Does second moment mean second-order Taylor terms?

No. It refers to variance information while the usual response approximation is first order.

Are input distributions fully required?

The moment calculation uses means and covariances, but additional distributional assumptions are needed for many probability claims.

Is the approximate output mean always exact?

No. Nonlinear curvature can make it differ substantially from the actual mean.

AI Glossary / Statistics and probability / Advanced

First-order second-moment method

Definition

The first-order second-moment method, often abbreviated FOSM, approximates the mean and variance of a model output using a first-order Taylor expansion and the input means, variances and covariances. In its common mean-value form, the expansion is made around the vector of input means. It is an uncertainty-propagation method, useful when a response model is expensive or a local approximation is adequate. The phrase second moment does not mean that a second-order Taylor expansion has been used.

What it means

A model output changes when uncertain inputs change. FOSM summarizes that propagation through local sensitivities instead of repeatedly evaluating every possible input combination. It can provide a quick approximation to output variability and help identify influential uncertainty contributions. However, means and variances do not determine a complete distribution, and a local linear approximation can miss curvature or asymmetric tails. An estimated mean and variance therefore do not automatically justify a particular failure probability, confidence level or normal-distribution assumption.

How it works

Let Y = f(X), with input mean vector mu and covariance matrix Sigma. Evaluate the response and its gradient at mu. The usual first-order approximations are E(Y) approximately f(mu) and Var(Y) approximately gradient(f)' Sigma gradient(f). For independent inputs, the variance expression becomes the sum of squared partial derivatives multiplied by input variances. With dependent inputs, covariance cross-terms must be included. Obtain derivatives analytically, through automatic differentiation or carefully checked numerical differences. Assess whether the response is sufficiently smooth and nearly linear across the relevant input range. This mean-centered calculation is distinct from FORM's boundary-centered reliability analysis.

Three practical examples

Measurement modeling: approximate how uncertainties in several measured inputs affect a derived quantity.
Simulation screening: identify which input variations contribute most to a response's local variance.
Model review: compare a quick moment approximation with analytical results or simulation before relying on it.

Think of it this way

Think of estimating how a surface rises or falls near one point using its local slopes. Small movements can be translated into approximate height changes without exploring the whole surface. If the surface bends sharply or the movements are large, those slopes can give a misleading picture of the resulting variation.

Worked example

Suppose independent inputs X and Z have means ten and five, variances one and 0.25, and output Y = XZ. The mean-value estimate is ten x five = 50. The derivatives at the means are dY/dX = five and dY/dZ = ten, so the FOSM variance is 5 squared x 1 + 10 squared x 0.25 = 50. Independence makes the exact mean 50, but the exact product variance is 25 x 1 + 100 x 0.25 + 1 x 0.25 = 50.25. The approximation misses the product-of-variances term. A more severe limitation occurs for Y = X squared with standard normal X: the derivative at mean zero is zero, so first-order FOSM predicts mean and variance zero, while the exact values are one and two. These analytical examples show both a modest error and a complete local-linearization failure.

Illustrative case study

Fictional case: Acacia Measurement Services uses FOSM to assess a derived production indicator. Analysts check units and covariance assumptions, then compare the approximation with simulated inputs over the observed uncertainty range. One nonlinear component has a near-zero derivative at the mean but substantial variation away from it, so they replace that local calculation with a method that represents curvature. This invented case illustrates checking approximation adequacy rather than assuming that a small derivative always means an input is unimportant.

Common misconceptions and mistakes

Do not drop covariance terms merely because inputs are listed in separate columns. Correlation can increase or decrease output variance depending on the signs of the sensitivities. Do not assume f(E[X]) equals E[f(X)] for nonlinear functions. Zero first derivative does not imply zero uncertainty. Numerical gradients can be unreliable when the step size is poorly chosen or the model is discontinuous. A variance approximation alone does not specify tail probabilities. FOSM and first-order reliability methods share local-approximation ideas but generally use different expansion points and answer different questions.

Business use

Businesses can use FOSM for transparent preliminary uncertainty budgets and computational screening. Record the response equation, input moments, covariance assumptions and derivative calculation. Check units: each variance contribution should have the squared units of the output. Compare the approximation with independent calculations where nonlinearities could influence decisions. Use sensitivity results to guide data collection cautiously, because local rankings can change across operating conditions. The method's simplicity is valuable when its limitations are understood and when the analysis does not demand unsupported precision in rare-event tails.

Check your understanding

Why is the product example's exact variance 50.25 rather than 50?

Answer: Independence yields an additional Var(X)Var(Z) term, equal to 0.25, which the first-order linearization omits.

Related terms

Taylor expansion
Covariance matrix
First-order reliability method
Monte Carlo simulation
uncertainty propagation
delta method
sensitivity coefficient

Learn next

derive the covariance cross-terms in a two-input uncertainty calculation, then compare first-order and higher-order approximations for a response with pronounced curvature.

Sources

Sources: NASA, uncertainty-propagation paper, First-Order Second Moment Method section: https://ntrs.nasa.gov/api/citations/20130011505/downloads/20130011505.pdf ; NASA, Real-Time Monitoring and Prediction of Airspace Safety Margins, analytical uncertainty methods: https://ntrs.nasa.gov/api/citations/20180006637/downloads/20180006637.pdf ; NIST, law of propagation of uncertainty: https://physics.nist.gov/cuu/Uncertainty/combination.html .

Previous: First-order reliability method Next: Fisher exact test