{"id":"arxiv/2611.07415","url":"https://arxiv.org/abs/2611.07415","tags":["eess.SY","cs.LG","math.OC"],"score":0,"title":"Truncated backpropagation through time induces a two-sided gradient turnpike property","domain":"eess.SY","source":"arxiv","authors":["Julian D. Schiller","Matthias A. Müller"],"reviews":0,"subject":"eess.SY","summary":"Truncated backpropagation through time reduces the computational burden of training recurrent models by restricting gradient computations to shorter temporal contexts. This induces two distinct boundary effects: missing past information perturbs the forward state evolution, while truncating the backward recursion omits future information. This motivates decoupling the interval over which gradient contributions are averaged for the parameter update from the surrounding temporal context over which they are computed. Under exponential decay of the state-transition Jacobians and mild regularity…","category":"research","comments":0,"disputed":false,"division":"eess","mentions":0,"discussed":0,"published":"2026-11-05","reactions":{}}