Biases in the Determination of Correlations Between Underground Muon Flux and Atmospheric Temperature

This paper demonstrates that while both unbinned and binned analysis methods are unbiased under ideal conditions, the binned method suffers significant bias from temperature uncertainties, prompting the authors to propose a novel stability-assessment procedure to ensure robust correlation estimates in underground muon flux studies.

The Gist

The Big Picture: A Cosmic Weather Report

Imagine you are standing in a deep underground cave (a physics detector). You are counting "cosmic muons"—tiny particles raining down from space. You notice something interesting: the number of muons hitting your detector changes with the seasons.

• Summer: The air up high is hot and puffy (less dense). Cosmic particles don't crash into the air as much; they survive longer and turn into muons. More muons reach the cave.
• Winter: The air is cold and dense. Particles crash and die before they can turn into muons. Fewer muons reach the cave.

Scientists want to measure exactly how much the muon count changes for every degree the temperature changes. This "connection strength" is called the correlation coefficient.

The Problem: Two Ways to Draw the Line

To find this connection, scientists have a list of data points: "On Day 1, Temp was X, Muons were Y." They want to draw a straight line through these dots to see the trend.

The paper compares two ways of doing this:

1. The "Unbinned" Method (The Individual Approach): You take every single day's data point and draw a line through all of them at once.
2. The "Binned" Method (The Grouping Approach): You group days together. For example, you take all days where the temperature was between 20°C and 21°C, calculate the average muon count for that group, and then draw a line through those averages.

The Twist: The "Fuzzy Glasses" Effect

The paper discovers a major problem with the Binned Method (Grouping).

Imagine you are trying to measure the relationship between a person's height and their shoe size.

• The Ideal World: You have perfect measurements. If you group people by shoe size, the average height for each group is perfect. Both methods work fine.
• The Real World: Your ruler is a little fuzzy. You can't measure the shoe size perfectly; sometimes you think a size 10 is a size 10.5.

Here is where the Binned Method breaks:
When you group people by their measured shoe size, you accidentally mix up people.

• A person who is actually Size 9.5 (but you measured as 10) gets put in the "Size 10" group.
• A person who is actually Size 10.5 (but you measured as 10) also gets put in the "Size 10" group.

Because the "fuzziness" (measurement error) is random, the group ends up with a weird mix of people. The paper shows that this mixing creates a fake "S-shape" curve. Instead of a straight line, the data bends. The result? The Binned Method consistently underestimates the connection. It tells you the temperature and muons are less related than they actually are.

The Unbinned Method is smarter. It looks at every single day individually. Even if the temperature measurement is a little fuzzy, the math (called "Weighted Total Least Squares") knows how to handle that fuzziness and still finds the true straight line.

The Catch: You Need to Know How Fuzzy Your Glasses Are

There is a catch for the Unbinned Method. To do the math correctly, you have to tell the computer: "My temperature measurements are fuzzy by exactly 0.4 degrees."

• If you guess right: The Unbinned Method gives the perfect answer.
• If you guess wrong:
  • If you say the error is smaller than it really is, the result is too low.
  • If you say the error is bigger than it really is, the result is too high.

The problem is, in real life, it is very hard to know exactly how "fuzzy" the temperature measurements are.

The Solution: The "Stability Test"

So, how do we fix this if we don't know the exact error? The authors propose a clever trick called Temporal Aggregation (merging data over time).

Imagine you are trying to guess the average height of a crowd, but your ruler is shaky.

• Day 1: You measure one person. Your guess is shaky.
• Day 7: You measure 7 people and take the average. The shakiness cancels out. The average is much more stable.

The paper suggests:

1. Take your daily data.
2. Combine 7 days into one "week" data point. Then combine 30 days into one "month" data point.
3. Run the Unbinned Method on these new, smoother data points.

The Magic Rule:

• If you guessed the error correctly, the result (the correlation line) will stay exactly the same whether you look at daily, weekly, or monthly data.
• If you guessed the error wrong, the result will drift as you change the time periods.

By adjusting your "fuzziness" guess until the result stops drifting, you find the correct error value and get the true correlation.

The Takeaway

1. Stop grouping data (Binned Method) when measuring muons and temperature. It creates a fake curve and gives you the wrong answer.
2. Use individual data (Unbinned Method), but you must be careful about how you estimate measurement errors.
3. Use the "Stability Test": If you aren't sure about your error estimates, merge your data into weeks or months. If your answer stays steady no matter how you group the time, you've found the right answer.

This paper essentially gives scientists a new, more reliable ruler to measure the cosmic connection between the weather above and the particles below.

Technical Summary

1. Problem Statement

Underground cosmic-ray muon detectors observe a seasonal modulation in muon rates that is positively correlated with fluctuations in the effective atmospheric temperature ($T_{eff}$). This correlation is typically quantified by a coefficient $\alpha$ in the linear relationship $\Delta R = \alpha \Delta T_{eff}$, where $\Delta R$ and $\Delta T_{eff}$ are relative variations in muon rate and temperature, respectively.

Two primary methodologies are used to determine $\alpha$:

1. Unbinned Method: Performing a linear regression (specifically Weighted Total Least Squares, WTLS) on individual daily (or hourly) data points.
2. Binned Method: Grouping data points into discrete bins based on $T_{eff}$, averaging the values within each bin, and performing regression on these binned averages.

The Core Issue: Previous studies (e.g., MINOS) have observed that the Binned Method systematically yields lower correlation coefficients than the Unbinned Method. The origin of this discrepancy was unclear. Furthermore, accurately quantifying the uncertainty of $T_{eff}$ is difficult due to model dependencies and varying atmospheric data quality. If the assigned uncertainties in the regression do not match the true measurement errors, the resulting correlation coefficient may be biased.

2. Methodology

The authors conducted a systematic investigation using Toy Monte Carlo (toyMC) simulations to isolate and analyze the sources of bias.

• Simulation Setup:

  • Generated synthetic datasets with a strict linear relationship between $\Delta R$ and $\Delta T_{eff}$ over a 3000-day period.
  • Both variables followed periodic (cosine) functions with a 365-day period, mimicking seasonal variations.
  • Independent random Gaussian errors were added to simulate measurement uncertainties for both muon rates and $T_{eff}$.
  • The "true" correlation coefficient was set to $\alpha_{true} \approx 0.359$.

• Analysis Techniques:

  • Unbinned Method: Applied WTLS directly to the daily data points.
  • Binned Method: Binned data by $\Delta T_{eff}$ (e.g., 0.2% intervals), calculated mean values per bin, and applied WTLS to the binned data.
  • Uncertainty Mismatch: The study varied the assigned uncertainty ($\sigma_{assigned}$) used in the fitting algorithm while keeping the true uncertainty ($\sigma_{true}$) in the simulation fixed, to test the sensitivity of the methods to error estimation errors.

• Mitigation Strategy:

  • Investigated temporal aggregation: Merging data over $n$ consecutive days (weekly, monthly) to reduce statistical noise and uncertainty mismatch ($\Delta \sigma \propto 1/\sqrt{n}$).
  • Proposed a stability test: Varying the assigned uncertainty and the aggregation window to find the point where the inferred $\alpha$ becomes stable (insensitive to binning size).

3. Key Contributions

1. Identification of Binning-Induced Bias: The paper demonstrates that the Binned Method introduces a systematic bias (underestimation of $\alpha$) whenever the predictor variable ($T_{eff}$) has non-zero measurement uncertainty. This is caused by "regression dilution" where x-axis smearing creates an artificial "S-shaped" distortion in the binned averages.
2. Validation of the Unbinned Method: The Unbinned Method is shown to be unbiased if and only if the assigned uncertainties accurately reflect the true measurement errors.
3. Sensitivity to Error Estimation: The study quantifies how the Unbinned Method becomes biased if $\sigma_{assigned} \neq \sigma_{true}$. Underestimating errors leads to an underestimated $\alpha$, while overestimating errors leads to an overestimated $\alpha$.
4. Practical Mitigation Framework: The authors propose a robust procedure to handle unknown or imperfectly estimated uncertainties:
   • Merge data over increasing time intervals ($n$ days).
   • Scan different values of assigned uncertainty.
   • Identify the specific uncertainty assignment that renders the correlation coefficient $\alpha$ stable (constant) regardless of the number of merged days ($n$).

4. Key Results

• Binned Method Failure: When $\sigma_{T} > 0$, the Binned Method consistently underestimates $\alpha$. The bias grows monotonically with the magnitude of the temperature uncertainty. For example, with $\sigma_T = 0.4$ K, the Binned Method yielded $\alpha \approx 0.327$ compared to the true $0.359$.
• Unbinned Method Robustness: When $\sigma_{assigned} = \sigma_{true}$, the Unbinned Method recovers $\alpha_{true}$ perfectly, regardless of the magnitude of the uncertainty.
• The "S-Shape" Distortion: The paper elucidates the geometric origin of the Binned Method's failure. Binning by the observed (noisy) temperature mixes points with different true values. At the extremes of the distribution, this preferentially mixes in points closer to the mean, creating a non-linear, S-shaped trend that flattens the fitted slope.
• Stability Criterion:
  • If the assigned uncertainty is incorrect, the inferred $\alpha$ will drift systematically as the data aggregation window ($n$) increases.
  • If the assigned uncertainty is correct, $\alpha$ remains stable as $n$ increases.
  • The authors demonstrated that even with a constant assigned uncertainty (ignoring day-to-day variations), finding the value that stabilizes $\alpha$ across aggregation windows yields an unbiased result.

5. Significance

• Resolution of Discrepancies: The paper resolves the long-standing tension between analysis methodologies in muon flux studies (e.g., the MINOS results), attributing the difference to the inherent bias of the Binned Method in the presence of measurement errors.
• Recommendation for Future Studies: The authors recommend the Unbinned Method as the standard approach for muon seasonal modulation studies.
• Framework for Uncertainty Quantification: Since $T_{eff}$ uncertainties are notoriously difficult to model precisely, the proposed stability-based mitigation strategy provides a practical, data-driven way to calibrate error models. Researchers can now determine the correct uncertainty assignment by observing the stability of the correlation coefficient across different time binning scales, rather than relying solely on theoretical error propagation.
• General Applicability: While focused on muon flux, the findings regarding errors-in-variables regression and the dangers of binning noisy predictor variables are broadly applicable to other fields in physics and statistics where similar correlation analyses are performed.