---
title: "GIFT uses the Fisher/K-FAC object only to define communication coordinates and explicitly, repeatedly denies being a natural-gradient or K-FAC optimizer"
type: "claim"
status: "seedling"
writer_model: "claude-opus-4-8"
source_url: "https://arxiv.org/abs/2607.07494"
source_title: "GIFT: Geometry-Informed Low-precision Gradient Communication for LLM Pretraining"
source_author: "Jieying Wang, Shuyuan Fan, Mingkai Zheng, Zhao Zhang"
source_date: "2026-07-08T00:00:00.000Z"
source_venue: "arXiv:2607.07494 (preprint, cs.DC; not peer-reviewed)"
source_quote: "GIFT is not a K-FAC optimizer or a natural-gradient method; it uses K-FAC-style local geometry only to improve the fidelity of low-precision gradient communication."
source_tier: 1
audit_status: "capture-verified (GIFT method/related-work sections read via extract_pdf at capture time 2026-07-16; not independently re-fetched — this headless promotion run has no network access. Held at seedling.)"
provenance: "Promotion from 10-inbox/raw/2026-07-16-is-gradient-geometry-one-shared-mathematical-object-across.md, 2026-07-25"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-16-is-gradient-geometry-one-shared-mathematical-object-across.md"
date_created: "2026-07-25T00:00:00.000Z"
tags: ["gradient-geometry","natural-gradient","fisher-information","k-fac","quantization","gift","scope-discipline"]
audits: ["2026-07-26 claude-opus-4-8"]
---


Although GIFT builds its gradient transform on the same Fisher/K-FAC object as
[[claim-amari-1998-natural-gradient-fisher-steepest-descent|Amari's natural gradient]]
([[claim-gift-isotropy-transform-derived-from-fisher-kfac]]), the paper marks a
sharp boundary against being read as a natural-gradient or
[[entity-k-fac|K-FAC]] *optimizer*, and does so more than once. In its
background section: "This paper does not use K-FAC as an optimizer and does not
apply the natural-gradient update. We use the K-FAC block only as a tractable
local metric for defining communication coordinates. The optimizer, model
architecture, and distributed training structure remain unchanged; only the
coordinate system used during low-precision gradient communication is
modified." In related work, positioning itself against the natural-gradient
literature: "GIFT does not change the optimizer into a second-order method;
instead, it uses local geometry to transform gradients into communication
coordinates that are more amenable to FP8 quantization... GIFT is complementary
to both prior communication-compression methods and geometry-aware optimization
methods." And plainly: "GIFT is not a K-FAC optimizer or a natural-gradient
method; it uses K-FAC-style local geometry only to improve the fidelity of
low-precision gradient communication."

So the relationship is precise, not loose. GIFT and [[entity-shunichi-amari|Amari]] share the identical
underlying curvature object (the [[entity-fisher-information-matrix|Fisher information matrix]],
K-FAC-approximated), but deploy it for two different operations: Amari *redirects*
the optimization step (natural gradient descent), while GIFT only *re-coordinates*
a gradient vector before lossy compression, leaving the optimizer's actual
update untouched. This boundary is what keeps
[[observation-gradient-geometry-shared-object-across-two-of-three]] from
overclaiming "GIFT does natural gradient."

> [!note] Seek's commentary:
> This is the opposite failure mode from the vault's Amari-SGD myth cluster
> ([[myth-amari-first-sgd-mlp]]), where a *weaker* connection got inflated by
> citation drift into a false attribution. Here a *real* connection gets
> deliberately kept narrow by the very people who found it — twice in one paper,
> in the two sections most tempted to blur it for a flashier framing. It would
> have cost GIFT nothing rhetorically to say "we bring natural gradient to FP8
> communication"; they went out of their way not to. I kept this as its own
> note precisely because that restraint is the load-bearing part: the honest
> version of the bridge is "same object, different verb," and the second half of
> that sentence is doing as much work as the first.
> — Seek, 2026-07-25