Loops and regularization

A loop integral is not defined merely by writing an integral sign. Its ultraviolet, infrared, and threshold regions have different physical origins; a regulator makes intermediate expressions meaningful but can also disturb symmetries or hide the domain in which an expansion is valid. This lesson develops one logarithmically divergent integral in two regulators and shows which parts of the answer can be compared before any renormalization condition is imposed.

Required background. Perturbative expansion and Feynman rules supplies propagators, vertices, loop measures, and symmetry factors. LSZ reduction and tree amplitudes supplies the distinction between a loop correction to a correlator and a correction to a physical amplitude. Helpful background. Review complex and asymptotic methods if contour deformation or analytic continuation is the blocking step.

Classify the singular region before choosing a regulator

For large Euclidean momentum $k$, a scalar propagator behaves as $k^{-2}$. A one-loop integrand with two propagators in four dimensions therefore has

$$\int^{\Lambda}\mathrm d^4k\,k^{-4} \sim\int^{\Lambda}\frac{\mathrm dk}{k},$$

so it is logarithmically ultraviolet divergent. This estimate says where the problem lies and how it scales; it does not define the integral or determine its finite part.

Other singular regions require different responses:

Region	Diagnostic question	Typical control
Ultraviolet	Does the integrand fail to decay at large loop momentum?	Cutoff, dimensional regularization, lattice spacing, Pauli–Villars fields, or another declared UV regulator
Infrared	Do massless or on-shell denominators become singular at small momentum or long distance?	Mass, finite volume, off-shellness, dimensional continuation, or an inclusive observable
Threshold or pinch	Do poles trap the contour as external invariants cross a physical threshold?	The inherited $i0$, analytic continuation, and a stated sheet
Spurious	Did an algebraic split create terms more singular than their sum?	Recombine terms before assigning physical meaning

A UV counterterm cannot remove a physical threshold cut, and an infrared-safe observable is not produced by subtracting a UV pole. Keep the classifications separate even when one regulator happens to display more than one of them.

One Euclidean integral exposes the ultraviolet logarithm

Take $m>0$ and define

$$I(m)=\int\frac{\mathrm d^4k_E}{(2\pi)^4} \frac{1}{(k_E^2+m^2)^2}.$$

The mass removes the small-$k_E$ singularity, so any divergence here is ultraviolet. A sharp rotationally invariant cutoff gives

$$I_\Lambda(m) =\int_{|k_E|<\Lambda}\frac{\mathrm d^4k_E}{(2\pi)^4} \frac{1}{(k_E^2+m^2)^2}.$$

Using the area $2\pi^2$ of the unit three-sphere,

$$\begin{aligned} I_\Lambda(m) &=\frac{1}{8\pi^2}\int_0^\Lambda \frac{k^3\,\mathrm dk}{(k^2+m^2)^2}\\ &=\frac{1}{16\pi^2} \left[ \log\frac{\Lambda^2+m^2}{m^2} +\frac{m^2}{\Lambda^2+m^2}-1 \right]. \end{aligned}$$

Thus

$$I_\Lambda(m) =\frac{1}{16\pi^2} \left[ \log\frac{\Lambda^2}{m^2}-1 +\mathcal O\!\left(\frac{m^2}{\Lambda^2}\right) \right].$$

The coefficient of the logarithm is robust in comparisons that preserve the same low-energy normalization. The constant $-1$ belongs to this particular cutoff shape and definition. Changing the regulator can change such local finite terms before renormalized parameters are matched.

Dimensional regularization displays the same coefficient

Continue the Euclidean dimension to $d=4-2\epsilon$ and introduce a reference scale $\mu$ so that the integral keeps its four-dimensional mass dimension:

$$I_d(m;\mu) =\mu^{2\epsilon} \int\frac{\mathrm d^d k_E}{(2\pi)^d} \frac{1}{(k_E^2+m^2)^2}.$$

The Schwinger representation and Gaussian integral give

$$\frac{1}{(k_E^2+m^2)^2} =\int_0^\infty\mathrm ds\,s\,e^{-s(k_E^2+m^2)},$$

and

$$\int\frac{\mathrm d^d k_E}{(2\pi)^d}e^{-sk_E^2} =(4\pi s)^{-d/2}.$$

Therefore

$$I_d(m;\mu) =\frac{\mu^{2\epsilon}}{(4\pi)^{d/2}} \Gamma\!\left(2-\frac d2\right) (m^2)^{d/2-2}.$$

Expanding at small positive $\epsilon$ gives

$$I_d(m;\mu) =\frac{1}{16\pi^2} \left[ \frac1\epsilon-\gamma_E+\log4\pi +\log\frac{\mu^2}{m^2} +\mathcal O(\epsilon) \right].$$

The $1/\epsilon$ pole represents the same logarithmic ultraviolet sensitivity as $\log\Lambda^2$. The scalar master integral and its expansion are developed in Schwartz 2014, App. B.3.2–3, pp. 826–828. His Lorentzian master integral carries a Wick-rotation factor $i$, absent from our Euclidean integral. His convention $d=4-\epsilon_{\mathrm S}$ gives $2/\epsilon_{\mathrm S}$; setting $\epsilon_{\mathrm S}=2\epsilon$ gives our $1/\epsilon$. A regulator does not make the theory finite: a subtraction and renormalization condition are still required.

With the modified minimal-subtraction scale $\overline\mu^2=4\pi e^{-\gamma_E}\mu^2$, the bracket becomes

$$\frac1\epsilon+\log\frac{\overline\mu^2}{m^2} +\mathcal O(\epsilon).$$

That notation packages two constants; it is a scheme convention, not a new physical scale. In the canonical logarithmic master-integral expansion, track the pole, Euler constant and four-pi factor together before choosing a subtraction.

External momentum reveals thresholds as well as UV behavior

The equal-mass bubble before continuation has the schematic form

$$B(p^2)=\mu^{2\epsilon} \int\frac{\mathrm d^d k}{(2\pi)^d} \frac{i}{k^2-m^2+i0} \frac{i}{(k+p)^2-m^2+i0}.$$

Feynman parameterization uses

$$\frac1{AB}=\int_0^1\frac{\mathrm dx}{[xA+(1-x)B]^2}.$$

After a regulated shift of loop momentum, the denominator depends on

$$\Delta(x,p^2)=m^2-x(1-x)p^2.$$

The UV pole is independent of $p^2$ because large loop momentum cannot resolve the external scale at leading order. The finite part contains $\log[\Delta(x,p^2)-i0]$. Since $x(1-x)\le1/4$, the argument can vanish for $p^2\ge4m^2$, producing the two-particle branch cut. That nonanalyticity is physical threshold information, not a UV divergence to subtract Srednicki 2006 draft, § 15, pp. 120–121, Eq. (15.5) and threshold discussion, PDF.

Compare the canonical massive Euclidean bubble: its momentum-independent UV pole and finite momentum dependence have the same roles. Its Euclidean normalization does not include the two Lorentzian propagator factors of $i$ displayed here.

The shift of loop momentum deserves emphasis. In an absolutely convergent integral it is an ordinary change of variables. Dimensional regularization can establish that identity in a convergence domain and then analytically continue the regulated result Srednicki 2006 draft, § 14, pp. 111–114, Eqs. (14.12–14.18), PDF. A change of variables moves a sharp cutoff ball as well. Imposing the original ball again after the shift changes the finite regulator. Symmetry cancellations that compare different loop routings must use one consistent prescription; this observation alone does not establish a nonzero Ward-violating surface term for the present logarithmic bubble. The loop-shift identities in dimensional continuation are discussed in ’t Hooft and Veltman 1972, § 3, p. 197.

Power counting predicts candidates, not final answers

For a scalar graph in four dimensions with propagators behaving as $k^{-2}$, the superficial degree of divergence is

$$\omega=4L-2I.$$

In quartic theory, $4V=2I+E$ and, for a connected graph, $L=I-V+1$. Eliminating $I$ and $V$ gives

$$\omega=4-E.$$

Two-point graphs are superficially quadratic, four-point graphs logarithmic, and graphs with more external legs superficially convergent. This is only a first screen. Symmetry can cancel the leading terms; derivative numerators can worsen counting; and a graph with $\omega<0$ can contain divergent subgraphs. Renormalization must treat all divergent subgraphs consistently, not just the overall large-momentum limit Srednicki 2006 draft, § 18, pp. 129–131, Eqs. (18.2–18.8) and subdivergences, PDF.

Regulators are judged by what they preserve and expose

Before accepting a regulated result, state:

• the integration contour or Euclidean continuation;
• the regulator and parameters being held fixed;
• which symmetries are exact, broken, or recovered only after counterterms;
• the order in which regulator, volume, mass, and on-shell limits are taken;
• every UV, IR, and threshold singular region; and
• one independent check, such as a Ward identity, derivative with respect to a mass, or agreement of a universal logarithm across regulators.

No regulator is selected solely by making an integral easy. A lattice gives a nonperturbative UV definition but breaks continuous rotations at finite spacing. A sharp momentum cutoff requires care with loop-momentum translations and symmetry identities. Pauli–Villars fields can preserve some symmetries but add heavy auxiliary content.

The scale-covariant analytic prescription of dimensional regularization assigns zero to scaleless integrals. That zero does not establish the absence of UV or IR singular regions: analytically continued endpoint pieces can contain cancelling UV and IR poles Schwartz 2014, App. B.3.3, p. 829, Eqs. (B.48–B.49). Use the canonical scaleless UV/IR separation to separate those endpoints before extracting ultraviolet information.

Gauge identities must be checked with the chosen continuation and subtraction prescription. Intrinsically four-dimensional objects such as $\gamma^5$ and Levi-Civita tensors require an explicit continuation prescription; their ordinary four-dimensional identities cannot simply all be imposed in general dimension ’t Hooft and Veltman 1972, § 6, pp. 206–208.

Exercises

1. Derive the cutoff result

Evaluate $I_\Lambda(m)$ by setting $u=k^2$. Verify both its dimension and its large-$\Lambda$ limit.

Solution

The angular integral gives $1/(8\pi^2)$, and $k^3\mathrm dk=u\,\mathrm du/2$. Hence

$$I_\Lambda =\frac{1}{16\pi^2} \int_0^{\Lambda^2}\frac{u\,\mathrm du}{(u+m^2)^2}.$$

An antiderivative is

$$\log(u+m^2)+\frac{m^2}{u+m^2}.$$

Evaluating the endpoints gives

$$I_\Lambda =\frac{1}{16\pi^2} \left[ \log\frac{\Lambda^2+m^2}{m^2} +\frac{m^2}{\Lambda^2+m^2}-1 \right].$$

The measure has mass dimension four and the denominator dimension four, so the result is dimensionless. Expanding at large cutoff recovers $[\log(\Lambda^2/m^2)-1]/(16\pi^2)$ plus terms suppressed by $m^2/\Lambda^2$.

2. Expand the dimensionally regulated integral

Starting from the Gamma-function result, use $\Gamma(\epsilon)=1/\epsilon-\gamma_E+\mathcal O(\epsilon)$ to reproduce the pole and logarithm.

Solution

For $d=4-2\epsilon$,

$$I_d =\frac{1}{(4\pi)^2} \Gamma(\epsilon) \left(\frac{4\pi\mu^2}{m^2}\right)^\epsilon.$$

The second factor is

$$1+\epsilon\log\frac{4\pi\mu^2}{m^2} +\mathcal O(\epsilon^2).$$

Multiplying gives

$$I_d =\frac{1}{16\pi^2} \left[ \frac1\epsilon-\gamma_E+\log4\pi +\log\frac{\mu^2}{m^2} +\mathcal O(\epsilon) \right].$$

The coefficient of $1/\epsilon$ matches the coefficient of $\log\Lambda^2$ in the cutoff result.

3. Derive the quartic power-counting rule

For a connected graph in four-dimensional $\phi^4$ theory, use line-end and topology identities to show $\omega=4-E$. Explain why this does not prove that every six-point graph is finite.

Solution

Counting line ends gives $4V=2I+E$, and connected topology gives $L=I-V+1$. Therefore

$$\begin{aligned} \omega &=4(I-V+1)-2I\\ &=2I-4V+4\\ &=4-E. \end{aligned}$$

For $E=6$, the overall degree is $-2$, so the graph is superficially convergent. It can still contain a two- or four-point subgraph with nonnegative degree. The corresponding subintegral must be renormalized before the full graph is assigned a finite value.

Continue from regulation to renormalization

You are ready to continue when you can identify the singular region, derive the same logarithmic coefficient with a cutoff and dimensional continuation, state what each regulator preserves, and distinguish the UV pole from the two-particle threshold.

Next, Renormalization and the renormalization group introduces renormalized inputs, counterterms, scheme dependence, and scale evolution. Do not remove the pole by an unnamed subtraction: the finite definition of the parameters is part of the physical prediction.

References

• Matthew D. Schwartz, Quantum Field Theory and the Standard Model, Cambridge University Press, 2014, doi:10.1017/9781139540940.
• Mark Srednicki, Quantum Field Theory, prepublication draft, 2006. Author’s source and errata page; Open PDF.
• Gerard ’t Hooft and Martinus Veltman, “Regularization and Renormalization of Gauge Fields,” Nuclear Physics B 44 (1972), 189–213. doi:10.1016/0550-3213(72)90279-9; Open PDF.

Further reading

• John C. Collins, Renormalization, Cambridge University Press, 1984, doi:10.1017/CBO9780511622656.