Adaptive kernel smoothing

Adaptive kernel smoothing is a highly effective technique for transforming the Poisson spatial point process of an ACIS image into a scalar real field that represents more clearly the true distribution of X-ray emission in the sky. Its advantage over standard Gaussian or boxcar smoothing is that the width of the Gaussian kernel is self-adjusted across the field to match the local density of events. The result is a smoothed image that shows simultaneously point sources and diffuse emission without smearing the former into the latter. See, e.g., Brian Silverman's monograph "Nonparametric Density Estimation" (Chapman & Hall, c. 1985) for a statistical discussion.

ACIS adaptive smoothing is achieved with a code developed by Harald Ebeling (Hawaii), and is implemented both within CIAO (asmooth) and as an IDL widget (csmooth). The method is complicated and generates a variety of ancillary files; users are strongly encouraged to read the CIAO manual entry. The formal reference is H. Ebeling et al. (2000, MNRAS).

An obvious complication with application to ACIS-I images is nonuniform exposure across the field, particularly the chip gaps which become smoothed along with the true photon distribution. The exposure map must thus also be appropriately smoothed. ACIS scientists, in communication with Harald, developed the following procedure:

  1. Run adaptive smoothing on the ACIS-I image:
  2. Smooth the exposure map with same spatial distribution of kernel widths developed for the image map (i.e., the sclmap file produced above is the input here):
  3. Divide the two smoothed maps. CIAO dmimgcalc will preserve headers and ancillary files:
However, a variety of potential problems have been noted, for which more investigation is needed: To assess this, P. Broos performed a test of asmooth using a synthesized random image of 100,000 Poisson deviates using a realistic ACIS exposure map (i.e. with gaps). The resulting smoothed image showed bright features in the four corners of the array with amplitude of several percent, suggesting that algorithmic artifacts can be present. In light of this result, H. Ebeling suggests that asmooth be run with the sliding-cell rather than FFT convolution option, as the FFT will always produce artifacts around the field edges/corners due to applied boundary conditions. The sliding-cell algorithm is bias-free but makes much greater demands on CPU time. But a single test of the sliding-cell option on a random image also showed diffuse features.

Note 1: Running csmooth on a full ACIS image seems to produce zeros in the kernel map (sclmap output file) which causes a failure in the code. On can check for and remove zeros as follows using CIAO and IDL: