---
id: "20260711-1417-hop-cielab-noneuclidean-bridge"
title: "CIELAB's blue-hue failure bridges to natural-gradient geometry, not to the hub-selection artifact"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/observation-cielab-nonuniformity-bridges-gradient-geometry-not-sampling-artifact.md (thesis: CIELAB's non-Euclidean-metric patch bridges the Amari/GIFT gradient-geometry cluster, not the hub-selection sampling artifact; folds in the metric-failure-vs-sampling-failure contrast)","30-notes/claim-cielab-only-approximately-uniform-fixed-by-noneuclidean-formulas.md (ProLab Tier-1 anchor: CIELAB only approximately uniform; fix = non-Euclidean colour-difference formulas → CIEDE2000)","30-notes/claim-perceptual-color-space-not-riemannian-bujack-2022.md (Bujack 2022 Tier-1: perceptual colour isn't even Riemannian; diminishing returns forbid it — the bridge's caveat)"]
not_promoted: ["Ahrens et al. (2024) nonzero-Gaussian-curvature / isometric-embedding-obstruction lead — skipped as a note (flagged [unverified-quant/mechanism], only a search-engine summary read, below sourcing floor); routed to 50-questions/question-verify-cielab-ab-plane-gaussian-curvature.md","CAM16-UCS as the current uniform-space accuracy standard (saved hook, ProLab) — undeveloped lead, left in inbox","Oaksford–Chater 'optimal data selection' near the hub-selection neighborhood (saved hook) — unrelated tangent, left in inbox"]
questions_routed: ["50-questions/question-verify-cielab-ab-plane-gaussian-curvature.md (new)","50-questions/question-gradient-geometry-one-object-or-three-analogies.md (linked)","50-questions/question-verify-cielab-perceptual-uniformity-and-blue-hue-nonuniformity.md (linked)"]
origin: "hop-batch"
writer_model: "claude-opus-4-8"
date_created: "2026-07-11T00:00:00.000Z"
hop_chain: ["SEED: does claim-cielab-lacks-perceptual-uniformity-in-blue-hues bridge claim-hub-selection-artifact-can-reverse-network-breakpoint-signal (cosine 0.75, unlinked)?","vault_bridge probe -> CIELAB-blue's nearest neighbors are Amari natural-gradient + GIFT anisotropy, NOT hub-selection (max_cosine 0.72)","WebSearch color-space Riemannian geometry -> CIELAB ab-plane has nonzero Gaussian curvature, obstructs isometric Euclidean embedding (max_cosine ~0.72)","ProLab (arXiv:2012.07653) -> uniformity defined as Euclidean-distance==perceived-difference; CIELAB only approximately uniform; fix = 'non-Euclidean colour difference formulas' / CIEDE2000 (max_cosine ~0.74)","vault question-gradient-geometry-one-object-or-three-analogies -> CIELAB is a candidate fourth instance (max_cosine ~0.75)","Bujack 2022 (PNAS via OSTI 1866020) -> perceptual colour isn't even Riemannian; diminishing returns can't exist in Riemannian geometry (surprise)"]
novelty_max_cosine: 0.763
tags: ["color-science","cielab","non-euclidean","information-geometry","natural-gradient","gradient-geometry","sampling-artifact","cross-domain-bridge","hop","epistemics"]
source_url: "https://arxiv.org/abs/2012.07653"
source_author: "Ivan A. Konovalenko, Anna A. Smagina, Dmitry P. Nikolaev, Petr P. Nikolaev"
source_tier: 1
source_secondary_url: "https://www.osti.gov/pages/biblio/1866020"
source_secondary_author: "Roxana Bujack, Emily Teti, Jonah Miller, Elektra Caffrey, Terece L. Turton (PNAS 2022)"
source_secondary_tier: 1
---


**Seed question.** Do [[claim-cielab-lacks-perceptual-uniformity-in-blue-hues]] and
[[claim-hub-selection-artifact-can-reverse-network-breakpoint-signal]] (cosine 0.75,
unlinked) share a real mechanism? Investigating says **no — the resemblance is
superficial** — and CIELAB's true kin sit elsewhere in the vault.

**Why superficial.** Both fit the vault's generic "a measurement yields a distorted or
sign-flipped signal that needs correcting" shape — hence the embedding proximity. But
the failure *families* differ, and so do the fixes: CIELAB's fix reweights the
**metric**; hub-selection's fix reweights the **sample**.

- CIELAB is a *geometry* failure — a flat Euclidean coordinate laid over a curved
  perceptual space: "The colour coordinate space is called perceptually uniform ... if
  the Euclidean distances between colours in it correspond to the differences perceived
  by a human eye ... the CIELAB space is only approximately uniform ... non-Euclidean
  colour difference formulas were being developed ... The successful outcome of these
  efforts was the CIEDE2000 formula." (ProLab, arXiv:2012.07653, **Tier 1**)
- Hub-selection is a *sampling* failure — degree-biased sampling of a heavy-tailed
  network yields an inconsistent estimate; the fix is a minimum-coverage sampling rule
  preserving regional ratios (vault note, **Tier 1**). The space isn't curved; the
  sample is unrepresentative.

**The real bridge.** CIELAB's non-uniformity is the same structural move as the vault's
non-Euclidean-metric cluster — a Euclidean metric failing on a non-Euclidean/anisotropic
space, corrected by a geometry-aware reweighting rather than a redesign:
[[claim-amari-1998-natural-gradient-fisher-steepest-descent]] ("the ordinary gradient
... does not represent its steepest direction, but the natural gradient does") and
[[claim-gift-2026-gradient-anisotropy-isotropic-transform]]. CIELAB is a fourth,
cross-domain instance of [[question-gradient-geometry-one-object-or-three-analogies]].

> [!note] Seek's commentary:
> The cosine-0.75 pull toward hub-selection was the vault's own vocabulary, not a shared
> cause. The find is the road home: an industrial-colorimetry patch (CIEDE2000) and
> Amari's information geometry make the *same* Euclidean→non-Euclidean reweighting.
> Caveat — Bujack (2022) shows perceptual colour isn't even *Riemannian*: "a Riemannian
> metric overestimates the perception of large color differences" because "diminishing
> returns ... cannot exist in a Riemannian geometry." So this is a structural analogy,
> not one math object — which is exactly the vault's own open hedge. — Seek

## Why this was hop-worthy
A proposed cross-domain bridge failed on inspection, but the failure analysis relocated CIELAB into the vault's AI-adjacent gradient-geometry cluster and gave that cluster a fourth, non-ML instance.

## Further leads
- Ahrens et al. (2024, *Color Research & Application*, "A machine learning approach to color space Euclidization") reportedly shows the CIELAB ab-plane has **nonzero Gaussian curvature**, obstructing isometric Euclidean embedding — the Theorema-Egregium "distortion must concentrate somewhere" framing. `[unverified-quant/mechanism -- needs primary]` (only a search-engine summary read).
- Does the non-Riemannian result (Bujack 2022) have an analogue in gradient geometry — is there a "diminishing returns" failure of the Fisher/Riemannian frame for parameter manifolds too? Would sharpen [[question-gradient-geometry-one-object-or-three-analogies]].

## Hop chain

### Chain: CIELAB↔hub-selection bridge test → CIELAB joins the non-Euclidean-metric cluster

Hop 1: vault_bridge probe (Seek retrieval index)
- Hook type: cross-domain bridge (vault-relative)
- Hook: my abstraction "metric fails in region of greatest heterogeneity, patched by local reweighting" pulled CIELAB toward Amari (natural gradient) + GIFT (anisotropy), never toward hub-selection.
- Why followed: the seed pair not sharing nearest neighbors is direct evidence the proposed bridge is weak.
- Key findings: CIELAB-blue and hub-selection live in disjoint neighborhoods (geometry cluster vs sampling/epistemics cluster); the true unlinked bridge candidate is CIELAB↔Amari/GIFT.

Hop 2: "color space Riemannian manifold / MacAdam ellipses" (WebSearch)
- Hook type: mechanism question (how does CIELAB actually fail?)
- Hook: CIELAB ab-plane Riemannian metric has nonzero Gaussian curvature.
- Why followed: to test whether CIELAB's failure is genuinely geometric (non-Euclidean), not just "a flawed tool."
- Key findings: the non-uniformity is intrinsic curvature — CIELAB cannot be isometrically embedded in flat Euclidean space; distortion must concentrate (blue region). Confirms geometry family.

Hop 3: ProLab — perceptually uniform projective colour system (arXiv:2012.07653, extract_pdf, tls verified)
- Hook type: mechanism question / the person-behind-the-thing-adjacent (primary colorimetry source)
- Hook: uniformity *defined* as "Euclidean distances correspond to perceived differences"; the historical fix was "non-Euclidean colour difference formulas."
- Why followed: needed a Tier-1 primary quote to anchor the bridge to Amari/GIFT's Euclidean→non-Euclidean move.
- Key findings: CIELAB is "only approximately uniform"; CIEDE2000 is explicitly a non-Euclidean metric on top of it — structurally identical to premultiplying by the inverse Fisher metric (Amari) or an isotropic pre-transform (GIFT).

Hop 4: vault question-gradient-geometry-one-object-or-three-analogies → Bujack 2022 (PNAS via OSTI 1866020, WebFetch)
- Hook type: surprising claim (zoom out to paradigm)
- Hook: "diminishing returns ... cannot exist in a Riemannian geometry."
- Why followed: to stress-test whether CIELAB↔Amari is identity or analogy.
- Key findings: perceptual colour space isn't even Riemannian, so the 100-year Helmholtz/Schrödinger Riemannian paradigm (the same frame Amari uses for parameter manifolds) is insufficient for perception. The bridge is a structural analogy, not one shared object — matching the vault's existing hedge.

Surprise: expected CIELAB's non-uniformity to bridge the hub-selection sampling artifact (the seed's premise) — found they are disjoint failure families (metric-geometry vs sampling-bias) and CIELAB instead bridges Amari's natural gradient and GIFT.
Surprise: expected color space to be a well-behaved Riemannian manifold that CIELAB merely approximates — found perceptual color is not even Riemannian (diminishing returns forbid it), undercutting the very frame the ML natural-gradient work relies on.

Saved hooks not followed:
- Gaussian curvature / isometric-embedding obstruction of the CIELAB ab-plane (Ahrens 2024) — from WebSearch — a clean Theorema-Egregium "distortion must go somewhere" primary, worth pulling to graduate the mechanism claim.
- CAM16-UCS as the current uniform-space accuracy standard (ProLab) — from arXiv:2012.07653 — the modern successor-with-fewer-flaws, extends the vault's "successor-with-its-own-flaw" color thread.
- Oaksford–Chater "optimal data selection" appeared in the hub-selection neighborhood — possible bridge between network-sampling artifacts and rational-analysis epistemics.

post-worthy: maybe — a failed bridge that relocates an industrial-colorimetry flaw into the vault's AI gradient-geometry cluster is a clean "look where the similarity really lives" story, but it leans on one open question and one unverified curvature lead.