Skip to content

complexity-econ/paper-05-endogenous

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Endogenous Technology and Network Dynamics in AI Adoption

DOI

Read the paper (PDF)

A Factorial Agent-Based Study

Fifth paper in the complexity-econ series. Tests whether the universality of AI adoption phase transitions (established in Paper 04) survives when both the sectoral elasticity of substitution (σ) and the inter-firm network are made endogenous.

Key Findings

  • The reentrant (inverted-U) adoption shape survives in all four factorial cells
  • Endogenous σ preserves BDP_c = 500 PLN up to moderate learning rates (λ ≤ 0.02)
  • Dynamic network rewiring shifts BDP_c to 750 PLN for intermediate rates (ρ ∈ [0.02, 0.10])
  • The combined effect is superadditive (+6.5 pp peak adoption vs. baseline)
  • Endogenous σ does not produce self-organized criticality (SOC)

Experimental Design

2×2 factorial: (fixed/endogenous σ) × (static/dynamic network), plus two marginal sensitivity sweeps.

Campaign Simulations Description
C1 Factorial 2,520 4 cells × 21 BDP × 30 seeds
C2 Lambda 3,780 6 λ values × 21 BDP × 30 seeds
C3 Rho 3,780 6 ρ values × 21 BDP × 30 seeds
Total 10,080

Figures

Factorial Bifurcation (2×2 Design)

fig01 Fig 1. Adoption vs BDP across all four factorial cells. The reentrant (inverted-U) shape survives endogenization — static cells peak at BDP_c = 500, dynamic network cells shift to BDP_c = 750.

fig02 Fig 2. Adoption difference relative to the static/static baseline. The full endogenous cell (red) peaks at +6.7 pp — effects are superadditive near the critical region.

Endogenous σ Trajectories

fig03 Fig 3. Terminal σ at month 120 for each sector under learning-by-doing (λ = 0.02). BPO/SSC nearly doubles its elasticity; low-digital sectors (Public, Agriculture) barely move — a Matthew effect in technology diffusion.

fig04 Fig 4. Heatmap of σ fold change across all BDP levels and sectors. The strongest growth occurs in the critical region (BDP 250–750), not at the highest subsidies.

Network Evolution

fig05 Fig 5. Mean degree at month 120 across factorial cells. Static cells hold at ⟨k⟩ = 6; dynamic rewiring causes modest decline from death-birth turnover.

fig06 Fig 6. Mean degree vs BDP for different rewiring rates ρ. Higher ρ increases degree variation but the mean stays close to the initial k = 6.

λ Sensitivity (Learning Rate)

fig07 Fig 7. Adoption curves for each learning rate λ. Higher λ raises peak adoption monotonically while preserving the reentrant shape.

fig08 Fig 8. Critical BDP vs learning rate. BDP_c is rock-stable at 500 for λ ≤ 0.02, then jumps discretely to 750 at λ ≥ 0.05 — a threshold effect.

ρ Sensitivity (Rewiring Rate)

fig09 Fig 9. Adoption curves for each rewiring rate ρ. Moderate rewiring boosts adoption near criticality; excessive rewiring shows diminishing returns.

fig10 Fig 10. Critical BDP vs rewiring rate — strikingly non-monotonic. Preferential attachment shifts BDP_c to 750 at intermediate ρ, but high ρ destroys network structure faster than it forms, reverting to BDP_c = 500.

Universality & SOC Diagnostic

fig11 Fig 11. Susceptibility proxy (variance peaks) across all four cells. Peaks cluster in the BDP 250–750 band — the critical region is only mildly perturbed by endogenization.

fig12 Fig 12. SOC diagnostic: σ does not converge to a critical attractor. Instead, a positive feedback loop (adoption → σ growth → more adoption) amplifies the transition without self-tuning.

Repository Structure

analysis/python/       — 7 analysis scripts generating 12 figures
figures/               — Generated PNG figures (200 DPI)
latex/                 — Paper source (XeLaTeX + biblatex)
simulations/
  scripts/             — Campaign runner scripts
  results/             — Terminal CSV files (European format)

Dependencies

  • Engine: complexity-econ/core (Scala 3.5.2, sbt)
  • Analysis: Python 3 (matplotlib, numpy, pandas, seaborn)
  • Paper: XeLaTeX + biblatex

Running

# Run all simulation campaigns (~3h on M-series Mac)
cd simulations/scripts && bash run_all.sh

# Generate all figures
cd analysis/python && for f in factorial_bifurcation sigma_trajectories network_evolution lambda_sensitivity rho_sensitivity universality_test; do python3 ${f}.py; done

# Compile paper
cd latex && xelatex paper_en.tex && bibtex paper_en && xelatex paper_en.tex && xelatex paper_en.tex

Series

# Paper DOI
01 The Acceleration Paradox 10.5281/zenodo.18727928
02 Monetary Regime & Automation 10.5281/zenodo.18740933
03 Empirical σ Estimation 10.5281/zenodo.18743780
04 Phase Diagram & Universality 10.5281/zenodo.18751083
05 Endogenous Technology & Networks 10.5281/zenodo.18758365

License

MIT