QFT Regulator Families and Their Tradeoffs

A regulator should be chosen by the structures that must remain trustworthy during the calculation. No single family simultaneously makes locality, gauge symmetry, chiral algebra, unitarity, scale separation, and nonperturbative computation manifest. The correct choice is therefore conditional: declare the observable and identities, select a regulator whose breakings are controlled, state the removal or extrapolation limit, and compare a non-input prediction after matching the same renormalized inputs.

Regularization is not renormalization. A regulator makes an intermediate object well defined; a subtraction prescription and renormalization conditions define finite parameters. Two regulators can describe the same theory only after their inputs and finite parts are matched. Changing the finite prescription reparametrizes a theory when its input data are translated consistently; changing those matched input data can change the theory.

Required background. Ultraviolet Sensitivity and the Renormalization Problem supplies the bare-to-observable distinction and the meaning of a regulator-removal test.

Helpful background. Regulated Jacobians and Measure Variation is useful when a symmetry acts nontrivially on the functional measure or when a regulator exposes an anomaly.

Regulator choice is a constraint problem

Before choosing a regulator, answer five questions.

1. Which object is being defined? A perturbative amplitude, an effective action, a composite insertion, a Wilsonian mode integral, and a nonperturbative path integral impose different requirements.
2. Which identities must hold before removal? Lorentz symmetry, translation invariance, a Ward or Slavnov–Taylor identity, chiral symmetry, reflection positivity, supersymmetry, or a topological quantization condition may be central.
3. Which singular regions must remain distinguishable? A regulator useful for pure ultraviolet subtraction may obscure the separation of ultraviolet from soft, collinear, or rapidity singularities.
4. What limit is actually available? One may take $\epsilon\to0$, $\Lambda\to\infty$, regulator masses $M_i\to\infty$, or lattice spacing $a\to0$; an EFT may instead retain a finite cutoff and test stability within its truncation error.
5. What independent check survives the method? A second regulator, a symmetry identity, a known analytic limit, or a non-input observable is needed to distinguish a controlled calculation from a regulator artifact.

The regulator must also be specified completely. “A cutoff” does not say whether it is sharp or smooth, Euclidean or spatial, imposed on loop momentum or an operator spectrum, covariantized or not, and applied before or after algebraic manipulations. Those choices can change finite local terms and which identities are manifest.

Decision table

The table gives default expectations, not universal theorems. A carefully engineered version of a family can preserve more structure than its simplest implementation, and an inconsistent implementation can preserve less.

Family	Regulated object and new scale	Usually manifest	Main costs or possible breakings	Removal or validation evidence	Best fit
Sharp momentum cutoff	Restrict $\lvert k\rvert<\Lambda$ or a momentum shell	Concrete UV domain; power sensitivity; Wilsonian scale separation	Loop-momentum shifts change the boundary; gauge and translation identities can acquire surface terms; a spatial cutoff also breaks Lorentz symmetry	Match finite inputs, include all allowed restoration terms, and show stability as $\Lambda$ grows or within the EFT window	Pedagogical integrals, Wilsonian shell arguments, cutoff EFTs with explicit scale hierarchy
Smooth momentum cutoff	Multiply propagators by $K(k^2/\Lambda^2)$ or add a smooth quadratic kernel	Differentiable flow; quasi-local expansion for suitable $K$; numerical stability	Results depend on kernel shape at finite truncation; naive kernels can break gauge or other symmetries	Vary admissible shapes, enlarge the truncation, test identities, and verify universal output	Wilsonian and functional RG, numerical continuum integrals
Pauli–Villars	Add weighted heavy auxiliary propagators with masses $M_i$	Lorentz covariance; explicit mass scale; improved large-$k$ decay	Auxiliary negative-metric or wrong-statistics fields are not physical; non-Abelian and chiral implementations are delicate; inconsistent factorization can break gauge identities	Satisfy cancellation moments, regulate the complete expression consistently, take $M_i\to\infty$, and check Ward identities	One-loop continuum calculations, anomaly diagnostics, covariant checks
Higher covariant derivatives	Add terms such as $D^2/\Lambda^2$ to kinetic operators	Locality; potentially gauge covariance when built from covariant derivatives; strong UV falloff	Extra finite-$\Lambda$ poles; residual lower-loop divergences may remain; unitarity is not manifest before removal	Remove $\Lambda$ after supplementary subtractions, verify identities, and show extra poles decouple	Symmetry-aware perturbative regularization and structural proofs
Dimensional regularization	Continue loop momenta and tensor algebra to $d=4-2\epsilon$	Lorentz covariance, translation invariance, loop-momentum shifts, and many gauge identities; efficient analytic continuation	No explicit mode cutoff; power divergences are represented analytically; scaleless integrals vanish while mixing UV/IR information; $\gamma_5$, Levi-Civita tensors, and intrinsically integer-dimensional objects require prescriptions	Label pole origins, state the continuation of all algebraic objects, subtract in a named scheme, and take $\epsilon\to0$	Perturbative relativistic amplitudes and mass-independent RG calculations
Analytic, zeta, or proper-time regularization	Continue propagator powers or spectral/proper-time integrals with a complex parameter or lower-time cutoff	Often covariant and efficient for determinants, heat kernels, and backgrounds	Multiplicative properties, phases, zero modes, boundaries, and anomalies need separate control; not every graph is regulated uniformly	State the analytic continuation and spectral assumptions, compare local coefficients, and test phase/zero-mode contributions	One-loop effective actions, curved backgrounds, spectral problems
Lattice regulator	Replace spacetime by sites/links with spacing $a$ and finite volume	Nonperturbative finite-dimensional definition; exact compact gauge invariance for link formulations; numerical sampling	Continuous translations and rotations are reduced; fermion doubling and chiral symmetry require special treatment; finite volume and discretization errors coexist	Tune required bare parameters, extrapolate $a\to0$ and volume to infinity, test restoration of continuum symmetries	Nonperturbative gauge theory and statistical field theory
Point splitting or position-space extension	Separate coincident points by a vector or extend singular distributions locally	Local support and short-distance geometry are explicit; can be covariantized with parallel transport	Direction or path dependence; gauge covariance requires Wilson lines; overlapping products and finite local ambiguities still require a prescription	Shrink the separation after local subtractions, average or covariantize as required, and compare extension choices	Currents, anomalies, curved backgrounds, local composite products

Dimensional regularization was designed to retain the shift and covariance properties needed for non-Abelian gauge calculations; in anomaly-free cases, the method supports the Ward-identity structure when the continued algebra is handled consistently ’t Hooft and Veltman 1972, § 3, p. 197. This advantage does not solve every chiral problem: a four-dimensional $\gamma_5$ cannot simply retain all of its familiar anticommutation and trace identities in general $d$ ’t Hooft and Veltman 1972, § 6, pp. 206–208.

The lattice provides a complementary tradeoff. Wilson’s link-variable construction preserves exact lattice gauge invariance and supplies a nonperturbative strong-coupling definition, while continuous Euclidean symmetry is recovered only in a controlled continuum limit Wilson 1974, pp. 2445–2459. Detailed actions, algorithms, finite-volume analysis, and continuum fits belong to Lattice and Hamiltonian Field Theory, not to this regulator comparison.

Three regulators on the same scalar loop

Take $m>0$ and real Euclidean external momenta $p,p_*$, with $p^2,p_*^2\geq0$ fixed as the regulator is removed. Write

$$\Delta_m(x,p^2)=m^2+x(1-x)p^2>0, \qquad 0\leq x\leq1.$$

The formal logarithmically divergent integral is

$$B(p^2;m^2) = \int\frac{d^4k}{(2\pi)^4} \frac{1}{(k^2+m^2)((k+p)^2+m^2)}.$$

This dimensionless loop is a building block: it occurs in a cubic scalar self-energy, with coupling and symmetry factors, and in channels of a quartic scalar four-point function. It is not by itself a scattering observable. The Euclidean subtraction and continuation of the self-energy example are developed in Collins 1984/2023, § 3.1, pp. 38–44.

For each regulator $\rho$, first define a finite $B_\rho$, and then impose the same momentum-subtraction condition:

$$\begin{aligned} B_{\rho,\mathrm R}(p^2;p_*^2) &=B_\rho(p^2)-B_\rho(p_*^2),\\ B_{\rho,\mathrm R}(p_*^2;p_*^2)&=0. \end{aligned}$$

Here $\rho$ is $\Lambda$, $M$, or $\epsilon$ for the three constructions below. The subtracted functions need not coincide before the respective removal limits.

Hard cutoff

Use one specified routing and one spherical domain:

$$B_\Lambda(p^2) = \int_{\lvert k\rvert<\Lambda} \frac{d^4k}{(2\pi)^4} \frac{1}{(k^2+m^2)((k+p)^2+m^2)}.$$

Both terms in $B_{\Lambda,\mathrm R}$ use that same $k$-centered domain. A direct radial check at $p=0$ gives

$$B_\Lambda(0) = \frac{1}{16\pi^2} \left[ \ln\!\left(1+\frac{\Lambda^2}{m^2}\right) -\frac{\Lambda^2}{\Lambda^2+m^2} \right].$$

For fixed $p,m$ at large $\Lambda$,

$$\begin{aligned} B_\Lambda(p^2) ={}&\frac{1}{16\pi^2} \left[ \ln\frac{\Lambda^2}{m^2}-1 -\int_0^1dx\,\ln\frac{\Delta_m(x,p^2)}{m^2} \right]\\ &+\mathcal O\!\left(\frac{m^2+p^2}{\Lambda^2}\right). \end{aligned}$$

The constant $-1$ belongs to this spherical prescription; other cutoff shapes can change local finite terms. A shift $k\mapsto k+xp$ moves the finite integration domain and must move its boundary too. It is therefore not a way to turn the raw cutoff bubble into an unrestricted radial integral. Instead, take the difference at $p$ and $p_*$ first: its large-$k$ leading term cancels and the difference is integrable on all of $\mathbb R^4$. Only after taking its $\Lambda\to\infty$ limit may unrestricted shifts be used.

The precise finite-cutoff remainder depends on that domain choice. Integrating the angles of the original sphere, without shifting $k$, gives

$$\begin{aligned} A(r,p^2)&=r^2+p^2+m^2,\\ B_\Lambda(p^2) &=\frac{1}{4\pi^2}\int_0^\Lambda dr\, \frac{r^3} {(r^2+m^2)\bigl[A(r,p^2)+\sqrt{A(r,p^2)^2-4r^2p^2}\bigr]}. \end{aligned}$$

Expanding its large-$r$ tail shows that the entire $1/\Lambda^2$ term is independent of $p$ and cancels in the MOM difference. For this construction,

$$\begin{aligned} B_{\Lambda,\mathrm R}(p^2;p_*^2)-B_{\mathrm R}(p^2;p_*^2) ={}&\frac{m^2(p^2-p_*^2)}{32\pi^2\Lambda^4}\\ &+\mathcal O\!\left( \frac{m^2\lvert p^2-p_*^2\rvert(m^2+p^2+p_*^2)} {\Lambda^6}\right). \end{aligned}$$

Here $B_{\mathrm R}$ is the common removal limit obtained below. Imposing a new sphere after a Feynman-parameter shift defines a different finite regulator and generally leaves a $1/\Lambda^2$ MOM residual. Neither rate should be transferred to another cutoff without checking its definition.

For a more divergent tensor integral, routing changes can leave local surface terms even in a removal limit. The appropriate symmetry identity must be checked with the complete expression and allowed counterterms.

Pauli–Villars masses

A basic Pauli–Villars propagator replacement is

$$\frac{1}{k^2+m^2} \longmapsto \sum_j c_j\frac{1}{k^2+M_j^2}, \qquad M_0=m,\quad c_0=1.$$

The coefficients and auxiliary masses cancel successive terms of the large-$k$ expansion. Depending on the degree of divergence, one needs conditions such as

$$\sum_jc_j=0, \qquad \sum_jc_jM_j^2=0.$$

For example, $1/(k^2+m^2)-1/(k^2+M^2)$ falls as $k^{-4}$. These cancellation conditions do not specify a complete loop prescription. In their vacuum-polarization application, Pauli and Villars require regulating the whole expression rather than factorizing its singular products Pauli and Villars 1949, § 1, pp. 434–435.

For the present scalar benchmark, choose the following complete equal-mass-bubble subtraction, with one auxiliary mass $M>m$:

$$\begin{aligned} B_{\mathrm{PV}}(p^2;m^2,M^2) =\int\frac{d^4k}{(2\pi)^4}\Bigg[ &\frac{1}{(k^2+m^2)((k+p)^2+m^2)}\\ -&\frac{1}{(k^2+M^2)((k+p)^2+M^2)} \Bigg]. \end{aligned}$$

The combined integrand falls as $k^{-6}$ at fixed $m,M,p$, so this is an absolutely convergent integral on $\mathbb R^4$. It is not the prescription obtained by replacing each internal propagator independently; that would produce mixed-mass terms. Nor does this scalar example establish a gauge-invariant prescription for another theory.

Combine the two terms before using Feynman parameters and shifting the unrestricted momentum. With $\Delta_M(x,p^2)=M^2+x(1-x)p^2$, the radial difference gives

$$B_{\mathrm{PV}}(p^2;m^2,M^2) = \frac{1}{16\pi^2} \int_0^1dx\,\ln\frac{\Delta_M(x,p^2)}{\Delta_m(x,p^2)}.$$

Its MOM-subtracted value is therefore

$$\begin{aligned} B_{M,\mathrm R}(p^2;p_*^2) ={}&-\frac{1}{16\pi^2} \int_0^1dx\,\ln\frac{\Delta_m(x,p^2)}{\Delta_m(x,p_*^2)} +r_M(p^2;p_*^2),\\ r_M(p^2;p_*^2) ={}&\frac{1}{16\pi^2} \int_0^1dx\,\ln\frac{\Delta_M(x,p^2)}{\Delta_M(x,p_*^2)}. \end{aligned}$$

The heavy-mass term is finite and generally nonzero. Since both Euclidean invariants are nonnegative, the mean-value theorem gives the useful bound

$$\lvert r_M(p^2;p_*^2)\rvert \leq \frac{\lvert p^2-p_*^2\rvert}{96\pi^2M^2}.$$

Thus $M\to\infty$ at fixed $m,p,p_*$ removes this artifact. Extra cancellation moments would be needed for more divergent objects; they are not a reason to change this explicitly defined scalar subtraction.

Dimensional continuation

Keep a fixed reference mass $\mu>0$, continue to $d=4-2\epsilon$, and define

$$B_\epsilon(p^2) = \mu^{2\epsilon} \int\frac{d^dk}{(2\pi)^d} \frac{1}{(k^2+m^2)((k+p)^2+m^2)}.$$

The Euclidean radial representation initially converges for $0<\operatorname{Re}d<4$. Its continuation, with the same measure normalization for both subtraction points, is

$$B_\epsilon(p^2) = \frac{\mu^{2\epsilon}\Gamma(\epsilon)} {(4\pi)^{2-\epsilon}} \int_0^1dx\,[\Delta_m(x,p^2)]^{-\epsilon}.$$

Expanding near $\epsilon=0$ gives

$$B_\epsilon(p^2) = \frac{1}{16\pi^2} \left[ \frac{1}{\bar\epsilon} -\int_0^1dx\,\ln\frac{\Delta_m(x,p^2)}{\mu^2} \right] +\mathcal O(\epsilon), \qquad \frac{1}{\bar\epsilon}=\frac{1}{\epsilon}-\gamma_E+\ln4\pi.$$

The pole cancels in $B_{\epsilon,\mathrm R}$ before $\epsilon\to0$. Minimal subtraction and modified minimal subtraction instead remove their named local pole terms; comparison with this MOM condition requires the corresponding finite translation. Collins derives the parameter integral and explains the scale and subtraction choices in Collins 1984/2023, §§ 3.5–3.6.1, pp. 53–57.

Dimensional continuation has no fixed sharp boundary: its shift identities belong to the analytically continued definition. For this massive example there is no infrared singularity, including at $p=0$. A massless scaleless integral is different: its vanishing can conceal cancellation between ultraviolet and infrared contributions, so it cannot replace the present massive test.

Match before comparing

The common function is the regulator-removal limit:

$$\begin{aligned} B_{\mathrm R}(p^2;p_*^2) &=\lim_{\Lambda\to\infty}B_{\Lambda,\mathrm R}(p^2;p_*^2)\\ &=\lim_{M\to\infty}B_{M,\mathrm R}(p^2;p_*^2)\\ &=\lim_{\epsilon\to0}B_{\epsilon,\mathrm R}(p^2;p_*^2)\\ &=-\frac{1}{16\pi^2} \int_0^1dx\,\ln\frac{\Delta_m(x,p^2)}{\Delta_m(x,p_*^2)}. \end{aligned}$$

The cutoff logarithm, heavy-mass logarithm and dimensional pole have different local terms. The identical subtraction condition removes those ambiguities, while finite-regulator residuals remain until removal. A change of finite prescription is a change of parameter coordinates when the same physical inputs are translated consistently Collins 1984/2023, § 3.4.1, pp. 50–52.

A non-input derivative detects an incorrect finite-mass equality. At $p_*^2=0$,

$$\begin{aligned} \left.\frac{\partial B_{\mathrm R}(p^2;0)}{\partial p^2}\right|_{p^2=0} &=-\frac{1}{96\pi^2m^2},\\ \left.\frac{\partial B_{M,\mathrm R}(p^2;0)}{\partial p^2}\right|_{p^2=0} &=-\frac{1}{96\pi^2} \left(\frac{1}{m^2}-\frac{1}{M^2}\right). \end{aligned}$$

The common input is exactly satisfied in both cases, but the derivatives differ by $1/(96\pi^2M^2)$. When the loop is used in an amplitude, its coupling and symmetry factors, local counterterms, analytic continuation and external normalization must also agree. Uncomputed higher-loop terms are a separate perturbative remainder; removing $\Lambda$, $M$ or $\epsilon$ does not remove that remainder. Regulator Removal and Renormalized Predictions develops this distinction for a complete scalar four-point coefficient.

Prediction-validation table

Compare each column at the same mass, subtraction point and other finite inputs. The table specializes the common validation questions to the constructions just defined; another regulated object needs its own domain and identity tests.

One massive Euclidean scalar loop with the same momentum-subtraction condition. The non-input function agrees after regulator removal; the finite-regulator residuals remain distinct.

Comparison	Hard cutoff	Pauli–Villars	Dimensional regularization
Regulated object	$B_\Lambda$: spherical $\lvert k\rvert<\Lambda$, fixed routing, $m>0$.	$B_{\mathrm{PV}}$: complete equal-mass bubbles subtracted on $\mathbb R^4$, $M>m>0$.	$B_\epsilon$: measure $\mu^{2\epsilon}d^dk/(2\pi)^d$, $d=4-2\epsilon$, $m>0$.
Auxiliary removal	$\Lambda\to\infty$ at fixed $m,p,p_*$.	$M\to\infty$ at fixed $m,p,p_*$.	$\epsilon\to0$ after subtraction, at fixed $m,p,p_*,\mu$.
Subtraction convention	$B_{\Lambda,\mathrm R}=B_\Lambda(p^2)-B_\Lambda(p_*^2)$ with the same domain.	$B_{M,\mathrm R}=B_{\mathrm{PV}}(p^2)-B_{\mathrm{PV}}(p_*^2)$ with the same auxiliary mass.	$B_{\epsilon,\mathrm R}=B_\epsilon(p^2)-B_\epsilon(p_*^2)$; translate MS or modified MS if used instead.
Symmetry identities	Euclidean rotations manifest; shifts move the boundary. Check any required Ward identity with allowed local restoration.	Rotations and momentum shifts hold for the combined scalar integral. A gauge application requires its own complete identity check.	Rotations and shifts hold in the continued definition. Specify the continuation of chiral or intrinsically integer-dimensional algebra.
Renormalized input	$B_{\Lambda,\mathrm R}(p_*^2;p_*^2)=0$.	$B_{M,\mathrm R}(p_*^2;p_*^2)=0$.	$B_{\epsilon,\mathrm R}(p_*^2;p_*^2)=0$.
Finite local parts	The spherical-cutoff constant $-1$ cancels in the subtraction.	The momentum-independent heavy-mass term cancels in the subtraction.	The pole and convention-dependent constants cancel; an MS-to-MOM finite map must be stated.
Non-input prediction	$B_{\mathrm R}(p^2;p_*^2)$ after $\Lambda\to\infty$.	The same $B_{\mathrm R}(p^2;p_*^2)$ after $M\to\infty$.	The same $B_{\mathrm R}(p^2;p_*^2)$ after $\epsilon\to0$.
Residual artifact	Leading term $m^2(p^2-p_*^2)/(32\pi^2\Lambda^4)$ for this fixed sphere, with the higher-order remainder given above.	The explicit $r_M$, bounded by $\lvert p^2-p_*^2\rvert/(96\pi^2M^2)$.	$\mathcal O(\epsilon)$ in the subtracted bubble before removal.
Independent check	Check the zero-momentum radial integral and the subtracted removal limit; retain boundary terms when testing another routing.	Vary the heavy mass and test the non-input derivative and its inverse-mass-squared residual.	Compare the finite MOM limit with the two four-dimensional constructions; distinguish ultraviolet from infrared poles.

The loop comparison separates regulator, subtraction convention, symmetry status, renormalized input, finite local choices and removal behavior. A perturbative amplitude using the loop must additionally state its retained order and omitted-order uncertainty. Regulator variation and finite-scheme variation are diagnostics, not automatic probability distributions.

Identity restoration and anomalies

For perturbative local Ward or BRST identities, when the regulated breaking satisfies the locality and consistency hypotheses, distinguish three outcomes:

1. The breaking vanishes with the regulator after ordinary counterterms are fixed.
2. The breaking is local and belongs to the trivial consistency class, so an allowed finite local counterterm restores the identity.
3. The breaking satisfies the consistency condition but lies in a nontrivial class; it is an anomaly and cannot be removed while preserving all assumptions.

One must not jump from “the regulator breaks the identity” to either “the theory is anomalous” or “the regulator is unusable.” The relevant questions are locality, quantum numbers, consistency, and available counterterms. The counterterm analysis belongs to Symmetry Constraints and the Space of Counterterms; the definition and matching of anomalies belong to Anomalies, Inflow, and Matching.

Dimensional regularization deserves a precise qualification. It often preserves gauge identities efficiently, but chirality and evanescent tensor structures can force an enlarged operator basis and finite restoration. A hard cutoff deserves the complementary qualification: it can make physical scale sensitivity transparent, but a raw $\Lambda^2$ term is not itself an observable or a proof of naturalness. Only a matched threshold correction or other invariant relation can carry that interpretation.

Common pitfalls

Calling the regulator a physical cutoff. An auxiliary $\Lambda$ sent to infinity is not automatically the EFT breakdown scale or the mass of new physics. State whether the scale is removed, retained as part of a Wilsonian definition, or identified with a physical threshold by matching.

Using a shifted hard-cutoff integral without a surface-term check. A finite integration domain is not invariant under $k\mapsto k+q$. Fix routing and test the identity that would normally justify the shift.

Interpreting a scaleless zero as absence of UV structure. Dimensional regularization can set a scaleless integral to zero through UV/IR cancellation. Introduce an infrared scale or label poles by regions before drawing a UV conclusion.

Treating auxiliary Pauli–Villars fields as states. Their wrong-sign or wrong-statistics contributions are part of the regulator and leave in the $M_j\to\infty$ limit. They are not a proposed physical spectrum.

Assuming exact lattice gauge symmetry implies exact continuum symmetry. Link gauge invariance can be exact at finite $a$, while rotations, translations, and chiral properties still require improvement, special formulations, tuning, and continuum tests.

Comparing finite answers before matching inputs. Regulator-dependent local constants encode different parameter coordinates. Apply the same renormalization condition or an explicit finite map before comparing a non-input observable.

Exercises

1. Shift test. Why can

$$\int_{|k|<\Lambda}d^4k\,f(k+q)$$

differ from the same integral with $q=0$ even after a change of variables?

Solution

The change of variables shifts the integration domain from a sphere centered at the origin to one centered at $q$. Their difference is a boundary or surface term. For sufficiently convergent integrals it vanishes as $\Lambda\to\infty$; for divergent tensor integrals it can leave a local term that matters to a Ward identity.

2. Pauli–Villars falloff. Show that

$$\frac{1}{k^2+m^2}-\frac{1}{k^2+M^2} = \frac{M^2-m^2}{(k^2+m^2)(k^2+M^2)}$$

falls as $k^{-4}$.

Solution

The numerator is independent of $k$, while each denominator factor grows as $k^2$. Their product grows as $k^4$, improving the original $k^{-2}$ propagator by two powers. Higher superficial degrees require more auxiliary terms and moment conditions.

3. Choose a family. You need a nonperturbative definition of a compact gauge theory and can afford a numerical continuum extrapolation. Which row is the natural start, and what evidence is still missing?

Solution

A gauge-invariant lattice link formulation is the natural start. Exact finite-$a$ gauge invariance is not enough: the calculation still needs finite-volume control, tuning or improvement where required, an $a\to0$ extrapolation, restoration of continuum rotational symmetry, and a checked renormalization of the target observable.

Selection rule

Choose the least intrusive regulator that makes the target object well defined and leaves the decisive identities either manifest or locally restorable. Then document

$$\text{regulator} + \text{subtraction} + \text{input conditions} + \text{identity checks} + \text{removal evidence}.$$

No regulator earns trust by name alone. Trust comes from a matched invariant output and a falsifiable artifact estimate.

Continue to Dimensional Regularization and Minimal Subtraction for the detailed continuation and MS/$\overline{\rm MS}$ conventions. Continue to Local Counterterms and Subdivergence Structure when locality, rather than regulator choice, is the central question. Return to the chapter overview for a route using a different scheme.

References

• Collins, John C. 1984; open-access reissue 2023. Renormalization: An Introduction to Renormalization, the Renormalization Group and the Operator-Product Expansion. Cambridge Monographs on Mathematical Physics. Cambridge University Press. DOI. Open PDF.

• Pauli, Wolfgang, and Felix Villars. 1949. “On the Invariant Regularization in Relativistic Quantum Theory.” Reviews of Modern Physics 21: 434–444. DOI.

• ’t Hooft, Gerard, and Martinus Veltman. 1972. “Regularization and Renormalization of Gauge Fields.” Nuclear Physics B 44: 189–213. DOI.

• Wilson, Kenneth G. 1974. “Confinement of Quarks.” Physical Review D 10: 2445–2459. DOI.