3. Probability theory foundations for structural reliability

Structural reliability analysis rests on a single, deceptively simple question: given that loads, material strengths, and geometry are all uncertain, what is the probability that a structure will perform its intended function? Answering that question rigorously requires a working command of random variables, probability distributions, statistical moments, and joint probability. This chapter builds that foundation using structural engineering objects—concrete compressive strength, wind pressure, yield stress, dead load—rather than abstract dice and coins, so that every concept maps directly onto quantities you will later feed into a limit state function.

Random Variables and Their Structural Meaning

A random variable (RV) is a function that assigns a real number to each outcome of an uncertain experiment. In structural reliability, the "experiment" is the act of observing a real structure or its environment. Two categories matter:

• Discrete random variables take countable values. Example: the number of overloaded trucks crossing a bridge span in one day, or the number of corroded tendons detected in a post-tensioned girder.
• Continuous random variables take any value in an interval. Example: concrete cylinder compressive strength f_c (MPa), yield stress f_y (MPa), annual maximum wind pressure q (kN/m²), or dead load intensity g (kN/m).

Each continuous RV is fully described by its cumulative distribution function (CDF), F_X(x) = P(X ≤ x), and, when it exists, its probability density function (PDF), f_X(x) = dF_X/dx. The PDF is not a probability; it is a density. Probability is obtained by integration: P(a < X ≤ b) = ∫_a^b f_X(x) dx. This distinction is the source of many early errors—students often read a PDF value as a probability, which is dimensionally wrong.

Distributions Used in Structural Reliability

Four distributions cover the majority of structural variables. Choosing the right one is a modeling decision with real consequences for the computed reliability index.

The lognormal distribution is especially important because it is strictly positive and its logarithm is normal. If f_c is lognormal with log-mean λ = ln(μ) − ½ζ² and log-std ζ, then the median is exp(λ) and the mean is exp(λ + ζ²/2). For a concrete batch with mean 35 MPa and coefficient of variation (COV) 0.15, ζ = √(ln(1 + 0.15²)) ≈ 0.149, and λ = ln(35) − ½(0.149²) ≈ 3.544. The 5th percentile—often used as a characteristic strength—is exp(λ − 1.645ζ) ≈ exp(3.544 − 0.245) ≈ 27.0 MPa. This is the kind of calculation that converts a design code's characteristic value into a probabilistic input.

Moments: Mean, Variance, and Beyond

The mean μ_X = E[X] = ∫ x f_X(x) dx is the first moment and represents the centroid of the distribution. The variance σ_X² = E[(X − μ_X)²] is the second central moment; its square root σ_X is the standard deviation. The coefficient of variation COV = σ_X/μ_X is dimensionless and is the standard way structural engineers express uncertainty: dead load COV ≈ 0.10, concrete strength COV ≈ 0.15–0.20, wind load COV ≈ 0.30–0.40, resistance model uncertainty COV ≈ 0.05–0.15.

Higher moments matter when distributions are skewed. The skewness γ₁ = E[(X − μ)³]/σ³ measures asymmetry; lognormal and Gumbel variables are positively skewed, which affects the tails where failure probabilities live. The kurtosis γ₂ = E[(X − μ)⁴]/σ⁴ − 3 measures tail heaviness. For reliability work, the tails are the whole point: a distribution that matches the mean and variance but misrepresents the tail can shift the reliability index β by 0.3 or more, which translates to an order-of-magnitude change in failure probability.

Joint Probability and Dependence

Real structures involve multiple random variables acting together. The joint CDF is F_X,Y(x,y) = P(X ≤ x, Y ≤ y), and the joint PDF f_X,Y(x,y) describes their simultaneous density. If X and Y are independent, f_X,Y(x,y) = f_X(x) f_Y(y). Independence is a strong assumption that is often violated: concrete strength and modulus are positively correlated; dead load and live load are usually treated as independent but may share a common model error.

Dependence is quantified by the covariance Cov(X,Y) = E[(X − μ_X)(Y − μ_Y)] and the correlation coefficient ρ_XY = Cov(X,Y)/(σ_Xσ_Y), with −1 ≤ ρ ≤ 1. In structural reliability, ρ is typically small (0.1–0.5) but not negligible. For a reinforced concrete beam, the correlation between concrete strength and steel yield stress is often taken as ρ ≈ 0.2–0.3 because both are influenced by common production quality. Ignoring positive correlation between resistance variables makes the structure appear safer than it is.

The conditional distribution f_Y|X(y|x) = f_X,Y(x,y)/f_X(x) is the workhorse for updating. If a proof load test is performed and the measured deflection is lower than predicted, the conditional distribution of resistance given that observation shifts upward—this is the basis of Bayesian updating in Chapter 12.

From Joint Probability to the Reliability Integral

With a resistance R and a load effect S, the failure event is {R − S ≤ 0}. The failure probability is the integral of the joint density over the failure domain:

P_f = P(R ≤ S) = ∫∫_{r ≤ s} f_{R,S}(r, s) dr ds

If R and S are independent:
P_f = ∫_{0}^{∞} F_R(s) f_S(s) ds

For the simple case R ~ N(μ_R, σ_R) and S ~ N(μ_S, σ_S), independent, the safety margin Z = R − S is normal with μ_Z = μ_R − μ_S and σ_Z = √(σ_R² + σ_S²). The reliability index is β = μ_Z/σ_Z, and P_f = Φ(−β).

Worked example (illustrative, not a measured result): A steel tension member has resistance R with μ_R = 300 kN, σ_R = 30 kN (COV 0.10), and load effect S with μ_S = 180 kN, σ_S = 36 kN (COV 0.20), both normal and independent. Then μ_Z = 120 kN, σ_Z = √(900 + 1296) = √2196 ≈ 46.9 kN, and β = 120/46.9 ≈ 2.56. The corresponding P_f = Φ(−2.56) ≈ 5.2 × 10⁻³. If the target β for this consequence class is 3.8 (P_f ≈ 7 × 10⁻⁵), the member is under-designed and the resistance mean must increase or the load COV must be reduced.

The figure shows the joint density of R and S as concentric ellipses centered at (μ_R, μ_S). The failure domain is the triangular region below the limit state line R = S. The design point—the point on the limit state surface closest to the origin in standard normal space—is where the joint density contributes most to the failure probability. The distance from the mean to the design point, measured in units of σ_Z, is the reliability index β. This geometric interpretation is the foundation of the First Order Reliability Method (FORM) developed in Chapter 6.

What You Should Be Able to Do After This Chapter

• Classify a structural variable as discrete or continuous and select an appropriate distribution family based on its physical support and tail behavior.
• Convert between mean/COV and distribution parameters for normal and lognormal variables, and compute characteristic percentiles.
• Compute μ_Z, σ_Z, β, and P_f for a linear safety margin with independent normal variables.
• Explain why correlation between resistance variables reduces the apparent reliability and how ρ enters the variance of a linear combination: σ_Z² = σ_R² + σ_S² − 2ρσ_Rσ_S.
• Recognize when the normal approximation is inappropriate and an extreme value or lognormal model is required.

These foundations—random variables, distributions, moments, and joint probability—are not preliminary mathematics to be skimmed. They are the vocabulary in which every subsequent reliability method is written. The next chapter introduces the limit state function formally and shows how the simple R − S margin generalizes to nonlinear, multi-variable performance functions.