Peter Li

Hi jianbo, I think this is the same as Martijn's request? I may have time to give this a try in the next month or two. Unfortunately I expect it to be significantly slower than just returning the partially-sorted output. This is a limitation of the way the C++ standard library is written and I'm not sure I can get around it. Bruno's version may have more flexibility to be efficient here.

I've noticed that the minimum spacing of random numbers generated by Boost Random 1.47 is greater than the minimum produced by Matlab; it appears to me that Matlab is using twice as many underlying integer seeds from the Twister to produce each double compared to Boost's behavior. So if you need very fine-grained doubles, this may be an issue for you.

As a follow up on WurmD's comment: "all" Matlab functions work on the first dimension by default so I think a "replacement" method should do so too.
The current behavior of fast_median breaks things like:
med = median(data(:));

Currently I can't use fast_median in my code as I need to be able to change the filtering function and I don't want to code in some special treatment for fast_median... :/

Would that be possible?

Hi, I am interested to have index, rather of the n-th element output, thx.

It appears that the minimum spacing for Boost Random 1.47 uniform_01 distribution (doubles between 0 and 1) is ~2.3e-10, which indicates that the RNG is simply taking each int generated by the Twister and (effectively) dividing it by the maximum int.

Matlab's RNG appears to use two int from the Twister for each [0,1] double, using the first int for the more significant bits and the second for the less significant bits.

So essentially using this library will give you faster RNG over Matlab at the expense of lower resolution.