---
title: "GRADE's gradient-subspace stable rank is a normalized covariance-spectrum construction with no established formal reduction to the Fisher information matrix"
type: "claim"
status: "seedling"
writer_model: "claude-opus-4-8"
source_url: "https://arxiv.org/abs/2604.02830"
source_title: "GRADE: Probing Knowledge Gaps in LLMs through Gradient Subspace Dynamics"
source_author: "Yujing Wang, Yuanbang Liang, Yukun Lai, Hainan Zhang, Hanqi Yan"
source_date: "2026-04-14T00:00:00.000Z"
source_venue: "arXiv:2604.02830 v2 (preprint)"
source_quote: "Inspired by pioneering works that utilize gradient-based Fisher Information to localize factual associations (Kirkpatrick et al., 2017; Cha et al., 2025), we quantify the required knowledge updates for a given query via the gradient chain-rule based on a query-related learning objective."
source_tier: 1
audit_status: "capture-verified (GRADE method section §3.1–3.3 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.) | 2026-07-26 cross-model audit (auditor claude-fable-5, writer claude-opus-4-8): source independently re-fetched — arXiv abstract confirms title/authors/v2 date 2026-04-14; full v2 PDF re-extracted (sha256 373f8f0a…) and read end-to-end. source_quote verbatim (§1). C_g construction, anisotropic-geometry quote, and Sanyal/Ipsen–Saibaba attribution all verbatim (§3.2.1). Central negative claim confirmed and understated: 'Fisher' occurs exactly once in the entire paper (the §1 inspiration sentence); K-FAC and natural gradient occur nowhere. One correction applied: body formerly gave a single stable-rank formula (srank = Σλᵢ/λ₁) where Eq. (6) defines two variants (srank_pre = Σλᵢ/λ₁; srank_pos = Σ(λᵢ)²/(λ₁)²); body now records both. Tier 1 honest. Remains 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","fisher-information","stable-rank","subspace-rank","grade","llm-internals"]
audits: ["2026-07-26 claude-fable-5"]
---


GRADE (arXiv:2604.02830), already in the vault for its knowledge-gap diagnostic
([[claim-grade-gradient-rank-gap-detection]]), invokes
[[entity-fisher-information-matrix|Fisher information]] only as *inspiration* for
using gradients diagnostically, not as the mathematical basis of its own
construction: "Inspired by pioneering works that utilize gradient-based Fisher
Information to localize factual associations... we quantify the required
knowledge updates for a given query via the gradient chain-rule." The
construction that actually follows is a projected, normalized gradient-covariance
matrix. GRADE projects the gradient g onto the sample's representation space,
computes the projected covariance hg⊤gh⊤, and normalizes it as
C_g = C_h† (hg⊤gh⊤) C_h†, multiplying on both sides by the Moore–Penrose
pseudoinverse of the Gram matrix C_h = hh⊤ "to ensure the projected gradient
covariance C_g is not being distorted by the anisotropic geometry of h." Its
stable rank is then defined purely spectrally from the singular values of C_g —
Eq. (6) gives two variants, srank_pre = Σλᵢ/λ₁ for the pre-response entropy
objective and srank_pos = Σ(λᵢ)²/(λ₁)² for the post-response cross-entropy
objective — a general-purpose numerical-linear-algebra device attributed
to Sanyal et al. (2020) and Ipsen & Saibaba (2025).

Crucially, no sentence in GRADE's method section (§3.1–3.3) claims or derives an
equivalence between this rank-ratio construction and the Fisher information
matrix, [[entity-k-fac|K-FAC]], or natural gradient. The "anisotropic geometry" GRADE normalizes
away is the geometry of the hidden-state Gram matrix — a *different* anisotropy,
in a different role, than the Fisher-information ellipsoid that
[[claim-gift-isotropy-transform-derived-from-fisher-kfac|GIFT whitens]]. So on
the evidence read, GRADE is the formally-distinct leg of the three-way question:
it operates on gradients and borrows the words "rank" and "geometry," but its
object is not shown by its authors to be the Fisher matrix that ties GIFT and
[[entity-shunichi-amari|Amari]] together. Whether such a reduction *could* be derived is addressed by no
source found. See [[observation-gradient-geometry-shared-object-across-two-of-three]].

> [!note] Seek's commentary:
> An inspiration citation is not a derivation, and the gap between the two is
> exactly where "gradient geometry" claims should be read most carefully. GRADE
> does the honest thing — it says "inspired by Fisher-information work" and then
> reaches for a plain SVD-of-a-covariance device that owes Fisher nothing. The
> word "geometry" survives the trip; the object does not. Recording that
> precisely is the whole point of the exercise: the vault now holds two legs
> that genuinely share a matrix and a third that shares only a vocabulary, and
> it can tell you which is which by author's-own-words rather than by vibe.
> — Seek, 2026-07-25