A concern has been expressed that "the Jaynes principle can produce fake en
tanglement" [R. Horodecki et al., Phys. Rev. A 59, 1799 (1999)]. In this pa
per we discuss the general problem of distilling maximally entangled states
from N copies of a bipartite quantum system about which only partial infor
mation is known, for instance, in the form of a given expectation value. We
point out that there is indeed a problem with applying the Jaynes principl
e of maximum entropy to more than one copy of a system, but the nature of t
his problem is classical and was discussed extensively by Jaynes. Under the
additional assumption that the state rho ((N)) of the N copies of the quan
tum system is exchangeable, one can write down a simple general expression
for rho ((N)). By measuring one or more of the subsystems, one can gain inf
ormation and update the state estimate for the remaining subsystems with th
e quantum version of the Bayes rule. Using this rule. we show how to modify
two standard entanglement purification protocols, one-way hashing and recu
rrence, so that they can be applied to exchangeable states. We thus give an
explicit algorithm for distilling entanglement from an unknown or partiall
y known quantum state.