Post-processing a POVM by a Markov kernel is a POVM:
ProvedBosonicLearn.fockPOVM_postprocessApplying a randomized classical map to the outcome of a measurement gives another measurement.
Let be an index set and let be a POVM on with outcomes in a measurable space . It consists of effects for measurable , with and , and of an outcome law for every density operator , a probability measure on given by the Born rule . Let be a Markov kernel from to a measurable space : is the probability that the post-processed outcome lies in when the measurement outcome is .
Then there is a POVM on with outcomes in such that for every density operator and every measurable
In terms of effects, .
Role. A measurement followed by a randomized estimator is thereby a single measurement whose outcome is the estimate. Together with BosonicLearn.adaptive_scheme_product_povm it shows that an adaptive single-copy learner is a special case of a collective one, which is how the proof of Theorem 3 applies the collective lower bound of Theorem 9 to single-copy schemes.
Formalization Note and are FockPOVMs: the effects are coefficient arrays whose quadratic forms on finitely supported vectors lie between and , and the Born rule is required on every density operator , with trace-class meant in the nuclear-expansion sense of IsDensity. The law of the post-processed outcome is Measure.bind. The identity is stated for every coefficient array ; on arrays that are not densities both sides are unconstrained data, so the content is the statement on densities. The source uses this fact without stating it.
Related platform material The space used here (Hilbert A) is the same Lean type as OAI.TensorLp.H A, and is the same type as OAI.EntropyPhotonNumber.Fock n. The published definitions OAI_Analysis_TensorLp (tensor products, slices and operator lifts on ) and OAI_MathematicalPhysics_Fock_Operators (OAI.CrossEnsemble.operator, the bounded operator of a family, with hasSum_operator and entry_hasSum) are sorry-free and can be imported. The proof of BosonicLearn.traceDist_le_one contains lemmas on nuclear expansions, their sums, and truncation to finite index sets.
import Mathlib import Definitions.Def_BosonicLearn_FockPOVM open MeasureTheory ProbabilityTheory open scoped ENNReal Matrix
namespace BosonicLearn
/-- **Classical post-processing of a POVM is a POVM.**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, Eq. (105), p. 40 ("the state estimator ρ̂(x) that allows internal randomness"); used in the proof of
Theorem 3, p. 67, where a single-copy scheme followed by its estimator is treated as one measurement.
Let `P` be a POVM on `ℓ²(I)` with outcomes in a measurable space `Ω`, and let `κ` be a Markov kernel from `Ω` to a
measurable space `Ω'` (a randomized map applied to the outcome). There is a POVM `P'` on `ℓ²(I)` with outcomes in
`Ω'` whose outcome law on every state is the law of the post-processed outcome:
`P'.law ρ = ∫ κ(x, ·) d(P.law ρ)(x)`.
**Formalization Note.** `P'` is a `FockPOVM`, so its effects are bounded between `0` and `1`, are normalized, and
satisfy the Born rule `P'.law ρ B = Re Tr(E'(B) ρ)` on every density operator `ρ`. The law of the post-processed
outcome is `Measure.bind`. The identity is stated for every coefficient array `ρ`; on arrays that are not densities
both sides are unconstrained data, so the content is the statement on densities. The source uses this without
stating it. -/
theorem fockPOVM_postprocess {I Ω Ω' : Type} [DecidableEq I] [MeasurableSpace Ω] [MeasurableSpace Ω']
(P : FockPOVM I Ω) (κ : Kernel Ω Ω') (hκ : IsMarkovKernel κ) :
∃ P' : FockPOVM I Ω', ∀ ρ : Matrix I I ℂ, P'.law ρ = (P.law ρ).bind κ := by sorry
end BosonicLearn