Neural simulation-based inference of the Higgs trilinear self-coupling via off-shell Higgs production

This paper proposes a hybrid neural simulation-based inference approach to constrain the Higgs trilinear self-coupling and other SMEFT operators using off-shell Higgs production at the High-Luminosity LHC, achieving near-theoretical-optimal sensitivity by combining matrix-element-enhanced training with classification-based background estimation.

The Gist

Imagine the universe is a giant, high-speed collision course where tiny particles smash into each other, creating a shower of new particles. At the center of this chaos sits the Higgs boson, a particle that gives mass to everything else. Physicists want to understand how the Higgs interacts with itself—specifically, how three Higgs particles might clump together. This is called the Higgs trilinear self-coupling.

Think of the Higgs field like a trampoline. If you bounce one ball on it, it's easy to understand. But if you throw three balls at once, how they bounce off each other tells you exactly how "springy" the trampoline is. If the bounce doesn't match our predictions, it means there's a hidden spring or a secret weight under the trampoline—evidence of New Physics beyond our current understanding.

The Problem: The "Ghost" Signal

Usually, scientists look for the Higgs when it's "on-shell," meaning it's produced as a real, stable particle that we can catch and measure. It's like trying to identify a specific singer by listening to their clear, recorded voice.

However, the Higgs can also be produced "off-shell." This is like the singer humming a note so briefly and faintly that it never fully forms a voice; it's a ghostly, fleeting vibration that disappears almost instantly. This "off-shell" signal is incredibly weak and gets drowned out by the noise of other particles (background noise) crashing into each other. Traditional methods of listening to this ghostly signal are like trying to hear a whisper in a hurricane using only a simple volume meter.

The Solution: A Neural "Super-Listener"

The authors of this paper built a Neural Simulation-Based Inference (NSBI) system. Think of this as a super-smart AI detective.

Instead of just counting how many times a signal happens (like a volume meter), this AI looks at the entire shape and pattern of the collision. It's like the difference between a security guard counting how many people enter a building versus a detective who analyzes the gait, clothing, and behavior of every single person to spot a specific suspect.

The AI was trained on massive computer simulations (like a flight simulator for particle physics) that included:

1. The Signal: The ghostly off-shell Higgs.
2. The Noise: The background particles that look similar.
3. The Interference: A tricky quantum effect where the signal and noise cancel each other out or amplify each other, like two sound waves meeting.

How They Tested It

The team simulated collisions at the High-Luminosity Large Hadron Collider (HL-LHC), which is the future, super-powered version of the current particle collider. They looked at two specific scenarios:

• The "Clean" Room (4 Leptons): Four charged particles (electrons or muons) fly out. This is like a clear, high-definition photo. The AI performed almost perfectly here, matching the theoretical "gold standard" of what is physically possible.
• The "Foggy" Room (2 Leptons + 2 Neutrinos): Two particles fly out, but two others (neutrinos) are invisible ghosts that escape detection. This is like trying to identify a suspect in a foggy room where half the people are invisible. The AI couldn't see the full picture, so its performance dropped, but it was still much better than just counting the total number of events.

The Results: Breaking the "Flat" Mystery

The main goal was to measure the "springiness" of the Higgs trampoline.

• Single Measurement: When looking at just the Higgs self-interaction, the off-shell method wasn't quite as sensitive as the traditional "on-shell" methods. It's like trying to measure the trampoline's springiness by listening to a faint hum; it's hard to get a precise number.
• The Real Win (The "Flat Direction"): The real magic happened when they looked at the Higgs along with other interactions (specifically how the Higgs talks to the top quark and how it's created by gluons).
  • Imagine trying to solve a puzzle where two pieces look identical. Traditional methods can't tell them apart; the solution is "flat" (you can't decide which is which).
  • The AI, by analyzing the subtle shapes of the data, could lift this flatness. It could distinguish between the different ways the Higgs interacts, effectively separating the "springiness" of the trampoline from the "weight" of the top quark.

The Bottom Line

This paper doesn't claim to have found new physics yet. Instead, it proves that AI can act as a powerful microscope for the faintest, most elusive signals in particle physics.

By using this neural network approach, physicists can:

1. Extract more information from the "ghostly" off-shell Higgs than ever before.
2. Break through "blind spots" where traditional math fails to distinguish between different theories.
3. Prepare for the future HL-LHC, ensuring that when the machine turns on, we are ready to spot the tiniest deviations from the Standard Model that could reveal a new universe.

In short: They built a smarter way to listen to the universe's faintest whispers, proving that even when the signal is hidden in the noise, a neural network can find the pattern.

Technical Summary

Problem Statement

The experimental determination of the Higgs trilinear self-coupling ($\lambda$) is a primary objective for the High-Luminosity Large Hadron Collider (HL-LHC). While current efforts focus on on-shell double-Higgs ($pp \to hh$) and single-Higgs production, these channels often suffer from degeneracies when other Beyond the Standard Model (BSM) effects are considered. For instance, in scenarios where both the Higgs trilinear coupling and the top-quark Yukawa coupling are modified, the double-Higgs cross-section exhibits a flat direction in parameter space, making it difficult to constrain these parameters independently.

Off-shell Higgs production via gluon-gluon fusion ($gg \to h^* \to ZZ \to 4\ell$ or $2\ell2\nu$) offers a complementary probe. However, extracting constraints from this channel is challenging due to:

1. Complex Kinematics: The signal involves quantum interference between the Higgs-mediated process, the continuum box background ($gg \to ZZ$), and the tree-level background ($q\bar{q} \to ZZ$).
2. SMEFT Complexity: Modifications are described by the Standard Model Effective Field Theory (SMEFT), involving dimension-six operators that affect the Higgs propagator, production, and decay.
3. Data Limitations: Traditional histogram-based analyses lose information by binning high-dimensional data, reducing sensitivity to subtle shape distortions caused by BSM physics.

Methodology: Hybrid Neural Simulation-Based Inference (NSBI)

The authors propose a hybrid Neural Simulation-Based Inference (NSBI) framework to construct an extended likelihood ratio for the $pp \to ZZ$ process. This approach leverages the training efficiency of matrix-element (ME) enhanced techniques while utilizing classification methods for background estimation.

1. Theoretical Framework and Simulation:

• SMEFT Parametrization: The analysis focuses on three dimension-six operators: $Q_H$ (modifying the Higgs potential/trilinear coupling), $Q_{tH}$ (modifying the top-quark Yukawa coupling), and $Q_{HG}$ (modifying the Higgs-gluon coupling).
• Monte Carlo Generation: The authors modified the MCFM 10.3 generator to include:
  • Full SM amplitudes for $gg \to h^* \to ZZ$, $gg \to ZZ$, and $q\bar{q} \to ZZ$.
  • SMEFT corrections at the one- and two-loop levels, including quantum interference effects between signal and background.
  • The code calculates squared matrix elements ($|M|^2$) for arbitrary Wilson coefficients ($C_H, C_{tH}, C_{HG}$).

2. The NSBI Architecture:
The method employs a probability mixture model to handle the inclusive $pp \to ZZ$ process, which consists of two distinct components:

• $gg$-initiated processes: These depend on the Wilson coefficients and include complex interference.
• $q\bar{q}$-initiated processes: These are independent of the Wilson coefficients (at the order considered) and act as a background.

To estimate the likelihood ratio $p(x|C)/p(x|0)$, where $x$ are observables and $C$ are Wilson coefficients, the authors train two types of Neural Networks (NNs):

• Regressor (for $gg$ processes): Trained to estimate the ratio of probability densities relative to the Standard Model (SM) hypothesis. It learns a Taylor expansion of the rate-normalized probability ratio up to the fourth order in the Wilson coefficients. This utilizes the exact squared MEs from the MC generator as "auxiliary information" to train the network on the latent-to-observable mapping.
• Classifier (for $q\bar{q}$ processes): Trained as a binary classifier to distinguish between $q\bar{q}$ and $gg$ initiated events (or specific background hypotheses), as these processes do not depend on the SMEFT parameters in the same way.

3. Calibration and Validation:
The authors implement rigorous calibration procedures to ensure the NN outputs are reliable probability ratios. This includes:

• Calibration Closure: Verifying that the classifier output matches the true fraction of events from the numerator hypothesis.
• Reweighting Closure: Comparing the NN-estimated reweighting of SM events to BSM scenarios against the exact parton-level ME ratios.
• Expectation Value Correction: Rescaling the probability ratio to ensure the best-fit parameters converge to the true values in Asimov datasets.

Key Contributions

• Hybrid NSBI Implementation: The paper introduces a novel method that combines the strengths of direct joint probability ratio estimation (feasible for $gg$ processes with ME access) and binary classification (for background processes), circumventing the limitations of applying a single NSBI technique to the entire inclusive process.
• MCFM 10.3 Extension: The authors provide a publicly accessible, modified version of MCFM 10.3 that includes full two-loop SMEFT corrections and interference effects for off-shell Higgs production, validated against previous implementations.
• Comprehensive Sensitivity Analysis: The study performs the first NSBI-based sensitivity analysis for off-shell Higgs production at the HL-LHC, covering both $4\ell$ and $2\ell2\nu$ final states and analyzing constraints on single and pairwise combinations of Wilson coefficients.

Results

The analysis assumes $\sqrt{s} = 14$ TeV and an integrated luminosity of $3 \text{ ab}^{-1}$.

1. Single-Parameter Constraints ($C_H$):

• $4\ell$ Channel: The NSBI approach achieves near-optimal sensitivity, closely matching the theoretical parton-level limit. This is because the full kinematics of the four leptons allow for complete reconstruction of the event.
• $2\ell2\nu$ Channel: Sensitivity is reduced compared to the parton-level limit due to the inability to fully reconstruct the neutrino momenta (projecting high-dimensional latent space to lower-dimensional observables). However, NSBI still outperforms rate-only analyses.
• Combined Constraint: The projected 95% CL (2$\sigma$) constraint on $C_H$ is $-43.5 < C_H < 26.9$ (assuming $\Lambda = 1$ TeV), corresponding to $-11.6 < \kappa_\lambda < 21.4$. This is weaker than current on-shell constraints from ATLAS and CMS but is derived in a model-independent context.

2. Multi-Parameter Constraints and Degeneracy Breaking:

• $C_H$ vs. $C_{tH}$: NSBI significantly tightens constraints on $C_{tH}$ compared to rate-only measurements, particularly for positive values, by exploiting shape distortions in the distributions.
• $C_{tH}$ vs. $C_{HG}$: The off-shell analysis partially lifts the "flat direction" ($C_{HG} \approx 4.9 \times 10^{-3} C_{tH}$) that exists in on-shell double-Higgs production. This is crucial because on-shell $pp \to hh$ depends linearly on $C_{HG}$ but quadratically on $C_{tH}$, whereas off-shell $pp \to ZZ$ offers different dependencies that help decouple these parameters.
• Complementarity: The paper demonstrates that while on-shell double-Higgs production provides stronger constraints on $C_H$ alone, the off-shell $pp \to ZZ$ channel is essential for constraining the $C_{tH}$ and $C_{HG}$ operators and resolving degeneracies in the multi-dimensional SMEFT parameter space.

Significance and Claims

The authors claim that their NSBI framework provides a robust, near-optimal method for extracting SMEFT constraints from complex processes involving interference and high-dimensional data.

• Robustness: The method successfully handles the mixture of signal and background processes and incorporates quantum interference effects that are often neglected in simpler analyses.
• Complementarity: The primary significance lies in demonstrating that off-shell Higgs production is not merely a sub-leading probe for the trilinear coupling but a vital tool for breaking degeneracies in global SMEFT fits, specifically between the Higgs self-coupling, top-Yukawa, and Higgs-gluon couplings.
• Future Outlook: The authors note that while systematic uncertainties (experimental and theoretical) are not included in this projection, the NSBI approach offers a path toward more comprehensive analyses. They suggest that future work could incorporate jet multiplicities, electroweak production channels, and on-shell single-Higgs differential information to further constrain the Higgs potential.

The paper concludes that including off-shell Higgs production in global SMEFT analyses is necessary to fully explore the parameter space of the Higgs trilinear self-coupling and related operators, particularly in scenarios where on-shell measurements alone are insufficient.