TITLE: Quantum fields and infinite tensor products
QUESTION [22 upvotes]: As I understand it, a naive interpretation of the state space of a quantum field theory is an infinite tensor product
$$\otimes_{x\in M} H_x,$$
where $x$ runs over the points of space. This corresponds to the fact that a field $\phi$ and the conjugate momentum $\pi$ can be viewed as a composite system of the array of $\phi(x)$ and $ \pi(x)$. Thus, again naively, the amplitude assigned by a quantum state $\Psi(\phi, \pi)$ to a classical initial condition $(\phi,  \pi)$  is a tensor product of the amplitudes $\Psi(\phi(x), \pi(x))\in H_x$.
Of course, this doesn't quite make sense, for many reasons including the fact that the infinite tensor product is quite badly behaved. Instead, the standard way of quantising, say a scalar field satisfying the Klein-Gordon equation, is to write it in terms of Fourier modes
$$\phi(x,t)=(2\pi)^{-3/2}\int [a(p)e^{i(px-\epsilon_pt)}+a^*(p)e^{-i(px-\epsilon_pt)}]\frac{d^3p}{2\epsilon_p}$$
with $\epsilon_p=\sqrt{p^2+m^2}$ (this being the KG equation). The canonical commutation relation for $\phi$ and $\pi$ become
$$[a(p), a^*(p')]=2\epsilon_p\delta(p-p'); \ [a(p), a(p')]=[a^*(p), a^*(p')]=0,$$
which can individually be quantised in a Segal-Bargmann fashion to act on a Hilbert space $H_p$. To quantise all these operators as we run over all momenta, we would again require an infinite tensor product
$$\otimes_p H_p.$$
This is avoided by imposing an additional condition, the existence of a vector $\Psi_0$ (interpreted as the vacuum), satisfying $$a(p)\Psi_0=0$$ for all $p$. After this, it all works out and we have a nicely quantised free field by putting the operators into the integral above. I think I sort of understand this procedure, with the level of uncertainty that I'm normally stuck with when thinking about physics.
However, I've come upon the following passage in the book by Streater and Wightman, page 86-87.

When, in fact, do non-separable Hilbert spaces appear in quantum mechanics? There are two cases which deserve mention. The first arises when one takes an infinite tensor product of Hilbert spaces... Infinite tensor products of Hilbert spaces are always non-separable. Since a Bose field can be thought of as a system composed of an infinitely of oscillators, one might think that such an infinite tensor product is the natural state space. However, it is characteristic of field theory that some of its observables involve all the oscillators at once, and it turns out that such observables can be naturally defined only on vectors belonging to a tiny separable subset of the infinite tensor product. It is the subspace spanned by such a subset that is the natural state space rather than the whole infinite tensor product itself. Thus, while it may be a matter of convenience to regard the state space as part of the infinite tensor product, it is not necessary.

My question is, how does one relate this passage to the usual quantisation procedure described above. In particular, what is the 'tiny separable subset' alluded to by Streater and Wightman?
Because the infinite tensor product picture is so intuitively compelling (this is emphasised, I think, by all authors on QFT), it would be nice to spell out the relation between it and standard quantisation with at least some level of mathematical clarity.

REPLY [3 votes]: There has been many answers, from many points of view on QFT: canonical quantization, algebraic QFT, etc. Let me add another perspective using the Euclidean path integral quantization, as emphasized in textbooks like "Quantum Field Theory and Critical Phenomena" by Jean Zinn-Justin.
For simplicity, I will take $M=\mathbb{R}^d$
In quantum mechanics, of say one particle in one dimension, the position is say $\phi\in\mathbb{R}$ and the momentum is say $\pi$. The physical Hilbert space is $L^2(\mathbb{R},d\phi)$ made of square integrable wave functions $\Psi(\phi)$, with respect to Lebsgue measure $d\phi$. When going to to QFT, the single number $\phi$, now becomes a classical field configuration $\phi=(\phi(x))_{x\in\mathbb{R}^d}$ and the physical Hilbert space should again be the $L^2$ space for a space of fields $\phi$ with respect to some measure $d\nu(\phi)$. Intuitively, one can think of this as a tensor product indexed by $x\in\mathbb{R}^d$ of $L^2$'s of $\mathbb{R}$, but trying to use this idea rigorously is probably counterproductive.
Suppose your theory is given by an honest Borel probability measure $\mu$ on $\mathscr{S}'(\mathbb{R}^{d+1})$, i.e., the Euclidean correlations or Schwinger functions
$$
S_n(x_1,t_1;\ldots;x_n,t_n)=\frac{1}{Z}\int 
\phi(x_1,t_1)\cdots\phi(x_n,t_n)\ e^{-S(\phi)}D\phi
$$
seen as distributions, namely, elements of $\mathscr{S}'(\mathbb{R}^{n(d+1)})$, are moments of a probability measure $\mu$. Suppose one can define a Borel-measurable restriction map $R:\mathscr{S}'(\mathbb{R}^{d+1})\rightarrow\mathscr{S}'(\mathbb{R}^{d})$ which sends the distribution $\phi(x,t)$ to $\phi(x,0)$. Then one can define a push-forward measure $\nu=R_{\ast}\mu$
and the physical Hilbert space should essentially be
$$
L^2(\mathscr{S}'(\mathbb{R}^d),d\nu)\ .
$$