Gaussian Copula Generation

Summary

Gaussian copula generation is a way to generate multivariate synthetic data when each variable can have its own marginal distribution, but the dependency between variables is controlled by a Gaussian correlation structure.

The key idea:

sample correlated Gaussian variables
convert them to uniform variables by the Gaussian CDF
convert each uniform variable to the target distribution by that variable's inverse CDF

So the Gaussian copula separates:

marginal distribution: what each variable looks like alone
dependency structure: how variables move together

What is a copula

A copula is a function/model that links individual marginal distributions into a joint multivariate distribution. In simple terms, it describes the dependency between variables separately from what each variable's own distribution looks like.

What problem does it solve

Often we want to simulate variables that are dependent, but not all normally distributed.

Example:

age may be approximately continuous and skewed
biomarker values may be log-normal
disease status may be binary
healthcare cost may be highly right-skewed

A plain multivariate normal model is too restrictive because it assumes all variables are Gaussian. A Gaussian copula lets us keep flexible marginals while still specifying a correlation matrix.

Intuition

Think of Gaussian copula as generating the rank-level dependency first.

Start with hidden correlated Gaussian variables
Turn them into percentiles between 0 and 1
Map those percentiles into whatever marginal distributions we want

So if two hidden Gaussian variables are positively correlated, high percentiles in one variable tend to appear with high percentiles in the other variable.

Generation steps

Suppose we want to generate $d$ variables.

1. Specify a correlation matrix

Choose a valid correlation matrix $R$ :

R = (\begin{matrix} 1 & ρ_{12} & \dots & ρ_{1 d} \\ ρ_{21} & 1 & \dots & ρ_{2 d} \\ ⋮ & ⋮ & ⋱ & ⋮ \\ ρ_{d 1} & ρ_{d 2} & \dots & 1 \end{matrix})

This controls the dependency structure. See Covariance & Correlation and Covariance Structure.

2. Sample from a multivariate normal distribution

Generate latent Gaussian variables:

Z = (Z_{1}, . . ., Z_{d}) \sim N (0, R)

Here, each $Z_{j}$ is standard normal, but the variables are correlated according to $R$ .

3. Convert Gaussian values to uniform values

Apply the standard normal CDF $Φ$ to each dimension:

U_{j} = Φ (Z_{j})

Then:

U_{j} \sim U n i f o r m (0, 1)

The values are now percentiles, but still dependent.

4. Convert uniforms to target marginals

For each target variable with CDF $F_{j}$ , use the inverse CDF:

X_{j} = F_{j}^{- 1} (U_{j})

Then $X_{j}$ follows the desired marginal distribution.

Example:

$X_{1} = F_{n o r m a l}^{- 1} (U_{1})$
$X_{2} = F_{l o g n o r m a l}^{- 1} (U_{2})$
$X_{3} = F_{b e r n o u l l i}^{- 1} (U_{3})$

Compact formula

The full transformation is:

X_{j} = F_{j}^{- 1} (Φ (Z_{j})), Z \sim N (0, R)

where:

$R$ : Gaussian correlation matrix
$Φ$ : standard normal CDF
$F_{j}^{- 1}$ : inverse CDF of the target marginal distribution

Why it is useful

Generate synthetic data with realistic dependency
Simulate non-normal variables jointly
Separate marginal modeling from dependency modeling
Useful in risk modeling, health data simulation, finance, and sensitivity analysis

Limitations

The dependency is still Gaussian-like, so it may miss tail dependence
The input correlation matrix must be valid / positive semi-definite
Correlation after transformation may not exactly equal the original $R$
Discrete variables can create ties and make correlation harder to control

Important

Gaussian copula is not saying the observed variables are Gaussian. It only uses a hidden Gaussian layer to create dependency.

Related Notes