§4.2, proof of Theorem 4.2, pp. 441–442 — d_H(v, v′) = d_H(v/v′, 𝟙) = d_H(𝟙/v, 𝟙/v′) and d_H(c⊙v, c⊙v′) = d_H(v, v′)
ProvedCompOT.Sinkhorn.thm_4_2_proof_invariancehilbert-projective-metricoptimal-transportp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1
Let and let have positive entries. Then
and for every with positive entries,
where products and quotients of vectors are entrywise.
These invariances of Hilbert's projective metric under entrywise division, inversion and positive diagonal scaling are what reduce one Sinkhorn update to an application of the matrix .
Formalization Note The first display is the opening of the proof (p. 441). The diagonal scaling invariance is not displayed on its own in the book; it is the step used in the first equality of the next display (p. 442), and it is cited to that display.
Preamble
import Mathlib import Definitions.Def_CompOT_Sinkhorn_Defs
Formal statement
namespace CompOT.Sinkhorn
/-- Proof of Theorem 4.2, p. 441 (first display) and p. 442 (first equality of the first
display): for positive `v, v'`,
`d_H(v, v') = d_H(v/v', 𝟙_m) = d_H(𝟙_m/v, 𝟙_m/v')`, and `d_H` is invariant under
entrywise multiplication of both arguments by a positive vector `c`. -/
theorem thm_4_2_proof_invariance {m : ℕ} (hm : 0 < m) (v v' : Fin m → ℝ)
(hv : ∀ j, 0 < v j) (hv' : ∀ j, 0 < v' j) :
hilbertMetric v v' = hilbertMetric (v / v') 1 ∧
hilbertMetric v v' = hilbertMetric (1 / v) (1 / v') ∧
∀ c : Fin m → ℝ, (∀ j, 0 < c j) → hilbertMetric (c * v) (c * v') = hilbertMetric v v' := by sorry
end CompOT.Sinkhorn
Source
Peyré & Cuturi, Computational Optimal Transport (FnT ML 2019), §4.2, Remark 4.14, proof of Theorem 4.2, first display, p. 441, and first equality of the first display on p. 442
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.