The standard LKJ family requires eta to be strictly positive.

Does eta equal to one make every correlation uniform?

Only in the two-dimensional case does the single correlation have that simple uniform distribution; higher-dimensional constraints change marginal behavior.

Why use a Cholesky factor?

It offers a convenient constrained parameterization and often better numerical behavior, provided the appropriate transformed distribution is used.

AI Glossary / Statistics and probability / Advanced

Lewandowski-Kurowicka-Joe distribution

Definition

The Lewandowski-Kurowicka-Joe distribution, usually called the LKJ distribution, is a probability distribution over positive-definite correlation matrices. It is widely used as a Bayesian prior for unknown correlation structure. For shape parameter eta greater than zero, its density is proportional to the determinant of the correlation matrix raised to eta minus one. The distribution respects the joint mathematical constraints among correlations, unlike independently choosing arbitrary pairwise values. It describes correlations rather than the marginal standard deviations needed to form a covariance matrix.

What it means

A multivariate model needs a coherent description of how variables vary together. Pairwise correlations must collectively form a valid matrix, so choosing them separately can produce an impossible combination. LKJ supplies a distribution over the valid space. Eta controls how strongly density favors correlation matrices near the identity versus near the boundary. Eta equal to one gives a uniform density over the correlation-matrix space under the usual parameterization. This does not mean that every individual correlation is independently uniform, especially when the matrix dimension exceeds two.

How it works

Specify the matrix dimension and a positive eta, then use the LKJ distribution for the correlation component of a multivariate model. Model marginal standard deviations separately if a covariance matrix is needed, combining them as Sigma = D R D, where R is the correlation matrix and D contains standard deviations. Computational implementations commonly parameterize R through a Cholesky factor for numerical efficiency and stability. The transformed density must account for the change of variables; a constant density on correlation matrices does not imply a constant density on Cholesky factors. Evaluate prior predictive implications and sensitivity to eta before interpreting posterior results.

Three practical examples

Hierarchical modeling: assign a coherent prior to correlations among group-level effects.
Multivariate outcomes: model dependence among several responses while keeping their marginal scales distinct.
Simulation studies: generate valid random correlation matrices to assess statistical procedures. The selected eta and dimension influence the simulated dependence patterns, so the generated matrices are not neutral examples of every possible application.

Think of it this way

Think of a correlation matrix as an interconnected structure whose pieces must fit together. LKJ places probability over complete valid structures instead of choosing each joint independently and hoping the assembly works. Eta adjusts the preference for structures with weaker overall dependence. It does not identify which particular positive or negative association the data should contain.

Worked example

For a two-by-two correlation matrix with off-diagonal correlation rho, the determinant is 1 - rho squared. With eta = 2, the normalized density on -1 < rho < 1 is 0.75(1 - rho squared). Its density at zero is 0.75, while at rho = 0.8 it is 0.27, a ratio of 0.36. Integrating over -0.5 to 0.5 gives probability 11/16 = 0.6875. Under eta = 1 in this two-dimensional case, rho is uniform and that probability is 0.5. If marginal standard deviations are 2 and 3 and rho = 0.5, the covariance matrix has diagonal entries 4 and 9 and off-diagonal entries 3. These original calculations show both the concentration effect of eta and the distinction between correlation and covariance. The uniform marginal statement here is specific to dimension two.

Illustrative case study

In a fictional case, Ash Research models several related performance outcomes across locations. Analysts use separate priors for outcome scales and an LKJ prior for correlations, implemented through a Cholesky factor. They simulate from the prior before fitting and discover that the initial specification implies dependence patterns that are too concentrated for their substantive expectations. They revise the prior transparently and compare posterior conclusions under plausible alternatives. The report describes the prior's role rather than presenting estimated correlations as entirely data-driven.

Common misconceptions and mistakes

Uniform over valid correlation matrices does not mean independently uniform pairwise correlations. Eta greater than one favors the identity in density but does not force all correlations to zero. The mean correlation matrix is the identity for every positive eta, so the mean alone does not reveal the distribution's concentration. LKJ is not a prior for variances, and it does not establish independence in a general non-Gaussian model merely because a correlation is zero. Dimension affects marginal correlation behavior, and transformed implementations require the correct density and Jacobian treatment.

Business use

Businesses using Bayesian multivariate models should document how dependence assumptions affect predictions and decisions. Choose eta through substantive reasoning, prior predictive checks and sensitivity analysis rather than treating a convenient default as universally uninformative. Keep scale priors separate and inspect whether the resulting covariance structures are plausible. Review sampling diagnostics and the stability of posterior correlations, especially when data are limited. A valid matrix is a necessary mathematical condition, not proof that the dependence model represents the operational system.

Check your understanding

Why is the covariance off-diagonal 3 when rho = 0.5 and standard deviations are 2 and 3?

Answer: Covariance equals correlation times both standard deviations: 0.5 times 2 times 3 = 3. LKJ supplies a distribution for the correlation component; the scale information comes from separate quantities.

Related terms

Correlation matrix
Covariance matrix
Regularization
LKJ prior
positive definiteness
Cholesky factor
hierarchical model
prior predictive check
multivariate distribution

Learn next

Study how covariance matrices can be decomposed into scales and correlations. Then examine dimension-dependent LKJ marginal distributions, Cholesky parameterization and prior sensitivity in hierarchical models with limited group-level information.

Sources

Stan Functions Reference, Correlation Matrix Distributions: https://mc-stan.org/docs/functions-reference/correlation_matrix_distributions.html . This documents the LKJ density, eta interpretation and Cholesky form and cites Lewandowski, Kurowicka and Joe's original work. The normalized two-dimensional calculation and fictional case above are original illustrations.

Previous: Leverage Next: Lexical ambiguity