Gaussian measure to handle asymptotic boundary condition

[wsdewitt]

The asymptotically constant boundary condition is funny when we consider the regularization integrals, because the standard choices (and my Bregman divergence) will diverge.

To cope with this I wave my hands and use a centered Gaussian measure on t instead of Lebesgue, so the infinite epoch does not blow up the regularization integral.

This is described in a bit more detail in regularization.ipynb, but it's a hack that should be scrutinized.

[kamdh]

I don't think this really makes much of a difference. In the integral formulation, true, the integrals are not defined. However, the regularization (as chosen) weights all intervals equally, so the point corresponding to \infty is weighted the same as any other.