org.replikativ.foerster.weights — org.replikativ/foerster 0.1.136

Shared population mass accounting. Normalized selection weights and total mass are separate; neither asserts posterior accuracy or unbiased evidence. A zero-mass population has no posterior and cannot be drawn or resampled.

Two drivers keep the ledger: the population-path interpreter (population/path: tempered SMC, θ-vector populations, population programs) holds a mass state (from-log, reweight, selected, resampled); the savepoint particle drivers (foerster.inference.smc: SMC, importance sampling, streaming, batched and conditional SMC, PGAS, and through them PIMH, particle Gibbs and IPMCMC), whose raw weights live in their particles' worlds, account each barrier with barrier. Both fold the same log mass: the normalizer before plus the log mean weight.

Drawing from weights is here too: sample-categorical and the resampling schemes (resample), which read the current generator (foerster.random).

barrier clj/s

(barrier xs log-normalizer)

The accounting of a savepoint particle population at a barrier: xs its raw log weights since its last resampling, log-normalizer the log mass folded in before them. Returns

:selection normalized selection weights (normalize; uniform among the impossible or the infinite, so that a degenerate population can still be resampled) :ess ess of the selection :log-mean the barrier's factor of the evidence, log mean exp of xs :log-mass the population's log mass, log-normalizer + :log-mean (from-log's), which a resampling folds into the normalizer as its children restart at equal weight :reset the log weight the children restart from: 0, or −∞ when the whole population is impossible (it stays impossible)

ess clj/s

(ess ws)

The effective sample size 1/Σw² of normalized selection weights ws.

from-log clj/s

(from-log xs)

(from-log xs log-normalizer)

Canonical accounting from raw log weights and an accumulated log normalizer. log-mass is log-normalizer + logmeanexp(raw). A zero-mass state stays explicit.

log-mean-exp clj/s

(log-mean-exp xs)

log-sum-exp clj/s

(log-sum-exp xs)

multinomial-resample clj/s

(multinomial-resample weights n)

n indices drawn independently from the normalized weights.

normalize clj/s

(normalize xs)

(normalize xs {:keys [zero infinite]})

Linear selection weights. Zero or infinite total mass requires an explicit :zero/:infinite :uniform policy (legacy selection inside existing engines).

normalize-log-weights clj/s

(normalize-log-weights log-weights)

Legacy engine selection: equal mass among impossible candidates, or among infinite-weight candidates. This does not make a zero-mass posterior valid.

resample clj/s

(resample scheme weights n)

n ancestor indices from the normalized weights by scheme: :systematic (default), :stratified, :residual or :multinomial.

resampled clj/s

(resampled state n)

residual-resample clj/s

(residual-resample weights n)

Residual resampling: ⌊n·wᵢ⌋ copies of each particle, the rest drawn multinomially from the residual weights.

reweight clj/s

(reweight state increments)

sample-categorical clj/s

(sample-categorical weights)

One index drawn from normalized weights, through random/uniform01.

selection clj/s

(selection state)

stratified-resample clj/s

(stratified-resample weights n)

Stratified resampling: one uniform draw in each of the n strata [i/n, (i+1)/n). Lower variance than multinomial, never more than systematic's failure modes.

systematic-resample clj/s

(systematic-resample weights n)

Systematic resampling algorithm from Anglican.

Low-variance resampling for particle filters. weights: normalized weights (sum to 1) n: number of particles to sample

Returns vector of indices into original particle vector.

uniform clj/s

(uniform n)

(uniform n log-mass)

Equal selection weights at the given total log mass.