> I don't understand why having less information is better (theoreticaly).

A ranked-choice ballot only encodes the orders the candidates against one another, whereas approval and score votes also encode the candidates' positions within the voter's range of subjective preferences. That is, if we have 3 candidates (A, B, C) and a few voters which each voters has a range from love to hate for each candidate, like so:

  Love                       Hate
   |-A--B-------------------C-|
   |-A-------------B-----C----|
   |-------------------A-B-C--|
   |-A-B---C------------------|

Under ranked choice voting, every one of these voters' ballots would look the same:

   1)A, 2)B, 3)C

Ranked choice voting encodes the ordering of the preferences, but the intensity of those preferences is lost when the ballot is cast. Whereas under approval and score voting, every one of these voters represents their preferences differently, because they're reflecting their personal response to each candidate:

     Approval | Disapproval
   |-A--B-----|-------------C-|
   |-A--------|----B-----C----|
   |----------|--------A-B-C--|
   |-A-B---C--|---------------|

Of course, some information is lost in the fact that we only have 2 values approval/disapproval to encode positional preferences. But I would argue this information is already more meaningful than a fully-expressed ranked ballot. And if necessary, score voting can capture more of that information by offering > 2 levels to divide the candidates into.

OK, this method recollect some information that the ranked choice voting ignores. But I still don't understand why the Arrow's theorem doesn't apply.

If in a hypothetic population everyone loves/hates each candidate equally spaced, then in that population it is possible to apply the Arrow's theorem and prove that for that population this method doesn't work. But the method should be useful for every population, even the pathological ones.