# Hamiltonian Monte Carlo Leapfrog: Practice Notes

Hamiltonian Monte Carlo Leapfrog shows up often in coding rounds. Here is how I talk through it when the interviewer is watching.

## Key idea

Simulate Hamiltonian dynamics with leapfrog steps using a gradient of the log-density, then accept or reject the endpoint with a Metropolis test.

## Steps to say out loud

1. Restate the input and the exact ask, and nail down constraints and overflow.
2. Name the pattern (binary search on the answer, monotonic stack or deque, interval / interval-DP, string hashing, graph traversal, or shortest paths) and why a nested scan or naive search fails.
3. Code it, then trace a short example plus one empty or already-solved edge case.

## Complexity

Each proposal costs L gradient evaluations; mixing is typically much faster than random-walk Metropolis in high dimension.

## Common mistake

Using a non-reversible integrator (or a single huge step) so the Metropolis test rejects almost every proposal.

## Practice tip

Time yourself: two minutes to explain, ten to code, three to test edges. Disclosure: I work on Linkjob AI. I also keep a practice cheat sheet at multilingual AI interview assistant.