As mentioned, in perturbation theory one initially only aims at defining the expectations (A.1) as a formal power series in the loop counting parameter , where the sum of all -loop contributions to a quantity is assigned a factor . For the reasons explained in  the loop expansion does not necessarily coincide with an expansion in powers of Planck’s constant . For example when massless fields are involved, 1-loop diagrams can contribute to the classical limit . The loop counting parameter refers to a set of free fields of mass such that the formal expansion of the exponential in Equation (A.1) gives expectations whose computation can be reduced to the evaluation of Gaussians. We denote the (quadratic) action of this set of free fields by . The interaction is described by a set of monomials , , which are “power counting renormalizable”. The latter means that their mass dimension is such that . It is also assumed that the are functionally independent, so that the corresponding couplings are essential. The so-called “bare” action functional then is , where are the essential “bare” couplings (including masses) corresponding to the interaction monomials . Inessential parameters are generated by subjecting to a suitable class of field redefinitions. In more detail one writes for a discussion).
The normalizations in Equation (2.1) can be chosen such that and , but one is really interested in the regime where . Inserting these parameterizations into gives an expression of the form
So far the counterterms in Equation (2.1) have been left unspecified. The point of introducing them is of course as a means to absorb the cut-off dependence generated by the regularized functional integral in Equation (A.1). Specifically, one replaces the Boltzmann factor by its power series expansion in , i.e. , and aims at an evaluation of multipoint functions as formal power series in . After inserting Equation (2.1) and the expansion of this reduces the problem to an evaluation of the free multipoint functions computed with the quadratic action on the field space with cutoff . The free multipoint functions will contain contributions which diverge in the limit . On the other hand via the parameterization (2.1, 2.2) the coefficients carry an adjustable dependence. In a renormalizable QFT the dependence in the coefficients can be chosen such as to cancel (for ) that generated by the multipoint functions . With this adjustment the limits
Since the renormalization scale is arbitrary, changing its value must not affect the values of observables. The impact of a change in can most readily be determined from Equation (2.1). The left-hand-sides are independent, so by differentiating these relations with respect to and extracting the coefficients in a power series in (say) and/or consistency conditions arise for the derivatives and . The ones obtained from the leading order are the most interesting relations. For the couplings one obtains a system of ordinary differential equations which define their renormalization flow under a change of . As usual it is convenient to work with dimensionless couplings , where is the mass dimension of . The flow equations then take the form
By construction the perturbative beta functions have a fixed point at , which is called the perturbative Gaussian fixed point. Nothing prevents them from having other fixed points, but the Gaussian one is built into the construction. This is because a free theory has vanishing beta functions and the couplings have been introduced to parameterize the deviations from the free theory with action . Not surprisingly the stability matrix of the perturbative Gaussian fixed point just reproduces the information which has been put in. The eigenvalues come out to be modulo corrections in the loop coupling parameter, where are the mass dimensions of the corresponding interaction monomials. For the eigenvectors one finds a one-to-one correspondence to the unit vectors in the ‘coupling direction’ , again with power corrections in the loop counting parameter. One sees that the couplings not irrelevant with respect to the stability matrix computed at the perturbative Gaussian fixed point are the ones with mass dimensions , i.e. just the power counting renormalizable ones.
The attribute “perturbative Gaussian” indicates that whenever in a nonperturbative construction of the renormalization flow in the same ‘basis’ of interaction monomials is also a fixed point (called the Gaussian fixed point), the perturbatively defined expectations are believed to provide an asymptotic (nonconvergent) expansion to the expectations defined nonperturbatively based on the Gaussian fixed point, schematically) and solves as an ordinary differential equation. One of the functions obtained, say , is used to eliminate in favor of and an integration constant (not to be confused with the cutoff, which is gone for good). So at this point carries a two-fold -dependence, the one which comes out of the renormalization procedure (2.3) and the one carried now by . For an observable quantity both dependencies cancel out, modulo terms of higher order, leaving behind a dependence on the integration constant . We write to indicate this dependence. One then uses the expectation of one, suitably chosen, observable to match its value (measured or otherwise known) with that of to a given small loop order (typically not larger than 2). For a well chosen this allows one to replace by a physical mass scale . Eliminating in favor of gives the perturbative predictions for all other observables. Apart from residual scheme dependencies (which are believed to be numerically small) this defines the right-hand-side of Equation (2.5) unambigously as a functional over the observables.
Nevertheless, except for some special cases, it is difficult to give a mathematically precise meaning to the ‘’ in Equation (2.5). Ideally one would be able to prove that perturbation theory is asymptotic to the (usually unknown) exact answer for the same quantity. For lattice theories on a finite lattice this is often possible; the problems start when taking the limit of infinite lattice size (see  for a discussion). In the continuum limit a proof that perturbation theory is asymptotic has been achieved in a number of low-dimensional quantum field theories: the superrenormalizable and theories [69, 43] and the two-dimensional Gross–Neveu model, where the correlation functions are the Borel sum of their renormalized perturbation expansion [87, 89]. Strong evidence for the asymptotic correctness of perturbation theory has also been obtained in the O(3) nonlinear sigma-model via the form factor bootstrap . In four or higher-dimensional theories unfortunately no such results are available. It is still believed that whenever the above is asymptotically free in perturbation theory, that the corresponding series is asymptotic to the unknown exact answer. On the other hand, to the best of our knowledge, a serious attempt to establish the asymptotic nature of the expansion has never been made, nor are plausible strategies available. The pragmatic attitude usually adopted is to refrain from the attempt to theoretically understand the domain of applicability of perturbation theory. Instead one interprets the ‘’ in Equation (2.5) as an approximate numerical equality, to a suitable loop order and in a benign scheme, as long as it works, and attributes larger discrepancies to the ‘onset of nonperturbative physics’. This is clearly unsatisfactory, but often the best one can do. Note also that some of the predictive power of the QFT considered is wasted by this procedure and that it amounts to a partial immunization of perturbative predictions against (experimental or theoretical) refutation.
So far the discussion was independent of the nature of the running of (which was traded for ). The chances that the vague approximate relation ‘’ in Equation (2.5) can be promoted to the status of an asymptotic expansion are of course way better if is driven towards by the perturbative flow. Only then is it reasonable to expect that an asymptotic relation of the form (2.5) holds, linking the perturbative Gaussian fixed point to a genuine Gaussian fixed point defined by nonperturbative means. The perturbatively and the nonperturbatively defined coupling can then be identified asymptotically and lie in the unstable manifold of the fixed point . On the other hand the existence of a Gaussian fixed point with a nontrivial unstable manifold is thought to entail the existence of a genuine continuum limit in the sense discussed before. In summary, if is traded for a running , a perturbative criterion for the existence of a genuine continuum limit is that the perturbative flow of is regular with . Since the beta functions of the other couplings are formal power series in without constant coefficients, the other couplings will vanish likewise as , and one recovers the local quadratic action at the fixed point. The upshot is that the coupling with respect to which the perturbative expansion is performed should be asymptotically free in perturbation theory in order to render the existence of a nonperturbative continuum limit plausible.
The reason for going through this discussion is to highlight that is applies just as well to a perturbative non-Gaussian fixed point. This sounds like a contradiction in terms, but it is not. Suppose that in a situation with several couplings the perturbative beta functions (which are formal power series in without constant coefficients) admit a nontrivial zero, . Suppose in addition that all the couplings lie in the unstable manifold of that zero, i.e. the flows are regular and . We shall call a coupling with this property asymptotically safe, so that the additional assumption is that all couplings are asymptotically safe. As before one must assign a numerical value in order to define the flow. Since the series in anyhow has zero radius of convergence, the ‘smallness’ of is not off-hand a measure for the reliability of the perturbative result (the latter intuition in fact precisely presupposes Equation (2.5)). Any one of the deviations , which is of order at some can be used as well to parameterize the original loop expansion. By a relabeling or reparameterization of the couplings we may assume that this is the case for . The original loop expansion can then be rearranged to read . However, if there is an underlying nonperturbative structure at all, it is reasonable to assume that it refers to a non-Gaussian fixed point,
Summarizing: In perturbation theory the removal of the cutoff can be done independently of the properties of the coupling flow, while in a non-perturbative setting both aspects are linked. Only if the coupling flow computed from the perturbative beta functions meets certain conditions is it reasonable to assume that there exists an underlying non-perturbative framework to whose results the perturbative series is asymptotic. Specifically we formulate the following criterion:
Criterion (Continuum limit via perturbation theory):
(PTC1) Existence of a formal continuum limit, i.e. removal of the UV cutoff is possible and the renormalized physical quantities are independent of the scheme and of the choice of interpolating fields, all in the sense of formal power series in the loop counting parameter.
(PTC2) The perturbative beta functions have a Gaussian or a non-Gaussian fixed point and the dimension of its unstable manifold (as computed from the perturbative beta functions) equals the number of independent essential couplings. Equivalently, all essential couplings are asymptotically safe in perturbation theory.
© Max Planck Society and the author(s)