arXiv:2608.07754 2026 Architecture 1 ideas extracted · analyzed Sep 1, 2026
The paper offers a complex-valued resonant state-space primitive in which leakage and oscillation are represented by the eigenvalue k = lambda + i omega. Unlike a purely real decay channel, this state preserves phase and can selectively amplify periodic structure while remaining stable under negative damping. The most direct transfer is a bank of learnable damped oscillators implemented as a diagonal state-space layer: use exact recurrence for streaming inference and the equivalent causal convolution for parallel training. This is a useful architecture idea, although its ingredients overlap substantially with existing diagonal complex SSMs and Fourier-style sequence layers.
Replace real diagonal state-space channels with complex damped oscillators whose hidden states encode both amplitude and phase. Train with parallel causal convolution and deploy with the equivalent one-step recurrence, allowing the same layer to support efficient batched training and low-memory streaming inference.