Bartlett Couplings of the Onion and Vine LKJ Samplers
arXiv:2608.06116
2026
Sampling
2 ideas extracted · analyzed Aug 31, 2026
What the math gives to ML
The paper provides an exact, rejection-free sampler for valid correlation matrices using row-normalized Bartlett factors: independent Gaussian off-diagonal entries and chi-squared diagonal entries are normalized row by row to produce an LKJ-distributed correlation matrix. This is transferable whenever a neural network needs a positive-semidefinite correlation structure but should avoid expensive generic matrix decompositions or invalid entrywise parameterizations. Two practical uses are correlated perturbations for ensembles or attention heads, and LKJ-constrained covariance factors for variational adapters or Bayesian neural-network blocks. The main engineering asset is a linear-in-dimension construction whose direct normalization avoids gamma-ratio redundancy and subtractive numerical cancellation.
Ideas from this paper
Unverified
2026
Use an LKJ correlation factor as the correlation component of a variational posterior over a compact adapter, LoRA factor, or Bayesian neural-network parameter block. The model learns marginal scales separately while the correlation matrix remains automatically positive semidefinite and unit-diagonal, avoiding unconstrained covariance matrices, invalid correlations, and fragile covariance decompositions.
Useful6/10
Difficulty6/10
Novelty6/10
Unverified
2026
Replace independent dropout or Gaussian perturbations across attention heads, ensemble members, or diffusion score replicas with a positive-semidefinite correlation matrix sampled from an LKJ distribution. The concentration parameter eta controls whether perturbations are nearly independent or strongly correlated in a controlled way, while the Bartlett construction guarantees a valid covariance without matrix rejection or projection.
Useful5/10
Difficulty4/10
Novelty7/10