UNNS Galaxy Structural Fingerprint Atlas — What the Data Output Shows
The project establishes a reproducible galaxy structural-fingerprint atlas across 121 SPARC rotation-curve galaxies and three chamber channels (ρ, ν, Aκ). The external transfer experiment shows that one static, whole-galaxy model architecture does not fully use that structure — while ρ(κ) emerges as the leading transfer-sensitive channel for a future local or regime-conditioned design.
Source: UNNS_GALAXY_STRUCTURE_AND_VALIDATION_v0_3 · 121 primary galaxies · Gate B (reproducibility) PASS · Gate C (representation sensitivity) mapped across 4 branches
1The fingerprint space has channel-specific internal architecture
The 121 galaxies do not scatter randomly. ρ(κ) and ν(Vκ)(κ) compress into 2–3 structural coordinates; Aκ deficit is more heterogeneous and does not compress as tightly.
Variance retained in first 3 PCs, by channel:
- ρ(κ): 90.8% — strongly compressible
- ν(Vκ)(κ): 88.5% — strongly compressible
- Aκ deficit: 55.9% — more heterogeneous
- All channels combined: 61.0%
ρ and ν form compact dominant response families, while Aκ carries a richer, more heterogeneous structure. The structural map also contains a dense central population plus pronounced outliers, rather than one smooth continuum — visible in the scatter below.
2Chamber regime is a major atlas coordinate
The dominant axis is doing two things at once: distinguishing individual curve shapes, and separating a discrete Weak Persistence regime from the main Stable Structure population.
Median structural-PC1, by persistence state:
- Stable Structure (110/121): ρ-PC1 ≈ −2.109, structural-PC1 ≈ −0.794
- Weak Persistence (11/121): ρ-PC1 ≈ +11.218, structural-PC1 ≈ +3.076
Weak Persistence is not a minor label variation; it is a pronounced multiscale response regime. Apparent galaxy "identity" distances can partly reflect distances between chamber regimes, so a stability audit should also evaluate identity within the same regime, not only across the full sample.
3The atlas carries both physical galaxy information and an observational layer
Some atlas coordinates organize galaxies by physical scale; other, different coordinates record how densely and over what radial range the galaxy was observed — a second, observational layer of the fingerprint atlas.
structural-PC2 organizes galaxies by broad physical scale: surface brightness, luminosity, rotation velocity, HI mass all correlate at ρ≈0.39–0.46.
Other directions encode sampling density and radial coverage — an observational layer that the thinning/deletion/truncation audit (Gate C) was built to characterize.
4Persistence, fragmentation, and connectivity are distinct
A galaxy can be highly persistent while remaining fragmented and never reaching full percolation on the native κ range — STRUC-I and STRUC-PERC-I measure genuinely different, largely independent dimensions of organization.
Stable Structure110 / 121
Hard fragmentation103 / 121
Both simultaneously94 / 121
Full percolation17 / 121
All 17 full-percolation galaxies reach it only in the adaptive extension (κ-connect > 1); none reaches full connectivity on the main κ grid. There is no detectable association between Stable/Weak Persistence and whether full percolation is eventually reached.
5Structural neighbours preserve conventional similarity
Median nearest-neighbour differences in the structural atlas are smaller than the all-galaxy-pair medians across five conventional properties — structural proximity is not arbitrary, though matches are far from exact.
The fingerprint atlas preserves recognizable galaxy organization while also revealing cross-class structural analogues among galaxies with different conventional classifications.
6Reproducibility (Gate B) vs. representation sensitivity (Gate C)
Across five unchanged-ladder STRUC-I executions, the complete curves form stable galaxy-specific structural identities: native chamber variation is much smaller than between-galaxy separation. Change how the galaxy was observed — radius deletion, thinning, coverage, mass-to-light prep — and the fingerprint responds measurably.
Top-1 source retrieval—
Top-5 source retrieval—
Median own/nearest-wrong ratio—
Median ICC (ρ PC1–3)—
Median top-5 Jaccard—
Gate B: five unchanged-ladder STRUC-I executions → perfect source recovery and neighbour stability. The atlas is not a product of native chamber randomness.
Gate C maps a measurable transformation of the fingerprint under point deletion, thinning, coverage change, and baryonic preparation. This does not erase the fingerprint — the current coordinates combine galaxy structure with the observational channel through which that structure is measured, and normalization is the next step.
7The tested static global gravity architecture did not transfer
The best held-out model was the simplest one — local baryonic acceleration plus normalized radius. Adding whole-galaxy summaries, structural or otherwise, made held-out prediction worse: one static whole-galaxy representation is not the correct transfer architecture for the structural fingerprints.
- M1 (local ḡbar + normalized radius) has the lowest galaxy-balanced held-out RMSE.
- Every model that adds whole-galaxy descriptors — even ordinary ones (M2), not just UNNS signatures (M-I, M-P, M-IP) — generalizes worse.
- M-IP's training R² improves (0.720 → 0.727) while its held-out R² falls (0.685 → 0.652), and calibration slope drops from 0.925 to 0.888 — consistent with overfit or sample-specific identity encoding, not a demonstrated leakage mechanism.
This result does not remove the structural signal: correct curve assignment beats shuffled assignment, and complete curves retain recoverable galaxy identity (see section 9).
8Ablation: Aκ is the most heterogeneous and most detrimental compact channel
Removing the compact Aκ summaries from the combined model improves galaxy-balanced RMSE the most of any single ablation — consistent with Aκ's low three-PC compression (55.9%) in section 1. Low compression does not by itself prove instability; here it co-occurs with the largest transfer cost.
9Full curves preserve identity and reveal unequal transfer channels
Correct galaxy-to-curve assignment beats shuffled assignment (p≈0.020), and the channels are not interchangeable: ρ(κ) is the leading channel-specific transfer indication (+1.14%, positive in 4/5 folds); Aκ carries weaker transfer structure (+0.46%, 3/5 folds); ν is counterproductive for this target (−1.52%, bootstrap interval entirely on the detrimental side).
Combining every channel is counterproductive: ρ alone outperforms the all-channel combination (+0.44% aggregate, 1/5 folds) because ν drags the combined result down. ρ is the strongest candidate for a new local or regime-conditioned transfer experiment.
10The aggregate result contains strong galaxy-level transfer heterogeneity
The compact combined model improved 53/121 galaxies; the full ρ curve improved 71/121 — but gains and losses are highly unequal and sometimes reverse between representations (e.g. NGC3972 worsens under compact signatures, improves under full curves).
Most consistently benefited (compact + full ρ)
Most consistently deteriorated
This heterogeneity points to structural regimes, profile-quality dependencies, or local transition phenomena — not a single global linear coefficient. It defines the case for a local or regime-conditioned successor design.