Split learning (SL) trains a neural network across devices by exchanging intermediate activations and gradients. In over-the-air SL, the wireless channel serves as an inter-layer linear map, and when the forward and backward channels match, reciprocity and conjugation supply the backward product without channel estimation at the base station (BS). Passive user motion, which the learning link does not control, breaks this match. Movable antennas let the antenna move relative to the user device, and we use this freedom to compensate the displacement. For a stationary BS and a propagation environment quasi-static within each exchange, returning the antenna to its forward-transmission position restores the channel and yields a conditionally unbiased gradient estimate. We also derive when discrete positioning or mechanical movement can achieve this in time. A residual displacement below a quarter wavelength keeps the real part of the channel correlation positive for any angular spectrum. At zero effective offset, independent Gaussian jitter attenuates this correlation without reversing its sign, whereas a persistent offset can reverse it, and in arrays the tolerance depends on the receive-combining direction. These effects enter a nonconvex convergence bound, and a local analysis shows that partial mechanical compensation saves cycle time to first order at only a second-order loss in correlation. In CIFAR-10 experiments, a user terminal without channel state information stays at chance level under uncompensated walking motion, while discrete positioning stays within about 0.3 percentage points of exact restoration, including at a speed beyond the modeled actuator's limit.

翻译：暂无翻译