ENH: Beamforming enhancements

• Fix high-SNR breakages (e.g., 1000 nAm dipoles)
• NAI (neural activity index) normalization from the mailing list.

in fact, I did not look into the details of the lcvm code. For what concerns dics, I think the output is the ratio of powers (data/noise), that is what is normally called "neural activity index" NAI. In don't think the 1997 Van Veen paper has defined the NAI (https://www.lsv.uni-saarland.de/fileadmin/teaching/dsp/ss15/BF_VeenBuckley.pdf). For a ref about NAI, you can look at this chapter:

http://www.ccs.fau.edu/~fuchs/pub/Beamer.pdf

Or the webpage of Fieldtrip:

http://www.fieldtriptoolbox.org/tutorial/beamformer#neural_activity_index

More generally, it would be good to have the possibility in MNE to have different normalization techniques (in addition to "None", in some case people may prefer not to normalise wrt noise). Keeping in mind that power values are not normally distributed (they are chi-squared), some ways to normalise are:

1) dB = 10 * log10(data/noise)

2) z-score = [data - mean(noise)] / std(noise)

3) z-score of logs = [log(data) - mean(log(noise))] / std(log(noise))  (the log-transform make power values approx. gaussian, as sqrt-transform)
4) ERD/S = [data - mean(noise)] / mean(noise)

... just to cite a few
Some inspiration here (lines 198-207 ):

https://github.com/fieldtrip/fieldtrip/blob/master/ft_freqbaseline.m
    ```

[larsoner]

I can bring some high-SNR phantom data to the sprint so we can interactively break things there.

[larsoner]

If it’s inherent, there is no known fix, and we should just learn to live with it, we should close