API/ENH: allow to do pure maximum statistic permutation tests without clustering

Maximum statistic tests, like implemented in mne.stats.permutation_t_test are popular choices in M/EEG http://onlinelibrary.wiley.com/doi/10.1111/j.1469-8986.2011.01273.x/full I feel we could easily make our clustering API perform such tests by skipping the clustering if connectivity=False. This would allow us to use the same API to perform to perform general max stat tests for 1 and multiple samples with arbitrary stat functions, beyond the 1 sample t-test in mne.stats.permutation_t_test.

cc @agramfort @larsoner @SherazKhan etc.

[larsoner]

+1 for adding adding connectivity=False to mean “no connectivity” (None means “regular lattice connectivity” now) and deprecating mne.stats.permutation_t_test

[dengemann]

if it is, speed is very rarely a limiting factor.

For benchmarking I implemented a generalized version that takes any stat_fun. When running it with ttest_1samp_no_p the loss in performance is not game changing. Another issue that this implementation has is memory as all permutations are precomputed, which on larger data consumes most of the time. It can also be done with a generator. But I agree that we should recycle permutation_t_test.

I agree with the common elements in Eric’s and Alex’ suggestions. Some temporary compromise would be ok. The next stable API should emerge with this GSOC.