PoissonRatioUQ: An R package for band ratio uncertainty quantification

This paper introduces `PoissonRatioUQ`, an R package designed for Bayesian modeling and uncertainty quantification of count ratios by treating them as ratios of Poisson means, offering various methods for spatial and non-linear intensity ratio problems.

The Gist

The Problem: The "Counting" Trap

Imagine you are standing in a forest, trying to figure out the ratio of Red Birds to Blue Birds. You spend an hour counting and find 10 Red Birds and 5 Blue Birds. Your brain immediately says, "The ratio is 2 to 1!"

But there’s a catch. You didn't actually see "ratios"; you saw counts. If you had stayed for another hour and seen 20 Red and 10 Blue, the ratio is still 2 to 1, but your certainty about that number changes. If you only saw 1 Red and 0 Blue, saying the ratio is "infinity" is mathematically true but physically ridiculous.

In science—like looking at stars through a telescope or measuring gases in the atmosphere—we often deal with "counts" (photons hitting a sensor). We want to know the ratio of the actual physical intensities (the true underlying "vibe" of the atmosphere), not just the messy, jittery numbers we counted.

The Solution: The PoissonRatioUQ Package

This paper introduces a new tool (an R package) that helps scientists move from "What did I count?" to "What is actually happening in nature?"

Here is how the authors approach this using three big ideas:

1. The "Ghostly Map" (The Permanental Process)

Instead of treating every bird sighting as an isolated event, the authors assume that nature has a "texture." If you see a Red Bird in one tree, there’s a higher chance you’ll see one in the tree right next to it.

Think of it like a heat map of ghosts. Even when you aren't looking, there is a "latent" (hidden) intensity of birds moving through the forest. The package uses a mathematical model called a permanental process to create a smooth, intelligent map of these hidden intensities. It doesn't just look at the dots; it looks at the "flow" of the forest.

2. The "Smart Guess" (Bayesian Modeling)

The package uses Bayesian statistics. Imagine you are playing a game of "Guess the Weight."

• The Data: You see a box move slightly.
• The Prior: You know from experience that most boxes weigh between 5 and 10 lbs.
• The Posterior: You combine your observation with your experience to make a "smart guess."

The package does this for ratios. It takes the jittery, random Poisson counts and combines them with the "spatial texture" (the forest map) to produce a much more stable and realistic estimate of the true ratio.

3. The "Safety Margin" (Uncertainty Quantification)

In science, saying "The ratio is 2.0" is dangerous if you don't know how much you might be wrong.

The package provides Uncertainty Quantification (UQ). Instead of giving you a single number, it gives you a "cloud of possibility" (called an HPD interval). It’s like a weather report: instead of saying "It will be exactly 72 degrees," it says, "It will be between 70 and 74 degrees, and we are 95% sure of that." This allows scientists to know exactly how much they can trust their data.

Why does this matter?

This isn't just for birdwatchers. This math is used for:

• Space Exploration: Measuring the chemical makeup of Earth's atmosphere from satellites.
• Astronomy: Figuring out the composition of X-ray sources in deep space.
• Satellite Safety: Understanding atmospheric drag so satellites don't crash.

In short: PoissonRatioUQ is a high-tech lens that turns blurry, jumpy "counts" into clear, reliable "maps" of the physical world, complete with a built-in "honesty meter" to tell you how much to trust the results.

Technical Summary

Technical Summary: PoissonRatioUQ: An R Package for Band Ratio Uncertainty Quantification

1. Problem Statement

In fields such as remote sensing (e.g., atmospheric composition retrieval), X-ray astronomy, and thermospheric temperature sensing, researchers often rely on band ratios to infer physical parameters. A common pitfall in these applications is treating the ratio of observed counts as the physical quantity of interest.

The authors argue that the fundamental physical variable is the ratio of Poisson means (latent intensities), not the ratio of the discrete counts themselves. Because the Poisson mean is a latent random variable, standard frequentist approaches often fail to provide robust uncertainty quantification (UQ), especially when spatial correlation is present or when the relationship between the ratio and the physical parameter is nonlinear.

2. Methodology

The paper proposes a Bayesian hierarchical modeling framework implemented in the R package PoissonRatioUQ. The methodology is built on three pillars:

• The Permanental Process: To account for spatial information, the authors utilize a permanental process, a type of Poisson Point Process (PPP) where the intensity function $\lambda(s)$ is driven by a latent Gaussian Process (GP). This model is particularly effective because it naturally captures "point clustering," a common feature in geophysical and astronomical data.
• Bayesian Estimation via Laplace Approximation:
  • The model uses a quadratic link function to relate the latent GP to the Poisson intensities.
  • The authors employ the Representer Theorem to transform the problem into an optimization of a weight vector $\psi$.
  • Optimization is performed using a conjugate gradient method with an analytic gradient.
  • The posterior distribution of the latent intensities is approximated using a Laplace Approximation (a Gaussian distribution centered at the Maximum A Posteriori (MAP) estimate).
• Generalized Beta Prime (BP) Distribution:
  • The ratio of two Gamma-distributed variables (which approximate the Poisson intensities) follows a generalized Beta Prime distribution.
  • The package provides closed-form solutions for the PDF, CDF, and quantile functions of this distribution.
  • For nonlinear forward models of the form $Z = (mT + z_0)^p$, the package derives a shifted generalized Beta Prime distribution to provide the posterior for the actual quantity of interest ($T$).

3. Key Contributions

• PoissonRatioUQ R Package: A specialized tool for Bayesian modeling of count ratios.
• Spatial and Non-Spatial Modes: The package offers two estimation paths: a spatially aware method using the permanental process and a pointwise method (zbetaprime) that neglects spatial correlation but allows for varying numbers of realizations per location.
• Nonlinear Retrieval Capability: The ability to propagate uncertainty through nonlinear transformations (e.g., from intensity ratios to temperature) using the derived Beta Prime properties.
• Advanced UQ Tools: Implementation of functions for calculating Highest Posterior Density (HPD) sets (including for multimodal distributions) and the Continuous Rank Probability Score (CRPS) for model scoring.

4. Results and Demonstration

The authors validated the package through two "toy problem" simulations:

• Ratio Estimation: The model successfully recovered a complex trigonometric ratio function. With 50 bins, it achieved a mean CRPS of ~0.12 and a relative Mean Absolute Error (MAE) of ~7%. The computational efficiency was high, processing 1,000 bins in approximately 15 seconds.
• Nonlinear Transformation: When estimating a quantity $T$ through a nonlinear power-law relationship, the model accurately captured the posterior distribution. The authors noted that nonlinearity can amplify estimation errors, which was reflected in the widening of the HPD intervals near the domain boundaries.

5. Significance

This work provides a mathematically rigorous bridge between raw photon counts and geophysical/astronomical parameters. By shifting the focus from "ratios of counts" to "ratios of Poisson means," the package allows scientists to:

1. Incorporate spatial context via Gaussian Processes.
2. Quantify uncertainty more accurately through Bayesian posterior distributions rather than simple error bars.
3. Handle complex forward models that are standard in remote sensing and astronomy.

Future Directions: The authors aim to extend the package to handle unbinned data (using Nyström approximations) and to optimize performance for large-scale datasets by exploiting sparsity in the kernel matrices.