Reimplement particle MCMC (SMC and PG) natively, removing AdvancedPS - #2853

Draft

yebai wants to merge 58 commits into

mainfrom

Commits on Jul 21, 2026

Replace AdvancedPS with a native particle MCMC implementation

committed
51549c8
Consolidate particle MCMC files, drop redundant fork step, add notes

committed
a2eb174
Default to stratified resampling

committed
3e0960b
Drop redundant prior/Jacobian accumulators from particles

committed
0620231
Reuse surviving parents on resample to cut allocations

committed
5d9d127
Clarity pass: high-level design note, sectioning, key-step notes

committed
ad5979b
Document particle-MCMC rewrite in HISTORY; add RNG-respected tests

committed
48e112a
Port extra particle-MCMC tests from PR 2848 (ck/smc)

committed
22daced
Note that resampler scheme types are not exported

committed
4789428
Minimise: drop unused step-counter generality

committed
cd6444a
Minimise: drop redundant threadsafe check in SMC step

committed
b904713
Derive numeric types instead of hardcoding Float64

committed
499b61a
Move import Random123 into particle_mcmc.jl

committed
87c0bb7
Release the particle MCMC rewrite as 0.47.0

committed
bcae152
Fix method ambiguity in TracedRNG seed!

committed
0223944
Make RNG seed splitting version-stable and decorrelated

committed
ad08bbf
Cite Andrieu, Doucet & Holenstein (2010) in the particle MCMC module

committed
716cc1e
Stabilise the gdemo CSMC+ESS test against seed variation

committed
541bc2a
Add opt-in within-sweep particle threading to SMC and PG

committed
160ab54

Commits on Jul 23, 2026

Keep particle log-score consistent with the likelihood accumulator

committed
75d3c0b
Rename ParticleMCMCContext to SMCContext

committed
50d2f40
Parametrise Particle on its RNG and weight types

committed
7247b13
Document why particle assume always draws from the prior

committed
39ab78e
Reuse Particle as the PG state (const PGState = Particle)

committed
8f6f700
Clarify particle parallelism: rename threaded to multithreaded, separate the two levels

committed
f80dc16
Move project instructions into AGENTS.md, include from CLAUDE.md

committed
280ade9
SMC: bundle the population directly, drop the step-loop, PGState, and weighting

committed
564b471
Add per-observation ESS to SMC output; rename evidence stat to log_normalizing_constant

committed
99a2c2b
Fix Inference.jl test for the log_normalizing_constant rename

committed
186282c
Rename resampler types with a Resampler suffix

committed
b55ea5a
Pin the CSMC reference to the retained trajectory by value

committed
96c11a5
Guard stratified/systematic resampling against weight undersum

committed
14bc29d
Warn when SMC is given initial_params

committed
1a917bb

Commits on Jul 24, 2026

Explain why the CSMC reference reproduces by value, with the math

committed
bf9357d

Commits on Jul 28, 2026

Finish the resampler rename in HISTORY.md

committed
118348b
Detect a changed reference execution trace instead of silently resampling it

committed
58a4e0a
Draw conditional-sweep ancestors multinomially, whatever scheme is named

committed
43f4694
Report which sampling keywords SMC actually honours

committed
8e74b4c

Commits on Jul 29, 2026

Let ess_per_step follow LogProbType instead of pinning it to Float64

committed
ac77e43
Build a CSMC reference from the retained particle, not from loose values

committed
8665e57
Record the measured CSMC+ESS margin, which is thinner than the comment claimed

committed
bee0e0c
Stop AGENTS.md claiming the particle samplers wrap AdvancedPS

committed
e8ef7a0
Delegate the accloglikelihood!! overload instead of reimplementing it

committed
f7af068
Assert MoGtest posterior means, not idealised cluster labels

committed
542b85c
Mention the reference-replay branch in the SMCContext docstring

committed
861d1ce
Skip Aqua's persistent_tasks check on Windows, where it is flaky

committed
7b12f5f
Drop the reference's seed-replay machinery, which value-pinning made dead

yebai
and
claude
committed
6c8247d
Rename ESSResampler and the weight-ESS helper away from the ESS sampler's name

yebai
and
claude
committed
9efc184
Say that PG's log_normalizing_constant is biased, and test that it is

yebai
and
claude
committed
2ff6b22
Check the particle samplers against two models with tractable posteriors

yebai
and
claude
committed
92b3638
Produce the particle weight from the accumulator, dropping the type piracy

yebai
and
claude
committed
8390a7c
Tidy the particle samplers and their tests

yebai
and
claude
committed
bbfc692
Derive the fixed-parameter test models, and give the files a header hierarchy

yebai
and
claude
committed
3f822ff
Drop the dead normalized_weights import from the particle tests

yebai
and
claude
committed
5cd20ea
Correct two comments that misdescribe how particles carry state

yebai
and
claude
committed
9dce003
Cover latents whose dimension varies between executions

yebai
and
claude
committed
e4628df
Widen the varying-dimension tolerance to allow for chain autocorrelation

yebai
and
claude
committed
6535094
Fold the normalizing-constant caveat into its bullet, for the formatter

yebai
and
claude
committed
8663adf