Default behaviour for `write_map` is to use lossy compression with floating-point numbers

[ziotom78]

The default value for the compress parameter to write_map is True, meaning that the code will use one of the compression algorithms specified by the FITS standard. Since these algorithms are lossy when applied to maps containing floating-point values, the code will silently introduce an error in the data written to disk, as the following example shows:

import numpy
import qubic.io

FILE_NAME = '/tmp/compressed_map.fits'

INPUT_PIXELS = numpy.random.normal(size=12 * 64 * 64)
qubic.io.write_map(FILE_NAME, INPUT_PIXELS, dtype=numpy.float64)
OUTPUT_PIXELS = qubic.io.read_map(FILE_NAME)

DIFF = OUTPUT_PIXELS - INPUT_PIXELS
print('Stddev of the difference: {0:.3f}'
      .format(numpy.std(DIFF)))

# Result:
# Stddev of the difference: 0.018

(Note that the code uses float64 instead of the default float32 type when calling write_map.) Setting compress=False in the above call to qubic.io.write_map makes the difference between INPUT_PIXELS and OUTPUT_PIXELS to vanish.

Since this behaviour might be surprising for people not versed with the FITS standard, I am suggesting a couple of workarounds:

1. Make compress default to False, and specify in the docstring for write_map that using True with floating-point maps will introduce an error in the pixel values (currently, the docstring does not mention the compress parameter anywhere but in the declaration);
2. Make the code emit a warning when compress is True and the datatype of the parameter map is a floating-point value.

[ziotom78]

Just for the record, using healpy.write_map and healpy.read_map with .fits.gz works seamlessy and produces relevant compression ratios for partial-coverage maps.