TODO

- [setup] Get rid of setup.py dependency on nose and cython
- [speed] Default to scs
- [cleanup] Remove support for internal cvxpy-old
  Not needed given the current state of cvxpy and PICOS and also extra slow
- [cleanup] Remove support for python 2.7 (eol by Jan 2020)
- [delsig] Look into quadrature modulators
- [delsig] implement findPattern
- [delsig] implement calculateSNR
- [delsig] implement simulateSNR
- [simulator] reuse common code between _simulateDSM_scipy_blas and
    _simulateDSM_cblas
- [simulator] verify speed difference between scipy_blas and cblas simulator
    possibly including an fblas simulator in the picture and trying
    different matrix representations (C, fortran)
- implement quadratic form
- implement scanner with quadratic form
- [sdtoolbox] implement calcSNR
- [sdtoolbox] sinusx
- [documentation] add tutorial
- [cython] in the cython simulator use np.double and np.intc for casting
- maybe move zpk and ba representations to named tuple model or to LTI
    representations
- avoid __import__ statements and follow
    http://docs.scipy.org/doc/scipy/reference/api.html
- [cython] use numpy.require to assure that arrays have the right properties
- [optimization] move to the newer cvxpy
- [examples] provide notebook version of the examples
- [setuptools] make docdist pass options to build_sphinx
- [documentation] Maybe use the versionadded/versionchanged/deprecated sphinx
    features to document api changes
- [*examples*] The numeric results in the ICECS 2012 paper are not consistent
    with what the sample script returns. Specifically, the SNR values
    evaluated via the NTF on the paper are better than those obtained by
    the code by a factor 2. In the original version of the code, the results
    where identical due to a missing factor 2 in the quantization noise gain
    evaluation. Probably, this error slipped in because it makes the NTF-based
    results and the time-domain-simulation results more similar. In fact, the
    'real' modulator behavior is better than predicted by the NTF model
    because in the specific example the quantization noise is not white,
    but for some reason slightly blue (more power at higher frequencies).
    A note should probably be added to the code and to the ArXiv version of
    the ICECS paper.
- [accuracy] A recent code change removes a matrix inversion noticing that
    the matrix is orthonormal, and that transposition should suffice.  However
    this seems to cause come accuracy issues on some platforms (AMD?)
- [clans] Consider the following suggestion for CLANS, that was submitted as
    a patch to the clans Matlab code:
    For better convergence, transform the errors in minimization functions
    (dsclansObj6a,dsclansObj6b) into quadratic errors.
    Howto:
    1. Open clans6.m
    2. in sub-function dsclansObj6a, after "f = abs(evalTF(H,exp(1i*pi/OSR)));"
       type "f=f*f".
    3. in sub-function dsclansObj6a, after
       "g = sum(abs(impulse(H,100))) -1 - Q;" type "g=g*g".
    This way the target value for Q will be reached quite accurately at low OSR.