Coaddition of Spectroperfect Reductions¶

This document covers coadd implementation details and mock-data tests. For details on the coadd dataflow and algorithms used for combining spectra, refer to DESI-doc-1056.

Implementation¶

Files¶

All spectra are grouped by brick. There are three types of brick file under $DESI_SPECTRO_REDUX/$PRODNAME//bricks/{brickid}/:

The brick files contains all exposures of every target that has been observed in the brick, by band.
The band coadd files contain the coadd of all exposures for each target, by band.
The global coadd files contain the coadd of band coadds for each target.

All files have the same structure with 4 HDUs:

HDU0: Flux vectors for each spectrum.
HDU1: Ivar vectors for each spectrum.
HDU2: The common wavelength grid used for all spectra.
HDU4: Binary table of metadata.

See the relevant data model descriptions for details (these are not in synch with the mock data challenge files as of 23-Mar-2015).

Wavelength Grids¶

All brick files have their wavelength grid in HDU2, as summarized in the table below. Note that we do not use a log-lambda grid for the global coadd across bands, since this would not be a good match to the wavelength resolution at the red ends of r and z cameras. See the note for details.

Band	Min(A)	Max(A)	Nbins	Size(A)	Files
b	3579.0	5938.8	3934	0.6	brick, band coadd
r	5635.0	7730.8	3494	0.6	brick, band coadd
z	7445.0	9824.0	3966	0.6	brick, band coadd
all	3579.0	9825.0	6247	1.0	global coadd

Metadata¶

The brick file and two co-add files all have a table with the same format in HDU4, with one entry per exposure of each object observed in the brick. Most of its columns are copied directly from the exposure fibermaps, except for the last three columns which identify the exposure and offset of the object’s spectrum in the other HDUs. The table columns are listed below.

Name	Description
FIBER	Fiber ID [0-4999]
POSITIONER	Positioner ID [0-4999]
SPECTROID	Spectrograph ID [0-9]
TARGETID	Unique target ID
TARGETCAT	Name/version of the target catalog
OBJTYPE	Target type [ELG, LRG, QSO, STD, STAR, SKY]
LAMBDAREF	Reference wavelength at which to align fiber
TARGET_MASK0	Targeting bit mask
RA_TARGET	Target right ascension [degrees]
DEC_TARGET	Target declination [degrees]
X_TARGET	X on focal plane derived from (RA,DEC)_TARGET
Y_TARGET	Y on focal plane derived from (RA,DEC)_TARGET
X_FVCOBS	X location observed by Fiber View Cam [mm]
Y_FVCOBS	Y location observed by Fiber View Cam [mm]
X_FVCERR	X location uncertainty from Fiber View Cam [mm]
Y_FVCERR	Y location uncertainty from Fiber View Cam [mm]
RA_OBS	RA of obs from (X,Y)_FVCOBS and optics [deg]
DEC_OBS	dec of obs from (X,Y)_FVCOBS and optics [deg]
MAG	magitude
FILTER	SDSS_R, DECAM_Z, WISE1, etc.
NIGHT	Date string for the night of observation YYYYMMDD
EXPID	Integer exposure number
INDEX	Index of this object in other HDUs

Programs¶

The following programs are used to implement the coadd part of the pipeline:

desi_make_bricks: Create brick files from all exposures taken in one night. Reads exposures from cframe files and adds metadata from the exposure fibermap.
desi_update_coadds: Update the coadds for a single brick. Reads exposures from brick files and writes the corresonding band coadd and global coadd files.

An additional program desi_inspect displays the information and creates a plot summarizing the coadd results for a single target.

Benchmarks¶

The rate-limiting step for performing coadds is the final conversion from Cinv and Cinv_f to flux, ivar and resolution in desispec.coaddition.Spectrum.finalize(). The computation time is dominated by one operation: solving the eigenvalue program for a large symmetric real-valued matrix using scipy.linalg.eigh(). The computation also involves inverting a real-valued resolution matrix using scipy.linalg.inv(), but this is relatively fast.

For a typical r-band coadd on a 2014 MacBook Pro, the total time for a single target is about 12 seconds, dominated by eigh (10.6s) and inv (1.0s). The time for a global coadd will be longer because of the larger matrices involved.

Interactive tests run on edison@nesrc indicate that it takes about 20s for each single-band coadd and 90s for the global coadd, for a total of about 150s per target. Note that the coadd step can be parallelized across bricks to reduce the wall-clock time required to process an exposure. The biggest speed improvement would likely come from using a sparse matrix eigensolver, or adjusting the algorithm to be able to use an incomplete set of eigenmodes (the scipy.sparse.linalg.eigsh() function can not calculate the full spectrum of eigenmodes).

Notes¶

The brick filenames have the format brick-{band}-{expid}.fits, where band is one of [rbz], which differs from the current data model (which is missing the {band}).
Bricks contain a single wavelength grid in HDU2, the same as current CFRAMES, but different from the CFRAME data model (where HDU2 is a per-object mask).
The NIGHT column in HDU4 has type i4, not string. Is this a problem?
The 5*S10 FILTER values in the FIBERMAP are combined into a single comma-separated list stored as a single S50 FILTER value in HDU4 of the brick file. This is a workaround until we sort out issues with astropy.io.fits and cfitsio handling of 5*S10 arrays.
The mock resolution matrices do not have np.sum(R,axis=1) == 1 for all rows and go slightly negative in the tails.
The wlen values in HDU2 have some roundoff errors, e.g., z-band wlen[-1] = 9824.0000000014425
Masking via ivar=0 is implemented but not well tested yet.
We need a way to programmatically determine the brick name given a target ID, in order to locate the relevant files. Otherwise, target ID is not a useful way to define a sample (a la plate-mjd-fiber or ThingID) and an alternative is needed for downstream science users.
The global coadd sometimes find negative eigenvalues for Cinv or a singular R.T. These cases need to be investigated.

Mock Data Tests¶

Warning

The description of environment setup and installation below may be out of date.

DESI Environment¶

Ssh to edison.nersc.gov (remember to use ssh -A to propagate your keys for github access) and:

source /project/projectdirs/desi/software/modules/desi_environment.sh

Installation¶

Clone the git package and select the co-add development branch (which should soon be merged into the master branch, making the last command unecessary):

git clone git@github.com:desihub/desispec.git
cd desispec
git checkout \#6

Per-Login Setup¶

Manually set paths for using this installation (assuming bash):

cd desispec
export PATH=$PWD/bin:$PATH
export PYTHONPATH=$PWD/py:$PYTHONPATH

Set pipeline paths:

export DESI_SPECTRO_REDUX=$DESI_ROOT/spectro/redux
export PRODNAME=sjb/cedar2a
export DESI_SPECTRO_SIM=$DESI_ROOT/spectro/sim
export DESI_SPECTRO_DATA=$DESI_SPECTRO_SIM/alpha-5

Run Tests¶

Convert mocks cframes and fibermaps into brick files using:

rm -rf $DESI_SPECTRO_REDUX/$PRODNAME/bricks
desi_make_bricks.py --night 20150211 --verbose

Note that the code is not yet smart enough to do the right thing for exposures that have already been added to brick files, hence the rm command above.

Update coadds for a single brick:

rm -rf $DESI_SPECTRO_REDUX/$PRODNAME/bricks/3587m010/coadd*
desi_update_coadds.py --brick 3587m010 --verbose

Look at a single target in this brick (this is an LRG):

desi_inspect.py --brick 3587m010 --target 3640213155238558158 --verbose

Inspect a brick file in iPython using, e.g.:

import os,os.path
import astropy.io.fits as fits
from astropy.table import Table
brick = fits.open(os.path.join(os.getenv('DESI_SPECTRO_REDUX'),os.getenv('PRODNAME'),'bricks','3587m010','brick-r-3587m010.fits'))
info = Table.read(brick,hdu=4)
print info
plt.errorbar(x=brick[2].data,y=brick[0].data[0],yerr=brick[1].data[0]**-0.5)

Run unit tests:

python -m desispec.resolution