Results

As mentioned above, two simulated images (stars centered on pixels, stars not necessarily centered on pixels) were reduced with five point-spread functions each (a PSF was obtained from each of the simulated ``science frames,'' one was obtained from actual WFC images, and three were obtained from Tiny TIM simulations). In each case I found that the PSFs obtained from the simulations themselves yielded the most accurate photometry. Tables 1 and 2 present actual values of derived using the PSFs from the synthetic frames themselves, for four subsets: (a) all stars; (b) all stars except the faintest two magnitudes (true < 2.00); (c) all stars more than 28 pixels from the center of the cluster; and (d) all stars brighter than true = 2.00 inside 28 pixels from the center of the cluster.

ALLSTAR does not seem to have a big problem with undersampling: with one exception, the results are not worse when the stars are allowed to lie anywhere within their pixels. And I believe that exception may be more an artifact of the simulation than of the reduction. In the process of generating the first test image, whenever two artificial stars had their centers in the same pixel, one was thrown away and only one was kept. In the latter simulation, any number of stars falling in the same pixel were all retained. Subsequently, the reduction software was unable to distinguish the numerous stars lying in a single pixel, and reduced the blend as one object. The cross-identification software then associated this object with only one of the artificial stars in the ``truth'' list, thereby implying a lower recovery rate and a higher typical photometric error.