Stein-based Sampling

ksd_u_gof test: https://arxiv.org/abs/1602.03253

ksd_v_gof test: https://arxiv.org/abs/1602.02964

fssd_gof test: https://arxiv.org/pdf/1705.07673

svgd: https://arxiv.org/abs/1608.04471

stein_thinning: https://arxiv.org/abs/2005.03952

stein_point: https://arxiv.org/abs/1803.10161

stein_point_mcmc: https://arxiv.org/abs/1905.03673

stein_based_importance_sampling: https://arxiv.org/abs/1610.05247

---

[ Layer 0: Core Infrastructure & Inputs ]

(The foundation providing gradients and kernel evaluations)

• Inputs:
  • Unnormalized Target / Score Function
  • External MCMC Samples (e.g., from rstan, coda)
• Engines:
  • [✓ Existing] Unified SteinKernel Framework (IMQ, Gaussian RBF, Custom)

    ↓↓ *(Kernel infrastructure powers all downstream tasks)*

[ Layer 1: Sample Generation ]

(Obtaining samples from complex or unnormalized distributions)

• [✓ Existing] Deterministic Approximation: SVGD, Stein Points, Stein Points MCMC

    ↓↓ *(Generated samples or External samples flow here)*

[ Layer 2: Quality Diagnosis ]

(Evaluating convergence and goodness-of-fit)

• [✓ Existing] Exact Kernel Discrepancies: KSD (U/V-statistics), FSSD

    ↓↓ *(Evaluated samples ready for refinement)*

[ Layer 3: Sample Compression ]

(Distilling core samples by discarding burn-in and highly correlated states)

• [✓ Existing] Stein Thinning (Greedy & Herding variants with multivariate preconditioners)