Bayesian and Monte Carlo approaches to estimating uncertainty for the measurement of the bound-state $β$ decay of $^{205}\mathrm{Tl}^{81+}$

This paper demonstrates that Bayesian and Monte Carlo methods provide comparable and robust estimates for the uncertainty in contaminant $^{205}\mathrm{Pb}^{81+}$ levels during the measurement of bound-state $\beta$ decay of $^{205}\mathrm{Tl}^{81+}$, recommending their adoption for future experiments facing similar statistical fluctuations.

The Gist

The Big Picture: A Cosmic Stopwatch

Imagine you are trying to build a cosmic stopwatch to measure how old our Solar System is. To do this, scientists use a specific element called Lead-205. But to know exactly how fast this "clock" ticks, they first need to measure how fast a different element, Thallium-205, decays into it.

The scientists in this paper successfully measured this decay rate. However, the real story of this paper isn't just the measurement itself; it's about how they handled the "noise" and "mistakes" in their data to make sure the result was trustworthy.

Think of it like trying to hear a whisper in a very loud, chaotic room. The scientists found the whisper, but they had to invent two new, sophisticated ways to prove that the background noise wasn't actually the whisper.

---

The Experiment: The "Ghost" in the Machine

The team fired a beam of heavy ions (atoms stripped of all their electrons) into a storage ring, which is like a giant, high-speed racetrack for atoms. They wanted to watch Thallium-205 atoms turn into Lead-205 atoms.

The Problem:
The racetrack wasn't perfectly clean. Along with the Thallium they wanted, a tiny bit of "ghost" Lead-205 was already there from the start (contamination).

• The Analogy: Imagine you are trying to count how many new blue marbles fall into a bucket every minute. But, the bucket already has a few blue marbles in it that you didn't put there, and you don't know exactly how many. If you just count the total at the end, you'll think more marbles fell in than actually did.

The scientists needed to subtract this "ghost" count to get the true answer. But here's the kicker: the amount of "ghost" Lead fluctuated slightly every time they ran the experiment, and they couldn't measure that fluctuation directly. It was a hidden variable.

The Challenge: The "Missing Uncertainty"

When they analyzed their data, they found something weird. The data points were scattered much more than their math predicted.

• The Analogy: Imagine you are throwing darts at a board. You expect them to land in a tight circle around the bullseye. Instead, they are scattered all over the board. You know your throwing technique (the physics), and you know your hand tremors (the known errors), but there is still a lot of extra scattering.
• The Question: Where is this extra scattering coming from? It turns out, it was likely due to tiny, invisible fluctuations in the magnetic fields of the machine that created the beam.

Because they couldn't measure this directly, they had to estimate it from the scatter of the data itself. This is where the paper gets interesting. They compared two different mathematical "detectives" to solve this mystery.

---

Detective #1: The Monte Carlo Method (The "Simulation Squad")

This method is like running a thousand different movie simulations of the experiment.

1. How it works: The computer takes the real data and adds random "noise" to it, simulating what might have happened if the magnetic fields fluctuated slightly differently. It does this one million times.
2. The Magic: In every single simulation, it asks, "If the noise was this big, does the math still work?" It builds a giant histogram of all possible answers.
3. The Result: It found that the "ghost" contamination varied by about 6%. By adding this "missing uncertainty" to their error bars, the scattered data points finally made sense.
4. The Catch: This method is great at handling complex correlations, but it's a bit rigid. If there was one really bad data point (an "outlier"), the whole simulation could get skewed, so the scientists had to manually throw that bad point out before running the simulations.

Detective #2: The Bayesian Approach (The "Flexible Judge")

This method is like a wise judge who listens to every piece of evidence but is also very skeptical of extreme outliers.

1. How it works: Instead of running simulations, it uses a mathematical framework that asks, "Given what we know, how likely is this result?" It treats the "missing uncertainty" not as a fixed number, but as something that can vary for each individual data point.
2. The Magic: It uses a "conservative" approach. If a data point looks weird (like a dart thrown way off the board), the Bayesian method says, "Okay, maybe the error bar for this specific point was just underestimated," rather than throwing the whole point away. It naturally inflates the uncertainty for that one point without ruining the rest of the data.
3. The Result: It handled the "bad" data point automatically, without the scientists needing to manually delete it. It arrived at almost the exact same answer as the Monte Carlo method.

---

The Showdown: Who Won?

The scientists compared the two detectives:

• Monte Carlo: Powerful and flexible, but requires you to manually clean up the data first. It's like a super-accurate scale that needs you to remove the dust before weighing.
• Bayesian: Smarter at handling "bad" data automatically. It's like a scale that can tell the difference between a heavy rock and a speck of dust and adjust itself.

The Verdict: Both methods agreed on the final answer! The decay rate of Thallium-205 is roughly 2.76 × 10⁻⁸ per second.

Why Does This Matter?

1. For the Solar System: This confirms how fast the "cosmic stopwatch" (Lead-205) ticks, helping us date the formation of our Solar System more accurately.
2. For Physics: It proves that when experiments get messy and data is scattered, you don't have to guess. You can use these advanced statistical tools to find the truth.
3. For the Future: The authors recommend using the Bayesian method (specifically a tool called NESTED FIT) for future experiments because it handles "bad data" automatically and is very robust.

The Takeaway

Science is often about dealing with the unknown. When you can't measure the "noise" directly, you have to use clever math to estimate it. This paper shows that whether you simulate a million scenarios (Monte Carlo) or use a flexible, skeptical judge (Bayesian), you can arrive at the same reliable truth, even when the data is messy. It's a victory for mathematical detective work.

Technical Summary

1. Problem Statement

The paper addresses the challenge of accurately estimating uncertainties in the measurement of the bound-state $\beta^-$ decay ($\beta_b$) of fully stripped $^{205}\text{Tl}^{81+}$ ions. This measurement is critical for:

• Refining the half-life of $^{205}\text{Tl}$ and $^{205}\text{Pb}$, which serves as a cosmochronometer for early Solar System studies.
• Supporting the LOREX project's low-energy measurement of the solar neutrino spectrum.

Core Issues:

• Contaminant Variation: The experiment relies on a secondary beam of $^{205}\text{Tl}^{81+}$ produced via projectile fragmentation. A significant contaminant, $^{205}\text{Pb}^{81+}$ (the decay daughter), is produced in the same reaction. Due to the small mass difference ($\sim 31$ keV), they cannot be fully separated, leading to an initial contamination level of $\sim 0.1\%$.
• "Missing Uncertainty": Standard error propagation based on detector noise and correction factors ("raw uncertainties") failed to explain the scatter in the experimental data. The $\chi^2$ of the fit was 303 (for 14 degrees of freedom), far exceeding the expected range, indicating unaccounted systematic variations, likely due to fluctuations in the Fragment Separator (FRS) magnetic fields affecting the contaminant ratio.
• Outliers: The dataset contained an outlier ($6.7\sigma$ from the fit) that, if included, biased the decay rate and doubled the uncertainty. Traditional $\chi^2$ minimization is highly sensitive to such outliers.
• Correlated Errors: Several correction factors (e.g., resonance, charge-changing cross-section) were applied globally to all data points, creating strong correlations that standard fitting methods struggle to handle.

2. Methodology

The authors compared two advanced statistical frameworks to handle missing uncertainty, correlations, and outliers:

A. Monte Carlo (MC) Approach

This method extends a Frequentist framework to handle complex error propagation:

1. Simulation: The experiment is simulated $10^6$ times.
2. Sampling:
   • Measured quantities (Schottky intensities, corrections) are sampled from their distributions.
   • Correlated Parameters: Global corrections are sampled once per run and applied to all 16 data points to preserve correlations.
   • Missing Uncertainty ($\sigma_{CV}$): Instead of using a fixed scaling factor (like the Birge ratio), the method maps the $\chi^2$ distribution to a contamination variation value ($\sigma_{CV}$). For each run, a random $\chi^2$ value is drawn, and the corresponding $\sigma_{CV}$ is added in quadrature to the data. This accounts for the uncertainty in estimating the missing uncertainty itself.
   • Poisson Statistics: Explicitly modeled to account for counting statistics of the ions, which contribute $\sim 3\%$ to the variance.
3. Fitting: A least-squares fit is performed on each simulated dataset to extract the decay rate ($\lambda_{\beta b}$).
4. Outlier Handling: The outlier was manually identified and excluded prior to the final MC analysis to prevent bias.

B. Fully Bayesian Approach

This approach uses the NESTED FIT code with nested sampling to explore the parameter space:

1. Likelihood Construction:
   • Standard Gaussian: Assumes data follows a normal distribution.
   • Conservative (Sivia) Approach: Treats the reported error bars as lower bounds. It integrates over unknown systematic uncertainties using a modified Jeffreys prior. This results in a likelihood function with "heavy tails" ($1/(y-y_i)^2$), making the fit naturally robust against outliers without manual exclusion.
2. Parameter Priors: Nuisance parameters (loss rates) are included with Gaussian priors based on experimental uncertainties.
3. Data Handling: The Bayesian method was tested with and without the outlier and with/without Poisson statistics inclusion.

3. Key Contributions

• Novel MC Strategy for Missing Uncertainty: The authors developed a method to self-consistently estimate the "missing uncertainty" by sampling the $\chi^2$ distribution rather than applying a global inflation factor (Birge ratio). This avoids the bias of assuming a specific target $\chi^2$ and respects the statistical constraints of small datasets ($N=16$).
• Explicit Poisson Modeling: The paper clarifies the contribution of Poisson counting statistics to the total uncertainty, distinguishing it from the FRS-induced contamination variation.
• Robust Bayesian Application: Demonstrated that the "conservative" Bayesian approach (Sivia method) can handle outliers and missing systematics simultaneously without manual data culling, providing a more automated and robust analysis pipeline.
• Comparative Analysis: Provided a rigorous side-by-side comparison of Frequentist (MC) and Bayesian methods for a complex nuclear physics experiment, validating that both yield consistent results despite fundamentally different assumptions.

4. Results

• Decay Rate ($\lambda_{\beta b}$):
  • MC Method (with outlier excluded): $\lambda_{\beta b} = 2.76(28) \times 10^{-8} \text{ s}^{-1}$.
  • Bayesian Method (Conservative, all data including outlier): $\lambda_{\beta b} = 2.62(32) \times 10^{-8} \text{ s}^{-1}$.
  • Consistency: The results agree within $0.5\sigma$. The inclusion of Poisson statistics in the MC analysis shifted the result by only $0.7\%$, confirming the robustness of the previous measurements.
• Uncertainty Sources: The dominant source of uncertainty was identified as the contamination variation from the FRS ($\sim 46\%$ of total variance), followed by Poisson statistics ($\sim 28\%$) and raw correction uncertainties ($\sim 26\%$).
• Outlier Impact:
  • In the MC method, the outlier had to be manually removed to obtain a reliable fit.
  • In the Bayesian conservative method, the outlier was naturally down-weighted by the heavy-tailed likelihood, yielding a consistent result without manual exclusion.

5. Significance

• Reliability of Cosmochronometers: The study confirms the half-life of $^{205}\text{Tl}$ with high precision and a rigorously validated uncertainty budget, strengthening the use of $^{205}\text{Pb}$ as a clock for early Solar System evolution.
• Methodological Advancement: The paper advocates for moving beyond the traditional Birge ratio for estimating missing uncertainties in small datasets. It demonstrates that:
  • Monte Carlo is superior when data correlations are complex and manual outlier identification is feasible.
  • Bayesian (Conservative) methods are superior for handling outliers and unknown systematics "out-of-the-box," reducing the need for subjective data selection.
• Future Applications: The authors recommend the open-source NESTED FIT code for future experiments dealing with unknown statistical fluctuations, while also endorsing the MC approach for scenarios with strong data correlations where Bayesian computation might be costly.

In conclusion, the paper successfully resolves the discrepancy between raw data scatter and theoretical uncertainties in a high-precision nuclear physics experiment, providing a blueprint for robust uncertainty quantification in the presence of complex systematics and outliers.