sns | R Documentation |
SNS is a Metropolis-Hastings MCMC sampler with a multivariate Gaussian proposal function resulting from a local, second-order Taylor series expansion of log-density. The mean of the Gaussian proposal is identical to the full Newton-Raphson step from the current point. During burn-in, Newton-Raphson optimization can be performed to get close to the mode of the pdf which is unique due to convexity, resulting in faster convergence. For high dimensional densities, state space partitioning can be used to improve mixing. Support for numerical differentiation is provided using numDeriv package. sns
is the low-level function for drawing one sample from the distribution. For drawing multiple samples from a (fixed) distribution, consider using sns.run
.
sns(x, fghEval, rnd = TRUE, gfit = NULL, mh.diag = FALSE , part = NULL, numderiv = 0 , numderiv.method = c("Richardson", "simple") , numderiv.args = list(), ...)
x |
Current state vector. |
fghEval |
Log-density to be sampled from. A valid log-density can have one of 3 forms: 1) return log-density, but no gradient or Hessian, 2) return a list of |
rnd |
Runs 1 iteration of Newton-Raphson optimization method (non-stochastic or 'nr' mode) when |
gfit |
Gaussian fit at point |
mh.diag |
Boolean flag, indicating whether detailed MH diagnostics such as components of acceptance test must be returned or not. |
part |
List describing partitioning of state space into subsets. Each element of the list must be an integer vector containing a set of indexes (between |
numderiv |
Integer with value from the set |
numderiv.method |
Method used for numeric differentiation. This is passed to the |
numderiv.args |
Arguments to the numeric differentiation method chosen in |
... |
Other arguments to be passed to |
sns
returns the sample drawn as a vector, with attributes:
accept |
A boolean indicating whether the proposed point was accepted. |
ll |
Value of the log-density at the sampled point. |
gfit |
List containing Gaussian fit to pdf at the sampled point. |
1. Since SNS makes local Gaussian approximations to the density with the covariance matrix of the Gaussian proposal being the log-density Hessian, there is a strict requirement for the log-density to be concave.
2. Proving log-concavity for arbitrary probability distributions is non-trvial. However, distributions generated by replacing parameters of a concave distribution with linear expressions are known to be log-concave. This negative-definiteness invariance as well as expressions for full gradient and Hessian in terms of derivatives of low-dimensional base distributions are discussed in the vignette. The GLM expansion framework is available in the R package RegressionFactory.
3. See package vignette for more details on SNS theory, software, examples, and performance.
Alireza S. Mahani, Asad Hasan, Marshall Jiang, Mansour T.A. Sharabiani
Mahani A.S., Hasan A., Jiang M. & Sharabiani M.T.A. (2016). Stochastic Newton Sampler: The R Package sns. Journal of Statistical Software, Code Snippets, 74(2), 1-33. doi:10.18637/jss.v074.c02
Hastings, W. K. (1970). Monte Carlo sampling methods using Markov chains and their applications. Biometrika, 57(1), 97-109.
Qi, Y., & Minka, T. P. (2002). Hessian-based markov chain monte-carlo algorithms. 1st Cape Cod Workshop on Monte Carlo Methods.
sns.run
, sns.fghEval.numaug
## Not run: # using RegressionFactory for generating log-likelihood and its derivatives library(RegressionFactory) loglike.poisson <- function(beta, X, y) { regfac.expand.1par(beta, X = X, y = y, fbase1 = fbase1.poisson.log) } # simulating data K <- 5 N <- 1000 X <- matrix(runif(N * K, -0.5, +0.5), ncol = K) beta <- runif(K, -0.5, +0.5) y <- rpois(N, exp(X %*% beta)) beta.init <- rep(0.0, K) # glm estimate, for reference beta.glm <- glm(y ~ X - 1, family = "poisson", start = beta.init)$coefficients # running SNS in non-stochastic mode # this should produce results very close to glm beta.sns <- beta.init for (i in 1:20) beta.sns <- sns(beta.sns, loglike.poisson, X = X, y = y, rnd = F) # comparison all.equal(as.numeric(beta.glm), as.numeric(beta.sns)) # trying numerical differentiation loglike.poisson.fonly <- function(beta, X, y) { regfac.expand.1par(beta, X = X, y = y, fgh = 0, fbase1 = fbase1.poisson.log) } beta.sns.numderiv <- beta.init for (i in 1:20) beta.sns.numderiv <- sns(beta.sns.numderiv, loglike.poisson.fonly , X = X, y = y, rnd = F, numderiv = 2) all.equal(as.numeric(beta.glm), as.numeric(beta.sns.numderiv)) # add numerical derivatives to fghEval outside sns loglike.poisson.numaug <- sns.fghEval.numaug(loglike.poisson.fonly , numderiv = 2) beta.sns.numaug <- beta.init for (i in 1:20) # set numderiv to 0 to avoid repeating # numerical augmentation inside sns beta.sns.numaug <- sns(beta.sns.numaug, loglike.poisson.numaug , X = X, y = y, rnd = F, numderiv = 0) all.equal(as.numeric(beta.glm), as.numeric(beta.sns.numaug)) ## End(Not run)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.