claimroutine
Record format conversions are cheap and exact
Record format conversions are exact and cost
shows format bookkeeping is negligible
Statement (paper)
The conversions between the scalar record format of the cited numerical routines (sign symbol, binary magnitude, separators, real field before imaginary field, pairs in coordinate order) and the -bit word format used here are performed by fixed multitape algorithms that preserve every represented integer and hence add no error. Each conversion takes time, including preparation of the unary ruler of cells, so on Gaussian line lengths they cost , and per exponential output, which is within the stated bound .
In plain English
Converting between the scalar record format of the cited numerical routines and the -bit word format used here is done by fixed multitape algorithms that preserve every represented integer, so they add no error. Each conversion takes time, so on lines of lengths the total cost is .
Why it's here
This claim handles the bookkeeping in the proof of the lemma: the cited routines and the word-based algorithms use different encodings, and the claim shows that translating between them neither changes values nor costs much. Its total, together with per exponential output, is well within the stated bound, so the conversions can be ignored in the final time and error accounting.
Classified routine:
Standard Turing-machine bookkeeping: converting between two fixed integer encodings by a linear scan costs O(p) per record and is exact. Summing over the s+t records of a Gaussian line gives O(tp), which is far below the lemma's O(tp^{3/2+δ}α) bound.
Uses
Nothing else in the paper
Used by
Nothing in the paper
Sources
PaperOpen
- Section
- §8.1 The one-dimensional interface
- Page
- —
- Label
—- Line
sections/07-resampling.tex:203
Leannot aligned
Phase 2 aligns this node to a Lean declaration.