# Эксперимент: Weighted-Volume Contractive Optimizer (#1378)

{
  "worked": false,
  "confidence": 9,
  "verdict": "Built a PyTorch MVP for weighted divergence, exact divergence, Hutchinson estimation, and a state-dependent rho controller on an anisotropic quadratic. The identity matched to 3.6e-15 and Hutchinson estimation was accurate, but the proposed optimizer performed worse: final loss was 0.004998 versus 3.2e-9 for fixed-step SGD, with peak loss 1391 versus 32. The more negative measured divergence did not produce stable or faster optimization.",
  "metrics": {
    "baseline": "8 seeds, 180 steps: final loss mean 3.203e-09, std 0; peak loss mean 32.0; mean divergence -0.88; mean rho 0.055",
    "idea": "8 seeds, 180 steps: final loss mean 4.998e-03, std 0; peak loss mean 1391.07; mean divergence -2.632; mean rho 0.1675; exact formula error 3.55e-15; Hutchinson divergence standard error 0.0434"
  },
  "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py",
  "files": [
    "experiment.py",
    "results.json"
  ],
  "limitations": "Only an anisotropic quadratic toy test was run, not MNIST/CIFAR or an MLP/ResNet. Optimization used explicit two-dimensional divergence rather than Hutchinson probes, did not match average update norms or FLOPs, and tested beta=0 only with one directional rho parameterization."
}