Bootstrap Markov chain Monte Carlo and optimal solutions for the Law of Categorical Judgment (Corrected)

A novel procedure is described for accelerating the convergence of Markov chain Monte Carlo computations. The algorithm uses an adaptive bootstrap technique to generate candidate steps in the Markov Chain. It is efficient for symmetric, convex probability distributions, similar to multivariate Gaussians, and it can be used for Bayesian estimation or for obtaining maximum likelihood solutions with confidence limits. As a test case, the Law of Categorical Judgment (Corrected) was fitted with the algorithm to data sets from simulated rating scale experiments. The correct parameters were recovered from practical-sized data sets simulated for Full Signal Detection Theory and its special cases of standard Signal Detection Theory and Complementary Signal Detection Theory.

💡 Research Summary

The paper introduces a new algorithm called Bootstrap Markov chain Monte Carlo (BMCMC) that dramatically speeds up the convergence of traditional Markov chain Monte Carlo (MCMC) methods. Conventional MCMC relies on a fixed proposal distribution V, which can be inefficient for high‑dimensional or poorly shaped target densities. BMCMC overcomes this limitation by exploiting the history of the chain: after an initial “warm‑up” phase using a pre‑specified multivariate Gaussian proposal V*, 90 % of subsequent proposals are generated by randomly selecting two previously accepted parameter vectors from an archive and taking their difference as the step direction. This bootstrap step approximates the convolution of the target density with itself; when the target is approximately multivariate Gaussian, the resulting proposal distribution has the same orientation and shape as the target but with twice the variance, leading to automatically aligned and appropriately scaled moves. The remaining 10 % of steps continue to be drawn from V* (with an adaptive diagonal scaling) to guarantee that the chain can escape any low‑dimensional subspace that the bootstrap might otherwise be confined to.

A key adaptive mechanism monitors the acceptance fraction fλ over a sliding window of length tλ. If fλ exceeds the target value of roughly ¼, the scaling factor λ is increased, shrinking step sizes; if fλ falls below the target, λ is decreased, enlarging steps. Although this adaptation temporarily violates the Markov property, the archive grows without bound, and the proposal distribution converges to the stationary target density, restoring asymptotic Markovian behavior.

BMCMC operates in two closely related modes. In optimization mode, the algorithm seeks the parameter vector that maximizes the log‑likelihood (or a Bayesian posterior). A simulated‑annealing‑like temperature schedule is employed: early iterations allow large decreases in log‑likelihood, while each accepted step reduces the temperature. When a new log‑likelihood exceeds the current maximum by more than half the temperature, a “reset” is triggered: the temperature is nudged upward, counters are cleared, the scaling window tλ is shortened, and the archive is pruned by discarding the lowest‑likelihood samples. After a reset, for the next 2 × Nₚ accepted steps (Nₚ = number of free parameters) the chain restarts from the best‑so‑far point, helping it escape local maxima. Termination requires (1) evidence that the drift of the parameter vector has ceased (checked by monitoring the angles between successive difference vectors) and (2) that a sufficient number of steps—estimated from a bent‑Gaussian model of the likelihood surface—have been taken to explore the longest axis of the posterior.

In sampling mode, after optimization has converged, the temperature is fixed at unity and the chain continues to draw samples from the posterior using the standard Metropolis acceptance rule. The archive now contains a large collection of draws that, once the final reset has occurred, constitute an accurate Monte‑Carlo approximation of the posterior distribution. Confidence intervals, means, variances, and any other integrals over the posterior are obtained by simple averages over the archived samples. If memory constraints arise, the oldest samples may be discarded without materially affecting the approximation because the remaining draws still represent the stationary distribution.

The authors validate BMCMC by applying it to the Law of Categorical Judgment (Corrected), a general psychophysical model for rating‑scale data. In this model each stimulus generates a Gaussian “signal” on an internal continuum, each response criterion is also Gaussian, and the observed response is determined by the smallest positive difference between signal and criterion. The model encompasses Full Signal Detection Theory, standard SDT, and Complementary SDT as special cases. Synthetic data sets were generated for each case, ranging from a few thousand to tens of thousands of trials, and BMCMC was used both to locate the maximum‑likelihood parameters and to obtain posterior confidence intervals. Across all conditions the algorithm recovered the true parameters within statistical error, even for problems with up to ~200 free parameters. Convergence was rapid: after a modest number of resets and adaptive λ adjustments the acceptance rate stabilized near the target, and the archive grew sufficiently large to guarantee asymptotic Markov behavior.

Key contributions of the paper include: (1) a novel bootstrap‑based proposal mechanism that automatically adapts to the shape and orientation of approximately Gaussian targets; (2) a rigorous adaptive scaling scheme that maintains a desirable acceptance rate while preserving eventual Markovian properties; (3) a unified framework that handles both deterministic optimization and full Bayesian sampling, providing parameter estimates and credible intervals in a single run; (4) practical strategies for dealing with hard constraints (walls) and for determining robust termination criteria based on drift analysis and a bent‑Gaussian exploration model; and (5) empirical evidence that the method scales to high‑dimensional problems and yields accurate inference for complex psychophysical models.

In summary, BMCMC offers a powerful, adaptive alternative to conventional MCMC, especially suited for symmetric, convex, near‑Gaussian posterior distributions common in psychological rating‑scale modeling and many other scientific domains. Its ability to accelerate convergence, handle constraints, and deliver both point estimates and uncertainty quantification makes it a valuable tool for researchers requiring efficient Bayesian or maximum‑likelihood inference in high‑dimensional settings.