Binned and Unbinned Transverse Single Spin Asymmetry Extraction, including Background Subtraction and Unfolding

This paper presents general methods for extracting transverse single-spin asymmetries using both binned and unbinned maximum likelihood approaches, specifically designed to address experimental challenges such as time-varying polarization, unequal spin-state luminosities, and the need for background subtraction and kinematic unfolding.

The Gist

Imagine you are a detective trying to solve a mystery: Why do some particles spin one way and others spin another?

In the world of subatomic physics, scientists shoot beams of particles at targets. Sometimes, they use "polarized" beams, meaning all the particles are spinning in the same direction (like a crowd of people all marching in step). When these spinning particles collide, they don't just bounce off randomly; they tend to scatter in a specific pattern based on their spin. This pattern is called the Transverse Single Spin Asymmetry.

The goal of this paper is to teach scientists how to measure this pattern accurately, even when the experiment is messy, the equipment is imperfect, and there are "fake" clues (background noise) hiding the real answer.

Here is the breakdown of their solution, using some everyday analogies:

1. The Problem: A Noisy, Shaky Camera

Imagine you are trying to take a photo of a spinning top to see which way it leans. But you have three big problems:

• The Light Flickers: Sometimes the light is bright, sometimes dim (this is like the luminosity changing).
• The Top is Unbalanced: Sometimes the top spins fast, sometimes slow, and sometimes the "up" spin is different from the "down" spin (this is polarization imbalance).
• The Lens is Blurry: Your camera lens distorts the image, making the top look like it's leaning in the wrong direction (this is detector smearing).
• There's a Ghost in the Room: There are other objects in the photo that look exactly like the top but aren't (this is background noise).

If you just take a simple average, your photo will be wrong. You need a smarter way to process the data.

2. The Two Main Tools: The Bucket Method vs. The Individual Method

The authors propose two ways to solve this puzzle.

Method A: The "Bucket" Approach (Binned Analysis)

Imagine you sort all your photos into buckets based on the angle the top was facing (0°, 10°, 20°, etc.).

• The Trick: You count how many tops are in each bucket for "Spin Up" and "Spin Down."
• The Cleanup: Since you know what the "Ghost" (background) looks like, you can estimate how many ghosts are in each bucket and subtract them.
• The Result: You get a clean count for every angle, and you can calculate the asymmetry.
• The Catch: If your camera is very blurry, the tops might fall into the wrong buckets. You have to use a mathematical "unfolding" trick to guess which bucket they really belonged to.

Method B: The "Individual" Approach (Unbinned Likelihood)

Instead of buckets, imagine looking at every single photo individually.

• The Trick: You don't just count; you calculate a "score" for every single event. Did this specific particle lean left or right? How strong was the light when it was taken?
• The Weights: If the light was dimmer for "Spin Up" photos, you give those photos extra weight (like giving a louder voice to a quiet speaker) so they count just as much as the bright "Spin Down" photos.
• The Cleanup: For the "Ghosts" (background), you give them a negative score. It's like saying, "This photo looks like a ghost, so it subtracts from the total."
• The Result: This method is more precise because it doesn't lose information by forcing data into buckets. It uses a mathematical formula (Maximum Likelihood) to find the perfect answer that fits all the individual clues.

3. The "Ghost" Hunter (Background Subtraction)

Sometimes, the "Ghosts" (background events) have their own spin patterns. If you just subtract them blindly, you might accidentally remove part of the real signal.

• The Solution: The authors use "Sidebands." Imagine the real signal is a mountain peak. The "Ghosts" are the hills on the sides. By measuring the hills, they can mathematically predict exactly how many ghosts are hiding under the mountain peak and subtract them perfectly.

4. The "Unfolding" Magic (Fixing the Blurry Lens)

If the camera is very blurry (strong smearing), the particles land in the wrong places.

• The Analogy: Imagine you are trying to guess the original shape of a clay sculpture, but someone has squished it through a sieve. The pieces are scattered.
• The Solution (OmniFold): The authors use a smart computer algorithm (like a neural network) that plays a game of "Guess and Check."
  1. It guesses what the original sculpture looked like.
  2. It simulates squishing it through the sieve to see what the result would look like.
  3. It compares the simulation to the actual messy photo.
  4. It adjusts the guess and repeats until the simulation matches the messy photo perfectly.
  5. The final "guess" is the true, unfolded shape of the particle's behavior.

5. The Verdict: Does It Work?

The authors tested their methods with millions of simulated "fake" experiments.

• They tried scenarios where the light flickered wildly.
• They tried scenarios where the "Ghosts" had their own spin.
• They tried scenarios where the camera was extremely blurry.

The Result: Both the "Bucket" method and the "Individual" method worked perfectly. They found the correct answer every time, even when the experiment was a mess. The "Individual" method (Unbinned) was slightly better at handling the messy, blurry data, but both are now reliable tools for physicists.

Summary

This paper is a user manual for cleaning up messy physics data. It tells scientists:

1. Don't just count things; weigh them based on how bright the light was and how strong the spin was.
2. If there is background noise, subtract it using the "sideband" trick.
3. If your detector is blurry, use a smart computer algorithm to "un-squish" the data.
4. Whether you sort data into buckets or look at every single event, you can now get the true answer without being fooled by the experiment's imperfections.

Technical Summary

1. Problem Statement

In nuclear and particle physics experiments involving polarized beams and/or targets, researchers aim to determine the transverse single-spin asymmetry ($A_N$). This asymmetry arises from spin-momentum correlations and manifests as an azimuthal modulation in the scattering angular distribution.

The extraction of $A_N$ faces several significant experimental challenges:

• Variable Polarization: The magnitude of polarization ($P$) often varies over time or differs between spin states (spin-up vs. spin-down).
• Luminosity Imbalances: Integrated luminosities ($L$) for different spin states or target materials may not be equal.
• Background Contamination: Signal events are often accompanied by background events with similar kinematic signatures but different asymmetries.
• Detector Smearing: Imperfect reconstruction of kinematic variables (like the azimuthal angle $\phi$) leads to "bin migration" or smearing, necessitating unfolding to recover true physics distributions.

Existing methods often struggle to simultaneously handle these factors without introducing bias or requiring restrictive assumptions (e.g., constant polarization or equal luminosities).

2. Methodology

The authors propose a comprehensive framework for extracting $A_N$ using both binned and unbinned statistical approaches, incorporating background subtraction and unfolding.

A. Binned Analysis (Without Unfolding)

• Yield Modeling: The total yield is modeled as the sum of foreground (signal) and background yields, each dependent on polarization, luminosity, and the asymmetry.
• Background Subtraction: A "sideband" method is used. Background yields are estimated from kinematic bins adjacent to the signal peak.
• Correction for Imbalances: The method derives a general formula for the experimental asymmetry that explicitly accounts for:
  • Different luminosities ($L^+$ vs. $L^-$).
  • Different polarization magnitudes ($P^+$ vs. $P^-$).
  • The ratio of background to foreground yields ($f_B$).
• Approximation: Under the assumption that polarization differences are small, the foreground/background ratio is approximated using the measured sideband asymmetry, allowing for the isolation of the foreground asymmetry $A_F$.

B. Unbinned Maximum Likelihood Analysis

• Likelihood Function: Instead of binning data, the authors maximize the log-likelihood function based on the probability distribution of individual events: $P_i \propto [1 + P_i A_N \sin(\phi_{pol, i} - \phi_i)]$.
• Experimental Weights: To correct for luminosity and polarization imbalances, the authors introduce specific experimental weights ($w^{exp}$) for each event. These weights are inversely proportional to the product of luminosity and polarization ($L \cdot P$) for that spin state.
  • Formula: $w^{\pm} = \frac{L^+ P^+ + L^- P^-}{2 L^{\pm} P^{\pm}}$.
• Background Subtraction: Background events are handled by assigning them negative weights in the log-likelihood sum.
• Small Asymmetry Approximation: For small $A_N$, the log-likelihood can be expanded, yielding a deterministic formula for $\hat{A}_N$ that avoids iterative maximization, though full maximization is also supported.

C. Unfolding Techniques

To address detector smearing, the paper explores three unfolding strategies:

1. Binned Unfolding: Standard techniques (e.g., Iterative Bayesian) applied to binned data, followed by asymmetry extraction.
2. Unbinned Unfolding + Binned Analysis: Unfolding the unbinned data, then binning and extracting asymmetry.
3. Unbinned Unfolding + Unbinned Likelihood: Using OmniFold (an iterative reweighting algorithm) to unfold unbinned data directly, followed by the unbinned likelihood extraction.
   • Likelihood Ratio Trick: The authors utilize a binary classifier (Neural Network) to estimate the likelihood ratio between source (truth) and target (reconstructed) distributions, effectively generating event weights to map distributions.

3. Key Contributions

• Generalized Weighting Scheme: Derivation of robust experimental weights that correct for simultaneous imbalances in luminosity and polarization magnitude in an unbinned maximum likelihood framework.
• Unified Background Treatment: A consistent method for background subtraction in both binned (sideband subtraction) and unbinned (negative weighting) analyses.
• Unfolding Integration: Demonstration of how to integrate unfolding (specifically OmniFold) with asymmetry extraction, comparing binned vs. unbinned approaches.
• Validation Framework: Development of a rigorous testing suite using Monte Carlo simulations with varying complexity (including cosine-dependent efficiencies, which mimic the physics signal).

4. Results

The methods were validated using 1000+ simulated datasets with varying conditions (polarization imbalance, luminosity imbalance, background asymmetry, and detector efficiency models).

• Accuracy: Both binned and unbinned methods successfully extracted the injected asymmetry ($A_N = 0.2$) within one standard deviation across all test cases.
• Robustness to Imbalance: The weighted unbinned method showed no bias even when polarization and luminosity were significantly imbalanced (e.g., $P^+=0.9, P^-=0.7$ and $L^+:L^- = 3:7$).
• Efficiency Independence: The method correctly recovered $A_N$ even when detector efficiency contained a $\cos \phi$ term (which mimics the physics signal), provided spin-flipping was employed and weights were applied. Without spin-flipping, the method failed in the presence of such efficiency biases.
• Unfolding Performance:
  • Weak Smearing: All three unfolding methods (binned, unbinned-to-binned, unbinned-to-unbinned) performed similarly.
  • Strong Smearing: The unbinned unfolding followed by unbinned likelihood analysis (Method 3) showed slightly better stability and lower systematic bias compared to binned approaches, particularly in complex scenarios with efficiency modulations.
  • Convergence: The unfolding algorithm converged in 4 iterations for weak smearing and 8 iterations for strong smearing.

5. Significance

This work provides a comprehensive, bias-free toolkit for analyzing transverse single-spin asymmetries in modern high-energy physics experiments. Its significance lies in:

1. Handling Real-World Data: It addresses the non-ideal conditions (fluctuating polarization, unequal luminosities) common in real experiments, which often force simplifications in standard analyses.
2. Statistical Efficiency: The unbinned maximum likelihood approach, combined with proper weighting, maximizes the statistical power of the data, which is crucial for rare processes or limited datasets.
3. Systematic Control: By explicitly modeling background subtraction and detector smearing (unfolding) within the likelihood framework, the method reduces systematic uncertainties associated with binning and reconstruction artifacts.
4. Versatility: The framework is applicable to a wide range of experiments involving polarized beams/targets, including those studying di-leptons or other processes with combinatoric backgrounds.

In conclusion, the authors demonstrate that with the correct weighting and unfolding strategies, precise extraction of spin asymmetries is possible even under challenging experimental conditions where traditional binned methods might fail or introduce significant bias.