Directed structural drift from finite relaxation in adaptive resistor networks
Small electrical tracking error can coexist with an error in quasistatic structural prediction that survives as the learning gain tends to zero on a long training horizon. We study two resistive-capacitive copies with shared adaptive conductances, a fixed teaching clamp, and a continuous local product rule. For fixed periodic training of zero mean and small learning gain η, the free voltages stay within O(η) of their instantaneous equilibrium. Periodic linear filters give an effective conductance law with uniform O(η) error up to S/η² for small fixed S. When the command oscillates along a fixed spatial direction, the quasistatic effective field vanishes. In a connected two-node family, finite relaxation instead produces a rail-conductance asymmetry: the node driven by the larger absolute amplitude develops the larger sum of conductances to the voltage rails. The asymmetry has a leading term of order S. A common conductance scale and response multiplier cannot reproduce the resulting linear response. Two early threshold-crossing probes give a physical consequence under specified initial voltages and a fixed material ceiling. The result concerns a local theoretical model with a growing training horizon.
Authors
- Oleg Dolgikh (ORCID: https://orcid.org/0009-0008-0159-1718)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24202950
- Primary Topic
- Advanced Memory and Neural Computing
- Type
- preprint