The way variability is analyzed in astronomical images has been deeply influenced by the development of the microlensing experiments (EROS, OGLE, MACHO, DUO). The enormous quantity of CCD images produced by these experiments has motivated the development of fast and accurate techniques to analyze variability. The first implementations of variability search were based on an analysis of catalogues of stellar objects. The production of the catalogues has been almost exclusively done using the DoPHOT software (Mateo & Schechter). This type of analysis is still used in microlensing experiments, and has proved very successful. However since we are concerned only with differential photometry, such data processing is certainly not optimal. Purely differential methods, like image subtraction, would appear to be better suited for study of variability in astronomical images. However the development of image subtraction techniques has been long delayed by the inherent difficulty and complexity of the method.
The first successful implementation of the method was by Tomaney &
Crotts (1996). Further progress has been made by the creation of the
fast Optimal Image Subtraction (OIS) method (Alard & Lupton 1998). Due
to its ability to solve the full least-square problem, the OIS method
proved to produce subtracted images and light curves with an accuracy
approaching closely the photon noise limit (Alard 1999). The OIS method
is very efficient for crowded stellar fields, like those encountered in
microlensing experiments, and could readily be used for massive
processing of microlensing data. The current implementation of the OIS
works by dividing the field into small sub-areas where constant kernel
solutions are derived. In dense crowded fields, sub-areas as small as
128
128 pixels can be used. At this scale kernel variations
can usually be ignored to a good approximation. However in case of very
bad optics or less dense fields, the constant kernel approximation does
not hold any more. For instance, in high latitude fields taken for
supernovae search, there are often insufficient bright objects per unit
area over which the kernel variations can be ignored. We are then
forced to try to make a self consistent fit of the kernel variations,
as described in Alard & Lupton (1998). It is important to notice
though that even if the spatial variations are fitted to order 1 only,
the cost to build the least-square matrix will be about 9 times larger
than for a constant kernel solution. Some situations might require
order 2 or 3 or even more. Order 3 requires roughly 100 times more
calculations than a constant kernel solution. Clearly the problem
quickly becomes intractable, and that one of the main advantages of the
OIS, the fast computing time, will be completely lost.
Fortunately, it will be demonstrated in this article that the fit of the spatial variations of the kernel can be achieved for little additional computing cost, given a reanalysis of the problem.
Copyright The European Southern Observatory (ESO)