I've carefully read about rainbow tables and can't get one thing. In order to build a hash chain a reduction function is used. It's a function that somehow maps hashes onto passwords. This article says that reduction function isn't an inverse of hash, it's just some mapping.
I don't get it - what's the use of a mapping that isn't even an inverse of the hash function? How should such mapping practically work and aid in deducing a password?
A rainbow table is "just" a smart compression method for a big table of precomputed hashes. The idea is that the table can "invert" a hash output if and only if a corresponding input was considered during the table construction.
Each table line ("chain") is a sequence of hash function invocations. The trick is that each input is computed deterministically from the previous output in the chain, so that:
The reduction function is the glue which turns a hash function output into an appropriate input (for instance a character string which looks like a genuine password, consisting only of printable characters). Its role is mostly to be able to generate possible hash inputs with more or less uniform probability, given random data to work with (and the hash output will be acceptably random). The reduction function needs not have any specific structure, in particular with regards to how the hash function itself works; the reduction function must just allow keeping on building the chain without creating too many spurious collisions.