consider jackknife approach to error analysis

[davidwhogg]

The S/N analysis we pair-coded today (2012-06-25) is very non-conservative; it assumes independent, Gaussian pixels and so on. A much more conservative option would be jackknife:

Take the N images (30, 100, 300, whatever), divide into 10 subsamples, make all 10 leave-one-out subsamples (the complements of the small subsamples), reoptimize the scene for each of these leave-one-out subsamples, and then inflate the rms among leave-one-out subsamples by a factor of 10 or 9 or whatever it is to get the jackknife rms.

This is a good idea, and very conservative.

[davidwhogg]

def jackknife(scene, dataStructure, njack = 10):
    """
    `scene` is the best-fit scene on *all* the data from The Thresher
    `dataStructure` is whatever your crazy-ass data structure is
    """
    Nimages = len(dataStructure)
    subsampleID = np.array((Nimages / njack) * range(njack)) # note brittleness
    assert subsampleID.size == Nimages
    np.random.shuffle(subsampleID)
    # replace this shit with a Pool(njack).map operation
    for id in range(njack):
        loo = dataStructure[subsampleID != id]
        loo_scene[id] = # re-optimize the scene using the loo sample
        photometry[id] = # re-measure photometry
    rms = # jackknife operation on photometry!
    return the, shizzle