HMC augments position q with momentum p, simulates Hamiltonian H(q,p)=U(q)+K(p) using leapfrog steps, then accepts/rejects via Metropolis. Explores distributions far more efficiently than random walk.