PHAS0083 Chap.5 Covariance, Correlation and Error Propagation
Covariance, Correlation and Error Propagation
Covariance, Correlation and Error Propagation measures how strongly variables depend on each other and uses that measure to carry uncertainties through calculations. Covariance is the expected product of deviations from the means, positive when the deviations tend to share a sign; the shortcut E(XY) minus the product of means makes it quick to compute.
Correlation divides by the two standard deviations so that it lies between minus one and one and reports linear association only. Independent variables are uncorrelated, but uncorrelated variables can be dependent, except in the bivariate normal case. The variance of a linear combination then includes a covariance term.
Error propagation is introduced as nothing more than the spread of a derived quantity computed from uncertain inputs, computed exactly by a change of variables, numerically by Monte Carlo, or approximately by the first-order delta method, which uses the gradient at the means and the covariance matrix.
The chapter shows where the delta method is reliable, quantifies its bias for a curved function, and warns that it can return a finite number when the true variance does not exist.
What this chapter covers
- 01
Covariance and its shortcut formula
- 02
Correlation and linear association
- 03
Uncorrelated versus independent
- 04
The bivariate normal
- 05
Variance of linear combinations
- 06
Delta method with a covariance matrix
- 07
Limits of first-order propagation
Worked example · free
Propagate two independent lengths into an area
- 1Value: A = ab = 10.00 square centimetres. This allocation is practice weighting, not a UCL marking scheme.
- 2For a product, relative variances add: (0.03/4.00)^2 + (0.02/2.50)^2 = 5.6e-5 + 6.4e-5 = 1.20e-4.
- 1Relative standard deviation 0.0110, so the area is 10.00 +/- 0.11 square centimetres.
Key terms
- Covariance
- The expected product of two variables' deviations from their means.
- Delta method
- A first-order Taylor approximation to the variance of a function of random inputs.
- Covariance matrix
- The matrix of variances on the diagonal and covariances off it for a random vector.
- Bivariate normal
- The two-variable normal distribution set by two means, two variances and one correlation.
- Linear association
- A tendency for one variable to change in proportion to another, which correlation measures.
Covariance, Correlation and Error Propagation FAQ
Should I propagate errors analytically or by simulation?
Use the delta method when the function is close to linear over the spread of the inputs, because it is quick and needs only means and covariances. Use an exact change of variables or a Monte Carlo simulation when the function curves strongly, when a denominator can approach zero, or whenever the result will be quoted with high precision.
Does zero correlation mean two variables are independent?
Not in general. Correlation detects only linear association, so variables can be strongly dependent with zero correlation, for example when one is a symmetric function of the other. For the bivariate normal, however, zero correlation does imply independence.
When is the delta method exact?
It is exact when the function is linear in its inputs, as for sums, differences and scaled values. For curved functions it is an approximation whose bias grows with the spread of the inputs relative to the curvature, so compare it with an exact or simulated result.
Why does error propagation fail for some ratios?
The ratio of two independent normal variables has no finite variance because the denominator can be close to zero. The delta method uses only means and covariances, so it still returns a finite number that looks reasonable but describes nothing real.
Exam move
Start with covariance on small discrete tables, computing it both from the definition and from the shortcut, and convert to correlation so the scale-free meaning becomes familiar. Produce one example of uncorrelated but dependent variables yourself. Then learn propagation as a recipe: write the function, evaluate the gradient at the means, assemble the full covariance matrix, and form the quadratic product.
Practise the relative form for products and powers and the absolute form for sums and differences until you can choose without thinking. For at least two curved functions, compare the delta-method answer with an exact or simulated one and note the size of the bias. Keep a short list of cases that need simulation, such as ratios with a denominator near zero and functions that curve strongly over the input spread.
Working through Covariance, Correlation and Error Propagation in PHAS0083? Sia is AskSia’s AI Statistics tutor — ask any PHAS0083 Covariance, Correlation and Error Propagation question and get a clear, step-by-step explanation grounded in how PHAS0083 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.