For example, PoW does not have absolute finality, only probabilistic finality; BFT has absolute finality, so does Casper FFG. YeeCo’s Tetris consensus also has absolute finality.
Use of 'absolute' is horribly misleading. There is no 'absolute' in any system which is not fully objective. Casper and all BFT-type consensus designs are subjective by nature.
You are correct that PoW only has probabilistic finality, but nothing else has 'absolute' finality, at best they have finality within the set of tolerances set out by the design, which usually include such things as an honest majority of bootstrapping nodes etc.
Absolute finality is just introduced to define
real finality that is not only a probabilistic one.
When we say a kind of consensue has absolute finality, it means that the
canonical chain will not change and then the transaction confirmed on it will not change, or there is a
hard fork.
While, on a PoW chain, the
canonical chain may change even after n confirmed, that's why we say PoW only has probabilistic finality.
The article also described :
From the above analysis we can learn that absolute finality can lead to two extremes. One is to reach a consistency that cannot be modified; the other is that if a consistency can’t be reached, it will be split into two different consistencies that can’t be modified, and the applications have to choose one between two and recognize the finality of the consistency it picked.