IPP Software Navigation Tools IPP Links Communication Pan-STARRS Links

Ignore:
Timestamp:
Jul 14, 2017, 5:55:11 AM (9 years ago)
Author:
eugene
Message:

rework paper to focus on the tree-rings

File:
1 edited

Legend:

Unmodified
Added
Removed
  • trunk/doc/release.2015/systematics.20140411/systematics.tex

    r39833 r40096  
    104104\section{Pan-STARRS1}
    105105
    106 The Pan-STARRS\,1 telescope (PS1) has recently completed observations
    107 for its first survey mission (REF).  PS1 is 1.8m telescope located on
    108 the summit of Haleakala on the Hawaiian island of Maui (REF).  This
    109 wide-field telescope images a 3.3 degree field of view on a 1.4
    110 gigapixel camera, with low distortion and generally good image quality
    111 (median \approx \note{1.1 arcseconds} in all filters except for g,
    112 with 1.3 arcseconds).  Available filters are \gps,\rps,\ips with
    113 bandpasses similar to the equivalent SDSS filters; \zps is somewhat
    114 bluer than the SDSS $z$, while \yps is somewhat redder (see, REF, for
    115 a complete specification).
    116 
    117 The first PS1 science survey mission began 2010 May and completed 31
    118 March 2014.  The bulk of the observing time (56\%) goes to the
    119 ``3$\pi$ Survey'', in which the \approx 30,000 square degrees
    120 observable from Hawaii (north of Dec = -30) were repeatedly observed
    121 over the nearly 4 year survey period.  The typical coverage at the end
    122 of the survey is \approx 10 observations per filter per point on the
    123 sky in the 3$\pi$ region.  The repeated observations allows for good
    124 characterization of systematic sources of uncertainty.
    125 
    126 The PS1 data will be released to the public in the Spring of 2015 via
    127 the Mikulsky Archive to Space Telescopes (MAST) at the Space Telescope
    128 Science Institute (STScI).
    129 
    130 \subsection{GigaPixel Camera\,1}
    131 
    132 The PS1 GigaPixel Camera\,1 (GPC1) was the largest astronomical camera
    133 until HypersuprimeCam was completed.  GPC1 is a mosaic camera
    134 consisting of 60 detectors (``chips'') arranged in an 8x8 grid with
    135 the 4 corners missing (FIGURE?).  The detectors in this camera are
    136 othogonal transfer array devices (REF), in which the single silicon
    137 chip is sub-divided into a checkerboard of 64 sub-arrays (``cells'').
    138 Each cell may be independently addressed and read, and the charge
    139 accumulated in the pixels may be moved in either the $x$ or $y$
    140 directions (thus the term ``orthogonal transfer'').  The full GPC1
    141 thus consists of a total of 3840 cells, each \note{608 x 610} pixels. 
    142 
    143 \subsection{Data Processing}
    144 
    145 Images obtained from PS1 are written to a set of computers located in
    146 the observatory dome (``pixel servers''), and notification is then
    147 sent via a web service to the Image Processing Pipeline (IPP)
    148 computers in Kihei.  The IPP cluster retrieves images from the summit
    149 pixel servers as they are available, though no summit hand-shaking
    150 takes place.  The summit computers have a buffer for data from several
    151 days' of observing so that data is not lost if the link to Kihei is
    152 down or the IPP computer cluster is offline.
    153 
    154 Raw images are saved on the IPP cluster and processed as needed.  In
    155 general, every science exposure is processed as soon as possible in a
    156 stream called the ``nightly science processing''.  As the survey has
    157 proceeded and improvements have been made in the analysis software,
    158 images have been reprocessed.  To date (2015 Jan), there have been 2
    159 complete reprocesssing runs (identified PV1 and PV2), and a third
    160 (PV3) is in progress.  As the software has been improved, the nightly
    161 science processing software has been updated to reflect those
    162 improvements.  The nightly science processing thus is heterogenous in
    163 the analysis and algorithms used.  The software for the PVx
    164 reprocessing is more homogenous.
    165 
    166 Raw images are stored as multi-extension FITS images, with each cell
    167 in a given chip stored as a separate image extension. The raw data are
    168 generally stored with lossless compression applied to the pixel
    169 arrays \note{CFITSIO REF}.  The data for the 60 chips in a given
    170 exposure are stored as separate FITS files.
    171 
    172 The data processing system performs a number of steps distributed
    173 across a cluster of linux-based computers.  The first stage (``chip
    174 processing'') is performed on individual chip image files
    175 independently in parallel.  At this stage, the images are detrended,
    176 including overscan, bias, dark, flat-field, and (for $y$-band)
    177 fringing corrections.  Areas of bad data (e.g., regions of poor charge
    178 transfer efficiency or non-linear dark glows) are masked.  Within the
    179 IPP processing code, masked pixels are identified as non-zero values
    180 in a separate 16-bit mask image, which is then carried as part of the
    181 processed image data products.  The chip analysis also generates a
    182 variance image which is used along with the signal image to specify
    183 the noise properities as a function of pixel.  Details of the
    184 chip-stage processing can be found in \note{Waters et al REF}.
    185 
    186 The chip-stage analysis also includes source detection and basic
    187 characterization.  After a sky background model is subtracted from the
    188 image, individual sources are detected via cross-correlation with a
    189 PSF model.  Sources detected in the image are used to define the PSF
    190 model.  Simple aperture based measurements (total flux, 2nd and
    191 higher-order moments, etc) are measured for all detections.  An
    192 estimate of the extendedness (non-PSF-nature) of the sources is used
    193 to distinguish ``stellar'' detections from ``non-stellar'' detections.
    194 Non-stellar detections above a minimum signal-to-noise \note{of 20}
    195 are fitted with a galaxy surface brightness model, while remaining
    196 sources are fitted with a PSF model, or a cluster of PSF models.  The
    197 end result of the chip-stage analysis is a collection of parameters
    198 for each detection, including the positions ($X$, $Y$), PSF model
    199 flux, aperture fluxes, 2nd moments.  Details of the source analysis
    200 and characterization can be found in \note{Magnier et al REF}.
    201 
    202 \section{Photometry Analysis}
    203 
    204 Lorem ipsum dolor sit amet, consectetur adipiscing elit. Vestibulum
    205 bibendum nisi id tristique posuere. Duis eu mollis nulla. Maecenas est
    206 turpis, mattis tempor urna vitae, placerat rhoncus sem. Lorem ipsum
    207 dolor sit amet, consectetur adipiscing elit. Sed quis velit
    208 nisl. Aliquam erat volutpat. Cras lacinia, nisl tristique auctor
    209 molestie, dolor nulla rhoncus purus, ac accumsan nunc nunc ac
    210 nibh. Maecenas vitae mollis mauris. Ut sollicitudin pulvinar purus,
    211 eget luctus lorem tincidunt vitae. Vestibulum eu mattis neque. Nulla
    212 in tortor id urna dapibus gravida a vel leo.
    213 
    214 % refer back to refcat, ubercal, tonry
    215 % discuss the PSF modeling
    216 % what are the limitations of the current photometric precision
    217 
    218 \section{Astrometry Analysis}
    219 
    220 Astrometric calibration defines a transformation from the raw
    221 instrumental positions (X,Y on a chip) to RA,DEC coordinates on the
    222 sky.  Astrometry calibration is performed at two separate stages in
    223 the IPP analysis.  First, every exposure is calibrated independently
    224 during the processing by comparison to a reference catalog.  This
    225 astrometric analysis is performed as part of the second processing
    226 stage, the ``camera'' analysis stage.  The output from the chip stage
    227 for all chips in an exposure are processed as a group in this
    228 analysis.  The initial guess for the astrometry comes from the
    229 telescope coordinates provided in the image headers. 
    230 
    231 The RA, DEC, and position angle of the telescope boresite are provided
    232 as part of the image metadata.  The telescope and camera systems also
    233 provide WCS header keywords for each chip separately.  However, in the
    234 early stages of the telescope commissioning, these values were not
    235 reliable.  Thus, within the IPP, and model for the camera layout is
    236 used so that the single triplet (RA,DEC,PA) can be used to predict the
    237 astrometric calibration of each chip.  This guess is generally good to
    238 10s of pixels, and need only be accurate to \approx 1/2 of a chip for
    239 the software to discover the true coordinates.
    240 
    241 The guess astrometry is used to select a set of likely reference stars
    242 from the reference database.  These stars are the cross-correlated
    243 with the measured source positions to find the likely match between
    244 stars in the reference catalog and detected sources in the image.
    245 
    246 Once a reliable match has been determined, astrometric transformations
    247 are determined to match the raw (X,Y) coordinates of the detections to
    248 sky coordinates of the corresponding reference stars.
    249 
    250 Within the IPP, two main classes of transformations may be used.  For
    251 single chips from non-mosaic cameras (or for the warp \& stack
    252 skycells), the astrometry calibration consists of a projection from the
    253 celestial sphere to a linearized coordinate system followed by a
    254 transformation from the projection coordiante frame to the pixel
    255 coordinate frame.  The projection may be one of several possible
    256 options (see Calabria et al REF), though in general the IPP uses the
    257 SIN (or TAN) projection (define?).  The transformation from the
    258 projection coordiates to the pixel coordinate may use an afine
    259 transformation or higher order polynomials (up to 3rd order).
    260 
    261 For chips from a mosaic camera, the astrometric transformation is
    262 defined as a multi-level operation.  As in single-chip astrometry, a
    263 projection (SIN or TAN) is used to convert the spherical celestial
    264 coordiates to a local linear system (the 'tanget plane' coorinate
    265 system).  There are then two level of cartesian transformations: a
    266 first set of polynomials (up to 3rd order) are used to transform the
    267 tangent plane coordinates to the ``focal plane'' coordinates.  This,
    268 relatively low-order, correction accounts for rotation of the camera
    269 and basic optical distortion coming from the optics and the
    270 atmosphere.  A second transformation is used to convert the focal
    271 plane coordinate to the pixel coordinates for each chip.  For the
    272 on-the-fly calibration, these
    273 transformations may consist of a set of polynomials up
    274 to 3rd order in X and/or Y (i.e, $x^i y^j$ where $i + j <= 3$)
    275 
    276 \note{add details on the transformation from focal-plane to tangent
    277   plane: fit is done on the gradient}.
    278 
    279 The on-the-fly astrometric calibration determines the astrometric
    280 transformation to an accuracy of at least 0.3 arcsec (chips with worse
    281 astrometry are rejected).  Data from these images may now be ingested
    282 into our internal database software for astronomical objects, the
    283 Desktop Virtual Observatory \citep[DVO,][]{PS1.IPP}.  This database
    284 software associates detections from images together based on their
    285 positions in the sky to define astrnomical ``objects''.  The database
    286 includes metadata to describe the links between objects and their
    287 multiple detections as well as metadata defining the images and their
    288 astrometric transformations. 
    289 
    290 After images have been ingested into the DVO software, calibration
    291 improvements may be made within the database.  A program called
    292 ``relastro'' is used to improve the astrometry iterative.  In one
    293 mode, relastro loops between improving the mean positions of objects
    294 and using the improved mean positions to improve the astrometric
    295 calibrations of the images.  In normal usage, relastro is only allowed
    296 to update the chip-to-focal plane transformations; it is not normally
    297 used to improve the focal-plane to tangent plane transformations.
    298 Also, within relastro, the transformation from chip-to-focal plane may
    299 be represented using either the polynomials as above or with a linear
    300 transformation plus bilinear interpolation of a grid of correction
    301 cells.  In the latter case, the grid may be sampled as finely as 6x6
    302 positions per chip. 
    303 
    304 % precision demonstration
    305 % koppenhoefer effect
    306 % MISSING: DCR!
    307 
    308 \section{Systematic Residuals}
    309 
    310 \subsection{Camera-Scale Trends}
    311 
    312 \subsection{Tree-Rings : An Example Chip}
     106\note{tidy up this section}
     107
     108The 1.8m Pan-STARRS\,1 telescope (PS1), located on the summit of
     109Haleakala on the Hawaiian island of Maui, has been surveying the sky
     110regularly since May 2010 \citep{chambers.2017}.  From May 2010 through
     111March 2014, PS1 was run under the aegis of the Pan-STARRS Science
     112Consortium to perform a set of wide-field science surveys; since March
     1132014, the telescope is operated by the Pan-STARRS New Science
     114Consortium (PSNSC).  Under the PS1SC, the largest survey, both in
     115terms of area of the sky covered and fraction of observing time
     116(56\%), was the \TPS\ in which the entire sky north of Declination
     117$-30$\degrees\ was imaged up \approx 80 times over the 4 years.  These
     118observations were distributed over five filters, \grizy, and have been
     119astrometrically and photometrically calibrated to good precision
     120\citep{magnier.2017.calibration}.
     121
     122The wide-field PS1 telescope optics \citep{PS1.optics} image a 3.3
     123degree field of view on a 1.4 gigapixel camera \citep[GPC1][]{PS1.GPC1}, with
     124low distortion and generally good image quality.  The median seeing
     125for the \TPS\ data vary somewhat by filter, with (\grizy) = (XXXX).
     126Routine observations are conducted remotely from the Advanced
     127Technology Research Center in Kula, the main facility of the
     128University of Hawaii's Institute for Astronomy operations on Maui.
     129
     130GPC1 \citep{PS1.GPCA}, currently the largest astronomical camera in
     131terms of number of pixels, consists of a mosaic of 60 edge-abutted
     132$4800\times4800$ pixel detectors, with 10~$\mu$m pixels subtending
     1330.258~arcsec. These \note{OTA51} detectors, manufactured by Lincoln
     134Laboratory, are \note{75$\mu$m}-thick back-illuminated CCDs with a
     135readout time of 7 seconds for a full unbinned image. \note{details
     136  about the voltages?}  Initial performance assessments are presented
     137in \cite{PS1.GPCB}. The active, usable pixels cover $\sim 80$\% of the
     138FOV.
     139
     140\subsection{Data Processing and Calibration}
     141
     142Images obtained by PS1 are processed by the Pan-STARRS Image
     143Processing Pipeline (IPP; \citealp{PS1_IPP,
     144  magnier.etal.2016.datasystem}).  All observations are processed
     145nightly, with results sent to groups within the science consortium
     146(i.e., PS1SC during the \TPS) performing short-term science projects
     147(e.g., searching for transient and moving objects).  In addition, the
     148\TPS\ dataset has been re-processed several times with improved
     149calibration and analysis techniques.  To date (2017 July), 3
     150re-processings starting from raw pixel data have been performed.  The
     151labels PV0, PV1, PV2, PV3 are used identify the nightly processing and
     152successive re-processing versions.  PV3 has been used for the public
     153release of the Pan-STARRS \TPS\ data via the {\it Barbara A. Mikulski
     154  Archive for Space Telescopes} (MAST) at the Space Telescope Science
     155Institute.\footnote{http//panstarrs.stci.edu}
     156
     157The data processing and calibration operations are discussed in detail
     158in elsewhere
     159\citep{magnier.etal.2017.analysis,magnier.etal.2017.calibration,waters.2017}.
     160We re-visit here a number of points that are of significance to this
     161study.  Images are processed following a fairly standard sequence of
     162image detrending, source detection, and initial calibration
     163(astrometric and photometric) of those detected sources.  Additional
     164standard processing critical to PS1 science operations includes
     165geometric transformation (`warping') and image combinations (summed
     166stacks and differences).  For the purposes of this analysis, we are
     167only considering the sources detected in the individual exposures from
     168the initial analysis steps.
     169
     170As discussed in \cite{waters.2017}, image detrending includes
     171flat-field processing with a single epoch flat-field image for each
     172filter.  The flat-field image used for this analysis has been
     173generated by median-combining dome flat-field images (after
     174pre-processing and pixel outlier rejections) and then multiplying by a
     175photometric flat-field correction image generated by the analysis of a
     176grid of images of a dense stellar field.  The purpose of this second
     177step is to correct the basic flat-field image for errors arising from
     178the non-uniformity of the illumination, from non-pixel uniformity due
     179to the varying optical distorition across the field, and any other
     180factors which may make the flat-field image inconsistent with stellar
     181photometry, e.g., SED, filter band-pass variations, etc
     182\citep[see][]{waters.2017,magnier.cuillandre,magnier.belgium}.  This
     183correction was made on a relatively coarse grid across the focal plane
     184in order to accumulate sufficient statistics from the stars in the
     185relatively small number of images available at the time.  We have
     186found that a single flat-field set can be used for all PS1
     187observations to yield photometric consistency at the level of \approx
     1882\% \note{use the ubercal flat stdev as a statistic}.  PS1 benefits in
     189this regard from the stability of having a single instrument which is
     190rarely removed. 
     191
     192Photometry of the PS1 images is performed using a
     193point-spread-function (PSF) model as well as multiple kinds of
     194apertures \citep{magnier.etal.2017.analysis}.  In this analysis, we
     195refer to aperture photometry performed using an aperture defined based
     196on the image quality observed for a given chip.  The aperture diameter
     197is set to be \note{XXX} times the FWHM for the image.
     198
     199To improve the photometric systematic errors beyond the level achieved
     200with a single (photometrically corrected) flat-field set, the PS1
     201photometry is re-calibrated within the databasing system based on the
     202properties of the measured photometry.  The calibration process is
     203discussed by \cite{ubercal,photladder,magnier.etal.2017.calibration}.
     204As part of this process, several flat-field corrections have been
     205determined.  For the PV2 analysis discussed here, a flat-field
     206correction determined during the ubercal analysis
     207\citep[see][]{ubercal} consisted of an $8\times 8$ grid of corrections
     208for each GPC1 chip and filter for each of 4 seasons.  The boundaries
     209of those seasons are \note{tentatively} identified with modifications
     210to the baffle structures or the system optics.  The critical point
     211here is that the final effective flat-field image for the PV2 dataset
     212is based on a dome-flat at the highest resolution, with very low
     213resolution corrections based on photometry, resulting in photometric
     214calibration with roughly 1 millimag consistency for each measurement
     215\note{better number from ubercal?}.
     216
     217For all objects, positions are measured from the PSF model for the
     218brighter sources (using a non-linear fitting process) and from a
     219simple centroid (1st moment) for the fainter source
     220\citep{magnier.etal.2017.analysis}.  These position measurements are
     221used in the astrometric analysis.  The astrometric calibration is
     222discussed by \cite{magnier.etal.2017.calibration}; for the PV2
     223dataset, the typical systematic floor is \approx 15 - 20
     224milliarcsecond for individual measurements of brighter stars.
     225
     226\section{Tree-Ring-Like Patterns}
    313227
    314228\begin{table}
     
    331245\end{table}
    332246
    333 We observe a number of low-level effects in different types of
    334 measurements which have a similar spatial structure on individual
    335 chips.  These structures have a circular pattern centered one corner
    336 of the affected chips.  \note{do all chips show all effects?  is the
    337   amplitude very different from chip to chip?}  We use measurements
    338 from chip XY40 to illustrate the spatial patterns and relationships
    339 between the different effects.  For all effects, we are measuring the
    340 mean value of the effect in 10x10 pixel boxes.  The resulting images
    341 are all constructed so that a given superpixel represents the same
    342 range of true GPC1 XY40 pixels.  Measurements were extracted from the
    343 ``nightly science'' DVO database for observations covering the region
    344 ($\alpha$,$\delta$) = (90\degree\ -- 150\degree, -25\degree\ --
    345 10\degree).  This region avoids the Galactic Plane where astrometric
    346 outliers have been more common.  We limit the analysis to good
    347 measurements (PSF\_QF $>$ 0.85) of likely stars ($|m_{psf} - m_{aper}|
    348 < 0.2$).  Only measurements with instrumental magnitude $< -8.0$
    349 ($-2.5\log \mbox{cts sec}^{-1} < -8.0$) are include to ensure
     247For many of the GPC1 OTA CCDs, we observe a pattern in the photometric
     248residuals which is similar in appearence to the Tree Rings described
     249in the Dark Energy Camera (DECam) by \cite{plazas.2014}.  This pattern
     250consists of systematic deviations which are consistent in a set of
     251circular arcs centered on the corner of the CCD, as shown in
     252Figure~\ref{fig:psfmags.by.filter}.  The details of the analysis used
     253to generate Figure~\ref{fig:psfmags.by.filter} are given below.  For
     254now, we note that the GPC1 CCDs are constructed by dividing the
     255circular silicon wafer into 4 inscribed squares.  Thus the corners of
     256the CCDs lie in the center of the silicon boule, just as the center of
     257the circular Tree Rings described by \cite{plazas.2014} match the
     258center of the boule from which they came.  This gives the impression
     259that a similar mechanism is responsible for the pattern observed in
     260the PS1 photometry and the DECam photometry, namely the diffusive
     261effects of lateral electric field variations in the detectors.  In the
     262next section, we will make the case that the patterns observed in the
     263PS1 residuals are {\em not} caused by this mechanism, but are instead
     264caused by variations in the {\em vertical} electric field (the field
     265direction perpendicular to the CCD surface). 
     266
     267First, in this section, we will describe how we have measured the
     268presence or absence of these tree-ring patterns in 5 types of data.
     269For all of these examples, we use a single GPC1 CCD (XY40) to
     270illustrate the effects in detail, but a similar set of effects are
     271seen in \note{many? most?} GPC1 detectors.  First, we show the
     272residual PSF photometry.  Second, we show the residual Aperture
     273photometry.  Third, we show the astrometric residual patterns.
     274Fourth, we show the patterns observed in the flat-field images.
     275Finally, we show measurements derived from the second-moments of the
     276stars.
     277
     278For all effects discussed below, we are measuring the mean value of
     279the effect in 10x10 pixel superpixels across the detector.  The
     280resulting images are all constructed so that a given superpixel
     281represents the same range of true GPC1 XY40 pixels regardless of the
     282type of measurement.  To generate the photometry, astrometry, or
     283second-moment measurements were extracted from the \note{PV0} DVO
     284database for observations covering the region ($\alpha$,$\delta$) =
     285(90\degree\ -- 150\degree, -25\degree\ -- 10\degree).  This region of
     286the sky provides a fairly high density of stars, but avoids the
     287Galactic Plane where confusion may potentially contaminate the
     288measurement.  We limit the analysis to good measurements
     289(\ippmisc{PSF_QF} $>$ 0.85) of likely stars ($|m_{psf} - m_{aper}| <
     2900.2$).  Only measurements with instrumental magnitude $< -8.0$
     291($-2.5\log \mbox{cts sec}^{-1} < -8.0$) are included to ensure
    350292reasonable signal-to-noise per measurement.  We require at least 2
    351293measurements in a given filter and 5 measurements total for any star
    352294included in the analysis.
    353 
    354 The following four different measurements show tree-ring structures
    355 (a) photometric residuals, (b) astrometric residuals, (c) a portion of
    356 the flat-field structure, and (d) variations in the second-moment of
    357 stars.  In the following section, we show the spatial patterns for
    358 these features and measure their intensity as a function of the
    359 different filters.  By comparing the spatial structures, we show that
    360 these effects are directly related.  We defer for now discussion of
    361 any causes of the observed effects.
    362295
    363296\subsubsection{Photometric Residuals}
     
    397330\end{figure*}
    398331
    399 The tree-ring structure is clearly seen in the PSF magnitude
    400 residuals.  In this case, we select PSF magnitude measurements for
    401 detections which fall in the given superpixel.  We subtract each
    402 measurement from the average magnitude for the object in the selected
    403 filter ($\delta m_{psf} = \overline{m}_{psf} - m_{psf}$) to determine the
    404 residual magnitude, excluding as bad any measurement with $|\delta
    405 m_{psf}| > 0.5$.  For a given superpixel, we measure the median of the
    406 $\delta m_{psf}$ distribution.  Figure~\ref{fig:psfmags.by.filter}
    407 shows the 2D patterns of $\delta m_{psf}$ for each filter (\grizy).
    408 The dynamic range of the color scale is from -20 to +20
    409 millimagnitudes for all 5 plots. 
     332Figure~\ref{fig:psfmags.by.filter} shows the 2D patterns of PSF
     333photometric residuals.  In this case, we select PSF magnitude
     334measurements for detections of stars which fall in the given
     335superpixel.  We subtract each measurement from the average magnitude
     336for the object in the selected filter ($\delta m_{psf} =
     337\overline{m}_{psf} - m_{psf}$) to determine the residual magnitude,
     338excluding as an outlier any measurement with $|\delta m_{psf}| > 0.5$.
     339For a given superpixel, we measure the median of the $\delta m_{psf}$
     340distribution.  The figure shows $\delta m_{psf}$ for each filter
     341(\grizy).  The dynamic range of the color scale is from -20 to +20
     342millimagnitudes for all 5 plots.
    410343
    411344The tree-ring pattern is clearly visible for the four blue filters,
     
    418351is comparable to the amplitude of the correlated structures, so we
    419352need to integrate along the radial structures to make stronger
    420 statements about these patterns.
    421 
    422 We have also performed the same measurement for aperture magnitudes,
    423 using the same selections.  The 2D patterns for the aperture
    424 magnitudes is shown in Figure~\ref{fig:apmags.by.filter}.  The finging
     353statements about these patterns. \note{hanging statement?}
     354
     355Figure~\ref{fig:apmags.by.filter} shows the equivalent measurement for
     356aperture photometry instead of PSF photometry.  The finging
    425357pattern again dominates the plot for \yps, but the tree-rings are not
    426358seen in any of the filters.  A diagonal pattern is visible in \gps
     
    450382\end{figure*}
    451383
    452 The tree-ring structure is also clearly seen in the astrometric
    453 residuals.  We use the same selection of measurements for astrometry
    454 as for photometry.  In this case, we extract the residual in both the
    455 RA and DEC directions ($\delta RA = \overline{RA} - RA_i$, $\delta DEC
    456 = \overline{DEC} - DEC_i$) and rotate these values to the chip
    457 coordinate system ($\delta X,\delta Y$).  We again exclude as bad any
    458 measurement with $|\delta X|$ or $|\delta Y| > 0.5$ arcsec before
    459 measuring the median values for each superpixel.  We have determined
    460 the approximate center of the circular tree-ring pattern as (-5,4960)
    461 for this particular chip.  Using this coordinate as the center, we
    462 have converted the $\delta X,\delta Y$ offsets into $\delta R,\delta
    463 \theta$ measurements ($\delta R$ : radial component away from the
    464 center, $\delta \theta$ : tangential component).
     384Figure~\ref{fig:astrom.by.filter} shows a similar type of measurement
     385for astrometric residuals.  To generate this plot, we use the same
     386selection of measurements for astrometry as for photometry.  In this
     387case, we extract the residual in both the RA and DEC directions
     388($\delta RA = \overline{RA} - RA_i$, $\delta DEC = \overline{DEC} -
     389DEC_i$) and rotate these values to the chip coordinate system ($\delta
     390X,\delta Y$) using our knowledge of the chip orientation on the sky.
     391We again exclude as bad any measurement with $|\delta X|$ or $|\delta
     392Y| > 0.5$ arcsec before measuring the median values for each
     393superpixel.  We have determined the approximate center of the circular
     394tree-ring pattern as (-5,4960) for this particular chip.  Using this
     395coordinate as the center of the pattern, we have converted the $\delta
     396X,\delta Y$ offsets into $\delta R,\delta \theta$ measurements
     397($\delta R$ : radial component away from the center, $\delta \theta$ :
     398tangential component).
    465399
    466400Figure~\ref{fig:astrom.by.filter} shows the 2D patterns of $\delta R$
    467401for each filter (\grizy).  The dynamic range of the color scale is
    468 from -20 to +20 milliarcseconds for all 5 plots.  The tree-ring
    469 pattern is visible for all five filters; the finging pattern is not
    470 apparent in the \yps\ astrometry.  \note{low-frequency structures? did
    471   that take off fringing?}  The per-pixel standard deviations of these
    472 plots is listed in Table~\ref{table:sigmas.by.filter}.  The
    473 signal-to-noise of these structures is again somewhat weak, but the
    474 pattern is clearly visible in these figures.
     402from -20 to +20 milliarcseconds for all 5 plots.  A tree-ring-like
     403pattern is visible for all five filters, with systematic structures
     404following a circular pattern centered on the chip corner.; the finging
     405pattern is not apparent in the \yps\ astrometry.  The per-pixel
     406standard deviations of these plots is listed in
     407Table~\ref{table:sigmas.by.filter}.  The signal-to-noise of these
     408structures is again somewhat weak, but the pattern is clearly visible
     409in these figures.
    475410
    476411\subsubsection{Flat-field Structures}
     
    493428\end{figure*}
    494429
    495 The tree-ring structure is also clearly seen in the flat-field
    496 pattern.  For this measurement, we have used a set of monochromatic
    497 flat-field images obtained with a tunable laser.  The laser is used to
    498 illuminate our flat-field screen which is then observed by the PS1
    499 telescope.  These flat-field images were obtained 2011 Feb 09 as part
    500 of a campaign to study the PS1 system response (Tonry et al REF).
    501 Flats were obtain in a set of 4nm steps, with XXnm band-pass.  To
    502 enhance the signal-to-noise, we have combined a set of 6 flats at the
    503 center of the corresponding filter.  \note{high-pass filtering}. 
     430Figure~\ref{fig:flats.by.filter} shows the high-spatial-frequency
     431structures in the flat-field images.  For this measurement, we have
     432used a set of monochromatic flat-field images obtained with a tunable
     433laser.  The laser is used to illuminate our flat-field screen which is
     434then observed by the PS1 telescope.  These flat-field images were
     435obtained 2011 Feb 09 as part of a campaign to study the PS1 system
     436response \citep{tonry.phot}.  Flats were obtain in a set of 4nm steps,
     437with \note{XXnm} band-pass.  To enhance the signal-to-noise, we have
     438median-combined a set of 6 flats at the center of the corresponding filter.
     439
     440In order to mask pixels which do not flatten well, we generate a
     441a copy of the image smoothed with a Gaussian kernel with
     442$\sigma = 1.5 pixels$.  Any pixels in the smoothed image which deviate
     443from the median value in the image by more than 4 standard deviations
     444is masked.  We generate the superpixel image by averaging the unmasked
     445pixels associated with each superpixel.  We then high-pass filter the
     446superpixel image by subtracting a copy smoothed with a Gaussian of
     447$\sigma = 3.0$. 
    504448
    505449Figure~\ref{fig:flats.by.filter} shows the remaining high-frequency
     
    513457measured flux in those pixels, and thus a {\em negative} deviation in
    514458$\delta m_{psf}$ as defined above.  The dynamic range of the color
    515 scale in these plots is -0.01 to +0.01.  The tree-ring pattern is
     459scale in these plots is -0.01 to +0.01.  The tree-ring-like pattern is
    516460strong in the (\gps,\rps,\ips) images, but nearly swamped by fringing
    517461in \zps, and completely lost to finging in \yps.  A diagonal banding
     
    556500\end{figure*}
    557501
    558 The tree-ring structure is also seen in the changes of the image size.
    559 To measure this effect, we extract the second moments for all
    560 detections, subject to the same selections as for the photometry and
    561 astrometry residuals (good stars, multiple detections).  The second
    562 moments are measured with a Gaussian weighting function, with the
    563 $\sigma_{w}$ scaled by the PSF size so that the $\sigma$ measured for
    564 PSF stars is \approx 60\% of $\sigma_{w}$.  (Note that, since the
    565 measured $\sigma$ of stellar objects is biased down by the weighting
    566 function, this is not quite the same as having $\sigma_{w} = 1.6$
    567 times the true PSF $\sigma$).  For each detection, we measure
    568 $M_{xx,xy,yy} = \sum F_i w_i (x^2, x y, y^2) / \sum F_i w_i$.  For
    569 each exposure, we find the mean second moments for PSF objects and
     502During the image analysis, the second moments are measured for all
     503stars.  The values can be used to assess changes in the shape of stars
     504on the image.  To measure changes in the shapes, we have extracted the
     505second moments for all stellar detections, subject to the same
     506selections as for the photometry and astrometry residuals (good stars,
     507multiple detections).  The second moments are measured with a Gaussian
     508weighting function, with the $\sigma_{w}$ scaled by the PSF size so
     509that the $\sigma$ measured for PSF stars is \approx 60\% of
     510$\sigma_{w}$.  (Note that, since the measured $\sigma$ of stellar
     511objects is biased down by the weighting function, this is not quite
     512the same as having $\sigma_{w} = 1.6$ times the true PSF $\sigma$, see
     513discussion in \citealt{magnier.etal.2017.analysis}).  For each stellar
     514detection, we extract the values $M_{xx,xy,yy} = \sum F_i w_i (x^2, x
     515y, y^2) / \sum F_i w_i$.  For each exposure, we find the mean second
     516moments ($\bar{M_{xx,xy,yy}}$) for PSF objects on this chip (XY40) and
    570517subtract that mean value from the instantaneous measurements of
    571518$M_{xx,xy,yy}$.  We then determine the median of the residual second
    572 moments for each superpixel.
     519moments for each superpixel, resulting in 3 images for each filter.
     520
     521\note{write out this math, check out psLibADD}
     522
     523Using the second moment images, we can construct certain interesting
     524combinations, inspired by discussions of lensing measurements \citep{kaiser.1995}:
     525\begin{eqnarray}
     526R^2 & = & \delta M_{xx} + \delta M_{yy} \\
     527e_1 & = & \delta M_{xx} + \delta M_{yy} \\
     528e_2 & = & 2 \delta M_{xy}
     529\end{eqnarray}
    573530
    574531Figure~\ref{fig:smear.by.filter} shows the spatial trend of the {\em
     
    576533\delta M_{yy}$.  This value corresponds to the increase or decrease in
    577534the circularly-symmetric component of the image size.  The dynamic
    578 range of these images is -0.3 to +0.3 pixel$^2$. The tree-ring pattern
    579 is visible for all 5 filters, though \yps is dominated by the fringing
    580 pattern.  Structures with relatively low spatial frequencies can also
    581 be seen.
     535range of these images is -0.3 to +0.3 pixel$^2$. A tree-ring-like
     536pattern is visible for all 5 filters, though \yps is dominated by the
     537fringing pattern.  Structures with relatively low spatial frequencies
     538can also be seen.
    582539
    583540We can also construct a measurement of the change in ellipticity
     
    590547ellipse orientation as a function of postion.  The length of the
    591548vectors corresponds to the value of $\sigma^2_{major} -
    592 \sigma^2_{minor}$.  The tree-ring structure is not apparent in this
     549\sigma^2_{minor}$.  The tree-ring-like structure is {\em not} apparent in this
    593550figure for any filter.  The spatial variations are low-frequency and
    594551unrelated to the radial trend from the upper-left corner.
    595552
    596 \subsubsection{Correlations Between Systematic Trends}
     553\subsubsection{Correlations Between Tree-Ring-Like Patterns}
    597554
    598555\begin{table}
     
    614571\end{table}
    615572
    616 As discussed above, the signal-to-noise per pixel in the plots of the
    617 systematic trends is relatively low (\approx 1.0).  While the tree
    618 rings are apparent in many of these figures, there are also
    619 some other systematic structures which may degrade the signal
    620 further. To quantatatively compare the tree-ring trends between
    621 filters and between systematic effects, we need to measure the
     573Tree-ring-like patterns are clearly seen in 4 of the measurement types
     574above: the PSF photometry, the astrometry, the flat-field, and the
     575smear terms.  As discussed above, the signal-to-noise per pixel in the
     576plots of the systematic trends is relatively low (\approx 1.0).  While
     577the tree-ring-like patterns are apparent in many of these figures,
     578there are also some other systematic structures which may degrade the
     579signal further.
     580
     581To quantatatively compare the tree-ring-like trends between
     582filters and between the types of measurements, we need to measure the
    622583tree-ring structure explicitly and filter out the other effects if
    623584possible.  To do this, we have applied a high-pass filter to all of
     
    630591the arc to minimize the error associated with the choice of the
    631592pattern center and to avoid several bad cells near the bottom of the
    632 chip.  \note{draw the arcs?}
    633 
    634 For a given trend, the systematic effect is strongly correlated
    635 between filters.  The strongest correlation is the smear term:
    636 Figure~\ref{fig:smear.trends} shows the correlation of the smear
     593chip.
     594
     595For a given type of measurement, the systematic effect is strongly
     596correlated between filters.  The strongest correlation is the smear
     597term: Figure~\ref{fig:smear.trends} shows the correlation of the smear
    637598pattern between \gps\ and the other four filters. Even \yps\ is
    638599strongly correlated with \gps\ despite the presence of the fringe
     
    645606flat-field residuals are generally correlated between filters, but
    646607both \zps\ and \yps\ are affected by fringing.  For \yps, the
    647 correlation is completely washed out by the very strong fringing pattern.
    648 
    649 For all four measurements, the slope of the fitted lines are listed in
    650 Table~\ref{table:correlation.by.filter}.  There is a consistency in
    651 the trend from \gps, with the strongest systematic tree-ring effects
    652 to \yps, with the weakest effects.  Note that the second moment smear
    653 and astrometry terms have different relative strength in
    654 \yps\ compared with \gps. 
     608correlation is completely washed out by the very strong fringing
     609pattern.
     610
     611For all four types of measurements, the slope of the fitted lines are
     612listed in Table~\ref{table:correlation.by.filter}.  There is a
     613consistency in the trend from \gps, with the strongest systematic
     614tree-ring effects to \yps, with the weakest effects.  Note that the
     615second moment smear and astrometry terms have different relative
     616strength in \yps\ compared with \gps.
    655617
    656618% smear trends by filter
     
    694656\end{figure*}
    695657
    696 An important question is the relationship between the different types
    697 of systematic effects.  Different models for the tree-ring structures
    698 will make different predictions about the correlations between
    699 different effects.  We find the following relationships hold.  First,
    700 the PSF magnitude residuals and the second-moment smear trends are
    701 strongly anti-correlated: regions which have larger PSFs than the mean
    702 tend to have smaller measured PSF fluxes than the mean (note that
     658An important question is the relationship of the tree-ring-like
     659pattern between the different types of measurements.  Different models
     660for the tree-ring structures make different predictions about the
     661correlations between different effects.  Note the very different
     662spatial structure between the different measurements in a given
     663filter: the radial variations do not all follow the same patterns.
     664Instead, we find the following relationships hold:
     665
     666First, the PSF magnitude residuals and the second-moment smear trends
     667are strongly anti-correlated: regions which have larger PSFs than the
     668mean tend to have smaller measured PSF fluxes than the mean (note that
    703669$\delta m_{psf}$ is defined so that positive values correspond to
    704670larger fluxes).  These trends are shown in
    705 Figure~\ref{fig:smear.vs.psfmag}.  Second, the radial derivative of
    706 the smear is anti-correlated with the radial component of the
    707 astrometric residuals: $\frac{\partial (\sigma^2_{major} +
    708   \sigma^2_{minor})}{\partial radius} \sim \delta R$ (see
    709 Figure~\ref{fig:dsmear.vs.astrom}.  Finally, the radial derivative of
    710 the radial component of the astrometric residual is anti-correlated
    711 with the flat-field residual errors: $\frac{\partial \delta
    712   R}{\partial radius} \sim \delta flat$ (see
    713 Figure~\ref{fig:dastrom.vs.flat}.  This last relationship is somewhat
    714 weakly measured.  Because of the periodic nature of the tree rings, it
    715 is also difficult to completely certain that the flat-field is
    716 proportional to the derivative of the astrometry residual, and not the
    717 other way around.  The correlation is somewhat weaker for derivative
    718 of the flat-field vs astrometry residual.  The correlation is very
    719 weak between the flat-field and the astrometry directly.  We are
    720 convinced that we have the sense of the derivative correct by the
    721 details of specific features.
     671Figure~\ref{fig:smear.vs.psfmag}. 
     672
     673Second, the radial derivative of the smear is anti-correlated with the
     674radial component of the astrometric residuals: $\frac{\partial
     675  (\sigma^2_{major} + \sigma^2_{minor})}{\partial radius} \sim \delta
     676R$ (see Figure~\ref{fig:dsmear.vs.astrom}. 
     677
     678Finally, the radial derivative of the radial component of the
     679astrometric residual is anti-correlated with the flat-field residual
     680errors: $\frac{\partial \delta R}{\partial radius} \sim \delta flat$
     681(see Figure~\ref{fig:dastrom.vs.flat}.  This last relationship is
     682somewhat weakly measured.  Because of the periodic nature of the tree
     683rings, it is also difficult to be completely certain that the
     684flat-field is proportional to the derivative of the astrometry
     685residual, rather than the astrometry residual being proportional to
     686the derivative of the flat-field.  The correlation is somewhat weaker
     687for derivative of the flat-field vs astrometry residual.  The
     688correlation is very weak between the flat-field and the astrometry
     689residual values without a derivative.  We are convinced that we have
     690the sense of the derivative correct by examination of specific
     691features in each imaage (e.g., \note{give example}).
    722692
    723693\begin{table}
     
    772742
    773743These trends help to illuminate the underlying causes of these
    774 different effects.  First, we can easily explain the relationship
    775 between the PSF photometry residuals and the observed smear.  In the
    776 photometry analysis, we model the PSF allowing for some spatial
    777 variation in the shape.  However, we limit the 2D variation to a 3x3
    778 grid for each chip: the PSF parameters may vary smoothly across the
    779 chip following the bilinear interpolation between the 3x3 grid points.
     744different effects. 
     745
     746\note{summarize what pure lateral electric fields would do}
     747
     748First, if we consider the smear pattern
     749(Figure~\ref{fig:smear.by.filter}), the measurement shows that the
     750intrinsic size of the stellar images is varying in a radial sense
     751between the different tree-ring regions.  Although images experience
     752an average image quality (due to seeing and focus) across the chip
     753which may vary substantially from exposure to exposure, stars landing
     754in the different tree-ring-like regions are consistently somewhat
     755larger or somewhat smaller than that average.
     756
     757Next, we can explain the relationship between the PSF photometry
     758residuals and the observed smear: In the photometry analysis, we model
     759the PSF, allowing for some spatial variation in the shape.  However,
     760we have a limited number of stars to measure any spatial variation.
     761Thus the 2D variation are sampled on a very coarse (e.g., 3x3) grid
     762for each chip: the PSF parameters may vary smoothly across the chip
     763following the bilinear interpolation between the 3x3 grid points.
    780764Thus, the spatial scale on which we model PSF variations is much
    781765larger than the spatial scale on which PSF variations are apparently
    782 occuring.  When the true PSF is larger than the model PSF, our model
    783 fits systematically underestimate the amount of flux in a given
    784 object.  Conversely, when the PSF is smaller, we overestimate the
    785 flux.  The slope of the trend depends on the mean typical seeing for
    786 the given filter.  For example, the \gps\ seeing is typically
    787 1.3\arcsec, corresponding to a Gaussian $\sigma$ of 2.15 pixels.  A
    788 smearing of $\sigma^2_{major} + \sigma^2_{minor} = 0.1$ pixels$^2$
    789 would increase the size by about 0.02 pixels, or 1\%, so roughly
    790 consistent with the observed photometric deviation of about 5 to 10
    791 millimags for this amount of smearing.  \note{model the 2D effect more
    792   explicitly}.
    793 
    794 Second, the relationship between the flat-field residual and the
    795 astrometric gradient is consistent with radial variations in the
    796 plate-scale.  The tree-rings observed by DES are completely attributed
    797 to effective plate scale changes.  Effective plate scale changes would
    798 result in flat-field deviations since the flat-field illumination is a
    799 source of constant surface brightness.  Pixels see a varying amount of
    800 flux depending on their effective area.  This changing plate scale
    801 also affects the astrometry since these variations occur on spatial
    802 scales much smaller than the astrometric model.  In such a model, the
     766occuring, as illustrated by the changes in the smear plot.  When the
     767true PSF is larger than the model PSF, our model fits systematically
     768underestimate the amount of flux in a given object.  Conversely, when
     769the PSF is smaller, we overestimate the flux -- this type of offset is
     770a typical effect when mis-estimating the PSF size.  The slope of the
     771trend depends on the mean typical seeing for the given filter.  For
     772example, the \gps\ seeing is typically 1.3\arcsec, corresponding to a
     773Gaussian $\sigma$ of 2.15 pixels.  A smearing of $\sigma^2_{major} +
     774\sigma^2_{minor} = 0.1$ pixels$^2$ would increase the size by about
     7750.02 pixels, or 1\%, roughly consistent with the observed photometric
     776deviation of about 5 to 10 millimags for this amount of smearing.
     777
     778The relationship between the flat-field residual and the astrometric
     779gradient is consistent with radial variations in the plate-scale.  The
     780tree-rings observed by DES are completely attributed to effective
     781plate scale changes.  Effective plate scale changes would result in
     782flat-field deviations since the flat-field illumination is a source of
     783constant surface brightness.  Pixels see a varying amount of flux
     784depending on their effective area.  This changing plate scale also
     785affects the astrometry since these variations occur on spatial scales
     786much smaller than the astrometric model.  In such a model, the
    803787flat-field deviations are $-1 \times \frac{\partial Pos}{\partial R}$.
    804788The slope of our relationship is \approx 0.5 in normalized units.
     
    8067902.  \note{looks like a slope of 1.0 would not be excluded by these
    807791  plots}
     792
     793\note{I need to use the relationship between the astrometry and the
     794  flat-field to calculate the amplitude of the lateral electric
     795  fields.}
    808796
    809797The fact that the PSF ellipticity changes are {\em not} correlated
     
    841829
    842830\end{document}
     831
     832Notes for paper re-work:
     833
     834* Paper focus is now only on the diffusion variations
     835  * strip out the discussion of other systematic effects
     836  * strip down the PS1 introduction discussion
     837
     838* tentative title:
     839  Evidence for Small-Scale Charge-Diffusion Variations in Pan-STARRS CCDs
     840
     841* outline
     842
     843 1. introduction
     844    * thick CCDs
     845    * tree rings == transverse field effects (see Plazas et al)
     846    * we see something else
     847
     848 4 model : diffusion variations due to E|| field variations
     849
     850 5 discussion (how to treat in calibration / analysis)
     851
     852 6 conclusions
     853
     854some possible refs to tree rings / charge diff:
     855
     856* http://adsabs.harvard.edu/abs/2016SPIE.9904E..2CW (Woods et al 2016; TESS)
     857* https://arxiv.org/pdf/1605.01001.pdf : plazas et al
     858* http://ieeexplore.ieee.org/document/1225293/?part=1 Altmannshofer et al 2003 (about thick Si)
     859
     860* plazas et al 2014 outline
     861
     862  1. intro: thick CCDs, transverse electric fields
     863  2. DES / DECam
     864
     865  2.1 flat-field tree rings (discussion of flat-field tree rings
     866  starting from the premise that they know the answer).
     867 
     868  3 impact on astrometry and photometry
     869
     870  4 improving calibrations given tree rings
     871
     872  5 summary and conclusions
     873
Note: See TracChangeset for help on using the changeset viewer.