← Home | Gravitational Collapse — Numerics & Analytics
Test Project · Local AI Infrastructure · Berlin · June 2026

Gravitational Collapse — Numerics & Analytics

Two complementary sub-projects: (1) Choptuik critical collapse via SpheriCo.jl, (2) ODE parameter study of Gravastar formation (Jampolski–Rezzolla). Both executed on Mac Studio M3 Ultra, fully locally.

SpheriCo.jl v0.1.0 · NRPy+ v2.1.1 · SFcollapse1D (arm64) · Julia 1.12.6 · Python 3.11 (nr_env) · SciPy · Mac Studio M3 Ultra (256 GB) · June 2026

Overview and Software Stack

ToolVersion / PathPurpose
SpheriCo.jlv0.1.0 · ~/.julia/packages/SpheriCo/VAgaD/Spherically symmetric scalar field collapse, Choptuik project
NRPy+v2.1.1 · ~/miniforge3/envs/nr_env/ (Python 3.11)Symbolic GR code generation; Jupyter kernel registered
SFcollapse1Darm64 binary · /Volumes/CLAUDE-DATA/SFcollapse1D/Scalar field collapse in C++ (GCC 15.2.0, -O2 -fopenmp)
HDF5v2.1.1 (Homebrew) + HDF5.jl v0.17.3Output format for SpheriCo.jl simulations

The GR-PHASE ODE system and Choptuik bisection were implemented in Python (SciPy Radau solver, rtol=1e-8, atol=1e-10) and Julia (SpheriCo.jl) respectively. All computations ran locally on Mac Studio M3 Ultra, without any cloud API calls.


Sub-project 1
Choptuik Critical Collapse — SpheriCo.jl

Starting point and method

Choptuik's (1993) critical collapse of a massless scalar field in spherical symmetry is the prototype for Type-II critical phenomena in general relativity: near the critical amplitude pc, black hole mass scales as MBH ∝ (p − pc)γ with universal γ ≈ 0.374. Discrete self-similarity (DSS) produces oscillations in MBH(p). The computation used SpheriCo.jl v0.1.0 with a Gaussian pulse initial profile (rc=0, width=2, rmax=12, Nr=259).

Result 1: Critical amplitude

pc = 0.165524768573898 (50 bisection steps, Δ=3.61×10−16 = machine precision)
Runtime: 11 minutes. SpheriCo.jl detects apparent horizon formation and terminates sub-critical runs cleanly.

Result 2: Mass-scaling exponent γ

Two methods were applied to the MBH(dp) data (dp = p − pc):

MethodγDeviation from lit. (0.37454)Remark
Minima-Envelope0.3681.8%Extracts slope from local minima of MBH(dp); avoids DSS noise. Physically reliable.
Naive linear fit0.63770%All 19 BH data points including far from pc; dominated by DSS oscillations. Not reliable.
Outcome: γ = 0.368 ± ~0.007 (Minima-Envelope), 1.8% below Choptuik reference γ=0.37454. DSS oscillations in MBH(dp) directly observed. The 70% deviation of the naive fit is a methodological artefact.
Limitation: High-resolution run (Nr=1025) yielded only 3 BH points before tmax boundary near pc. Higher tmax or adaptive stepping needed to reduce the 1.8% discrepancy further.

Scripts and data


Sub-project 2
GR-PHASE: Phase Diagram of Gravastar Formation (Jampolski–Rezzolla)

Paper and model

Based on Jampolski & Rezzolla, arXiv:2509.15302. Three spacetime regions: Region I (de Sitter, interior, pI=−eI), Region II (dust shell), Region III (Schwarzschild exterior). Master ODE governs junction dynamics. Three outcomes: Black Hole (R2<2M before ρ12), Gravastar (equilibrium at R2=2M), No-Equilibrium. Free parameters: eI/eII(initial), |kI|/kII, R̅2/M (compactness C=M/R̅2≤3/8). Implementation: grphase_core.py (v3), SciPy Radau (rtol=1e-8, atol=1e-10).

Result 1: Phase diagram (parameter scan)

2/MC=M/R̅2Black HolesGravastarsNo-Equilibrium
9.00.11100625
4.50.22212622
3.00.3337843504

Gravastars form preferentially at high compactness C near 3/8. At low compactness (R̅2=9, C=0.111), No-Equilibrium dominates. This confirms Jampolski & Rezzolla's central finding. Data: /Volumes/CLAUDE-DATA/GR-PHASE/results/scan/scan_Rb{3.00,4.50,9.00}.json

Result 2: Separatrix scaling exponent γ ≈ 1

At fixed R̅2=3.0, separatrix located at 5 kI values by bisection (precision ≤4×10−7), then power-law fit |δ| ∝ |eI−eIc|γ:

kIeIcγBHγNo-Eq
0.0800.262820961.03530.9990.97541.000
0.1500.159149231.03300.9990.97631.000
0.2500.092082691.01120.9990.99820.999
0.4000.042972631.02451.0000.98021.000
0.6000.009228270.98881.0001.01540.999
γBH = 1.019 ± 0.017  |  γNo-Eq = 0.989 ± 0.016 (N=5, R²>0.999 in all cases)

Comparison: Choptuik 1993 (scalar field, D=4): γ=0.374, Type-II DSS · Boson star (CSS, D=4): γ=0.26, Type-II CSS · Jampolski-ODE (BH↔Gravastar): γ=1.0, Landau-type, generic

Physical conclusion: γ≈1.0 indicates Landau-type (mean-field) critical behavior. The BH↔Gravastar separator is not a Type-II critical point: no DSS/CSS echoing, no universal scaling. The transition is controlled by a codimension-1 surface in parameter space with smooth (non-universal) structure.

Result 3: Separatrix geometry — no simple power law

Fit of eIc(kI) = A × kIα for four R̅2 values:

2/Mα (free fit)RMS (α=−3/2)RMS (α=−π/2)
2.8−0.886 ± 0.0880.9110.9741.068
3.0−0.837 ± 0.0770.9440.7500.823
3.5−0.933 ± 0.0960.9600.4280.476
4.0−1.393 ± 0.2300.9250.2440.258

Conclusion: α varies from −0.84 to −1.39 with R̅2. No universal exponent. α=−3/2 consistently better than α=−π/2 but both inferior to free fit. Variation too large for single-parameter description.

Physical reason: The Gravastar condition involves T(eI, kI) = (1/HI) arcsinh(2M HI/√kI) and τ≈3.99. Junction conditions couple T and τ non-trivially, precluding a closed-form eIc(kI).

Scripts and data


Sub-project 3
EEG-DSS: Analytical DSS Solutions in the Large-D Limit (Ecker–Ecker–Grumiller)

Based on Ecker, Ecker & Grumiller, arXiv:2601.14358 (PRL 2026). LO+NLO analytical solutions implemented in Python (SciPy, NumPy, Mac Studio M3 Ultra). Four computational phases executed on 2026-06-14 in under 1 second. Status: executed.

Mathematical framework

Coordinate transformation τ=−ln(−t), x=−r/t, small parameter ε=1/(D−1). DSS condition: fields periodic in τ with period Δ. Free integration function β(τ+Δ)=β(τ). LO solutions (Eqs. 7–13): ΠLO=β/√(1+β²x²), fLO=√((1+β²x²)/(1+β²)). NLO correction (Eqs. 17–18): ΠNLO=[4β²x²(1+β²)(β+β')²+p1ln(1+β²x²)]/[2β³(1+β²x²)5/2] with explicit polynomials p1–p4. NLO consistency condition (Eq. 20): Δ=|β''|/(3|β'|) at β=0 — uniquely fixes the echo period.

Result 1: Paper example reproduced (Phase 0, D=300)

NEC vertex: τ0 = 0.24018424593  |  Δimplied = 1.0000000 ± 3.6×10−8 (machine precision)

Field ranges and NLO corrections at D=300 (ε=0.00334):
  Π: LO ∈ [−1.016, +1.016], max|εΠNLO| = 0.074 (7.4% correction)
  f: LO ∈ [0.701, 1.000], max(ftotal−1) = −3.2×10−8 → convexity f≤1 satisfied ✓

NEC lines: LO: τ=0.240184 (horizontal) · NLO: τ=0.240184+0.00334x (tilt εx) · SSH shift: Δτ=0.00334

Result 2: Consistency landscape (Phase 1a, 50×50 grid)

120/2500 points satisfy |Δimplied−1|<5%. The computed consistency line reveals a simple structure: A3≈A1/15.95 (linear). The paper constant A=15.9476 is universal for the 2-mode Fourier class. Two branches exist (positive and negative A3), corresponding to qualitatively different DSS solution families.

Result 3: Dmin map and optimal β (Phase 1b/1c)

A1A3 (consistent)Dmin (LO+NLO)
0.400.025180 ← minimum found
0.500.031482
0.750.047090
1.00 (paper)0.0627≈100 (LO+NLO) vs. 52 (paper NNLO)
2.500.1568287
Discrepancy vs. paper: Our LO+NLO gives Dmin≈100 for A1=1; the paper finds 52 from NNLO (Supplement Eqs. S10–S18, not implemented here). Factor ~2 gap is the expected NNLO contribution. Qualitative trend correctly captured: smaller amplitude → smaller Dmin.

Result 4: Convergence of the 1/D series (Phase 2)

Dεmax|εΠNLOLO|max(ftotal−1)
4 (Choptuik)0.33317.4+2.27 — divergent!
52 (paper Dmin)0.01961.026+0.019
1000.01010.529+1.2×10−4
3000.003340.175−1.5×10−6 (≤1 ✓)
Convergence finding: The 1/D series requires D≥~200 for f≤1 globally (17% NLO correction at D=300). Converges fast near center x≈0, slowly near SSH x=1. At D=4 (Choptuik): NLO/LO=17, f=1+2.27 — 1/D perturbation theory is completely inapplicable to the physical Choptuik problem. This confirms that EEG solutions are large-D objects with no perturbative connection to D=4 critical collapse.

Scripts and data


Overall Assessment

Sub-projectTypeOutcome
1 — Choptuik (SpheriCo.jl)Numerical GRpc=0.16552476857... (machine precision) · γ=0.368 (1.8% from lit.) · DSS oscillations observed
2 — GR-PHASE ODE (Jampolski-Rezzolla)ODE parameter studyPhase diagram (3×625 pts) · γ≈1.0 Landau-type (NOT Choptuik) · Separatrix: no universal power law
3 — EEG-DSS (Ecker et al.)LO+NLO Python (4 phases)τ0=0.24018, Δ=1.000000 ✓ · Dmin,opt=80 · 1/D diverges at D=4 (NLO/LO=17)
T. Riepe, Berlin. June 2026. Numerical relativistic computations on Mac Studio M3 Ultra (256 GB). Papers: Jampolski & Rezzolla arXiv:2509.15302; Ecker, Ecker & Grumiller arXiv:2601.14358; Choptuik (1993) Phys. Rev. Lett. 70, 9. Software: SpheriCo.jl (Giannakopoulos et al., arXiv:2412.19722), NRPy+, SFcollapse1D.