Coherent copying and uncomputation: exact error identity and square-root failure bound
ProvedInfoLossLAR.WorstCase.coherent_decoder_copy_errorcoherent-decodinginformation-loss-larlatticesquantum-computing
Let and be finite sets, let be an arbitrary amplitude vector, and let be the computed and desired classical answers. The register indexed by retains all workspace and any query label. Copy the answer into a fresh register, obtaining
For any complex-linear isometry into a seminormed complex vector space, if the failure mass satisfies , then
Normalization is unnecessary for the identity; for a normalized decoder state, is precisely its classical failure probability. Taking to be uncomputation establishes the semantic error estimate needed in the coherent-decoder step of the Gaussian LWE worst-case reduction. The statement does not assert an efficient implementation of or construct the Regev reduction.
Preamble
import Mathlib set_option autoImplicit false
Formal statement
theorem InfoLossLAR.WorstCase.coherent_decoder_copy_error
{α β E : Type*} [Fintype α] [Fintype β] [DecidableEq β]
[SeminormedAddCommGroup E] [NormedSpace ℂ E]
(U : EuclideanSpace ℂ (α × β) →ₗᵢ[ℂ] E)
(ψ : α → ℂ) (readout desired : α → β) (ε : ℝ)
(hfailure : (∑ x : α, if readout x = desired x then 0 else ‖ψ x‖ ^ 2) ≤ ε) :
let actual : EuclideanSpace ℂ (α × β) :=
WithLp.toLp 2 (fun p => if p.2 = readout p.1 then ψ p.1 else 0)
let ideal : EuclideanSpace ℂ (α × β) :=
WithLp.toLp 2 (fun p => if p.2 = desired p.1 then ψ p.1 else 0)
‖U actual - U ideal‖ ^ 2 =
(2 * ∑ x : α, if readout x = desired x then 0 else ‖ψ x‖ ^ 2) ∧
‖U actual - U ideal‖ ≤ 2 * Real.sqrt ε := by sorrySource
Auxiliary formalization of the coherent compute-copy-uncompute estimate used in Chen Qian, Information Loss Is All You Need, arXiv:2611.05136v1, Appendix I.1, pp. 132–133. https://arxiv.org/html/2611.05136v1