Uploaded image for project: 'Data Management'
  1. Data Management
  2. DM-16875

Test two-component Gaussian mixture models on COSMOS galaxies

    XMLWordPrintable

    Details

    • Type: Story
    • Status: Done
    • Resolution: Done
    • Fix Version/s: None
    • Component/s: None
    • Labels:
      None
    • Story Points:
      2
    • Sprint:
      DRP S19-6b
    • Team:
      Data Release Production

      Description

      Test galaxy models consisting of two elliptical components with their radial profiles describe by N Gaussians on COSMOS (HST, then HSC).

        Attachments

          Activity

          Hide
          dtaranu Dan Taranu added a comment - - edited

          The model specifications that I'm fitting now include a number of such models, of which "mg4x2px" is the most general (two free ellipses, each described by four Gaussians). The table in this notebook suggests (unsurprisingly) that it is the best-fitting model for most galaxies out of a sample of ~2000:

          profit.hst.mg4x2px.chisqred.1 1417
          profit.hst.mg8bpx.chisqred.1 194
          profit.hst.mg4bpx.chisqred.1 143
          profit.hst.mg8serserbpx.chisqred.1 74
          profit.hst.serb.chisqred.0 67
          profit.hst.serbpx.chisqred.0 58
          profit.hst.mg4serserbpx.chisqred.1 11

          There are also some plots with direct comparisons to double Sersic fits, showing that mg4x2px provides small but non-negligible improvements in reduced chi-squared for most galaxies. I expect the improvements would be larger in multi-band fitting as Gaussian mixtures allow for color gradients within components.

          Having said that, mg4x2px has 30-odd free parameters and is extremely expensive to fit. The fact that 'mg4bpx' - which is the equivalent model with just one component - is a better fit ~10% of the time despite being strictly less complex also indicates that the two-component variant is difficult to fit robustly (unsurprisingly). While there are less complex variations that we could test (e.g. with a scale radius and fixed size ratios instead of four free sizes, as implemented in DM-18585), we are focusing on choosing and optimizing single-component models for the near future. After that, we might take a closer look at using these more complex models as the ground truth, especially for multi-band fitting.

          Show
          dtaranu Dan Taranu added a comment - - edited The model specifications that I'm fitting now include a number of such models, of which "mg4x2px" is the most general (two free ellipses, each described by four Gaussians). The table in this notebook suggests (unsurprisingly) that it is the best-fitting model for most galaxies out of a sample of ~2000: profit.hst.mg4x2px.chisqred.1 1417 profit.hst.mg8bpx.chisqred.1 194 profit.hst.mg4bpx.chisqred.1 143 profit.hst.mg8serserbpx.chisqred.1 74 profit.hst.serb.chisqred.0 67 profit.hst.serbpx.chisqred.0 58 profit.hst.mg4serserbpx.chisqred.1 11 There are also some plots with direct comparisons to double Sersic fits , showing that mg4x2px provides small but non-negligible improvements in reduced chi-squared for most galaxies. I expect the improvements would be larger in multi-band fitting as Gaussian mixtures allow for color gradients within components. Having said that, mg4x2px has 30-odd free parameters and is extremely expensive to fit. The fact that 'mg4bpx' - which is the equivalent model with just one component - is a better fit ~10% of the time despite being strictly less complex also indicates that the two-component variant is difficult to fit robustly (unsurprisingly). While there are less complex variations that we could test (e.g. with a scale radius and fixed size ratios instead of four free sizes, as implemented in DM-18585 ), we are focusing on choosing and optimizing single-component models for the near future. After that, we might take a closer look at using these more complex models as the ground truth, especially for multi-band fitting.
          Hide
          yusra Yusra AlSayyad added a comment -

          Recommendation?

          Show
          yusra Yusra AlSayyad added a comment - Recommendation?

            People

            Assignee:
            dtaranu Dan Taranu
            Reporter:
            dtaranu Dan Taranu
            Reviewers:
            Yusra AlSayyad
            Watchers:
            Dan Taranu, Yusra AlSayyad
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

              Dates

              Created:
              Updated:
              Resolved:

                Jenkins

                No builds found.