruction via cDDPM:
To overcome the neutrino reconstruction challenge, the authors employ Conditional Denoising Diffusion Probabilistic Models (cDDPM).

• Target: The model targets the 6-dimensional helicity angles of the charged leptons in the W rest frames directly, rather than reconstructing the full 8-dimensional neutrino four-momenta.
• Conditioning: The model is conditioned on detector-level lepton four-momenta, missing transverse momentum, and an explicit one-hot process tag (identifying the physics process: $HWW$, $WW$, $t\bar{t}$, or $Z\to\tau\tau$).
• Advantage over Alternatives: Unlike reweighting-based methods like OmniFold, which require truth-level labels for all events and struggle in multi-background environments, cDDPM operates on individual events. It can be applied directly to the measured dataset (including backgrounds) without requiring truth-level labels during inference.
• Performance: The diffusion model is initialized with an analytical reconstruction solution (warm-start) to correct residual distortions. The authors demonstrate that this method preserves the shape of the $I_3$ distribution with high fidelity (Kolmogorov-Smirnov distance < 0.05 relative to truth) and maintains per-event correlations necessary for hypothesis testing.

3. Event Selection:
To improve the signal-to-background ratio without biasing the angular distributions, the authors utilize a Boosted Decision Tree (BDT) selection based strictly on non-angular kinematic variables (e.g., transverse momenta, invariant masses, missing energy). They explicitly exclude angular variables to prevent sculpting the $I_3$ distribution. This selection improves the signal-to-background ratio ($S/B$) from 4.5% to 13.9% while preserving the shape of the entanglement observable.

Key Contributions

• Continuous Formulation: The paper introduces a continuous formulation of the CGLMP inequality that enables standard statistical hypothesis testing, avoiding the statistical pitfalls of expectation-value-based approaches in collider environments.
• Diffusion-Based Unfolding: It demonstrates the application of conditional denoising diffusion models for multidimensional, object-wise unfolding in a high-energy physics context, specifically for neutrino reconstruction in a mixed-background environment.
• Systematic Propagation: The analysis fully propagates systematic uncertainties, including background normalizations and the unfolding shape bias, using a bootstrap-PCA template statistical uncertainty model that respects bin-to-bin correlations.
• Feasibility Study: It provides a comprehensive feasibility study showing that quantum entanglement in Higgs decays can be tested with existing LHC analysis tools and datasets.

Results

The study projects the sensitivity of the proposed method using the full systematic model and BDT-enhanced selection:

• 139 fb$^{-1}$ (Run 2): The expected Asimov significance for rejecting the separable hypothesis is 1.7$\sigma$ (inclusive) and 2.3$\sigma$ (BDT-enhanced).
• 3$\sigma$ Evidence: Projected to be achievable at approximately 555 fb$^{-1}$ (roughly 550 fb$^{-1}$).
• 5$\sigma$ Observation: Projected to be achievable at approximately 1600 fb$^{-1}$.
• HL-LHC (3000 fb$^{-1}$): The expected significance reaches 6.7$\sigma$.

The analysis identifies background normalization uncertainties (particularly for the irreducible $WW$ background) and unfolding shape uncertainties as the dominant systematic limitations, rather than statistical precision. The BDT selection improves sensitivity by a factor equivalent to tripling the integrated luminosity, primarily by rejecting reducible backgrounds ($t\bar{t}$ and $Z\to\tau\tau$) rather than the irreducible $WW$ background.

Significance and Claims

The paper claims that the proposed strategy provides a robust and practical path to measuring quantum entanglement in Higgs boson decays.

• Robustness: By shifting from expectation values to continuous distribution shapes, the method is less sensitive to detector effects and selection biases that previously hindered such measurements.
• Feasibility: The results demonstrate that observing quantum entanglement in $H \to WW^*$ is well within the reach of the High-Luminosity LHC (HL-LHC) luminosity targets, provided that systematic uncertainties (specifically background normalizations) are managed via data-driven control regions.
• Extension to New Physics: The authors note that the per-event correlation matrix ($c_{ij}$) reconstructed via this method is not limited to entanglement testing. It can be used to construct observables sensitive to CP-violating and anomalous CP-even couplings in the Higgs sector (SMEFT operators), allowing a single dataset to simultaneously test for entanglement, CP violation, and new physics.

The paper concludes that while the current sensitivity is limited by systematic floors, the continuous hypothesis testing framework combined with diffusion-based unfolding offers a mature, well-validated methodology for precision quantum information science at colliders.