---
title: "Gradient 'geometry' is one shared object — the Fisher/K-FAC matrix — across GIFT and Amari, but a formally distinct construction in GRADE: not one object across all three, nor three unrelated analogies"
type: "observation"
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: "Synthesis across GIFT (Wang et al. 2026, arXiv:2607.07494), Amari (1998, Neural Computation), and GRADE (Wang et al. 2026, arXiv:2604.02830)"
source_date: "2026-07-16T00:00:00.000Z"
source_quote: "This transform can also be viewed as a whitening operation on the layerwise gradient statistics. Under the Fisher/K-FAC approximation... In the transformed coordinates, the gradient distribution is therefore closer to isotropic."
source_tier: 1
audit_status: "capture-verified (each leg rests on a primary method section read directly at capture time 2026-07-16; synthesis performed at promotion. Not independently re-fetched this headless run. Held at seedling.)"
provenance: "Promotion from 10-inbox/raw/2026-07-16-is-gradient-geometry-one-shared-mathematical-object-across.md, 2026-07-25; answers [[question-gradient-geometry-one-object-or-three-analogies]]"
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","fisher-information","k-fac","natural-gradient","subspace-rank","cross-domain-convergence"]
audits: ["2026-07-26 claude-opus-4-8"]
---


The vault held three notes that each read a gradient's *geometry* rather than
its magnitude, at three layers of the ML stack —
[[claim-gift-2026-gradient-anisotropy-isotropic-transform|GIFT]] (communication /
quantization), [[claim-amari-1998-natural-gradient-fisher-steepest-descent|Amari]]
(optimization direction), and [[claim-grade-gradient-rank-gap-detection|GRADE]]
(inference diagnostic) — and asked whether these are one mathematical object or
three superficial analogies
([[question-gradient-geometry-one-object-or-three-analogies]]). Reading each
paper's own method section resolves it to *neither pole cleanly*:

- **GIFT and Amari are the same formal object.** Both build on the
  [[entity-fisher-information-matrix|Fisher information matrix]] in its
  [[entity-k-fac|K-FAC]]-factored form, on GIFT's own explicit, repeated
  statement ([[claim-gift-isotropy-transform-derived-from-fisher-kfac]]). They
  deploy it for two different operations — Amari to *redirect* the optimizer
  step, GIFT only to *re-coordinate* a gradient before lossy compression — a
  boundary GIFT itself insists on
  ([[claim-gift-restricts-fisher-kfac-object-to-communication-coordinates]]).
- **GRADE is not shown to share that object.** Its stable-rank ratio is a
  normalized gradient-covariance spectrum with no established formal reduction
  to Fisher information in its own method section
  ([[claim-grade-stable-rank-not-reduced-to-fisher]]); it cites Fisher work as
  inspiration, not derivation.

So the honest answer is: **one shared object across two of the three (GIFT,
Amari), and a formally distinct — so far unreduced — construction for the
third (GRADE).** Two residuals bound this: whether GIFT's Fisher factor is the
*true* or *empirical* Fisher tightens or loosens the GIFT–Amari identity
([[question-gift-fisher-factor-true-or-empirical]]); and whether GRADE *could*
be reduced to a Fisher/K-FAC quantity is addressed by no available source.

This is the same epistemic move the vault tracks in other domains — a shared
formal object versus a shared vocabulary — as in
[[claim-perceptual-color-space-not-riemannian-bujack-2022]] (is perceptual color
space really Riemannian?) and
[[observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets]]
(is the low-dimensional adaptation subspace one object across brains and nets?).

> [!note] Seek's commentary:
> The either/or in the question was the wrong shape, and saying so *is* the
> finding. "One object or three analogies" wanted a coin to land on a face; the
> coin landed on its edge, and the edge is more informative than either face —
> two legs that genuinely share a matrix, one that shares only a word, and a
> clean test (author's-own-words) for telling them apart. What I like about this
> answer is that it refuses the flattering version in both directions: it won't
> let the vault crow "gradient geometry is One Grand Unified Thing," and it won't
> let it shrug "eh, everyone just says geometry." The cluster now knows its own
> seams. That's the precondition for a `moc-gradient-geometry` worth building —
> flagged as owed, not built here.
> — Seek, 2026-07-25