This document presents the Monte Carlo study used to validate the
implementation of the Rtwalk package. The objective is to
demonstrate that the sampler behaves as expected across a variety of
challenging scenarios, replicating classic tests from the MCMC
literature.
calculate_diagnostics <- function(samples, burnin_frac = 0.2, param_names = NULL, title = "") {
n_iter <- nrow(samples)
n_burnin <- floor(burnin_frac * n_iter)
post_burnin_samples <- samples[(n_burnin + 1):n_iter, , drop = FALSE]
n_param <- ncol(samples)
if (is.null(param_names)) {
param_names <- paste0("theta", 1:n_param)
}
means <- apply(post_burnin_samples, 2, mean)
sds <- apply(post_burnin_samples, 2, sd)
quantiles <- apply(post_burnin_samples, 2, quantile, probs = c(0.025, 0.5, 0.975))
chain <- coda::as.mcmc(post_burnin_samples)
ess <- coda::effectiveSize(chain)
results_table <- data.frame(
Parameter = param_names,
Mean = round(means, 4),
SD = round(sds, 4),
Q2.5 = round(quantiles["2.5%", ], 4),
Median = round(quantiles["50%", ], 4),
Q97.5 = round(quantiles["97.5%", ], 4),
ESS = round(ess, 0)
)
cat("\nCONVERGENCE DIAGNOSTICS -", title, "\n")
print(results_table, row.names = FALSE)
return(invisible(results_table))
}
visualize_results <- function(samples, true_values = NULL,
title = "Results", burnin_frac = 0.2,
true_covariance = NULL, show_acf = TRUE) {
n_total <- nrow(samples)
start_index <- ceiling(n_total * burnin_frac) + 1
analysis_samples <- samples[start_index:n_total, , drop = FALSE]
n_dim <- ncol(analysis_samples)
oldpar <- par(no.readonly = TRUE)
on.exit(par(oldpar))
colors <- c("#2E86AB", "#A23B72", "#F18F01", "#C73E1D", "#7209B7", "#2F9599")
bg_color <- "#FAFAFA"
if (n_dim == 1) {
layout(matrix(c(1, 2, 3, 3), 2, 2, byrow = TRUE))
par(mar = c(4, 4, 3, 1), bg = bg_color, col.axis = "gray30", col.lab = "gray30", col.main = "gray20")
hist(analysis_samples[,1], main = paste(title, "- Distribution"),
xlab = bquote(theta[1]), freq = FALSE, col = colors[1], border = "white",
breaks = 30, las = 1)
plot(samples[start_index:n_total, 1], type = 'l', col = colors[1], lwd = 0.8,
main = paste(title, "- Trace"), xlab = "Iteration", ylab = bquote(theta[1]), las = 1)
if (show_acf) {
acf(analysis_samples[,1], main = paste(title, "- Autocorrelation"),
col = colors[2], lwd = 2, las = 1)
}
} else if (n_dim == 2) {
layout(matrix(c(1, 2, 3, 4), 2, 2, byrow = TRUE))
par(mar = c(4, 4, 3, 1), bg = bg_color, col.axis = "gray30", col.lab = "gray30", col.main = "gray20")
plot(analysis_samples, pch = ".", col = adjustcolor(colors[1], alpha = 0.6),
main = title, xlab = bquote(theta[1]), ylab = bquote(theta[2]), las = 1,
cex = 1.2)
if (!is.null(true_covariance) && requireNamespace("ellipse", quietly = TRUE)) {
lines(ellipse::ellipse(true_covariance, centre = true_values), col = colors[2], lwd = 2)
}
plot(samples[start_index:n_total, 1], type = 'l', col = colors[1], lwd = 0.8,
main = bquote(paste("Trace ", theta[1])), xlab = "Iteration", ylab = bquote(theta[1]), las = 1)
plot(samples[start_index:n_total, 2], type = 'l', col = colors[2], lwd = 0.8,
main = bquote(paste("Trace ", theta[2])), xlab = "Iteration", ylab = bquote(theta[2]), las = 1)
dens1 <- density(analysis_samples[,1])
plot(dens1, main = bquote(paste("Marginal Density ", theta[1])), xlab = bquote(theta[1]),
col = colors[1], lwd = 3, las = 1, zero.line = FALSE)
polygon(dens1, col = adjustcolor(colors[1], alpha = 0.3), border = NA)
} else {
n_plots <- min(6, n_dim)
if (n_plots <= 4) {
layout(matrix(1:4, 2, 2, byrow = TRUE))
} else {
layout(matrix(1:6, 3, 2, byrow = TRUE))
}
par(mar = c(4, 4, 3, 1), bg = bg_color, col.axis = "gray30", col.lab = "gray30", col.main = "gray20")
for (i in 1:n_plots) {
if (i <= 3) {
plot(samples[start_index:n_total, i], type = 'l', col = colors[i], lwd = 0.8,
main = bquote(paste("Trace ", theta[.(i)])),
xlab = "Iteration", ylab = bquote(theta[.(i)]), las = 1)
} else {
dens <- density(analysis_samples[,i])
plot(dens, main = bquote(paste("Density ", theta[.(i)])),
xlab = bquote(theta[.(i)]), col = colors[i], lwd = 3, las = 1, zero.line = FALSE)
polygon(dens, col = adjustcolor(colors[i], alpha = 0.3), border = NA)
}
}
if (n_dim > 6) {
cat("\nNote: Only showing first 6 dimensions. Total dimensions:", n_dim, "\n")
}
}
}N_ITER_VIGNETTE <- 5000
BURN_FRAC <- 0.2
# --- TEST 1: Standard Univariate Normal ---
log_posterior_1 <- function(x) dnorm(x, log = TRUE)
result_1 <- twalk(log_posterior_1, n_iter = N_ITER_VIGNETTE, x0 = -2, xp0 = 2)
calculate_diagnostics(result_1$all_samples, BURN_FRAC, "theta", "Standard Normal")
visualize_results(result_1$all_samples, 0, "Test 1: Standard Normal")
# --- TEST 2: Correlated Bivariate Normal ---
true_mean_2 <- c(1, -0.5)
true_cov_2 <- matrix(c(4, 0.7*2*1.5, 0.7*2*1.5, 2.25), 2, 2)
log_posterior_2 <- function(x) mvtnorm::dmvnorm(x, mean = true_mean_2, sigma = true_cov_2, log = TRUE)
result_2 <- twalk(log_posterior_2, n_iter = N_ITER_VIGNETTE, x0 = c(0,0), xp0 = c(2,-1))
calculate_diagnostics(result_2$all_samples, BURN_FRAC, c("theta1", "theta2"), "Bivariate Normal")
visualize_results(result_2$all_samples, true_mean_2, "Test 2: Bivariate Normal", true_covariance = true_cov_2)
# --- TEST 3: Funnel Distribution ---
log_posterior_3 <- function(x) {
x1 <- x[1]
x2 <- x[2]
log_prior_x1 <- dnorm(x1, mean = 0, sd = 3, log = TRUE)
log_lik_x2 <- dnorm(x2, mean = 0, sd = exp(x1 / 2), log = TRUE)
return(log_prior_x1 + log_lik_x2)
}
result_3 <- twalk(log_posterior_3, n_iter = N_ITER_VIGNETTE, x0 = c(-1.5, -0.2), xp0 = c(1.5, -0.2))
calculate_diagnostics(result_3$all_samples, BURN_FRAC, c("theta1", "theta2"), "Funnel")
visualize_results(result_3$all_samples, NULL, "Test 3: Funnel")
# --- TEST 4: Rosenbrock Distribution ---
log_posterior_4 <- function(x) {
x1 <- x[1]
x2 <- x[2]
k <- 1 / 20
return(-k * (100 * (x2 - x1^2)^2 + (1 - x1)^2))
}
result_4 <- twalk(log_posterior_4, n_iter = 100000, x0 = c(0,0), xp0 = c(-1,1))
calculate_diagnostics(result_4$all_samples, BURN_FRAC, c("theta1", "theta2"), "Rosenbrock")
visualize_results(result_4$all_samples, c(1,1), "Test 4: Rosenbrock")
# --- TEST 5: Gaussian Mixture ---
weight1 <- 0.7; mean1 <- c(6, 0); sigma1_1 <- 4; sigma1_2 <- 5; rho1 <- 0.8
cov1 <- matrix(c(sigma1_1^2, rho1*sigma1_1*sigma1_2, rho1*sigma1_1*sigma1_2, sigma1_2^2), nrow=2)
weight2 <- 0.3; mean2 <- c(-3, 10); sigma2_1 <- 1; sigma2_2 <- 1; rho2 <- 0.1
cov2 <- matrix(c(sigma2_1^2, rho2*sigma2_1*sigma2_2, rho2*sigma2_1*sigma2_2, sigma2_2^2), nrow=2)
log_posterior_5 <- function(x) {
log(weight1 * mvtnorm::dmvnorm(x, mean1, cov1) + weight2 * mvtnorm::dmvnorm(x, mean2, cov2))
}
result_5 <- twalk(log_posterior_5, n_iter = 100000, x0 = mean1, xp0 = mean2)
calculate_diagnostics(result_5$all_samples, BURN_FRAC, c("theta1", "theta2"), "Gaussian Mixture")
visualize_results(result_5$all_samples, NULL, "Test 5: Gaussian Mixture")
# --- TEST 6: High Dimensionality (10D) ---
n_dim_6 <- 10; true_mean_6 <- 1:n_dim_6; rho <- 0.7
true_cov_6 <- matrix(rho^abs(outer(1:n_dim_6, 1:n_dim_6, "-")), n_dim_6, n_dim_6)
log_posterior_6 <- function(x) mvtnorm::dmvnorm(x, mean = true_mean_6, sigma = true_cov_6, log = TRUE)
result_6 <- twalk(log_posterior_6, n_iter = N_ITER_VIGNETTE, x0 = rep(0, n_dim_6), xp0 = rep(2, n_dim_6))
calculate_diagnostics(result_6$all_samples, BURN_FRAC, paste0("theta", 1:n_dim_6), "10D Normal")
visualize_results(result_6$all_samples, true_mean_6, "Test 6: 10D Normal")
# --- TEST 7: Bayesian Logistic Regression ---
set.seed(123); n_obs <- 2000
true_beta <- c(0.5, -1.2, 0.8)
X <- cbind(1, rnorm(n_obs, 0, 1), rnorm(n_obs, 0, 1.5))
eta <- X %*% true_beta; prob <- 1 / (1 + exp(-eta)); y <- rbinom(n_obs, 1, prob)
log_posterior_7 <- function(beta, X, y) {
eta <- X %*% beta
log_lik <- sum(y * eta - log(1 + exp(eta)))
log_prior <- sum(dnorm(beta, 0, 5, log = TRUE))
return(log_lik + log_prior)
}
result_7 <- twalk(log_posterior_7, n_iter = N_ITER_VIGNETTE, x0 = c(0,0,0), xp0 = c(0.2,-0.2,0.1), X = X, y = y)
calculate_diagnostics(result_7$all_samples, BURN_FRAC, c("beta0", "beta1", "beta2"), "Logistic Regression")
visualize_results(result_7$all_samples, true_beta, "Test 7: Logistic Regression")The results from the test battery demonstrate that this implementation of the t-walk is robust and behaves as expected. The sampler was able to converge and efficiently explore distributions with high correlation and multimodality without the need for manual tuning, validating its utility as a general-purpose tool for Bayesian inference.