Location of Derivatives

[jenniferyee]

When we implemented derivatives for calculating the chi2 gradient (e.g. dA_dparams), we put them in FitData and it led to a lot of recalculating of magnification curves and trajectories. I propose we move all the derivatives to be properties of MagnificationCurve and introduce lazy loading for a lot of parameters (e.g. a_pspl, b0), so they can be calculated just once.

@rpoleski

[jenniferyee]

See commits e095289 and 4907b8e

[jenniferyee]

Related question: right now, we calculate the gradient with respect to rho based on the whole light curve, even if the finite source method is only applied to some of the data. This seems problematic, but is related to the difficulty of accessing db0, db1, etc. from MagnificationCurve (or FitData), because they are read when the Lens Class is created. Related to the fact that Model and PointLens exists without a specific set of reference times, but MagnificationCurve does not.

[jenniferyee]

The initial premise here is not quite correct. We take into account the different magnification methods when we calculate the derivative of B0 and B1. However the underlying problem remains the same how do we avoid recalculating z, B0, and the PSPL magnification?

[jenniferyee]

maybe instead of lens = PointSourcePointLens(parameters), we need lens_magnification = PointSourcePointLensMagnification(u, parameters), i.e., something that stores all the information for a specific trajectory.

[jenniferyee]

IDEA: in MagnificationCurve, instead of storing _methods_per_epoch, what about storing a list of "Magnification" objects that each take a Trajectory defined for the relevant epochs?

[jenniferyee]

See Version3_1 aebbc72