That no one seems to have produced (or pointed at) an asymptotically better solution in any language, I think that's some small evidence that there isn't such a solution and substantially more evidence that it's not easy to find.
Although I don't know how one would go about proving or disproving if this is asymptotically better or just a (substantial) constant-factor speedup.
I had initially skimmed the article, and assumed he was already handling uniqueness. With the correct definition of StateL, the solution presented is equivalent to what you describe here, but apparently that definition is deferred to a future article.
The idea would be that, for StateL, for_each would pick one value and pass that value to the lambda, interpreting the result in a context where the sel only sees the remaining values in the list.
Nope. It looks like it at first glance, but it isn't.
Look at what happens when n == b/C:
b^(b/C) / (b!/(b - b/C)!) ==
b^(b/C) / (b!/(b*(C-1)/C)!) ==
e^(log(b^(b/C)) + log((b/C)!) - log((b*(C-1)/C)!)) ~=
e^(b/C log(b) + [b*(C-1)/C] log (b*(C-1)/C) - [b*(C-1)/C] - b log b + b) ==
e^((b+b (-1+C) log((-1+C)/C))/C) ==
e^(b/C* (1+(C-1)log((C-1)/C)))
(I used Stirling's Approximation, namely the form ln(x!) ~= x ln x - x for large x)However, note that (1+(C-1)log((C-1)/C)) / C is greater than zero for all C > 1, as 1+(C-1)log((C-1)/C) > 0 for all C > 1. As such, this diverges for all C > 1.
In any case, I make the assertion that the algorithm proposed in the article is in fact O(b!/(b-n)!), with the intended definition of StateL.
Edit: edited.
However, this doesn't mean we have to punt the easy cases just because there are hard cases.