Log Gaussian Cox Process Background Modeling in High Energy Physics

This paper introduces a novel Log Gaussian Cox Process (LGCP) method for modeling smooth backgrounds in high energy physics that minimizes assumptions about the underlying shape by utilizing a Gaussian process for the intensity function and Markov Chain Monte Carlo for optimization, demonstrating its effectiveness through synthetic experiments against traditional analytic functional forms.

The Gist

Imagine you are a detective trying to find a specific, rare criminal (a new particle) hiding in a massive crowd of innocent bystanders (background noise) at a giant, chaotic concert (the Large Hadron Collider).

The problem? The crowd is huge, and the "innocent" people are moving in a smooth, predictable pattern. If you just look at the crowd, you might think a sudden bump in the crowd density is the criminal, when it's actually just a random fluctuation or a weird shape in the crowd's natural movement.

For decades, physicists have tried to solve this by guessing a mathematical formula (like a specific type of curve) to describe how the innocent crowd moves. They fit this curve to the data and look for a "bump" that doesn't fit the curve. But this is risky: if you guess the wrong curve, you might miss the criminal or falsely accuse an innocent person.

This paper introduces a new, smarter detective tool called the Log Gaussian Cox Process (LGCP). Here is how it works, using simple analogies:

1. The Old Way: The "Rigid Blueprint"

Imagine you are trying to trace the outline of a cloud. The old method forces you to use a ruler and a set of pre-made stencils (circles, squares, triangles). You have to pick the stencil that looks closest to the cloud.

• The Problem: If the cloud is weirdly shaped, your stencil won't fit perfectly. You might force the circle to look like a cloud, creating a fake "bump" where there isn't one, or missing a real bump because your circle is too rigid.

2. The New Way: The "Flexible Rubber Sheet" (LGCP)

The LGCP method doesn't use rigid stencils. Instead, imagine a giant, stretchy rubber sheet.

• How it works: You drop pins on the sheet wherever you see data points (events). The sheet naturally stretches and settles into a shape that fits the pins perfectly, without you forcing it into a circle or square.
• The "Log Gaussian" part: This is just the physics-speak for "a very smart, flexible sheet that knows how to stretch smoothly." It assumes the background noise is random (like rain falling) but that the intensity of the rain follows a smooth, wavy pattern.
• The Benefit: It doesn't need to guess a formula. It just learns the shape of the background directly from the data.

3. The "Spurious Signal" Problem (The False Alarm)

In the paper, the authors test their new method against the old one using "Toy Datasets" (fake data generated by computers).

• The Test: They create a crowd of innocent people and ask the detective: "Is there a criminal here?"
• The Old Method (MLE): Sometimes, the rigid stencil fits the crowd so poorly that a random wobble in the crowd looks like a criminal. This is called a "spurious signal" (a false alarm).
• The LGCP Method: Because the rubber sheet is so flexible, it hugs the crowd's natural shape very well. It rarely mistakes a random wobble for a criminal.

4. The Catch: The "Edge of the Map"

The paper found one weakness. The rubber sheet works great in the middle of the crowd, but near the very edges (the boundaries of the data), it sometimes gets a little confused and stretches too far.

• Analogy: If you stretch a rubber sheet over a table, the middle is smooth, but the edges might curl up weirdly.
• The Fix: The authors suggest simply ignoring the very edges of the data or using a wider area to fit the sheet, then cutting off the weird edges.

5. Finding the Real Criminal (Signal Injection)

Finally, they tested if the new method could actually find a real criminal if one was planted in the crowd.

• Result: The LGCP method was excellent at spotting the real criminal (up to a certain size) without getting confused by the background noise.
• Comparison: Another flexible method (called GPR) was good at smoothing the background but was sometimes too smooth, effectively "hiding" the criminal by smoothing them out into the background. The LGCP was just right: flexible enough to fit the background, but sharp enough to spot the bump.

The Bottom Line

This paper proposes a new way to model background noise in particle physics. Instead of forcing the data into a rigid mathematical box, it uses a flexible, data-driven "rubber sheet" approach.

• Why it matters: It reduces false alarms (accusing innocent particles of being new physics) and improves the chances of finding real new particles.
• The Verdict: It's a powerful new tool for the "bump hunt," provided you are careful about the edges of your data. It makes the search for new physics faster, more accurate, and less dependent on guessing the right formula.

Technical Summary

1. Problem Statement

In High Energy Physics (HEP), specifically in searches for Beyond the Standard Model (BSM) particles at the Large Hadron Collider (LHC), a primary strategy involves identifying localized "bumps" (resonances) in invariant mass spectra against a smooth, continuous background.

• Current Limitations: Traditional background modeling relies on fitting analytic functional forms (e.g., polynomials, exponentials) to sideband data. This approach faces significant challenges:
  • Model Bias: Choosing the wrong functional form can lead to "spurious signals" (fitting statistical fluctuations as new physics) or masking real signals.
  • Uncertainty Quantification: Deriving robust uncertainties for the choice of functional form is difficult. Methods like the "spurious signal test" or "discrete profiling" require large simulated templates or extensive testing of function families, which can be computationally expensive and statistically unstable with small datasets.
  • Low Statistics: Analytic forms often struggle when data statistics are low, requiring arbitrary constraints that may bias the result.
• Alternative Limitations: Gaussian Process Regression (GPR) offers a non-parametric alternative but requires binned data (losing information from unbinned events) and assumes Gaussian bin uncertainties, which introduces bias in low-statistics regimes (typically <10 events per bin).

2. Methodology: Log Gaussian Cox Process (LGCP)

The authors propose using Log Gaussian Cox Processes (LGCP) as a flexible, non-parametric framework for modeling smooth backgrounds directly on unbinned data.

• Core Assumptions:

  1. Non-homogeneous Poisson Process: The observed event samples $x$ are drawn from a Poisson process with an intensity function $\lambda(x)$.
  2. Log-Gaussian Intensity: The intensity function is defined as $\lambda(x) = N_E \cdot \exp(Z(x))$, where $N_E$ is the total expected number of events and $Z(x)$ is a Gaussian Process (GP).
  3. Gaussian Process Prior: $Z(x) \sim \mathcal{GP}(\mu(x), K(x, x'))$. The authors typically set the mean $\mu(x) = 0$ and use a Radial Basis Function (RBF) or Gibbs kernel for the covariance $K$.

• Inference via Markov Chain Monte Carlo (MCMC):
  Since the marginal likelihood is intractable to integrate analytically, the authors employ a two-stage MCMC approach:

  1. Hyperparameter Optimization: The Metropolis-Hastings algorithm is used to optimize the GP hyperparameters (length scale $\ell$ and variance $\sigma^2$) by maximizing the marginal likelihood estimated via Monte Carlo integration.
  2. Posterior Sampling: A second MCMC chain samples the function $Z(x)$ from the posterior distribution $p(Z(x)|X, \Theta)$. The final background estimate is the median of this chain, with uncertainty bands derived from the 16th and 84th percentiles.

• Signal+Background Extension:
  To detect signals, the intensity function is modified to include a known signal PDF $S(x)$:
  $$\lambda(x) = (N_E - N_S) \cdot \exp(Z(x)) + N_S \cdot S(x)$$
  Here, $N_S$ (signal yield) is treated as an additional hyperparameter optimized within the MCMC chain.

3. Key Contributions

• Unbinned Non-Parametric Modeling: Unlike GPR, LGCP operates directly on unbinned event data, preserving maximum information and avoiding biases associated with binning and Gaussian uncertainty assumptions in low-statistics regions.
• Robust Uncertainty Estimation: The Bayesian framework naturally provides posterior uncertainties for the background shape without relying on the "spurious signal" template method used in current ATLAS/CMS analyses.
• Comparative Framework: The paper provides a rigorous benchmark comparing LGCP against:
  • Standard unbinned Maximum Likelihood Estimators (MLE) with both "optimal" (true) and "estimated" (mismatched) functional forms.
  • Binned Gaussian Process Regression (GPR).

4. Results

The authors tested the methods on synthetic "toy" datasets generated from two complex background shapes ($F_1$: smooth falling; $F_2$: turn-on with falling tail) across three statistics regimes (100, 1,000, and 10,000 events).

• Background-Only Fits (Pull Plots):

  • Low Statistics (100 events): LGCP and GPR outperformed MLEs, which showed instability when the assumed functional form did not match the truth.
  • High Statistics (10,000 events): LGCP and GPR showed mild biases near the edges of the kinematic range (edge effects), whereas MLEs performed well in the center. However, LGCP uncertainties were sometimes under-estimated (pulls > 1$\sigma$).
  • Complex Shapes ($F_2$): LGCP and GPR significantly outperformed MLEs when the assumed analytic form was sub-optimal, demonstrating the value of non-parametric flexibility.

• Spurious Signal Tests:

  • GPR was the most resilient to statistical fluctuations, rarely identifying false signals (closest to 0 spurious signal).
  • LGCP showed some bias near edges and at "turn-on" features, occasionally misinterpreting these shapes as signals, though generally remaining within acceptable limits (<2% of total events).
  • MLE (with mismatched forms) produced large spurious signals.

• Signal Injection Tests:

  • Sensitivity: LGCP successfully captured injected Gaussian signals up to ~5% of the total event yield.
  • Performance:
    • LGCP: Performed well in capturing signal magnitude in the center of the spectrum but underestimated signal yields at high injection levels (>5%) and near edges.
    • GPR: Significantly underestimated injected signals, particularly in low-statistics regimes, often absorbing the signal into the background model.
    • MLE: Performed best when the functional form was known, but failed when the form was mismatched.

5. Significance and Conclusion

The paper demonstrates that LGCP is a viable and powerful alternative for background modeling in HEP, particularly for "bump-hunt" analyses.

• Advantages over GPR: LGCP handles unbinned data and low-statistics regimes better than GPR, which suffers from Gaussian binning assumptions.
• Advantages over Analytic Forms: LGCP removes the need to guess the correct functional form, reducing model bias and the associated systematic uncertainties.
• Practical Application: While LGCP exhibits some edge effects, the authors suggest these can be mitigated by fitting a wider sideband region and truncating the edges.
• Conclusion: The LGCP method is recommended for both background-only and signal-plus-background modeling. It offers a more automated, assumption-light approach that can improve the efficiency and accuracy of future LHC analyses, provided the sideband regions are sufficiently wide to avoid edge distortions.