Tom Stephens

1

Sliding Box Source Detection
Tom Stephens GSSC Database Programmer
2
Outline

• Motivation
• Algorithm
• Some tests
• Initial DC2 results
• Future directions

3
Motivation

• Simple
• To understand
• To implement (at least at first)
• About equal to the other methods in DC1
• Similar number of detections
• Similar number of false positives
• Requires no a priori knowledge of the sky
• Serves as a baseline to compare other methods
• Simplest thing that could possibly work
• More complicated methods should be more
  successful.
• Plenty of room to grow

4
Basic Method

• Pick a position on the sky
• Count up all the photons in a set aperture around
  that position
• Count up all the photons in an annulus around the
  aperture
• Divide the number of counts in the annulus by its
  area to get local average background counts
• Compute expected background counts for the
  aperture
• Compute the significance of the count excess for
  that position
• Build a map of the significance
• Search map for peaks

5
My Implementation

• Works directly with the events, no binning
• Photons loaded into a spatial grid based on HTM
  pixelization
• Allows quick access to only those photons near
  the point of interest
• Trial positions selected on a Cartesian grid of
  RA and Dec with 0.1 spacing
• Distance from trial point calculated for each
  nearby photon
• All calculations done in spherical coordinates
• Counts added to appropriate bins (aperture or
  annulus) or dropped if within neither
• Significance calculated and written to output
• Computation time scales as number of events and
  size of outer radius of the background annulus

6
Sample Significance Map
7
Picking Targets

• Once the significance map is created we scan
  through it to pick out the peaks
• Currently use a simple algorithm that find the
  most significant pixel within a specified region
• Can specify the size of the region to consider
• Region size is adjusted to account for distortion
  near the poles
• This gives a list of potential sources to be fed
  into the likelihood pipeline.
• On the previous map, there are 1463 targets with
  a significance greater than 4 sigma

8
Optimizing parameters

• Need to optimize the choice of 3 parameters
• Aperture radius
• Offset of inner background annulus radius from
  aperture
• Width of background annulus
• Ran a series of tests using sample data created
  by Jim Chiang
• Field containing 35 sources
• 7 different fluxes
• 5 different spectral indicies
• Isotropic background
• Goal to find the combination of parameters that
  detected the greatest number of sources
• Ran source detection routine on data set with 160
  combinations of the parameters and 5 different
  energy bands

9
Best Aperture size

• Number of detections averaged over all choices of
  position and size of annulus for each energy band
• Surprisingly, best aperture size was 0.25 (the
  smallest tried) regardless of energy band even at
  low and mid energies where PSF is larger

10
Annulus Size and Position

• Examined each of these parameters by averaging
  the number of detected sources over the other
  parameter for each of the different energy bins
  and the two smallest aperture sizes
• The dependence of detection efficiency on the
  size or offset of the background annulus was
  small
• 2 sources at most
• Typically 1 source or less
• Best choices seems to be
• Annulus offset from Aperture 0.75
• Width of Aperture 1.5

11
Looking at the DC2 data

• Used just the class A events
• Ran source detection using the optimized
  parameters
• Aperture 0.25 radius
• Annulus offset 0.75
• Annulus width 1.5
• Ran over all 5 energy bands
• Will show results from the analysis of the data
  without energy cuts

12
All Event Significance Map
13
Detected Sources
14
DC2 Catalog targets
15
Matched Sources
16
Other Possible Sources
17
Future Work

• Many improvements possible
• Computing Significance
• A lot of time wasted at the poles better
  selection of trial points
• Binned data
• Eliminate computation time dependence on number
  of events
• May help improve background calculation
• Use background model
• Correlate between different runs to eliminate
  false positives
• Energy bins
• Time bins
• Different parameter on/off galactic plane
• Target Selection
• Improve algorithm to select sources
• Improve localization