Eq. (65) — the target's marginal on the past registers is the learner's input
ProvedQIndInf.target_marginal_eqlearning-theoryp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1quantum-information
Let be a quantum source, instruments and a history with outcomes . Discarding the next register and its outcome from the prediction target gives the learner's input:
This is the first identity of Eq. (65) (stated on p.19 as "Its marginal on is exactly the learner's input state"). It is what lets the recovery inequality, applied to the channel , compare the learner's output with the target.
Differences from the source:
- generalization: ρ_{R1…Rm} ∈ C_{n,m}[s,t] → any quantum source (consistent family of density operators); reason: the identity uses only consistency of the source and trace preservation of the instrument; the paper's sources are consistent sources
- encoding: registers R_i of n qubits; outcome register written first in M_I → Fin (2^n) per register, register R_i = coordinate i−1; output indexed (past registers) × (next register × outcome); reason: relabelling of the index set; changes no trace, trace distance or relative entropy
- encoding: conditional states defined for histories of positive probability → 0⁻¹ = 0 makes them the zero matrix on histories of probability zero; reason: those histories have weight zero in every sum; only undefined cases are affected
Preamble
import Mathlib import Definitions.Def_QIndInf_target import Definitions.Def_WildeQIT_partialTrace set_option autoImplicit false
Formal statement
namespace QIndInf
open QAlgorithms.CopyMeasurements
/-- **The target's marginal on the past registers is the learner's input** (p.19 after
Eq. (55); first identity of Eq. (65)). For every quantum source `ρ`, every round with `j` past
registers, all instruments `I₁, …, I_{j+1}` and every history `x`,
`N_i(τ_{≤i|h<i}) = Tr_{R_i X_i} τ_{≤i|h<i} = ρ_{<i|h<i}`.
arXiv:2609.36912 (PDF of 2026-11-02), §4.1, p.19 ("Its marginal on (R₁, ..., R_{i−1}) is exactly
the learner's input state ρ_{<i|h<i}") and §4.2, Eq. (65), p.21.
**Formalization Note.** `N_i` is `WildeQIT.partialTraceRight` over the factor `R_i × X_i`.
Round `i` has `j = i − 1` past registers. On a history of probability zero both sides are the zero
matrix. -/
theorem target_marginal_eq {n j : ℕ} (ρ : QuantumProcess n) {k : Fin (j + 1) → ℕ}
(I : (l : Fin (j + 1)) → QuantumInstrument (Reg n) (Fin (k l)))
(x : (l : Fin j) → Fin (k l.castSucc)) :
WildeQIT.partialTraceRight (target ρ.state I x) = condPast ρ.state (pastInstr I) x := by sorry
end QIndInf
Source
Hiroka, Hsieh, Shirakawa, arXiv:2609.36912 (PDF of 2026-11-02), §4.1, p.19 (after Eq. (55)); §4.2, Eq. (65), p.21
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.