A likelihood analysis for gamma-ray background models

This paper presents a likelihood-based comparison of locally constructed empirical and theoretically-motivated gamma-ray background models, finding that empirical descriptions provide statistically competitive fits to data on degree scales in high-latitude regions relevant for indirect dark matter searches.

The Gist

Imagine you are trying to hear a specific whisper in a very noisy, crowded room. That whisper is Dark Matter. The noise is everything else in the universe that makes light (gamma rays), like stars, gas clouds, and black holes.

Scientists want to find that whisper. But to do that, they first need to understand the noise. If they don't know exactly what the background noise sounds like, they might mistake a loud cough for the whisper they are looking for.

This paper is about three different ways to figure out what that "background noise" sounds like, so they can subtract it and find the Dark Matter.

The Three Competitors

The researchers tested three different "noise-canceling headphones" (models) to see which one worked best.

1. The "Look-Around" Method (Model E1)

• The Idea: If you want to know how loud the room is at your table, you look at the tables right next to you.
• How it works: They look at a patch of sky near the target (but not on the target). They count the gamma rays in different energy "buckets." They assume each bucket is separate from the others.
• The Analogy: Imagine you are counting apples in baskets. You count the red apples in Basket A, and the green apples in Basket B. You assume the number of red apples has nothing to do with the green apples.

2. The "Connected Baskets" Method (Model E2)

• The Idea: Sometimes, things in the room are connected. If the music gets louder, everyone talks louder.
• How it works: This method also looks at the sky nearby, but it realizes that the "buckets" of energy are actually talking to each other. If there are more high-energy rays, there might be more low-energy rays too.
• The Analogy: You realize that if Basket A is full of apples, Basket B is probably full too because they came from the same truck. You count them together as a team.

3. The "Weather Forecast" Method (Model FT)

• The Idea: Instead of looking at the room, you use a computer model to predict what the room should sound like based on physics.
• How it works: This uses complex physics theories and maps of the galaxy to calculate where the gamma rays should come from.
• The Analogy: Instead of listening to the room, you check the weather app. It tells you, "Based on the wind and pressure, the noise level should be 60 decibels."

The Big Test

To see which method was best, the scientists played a game of "Blind Taste Test."

They picked 100 random spots in the sky that were supposed to be empty of bright stars or galaxies (they called these "blank-sky" regions). They asked: Which model predicted the actual data best?

They used a scoring system that rewards accuracy but punishes you if your model is too complicated (like a "complexity tax"). If you need a million knobs to tune your model to fit the data, you get a penalty.

The Results

Here is what they found:

1. The "Look-Around" and "Connected Baskets" methods were tied.
   In most of the empty spots, the simple data-driven methods (E1 and E2) worked just as well as the complex physics model. It turns out, sometimes it's easier to just look at the data than to build a complex theory.

2. The "Weather Forecast" won when there was a loud neighbor.
   If there was a bright star or a known source close by, the theoretical model (FT) was better. Why? Because the theoretical model could specifically account for that bright star. The "Look-Around" method gets confused if a bright neighbor is leaking light into the area you are studying.

3. Simplicity wins in quiet neighborhoods.
   In the cleanest parts of the sky (far away from bright stars), the simple "Look-Around" method was preferred. The complex theoretical model tried to use too many "knobs" to fit the data, and the scoring system penalized it for being too complicated.

The Bottom Line

There is no single "best" way to model the background noise.

• If the sky is quiet and empty: A simple, data-driven look-around method works great and is less complicated.
• If there are bright, nearby sources: You need the complex physics model to account for them.

Why does this matter?
This helps scientists hunting for Dark Matter. By knowing which "noise-canceling headphone" to use for a specific galaxy, they can be more confident that if they hear a whisper, it's actually Dark Matter and not just a glitch in their background model. It’s about making sure you don't mistake a cough for a secret message.

Technical Summary

Here is a detailed technical summary of the paper "A likelihood analysis for gamma-ray background models" by Hoskinson, Kumar, and Sandick.

1. Problem Statement

Indirect searches for dark matter (DM) often target dwarf spheroidal galaxies (dSphs), which are DM-dominated and expected to produce minimal astrophysical gamma-ray background. However, detecting a potential DM signal requires accurately modeling the diffuse gamma-ray background arising from standard astrophysical processes along the line of sight. Current theoretical models (e.g., the Fermi-LAT diffuse model) carry systematic uncertainties of approximately 3% or higher in specific regions. Conversely, purely empirical models constructed from off-axis data avoid template assumptions but may lack physical motivation. The authors aim to determine which approach—empirical data-driven modeling or theoretical template-based modeling—provides a statistically superior description of gamma-ray data in "blank-sky" regions, thereby informing the systematic uncertainties inherent in indirect DM searches.

2. Methodology

The authors employ a likelihood-based framework to compare three distinct background modeling approaches across 100 randomly selected 1° sky regions of interest (ROIs) in two different sets (Set A and Set B).

• Models Evaluated:

  1. Model E1 (Independent Binning): An empirical model where photon counts in 16 logarithmically spaced energy bins (1–100 GeV) are treated as independent. Probability mass functions (PMFs) are derived from normalized histograms of off-axis sample regions.
  2. Model E2 (Covariance): An empirical model that accounts for cross-energy correlations. It approximates the joint photon count distribution as a multivariate Gaussian, utilizing a covariance matrix inferred from off-axis data.
  3. Model FT (Fermi Tools): A theoretically motivated model using the standard Fermi-LAT analysis pipeline (Fermitools/FermiPy). It includes Galactic diffuse emission, isotropic background, and catalog point sources. Parameters are optimized over a 20°×20° region, with specific normalizations freed for sources within 3° of the ROI center.

• Data Selection:

  • ROIs: 1° opening angle, centered away from the Galactic Center, Galactic plane, and known point sources.
  • Sets:
    • Set A: Less restrictive ($|b| \ge 15^\circ$, 1° separation from sources).
    • Set B: More restrictive ($|b| \ge 30^\circ$, 3° separation from sources).
  • Sample Regions: Off-axis regions used for empirical modeling are drawn within 10° of the ROI center, excluding overlaps with the ROI or point sources.

• Statistical Comparison:

  • Because the models are not nested and have different parameter counts (Empirical models have $k=0$; FT has free parameters), raw likelihoods cannot be directly compared.
  • Information Criteria: The authors use the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) to penalize model complexity.
  • Metric: Pairwise comparisons of $\Delta$BIC and $\Delta$AIC determine which model is preferred in each ROI.

3. Key Contributions

• Systematic Comparison: Provides a rigorous statistical comparison between purely empirical background descriptions and theoretical templates using information criteria to account for model complexity.
• Covariance Modeling: Introduces and tests a covariance-based empirical model (E2) to capture cross-energy correlations, comparing it against the standard independent-binning approach (E1).
• Quantification of Systematics: Quantifies the fraction of the sky where empirical models outperform theoretical models, offering insight into the systematic uncertainties of DM searches.
• Parameter Penalty Analysis: Demonstrates how the statistical penalty for floating parameters in theoretical models affects their performance relative to parameter-free empirical models.

4. Results

• Empirical vs. Empirical (E1 vs. E2):

  • There is no strong preference between the independent-binning (E1) and covariance (E2) models.
  • E1 is marginally preferred in roughly 2/3 of ROIs, but strong preference occurs in only ~18% (Set A) to 6% (Set B) of cases.
  • This suggests that while cross-energy correlations exist, the assumption of independence (E1) is often sufficient, or the Gaussian approximation in E2 introduces its own limitations.

• Empirical vs. Theoretical (E1/E2 vs. FT):

  • Set A (Closer to sources): The empirical model E1 provides a better description of the data than FT in ~80% of ROIs, with a strong preference in ~55%. This is largely because FT incurs a statistical penalty for floating parameters (normalizations of nearby sources) that does not sufficiently compensate for the likelihood improvement.
  • Set B (Cleaner regions): FT provides a better description in 60% of ROIs, but the preference is mostly weak (50%). Strong preference for FT occurs in only ~10% of ROIs.
  • AIC vs. BIC: Using AIC (which penalizes parameters less heavily) reduces the preference for empirical models in Set A, but E1 remains preferred in ~75% of ROIs where a strong preference exists.

• Outliers:

  • In specific outlier ROIs containing bright, nearby sources (e.g., blazar 3C 279), FT is strongly preferred. Explicitly modeling the source's spatial and spectral profile outweighs the statistical penalty.
  • In regions with extended spatial features, E2 and FT outperform E1, indicating that covariance or template modeling captures extended structures better than independent binning.

5. Significance

• Validation of Empirical Models: The study validates that locally constructed empirical background models are statistically competitive with, and often superior to, global theoretical templates for gamma-ray data on degree scales in high-latitude regions.
• DM Search Implications: For indirect DM searches in dSphs, empirical background models can mitigate systematic uncertainties associated with theoretical diffuse emission templates, provided nearby point sources are adequately masked.
• Model Complexity: The results highlight that the statistical penalty for additional parameters in theoretical models often outweighs the likelihood gains unless the sources are particularly bright or the region is very clean.
• Future Directions: The authors suggest that future empirical models could improve by using the covariance matrix as a kernel for kernel density estimation (to avoid Gaussian assumptions) or by dynamically adjusting masking radii based on source brightness and spectral type.