Kai Wohlfahrt

I don't see how an `approx` attribute will work for the multidimensional case. To specify the transform, it needs to know whether it is approximate or detail in every dimension,...

Yes, the current implementation requires convolving both coefficients again with the filter, and adding the two coefficients together and dividing by 2. The result will be circularly shifted by `2**(level-1)`,...

R uses a 'MODWT' transform, whereas MATLAB uses the "algorithme à trous" to perform shift-invariant transforms. The difference is in how the 'stretched' wavelet filters are calculated. MATLAB uses something...

Note that the MODWT filters should allow one to directly compute the `j`th level DWT (with appropriate downsampling). It doesn't actually reduce the number of operations (the filter is `2**j`...

Re-naming and re-categorizing.

There has been more recent discussion about implementing a MODWT (yay!), with a DFT approach being suggested in #200. My first instinct would be that this could be slower than...

OK, changing the issue title, looks like there's a lot to be done about documentation (sigh). I've made a list in the original report, feel free to mention any other...

Notes on `doc/source/regression`. I found the chapters on wavelet packets to be much more useful than the API documentation, so there is some value in them. Perhaps a "Getting started"...

Notes on `demo` - `benchmark.py`: benchmark for several sizes and wavelet of 1D transform - I don't think this needs to be preserved, fairly self explanatory - `dwt2_dwtn_image.py`: Show round-trip...

@rgommers: Do you know what the `demo/dwt_signal_decomposition.py` demo is showing? I understand the decomposition part, but then before plotting it runs the coefficients through `waverec`. If this is a common...