claimroutineinferred_dep
Guarded evaluation with single final truncation
Guard width is , so one final truncation gives error
gives the error bound and output format
Statement (paper)
The guard width is uniformly over the allowed , so every intermediate coefficient of the exact evaluation of is represented exactly in the stated width. Every row of has absolute row sum one, so lies in the complex unit disk and so does . The single final truncation satisfies , and the truncated numerators have magnitude at most , so they fit in the supplied component width .
In plain English
The guard width is uniformly over allowed , so all intermediate coefficients of the exact evaluation of fit exactly in width. Since has absolute row sums one, and lie in the unit disk, the single final truncation has error below in sup norm, and the truncated numerators are at most so they fit in width .
Why it's here
Having converted the computed exact -layer into , this step controls precision: because and , the guard width is negligible next to , so nothing is rounded until the one final truncation. This gives the proposition its error bound and shows the output format fits in the supplied component width.
Classified routine:
Direct precision bookkeeping: since εC_1<1 and d ≤ b_d p^ε, the guard width is o(p). The unit row-sum bound and the single componentwise rounding estimate (error √2·2^{-p}) follow from elementary norm inequalities and the earlier exact-evaluation claim.
Uses
Used by
Nothing in the paper
Sources
PaperOpen
- Section
- §5.4 An explicit guard-width bound
- Page
- —
- Label
—- Line
sections/05-layers.tex:979
Leannot aligned
Phase 2 aligns this node to a Lean declaration.
Friction
- infoinferred_depoutline: "Exact guarded evaluation and single truncation error" rests on claim-prop-simultaneous-layer-3