4. Structural reliability fundamentals

Structural reliability analysis answers a deceptively simple question: given that loads, material strengths, and geometry are all uncertain, what is the probability that a structure will fail to perform its intended function? To answer this quantitatively, we need three ingredients — a mathematical definition of "failure," a probabilistic description of the uncertain variables, and a scalar measure that summarizes how far the design sits from the failure boundary. These three ingredients are the limit state function, the failure probability, and the reliability index.

Limit State Functions

A limit state is any condition at which the structure ceases to satisfy a design requirement. The most common is the ultimate limit state, expressed as a resistance-minus-load form:

g(X) = R(X) − S(X)

Here R is resistance (e.g., flexural capacity, buckling strength), S is the load effect (e.g., bending moment, axial force), and X = (X₁, X₂, …, Xₙ) is the vector of random variables. The sign convention is critical and must be fixed before any analysis: g > 0 denotes the safe domain, g = 0 is the limit state surface, and g < 0 is the failure domain. Every subsequent calculation, sensitivity study, and design check in this course uses this same convention.

For a simply supported steel beam under a uniformly distributed load q and a concentrated mid-span load P, a serviceability limit state on mid-span deflection could be written as:

g = δ_allow − δ_max
  = L/250 − (5qL⁴)/(384EI) − (PL³)/(48EI)

where L, E, and I may themselves be random. Note that the limit state function need not be linear in the basic variables; the deflection expression above is nonlinear in L and I, which has direct consequences for the solution methods discussed later.

Failure Probability

The failure probability is the probability mass of the failure domain:

P_f = P[g(X) ≤ 0] = ∫_{g(X)≤0} f_X(x) dx

where f_X(x) is the joint probability density of the random variables. For more than two or three variables, this integral is analytically intractable, and even numerical quadrature becomes impractical. This is why reliability theory relies on transformations and approximations rather than direct integration.

A useful special case is the linear limit state with independent normal variables. If g = a₀ + a₁X₁ + … + aₙXₙ and all Xᵢ ~ N(μᵢ, σᵢ), then g is itself normal with:

μ_g = a₀ + Σ aᵢμᵢ
σ_g = sqrt( Σ aᵢ²σᵢ² )
P_f = Φ(−μ_g / σ_g)

where Φ is the standard normal CDF. This closed-form result is the foundation on which the reliability index is built.

The Reliability Index β

Cornell's reliability index is defined as the number of standard deviations separating the mean of the safety margin from the limit state surface:

β = μ_g / σ_g

For the linear-normal case, P_f = Φ(−β) exactly. The mapping is highly nonlinear at the tails: β = 2.0 gives P_f ≈ 2.3 × 10⁻², β = 3.0 gives ≈ 1.3 × 10⁻³, β = 3.8 gives ≈ 7.2 × 10⁻⁵, and β = 4.5 gives ≈ 3.4 × 10⁻⁶. Target β values in codes reflect this — for instance, a one-year target of β ≈ 4.2 and a fifty-year target of β ≈ 3.8 are typical for ultimate limit states in structural codes, corresponding to different reference periods.

Cornell's β is exact only when g is linear and all variables are normal. When g is nonlinear or variables are non-normal, β loses its invariance — different but mechanically equivalent formulations of the same limit state can yield different β values. The Hasofer–Lind index β_HL resolves this by transforming all variables to independent standard normals U and defining β as the minimum distance from the origin to the limit state surface in U-space:

β_HL = min ‖u‖  subject to  g(u) = 0

The point u* achieving this minimum is the design point (or most probable failure point), and it identifies the combination of variables most likely to cause failure. This is the geometric interpretation that makes the reliability index both a probability measure and a sensitivity indicator.

From β to P_f: The First-Order Approximation

The First-Order Reliability Method (FORM) linearizes g at the design point and approximates P_f ≈ Φ(−β_HL). The algorithm proceeds iteratively:

1. Transform all basic variables X to independent standard normals U (e.g., via the Nataf or Rosenblatt transformation).
2. Start from an initial guess u₀, often the origin or the mean-value point.
3. Evaluate g(uₖ) and its gradient ∇g(uₖ).
4. Update using the HL-RF iteration: uₖ₊₁ = [uₖᵀαₖ + g(uₖ)/‖∇g(uₖ)‖] αₖ, where αₖ = −∇g(uₖ)/‖∇g(uₖ)‖.
5. Repeat until ‖uₖ₊₁ − uₖ‖ < ε (typically 10⁻⁴).

For the deflection limit state above with L ~ N(6000, 100) mm, E ~ N(210000, 10500) MPa, I ~ N(8.5×10⁷, 4.25×10⁶) mm⁴, q ~ N(12, 2.4) N/mm, and P ~ N(20, 4) kN, a FORM analysis converges to β ≈ 3.1, giving P_f ≈ 9.7 × 10⁻⁴. The design point typically shows that I and q dominate the failure contribution, with sensitivity factors α_I ≈ −0.62 and α_q ≈ 0.55. These α values tell the engineer where to focus quality control or where to add conservatism.

Second-Order Effects and Simulation

When the limit state surface is strongly curved, FORM's linearization can be inaccurate. The Second-Order Reliability Method (SORM) adds a curvature correction, typically using Breitung's formula:

P_f ≈ Φ(−β) · Π (1 + βκᵢ)^(−1/2)

where κᵢ are the principal curvatures at the design point. For most structural problems with moderate nonlinearity, the FORM–SORM difference is small, but for problems involving buckling or plastic mechanisms, SORM can change P_f by an order of magnitude.

As an independent verification, Monte Carlo simulation with N = 10⁶ samples gives a coefficient of variation on P_f of approximately 1/√(N·P_f). For P_f ≈ 10⁻³, this is about 3.2%, which is adequate for validating FORM results. Importance sampling centered at the design point reduces the required sample size by one to two orders of magnitude for rare events.

Summary of Key Relationships

The reliability index is not merely a computational artifact — it is a design currency. Once β is established for a given limit state, it can be compared directly against code targets, used to calibrate partial safety factors, and tracked across design iterations. The next step in practice is to connect these probabilistic measures to the load and resistance factor format used in everyday design checks, which is where the theory becomes directly actionable.