Is the thermal expectation value of a square of Hermitian operator always finite? If $\mathcal{O}$ is an hermitian operator in a system given by Hamiltonian $H$ and inverse temperature $\beta$, is $$\langle \mathcal{O} \mathcal{O} \rangle =  Tr (e^{-\beta H} \mathcal{O} \mathcal{O})$$ always convergent, or do we have to regulate it ? I am assuming that the system is a very general quantum field theory and the hamiltonian is bounded from below.  
 A: In general, the expectation value $tr(A e^{-\beta H})$ is not defined for a generic selfadjoint operator $A$ (of the form ${\cal O}^2$ or not) if it is unbounded as it is the standard situation in QFT.
$tr(A e^{-\beta H})$ however converges if (with the written order!) the range of $e^{-\beta H}$ belongs to the domain of $A$ and the composition is trace class.
As an example, the hypotheses are valid if $H$ is selfadjoint and bounded below, with pure point spectrum and the dimension of eigenspaces is finite and bounded. In this case, for instance, $A= H^n$ for $n=0,1,2, \ldots$ satisfies all conditions.
When the expectation value is not defined with the formula above,  one may try to regularize taking some thermidynamical limit or directly use a more advanced formalism like the algebraic one,  together with the KMS condition to characterize thermal states.
In QFT in Minkowski spacetime, for a free theory, $e^{-\beta H}$ though bounded is never trace class, since its spectrum is not a pure point spectrum. Therefore you always must regularize, independently from the choice of $A$, barring forced unphysical choices ($A$ of trace class).
A: No.
Consider for example observables of the form $\mathcal O = e^{\beta H/2}$; the value exponentially increases as the probability exponentially decreases and you can get what's in principle an infinite sum of constant terms.
A: Now I have a clear reason why thermal expectation value of a square of an operator is not convergent. In any generic field theory I can express the expectation value as 
$$ Tr (e^{-\beta H} O O) = \sum_{i,k} e^{-\beta e_{i} }O_{ik} O^*_{ik}$$ Since there is no damping factor in the summation k, the sum is not convergent for any theory and operator. 
