Эксперимент: Recursive Nonlocal Edge Feedback GNN (#1120)

{ "worked": true, "confidence": 8, "verdict": "Built a sparse-ring Recursive Nonlocal Edge Feedback GNN with recurrent per-edge contexts, pooled global feedback, endpoint-conditioned edge flows, and signed sparse aggregation. The mechanism checks confirmed all three quantitative predictions: the contraction boundary occurred exactly at gamma*Lambda=1, geometric decay matched (0.8)^12 with zero error, and nonlocal response was linear in coupling alpha with slope error below 1e-15. The recursive model slightly improved one-step RMSE (0.12498 vs 0.12585) and 8-step RMSE (0.43946 vs 0.44356), but used 5605 parameters versus 452, so this is evidence for the mechanism rather than a demonstrated efficiency win.", "metrics": { "baseline": "Local sparse GNN: one-step RMSE 0.125847, 8-step RMSE 0.443560, rollout Jacobian rho 0.955725, 452 parameters.", "idea": "Recursive feedback GNN: one-step RMSE 0.124978, 8-step RMSE 0.439464, rollout Jacobian rho 0.906252, 5605 parameters. Mechanism checks: boundary product 1.0 exactly; decay prediction/observation 0.0687195/0.0687195; linearity slope prediction/observation 1.7/1.7." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a tiny synthetic ring with one distant coupling was tested; no dense-GNN comparison, equal-parameter comparison, multi-seed confidence interval, explicit learned stability penalty, or larger graph/scaling benchmark was run. The mechanism sweeps use the exact scalar linearized recurrence, so they verify the claimed stability signature but do not establish that training reliably reaches the favorable regime." }