Elman networks show ~15% higher finite-lag transport trace than GRU/LSTM
measured in 1 paper- Reddy proves an exact decomposition of the finite-lag transport tensor into a conditional-spread trace, a coherent-displacement trace, and an antisymmetric circulation statistic, estimated from gradient-trained Elman/GRU/LSTM networks on a repeat-copy task (recall 0.9999-1.000). [reddy-2026-finite-lag-operator-geometry] - Elman's transport trace runs ~15% higher than GRU's and LSTM's (tr(G_Delta) Elman 1.0085+/-0.0008 vs GRU 0.8673+/-0.0076, LSTM 0.8815+/-0.0035), with a larger 16-26% gap in the coherent-displacement trace. [reddy-2026-finite-lag-operator-geometry] - The gap survives a capacity-matched control (Elman-64 vs GRU-36: 1.013 vs 0.858), ruling out parameter count, and localizes to specific task phases (write vs recall). [reddy-2026-finite-lag-operator-geometry] - Observational (an empirical trajectory statistic), not a causal intervention. [reddy-2026-finite-lag-operator-geometry]