GitHub

Solvers

Fundamentally, all interesting quantities of GMRFs (samples, marginal variances, posterior means, ...) must be computed through sparse linear algebra. GaussianMarkovRandomFields.jl provides a number of so-called solvers which perform the underlying computations in different ways and with different trade-offs. Our goal is to provide sane defaults that "just work" for most users, while still allowing power users to customize the behavior of the solvers. Read further to learn about the available solvers and how to use them.

If you don't have time for the details and you're just looking for a recommendation, skip to Choosing a solver.

If you're interested in the interface of solvers, skip to Solver interface.

Cholesky (CHOLMOD)

CHOLMOD is a state-of-the-art direct solver for sparse linear systems. CholeskySolver uses CHOLMOD to compute the sparse Cholesky factorization of the precision matrix, which it then leverages to draw samples and compute posterior means. Marginal variances are computed using a user-specified variance strategy, which defaults to a sampling-based approach.

GaussianMarkovRandomFields.CholeskySolverBlueprint — Type
CholeskySolverBlueprint(; var_strategy = RBMCStrategy(100), perm = nothing)

A blueprint for a direct solver that uses a sparse Cholesky decomposition computed through CHOLMOD.

Keyword arguments

  • var_strategy::AbstractVarianceStrategy: Strategy for computing the marginal variances of the GMRF. Defaults to RBMCStrategy(100).
  • perm::Union{Nothing,Vector{Int}}: Permutation / node reordering to use for the Cholesky decomposition. Defaults to nothing, which means that the solver will compute its own permutation.
source

Pardiso

Pardiso is a state-of-the-art direct solver for sparse linear systems. Its main benefit over the CholeskySolver is its potential for better performance on large-dimensional GMRFs, as well as its highly efficient method for marginal variance computation.

To use this solver, you need to set up and load Pardiso.jl. The code for our Pardiso solver will then be loaded automatically through a package extension.

CG

The Conjugate Gradient (CG) method is an iterative method for solving symmetric positive-definite linear systems. CGSolver uses CG to compute posterior means and draw samples.

GaussianMarkovRandomFields.CGSolverBlueprint — Type
CGSolverBlueprint(; reltol, abstol, maxiter, preconditioner_strategy,
                    var_strategy, mean_residual_guess)

A blueprint for a conjugate gradient-based solver.

Keyword arguments

  • reltol::Real = sqrt(eps(Float64)): Relative tolerance of CG.
  • abstol::Real = 0.0: Absolute tolerance of CG.
  • maxiter::Int = 1000: Maximum number of iterations.
  • preconditioner_strategy::Function: Maps a GMRF instance to a preconditioner.
  • var_strategy::AbstractVarianceStrategy: A variance strategy.
  • mean_residual_guess::Union{Nothing,AbstractVector}: An initial guess for the mean residual.
source

Variance strategies

Fundamentally, computing marginal variances of GMRFs is not trivial, as it requires computing the diagonal entries of the covariance matrix (which is the inverse of the precision matrix). The Takahashi recursions (sometimes referred to as a sparse (partial) inverse method) are a highly accurate and stable method for computing these variances. TakahashiStrategy uses the Takahashi recursions to compute marginal variances. PardisoSolver also uses this algorithm internally, but with a highly optimized implementation.

RBMCStrategy is a sampling-based approach to computing marginal variances. It is less accurate than the Takahashi recursions, but can be much faster for large GMRFs.

GaussianMarkovRandomFields.RBMCStrategy — Type
RBMCStrategy(n_samples; rng)

Rao-Blackwellized Monte Carlo estimator of a GMRF's marginal variances based on [3]. Particularly useful in large-scale regimes where Takahashi recursions may be too expensive.

Arguments

  • n_samples::Int: Number of samples to draw.

Keyword arguments

  • rng::Random.AbstractRNG = Random.default_rng(): Random number generator.
source
GaussianMarkovRandomFields.BlockRBMCStrategy — Type
BlockRBMCStrategy(n_samples; rng, enclosure_size)

Block Rao-Blackwellized Monte Carlo estimator of a GMRF's marginal variances based on [3]. Achieves faster convergence than plain RBMC by considering blocks of nodes rather than individual nodes, thus integrating more information about the precision matrix. enclosure_size specifies the size of these blocks. Larger values lead to faster convergence (in terms of the number of samples) at the cost of increased compute. Thus, one should aim for a sweet spot between sampling costs and block operation costs.

Arguments

  • n_samples::Int: Number of samples to draw.

Keyword arguments

  • rng::Random.AbstractRNG = Random.default_rng(): Random number generator.
  • enclosure_size::Int = 1: Size of the blocks.
source

Choosing a solver

The default solver uses the CholeskySolver for small to medium-sized GMRFs and switches to the CGSolver for large GRMFs. Similarly, the default variance strategy is TakahashiStrategy for small to medium-sized GMRFs and RBMCStrategy for large GMRFs. See:

Indeed, this matches our general recommendations: Direct solvers combined with the Takahashi recursions are highly accurate and stable and should be used whenever possible. However, direct solvers become prohibitively expensive for very large-scale GMRFs, both in terms of compute and memory use. In these regimes, CG with a good preconditioner may still be a viable option.

If you have access to both strong parallel computing resources and a Pardiso license, we recommend the use of the PardisoGMRFSolver. In particular, Pardiso's Takahashi recursions are highly optimized and much faster than the implementation used for our TakahashiStrategy.

Solver interface