Effective Field Theory as a Controlled Expansion

An effective field theory is controlled when its error can be reduced systematically by adding the next terms required by a declared expansion. The ultraviolet completion may be known, partly known, or entirely unknown; predictivity instead rests on a domain, active degrees of freedom, symmetry, power counting, and a finite set of coefficients at each target accuracy. This page makes that control statement explicit and tests it against an exactly known heavy-mediator amplitude.

Required background. Regulator Removal and Renormalized Predictions supplies the distinction between a regulator and a physical prediction, while Local and Composite Operator Insertions supplies the local-operator grammar. Helpful background. Free-Field OPE Preview previews how short-distance effects can be organized by local operators.

A finite prediction from an infinite action

Let $Q$ denote the low scale probed by an observable and let $\Lambda_b$ denote the scale at which the chosen effective description breaks down. A one-parameter expansion begins with

$$q\equiv\frac{Q}{\Lambda_b}<1.$$

The EFT contains every local operator allowed by its field content and symmetries,

$$\mathcal L_{\mathrm{EFT}} = \sum_i C_i(\mu)\,\mathcal O_i(\mu).$$

This sum is generally infinite. It becomes predictive only after a power counting assigns an order $\nu_i$ to each insertion and to the loops built from them. At a requested order $\nu_{\max}$, the calculation retains all contributions with $\nu\leq\nu_{\max}$ and estimates the remainder. The finite object is therefore not the exact Lagrangian but the set of contributions needed for a specified observable and accuracy.

Canonical dimension supplies part of the information. If $\mathcal O_i$ has mass dimension $d_i$ in $d$ spacetime dimensions, one convenient normalization is

$$\mathcal L_{\mathrm{EFT}} \supset \frac{c_i(\mu)}{\Lambda_b^{d_i-d}}\mathcal O_i,$$

with dimensionless $c_i$. But a genuine counting may also include weak couplings, loop factors, symmetry-breaking spurions, light masses, velocity factors, large occupation numbers, or infrared enhancements. “Dimension six” does not by itself mean “next-to-leading order.”

For an observable $X$, a control declaration can be written as

$$X(Q) = X_{\mathrm{ref}}(Q) \left[ \sum_{n=0}^{k}c_n(Q)q^n +R_{k+1}(Q) \right],$$

where the counting predicts which powers occur and the error model states what is assumed about $R_{k+1}$. A calculation is systematically improvable when increasing $k$ adds a known finite set of terms, renormalizes them consistently, and pushes the residual to the predicted next order.

Weinberg’s general prescription is precisely to write the most general local action consistent with the symmetries and organize its contributions by an expansion, rather than to reject interactions of dimension greater than four Weinberg 1979, pp. 327–331.

What must be declared before calculating

A reproducible EFT prediction fixes the following data.

Item	Question that must have an answer	Typical failure if omitted
Domain	Which energies, momenta, backgrounds, and particle multiplicities are admitted?	Extrapolation through a threshold or into a new phase
Degrees of freedom	Which states can propagate over resolved distances?	A light or nearly on-shell mode is encoded incorrectly as a local coefficient
Symmetry realization	Which symmetries are exact, broken by spurions, anomalous, or nonlinear?	Missing operators or forbidden interactions
Expansion parameters	Which ratios are small, and how are several ratios correlated?	Canonical dimension substitutes for a power counting
Normalization and scheme	How are fields, operators, coefficients, and $\mu$ defined?	Coefficients from incompatible conventions are combined
Target order	Which trees, loops, insertions, and parameter corrections enter?	An equally large contribution is omitted
Error model	What sets the first omitted term, and how is it tested?	Scale variation is mistaken for a complete uncertainty

Three different scales should not be conflated:

• $\Lambda_b$ is a physical or dynamical boundary of the EFT’s usefulness;
• $\Lambda_{\mathrm{reg}}$ is a regulator parameter introduced to define intermediate expressions; and
• $\mu$ is a renormalization or matching scale used to divide logarithms between coefficients and matrix elements.

A good calculation removes or controls $\Lambda_{\mathrm{reg}}$, cancels $\mu$ dependence through the retained order, and never claims validity beyond $\Lambda_b$. The numerical values can be comparable in a convenient implementation, but the concepts remain distinct.

The scale/content map below shows the relationship to inspect. The heavy state is present above its threshold, absent from the low-energy Hilbert space, and still represented through Wilson coefficients below the threshold. The nearest singularity bounds the local expansion.

Scale separation determines content and locality. Below the schematic threshold $M\simeq\Lambda_b$, the resolved field $\phi$ remains active while the heavy field $H$ is encoded by coefficients $C_4,C_6,C_8,\ldots$ ordered in $q=Q/\Lambda_b$. The geometric propagator expansion is local only for $|q^2|<M^2$; crossing the threshold or omitting a massless nonanalytic contribution requires different degrees of freedom. The diagram is schematic and not to scale.

Exact heavy exchange and its local expansion

Consider a massless real scalar $\phi$ and a real scalar $H$ of mass $M$,

$$\mathcal L = \frac12(\partial\phi)^2 +\frac12(\partial H)^2 -\frac12M^2H^2 -\frac g2H\phi^2.$$

For $\phi\phi\to\phi\phi$, define the reduced tree amplitude by removing the overall Feynman-rule phase. The full result is

$$\frac{\mathcal A_{\mathrm{full}}}{g^2} = \frac{1}{M^2-s} +\frac{1}{M^2-t} +\frac{1}{M^2-u}.$$

Below all heavy poles, each channel has the geometric expansion

$$\frac{1}{M^2-x} = \frac1{M^2} +\frac{x}{M^4} +\frac{x^2}{M^6} +\frac{x^3}{M^8} +\cdots, \qquad |x|<M^2.$$

Keeping the local series through $1/M^6$ gives

$$\frac{\mathcal A_{\mathrm{EFT}}^{(6)}}{g^2} = \frac3{M^2} +\frac{s+t+u}{M^4} +\frac{s^2+t^2+u^2}{M^6}.$$

For massless on-shell $2\to2$ scattering, $s+t+u=0$. The nominal $1/M^4$ contribution therefore cancels in this observable even though the corresponding derivative operators still belong to the action and can matter off shell or in other processes. An accidental kinematic zero does not license deleting an operator from the EFT.

At fixed center-of-mass angle $\cos\theta=1/3$, use

$$s=E^2, \qquad t=-\frac{E^2}{3}, \qquad u=-\frac{2E^2}{3}.$$

Then

$$\begin{aligned} s^2+t^2+u^2&=\frac{14}{9}E^4,\\ s^3+t^3+u^3&=\frac23E^6. \end{aligned}$$

The first omitted contribution is therefore

$$\frac{\Delta\mathcal A^{(8)}}{g^2} = \frac{2E^6}{3M^8} +O\!\left(\frac{E^8}{M^{10}}\right).$$

Relative to the leading $3g^2/M^2$ amplitude,

$$\frac{|\mathcal A_{\mathrm{full}}-\mathcal A_{\mathrm{EFT}}^{(6)}|} {|\mathcal A_{\mathrm{full}}|} = \frac29\left(\frac EM\right)^6 +O\!\left(\frac{E^8}{M^8}\right).$$

This is a sharper control statement than “$E$ is small”: a log–log plot of the residual against $E/M$ must approach slope six.

Numerical residual check

For the chapter fixture $M=10$, $g=1$, the exact amplitude and the expansion through $1/M^6$ give:

$E/M$	$\mathcal A_{\mathrm{full}}$	$\mathcal A_{\mathrm{EFT}}^{(6)}$	Relative residual
0.050	0.03000009733	0.03000009722	$3.49\times10^{-9}$
0.075	0.03000049339	0.03000049219	$4.00\times10^{-8}$
0.100	0.03000156234	0.03000155556	$2.26\times10^{-7}$
0.150	0.03000795409	0.03000787500	$2.64\times10^{-6}$
0.200	0.03002534746	0.03002488889	$1.53\times10^{-5}$
0.250	0.03006258503	0.03006076389	$6.06\times10^{-5}$
0.300	0.03013171112	0.03012600000	$1.90\times10^{-4}$

An unweighted linear fit to $\ln r$ versus $\ln(E/M)$ over these seven points gives slope $6.080$. The small excess above six is expected because the exact residual contains higher powers. Restricting the fit toward smaller $E/M$ approaches the analytic slope six. The check would fail if the $1/M^6$ term were omitted, if the axes used $E$ rather than $E/M$ without recording $M$, or if the residual were normalized by an inconsistent amplitude.

A reproducible calculation is designed around this deterministic benchmark and adds explicit adversarial cases. The table above preserves the reference claim independently of an interactive implementation.

Top-down and bottom-up EFTs

When the full theory is known, one can integrate out or match its heavy degrees of freedom. The coefficients $C_i$ are then calculable functions of the heavy masses, couplings, renormalization scheme, and matching scale. The exact heavy-exchange calculation is such a top-down construction.

When the ultraviolet completion is unknown, locality and symmetry still determine the operator structures. Their coefficients are independent low-energy parameters to be measured or constrained. This bottom-up construction remains predictive because only finitely many combinations contribute at a declared order. Measurements can then reveal which symmetries are approximate, estimate $\Lambda_b$, and test whether the assumed counting is self-consistent.

The two viewpoints share the same low-energy action. They differ in how its coefficients are obtained and what correlations among them are justified. A bottom-up coefficient should not be assigned a heavy-mediator relation merely because that relation holds in one possible completion. Conversely, matching a known full theory without including every EFT operator required at the retained order makes the comparison incomplete.

Burgess formulates the low-energy generator, Wilson action, and scaling logic without requiring that every short-distance detail be accessible Burgess 2021, §§ 2.2–2.4, pp. 26–44.

What systematic improvement does and does not promise

Systematic improvement means that a larger calculation comes with a sharper residual prediction. It does not mean that every series converges indefinitely. Perturbative coefficients may eventually grow, several small parameters may compete, or a new threshold may reorganize the degrees of freedom. The useful claim is local in theory space and kinematics: within a tested window, the retained hierarchy orders contributions and the residual behaves accordingly.

The statement also does not guarantee naturally sized coefficients. A symmetry can suppress a coefficient; a resonance can enhance one; an infrared fine tuning can promote an interaction; and a poor normalization can make dimensionless coefficients appear large. Coefficient sizes become evidence only relative to a specified normalization and mechanism.

Finally, an EFT does not erase ultraviolet physics. Heavy effects survive in Wilson coefficients, threshold corrections, anomalies, and the values of relevant parameters. What decouples is the need to resolve the detailed heavy dynamics in every low-energy calculation. The assumptions behind that statement are examined in Decoupling Theorems and Threshold Corrections.

Common pitfalls

“Nonrenormalizable” means unpredictive. An interaction with canonical dimension greater than four requires additional counterterms, but an EFT includes them in a hierarchy. Predictivity is an order-by-order claim, tested by Loops, Counterterms, and Closure of an EFT Expansion.

The regulator cutoff is the breakdown scale. A regulator is an intermediate definition; $\Lambda_b$ is a property of the chosen physical description. Cutoff variation can diagnose missing counterterms, but setting the regulator equal to a heavy mass does not prove the EFT is valid up to that mass.

A small correction proves the expansion. A single accidental cancellation can make one order tiny. Control requires the predicted scaling across energies or observables and stability under the complete contribution set.

Integrating out means deleting a field. Eliminating $H$ produces a generally nonlocal functional of $\phi$; the local EFT is its low-energy expansion. Loop-level elimination also produces a determinant, developed in Integrating Out Heavy Fields.

Exercises

For the fixed-angle heavy-exchange fixture, truncate the amplitude after the $1/M^4$ term. What residual slope is expected even though $s+t+u=0$?

Solution

On shell the entire $1/M^4$ contribution cancels, so this truncation equals the leading result $3g^2/M^2$. The first nonzero correction is

$$\frac{g^2(s^2+t^2+u^2)}{M^6} = \frac{14g^2E^4}{9M^6}.$$

Relative to $3g^2/M^2$, the residual begins as $14(E/M)^4/27$. The expected log–log slope is four, not two. The absent quadratic correction is an observable-specific kinematic cancellation.

Suppose a full amplitude contains $\ln(-s/m^2)$ from a light-particle cut. Can its entire effect be absorbed into a momentum-independent Wilson coefficient after the light particle is removed?

Solution

No. The logarithm is nonanalytic at the light threshold and represents propagation over resolved distances. A momentum-independent coefficient is analytic in the low external invariants. The light degree of freedom, or an equivalent nonlocal structure, must remain in the effective description; only the hard analytic part can be assigned to local Wilson coefficients.

References

• Burgess, C. P. Introduction to Effective Field Theory: Thinking Effectively about Hierarchies of Scale. Cambridge: Cambridge University Press, 2021. DOI
• Weinberg, Steven. “Phenomenological Lagrangians.” Physica A: Statistical Mechanics and its Applications 96 (1979): 327–340. DOI