A response-matrix-centred approach to presenting cross-section measurements

This paper introduces a response-matrix-centred forward-folding approach as an alternative to traditional unfolding for presenting cross-section measurements, accompanied by the ReMU Python software framework to facilitate the direct comparison of model predictions with reconstructed data while robustly handling detector effects and systematic uncertainties.

The Gist

Imagine you are trying to figure out what a mysterious object looks like, but you can only see it through a very foggy, distorted window.

In the world of particle physics, scientists do this all the time. They smash particles together, but the "true" particles (the object) are invisible. All they see are the "reconstructed" particles (the blurry reflection) bouncing off the walls of their giant detectors.

The Old Way: "Unfolding" the Fog

Traditionally, physicists tried to mathematically "unfog" the window. They would take the blurry picture and try to reverse-engineer exactly what the object looked like.

• The Problem: This is like trying to un-bake a cake to get the raw eggs and flour back. It's mathematically messy. If there's even a tiny bit of noise in the blurry picture, the "unfogged" result can look completely wrong. It's an unstable process that often requires making a lot of guesses about what the cake should have tasted like (the physics model) to make the math work.

The New Way: The "Response Matrix" (The Translation Guide)

This paper proposes a smarter, more honest approach. Instead of trying to un-fog the window, the authors say: "Let's just describe the window perfectly, and let anyone else try to guess what's behind it."

They call this the Response-Matrix-Centred Approach.

Here is how it works, using a simple analogy:

1. The "Translation Dictionary" (The Response Matrix)

Imagine you have a dictionary that translates between two languages: Truth (what actually happened) and Reco (what the detector saw).

• Truth: "A particle with 5 units of energy."
• Reco: "The detector saw a blurry blob with 4.8 units of energy."

The Response Matrix is a giant spreadsheet (or a translation dictionary) that says: "If a particle has 5 units of energy, there is a 90% chance the detector will see 4.8, a 10% chance it will see 5.2, and a 0% chance it will see 10."

This dictionary is built using supercomputer simulations of the detector. Crucially, the authors argue this dictionary should be model-independent. It shouldn't care what kind of particle you are throwing at it; it only cares about how the detector reacts to it.

2. The "Forward-Folding" Game

Once you have this dictionary and the actual blurry data (the "Reco" numbers), you don't try to guess the Truth. Instead, you play a game of "Forward-Folding":

1. A theorist (someone who makes predictions) says, "I think the particles look like this."
2. You take their prediction and run it through your Response Matrix (the dictionary).
3. The matrix translates their "Truth" prediction into "Reco" predictions.
4. You compare the theorist's "Reco" prediction directly to the actual blurry data you measured.

If the prediction matches the blurry data, the theorist is right. If not, they are wrong.

3. Why is this better?

• No More Guessing: In the old way, if you wanted to test a new theory, you had to re-run the whole complex "unfogging" math. With this new way, you just plug the new theory into the dictionary. It's instant.
• Honesty: It admits that the detector is imperfect. It doesn't pretend to know the "True" answer perfectly; it just provides the tools for others to find the best answer.
• Future-Proof: If a new, better theory is invented ten years from now, scientists can just take the old data and the old dictionary and test the new theory immediately. They don't need to go back and re-collect data.

The "Software Toolkit" (ReMU)

The author also built a free software tool called ReMU (Response Matrix Utilities). Think of this as a "DIY kit" for physicists.

• Before, building this dictionary required expensive, secret software that only the big experimental teams had.
• Now, ReMU allows anyone with a laptop and Python to build these dictionaries, test theories, and share results without needing the secret "insider" tools.

Handling the "Background Noise"

Sometimes, the detector sees things that aren't the signal (like background radiation or other particles).

• Old way: You try to subtract the noise from your data. This is risky because if you subtract too much or too little, you ruin the data.
• New way: You give the "noise" its own column in the dictionary. You tell the computer, "Here is how the noise looks in Truth, and here is how it looks in Reco." The math then figures out how much of the blurry picture is signal and how much is noise, without ever deleting a single data point.

The Big Picture

This paper is essentially saying: "Stop trying to clean the window. Just give everyone a perfect manual on how the window distorts things, and let them figure out what's behind it."

It turns the difficult, unstable job of "unfolding" data into a simple, stable game of "translation," making science more open, faster, and more reliable for everyone.

Technical Summary

1. Problem Statement

The current canonical method for publishing cross-section measurements involves unfolding reconstructed data distributions to recover "true" event properties. This process attempts to mathematically reverse detector effects (efficiency and smearing). However, this approach faces significant challenges:

• Ill-posed Nature: Unfolding is an ill-posed inverse problem where small statistical fluctuations in reconstructed data can lead to large, unstable variations in the unfolded spectra.
• Model Dependence: Unfolding often requires assumptions about the underlying physics model to regularize the solution. If the model used for unfolding differs significantly from reality, the results can be biased or even become a "carbon copy" of the model assumptions.
• Reinterpretation Difficulty: Unfolded results are often difficult to reuse for future theoretical developments because they are tied to specific model assumptions or detector configurations used during the analysis.
• Statistical Limitations: Multi-dimensional differential measurements (to ensure model independence) require vast amounts of data to populate all bins, which is often unavailable.

2. Methodology: The Response-Matrix-Centred Approach

The paper proposes a forward-folding approach centered on the Response Matrix. Instead of unfolding data to truth space, this method projects theoretical model expectations from truth space into reconstructed (reco) space to compare directly with raw data.

Core Philosophy

1. Linearity: There is a linear relationship between true physics expectations (generator level) and measured event counts, described by a response matrix.
2. Imperfect Knowledge: The response matrix is known only with finite precision due to systematic and statistical uncertainties.
3. Data Primacy: The raw measured data is treated as exact counts; uncertainties arise only when interpreting these counts as expectation values for future experiments.

The Response Matrix ($R$)

• Definition: A matrix $R_{ij}$ where element $R_{ij}$ represents the probability that an event generated in truth bin $j$ is reconstructed in bin $i$.
• Construction: Built using Monte Carlo (MC) simulations. It encapsulates both selection efficiency (probability of an event being detected) and smearing (migration between bins).
• Model Independence: The matrix is constructed to be independent of the specific physics model used for testing. Different models predict different truth distributions ($\mu$), but the matrix $R$ remains constant.
• Handling Uncertainties:
  • Systematic Uncertainties: Instead of a single matrix, a set of "toy" matrices ($R^t$) is generated by varying detector parameters (e.g., momentum resolution, efficiency) according to their uncertainty distributions.
  • Statistical Uncertainties: The finite number of MC events introduces statistical noise. The paper proposes a Bayesian-inspired three-step process to model this:
    1. Efficiency: Modeled using a Beta distribution (conjugate prior for binomial processes).
    2. Smearing: Modeled using a Dirichlet distribution (conjugate prior for multinomial processes).
    3. Weight Corrections: Modeled using normal distributions for event weights.
  • Likelihood Calculation: The total likelihood is calculated by marginalizing over the set of toy matrices (or using a profile likelihood), integrating the systematic and statistical uncertainties directly into the statistical inference.

Handling Backgrounds

The paper outlines three strategies for background events without subtracting them from the data vector (which would violate Poissonian assumptions):

1. Irreducible Background: Treated as part of the signal in truth bins; purity is determined by the tested model.
2. "Physics-like" Background: Assigned its own truth bins and response matrix columns, allowing the background shape to evolve with new physics models.
3. Detector-specific Background: Encoded as fixed templates in the response matrix columns, with a single truth bin controlling the overall strength/weight.

Binning Strategy

• Truth Space: Must be binned finely enough to capture all variables affecting detector performance (to ensure model independence) but coarse enough to have sufficient MC statistics.
• Reco Space: Binned according to physics goals and detector resolution.
• Optimization: The paper suggests an algorithm to merge truth bins until statistical requirements are met while keeping intra-bin variations of the response matrix below a specific threshold.

3. Key Contributions

• ReMU Framework: Development of ReMU (Response Matrix Utilities), a Python package distributed via PyPI. It is experiment-agnostic, relying only on standard scientific libraries (NumPy, SciPy, PyMC).
  • Capabilities: Building response matrices from MC, forward-folding model predictions, and performing Bayesian or frequentist statistical inference.
  • Data Formats: Uses standard, human-readable formats (YAML for binning, CSV/NumPy for data) to ensure long-term accessibility and interoperability with frameworks like NUISANCE and Rivet.
• Statistical Formalism: A rigorous method for propagating both systematic and MC statistical uncertainties through the response matrix using toy matrices and Bayesian priors.
• Background Handling: A novel framework for including backgrounds directly in the likelihood calculation without data subtraction.
• Model Independence Tests: A procedure using the Mahalanobis distance to test if response matrices generated from different physics models are statistically compatible, ensuring the matrix is truly model-independent.

4. Results and Example Analysis

The paper demonstrates the approach using a mock experiment with two variables ($x$ and $y$), where $x$ is smeared and detection efficiency depends on $y$.

• Matrix Construction: The response matrix was successfully built and tested for compatibility between different input models (Model A and Model B).
• Statistical Testing: The framework was used to calculate log-likelihoods, maximize likelihoods over free parameters (normalization), and compute p-values (goodness-of-fit and likelihood ratios).
• Comparison with Unfolding: The example showed that the forward-folding approach could distinguish between models with different correlations in $y$ (which affect efficiency) more effectively than an unfolded approach. Unfolding would have required inflating systematic errors to account for model-dependent efficiencies, whereas the forward-folding approach naturally incorporates these differences via the response matrix.

5. Significance

• Future-Proofing: By publishing the raw data and the response matrix (rather than unfolded results), experimentalists enable theorists to test new models against old data without needing access to proprietary detector simulation software.
• Efficiency: It significantly reduces the time required for model testing, as the computationally expensive detector simulation is replaced by a simple matrix multiplication.
• Robustness: It avoids the instabilities associated with unfolding and provides a more powerful statistical test by comparing models in the space where the data actually exists (reco space).
• Global Fits: The approach facilitates the inclusion of experimental results in global fits (e.g., for neutrino oscillation parameters) by providing a standardized, model-independent interface for data interpretation.

In conclusion, the response-matrix-centred approach offers a robust, flexible, and statistically sound alternative to traditional unfolding, bridging the gap between experimental data and theoretical model development.