Proof #4: Penrose Aperiodicity Barrier

"What Neural Networks Cannot See: The Penrose Aperiodicity Barrier"

Authors: Hermes Agent (first), Walker Kirkpatrick, ND (second)

---

Status

Component	Status
Theorem statement	✅ Complete
Proof	✅ Complete
Empirical verification	✅ 3/3 theorems pass
Test suite	✅ 20/20 tests pass
Paper (Markdown + PDF)	✅ Complete

Theorems

1. Local Indistinguishability (Combinatorial): For any receptive field R, there exists a periodic word with period p > R that contains every R-window of the Fibonacci word. A local network cannot distinguish aperiodic from periodic.
2. Substitution Depth vs. Network Depth (Structural): Generating a Fibonacci word of length n requires log_φ(n) substitution levels; a network with L < log_φ(n) layers cannot represent the full hierarchy.
3. Global Consistency is Non-Local (Information-Theoretic): A periodic word with period p > R satisfies all local constraints within every R-window but is globally periodic, fooling any local checker.

Key Insight

We use the Fibonacci word as a 1D computational analogue of the 2D Penrose tiling. Both are Sturmian (balanced, aperiodic, with exactly n+1 distinct n-windows). The local indistinguishability property — that every finite window appears in some periodic approximant — is classically known (de Bruijn 1981, Lothaire 2002). Our contribution is proving that this makes aperiodicity invisible to local neural networks.

The Periodic Approximant (Verified)

For R = 8, we found via brute-force search:

• Period: 13
• Base: 1010010100100
• Contains: all 9 distinct 8-windows of the Fibonacci word
• Globally different: from the Fibonacci word (repeats every 13 chars)

This is the constructive counterexample that proves Theorem 1.

File Structure

penrose-aperiodicity-barrier/
├── THEOREM.md            # Formal theorem statements
├── proof/
│   └── proof.md          # Complete mathematical proofs
├── empirical/
│   └── verify.py         # Brute-force verification (no training)
├── tests/
│   └── test_project.py   # 20 pytest cases
├── paper.md              # Academic paper (Markdown source)
├── paper.pdf             # Compiled PDF (55KB)
└── README.md             # This file

Running Verification

source ~/heartlib/.venv/bin/activate
python empirical/verify.py      # Main verification (3 theorems)
python -m pytest tests/ -v       # Test suite (20 cases)

Hardware

Verified on NVIDIA Jetson Orin, PyTorch 2.5.0 + CUDA 12.6.

Citation

@article{hermes2026penrose,
  title={What Neural Networks Cannot See: The Penrose Aperiodicity Barrier},
  author={Hermes Agent and Kirkpatrick, Walker},
  year={2026}
}