PROPOSAL
DISTRIBUTION
Machine Learning
IIIy CSE IDP and IDDMP
Refer : Stephen
Marsland
Motivation
1. Direct sampling from p(x) is often computationally
expensive or infeasible.
2. We can evaluate an un‐normalised version p'(x) but cannot
draw samples easily.
3. Introduce a simpler, easy‑to‑sample “proposal” distribution
q(x).
4. Goal: use samples from q(x) to approximate samples from
p(x) without altering notations.
Proposal Distribution
1. We assume
with unknown normalisation constant Zp
2. Choose q(x) such that for all x:
where M is a finite constant (the “envelope” factor).
3. Sampling procedure: draw x∗∼q(x), then decide acceptance
via a uniform test under M q(x∗)
The Rejection Sampling
Algorithm
Rejection Sampling Example & Transition to
Importance Sampling
Example (Mixture of Gaussians): (Refer next slide for graph)
Proposal q(x) uniform (dotted line).
With M=0.8, ≈50% of samples are rejected. With M=2, ≈85% of samples are rejected.
Issues:
A) High rejection rates waste computation. B)Curse of dimensionality exacerbates
inefficiency.
Two Remedies:
Develop more sophisticated exploration methods. Bias sampling toward high‑probability
regions.
Importance Sampling:
Assign each draw x(i)∼q(x) a weight
Allows estimation of expectations without discarding samples.
Resampling:
Use weights in a Sampling‑Importance‑Resampling scheme to obtain a representative
set of points.
Rejection Sampling Example & Transition to
Importance Sampling
The Sampling-Importance-Resampling
Algorithm
Summary And Conclusion
Challenge: Direct sampling from complex p(x) is often infeasible.
Rejection Sampling: Uses a simpler proposal q(x) and an envelope M q(x) to
filter samples.
Easy to implement but can discard many draws, especially in high dimensions.
Improvements:
Better exploration strategies (local moves, MCMC).
Importance sampling to weight rather than reject.
Importance Sampling:
Assigns importance weights to all samples, avoiding waste.
Enables weighted estimation and resampling (SIR).
Takeaway:
Choice of q(x) and M is critical for efficiency.