In the final chapter of his book "An Interpretive Introduction to Quantum Field Theory", Paul Teller writes about three interpretations of renormalization in quantum field theory. In particular, Teller denotes one of these approaches "Real-Infinities Renormalization," which eschews regularization and as such, he concludes, doesn't make much sense. However, Teller suggests the "Real-Infinities" approach could possibly be made rigorous using Conway/surreal numbers. The relevant section of Teller's book can be found on Google Books here: Teller, pg. 162.

My question is this: has anyone tried to formulate renormalization in quantum field theories without the use of regulators, in terms of Conway/surreal numbers as suggested by Teller?

| cite | improve this question | | | | |
  • $\begingroup$ Have you made any progress on this? $\endgroup$ – meowzz Jun 26 '18 at 9:17

One of the divergence problem in quantum field theory stems from the use of products of distributions, which are not well defined (due to Schwartz's impossibility theorem). The simplest example of this is the Hamiltonian density of a free scalar field. Consider this :

The Hamiltonian of a free scalar field is

\begin{eqnarray} H &=& \frac{1}{2} \int d^3x (\pi^2 + (\nabla \phi)^2 + m^2 \phi^2)\\ &=& \int \frac{d^3p}{(2\pi)^3} \frac{\omega_{\vec{p}}}{2} [a_{\vec{p}} a^\dagger_{\vec{p}} + a^\dagger_{\vec{p}} a_{\vec{p}} ]\\ &=& \int \frac{d^3p}{(2\pi)^3} \frac{\omega_{\vec{p}}}{2} [a^\dagger_{\vec{p}} a_{\vec{p}} + (2\pi)^3 \delta^{(3)}(0)] \end{eqnarray}

This isn't terribly rigorous for a number of reasons (also we are using generalized eigenstates as a basis, complicating things further), but importantly we are performing the following product : for the operator-valued distribution $a$, we are trying to compute $\hat{a}(x) \hat{a}^\dagger(x)$, but as we saw earlier, this quantity isn't well defined (as their commutator is a delta function it's easy to show that they can't be non-singular distributions), we're essentially trying to compute the product of two delta functions, which if we try to force it by going to function sequences can be shown to be divergent :

\begin{eqnarray} \int dx [\delta(x)]^2 &=& \lim_{\varepsilon \to 0} \int dx ({\frac {1}{\sqrt {2\pi \varepsilon }}}\mathrm {e} ^{-{\frac {x^{2}}{2\varepsilon }}})^2\\ &=& \lim_{\varepsilon \to 0} \int dx {\frac {1}{{2\pi \varepsilon }}}\mathrm {e} ^{-{\frac {x^{2}}{\varepsilon }}}\\ &=& \lim_{\varepsilon \to 0} (2\sqrt{\varepsilon})^{-1} \\ \end{eqnarray}

which is indeed divergent. The quantity is well-enough defined at two different points (from the identity $[a_p, a^\dagger_q] = (2\pi \delta^{(3)}(p - q))$), when their singular support are disjoint (the singularities are at two different points), but it ceases to be so if we try to evaluate them at the same point.

There are many physicist tricks to get around this, but the "proper" way for the product of two distributions is to use a wider space containing distributions, so-called generalized functions. The most common ones are Colombeaux algebras, which are algebras of distributions (which have been used for QFT a few times, including by Colombeau himself).

The trick with Colombeau algbras is the following : rather than considering a distribution as a linear functional on a function space, we consider equivalency classes of suitably mollified distributions (that is, distributions convoluted with some appropriate bump function). For some mollifier $\eta$, and some distribution $f$, we consider the class of functions

$$f_\varepsilon = \frac{1}{\varepsilon} \eta (-\frac{x}{\varepsilon}) * f$$

The equivalency class is defined using the set of moderate functions $\mathcal{E}_M$ (generalized functions such that for some $n \geq 0$, $f_\varepsilon = \mathcal{O}(\varepsilon^{-n})$), and the set of negligible functions $\mathcal{N}$ (for all $q > 0$ , $f_\varepsilon - f = \mathcal{O}(\varepsilon^q)$). The basic idea is that negligible functions are "infinitesimal" while moderate functions are not. The Colombeau algebra is then the equivalence class of moderate functions up to negligible functions

$$\mathcal{G} = \mathcal{E}_M / \mathcal{N}$$

In other words, for a distribution $f$, its representatives are of the form $f_\varepsilon + \mathcal{N}$.

The Colombeau algebra is strictly greater than the space of distributions, as we've seen that the product of two distributons will not necessarily be one. It can be defined as a function space on what are called generalized numbers $\tilde{\mathbb{R}}$, which are where the "real infinities" come in. we can consider the algebra of generalized functions at specific points $x$ (since unlike distributions, generalized functions are just functions from the space $\mathbb{R}^n$). Generalized numbers only form a ring (there are some elements which do not admit an inverse), and as can be expected, they are not archimedean (there are infinite and infinitesimal values).

Once you have those elements, you can perform this analysis on QFT : take the convolution of field operators with some mollifier and consider their equivalency classes. The Colombeau-Gsponer paper does this, which, after a fairly long calculation, finds that the Hamiltonian is a sum of four operator-valued generalized functions : two negligible functions, a moderate function (that is just the energy part of the Hamiltonian), and the vacuum energy term which is of the form

\begin{eqnarray} E &=& \frac{1}{2} \int d^3p_1 d^3p_2 \hat \rho(\varepsilon p_1) \hat \rho^*(\varepsilon p_2) \exp{[i(E_{p_1} - E_{p_2}) (t - t\xi)]} \\ && \frac{(E_{p_1} E_{p_2} + p_1 \cdot p_2 + m^2)(E_{p_1} + E_{p_2})}{4E_{p_1} E_{p_2}} |\frac{1}{\varepsilon^3} \chi (\frac{p_1 - p_2}{\varepsilon})|^2 \end{eqnarray}

while you can compute that integral, the important point is that it contains a term in $\mathcal{O}(\varepsilon^{-3})$, ie it is divergent. But this term is just a generalized number, it does not depend on any position in space. You can then redefined the Hamiltonian as simply

$$\hat{H}_R = \hat{H} - E$$

The Colombeau algebra isn't the only algebra for distributions. There are other, quite possibly more fitting algebras that can be used for it, including the algebra of distributions with the hyperreal numbers. It's quite easy to build a delta function from most of the non-archimedean fields, such as the hyperreal numbers or the asymptotic numbers, simply by picking for instance the Gaussian

$$\delta_\varepsilon(x) = \frac{1}{\sqrt{2\pi \varepsilon}} {e} ^{-{\frac {x^{2}}{2\varepsilon }}}$$

with $\varepsilon$ some infinitesimal number in the hyperreal, in which case any distribution mapped to the hyperreal (via Łoś's theorem) can become a simple function over the hyperreal via

$$f_\varepsilon = f * \delta_\varepsilon$$

The basic idea is the same as the Colombeau method, although the hyperreals and related sets have the benefit of being a field rather than a ring. More details can be found on the topic for instance in Todorov's papers.

There really is nothing magical about using "infinite" numbers for renormalization. Both the generalized numbers and hyperreal can be defined as some sequence of numbers, for instance an infinitesimal hyperreal can be written as

$$\varepsilon = (1, 1/2, 1/3, ...)$$

while an infinite number will be

$$\omega = (1, 2, 3, ...)$$

In the end, as far as the fundamental idea goes, this is the same as renormalization via the usual means : you consider the divergent quantity as a divergent sequence of finite objects from which you remove a finite quantity at each point of the sequence.

| cite | improve this answer | | | | |
  • $\begingroup$ Nice answer. While there is good info here, it doesn't explicitly reference conway/surreal numbers. Definitely on the right track tho! $\endgroup$ – meowzz Jul 9 '18 at 13:53
  • 1
    $\begingroup$ As far as I know, there are no elaborate usage of surreal numbers for renormalization, due to the fact that overall, the need for very elaborate infinities isn't strictly necessary for it (the distribution algebras tend to use asymptotic numbers rather than hyperreals directly for instance, where only powers of a specific infinite scale are used). The important part is more that they can replicate a divergent quantity, although using a similar method would certainly give the same kind of results, assuming the analysis on surreal number (cf arxiv.org/abs/1307.7392) allows such things. $\endgroup$ – Slereah Jul 9 '18 at 13:57
  • $\begingroup$ In terms of surreals, epsilon={0|1,1/2,1/3,...} & omega={0|1,2,3,...}. While there may not be elaborate usage of surreals for renormalization, I am still interested in how it would work. Also, I am particularly interested in star, games & surcomplex numbers. $\endgroup$ – meowzz Jul 9 '18 at 15:08
  • 1
    $\begingroup$ Well, the procedure for all these methods is usually just to find some embedding of distributions to functions over the non-archimedean set $$\iota : D^* \to C^\infty(\mathbb{R},S)$$some projection back to real numbers where applicable $\pi : S \to \mathbb{R}$ and from there use the natural algebraic structure take over. The difficulty for surreal numbers might be that the usual analysis tools may not adequately suffice for such things (although the linked paper may differ). What I feel may happen is simply that in the end you'll simply use a subset of surreal numbers equivalent to hyperreals. $\endgroup$ – Slereah Jul 9 '18 at 15:14
  • $\begingroup$ "subset of surreals equivalent to hyperreals" sounds about right. $\endgroup$ – meowzz Jul 9 '18 at 15:17

There are attempts to use nonstandard analysis (e.g., Albeverio) or Colombeau algebras, but these haven't been developed very far. I haven't seen anything in terms of surreal numbers, but they may probably substitute for the infinitesimals in nonstandard analysis.

| cite | improve this answer | | | | |
  • $\begingroup$ Can you provide references? $\endgroup$ – sferics Nov 9 '15 at 21:25
  • 1
    $\begingroup$ I wrote this from memory and would have to search for references. But I cannot recommend this approach, hence have no motivation to look up the references. With the hintsin my answer and appropriate key words you might conduct your own search. $\endgroup$ – Arnold Neumaier Nov 10 '15 at 5:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.