Skip to content

Sampling

These examples draw samples from a fixed, known target. Nothing is learned; the subject is the dynamics and the numerics.

  • MCMC: Langevin dynamics and Hamiltonian Monte Carlo, with chains as a batch dimension, so one call runs thousands of them.
  • Integrators: the numerical engines inside the samplers, with their order of accuracy measured against an exact solution.
  • Flow: continuous-time generation, where FlowSampler integrates a velocity field as an ODE or an SDE using those same integrators.

Start with Langevin Dynamics 101; the integrator and flow examples show the numerics that both MCMC and generative sampling stand on.

Theory: Sampling and Integration.

Next: Training, where the target is learned from data.

Example Summary Level
Langevin Dynamics 101 Sample a 2D energy with Langevin; trade step size against noise. intro
HMC 101 Hamiltonian Monte Carlo on a correlated Gaussian; trade leapfrog steps against gradient cost and chain autocorrelation. intro
Parallel Chains 10,000 vectorized Langevin chains in one call; population statistics without long single-chain runs. intro
Integrator Comparison Euler, Heun, and RK4 against the exact solution of a harmonic oscillator; measure order of accuracy directly. intermediate
FlowSampler ODE 101 Integrate a closed-form velocity field from noise to a Gaussian target; study steps vs fidelity without training. intermediate