Experimental Uncertainty & Error Propagation

No physical measurement is ever exact. From the earliest astronomical observations of Tycho Brahe to the nanoscale metrology of modern semiconductor fabrication, experimenters have wrestled with the same fundamental question: how much can we trust our numbers? The formal study of experimental uncertainty grew out of centuries of effort to reconcile conflicting observations in astronomy, geodesy, and physics. Understanding how uncertainties arise—and how they combine when measurements feed into calculations—is not a pedantic bookkeeping exercise; it is the backbone of scientific credibility. Without a rigorous framework for error propagation, we cannot meaningfully compare theory with experiment, replicate results, or design instruments that push the frontiers of precision.

The central question that this lesson addresses is deceptively simple: if the inputs to a calculation each carry some uncertainty, what is the uncertainty of the output? Answering that question requires us first to classify the kinds of error that afflict measurements, then to develop a mathematical machinery—rooted in calculus—for tracking how uncertainties flow through formulas. By the end, you will be equipped to report results that honestly reflect the limits of your apparatus and methodology.

Before diving into formulas, it is essential to build a precise vocabulary. Physicists distinguish between the true value of a measurand (the quantity being measured), the measured value obtained in practice, and the uncertainty that quantifies how far the measured value might plausibly lie from the true value. Crucially, uncertainty is not a mistake; it is an intrinsic feature of every measurement process.

The classic target analogy for accuracy and precision is well-known, but a more quantitatively useful visualization is the Gaussian distribution of repeated measurements. When random errors dominate, repeated measurements of the same quantity cluster around the mean in a bell curve whose width is characterized by the standard deviation σ. The following diagram shows how the distribution relates to the concepts of accuracy, precision, and the standard deviation of the mean.

In this diagram, the width of the distribution reflects precision: a narrow bell curve means tightly clustered readings (high precision), while a broad curve indicates large scatter. The horizontal offset between the dashed cyan line (x̄) and the dashed amber line (true value) represents systematic bias—the measure of inaccuracy. Importantly, taking more data points narrows the standard deviation of the mean (σ_x̄ = σ/√N), but it cannot reduce systematic bias—only careful calibration or redesign can do that.

The mathematical heart of this topic lies in answering the question: given a derived quantity q = f(x, y, z, …) where each input variable carries an uncertainty, what is δq? The general answer comes from a first-order Taylor expansion of f about the measured values, assuming that the uncertainties are small relative to the values themselves. We begin with the standard deviation of a single set of measurements, build up to the standard deviation of the mean, and then derive the general propagation formula.

Common Special Cases

The general formula simplifies to memorable rules for the two most common function forms. For sums and differences (q = x ± y), the absolute uncertainties add in quadrature: δq = √(δx² + δy²). For products and quotients (q = xy or q = x/y), the relative (fractional) uncertainties add in quadrature: δq/|q| = √( (δx/x)² + (δy/y)² ). For a quantity raised to a power (q = xⁿ), the relative uncertainty scales linearly with the exponent: δq/|q| = |n| × (δx/x). These rules are not separate axioms—they are all derivable from the general propagation formula via straightforward partial differentiation.

To solidify your understanding, the diagram below maps out the decision tree you should follow when propagating uncertainties through any formula. The key insight is to decompose complex expressions into a chain of elementary operations (addition, multiplication, powers, functions) and apply the appropriate rule at each stage, combining intermediate uncertainties in quadrature.

Suppose you measure the mass and dimensions of a small aluminum cylinder to determine its density. You find: mass m = 45.2 ± 0.3 g, diameter d = 2.10 ± 0.02 cm, and height h = 3.90 ± 0.03 cm. What is the density and its propagated uncertainty?

The quadrature-based propagation framework is remarkably powerful, but it rests on assumptions that you should be aware of. Knowing when the standard approach works well—and when it breaks down—is a mark of experimental sophistication.

The propagation rules you have learned constitute a first-order, analytic approach that works beautifully for most introductory physics scenarios. However, modern metrology and experimental research often face situations where these rules are insufficient. Three advanced frameworks extend the basic approach: Monte Carlo simulation, Bayesian inference, and the full GUM framework with correlation. Understanding where the basic approach fits in the landscape of uncertainty analysis will prepare you for upper-division lab courses and research.

In a Monte Carlo uncertainty analysis, you generate thousands (or millions) of synthetic data sets by randomly sampling each input variable from its probability distribution, compute the output for each set, and then examine the resulting distribution of outputs. The standard deviation of that output distribution gives δq without any Taylor-expansion approximation. Bayesian methods go further by incorporating prior knowledge about parameters and producing full posterior probability distributions. These approaches are increasingly accessible through computational tools like Python's uncertainties package or dedicated Monte Carlo libraries. As you advance in your physics career, these techniques will become indispensable, but the analytic propagation rules remain the essential foundation upon which they are built.