φ-register Rosetta Stone
Five domains, one structure — the translation dictionary.
The constraint
| Structural concept | φ-register | CSD accelerators | QDI circuits | Neural CA / rewriting | Coding theory |
|---|---|---|---|---|---|
| No two adjacent elements active | No consecutive 1s (Zeckendorf) | No adjacent nonzero signed digits (canonic recoding) | Mutual exclusion between adjacent rails (dual-rail encoding) | Neighborhood constraint on update mask | Forbidden word '11' in golden mean shift |
| The generator of the constraint | M = [[1,1],[1,0]], φ² = φ+1 | Booth recoding scan of adjacent bit pairs | Muller C-element gate topology | Learned local update kernel | Transition matrix of the shift space |
The process
| Structural concept | φ-register | CSD accelerators | QDI circuits | Neural CA / rewriting | Coding theory |
|---|---|---|---|---|---|
| Local rewrite that resolves a violation | 011 → 100 (carry rule) | Replace digit pair (1,1) with (−1,0,1) or equivalent | Handshake: req/ack exchange between neighbors | Cell applies update function to local neighborhood | Context-free substitution rule in the transducer |
| Intermediate state that violates the constraint | Digit 2, or adjacent 1s before normalization | Non-canonic representation (adjacent nonzeros present) | Spacer / invalid codeword in dual-rail | Unstable cell configuration | Non-prefix-free codeword; ambiguous parse |
| Propagation direction | Bidirectional: +1 left, −2 right (Fenwick) | Left-to-right scan; borrow propagates right | Completion tree: leaves → root (upward) | Neighborhood radius determines propagation extent | Berstel transducer: left-to-right, 4 states |
The result
| Structural concept | φ-register | CSD accelerators | QDI circuits | Neural CA / rewriting | Coding theory |
|---|---|---|---|---|---|
| Global consistency achieved | Zeckendorf normal form | Canonic Signed Digit form (minimal Hamming weight) | All handshakes complete; output valid | Stable fixed point of iterated updates | Valid prefix-free encoding; delimiter '11' placed |
| Completion signal | No cell contains 2; no adjacent 1s | All digit pairs satisfy adjacency check | Completion tree output fires (matched rising/falling) | State delta < ε (convergence criterion) | '11' delimiter marks end-of-codeword |
| Error / corruption detection | Cassini identity: F(n-1)·F(n+1) − F(n)² = (−1)ⁿ | Non-canonic digits in output (should never appear) | Dual-rail: both rails high = error (1-of-2 violation) | Divergence or oscillation = failed training | Constraint violation mid-stream → resync within 3 tokens |
The cost
| Structural concept | φ-register | CSD accelerators | QDI circuits | Neural CA / rewriting | Coding theory |
|---|---|---|---|---|---|
| Capacity per element | log₂(φ) ≈ 0.694 bits/cell | ~0.67 nonzero digits per position (avg. Hamming density) | 2 wires per bit (dual-rail overhead) | Effective receptive field per parameter | 1.44 × log₂(n) bits per integer |
| What you trade for the constraint | ~44% more cells per integer vs. binary | Signed digit storage (2 bits per trit) | 2× wire count; area overhead 10-40% | Convergence iterations before stability | ~44% longer codewords vs. optimal binary |
| What you gain | Free error detection, gentle overflow, self-delimiting | 77% multiplier area reduction; carry-free addition | No clock; data-dependent timing; glitch-free | Self-organizing; no global coordination needed | Self-synchronizing; error-resilient; no parameter tuning |
Each row is one structural concept. Each column is one domain's name for it. Every cell in a row describes the same mathematical object — not an analogy.
φ-register × AI engineering: research tiers map
Companion to the Rosetta Stone dictionary. The dictionary translates vocabulary across domains. This map shows what's been done at each intersection and what hasn't, with literature pointers.
Tier 1 — Established and deployed
Results with working hardware, production systems, or mature publication records.
| Intersection | What exists | Key literature |
|---|---|---|
| CSD multipliers for CNNs | Canonic Signed Digit approximation of CNN filter weights. 77% logic area reduction, 5.63% top-1 accuracy drop on VGG-16/ImageNet, no retraining. The adjacency constraint (no adjacent nonzero digits) is structurally identical to Zeckendorf. | Riaz et al., "CAxCNN," IEEE Access 2020; MA4C, ScienceDirect 2022 (4.2× FPGA logic reduction); RNS-CSD CNN accelerator at 22nm (1.54× energy efficiency) |
| CSD in general processors | CSD encoding in a RISC-V pipeline achieves 2.41× clock frequency increase over standard binary, >20% from the encoding itself. | Reichenbach et al., "CSD-Driven Speedup in RISC-V Processor," Springer 2025; companion project "RISC-V3" |
| Carry-save in TPU systolic arrays | Google TPU patent describes carry-save adders (redundant binary) within MAC units, passing carries between units in partially redundant form. Redundant number arithmetic inside the most deployed AI accelerator. | US Patent US20230015148A1, Google 2023 |
| Ternary weight networks | Weights constrained to {−1,0,+1} — structurally identical to balanced ternary digits. 16× model compression, minimal accuracy loss. Mature field with memristor implementations. | Li & Liu, "Ternary Weight Networks," arXiv:1605.04711, 2016; Zhu et al., "Trained Ternary Quantization," ICLR 2017; memristor TWN, IEEE 2024–2025 |
| QDI neuromorphic circuits | Quasi-delay-insensitive design where local handshaking settles to global consistency — structurally identical to Fibonacci normalization. Working silicon. | IBM TrueNorth, ASYNC 2012; SynSense Speck SoC, Nature Communications (0.42mW, 4-phase QDI dual-rail) |
| Fibonacci lattice sampling | Golden-angle-based Fibonacci spiral as optimal quasi-uniform spherical point distribution. Used in 3D vision rotation sampling and physics ML. | Alexa, "Super-Fibonacci Spirals," CVPR 2022; Song et al., "ML on Spherical Fibonacci Lattice," Phys. Rev. Research 2022 |
| Golden mean shift capacity | Channel capacity of binary sequences with no consecutive 1s = log₂(φ) ≈ 0.694 bits/symbol. Classical, rigorous, derived directly from Q-matrix eigenvalue. | MacKay, Information Theory, Inference, and Learning Algorithms, Ch. 17; Cover & Thomas |
| Fibonacci hashing in ML infra | Knuth's multiplicative hash using floor(2^w/φ). Implemented in TensorFlow hash table operations. Golden ratio's irrationality ensures optimal key distribution. | Knuth, TAOCP Vol. 3; TensorFlow source |
| Quasi-random hyperparameter search | Golden-ratio-based low-discrepancy sequences for hyperparameter optimization. | Google Deep Learning Tuning Playbook (quasi-random search section) |
Tier 2 — Published but niche
Peer-reviewed or preprint results that directly bridge φ-register structures and AI, but with limited follow-up.
| Intersection | What exists | Key literature |
|---|---|---|
| Fibonacci weight encoding for hardware inference | Neural network weights retrained to Fibonacci-encoded values (no consecutive 1s in binary), then computed with an inexact multiplier that produces exact results for Fibonacci inputs. 0.4% accuracy drop on SqueezeNet 1.0, 73% multiplier area reduction, 43% power-delay-product reduction. | "Exact Neural Networks from Inexact Multipliers via Fibonacci Weight Encoding," IEEE/ACM DATE 2021 |
| Fibonacci codeword quantization (FCQ) | Rounds NN weights to nearest Fibonacci codeword (no consecutive 1s), then exploits redundancy for lossless compression. Applied to neural radio receivers with two novel compression algorithms. | arXiv:2511.01921, "Fibbinary-Based Compression and Quantization for Efficient Neural Radio Receivers," 2025 |
| Zeckendorf pooling (Z pooling) | CNN pooling operator that encodes features via non-consecutive Fibonacci weights. Replaces max pooling with Zeckendorf-decomposition-based operator. Rotation-invariant neighborhoods. Evaluated on BSD500 and glioblastoma imaging. | Vigneron, Maaref & Syed, "Z Pooling," Entropy 2021 |
| Fibonacci Network architecture | Each block takes input from two previous blocks plus original, mirroring f(n) = f(n-1) + f(n-2). Used for signal reconstruction. Outperforms positional encoding in noisy settings. | arXiv:2411.05052, "The Fibonacci Network: A Simple Alternative for Positional Encoding," 2024 |
| Stakhov's ternary-Fibonacci-Bergman bridge | Explicitly connects balanced ternary, base-φ, and golden-proportion codes. Ternary mirror-symmetrical arithmetic for pipeline DSP. Argues binary has inherent vulnerability that Fibonacci p-codes resolve via built-in error detection. | Stakhov, "Brousentsov's Ternary Principle, Bergman's Number System and Ternary Mirror-Symmetrical Arithmetic," The Computer Journal 2002; Stakhov, The Computer Journal 2018 |
| Multifractal loss landscapes | Neural network loss surfaces modeled as multifractals with varying Hölder exponents. Unifies clustered minima, edge-of-stability, anomalous diffusion. Fractal dynamics established but not φ-specific. | Ly & Gong, Nature Communications 2025 |
| Fractal trainability boundaries | Boundary between successful and failed training in hyperparameter space has Mandelbrot/Julia-set fractal structure. Rigorous result, not φ-specific. | Sohl-Dickstein, arXiv 2024 (Google DeepMind) |
| Neural rewriting systems | Learned convergent term rewriting with modular match-select-apply cycle. Outperforms GPT-4o on formula simplification. Fibonacci normalization is a specific convergent rewriting system (Church-Rosser). | Petruzzellis et al., arXiv:2507.19372, July 2025 |
| Growing Neural Cellular Automata | Learned differentiable local update rules (~8,000 params). Same structural pattern as 011→100: local rule, iterated application, convergence to stable state. | Mordvintsev et al., Distill 2020; Béna et al., "Path to Universal NCA," GECCO 2025; Miotti et al., "Differentiable Logic CA," Google Research 2025 |
| FractalNet | Self-similar architecture: f_{n+1}(z) = (f_n ∘ f_n)(z) ⊕ f₁(z). Binary doubling, not Fibonacci recurrence. 22.85% error CIFAR-100, matching ResNets without residual connections. | Larsson, Maire & Shakhnarovich, ICLR 2017 |
| Hausdorff dimension generalization bounds | SGD trajectory fractal dimension bounds generalization: heavier tails (lower dimension) generalize better. Rigorous, not φ-specific. | Şimşekli et al., NeurIPS 2020 |
Tier 3 — Genuinely unexplored
No paper exists. The structural connection is clear from the dictionary but nobody has made it.
| Intersection | What's missing | Why it's plausible | Dictionary row it maps to |
|---|---|---|---|
| Zeckendorf-constrained pruning masks | No paper frames "no consecutive 1s" as a structured sparsity pattern for neural network weight or activation pruning. | NVIDIA's 2:4 structured sparsity proves hardware-enforced masks work. Zeckendorf's constraint is a (1,∞)-RLL pattern enforcing ≤50% density with a precise capacity bound (0.694n bits per n-width layer) that 2:4 lacks. | "No two adjacent elements active" + "Capacity per element" |
| Fibonacci coding in PAC-Bayes bounds | PAC-Bayes compression bounds require choosing a coding scheme. Everyone uses Huffman or arithmetic coding. Nobody has tried Fibonacci coding. | Fibonacci's self-synchronization means bounds might degrade gracefully under weight perturbation rather than catastrophically. Self-delimiting property removes framing overhead. | "Error / corruption detection" + "What you gain" |
| Q-matrix as a stabilized state space kernel | M=[[1,1],[1,0]] is a companion matrix with eigenvalues φ and −1/φ. As-is it's an unstable SSM (|φ| > 1). Nobody in S4/Mamba has studied this spectral structure. | Discretize or project onto unit circle and you get a recurrence with optimal irrationality baked in. The diagonalization parallels Diagonal State Space approaches. Fast doubling (M^2n = (M^n)²) parallels parallel scan in linear RNNs. | "The generator of the constraint" |
| Fibonacci-variant FractalNet | Existing FractalNet merges fn ∘ f_n (binary doubling). Nobody has published a variant merging f{n-1} and f_{n-2}, which gives golden-ratio-scaled subpath lengths instead of power-of-2 lengths. | FractalNet already proves self-similar architectures compete with ResNets. The Fibonacci variant is a one-line change to the expansion rule. | "Local rewrite that resolves a violation" (self-similar structure) |
| 0.694 bits/symbol as a neural capacity bound | The information-theoretic cost of the adjacency constraint is exactly log₂(φ) per symbol. Nobody has applied this as a capacity bound on constrained neural network layers. | Directly computable, rigorous, and would give tighter bounds than unconstrained entropy for any layer operating under a Zeckendorf-like sparsity mask. | "Capacity per element" + "What you trade" |
| Cassini identity as runtime error detection in inference | F(n-1)·F(n+1) − F(n)² = (−1)ⁿ gives free corruption detection. Nobody has applied this to detect bit-flip errors during neural network inference on Fibonacci-encoded weights. | The DATE 2021 Fibonacci weight encoding paper already constrains weights to Fibonacci form. Adding Cassini checks is a natural extension for safety-critical inference. | "Error / corruption detection" |
| Normalization-as-settling for learned rewrite rules | NCA/NRS learn local rules that converge. Fibonacci normalization has proven Church-Rosser convergence. Nobody has used Fibonacci's guaranteed termination properties to constrain or initialize learned rewriting systems. | NCA convergence is currently empirical (train and hope). Fibonacci normalization provides a known-convergent starting point that could warm-start or regularize learned rules. | "Global consistency achieved" + "Completion signal" |
Reading order: start with the Rosetta Stone dictionary to learn the shared vocabulary, then use this map to find where you want to work. Tier 1 entries are tools you can use today. Tier 2 entries are papers to cite and build on. Tier 3 entries are open contributions.