In classical mechanics one thinks of points in phase space as states, all the while knowing that for the precise description of any real physical system one will inevitably have to deal with uncertainties and thus will have to talk about probability distributions on phase space. One assumes that one can measure observables to infinite precision although in practice this is never possible.
In quantum mechanics "states" of definite momentum (or position) aren't states - they are not represented in the Hilbert space. They are used a lot but always bearing in mind that one is actually considering very narrow peaks rather than delta distributions, for example. Also strictly speaking momentum measurements are impossible, only measurements of they type "Was the particle's momentum uppon measurement contained in the Borel set $A$?" The mathematical machinery (Gelfand triples) is in place but one does not physically allow the generalized states (many non-normalizable functions as well as distributions) to be states. Why would it be wrong do define objects as states of definite momentum to be actual quantum states?
I can think of the following arguments:
One can think of the infinite space model as the limit (i.e. an idealization) of a very large "box with boundary conditions" where momenta are discrete. Here distributions used to correspond to actual states before the limit. A common argument is that it should be impossible to distinguish a system in a very large box from a system in infinite space. Why does this not imply that one should be allowed to think of definite momentum states as actual states?
In such a generalized state most observables won't have a probability distribution for the values one could obtain measuring them (a definite momentum state will not have a position distribution). I don't see how this is a problem, since there are vectors in the Hilbert space representing actual states for which the same is true (them not being in the domain of definition of certain observables). Looking at examples there is no obvious reason to dismiss these states as pathological. Now domains of observables being dense, one can always find a state that is arbitrarily norm-close to some given state and in the domain of a collection of arbitrary observables. However, norm-closeness has nothing to do with "closeness of physical states" (a notion I'm not completely sure of, which should probably be defined in terms of observables. I assume the projective Hilber space topology actually fails do provide such a notion.) so this argument is void as well, furthermore it would also apply to delta distributions in a generalized sense (those can be approximated by test functions arbitrarily well).
Making some general assumptions about statistics and defining properties of a propositional lattice that should discribe QM one can show that this is equivalent to the lattice of closed subspaces of some Hilbert space. This clearly rules out the distributions as states. However, this step is not completely transparent to me and the conclusion seems conceptually very far from where the argument starts. I feel there should be some more intuitive way to provide a connection between these aspects of quantum logic and there not being such states.
Does anyone know a good argument for this or can explain the mistakes in the ideas above? Any sources on this issue?