Effective field theory and matching

An effective field theory (EFT) describes a chosen range of energies without pretending to resolve shorter-distance physics. It is predictive when its degrees of freedom, symmetries, expansion parameters, operator basis, and breakdown conditions are stated together. The result is not merely a convenient approximation: it is an ordered calculation whose first omitted terms can be identified and tested.

Required background. Use perturbative rules to translate local operators into amplitudes, loops and regularization to separate momentum regions, and renormalization and the RG to distinguish matching from scale evolution. Helpful background. The complex and asymptotic methods review is useful when an expansion is being mistaken for an exact identity.

Why a low-energy theory can predict

Suppose a process probes momenta and light masses of a common scale $Q$, while a particle or excitation of mass $M$ is not produced and $Q\ll M$. Its virtual effects can often be expanded in powers of

$$\epsilon=\frac{Q}{M}.$$

At low energy, short-distance propagation then appears as a sequence of local interactions with increasing numbers of derivatives. The EFT keeps the light fields explicitly and represents the removed physics through Wilson coefficients:

$$\mathcal L_{\mathrm{EFT}} =\mathcal L_{\mathrm{light}} +\sum_{d,i}\frac{C_i^{(d)}(\mu)}{M^{d-4}}\,\mathcal O_i^{(d)}(\mu).$$

The index $i$ distinguishes operators of the same dimension; $\mu$ is the renormalization scale. This notation is only a starting point. Canonical dimension alone does not determine importance when a problem contains small couplings, nonrelativistic velocities, symmetry breaking, large logarithms, or several low scales. A usable EFT needs a power counting that assigns an order to every derivative, mass insertion, coupling, loop, and operator insertion.

The locality expansion has a boundary. A heavy pole, a threshold for a removed state, or the loss of the hierarchy $Q/M\ll1$ cannot be repaired by adding a few more local terms. The domain must therefore be specified before the Lagrangian is truncated. The local action reproduces the declared low-energy orders of the full calculation; it does not reproduce a removed pole at any finite derivative order Burgess 2021, § 1.2, pp. 10–13.

Worked matching: remove a heavy scalar

Consider a light real scalar $\phi$ and a heavy real scalar $H$ in four spacetime dimensions. Use a complete tree-level model,

$$\mathcal L =\frac12(\partial\phi)^2-\frac12m^2\phi^2 +\frac12(\partial H)^2-\frac12M^2H^2 -\frac{g}{2}H\phi^2-\frac{\lambda_{\mathrm{full}}}{4!}\phi^4, \qquad m,Q\ll M.$$

where $m^2,M^2>0$, $g$ is real with mass dimension one, and $\lambda_{\mathrm{full}}>3g^2/M^2$. The last condition is a sufficient classical stability check:

$$\begin{aligned} U(\phi,H) ={}&\frac{M^2}{2}\left(H+\frac{g\phi^2}{2M^2}\right)^2 +\frac{m^2\phi^2}{2}\\ &+\frac{\lambda_{\mathrm{full}}-3g^2/M^2}{24}\phi^4. \end{aligned}$$

Every term is nonnegative and the minimum is at $\phi=H=0$. Without the light quartic, minimizing over $H$ leaves a negative quartic in $\phi$ and an unbounded potential. Thus the stabilizing contact interaction must be retained in both scattering calculations. Take $\lambda_{\mathrm{full}}/(16\pi^2)$ and $g^2/(16\pi^2M^2)$ small when using perturbation theory. This tree calculation does not prove quantum continuum existence or include loop threshold corrections. $H$ is an internal mediator and need not be a stable asymptotic particle.

After integrating the heavy kinetic term by parts, its contribution is

$$\mathcal L_H =-\frac12H(\Box+M^2)H-\frac{g}{2}H\phi^2.$$

The classical heavy-field equation is

$$(\Box+M^2)H=-\frac{g}{2}\phi^2,$$

so substituting its solution with the scattering boundary prescription gives the exact tree-level nonlocal interaction

$$\Delta\mathcal L_{\mathrm{tree}} =\frac{g^2}{8}\, \phi^2\frac{1}{\Box+M^2}\phi^2.$$

The inverse has the Feynman boundary prescription; its momentum denominator is $M^2-q^2-i0$. For field configurations whose exchanged invariants obey $|q^2|<M^2$,

$$\frac{1}{\Box+M^2} =\frac{1}{M^2} \left(1-\frac{\Box}{M^2} +\frac{\Box^2}{M^4}-\cdots\right),$$

and hence

$$\Delta\mathcal L_{\mathrm{EFT}} =\frac{g^2}{8M^2}\phi^4 -\frac{g^2}{8M^4}\phi^2\Box(\phi^2) +O\!\left(\frac{g^2Q^4}{M^6}\phi^4\right).$$

On a momentum mode, $\Box\to-q^2$, so this reproduces the geometric expansion

$$\frac{g^2}{M^2-q^2} =\frac{g^2}{M^2} \left[1+\frac{q^2}{M^2} +\left(\frac{q^2}{M^2}\right)^2+\cdots\right].$$

This is tree-level matching: coefficients of local EFT operators are chosen so that low-energy amplitudes in the full and effective theories agree to a declared order. The matching calculation determines the coefficients; the power counting determines which coefficients are needed.

The local quartic includes both the retained contact and the heavy-induced term. In the unreduced operator basis above, its leading matched coupling is

$$\lambda=\lambda_{\mathrm{full}}-\frac{3g^2}{M^2}>0.$$

The factor three follows from converting $g^2\phi^4/(8M^2)$ to the normalization $-\lambda\phi^4/4!$. For identical light scalars, the full on-shell tree amplitude contains all three exchanges:

$$\begin{aligned} \mathcal M_{\mathrm{full}} &=-\lambda_{\mathrm{full}}+A_{\mathrm{ex}},& A_{\mathrm{ex}}&=g^2\sum_{x\in\{s,t,u\}}\frac1{M^2-x},\\ \mathcal M_{\mathrm{EFT}} &=-\lambda+\frac{g^2}{M^4}(s+t+u) =-\lambda+\frac{4g^2m^2}{M^4}. \end{aligned}$$

Here the EFT keeps the first derivative term. The vertex product $(-ig)^2i/(x-M^2+i0)$ gives the sign of each exchange. The identity $s+t+u=4m^2$ makes this derivative contribution constant on the elastic mass shell; it does not make the off-shell operator vanish. Compare this massive, first-order expansion with the massless example retaining quadratic invariant terms in the controlled-expansion treatment; the scalar capstone carries this massive example through a numerical comparison.

The example also exposes breakdown cleanly. The local series cannot reproduce the pole at $q^2=M^2$ at any finite order. Near that pole, $H$ must be restored as an active degree of freedom, or a different description must be used.

Choose an operator basis

All local operators allowed by the retained symmetries should be considered through the requested order. The word “all” is qualified by equivalences:

• integration by parts moves derivatives between fields without changing the action when the boundary term vanishes;
• algebraic identities can relate apparently different contractions;
• the leading equations of motion remove operators that differ by perturbative field redefinitions; and
• flavor, discrete-symmetry, and gauge identities further reduce the list.

These choices define an operator basis. Coefficients quoted in different bases are not directly comparable. A field redefinition can change off-shell Green functions and individual Wilson coefficients while leaving consistently computed on-shell observables unchanged, provided the induced terms are retained through the requested order Burgess 2021, § 2.5, pp. 45–46. The local-operator overview defines composite insertions, while the operator-basis chapter develops operator reduction and equivalence.

Power counting is separate from basis reduction. In the heavy-scalar example, every extra pair of derivatives costs $Q^2/M^2$ relative to the previous term. In a gauge theory or many-body system, symmetry and kinematics may force a different order. The test is operational: at a fixed order, the counting must produce a finite list of terms, loops built from lower-order vertices must be absorbed by operators allowed at that or higher order, and successive predictions should improve while the declared expansion parameters remain small Burgess 2021, § 2.4, pp. 39–44.

Match at the high scale and run to the low scale

Matching and running answer different questions:

1. Matching compares the full and effective theories at a scale $\mu_M$ usually chosen near $M$. It fixes the short-distance coefficients.
2. Running evolves those coefficients between scales within the EFT. It sums logarithms and compensates the scale dependence of renormalized operators.
3. Low-energy calculation combines the evolved coefficients with EFT matrix elements at a scale suited to the observable.

If the renormalized operators obey

$$\mu\frac{\mathrm d}{\mathrm d\mu}\mathcal O_i =-\gamma_{ij}\mathcal O_j,$$

then scale independence of $\sum_iC_i\mathcal O_i$ requires

$$\mu\frac{\mathrm d}{\mathrm d\mu}C_i =\gamma_{ji}C_j.$$

Operator mixing therefore makes the Wilson coefficients a coupled vector rather than a collection of independent numbers. A physical amplitude has the schematic form

$$\mathcal A(Q) =\sum_i C_i(\mu)\, \langle f\lvert\mathcal O_i(\mu)\rvert i\rangle,$$

and its $\mu$ dependence cancels through the calculated order. Residual scale dependence is a diagnostic of omitted terms, not a physical dependence on an arbitrary scale.

At loop level, compare the same renormalized object in both theories through the same order, with compatible field normalizations, regulators, gauges, subtraction schemes, and infrared prescriptions. The EFT’s retained light-field loops must be included before the common low-momentum terms cancel in the matching difference; the remaining short-distance terms fix the Wilson coefficients. This separates the Wilson action from the complete low-energy calculation Burgess 2021, §§ 2.2–2.3, pp. 26–39. A scaleless EFT integral that vanishes in dimensional regularization does not by itself imply a zero matching coefficient. Matching beyond tree level develops the aligned subtraction and its infrared conditions.

Estimate omissions without overclaiming

Assume an observable has an expansion

$$X=X_0\sum_{n=0}^{\infty}c_n\epsilon^n, \qquad \lvert\epsilon\rvert<1.$$

After retaining terms through order $k$, a first truncation estimate is the size expected for $X_0c_{k+1}\epsilon^{k+1}$. This statement needs a coefficient model: “$c_{k+1}$ is of order one,” a bound inferred from symmetry, or a distribution calibrated from a relevant class of calculations. Without such a model, the first omitted power supplies an order estimate, not a confidence interval.

One heavy-exchange channel gives an unusually transparent check. Let $x=q^2/M^2$ and truncate $1/(1-x)$ after $x^N$. The exact relative error of that exchange contribution is

$$\frac{\left\lvert(1-x)^{-1}-\sum_{n=0}^{N}x^n\right\rvert} {\left\lvert(1-x)^{-1}\right\rvert} =\lvert x\rvert^{N+1}.$$

For $x=0.1$, the leading and next-to-leading relative errors are $10\%$ and $1\%$. For $x=0.5$, they are $50\%$ and $25\%$. The exact model confirms the expected next-power scaling and also shows why the expansion becomes unhelpful near the pole.

This is not automatically the relative error of a full amplitude. In the three-channel scalar calculation above, the common contact cancels in the absolute remainder:

$$\begin{aligned} R&=\mathcal M_{\mathrm{full}}-\mathcal M_{\mathrm{EFT}} =\frac{g^2}{M^4}\sum_{x\in\{s,t,u\}}\frac{x^2}{M^2-x},\\ |R|&\le \frac{g^2(s^2+t^2+u^2)}{M^6(1-\rho)},& \rho&=\frac{\max(|s|,|t|,|u|)}{M^2}<1. \end{aligned}$$

This follows by applying $1/(1-y)-(1+y)=y^2/(1-y)$ to each channel; every denominator is positive for real invariants in this domain. If $r_{\mathrm{ex}}=|R|/|A_{\mathrm{ex}}|$, then the full-amplitude relative error is

$$r_{\mathrm{full}} =\frac{|R|}{|-\lambda_{\mathrm{full}}+A_{\mathrm{ex}}|} =r_{\mathrm{ex}}\, \frac{|A_{\mathrm{ex}}|}{|-\lambda_{\mathrm{full}}+A_{\mathrm{ex}}|}.$$

The ratios are defined only for nonzero denominators. A cancellation between contact and exchange can make $r_{\mathrm{full}}$ large while $r_{\mathrm{ex}}$ is small. A cross section requires the full amplitude squared, including interference with the contact, and has its own propagated error. The bound on $R$ compares these two tree amplitudes; it does not bound omitted heavy loops, light loops, or running effects.

A credible EFT result reports at least four distinct limitations:

• parametric truncation: powers, couplings, or loops not retained;
• matching and running: perturbative order, scheme, scale, and basis choices;
• inputs and numerics: parameter covariance, discretization, sampling, and integration error; and
• domain: thresholds, new light modes, large logarithms, or other failures of the assumed hierarchy.

Do not combine these automatically in quadrature. State correlations and distinguish probabilistic uncertainties from scale variations, bounds, and qualitative diagnostics.

Common pitfalls

“Nonrenormalizable” means nonpredictive. An EFT usually contains infinitely many symmetry-allowed operators, but only finitely many contribute at any fixed order in a valid power counting. Predictivity comes from that ordering, not from keeping a finite list for all energies.

Heavy effects must vanish from every coefficient. In the stable, weak-coupling example above, the matching shift has magnitude $3g^2/M^2$. It vanishes as $M$ grows at fixed $g$, but remains finite when $g/M$ is held fixed. Matching absorbs this finite shift into the light quartic; it does not contradict decoupling of low-energy predictions after matching. The derivative terms retain their declared inverse-mass suppression.

Matching and running are interchangeable. Matching supplies boundary data at a threshold; running transports that data within one theory. Running cannot discover a finite hard threshold contribution that was omitted from matching.

One higher-dimension operator is enough. A calculation must include every independent operator and loop contribution at the declared order. Choosing only the most familiar term destroys the stated accuracy.

Exercises

1. Complete the square

Starting from $\mathcal L_H=-\tfrac12HAH-JH$, where $A=\Box+M^2$ and $J=g\phi^2/2$, show that eliminating $H$ gives $\Delta\mathcal L=\tfrac12JA^{-1}J$.

Solution

Write

$$-\frac12HAH-JH =-\frac12(H+A^{-1}J)A(H+A^{-1}J) +\frac12JA^{-1}J.$$

The classical solution is $H=-A^{-1}J$, so the squared term vanishes there. With $J=g\phi^2/2$, the remaining interaction is $g^2\phi^2A^{-1}\phi^2/8$, including the positive sign and the factor $1/8$.

2. Test the truncation

For $x=q^2/M^2$, prove the single-channel relative-error formula above and find the largest positive $x$ for which retaining $1+x$ is accurate to better than $5\%$ in that exchange contribution. Does the same percentage bound the full amplitude with its contact term?

Solution

The finite geometric sum is

$$S_N=\sum_{n=0}^{N}x^n =\frac{1-x^{N+1}}{1-x}.$$

Thus $(1-x)^{-1}-S_N=x^{N+1}/(1-x)$. Dividing its magnitude by $\lvert(1-x)^{-1}\rvert$ gives $\lvert x\rvert^{N+1}$. Retaining $1+x$ means $N=1$, so $x^2<0.05$ and therefore $0\le x<\sqrt{0.05}\simeq0.224$. This numerical boundary is an accuracy choice, not the physical pole at $x=1$.

The full-amplitude percentage need not obey this bound. Its denominator includes the common contact and can approach zero without changing the heavy-exchange remainder. Use the full denominator or report the absolute bound instead.

3. Identify a redundant operator

Show that $\phi^2(\partial\phi)^2$ and $-\phi^3\Box\phi/3$ differ by a total derivative. For a leading Lagrangian $\mathcal L_0=\tfrac12(\partial\phi)^2-\tfrac12m^2\phi^2-\lambda\phi^4/4!$, use the leading equation of motion to express $\phi^3\Box\phi$ in terms of nonderivative operators.

Solution

The product rule gives

$$\partial_\mu(\phi^3\partial^\mu\phi) =3\phi^2(\partial\phi)^2+\phi^3\Box\phi.$$

After integration, the left side is a boundary term, so the two derivative operators are equivalent in the action under the stated boundary conditions. The leading equation of motion is

$$\Box\phi+m^2\phi+\frac{\lambda}{3!}\phi^3=0.$$

Consequently,

$$\phi^3\Box\phi =-m^2\phi^4-\frac{\lambda}{3!}\phi^6$$

up to terms beyond the order at which the leading equation is valid. This is a basis relation, not permission to discard the corresponding physical effect.

References

• Burgess, C. P. Introduction to Effective Field Theory: Thinking Effectively about Hierarchies of Scale. Cambridge: Cambridge University Press, 2021. DOI.

Further reading

• Weinberg, Steven. “Phenomenological Lagrangians.” Physica A: Statistical Mechanics and its Applications 96, nos. 1–2 (1979): 327–340. DOI.

For the scalar sequence, complete the worked capstone: it combines the scattering rate and running coupling with a stable heavy-field model and an exact tree-level remainder bound.

For the full graduate core, continue to infrared-safe observables and synthesis to turn amplitudes and scale-separated ingredients into a measurable prediction. If you are following the displayed roadmap, QED and Yang–Mills theory is the preceding application; EFT itself can be entered directly after loops and renormalization-group reasoning.