Last-customer rule for
ProvedNestedCRP.last_customer_rulebayesian-nonparametricsp2o-gran-per-chapterp2o-plan-paperp2o-v1topic-models
Let be trees with the same restriction to the tourists other than . Then
where is the mass of the path of when arrives last.
So, as a function of the path of , the nested CRP mass is proportional to , which is the prior on paths in Eq. (4).
Differences from the source:
- reading: p(c_d | c_{-d}, γ) "as the prior on paths implied by the nested CRP" → the prior on paths is the last-arrival mass, and ncrp is proportional to it over trees with the same restriction; reason: the conditional of the law under exchangeability; this is the only meaning of the phrase
- encoding: a conditional probability → a cross-multiplied identity; reason: equivalent since all masses are positive; no division appears
Preamble
import Mathlib import Definitions.Def_NestedCRP_ncrpMass
Formal statement
namespace NestedCRP
/-- The last-customer rule: for trees with the same restriction to the tourists other than `d`,
the ratio of their nCRP masses is the ratio of `d`'s path masses as the last arrival. This is
`p(c_d | c_{-d}, γ)` as "the prior on paths implied by the nested CRP".
Blei–Griffiths–Jordan, *The nested Chinese restaurant process and Bayesian nonparametric
inference of topic hierarchies*, J. ACM 57(2), Article 7 (2010), §5.2, p.7:17. -/
theorem last_customer_rule {D : ℕ} {γ : ℝ} (hγ : 0 < γ) (d : Fin D) (T T' : Tree D)
(h : T.AgreesOff d T') :
ncrpMass γ T * lastMass γ T' d = ncrpMass γ T' * lastMass γ T d := by sorry
end NestedCRP
Source
Blei–Griffiths–Jordan, The nested Chinese restaurant process and Bayesian nonparametric inference of topic hierarchies, J. ACM 57(2), Article 7 (2010), https://doi.org/10.1145/1667053.1667056, §5.2, p.7:17
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.