r_test / packages / r-sns

• 0

Stochastic Newton Sampler (SNS) is a Metropolis-Hastings-based, Markov Chain Monte Carlo sampler for twice differentiable, log-concave probability density functions (PDFs) where the proposal density function is a multivariate Gaussian resulting from a second-order Taylor-series expansion of log-density around the current point. The mean of the Gaussian proposal is the full Newton-Raphson step from the current point. A Boolean flag allows for switching from SNS to Newton-Raphson optimization (by choosing the mean of proposal function as next point). This can be used during burn-in to get close to the mode of the PDF (which is unique due to concavity). For high-dimensional densities, mixing can be improved via 'state space partitioning' strategy, in which SNS is applied to disjoint subsets of state space, wrapped in a Gibbs cycle. Numerical differentiation is available when analytical expressions for gradient and Hessian are not available. Facilities for validation and numerical differentiation of log-density are provided.

• Conda
• Files
• Labels
• Badges