Spontaneous Symmetry Breaking of Reference Frames as a Computational Cost Minimization Strategy

Version 2 adds a Numerical Demonstration section (Section 7) containing a sparse rotating-support linear prediction model with D=64 ambient dimensions and m=8 signal dimensions. A budget-constrained selector agent using signed gradient EMA is compared against a random-k baseline across 12 budget levels and 3 drift rates. The simulation reproduces two signatures predicted by the theoretical analysis:

(a) attention entropy collapses away from ln D toward O(ln m) as the agent concentrates on the signal subspace;

(b) the budgeted selector achieves lower prediction error than the random baseline for k up to roughly 3m, with a slight negative dip at larger k reflecting commitment cost.

All figures and data can be reproduced from the included Python script (paper2_kstar_scaling_demo.py, NumPy + Matplotlib only, fixed seeds).                                                                                                                                                                 Additional changes: updated Paper I reference to concept DOI; tightened cross-series language throughout; added reproducibility file list; removed orphan bibliography entries.