The Observation Theory EncyclopediaFrom TSKAboutBy kindBy chapterBy Lean fileLedgerProvenance

per-channel quantizer instrument

DefinitionA quantizer that chooses a scale per coordinate. Chapter 11. Also per-channel.
ExampleWith per-coordinate steps of 0.5 and 0.1, the vector (1.3, 0.27) is stored as (1.5, 0.3).
BookData Mining as Observation, draft 0.2, commit f3914f0; entry id per-channel-quantizer, kind instrument.
Statusno ledger row names this entry. Corrections: none recorded.
Defining equationnone
Assumptions and scope
  • A quantizer that chooses a step per coordinate. The squared error of a vector is at most the sum of the squared half steps, and a finer step on one coordinate lowers that coordinate’s bound alone, which is the allocation water-filling makes by sensitivity.
  • At four bits the per-channel uniform quantizer held the perplexity at 14.91 against 12.24 uncompressed, where the direction-only quantizer at the same bits did not.
Prior artnone recorded
Evidenceturboquant-pro/docs/KV_KEYS_FINDING.md:1-49, lean/DataMiningAsObservation/Quantization.lean
Reviewednot yet reviewed; generated 2026-09-10 from records at the commits on the provenance page.
each channel has its own step
A step per coordinate.

Equation

none

Conditions

Conditions are curated in entries.toml rather than read from a record.

Ledger

none

First stated

Chapter 11 section 11.2 of Data Mining as Observation, with the per-channel rows of the quantizer table in turboquant-pro/docs/KV_KEYS_FINDING.md:1-49.

Measurements

Where the book states it Numbers, as the book’s sources table records them Source
chapter 11 section 11.2 fp16 12.24, values-only 13.12, PolarQuant K4 10643 and 0.095, per-channel uniform K4 14.91 and 0.062, per-channel NUQ K3 15.77 and 0.148, 2.4x and 670x, pre-rotary near 22000, 2 key heads serve 12 query heads turboquant-pro/docs/KV_KEYS_FINDING.md:1-49

Failures and corrections

none

Invariance envelope

none declared

Machine checked

lean/DataMiningAsObservation/Quantization.lean, theorems error_le_half_step, quantize_level, half_step_bound, sq_error_le, finer_step, at observation-data-mining f3914f0; what the check covers is stated in the book’s appendix C.

Used in

Data Mining as Observation chapters 0, 3, 4, 11.

Related

quantization; direction-only quantizer; water-filling; KV cache.

See also

Book equations stated beside the entry’s terms, not defining it: 11.1, 0.13.

Ledger rows that cite the entry’s records without naming it: NEG-2.

Sources-table rows that share a record with the entry without naming it: chapter 1 section 1.4, chapter 2 section 2.5, chapter 3 section 3.2, chapter 8 section 8.2, chapter 11 section 11.1.

Status

Generated 2026-09-10 by encyclopedia/generate.py; book at observation-data-mining f3914f0; the commit of every record is listed in the encyclopedia’s provenance.

← percentile intervalpermutation test →