Nataf Transformation

Overview

The Nataf Transformation is a widely utilized isoprobabilistic transformation in structural reliability analysis. Its purpose is to transform random vectors with correlated non-normal marginals $\vec{X}$ into random vectors with uncorrelated standard normal marginals $\vec{U}$. This transformation was first introduced by André Nataf in 1962 Nataf (1962).

The Nataf Transformation $\vec{U} = T^{N}(\vec{X})$ is composed of two transformations:

\[\vec{U} = T^{N}(\vec{X}) = (T_{2}^{N} \circ T_{1}^{N})(\vec{X})\]

  • The first transformation $\vec{Z} = T_{1}^{N}(\vec{X})$ transforms random vector with correlated non-normal marginals $\vec{X}$ (with correlation matrix $\rho^{X}$) into random vector with correlated standard normal marginals $\vec{Z}$ (with correlation matrix $\rho^{Z}$). Here, $\Phi^{-1}(\cdot)$ is the inverse of the cumulative density function of a standard normal random variable and $F_{X_{i}}(\cdot)$ is the cumulative density function of a marginal$X_{i}$.

\[\vec{Z} = T_{1}^{N}(\vec{X}) = \begin{bmatrix} \Phi^{-1}(F_{X_{1}}(X_{1})) \\ \Phi^{-1}(F_{X_{2}}(X_{2})) \\ \vdots \\ \Phi^{-1}(F_{X_{n}}(X_{n})) \end{bmatrix} \]

  • The second transformation $\vec{U} = T_{2}^{N}(\vec{Z})$ transforms random vector with correlated standard normal marginals $\vec{Z}$ into random vector with uncorrelated standard normal marginals $\vec{U}$. Here, the matrix $\Gamma$ is used to decorrelate the standard normal marginals of random vector $\vec{Z}$ and can be chosen as any square-root matrix of the correlation matrix $\rho^{Z}$. Fortuna.jl uses the Cholesky factor of the inverse of the correlation matrix $(\rho^{Z})^{-1}$ as the matrix $\Gamma$.

\[\vec{U} = T_{2}^{N}(\vec{Z}) = \Gamma \vec{Z}\]

The first transformation $\vec{Z} = T_{1}^{N}(\vec{X})$ causes so-called correlation distortion. The correlation distortion causes the correlation coefficient between two standard normal marginals $Z_{i}$ and $Z_{j}$, denoted by $\rho_{ij}^{Z}$, to distort and differ from the original correlation coefficient between the corresponding non-normal marginals $X_{i}$ and $X_{j}$, denoted by $\rho_{ij}^{X}$, such that $\rho_{ij}^{Z} \neq \rho_{ij}^{X}$. The relationship between the components of the correlation matrices $\rho_{ij}^{Z}$ and $\rho_{ij}^{X}$ is given by

\[\rho_{ij}^{X} = \dfrac{1}{\sigma_{X_i} \sigma_{X_j}} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} (F_{X_i}^{-1}(\Phi(z_i)) - \mu_{X_i}) (F_{X_j}^{-1}(\Phi(z_j)) - \mu_{X_j}) \phi_2(z_i, z_j, \rho_{ij}^{Z}) dz_i dz_j\]

where $\phi_2(\cdot)$ is the bivariate standard normal probability density function. Generally, this integral cannot be inverted analytically to solve for the coefficients of the distorted correlation matrix $\rho_{ij}^{Z}$. In order to compute these coefficients, Fortuna.jl package (1) employs a two-dimensions Gauss-Legendre quadrature implemented in FastGaussQuadrature.jl package to expand the integral into a finite summation using Gauss–Legendre quadrature and (2) utilizes NonlinearSolve.jl package to find values of the coefficients of the correlation matrix $\rho_{ij}^{Z}$ that satisfy the resulting expression.

API

Fortuna.NatafTransformation — Type
NatafTransformation <: AbstractTransformation

Type used to perform Nataf Transformation.

  • X::AbstractVector{<:Distribution{Univariate, Continuous}}: Random vector $\vec{X}$

  • ρˣ::AbstractMatrix{<:Real}: Correlation matrix $\rho^{X}$

  • ρᶻ::AbstractMatrix{<:Real}: Distorted correlation matrix $\rho^{Z}$

  • L::AbstractMatrix{<:Real}: Lower triangular matrix of the Cholesky decomposition of the distorted correlation matrix $L$

  • L⁻¹::AbstractMatrix{<:Real}: Inverse of the lower triangular matrix of the Cholesky decomposition of the distorted correlation matrix $L^{-1}$

Fortuna.getdistortedcorrelation — Method
getdistortedcorrelation(X::AbstractVector{<:Distributions.UnivariateDistribution}, ρˣ::AbstractMatrix{<:Real})

Function used to compute the distorted correlation matrix $\rho^{Z}$.

Fortuna.transformsamples — Method
transformsamples(TransformationObject::NatafTransformation, Samples::AbstractVector{<:Real}, TransformationDirection::Symbol)

Function used to transform samples from $X$- to $U$-space and vice versa.
If `TransformationDirection is:

  • :X2U, then the function transforms samples $\vec{x}$ from $X$- to $U$-space.
  • :U2X, then the function transforms samples $\vec{u}$ from $U$- to $X$-space.
Fortuna.getjacobian — Method
getjacobian(TransformationObject::NatafTransformation, Samples::AbstractVector{<:Real}, TransformationDirection::Symbol)

Function used to compute the Jacobians of the transformations of samples from $X$- to $U$-space and vice versa.
If TransformationDirection is:

  • :X2U, then the function returns the Jacobians of the transformations of samples $\vec{x}$ from $X$- to $U$-space.
  • :U2X, then the function returns the Jacobians of the transformations of samples $\vec{u}$ from $U$- to $X$-space.
Distributions.pdf — Method
pdf(TransformationObject::NatafTransformation, x::AbstractVector{<:Real})

Function used to compute the joint PDF in $X$-space.