Bayesian inference usually needs sampling algorithms to approximate complex posterior distributions. Discrete-time Markov chain Monte Carlo (MCMC) methods, especially the No-U-Turn Sampler (NUTS), are widely used in modern Bayesian computation. However, MCMC methods still have limitations related to reversibility and complex posterior geometry. Recently developed Piecewise Deterministic Markov Process (PDMP) samplers are continuous-time and non-reversible alternatives. In theory, they may improve posterior exploration, but their practical performance across different posterior distributions still lacks systematic comparison. 

This study constructed a reproducible and standardized benchmark to systematically compare NUTS and three PDMP samplers: the ZigZag sampler, the Bouncy Particle Sampler (BPS), and the Boomerang sampler. We selected a posterior set from an open database and classified it into three posterior complexity groups: simple, moderate, and complex. We evaluated algorithm performance from two aspects: accuracy and efficiency. 

The results showed that complex posteriors usually had worse accuracy and efficiency. NUTS performed best, while BPS often performed worse and had larger variability on some posteriors. ZigZag and Boomerang usually performed between NUTS and BPS, but the difference between them became clearer for complex posteriors. 

Overall, in this benchmark, the theoretical advantages of PDMP samplers did not necessarily translate into better practical performance. Posterior complexity affected sampler performance, but a single condition number could only reflect part of this complexity. By using the same posterior set, performance metrics, and evaluation procedure, this study provides comparative evidence on NUTS and PDMP samplers and offers a reusable benchmark framework for evaluating future algorithms.