Bi-Symmetric Autocorrelation & such

T. Jarrett, IPAC
(971001)

In the near-infrared, most galaxies appear smoothly symmetric about the major and minor axis (as opposed to the optical, where the younger population, preferentially distributed in spiral arms, dominate the light, giving a "grainy" appearance to their profiles). Exceptions: irregular galaxies (rare), and galaxies with local stellar contamination (exclusive to high stellar density regions). For the most part, galaxies appear two-fold symmtric, while multiple star groups appear asymmetric.

We can exploit this "feature" to separate galaxies from "false" galaxy detections, including double stars and triple stars. Double stars (and most triples) appear asymetric across the minor axis -- that is to say, if you center the elliptical axis on the primary component of the double star, the resultant profile (one-D or two-D) is asymetric from one side of the major axis to the other. This is also generally the case for triple stars (although you can have configurations of three stars in which the alignment is symmetric across both the minor and major axis -- but these configs are much less common).

One way to measure the "symmetry" of an object is to perform a bi-symmetric spatial autocorrelation. Divide the object in two halfs, given by the minor axis. Rotate one-half 180 degrees and multiply the resultant pieces. The autocorrelation is then normalized by the original galaxy (squared). Mathematically, this is equivalent to:

We then have a pixel ratio distribution (ratio (i,j)) which can be evaluated statistically. In our case, we will compute the mean & median of the distribtion after throwing out the high and low values. We avoid pixels in which we suspect contamination from local stars, and we avoid the inner 3 arc seconds (i.e., the PSF).

In order to use low SNR points (which are significant in this exercise, since an absense of flux is meaningful), we account for the flux uncertainty in each pixel. This is done by computing a family of ratios for a given pixel:

The final ratio that comes from this set is the one closest to unity. Thus, for cases in which p and p* are lost in the noise, the ratio will be close to unity (thus, we err on the conservative side for this low SNR case), but for the case in which p or p* is high snr AND its counterpoint, p or p* accordingly, is low snr, the final ratio can still depart significantly from unity -- a signature of asymmetry.

In addition to the autocorrelation, we can also compute the ratio of total (integrated) flux between symmtric half pieces. This is a sort of "reduced" autocorrelation between the half pieces.

The final piece of information to glean from bi-symetric cross-correlation is a chi-square test (as proposed by Steve Schneider). The reduced chi-square can be written as:

In the words of Schneider,
"This has several advantages: (1) the distribution is well understood statistically with tabulated confidence ranges, (2) there are no asymmetries in the distribution like those introduced in a ratio comparison, (3) there are pretty-well understood schemes for rejecting outliers [which might be useful for double-nucleus sources and sources with superimposed stars], and (4) no data needs to be discarded because it is too close to zero (or negative)."


Case 1: 12th mag galaxy in the galactic plane

The following gif images shows the object, a real galaxy located near the galactic plane. The first panel shows the raw image (with previously processed objects blanked from the image), and the second panel shows the resultant image after stars have been subtracted. The regions in which stars have been subtracted, and regions in which there was a previously processed object, are not used in the autocorrelation or in the integrated flux ratio measurement.

What does the two-D elliptical profile look like for this object? The following plot shows the derived elliptical aperture for the J band image. The axis ratio is between 0.6 and 0.70, the position angle between 50 and 60 degrees, and the semi-major axis from 12.8 to 13.5 pixels (corresponding to the Kron radius). The red and the green points delineate the two symmetric halfs.

This is not the final aperture used, however. We want to avoid contamination from nearby stars or objects (re: galaxies) previously processed. So we mask those areas with suspected contamination (see first gif image). The red and the white points delineate the two symmetric halfs used in the autocorrelation and integrated flux operations. (note also that we avoid pixels with values lost in the noise -- not shown in the gif below).

Results

The autocorrelation (p*/p) distribution is shown in the histogram below.

This object -- a real galaxy, which appears to the eye quite symmetric -- is indeed spatially bisymmetric with a mean correlation ratio of 0.99 to 1.0. The integrated flux ratios are not quite so good (except for J), with ratios ranging from 0.76 to 0.99.

****** Alternate Method *************

Here we attempt a different filter method to remove extraneous low snr points. The approach is to use a integrated flux threshold criterion. Instead of checking each pixel to see if it has an SNR greater than some threshold, we integrate the pixel plus its four nearest neighbors and use the sum to compare with an N-sigma threshold. If sum(p) or sum(p*) greater than N-sigma, then the pixel combination (p,p*) is used in the ratio computation. Thus for galaxies, even low snr points will be included in the ratio computation because they "integrate up", whereas noise bumps around stars may not integrate up (depending on the severity of the N-sigma threshold).

Although this method should eliminate most low snr points (particularly around non-galaxies), some will still remain, which can skew the histogram (since we are taking the ratio of very small numbers). Consequently; as in the previous method, we allow a range in values for the ratio (with the range set by the background noise) -- see the intro of this memo for more details.


The autocorrelation (p*/p) distribution is shown in the histogram below.


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 1.0482
    H band: 1.0289
    K band: 1.1959


Case 2: 12.9 mag galaxy in the galactic plane

Results

    Histogram Statistics:
      J band : mean = 0.97 +- 0.10 , median = 0.99
      H band : mean = 1.05 +- 0.21 , median = 1.00
      K band : mean = 0.94 +- 0.21 , median = 0.99

    Integrated flux ratio:

      J band : 0.59
      H band : 0.78
      K band : 0.71


****** Alternate Method *************


The autocorrelation (p*/p) distribution is shown in the histogram below.

    Histogram Statistics:
      J band : mean = 0.96 +- 0.13 , median = 0.99
      after min/max rejection: J band : mean = 0.97 +- 0.10 , median = 1.00

      H band : mean = 1.05 +- 0.31 , median = 1.00
      after min/max rejection: H band : mean = 1.05 +- 0.21 , median = 1.00

      K band : mean = 0.94 +- 0.26 , median = 0.99
      after min/max rejection: K band : mean = 0.94 +- 0.20 , median = 0.99


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 1.0202
    H band: 1.0655
    K band: 1.1457


Case 3: 13.5 mag galaxy in the galactic plane

Results

    Histogram Statistics:
      J band : mean = 0.97 +- 0.19 , median = 1.00
      H band : mean = 0.99 +- 0.05 , median = 0.99
      K band : mean = 0.83 +- 0.53 , median = 0.98

    Integrated flux ratio:

      J band : 0.77
      H band : 0.78
      K band : 0.29


****** Alternate Method *************


The autocorrelation (p*/p) distribution is shown in the histogram below.

    Histogram Statistics:
      J band : mean = 0.94 +- 0.28 , median = 1.01
      after min/max rejection: J band : mean = 0.97 +- 0.18 , median = 1.00

      H band : mean = 0.99 +- 0.07 , median = 0.99
      after min/max rejection: H band : mean = 0.99 +- 0.05 , median = 1.00

      K band : mean = 0.83 +- 0.57 , median = 0.99
      after min/max rejection: K band : mean = 0.84 +- 0.52 , median = 0.99


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 0.7643
    H band: 0.6243
    K band: 1.1349


Case 4: Double Star in the plane

Results

    Histogram Statistics:
      J band : mean = 0.70 +- 0.38 , median = 0.88
      H band : mean = 0.68 +- 0.44 , median = 0.87
      K band : mean = 0.80 +- 0.35 , median = 0.97

    Integrated flux ratio:

      J band : 0.17
      H band : 0.13
      K band : 0.38


****** Alternate Method *************


The autocorrelation (p*/p) distribution is shown in the histogram below.

    Histogram Statistics:
      J band : mean = 0.71 +- 0.46 , median = 0.76
      after min/max rejection: J band : mean = 0.69 +- 0.38 , median = 0.76

      H band : mean = 0.69 +- 0.53 , median = 0.86
      after min/max rejection: H band : mean = 0.67 +- 0.44 , median = 0.86

      K band : mean = 0.80 +- 0.35 , median = 0.97
      after min/max rejection: K band : mean = 0.81 +- 0.33 , median = 0.97


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 3.5732
    H band: 3.5165
    K band: 1.6800


Case 5: Double Star in the plane

Results

    Histogram Statistics:
      J band : mean = 0.85 +- 0.29 , median = 0.97
      H band : mean = 0.91 +- 0.44 , median = 0.97
      K band : mean = 0.64 +- 0.88 , median = 0.97

    Integrated flux ratio:

      J band : 0.53
      H band : 0.53
      K band : 0.50


****** Alternate Method *************


The autocorrelation (p*/p) distribution is shown in the histogram below.

    Histogram Statistics:
      J band : mean = 0.86 +- 0.31 , median = 0.98
      after min/max rejection: J band : mean = 0.86 +- 0.29 , median = 0.97

      H band : mean = 0.94 +- 0.45 , median = 0.98
      after min/max rejection: H band : mean = 0.67 +- 0.44 , median = 0.86

      K band : mean = 0.63 +- 0.95 , median = 0.97
      after min/max rejection: K band : mean = 0.64 +- 0.89 , median = 0.97


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 1.7793
    H band: 1.8403
    K band: 1.5173


Case 6: Triple Star in the plane: compact config

Results

    Histogram Statistics:
      J band : mean = 1.01 +- 0.25 , median = 1.00
      H band : mean = 1.05 +- 0.35 , median = 1.00
      K band : mean = 0.65 +- 0.93 , median = 0.93

    Integrated flux ratio:

      J band : 0.95
      H band : 0.98
      K band : 0.97


****** Alternate Method *************


The autocorrelation (p*/p) distribution is shown in the histogram below.

    Histogram Statistics:
      J band : mean = 1.01 +- 0.28 , median = 1.00
      after min/max rejection: J band : mean = 1.01 +- 0.25 , median = 1.00

      H band : mean = 1.06 +- 0.41 , median = 1.00
      after min/max rejection: H band : mean = 1.04 +- 0.34 , median = 1.00

      K band : mean = 0.64 +- 1.63 , median = 0.94
      after min/max rejection: K band : mean = 0.75 +- 0.94 , median = 0.94


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 2.0325
    H band: 2.2965
    K band: 2.8208


Case 7: Triple Star in the plane: quasi-symmetric config

Results

    Histogram Statistics:
      J band : mean = 0.95 +- 0.11 , median = 1.00
      H band : mean = 0.96 +- 0.12 , median = 0.99
      K band : mean = 0.94 +- 0.14 , median = 0.99

    Integrated flux ratio:

      J band : 0.78
      H band : 0.76
      K band : 0.49


****** Alternate Method *************


The autocorrelation (p*/p) distribution is shown in the histogram below.

    Histogram Statistics:
      J band : mean = 0.95 +- 0.12 , median = 0.99
      after min/max rejection: J band : mean = 0.95 +- 0.11 , median = 0.99

      H band : mean = 0.95 +- 0.14 , median = 0.99
      after min/max rejection: H band : mean = 0.96 +- 0.13 , median = 0.99

      K band : mean = 0.93 +- 0.15 , median = 0.99
      after min/max rejection: K band : mean = 0.94 +- 0.14 , median = 0.99


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 1.3285
    H band: 1.2107
    K band: 1.4070


Case 8: Triple Star+ in the plane

Results

    Histogram Statistics:
      J band : mean = 1.33 +- 1.10 , median = 1.00
      H band : mean = 0.98 +- 0.68 , median = 0.99
      K band : mean = 1.22 +- 1.15 , median = 0.99

    Integrated flux ratio:

      J band : 0.63
      H band : 0.86
      K band : 0.35


****** Alternate Method *************


The autocorrelation (p*/p) distribution is shown in the histogram below.

    Histogram Statistics:
      J band : mean = 1.35 +- 1.29 , median = 1.00
      after min/max rejection: J band : mean = 1.31 +- 1.06 , median = 0.99

      H band : mean = 0.99 +- 0.76 , median = 0.99
      after min/max rejection: H band : mean = 0.98 +- 0.68 , median = 0.99

      K band : mean = 1.24 +- 1.37 , median = 1.00
      after min/max rejection: K band : mean = 1.21 +- 1.13 , median = 1.00


Chi-Sqr
The reduced chi is given below (square root of the chi-sqr):

    J band: 2.7638
    H band: 1.9643
    K band: 2.2148


Summary Plots

The max departure is computed by subtracting (or adding, case being) the errorbar value to the mean ratio value, and subtracting this result from unity. Thus it measures the departure from unity assuming a 1-sigma deviation from its mean value. This seems to work because galaxies have fairly symmetric ratios (i.e., close to unity) while their error bars are small. Doubles and trips have large errorbars (except the one case in which the trip is configured in a rather symmetric fashion) and have mean ratios significantly departing from unity (especially doubles).


****** Alternate Method *************


Reduced Chi (derived from the square root of the chi-sqr)



Algor Run on a Large Data Set

A data set consisting of 7 scans of a high source density field (glat ~ 5 degrees) have been processed with the bi-symmetric autocorr algorithm. This field is chock full of double and triple stars (and quad+ groupings of stars). Only a few galaxies were extracted, so the statistics are on the short side. Also, since the source density is high, the chance of stellar contamination to one of the galaxies is high -- thus we expect some degree of asymetry in the detected galaxies. Some relevant results are given below. For the full details, see GALWORKS Performance on a Low GLAT Field .

  • Flux Ratios
    galaxies == white filled circles
    doubles = red points
    triples = blue points

The next step is to apply a sliding scale to the flux ratio computation. Similar to the autocorrelation ratios, we allow some range in the integrated flux measurements based on the predicted uncertainty in the measurement.

    old flux measurement: flux ratio == INT ( p*) / INT (p)
    where INT (p) is the integrated flux in one of the symmetric pieces.

    new flux measurement:
    family (flux ratio) == INT ( p*) +- del_P* / INT (p) +- del_P
    where del_P is the expected uncertainty in the integrated flux measurement.

The flux ratio is then the closest measurement ( of the family of values) to unity.

Reduced Chi