Uploaded image for project: 'Data Management'
  1. Data Management
  2. DM-13800

Astrometry model should fix one sensor instead of one visit

    Details

      Description

      Jointcal's current astrometry model fixes one visit to be the identity, to break the sensor/visit model degeneracy. This leaves one visit unable to correct for focal plane/sky distortions. We should instead fix one sensor, fitting the rest of the sensors to that one.

      This may mean making one sensor's transform be the identity, or it may mean taking canonical values for one sensor but not letting them vary during the fit.

      Because I'm digging into the astrometry code now, I'm going to try to do this as part of DM-13272.

        Attachments

          Issue Links

            Activity

            Hide
            Parejkoj John Parejko added a comment -

            The idea is to separate out the chip positions/rotations (1st degree chip map: the chip motion should be purely affine) from the optics/atmosphere (>=3rd degree visit map). This does blend the hopefully stable optics with the varying atmosphere, but the eventual goal is to put in a third fixed (preferably radial) polynomial for the optics. Constraining the chips separately from the optics+atmosphere is still a constrained model: the atmosphere is smoothly varying across the whole focal plane, while the chips positions should be fixed with time (modulo the instrument being opened up).

            The above is the model that meas_mosaic uses to good effect for HSC: affine for the chips, and 5th (by default) degree for the visits.

            This model is different from Gary Bernstein's DECam model which has a single 4th degree polynomial for optics that they claim also incorporates the chip positions: I'm a bit skeptical of that, because the chip-to-chip positions aren't a smooth function, but really are discrete per chip.

            Show
            Parejkoj John Parejko added a comment - The idea is to separate out the chip positions/rotations (1st degree chip map: the chip motion should be purely affine) from the optics/atmosphere (>=3rd degree visit map). This does blend the hopefully stable optics with the varying atmosphere, but the eventual goal is to put in a third fixed (preferably radial) polynomial for the optics. Constraining the chips separately from the optics+atmosphere is still a constrained model: the atmosphere is smoothly varying across the whole focal plane, while the chips positions should be fixed with time (modulo the instrument being opened up). The above is the model that meas_mosaic uses to good effect for HSC: affine for the chips, and 5th (by default) degree for the visits. This model is different from Gary Bernstein's DECam model which has a single 4th degree polynomial for optics that they claim also incorporates the chip positions: I'm a bit skeptical of that, because the chip-to-chip positions aren't a smooth function, but really are discrete per chip.
            Hide
            jbosch Jim Bosch added a comment - - edited

            As John Parejko said, the primary reason I wanted to try this was because it's what meas_mosaic does.  Beyond that, I think there are good motivations for both approaches:

            • low-order visit fit and high-order ccd fit: separates terms that should be constant (e.g. optics) from those that should not (atmosphere), but (to remove degeneracies) imposes a constraint that the atmospheric component average down to the identity, which may only be true when the numbers of visits is large.  Also does not make use of our knowledge that the optical component is smooth over the full focal plane.
            • high-order visit fit and low-order ccd fit: separates terms that are smooth across the full focal plane from those that are not.  Because this approach does not try to break the degeneracy between the optics and atmosphere (which are perfectly degenerate when only a single visit is fit and there is no prior information about their spatial scales), I think this should have fewer problems with degeneracies.  However, it of course does not try to separate constant terms from those that vary per-epoch.

            I imagine we all agree that the best model would one with (at least) three terms: one for optics (full focal plane, does not change between visits), one for sensor offsets/rotations (does not change between visits), and one for the atmosphere (full focal plane, does change between visits), if we have a way to break the degeneracies in that model.

            Show
            jbosch Jim Bosch added a comment - - edited As John Parejko said, the primary reason I wanted to try this was because it's what meas_mosaic does.  Beyond that, I think there are good motivations for both approaches: low-order visit fit and high-order ccd fit: separates terms that should be constant (e.g. optics) from those that should not (atmosphere), but (to remove degeneracies) imposes a constraint that the atmospheric component average down to the identity, which may only be true when the numbers of visits is large.  Also does not make use of our knowledge that the optical component is smooth over the full focal plane. high-order visit fit and low-order ccd fit: separates terms that are smooth across the full focal plane from those that are not.  Because this approach does not try to break the degeneracy between the optics and atmosphere (which are perfectly degenerate when only a single visit is fit and there is no prior information about their spatial scales), I think this should have fewer problems with degeneracies.  However, it of course does not try to separate constant terms from those that vary per-epoch. I imagine we all agree that the best model would one with (at least) three terms: one for optics (full focal plane, does not change between visits), one for sensor offsets/rotations (does not change between visits), and one for the atmosphere (full focal plane, does change between visits), if we have a way to break the degeneracies in that model.
            Hide
            astier Pierre Astier added a comment -

            I agree that conceptually going from pixels to focal plane, from focal plane angles in the pupil, and then to angles on the sky  makes a lot of sense conceptually. However, I have never seen a proposal to lift degeneracies. The advantage I see of separating the focal plane positions form from the optical distortions is that the parametrization of the latter could take advantage of the approximate azimuthal symmetry (not easy because it has to be approximate). However, this requires to fit the position of the optical center which is highly non-linear. In the end, the "constrained model" has less parameters per input calexp than a regular WCS fit. We can certainly reduce the number of parameters used to map pixels onto the tangent plane (the static part of the mappings), but I am not sure to anticipate what the practical benefit is.

            Show
            astier Pierre Astier added a comment - I agree that conceptually going from pixels to focal plane, from focal plane angles in the pupil, and then to angles on the sky  makes a lot of sense conceptually. However, I have never seen a proposal to lift degeneracies. The advantage I see of separating the focal plane positions form from the optical distortions is that the parametrization of the latter could take advantage of the approximate azimuthal symmetry (not easy because it has to be approximate). However, this requires to fit the position of the optical center which is highly non-linear. In the end, the "constrained model" has less parameters per input calexp than a regular WCS fit. We can certainly reduce the number of parameters used to map pixels onto the tangent plane (the static part of the mappings), but I am not sure to anticipate what the practical benefit is.
            Hide
            Parejkoj John Parejko added a comment -

            Done in DM-13272

            Show
            Parejkoj John Parejko added a comment - Done in DM-13272
            Hide
            Parejkoj John Parejko added a comment -

            Note that although I've implemented this as part of DM-13272 (now in review), please don't mean that should be the end of the conversation.

            In particular, I am curious how we would deal with the focal->optics vs. optics->sky degeneracy that Pierre mentions above. I think it's equivalent to the "Exposure/Instrument trades" (section 2.3.5) in Bernstein et al., but I think that's also equivalent to the degeneracy we break by fixing one of the chip transforms, and we can't do that twice. I do wonder whether a radial polynomial would be non-degenerate enough by itself, but I also haven't looked at how difficult a radial polynomial would make the model derivatives.

            Show
            Parejkoj John Parejko added a comment - Note that although I've implemented this as part of DM-13272 (now in review), please don't mean that should be the end of the conversation. In particular, I am curious how we would deal with the focal->optics vs. optics->sky degeneracy that Pierre mentions above. I think it's equivalent to the "Exposure/Instrument trades" (section 2.3.5) in Bernstein et al., but I think that's also equivalent to the degeneracy we break by fixing one of the chip transforms, and we can't do that twice. I do wonder whether a radial polynomial would be non-degenerate enough by itself, but I also haven't looked at how difficult a radial polynomial would make the model derivatives.

              People

              • Assignee:
                Parejkoj John Parejko
                Reporter:
                Parejkoj John Parejko
                Watchers:
                Dominique Boutigny, Jim Bosch, John Parejko, John Swinbank, Pierre Astier
              • Votes:
                0 Vote for this issue
                Watchers:
                5 Start watching this issue

                Dates

                • Created:
                  Updated:
                  Resolved:

                  Summary Panel