# Эксперимент: Permutation-Mixed Orthogonal Quantization (#972)

{
  "worked": false,
  "confidence": 8,
  "verdict": "Built a NumPy MVP for signed-permutation Hadamard preconditioning, exact reconstruction, quantization, coherence, and coordinate-sampling tests. Orthogonality was exact, transformed sparse vectors had maximum coordinates scaling downward from 0.335 at d=64 to 0.191 at d=256, and extreme sampling at d=128,m=4 matched the predicted NMSE (30.95 observed versus 31 predicted). However, max-scaled 4-bit quantization was consistently worse after mixing, so the promised general compression win was not observed.",
  "metrics": {
    "baseline": "Per-vector max-scaled 4-bit quantization NMSE averages: 6.92e-5 at d=64, 4.67e-5 at d=128, and 2.97e-5 at d=256; identity coordinate sampling at d=128,m=4 gave 35.24 NMSE.",
    "idea": "Mixed AΣB 4-bit quantization NMSE averages: 2.01e-4 at d=64, 1.14e-4 at d=128, and 6.13e-5 at d=256; mixed sampling at d=128,m=4 gave 30.95 NMSE, with maximum transformed coordinate energy 0.0625 versus 1.0 for identity."
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json",
    "run_output.txt",
    "run_output2.txt"
  ],
  "limitations": "Only toy synthetic vectors and NumPy Hadamard transforms were tested; no Transformer, WikiText, KV cache, latency, memory traffic, percentile or fixed-scale calibration, or GPU experiment was evaluated. The coherence sweep used finite dimensions and structured Hadamard factors rather than independently sampled dense orthogonal matrices."
}