- Python 100%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| code | ||
| data | ||
| .gitignore | ||
| CITATION.cff | ||
| LICENSE | ||
| README.md | ||
| requirements.txt | ||
Procrustes shape-matching and null tests for The Starry Night (F612) and benchmark nocturnes
Code and data accompanying "The actual sky of Van Gogh's Starry Night: a multidisciplinary reconstruction" (Supplementary Notes 3–4).
The analysis quantifies how well the central asterism of The Starry Night (F612) matches the Hyades cluster, and tests whether that match could arise by chance, by an arbitrary choice of points, or by an arbitrary labelling.
Identity, not date.
rho_CSis scale-invariant and rotation-free, so it measures pure shape and is therefore date-invariant. Every test here probes identification ("are these points the Hyades?"), never the calendar date. No date is claimed or tested by this code.
Contents
data/
F612_measurements.csv 11 central stars of F612, three CAD measurements (mm)
F612_mean.csv their means (analysis input); Sn8 = Venus
sky_hyades.csv Hyades V-asterism, J2000.0 gnomonic, arcmin, centred
sky_aries.csv Aries (non-diagnostic comparison)
replica_F1540_centered.csv replica drawing F1540, five central points, centred (mm)
benchmark_rhone.csv Big Dipper benchmark (Rhone, 1888): canvas + sky, k=7
benchmark_cafe.csv Summer Triangle benchmark (Cafe, 1888): canvas + sky, k=5
benchmark_millet.csv Winter-asterism transferability fit (Millet, ~1850-65): canvas + sky, k=8
correspondences.csv anatomical point-to-star correspondences (F612)
code/
procrustes.py closed-form 2D Procrustes engine + CSV loaders
benchmark_fits.py 1888 validation (Rhone, Cafe) + Millet transferability fits + their MC null
montecarlo_null.py random-placement null test (pillar A)
permutation_tests.py label (B1), subset (B2), direction (C) tests
reproduce_all.py runs everything and checks the headline numbers
Method
The optimal rotation, isotropic scale and translation that map a fixed sky
constellation onto a canvas configuration are obtained in closed form from the
cross-product matrix M = Xc^T Tc:
theta = atan2(M21 - M12, M11 + M22) scale = (sigma0 + sigma1) / Tr(T^T T)
This is mathematically identical to the SVD solution R = U V^T, and being built
from cos/sin it is always a proper rotation (det = +1), so reflections are
excluded. The shape-mismatch index is the centroid-size-normalised RMS residual,
rho_CS = 100 * RMS(r_i) / CS, with CS = sqrt(sum ||Xc_i||^2).
Coordinate registration. The canvas is measured in image coordinates (the y
axis runs downward); the sky is a gnomonic projection (declination runs upward).
Because the fit allows proper rotation only, the two coordinate systems must
share the same handedness, so one axis is negated to register them. In the
stored data this is folded into sky_eta (declination axis already negated),
so every fit is simply procrustes(canvas, sky) with no per-fit flip. This is a
fixed coordinate-system registration, not a fitted reflection.
Reproduce
pip install -r requirements.txt
cd code
python reproduce_all.py # ~ seconds; prints results and PASS/FAIL checks
python reproduce_all.py 1000000 # optional: larger Monte Carlo sample
Expected headline values:
Fits are listed in manuscript order (transferability, then 1888 validation, then the Hyades application and its replica, with the reference-only Aries last):
| quantity | value |
|---|---|
| Millet winter-asterism transferability fit (k = 8) | rho_CS = 3.85 %, rho_max = 7.22 % (Rigel), CS = 497.18 mm |
| Rhone Big Dipper benchmark (k = 7) | rho_CS = 6.54 % |
| Cafe Summer Triangle benchmark (k = 5) | rho_CS = 5.46 % |
| F612 Hyades fit (k = 5) | rho_CS = 8.17 %, rho_max = 10.59 %, theta = 46.5°, CS = 472.93 mm |
| F1540 replica fit (k = 5) | rho_CS = 6.62 %, rho_max = 7.85 %, CS = 317.34 mm |
| F612 Aries fit (k = 3) | rho_CS = 5.36 % (non-diagnostic / reference-only, see below) |
| Monte Carlo, Millet / Rhone / Cafe | p < 5 × 10⁻⁶ (0 of 200,000) |
| Monte Carlo, F612 Hyades | p ≈ 3 × 10⁻⁵ (null median ≈ 40 % at k = 5) |
| Monte Carlo, F1540 replica | p ≈ 5 × 10⁻⁶ |
| Monte Carlo, Aries (k = 3) | p ≈ 0.008 (k = 3 null is permissive: 5th percentile 13.2 % vs 26.8 % at k = 5) |
| Label permutation (B1) | anatomical labelling ranks 1 / 120, p ≈ 1/120 ≈ 0.008 |
| Subset enumeration (B2) | {Sn2,Sn3,Sn6,Sn7,Sn10} ranks 1 / 462; winner excludes Sn8 (Venus) |
| Direction gate (C) | Hyades first; runner-up 8.85 % is Venus-contaminated |
Reproducibility of the Monte Carlo p-values. The random-placement null draws
k points uniformly in the unit square [0,1]^2; the RNG seed (20250722) and
sample size (200,000) are fixed, so the p-values are exactly reproducible. With
N = 200,000 the smallest measurable non-zero p is 5 × 10⁻⁶. A fit at rho_CS =
8.17 % (the F612 Hyades fit) has a true tail probability of order 10⁻⁵, so it
returns a small non-zero count (p ≈ 3 × 10⁻⁵) rather than the sampling floor;
lower-residual fits (≤ 6.6 %) reach the floor (p < 5 × 10⁻⁶). Raw null draws are
not stored because they are regenerable from the seed.
Why Aries is non-diagnostic. Its raw rho_CS (5.36 %) is low only because k = 3
has two degrees of freedom; the permutation ceiling for k = 3 is 1/3! ≈ 0.167,
so no k = 3 fit can be statistically distinguished from chance regardless of its
residual.
Data sources
Celestial coordinates are from the Hipparcos catalogue (J2000.0). The painting and drawing images analysed here are held by the Van Gogh Museum (F612, F1540) and are not redistributed in this repository.
License
Code and data are released under the MIT License (see LICENSE).