Renormalization and the renormalization group

Regularization makes loop integrals well defined; renormalization replaces regulator-dependent parameters by a finite set of declared inputs; the renormalization group (RG) ensures that an arbitrary renormalization scale does not alter a physical prediction. These are three related operations, not three names for removing an infinity.

This lesson carries one dimensionless quantity through that sequence. It shows how counterterms and renormalization conditions define finite parameters, how explicit and implicit scale dependence cancel, and which statements about running couplings, fixed points, and generated scales are actually justified.

Required background. Loops and regularization supplies the ultraviolet poles and logarithms that must be renormalized.

Helpful background. Perturbative expansion and Feynman rules explains why loop graphs and counterterm graphs must be kept at the same perturbative order.

From a regulated action to finite inputs

Consider a real scalar field in $d=4-2\epsilon$ dimensions. Write the bare Lagrangian as

$$\mathcal L_0 = \frac12(\partial_\mu\phi_0)(\partial^\mu\phi_0) -\frac12m_0^2\phi_0^2 -\frac{\lambda_0}{4!}\phi_0^4.$$

Introduce a renormalized field and additive counterterms through

$$\begin{aligned} \phi_0&=Z_\phi^{1/2}\phi, & \delta Z_\phi&=Z_\phi-1, \\ Z_\phi m_0^2&=m^2+\delta m^2, & Z_\phi^2\lambda_0&=\mu^{2\epsilon}(\lambda+\delta\lambda). \end{aligned}$$

After substitution, the same regulated Lagrangian is

$$\mathcal L_0=\mathcal L_{\rm ren}+\mathcal L_{\rm ct},$$

with

$$\begin{aligned} \mathcal L_{\rm ren} &= \frac12(\partial\phi)^2 -\frac12m^2\phi^2 -\mu^{2\epsilon}\frac{\lambda}{4!}\phi^4, \\ \mathcal L_{\rm ct} &= \frac12\delta Z_\phi(\partial\phi)^2 -\frac12\delta m^2\phi^2 -\mu^{2\epsilon}\frac{\delta\lambda}{4!}\phi^4. \end{aligned}$$

This is an exact reparametrization of the regulated theory. In perturbation theory, each $\delta$ is expanded in loop order. Its pole part cancels the ultraviolet pole of loop graphs, while its finite part is fixed by a renormalization prescription. Locality and the symmetries determine which counterterm operators are allowed; the calculation determines their coefficients. The systematic organization of loop graphs and counterterm insertions is developed in Collins 1984, §§5.6–5.7, pp. 112–125.

The renormalized parameters become useful only after conditions give them a meaning. Examples include:

• an on-shell condition, which fixes a stable particle’s propagator pole and residue and defines a coupling from a stated physical process;
• momentum subtraction, which fixes two- and higher-point functions at a specified nonexceptional momentum configuration; and
• minimal subtraction or $\overline{\mathrm{MS}}$, which removes a prescribed pole package and determines the parameters only after they are matched to physical inputs.

Different conditions can assign different numerical values to $m(\mu)$ and $\lambda(\mu)$ while predicting the same non-input observable after all parameters and coefficient functions are translated consistently. This finite-redefinition principle is established in Collins 1984, §7.1, pp. 169–176.

At a fixed regulator, bare quantities are held fixed when the renormalization scale is varied. When the regulator is removed at fixed physical inputs, the bare parameters and counterterms may separately be singular. Their separate values are not measurements; the test is whether the renormalized prediction has a finite regulator-free limit.

Four scales and choices that must not be conflated

Object	Example	What changing it means	What a sound calculation does
Regulator	$\epsilon$ in $d=4-2\epsilon$, or an auxiliary cutoff $\Lambda_{\rm reg}$	Changes the temporary definition of divergent integrals	Cancels regulator dependence and removes the regulator at fixed inputs
Renormalization scheme	$\overline{\mathrm{MS}}$, momentum subtraction, on shell	Changes the finite definition of renormalized parameters and fields	Translates parameters and coefficient functions together
Renormalization scale	$\mu$	Changes where the renormalized parameters are defined within a scheme	Evolves the parameters so complete predictions are $\mu$ independent
Physical scale	momentum transfer $Q$, a mass $M$, or a temperature $T$	Changes the physical question or kinematics	Retains the resulting physical dependence

The choice $\mu=Q$ is often convenient, but it does not make $\mu$ a physical momentum. Likewise, an effective theory’s breakdown scale is physical information about omitted degrees of freedom; it is not an auxiliary ultraviolet regulator.

One-loop scale cancellation

Let $F(Q)$ be a finite dimensionless physical quantity with no additional running masses or operator factors in the present example. Suppose its one-loop expansion in a scheme $\mathcal S$ is

$$F(Q) = g^2(\mu) +g^4(\mu) \left[ A\ln\frac{Q^2}{\mu^2}+C_{\mathcal S} \right] +O(g^6).$$

Define the beta function by differentiating at fixed bare data,

$$\beta(g) \equiv \left.\mu\frac{dg}{d\mu}\right|_0 =b g^3+O(g^5).$$

The RG equation for this quantity is

$$0 = \left. \mu\frac{dF}{d\mu} \right|_0 = \left( \mu\frac{\partial}{\partial\mu} +\beta(g)\frac{\partial}{\partial g} \right)F.$$

There are two contributions at order $g^4$. Running the tree term gives $2g\beta=2b g^4$, whereas differentiating the logarithm gives $A g^4(-2)$. Running $g^4$ inside the one-loop term first contributes at $O(g^6)$. Therefore

$$\left.\mu\frac{dF}{d\mu}\right|_0 =2(b-A)g^4+O(g^6),$$

so scale independence requires

$$A=b.$$

The same sign can be checked by solving the flow. Integrating $dg/d\ln\mu=b g^3$ gives

$$\frac{1}{g^2(\mu)} = \frac{1}{g^2(\mu_0)} -2b\ln\frac{\mu}{\mu_0}.$$

For a scale $Q$ within the perturbative domain,

$$g^2(Q) = g^2(\mu_0) +2b g^4(\mu_0)\ln\frac{Q}{\mu_0} +O(g^6).$$

Evaluating the fixed-order expression at $\mu=\mu_0$ and using $\ln(Q^2/\mu_0^2)=2\ln(Q/\mu_0)$ now gives

$$F(Q) =g^2(Q)+C_{\mathcal S}g^4(Q)+O(g^6),$$

which is exactly the result obtained by choosing $\mu=Q$ and re-expanding to the same order. This is scale cancellation, not exact scale independence of a truncated expression: the residual derivative begins at the first omitted order. The fixed-bare derivation and its extension to masses and anomalous dimensions are given in Collins 1984, §§7.3.1–7.3.3, pp. 180–185.

A finite scheme change and its inverse

The constant $C_{\mathcal S}$ is not generally invariant. Consider a finite, locally invertible redefinition

$$g'=g+c g^3+O(g^5).$$

Its perturbative inverse is

$$g=g'-c g'^3+O(g'^5).$$

Substitution into the prediction gives

$$\begin{aligned} F(Q) &=g'^2-2c g'^4 +g'^4\left[b\ln\frac{Q^2}{\mu^2}+C_{\mathcal S}\right] +O(g'^6) \\ &=g'^2 +g'^4\left[b\ln\frac{Q^2}{\mu^2}+C_{\mathcal S'}\right] +O(g'^6), \end{aligned}$$

where

$$C_{\mathcal S'}=C_{\mathcal S}-2c.$$

The inverse round trip is immediate: insert $g'=g+c g^3$ and $C_{\mathcal S'}=C_{\mathcal S}-2c$ into the primed expression. The $+2c g^4$ from $g'^2$ cancels the $-2c g^4$ in the finite coefficient, and the original expansion is recovered through $O(g^4)$.

The beta function transforms by the chain rule,

$$\beta'(g') =\frac{dg'}{dg}\,\beta(g) =b g'^3+O(g'^5).$$

Thus the leading coefficient $b$ is unchanged under this analytic map with unit linear term, while the finite coefficient moves. Higher-order claims require the full beta function and the precise class of schemes being compared. Couplings and coefficient functions are coordinates; their consistently combined prediction is the invariant target.

Running, fixed points, and dimensional transmutation

The beta function is a velocity on the space of dimensionless renormalized parameters. A fixed point $g_*$ satisfies $\beta(g_*)=0$. Near an isolated one-coupling fixed point,

$$\frac{d(g-g_*)}{d\ln\mu} \simeq \left. \frac{d\beta}{dg} \right|_{g_*}(g-g_*).$$

With $\ln\mu$ increasing toward the ultraviolet, a negative slope is ultraviolet-attractive and a positive slope is infrared-attractive. Under a regular exact scheme transformation, a zero maps to a zero and the linearized eigenvalues agree. A zero found only after truncating a perturbative beta function can move substantially or disappear when higher orders are included; it is evidence for investigation, not by itself a construction of a complete fixed-point theory. The physical interpretation of RG fixed points and scaling is reviewed in Wilson and Kogut 1974, §§2–4.

For an asymptotically free one-loop flow, write $b=-b_0$ with $b_0>0$. Then

$$\frac{1}{g^2(\mu)} = \frac{1}{g^2(\mu_0)} +2b_0\ln\frac{\mu}{\mu_0}.$$

The boundary value can equivalently be encoded in

$$\Lambda_{\mathcal S} = \mu\exp\left[-\frac{1}{2b_0g^2(\mu)}\right],$$

because

$$\frac{d\ln\Lambda_{\mathcal S}}{d\ln\mu} = 1+\frac{\beta(g)}{b_0g^3} =0$$

at this order. The exponential is dimensionless, so $\Lambda_{\mathcal S}$ has the same mass dimension as $\mu$. Replacing the boundary datum $g(\mu_0)$ by $\Lambda_{\mathcal S}$ is dimensional transmutation: within a specified scheme and perturbative order, a dimensionless coupling at a reference scale is traded for a dimensionful integration constant.

This statement has limits. The one-loop formula is controlled only while $g(\mu)$ is small, and $\mu\sim\Lambda_{\mathcal S}$ usually marks the loss of that control. It does not by itself prove confinement, a mass gap, or the value of any hadron mass. A finite scheme change can rescale the numerical value of $\Lambda_{\mathcal S}$ by a constant; matched physical predictions remain unchanged. For $b>0$, the same one-loop solution has a formal ultraviolet Landau scale. That singularity marks the failure of the truncated flow before it establishes the ultraviolet fate of the theory.

Large logarithms and the domain of the method

If $|b g^2\ln(Q/\mu)|$ is not small, powers of the logarithm can spoil the ordering of fixed-order perturbation theory even when $g$ itself is small. Choosing $\mu$ near $Q$ and evolving the coupling from a reference scale resums a class of logarithms. A reliable use of this procedure states the beta-function order, matching order, active degrees of freedom, and scale range. Residual $\mu$ variation can diagnose sensitivity to omitted terms, but it is not a probability distribution for the theoretical error.

Several common situations require more structure than the one-coupling example:

• masses, fields, and composite operators introduce anomalous dimensions;
• several couplings produce coupled beta functions and a stability matrix;
• heavy-particle thresholds require matching between theories with different active fields;
• several widely separated physical scales may require factorization rather than a single choice of $\mu$; and
• strong coupling can invalidate the perturbative beta function.

The QED and Yang–Mills theory lesson applies these ideas to gauge couplings and Ward identities. The effective field theory and matching lesson uses threshold matching and RG evolution to separate short- and long-distance physics.

Exercises

1. Why the quartic coupling needs a scale factor

In $d=4-2\epsilon$, determine the engineering dimensions of $\phi$ and the coefficient of $\phi^4$. Explain the factor $\mu^{2\epsilon}$ in the renormalized scalar Lagrangian.

Solution

The action is dimensionless, so $[\mathcal L]=d$ and the kinetic term gives

$$2+2[\phi]=d.$$

Hence

$$[\phi]=\frac{d-2}{2}=1-\epsilon.$$

The coefficient of $\phi^4$ must have dimension

$$d-4[\phi] =(4-2\epsilon)-4(1-\epsilon) =2\epsilon.$$

Defining $\lambda$ to remain dimensionless therefore requires $\lambda_0\propto\mu^{2\epsilon}\lambda$. Since $[\mu]=1$, the factor has dimension $2\epsilon$, exactly as required. A factor $\mu^\epsilon$ would have the wrong dimension for a quartic coupling, although it is appropriate for some couplings with a different continuation to $d$ dimensions.

2. Check explicit–implicit scale cancellation

For

$$F=g^2+g^4\left[A\ln\frac{Q^2}{\mu^2}+C\right]+O(g^6), \qquad \beta(g)=b g^3+O(g^5),$$

compute $\mu\,dF/d\mu$ at fixed bare data through $O(g^4)$. Then expand $g^2(Q)$ in terms of $g(\mu_0)$ and recover the logarithm in $F(Q)$.

Solution

The implicit derivative of the tree term is

$$\beta\frac{\partial g^2}{\partial g} =2g\beta =2b g^4+O(g^6).$$

The explicit derivative is

$$\mu\frac{\partial}{\partial\mu} \left(A g^4\ln\frac{Q^2}{\mu^2}\right) =-2A g^4+O(g^6).$$

Terms from differentiating $g^4$ are $O(g^6)$. Thus

$$\mu\frac{dF}{d\mu} =2(b-A)g^4+O(g^6),$$

and scale independence requires $A=b$.

The integrated flow gives

$$g^2(Q) =g^2(\mu_0) +2b g^4(\mu_0)\ln\frac{Q}{\mu_0} +O(g^6).$$

Therefore

$$g^2(Q)+C g^4(Q) =g^2(\mu_0) +g^4(\mu_0) \left[ b\ln\frac{Q^2}{\mu_0^2}+C \right] +O(g^6),$$

which recovers the fixed-order logarithm with the same sign.

3. Complete the scheme-change round trip

Let $g'=g+c g^3+O(g^5)$. Find the inverse map, transform the finite coefficient in $F$, and then substitute back to verify that the original expression is recovered through $O(g^4)$. Check the leading beta-function coefficient as well.

Solution

Assume $g=g'+a g'^3+O(g'^5)$. Substitution into the forward map gives $g'=g'+(a+c)g'^3+O(g'^5)$, so $a=-c$ and

$$g=g'-c g'^3+O(g'^5).$$

Consequently,

$$g^2=g'^2-2c g'^4+O(g'^6), \qquad g^4=g'^4+O(g'^6).$$

The primed finite coefficient is therefore $C'=C-2c$. Returning to the unprimed coordinate,

$$\begin{aligned} g'^2+g'^4(bL+C') &=g^2+2c g^4+g^4(bL+C-2c)+O(g^6) \\ &=g^2+g^4(bL+C)+O(g^6), \end{aligned}$$

where $L=\ln(Q^2/\mu^2)$. Finally,

$$\beta'(g') =\frac{dg'}{dg}\beta(g) =(1+3c g^2)b g^3+O(g^5) =b g'^3+O(g'^5).$$

The round trip preserves $F$ to the stated order and preserves the leading beta-function coefficient under this allowed map.

4. Construct and test a transmuted scale

For $\beta(g)=-b_0g^3$ with $b_0>0$, solve for $g(\mu)$, construct an RG invariant scale, check its mass dimension and derivative, and state where the one-loop expression ceases to be trustworthy.

Solution

Since

$$\frac{d}{d\ln\mu}\frac{1}{g^2} =-\frac{2\beta(g)}{g^3} =2b_0,$$

integration gives

$$\frac{1}{g^2(\mu)} =\frac{1}{g^2(\mu_0)} +2b_0\ln\frac{\mu}{\mu_0}.$$

Define

$$\Lambda =\mu\exp\left[-\frac{1}{2b_0g^2(\mu)}\right].$$

Its exponential is dimensionless, so $[\Lambda]=[\mu]=1$. Moreover,

$$\frac{d\ln\Lambda}{d\ln\mu} =1-\frac{1}{2b_0} \frac{d(g^{-2})}{d\ln\mu} =1-1 =0.$$

Equivalently,

$$\frac{1}{g^2(\mu)}=2b_0\ln\frac{\mu}{\Lambda}.$$

The weak-coupling solution requires $\mu\gg\Lambda$. As $\mu$ approaches $\Lambda$, the one-loop expression predicts strong coupling, so the perturbative derivation has reached its boundary. It cannot determine the infrared spectrum or prove a mass gap.

Check before continuing

Continue when you can:

• write a bare Lagrangian as renormalized terms plus local counterterms and name the condition that fixes each finite input;
• distinguish the regulator, scheme, renormalization scale, and physical kinematic scale in a calculation;
• show the sign-by-sign cancellation of explicit and implicit $\mu$ dependence through the calculated order;
• translate a finite coupling redefinition in both directions without changing the retained prediction; and
• solve a one-coupling flow while stating where its perturbative interpretation ends.

If the regulator pole still survives after combining diagrams, return to loops and regularization. If counterterm graphs or perturbative orders are unclear, return to perturbative expansion and Feynman rules.

References

• John C. Collins, Renormalization, Cambridge University Press, 1984, doi:10.1017/CBO9780511622656.
• Kenneth G. Wilson and John Kogut, “The Renormalization Group and the $\epsilon$ Expansion,” Physics Reports 12, no. 2 (1974), 75–199, doi:10.1016/0370-1573(74)90023-4.