✨ Developing PerturbationMatrix class

[christinahedges]

This PR is for developing a new PerturbationMatrix class which will ultimately factor our a lot of the headaches we're having inside machine

To Do

• Documentation (docstrings!)
• Merge into machine.py
• Write some tests for Kepler/K2/TESS data

[christinahedges]

Roughly how to use:

If flux and flux_err are arrays with dimensions (ntimes x npixels)

P = PerturbationMatrix3d(time=time, dx=dx, dy=dy)
P.fit(flux, flux_err)
model = P.model()

model will then also have (ntimes x npixels)

[jorgemarpa]

This is great! I think the basic functionalities are implemented here.

Something that will be good to have is a pixel_mask argument in PerturbationMatrix3D so we can create matrix with the source_mask (for fitting) pixels and the uncontaminated_source_mask (for building) in one perturbation object.

I did a couple of "performance" test with TPF data and turned out that the self.model method runs really slow, about 1.7 sec/iter. We need to find why this is so slow.

[christinahedges]

I added a PCA method so you can do something like:

y = [whitened_timeseries, whitened_timeseries]
p = PerturbationMatrix(...)
p.pca(y, ncomponents=10)
p.fit(y)

[jorgemarpa]

@christinahedges I found the bug with the inconsistent shape when doing self.model(). The problem is when using PerturbationMatrix through PerturbationMatrix3D that some attributes from the latter are not updated to reflect the addition of PCA components in self.vectors. See my comments for more details.

[christinahedges]

Any vectors that aren't the time polynomial are now set to zero mean in each "segment" during _clean_vectors method. The pca method can now be repeated and it will just update the pca components, not append ad infinitum

[jorgemarpa]

This looks great! I tested a handful of component combinations and other parameters to check the matrices look Ok.
It fulfills what we always thought about the perturbation matrix API.

I just added a couple of comments regarding docstrings. Besides those, I think this is ready to be merged.

[jorgemarpa]

to be consistent with the rest

Suggested change

	def pca(self, y, ncomponents=5, smooth_time_scale=0):
	def pca(self, y, ncomponents=3, smooth_time_scale=0):

[jorgemarpa]

Suggested change

	Will add two time scales of principal components, definied by `long_time_scale`
	and `med_time_scale`.

[jorgemarpa]

What is the scale of smooth_time_scale?

Suggested change

	long_time_scale: float
	The time scale where variability is considered "long" term. Should be
	in the same units as `self.time`.
	med_time_scale: float
	The time scale where variability is considered "medium" term. Should be
	in the same units as `self.time`. Variability longer than `long_time_scale`,
	or shorter than `med_time_scale`, will be removed before building components
	smooth_time_scale: float
	Amount to smooth the components, using a spline in time.
	If 0, the components will not be smoothed.

[jorgemarpa]

maybe we can still keep this function in case we want to remove the temporal trend for other components. But it isn't necessary with the current API.