The assumption that all energy eigenstates with the same energy are equally likely is simply stated as an axiom in typical books, [4, p. 92], [18, p. 1], [25, p. 230], [51, p. 177]. Some of these sources quite explicitly suggest that the fact should be self-evident to the reader.
However, why could not an energy eigenstate, call it A, in which all particles have about the same energy, have a wildly different probability from some eigenstate B in which one particle has almost all the energy and the rest has very little? The two wave functions are wildly different. (Note that if the probabilities are only somewhat different, it would not affect various conclusions much because of the vast numerical superiority of the most probable energy distribution.)
The fact that it does not take any energy to go from one state to the other [18, p. 1] does not imply that the system must spend equal time in each state, or that each state must be equally likely. It is not difficult at all to construct nonlinear systems of evolution equations that conserve energy and in which the system runs exponentially away towards specific states.
However, the coefficients of the energy eigenfunctions do not satisfy
some arbitrary nonlinear system of evolution equations. They evolve
according to the Schrödinger equation, and the interactions between
the energy eigenstates are determined by a Hamiltonian matrix of
coefficients. The Hamiltonian is a Hermitian matrix; it has to be to
conserve energy. That means that the coupling constant that allows
state A to increase or reduce the probability of state B is just as
big as the coupling constant that allows B to increase or reduce the
probability of state A. More specifically, the rate of increase of
the probability of state A due to state B and vice-versa is seen to be
This note would argue that under such circumstances, it is simply no longer reasonable to think that the difference in probabilities between eigenstates of the same energy is enough to make a difference. How could energy eigenstates that readily and randomly exchange probability, in either direction, end up in a situation where some eigenstates have absolutely nothing, to incredible precision?
Feynman [18, p. 8] gives an argument based on time-dependent perturbation theory, chapter 11.10. However, time-dependent perturbations theory relies heavily on approximation, and worse, the measurement wild card. Until scientists, while maybe not agreeing exactly on what measurement is, start laying down rigorous, unambiguous, mathematical ground rules on what measurements can do and cannot do, measurement is like astrology: anything goes.