Monte Carlo Methods

The traditional approach to evaluating measurement uncertainty, as defined in the core Guide to the Expression of Uncertainty in Measurement (GUM), relies on the Law of Propagation of Uncertainty (LPU). This analytical method uses a first-order Taylor series approximation to propagate summary statistics (the mean and variance) of the input quantities to the final result.

However, the LPU approach has limitations. To address these, the Joint Committee for Guides in Metrology (JCGM) published Supplement 1 to the GUM (JCGM 101). This document formalizes the use of Monte Carlo Methods (MCM) for the propagation of distributions, representing a paradigm shift in advanced uncertainty analysis.

Limitations of the Traditional GUM Framework

The standard LPU framework can break down or become highly inaccurate under specific mathematical conditions:

• Strong Non-Linearity: If the measurement model $Y = f(X_1, X_2, ...)$ is highly non-linear, a first-order Taylor series approximation is insufficient. Higher-order terms become significant, making the analytical calculation extremely complex.
• Asymmetric Distributions: The traditional GUM framework implicitly assumes the resulting distribution of the measurand can be characterized by a Gaussian (Normal) or scaled Student's t-distribution to determine expanded uncertainty. If the input quantities have heavily asymmetric distributions, or if the model itself induces asymmetry, the resulting coverage intervals calculated using a standard $k=2$ factor may not accurately represent a 95% confidence level.
• Large Uncertainties: If the magnitude of the uncertainties of the input quantities is very large relative to their expected values, the first-order approximation becomes invalid.

How the Monte Carlo Method Works

The implementation of MCM for uncertainty analysis typically follows these steps:

1. Define the Measurement Model: Establish the mathematical equation relating the output quantity $Y$ to the input quantities $X_i$.
2. Assign Probability Density Functions (PDFs): Based on available knowledge (Type A and Type B evaluations), assign a specific PDF to every input quantity $X_i$.
3. Generate Random Samples: Using a computerized pseudo-random number generator, generate a random sample value for each input quantity based strictly on its assigned PDF. This creates one discrete "state" of the system.
4. Evaluate the Model: Calculate the output value $y$ using the measurement model and the randomly generated input values.
5. Iterate: Repeat steps 3 and 4 a very large number of times (denoted as $M$). $M$ is typically $10^5$ to $10^6$ trials, depending on the required numerical tolerance.
6. Analyze the Results: Sort the resulting $M$ output values into an array. This array forms a discrete representation of the PDF of the output quantity $Y$.

Extracting Expanded Uncertainty from MCM

One of the greatest advantages of the Monte Carlo Method is how elegantly it handles expanded uncertainty and coverage intervals.

In the traditional framework, you must calculate effective degrees of freedom ($\nu_{eff}$) using the Welch-Satterthwaite equation and look up a coverage factor $k$ from a t-table.

With MCM, this entire process is bypassed. Because you have generated a massive, sorted list of resulting $y$ values, determining the 95% coverage interval simply involves identifying the values at the 2.5th percentile and the 97.5th percentile of your sorted array. The space between these two values represents the shortest 95% coverage interval, regardless of whether the resulting distribution is normal, skewed, or completely asymmetrical.