DarMM said:
And what would you think of the fact that fields don't have well defined values at a point, i.e. ##\phi(x)## is undefined? That the fundamental things are the smeared fields, ##\int_{\mathcal{M}}{\phi(x)f(x)d^{4}x}##?
Those ##\phi##'s really need an indication that they're operators, indexed by a set of
test functions, ##\hat\phi_f##, which, taken independently of any other observables we can take to represent a random variable, with the vacuum state providing a probability density. We can't measure ##\hat\phi_f## at a point, for ##f## a delta function, insofar as the variance is infinite (or, better, undefined, as you say) even for the free field, but, thinking very loosely, for any finite region an infinite sum of infinite variance random fields can perfectly well be finite and finite variance.
I find it helpful to use signal analysis language, with the test functions performing a function very close to that performed by "
window functions" (signal analysis), Chris Fewster calls them "
sampling functions" (and perhaps others do, but I haven't seen it from others). In signal analysis terms, we can think of ##\hat\phi_g|0\rangle## as a multiplicative modulation of the vacuum state, so that when a test function is used in this way it's appropriate to call a test function a "
modulation function" (so signal analysis again).
In signal analysis terms, there's no
a priori reason to think that the (pre-)inner product ##\langle 0|\hat\phi_f^\dagger\hat\phi_g|0\rangle## has to be a linear functional of ##f## and ##g## at all length scales and at all amplitudes, provided it's complex-linear in ##\langle 0|\hat\phi_f^\dagger## and ##\hat\phi_g|0\rangle##, indeed our usual experience of nonlinearity in signal analysis suggests that
we should expect it not to be, and yet the Wightman axioms insists it must be (for no physically justified principle), and the Haag-Kastler axioms, to approximately the same effect, insist on Additivity. Loosening this axiom results in a plethora of (what I find) interesting nonlinear models, as a result of which we can naturally construct multi-point operators (by polarization) that can be used to represent bound states (the account I've given here is obviously much too fast: a development that is as good as I could manage a few years ago can be found in
arXiv:1507.08299, still very early days yet). FWIW, I see a connection between this account and the discussion in this thread of bound systems, with apparently no resolution, whereas for me this kind of approach offers at the least some possibilities — which, moreover, are moderately principled and empirically grounded in signal analysis concerns.