[Paper Review] A Conceptual Introduction to Hamiltonian Monte Carlo

Michael Betancourt|arXiv (Cornell University)|Jan 10, 2017

Markov Chains and Monte Carlo Methods21 references548 citations

TL;DR

The paper provides a principled, intuition-focused account of Hamiltonian Monte Carlo (HMC), its theoretical foundations, practical implementation, robustness, and diagnostics, aimed at making the method accessible beyond differential geometry.

ABSTRACT

Hamiltonian Monte Carlo has proven a remarkable empirical success, but only recently have we begun to develop a rigorous understanding of why it performs so well on difficult problems and how it is best applied in practice. Unfortunately, that understanding is confined within the mathematics of differential geometry which has limited its dissemination, especially to the applied communities for which it is particularly important. In this review I provide a comprehensive conceptual account of these theoretical foundations, focusing on developing a principled intuition behind the method and its optimal implementations rather of any exhaustive rigor. Whether a practitioner or a statistician, the dedicated reader will acquire a solid grasp of how Hamiltonian Monte Carlo works, when it succeeds, and, perhaps most importantly, when it fails.

Motivation & Objective

• Motivate why high-dimensional expectations are hard and how typical set geometry guides sampler design.
• Develop an intuition for Hamiltonian dynamics as a way to explore the typical set efficiently.
• Bridge theory and practice by outlining practical implementation choices (symplectic integrators, tuning) and diagnostics.
• Explain when Markov chain Monte Carlo (MCMC) methods succeed or fail, and how HMC addresses common pathologies.

Proposed method

• Introduce geometric intuition for high-dimensional spaces and the typical set to motivate efficient sampling.
• Derive and describe the Hamiltonian framework as a mechanism to move coherently through the typical set.
• Present the idealized Hamiltonian Markov transition and its connections to Metropolis-Hastings style corrections.
• Discuss practical implementation aspects including symplectic integrators and error correction.
• Outline diagnostics and robustness considerations, including geometric ergodicity and the split R-hat statistic.

Experimental results

Research questions

• RQ1 What geometric properties of high-dimensional distributions define the challenging typical set for sampling?
• RQ2 How does Hamiltonian dynamics provide coherent, large-move exploration of the typical set compared to diffusive MCMC methods?
• RQ3 What are the practical components and tuning choices (like kinetic energy, integration time, symplectic integrators) that make HMC effective?
• RQ4 How can diagnostics detect and diagnose failures or pathologies in HMC (e.g., regions of high curvature, inappropriate kinetic energy)?

Key findings

• Hamiltonian Monte Carlo leverages phase-space dynamics to explore the typical set more efficiently than random-walk approaches.
• The effectiveness of HMC depends on choosing appropriate kinetic energy structure and integration parameters, and on using symplectic integrators with corrections.
• Diagnostics such as split R-hat are essential for diagnosing non-ergodicity or poor mixing and for signaling potential pathologies.
• The No-U-Turn Sampler (NUTS) and related practical implementations (e.g., in Stan) operationalize these ideas for robust usage.
• Pathologies arise when transitions interact poorly with target geometry, highlighting the need for robust tuning and diagnostics.