At one point xt, every data point proposes a simple conditional arrow. Their average is the marginal velocity. Sampling one arrow at a time recovers it.
conditional arrows (one per data point) sampled this step running average (what least squares learns) true marginal