Assume a massless scalar field in 3+1 dimensions which can be written as $$ \phi(t,\vec{x})=\int\frac{d^3k}{\sqrt{(2\pi)^32k}}\left(a(\vec{k})e^{-ikx}+a^\dagger(\vec{k})e^{ikx}\right)\, , $$ where $\vec{k}$ is the momentum (natrual units), $k=\vert\vec{k}\vert$, $kx=k_\mu x^\mu$ is the inner product of the four-vectors and $a(\vec{k})$ and $a^\dagger(\vec{k})$ obey the familiar commutator relations for the creation and annihilation operators.

Now I am interested in the equal-time 2-point correlator function $\langle 0\vert\phi(t,\vec{x}_1)\phi(t,\vec{x}_2)\vert 0\rangle$. If we pull out a piece of paper we can find $$ \langle 0\vert\phi(t,\vec{x}_1)\phi(t,\vec{x}_2)\vert 0\rangle =\int\frac{d^3k}{(2\pi)^32k}e^{i\vec{k}\cdot(\vec{x}_1-\vec{x}_2)}\, . $$ My approch to solve this integral explicitly is to transform it into spherical coordinates. Let $r:=\vert\vec{x}_1-\vec{x}_2\vert$. Thus, one can find $$ \langle 0\vert\phi(t,\vec{x}_1)\phi(t,\vec{x}_2)\vert 0\rangle =\int\limits_0^\infty dk\, k^2\int\limits_{-1}^1d\cos(\theta)\int\limits_0^{2\pi}d\varphi\,\frac{1}{(2\pi)^32k}e^{ikr\cos(\theta)}\\ =\frac{2\pi}{2(2\pi)^3ir}\int\limits_0^\infty dk\, (e^{ikr}-e^{-ikr})\, .\,\,\,\,\,\,\,\,\,(1) $$ Note, that the exponentials can be decomposed into a $2i\sin(kr)$. This integral looks divergent to me. But someone told me that I should define a new integral with a regulator $\epsilon>0$: $$ I(\epsilon)=\int\limits_0^\infty dk\, (e^{ikr}-e^{-ikr})e^{-\epsilon k}\, . $$ This integral can be calculated rather easily. Now I can take the limit $\epsilon\rightarrow 0$ and see that $I(0)$ converges and hence the two-point correlator from above converges.

My questions are:

  • Why is the interchange of limits possible such that I can even use $I(\epsilon\rightarrow 0)$ to express the two-point correlator?
  • Is this integral still a Riemann-integral? Since we made, in at least my eyes, the divergent integral in eq. (1) converge somehow.

2 Answers 2


Your original integral diverges, and hence the equation defining $\phi(x,t)$ is not well defined as written. As you've shown, there's simply no way around that. Since your result when considering $I(\epsilon)$ is finite, it simply cant be equal to what you get if you use your original expression for $\phi$.

I believe that the correct way to think about it is to simply go back and re-define what you mean by the field $\phi$. That is, the real definition of the $\phi$ is

$$ \phi(t,\vec{x})= \lim_{\epsilon\to 0} \int\frac{d^3k}{\sqrt{(2\pi)^32k}}\left(a(\vec{k})e^{-ikx}+a^\dagger(\vec{k})e^{ikx}\right)e^{-\epsilon k/2}. $$

This works since, if the integral were actually convergent and sufficiently well behaved, then you could exchange the integral and the limit and do away with the $\epsilon$ from the start—leaving you with an answer that would agree with the original definition. In the case that it diverges, then you've worked your exponential regulator into the theory.

  • 2
    $\begingroup$ Sorry but this is wrong. OP has not specified the decay rate of $a(\vec k)$ so there is no way to check whether the original integral diverges or not. The correct answer is that $\phi$ is a distribution and therefore the integrals are just a schematic notation. The two-point function is only defined when evaluated on suitable test functions, in which case the regularization is justified by the fact that the integral is absolutely convergent. $\endgroup$ Nov 16, 2021 at 21:45
  • $\begingroup$ I agree on that @AccidentalFourierTransform - the scalar field is indeed a distribution. Just to clear thing up: Equation (1) is a distribution and thus the integral is not a Riemann-integral. So we have to work with test functions, which is the regularization $e^{-\epsilon k}$ in this case. Then I can calculate $I(\epsilon\rightarrow 0)$. Formally, this can be used to express the original "integral" in eq. (1). Right? But one could have chosen any other suitable test function and have to obtain the same result? $\endgroup$
    – BraKet
    Nov 17, 2021 at 8:09

You have actually obtained the right answer in $(1)$ which is $$\langle0|\phi(x_1)\phi(x_2)|0\rangle=\frac{1}{8\pi |x_1-x_2|}\big[\delta(|x_1-x_2|)-\delta(|x_2-x_1|)\big]$$ Schwartz $(12.75)$. Looking at above expression you can see the correlator function is not analytical so special care has to be taken while using a regularization scheme. The way you have deformed UV region of your theory you obtain something like $\frac{1}{ir\pm\epsilon}$, since $\epsilon\to0$ any multiple of $r$ with it can be redfined to $\epsilon$, which have delta function hidden inside them.

You could have gone to another regularization scheme $m\to0$ where you obtain the same issue of $\frac{1}{r}$ and no $\delta(r)$. Sadly, I couldn't find the missing piece of $\delta(r)$ in this calculation.

The crux lies in non-analytical behavior of the above correlator, Schwartz sec $12.6$.

  • $\begingroup$ I can see your point. In the solution to the problem I gave above it is written that (1) ist just equal to a real number: $(1)=1/(4\pi^2r^2)$. This now confuses me since (1) is a distribution. With your "hidden delta function" (thanks for that, I didn't know that!) I obtain a different solution. So what is going on here? $\endgroup$
    – BraKet
    Nov 17, 2021 at 8:47

Your Answer

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