Infrared-safe observables and synthesis

A hard amplitude becomes a prediction only after a measurement has specified which final states are counted and how unresolved radiation is treated. In a massless theory, that definition determines whether soft and collinear singularities cancel, which logarithms remain large, and which long-distance inputs are still required. This lesson follows one event shape from its measurement function through next-to-leading-order (NLO) subtraction, factorization, resummation, matching, and a defensible uncertainty statement.

Required background. LSZ and tree amplitudes fix the rate normalization; loops and regularization separate ultraviolet from infrared poles; renormalization and the RG explain scale evolution; and QED and Yang–Mills theory supplies the massless gauge radiation. Effective field theory and matching is especially useful for the factorized description of separated scales.

A measurement function defines the prediction

For a fixed incoming state, let $F_n$ assign a weight to every physical $n$-particle final state. With relativistic state normalization, a weighted cross section can be written

$$\begin{aligned} \sigma[F] ={}&\frac{1}{\mathcal F} \sum_n\frac{1}{S_n} \int \mathrm d\Phi_n(P;\{p\}_n)\\ &\times \overline{\lvert\mathcal M_n(\{p\}_n)\rvert^2} F_n(\{p\}_n). \end{aligned}$$

$\mathcal F$ is the incoming flux, $S_n$ removes overcounting of identical final particles, $\mathrm d\Phi_n$ is Lorentz-invariant phase space at total momentum $P$, and the bar denotes the declared sums over final labels and averages over initial labels. The family $F_n$ contains the bin, cuts, recombination rules, accepted species, and any normalization. For example, $F_n=1$ gives an inclusive rate, while a bin $B$ of an observable $v_n$ uses $F_n^{B}=\mathbf 1_B(v_n)$. A differential distribution is best understood through such bins or smooth test functions, because endpoint delta functions are distributions rather than ordinary pointwise values. This measurement- function formulation, including its normalization and unresolved limits, is developed in Catani and Seymour 1997, § 2.1, pp. 297–299.

For massless final-state radiation, the elementary infrared-and-collinear (IRC) safety tests are

$$F_{n+1}(\{p\}_n,\lambda q) \longrightarrow F_n(\{p\}_n) \qquad (\lambda\to0^+)$$

and

$$F_{n+1}(\ldots,zp,(1-z)p,\ldots) \longrightarrow F_n(\ldots,p,\ldots), \qquad 0<z<1.$$

The first condition says that a zero-energy particle cannot change the measurement; the second says that exactly collinear daughters are indistinguishable from their parent. They must hold for the complete algorithm, including axes, flavor labels, cuts, and tie-breaking rules. A finite change in either limit weights degenerate states differently and can leave an uncancelled singular integral.

IRC safety is necessary for ordinary order-by-order partonic predictions of this kind, but it is not a claim that corrections are small. It does not remove initial-state collinear structure, endpoint logarithms, hadronization, detector response, or possible failures of factorization.

Thrust passes both unresolved-limit tests

Consider massless final states in $e^+e^-$ annihilation at center-of-mass energy $Q$. Define thrust and its two-jet variable by

$$T=\max_{\lvert\widehat{\boldsymbol n}\rvert=1} \frac{\sum_i \lvert\boldsymbol p_i\cdot\widehat{\boldsymbol n}\rvert} {\sum_i\lvert\boldsymbol p_i\rvert}, \qquad \tau=1-T.$$

For every trial axis, adding $\boldsymbol k=\lambda\boldsymbol q$ changes the numerator by at most $\lvert\boldsymbol k\rvert$ and changes the denominator by exactly that amount. The maximum may move—even discontinuously when two axes are degenerate—but its value changes only by $O(\lambda/Q)$. Hence $T_{n+1}\to T_n$ in the soft limit.

Now replace one massless momentum by exactly collinear daughters, $\boldsymbol p\to z\boldsymbol p+(1-z)\boldsymbol p$. For every trial axis,

$$\begin{aligned} &\lvert z\boldsymbol p\cdot\widehat{\boldsymbol n}\rvert +\lvert(1-z)\boldsymbol p\cdot\widehat{\boldsymbol n}\rvert\\ &\qquad=\lvert\boldsymbol p\cdot\widehat{\boldsymbol n}\rvert, \end{aligned}$$

and the daughters’ momentum magnitudes also sum to $\lvert\boldsymbol p\rvert$. The maximized ratio is therefore exactly unchanged. Thrust is IRC safe.

The limiting values provide useful checks. Two back-to-back massless particles have $T=1$ and $\tau=0$. Three equal-energy momenta separated by $120^\circ$ in a plane have $T=2/3$ and $\tau=1/3$. More generally, $1/2\leq T\leq1$ for a center-of-mass final state: the upper bound follows term by term, while the angular average of each absolute projection is half its momentum magnitude, so the maximum cannot be smaller than $1/2$. Therefore $0\leq\tau\leq1/2$. Near $\tau=0$, however, fixed-order coefficients contain large logarithms of $\tau$ and long-distance effects become increasingly important. Finiteness does not imply uniform fixed-order accuracy.

Real–virtual cancellation and local subtraction

Virtual loops and real emission describe different final-state multiplicities and are separately infrared divergent. Cancellation occurs only after the observable assigns compatible weights to states that become degenerate. The Kinoshita–Lee–Nauenberg result makes this inclusiveness conditional on the specified initial and final sums, not on an arbitrary exclusive amplitude Kinoshita 1962, pp. 650–677 and Lee and Nauenberg 1964, pp. B1549–B1562.

Subtraction makes the cancellation numerically usable. Work in dimensional regularization, $d=4-2\epsilon$, and first assume that all unresolved singularities are in the final state. Let $\mathrm d\sigma^B$, $\mathrm d\sigma^V$, and $\mathrm d\sigma^R$ denote the Born, ultraviolet- renormalized virtual, and real contributions. Choose a local counterterm $\mathrm d\sigma^A(\Phi_{n+1})$ and an exact map $\Phi_{n+1}\mapsto\widetilde\Phi_n$ such that the counterterm reproduces the real contribution in every singly unresolved limit. Then

$$\begin{aligned} \sigma_{\mathrm{NLO}}[F] ={}&\int_n \left[ \mathrm d\sigma^B+\mathrm d\sigma^V +\int_1\mathrm d\sigma^A \right]_{\epsilon=0}F_n(\Phi_n)\\ &+\int_{n+1} \left[ \mathrm d\sigma^R F_{n+1}(\Phi_{n+1}) -\mathrm d\sigma^A F_n(\widetilde\Phi_n) \right]_{\epsilon=0}. \end{aligned}$$

The second line is integrable because both the singular matrix element and the measurement approach their mapped lower-multiplicity limits. Integrating $\mathrm d\sigma^A$ over the unresolved variables exposes explicit $1/\epsilon^k$ poles that cancel those of the virtual term in the first line. This is an exact add–subtract identity for a coherent subtraction scheme; the counterterm need only agree with the real term in its singular limits. Catani and Seymour 1997, §§ 2.1 and 7.1–7.2, pp. 297–298 and 343–346 give a fully specified NLO construction.

For massless incoming partons, add the scheme-dependent initial-state mass-factorization counterterm on the $n$-particle phase space. At higher orders, multiple unresolved limits require additional counterterms and overlap organization; the two-line NLO formula is not an NNLO prescription.

A trustworthy implementation checks more than the final integral:

• along every soft and collinear trajectory, $R/A\to1$ or the appropriately normalized subtracted residual tends to zero;
• every regulator pole cancels among the virtual, integrated-subtraction, and, where needed, mass-factorization terms;
• the map preserves momentum, on-shell conditions, and its phase-space Jacobian;
• auxiliary slicing or restriction parameters leave a stable limit; and
• an independent phase-space point or inclusive normalization is reproduced.

Monte Carlo convergence cannot repair a wrong map or a missing finite term. Record the calculation version, input parameters, subtraction scheme, precision, seed policy, sample count, and benchmark values so that the numerical claim can be reproduced.

Factorization is more than naming momentum regions

Finding hard, collinear, and soft scalings identifies candidate leading regions. A factorization theorem must additionally show that approximations to those regions reproduce the leading contribution, define gauge-invariant operators and their measurements, subtract overlaps, control regulators, and either cancel or include Glauber exchange. The convolution and its power correction are part of the theorem. The pinch-surface, Ward-identity, and remainder logic is developed in Collins, Soper, and Sterman 2004, §§ 8–9, pp. 76–95, PDF.

For the two-jet limit of thrust, a representative leading-power statement is

$$\begin{aligned} \frac{1}{\sigma_0}\frac{\mathrm d\sigma}{\mathrm d\tau} ={}&H(Q^2,\mu) \int \mathrm ds_n\,\mathrm ds_{\bar n}\,\mathrm dk\\ &\times J_n(s_n,\mu)J_{\bar n}(s_{\bar n},\mu)S(k,\mu)\\ &\times\delta\!\left( \tau-\frac{s_n+s_{\bar n}}{Q^2}-\frac{k}{Q} \right) +\frac{1}{\sigma_0} \frac{\mathrm d\sigma_{\mathrm{nons}}}{\mathrm d\tau}. \end{aligned}$$

Here $H$ contains fluctuations of virtuality $Q^2$, the jet functions describe collinear hemisphere invariant masses, and $S$ is defined by soft Wilson lines with the thrust measurement. The convolution is the leading-power singular contribution. The term $\mathrm d\sigma_{\mathrm{nons}}/\mathrm d\tau$ is integrable and suppressed relative to the $1/\tau$ singular structure as $\tau\to0$; it need not be numerically small away from that limit, and fixed-order matching restores it. The formula applies to the stated $e^+e^-$ two-jet limit; it is not a template that can be transferred merely by renaming functions. Operator definitions, Wilson-line directions, color representations, measurement support, overlap subtractions, and the treatment of rapidity or Glauber regions are process dependent. A systematic derivation and the thrust scale hierarchy appear in Becher, Broggio, and Ferroglia 2015, §§ 4.5–7.

Several auxiliary scales must not be conflated:

• $\mu_R$ is the scale at which couplings and ultraviolet-renormalized operators are defined;
• in a hadronic process, $\mu_F$ separates incoming collinear physics absorbed into parton distributions from the short-distance coefficient;
• sector scales such as $\mu_H$, $\mu_J$, and $\mu_S$ minimize logarithms in the corresponding factorized functions; and
• a rapidity scale $\nu$ appears only when the chosen factorization has a rapidity divergence not regulated by dimensional regularization.

For $e^+e^-$ thrust there are no incoming hadron parton distributions and no PDF factorization scale $\mu_F$. In a hadronic prediction, the $\mu_F$ dependence of coefficient functions cancels that of the evolved parton distributions through the retained order. One may set $\mu_R=\mu_F$ numerically, but they answer different questions.

RG evolution, resummation, and matching

In the two-jet region the natural virtuality scales are

$$\mu_H\sim Q, \qquad \mu_J\sim Q\sqrt\tau, \qquad \mu_S\sim Q\tau.$$

Evaluating every function at one of these scales leaves large logarithms in the others. Instead, compute each function near its natural scale and evolve all sectors to a common scale. RG consistency requires their anomalous dimensions, interpreted as convolution kernels where necessary, to cancel in the physical cross section. The evolution resums towers $\alpha_s^n L^m$ with $L=\ln(1/\tau)$; its accuracy label must state which anomalous dimensions, boundary terms, and running-coupling order were used. Resummation reorganizes a factorization theorem—it does not prove one.

Away from $\tau=0$, fixed order contains nonsingular terms that the resummed limit does not. If $\sigma_{\mathrm{FO}}^{[k]}$ is known through order $k$ and $[\sigma_{\mathrm{res}}]_{[k]}$ is the expansion of the resummed result through that same order, additive matching gives

$$\sigma_{\mathrm{match}} =\sigma_{\mathrm{res}} +\sigma_{\mathrm{FO}}^{[k]} -[\sigma_{\mathrm{res}}]_{[k]}.$$

The subtraction prevents double counting, and direct expansion verifies

$$[\sigma_{\mathrm{match}}]_{[k]} =\sigma_{\mathrm{FO}}^{[k]}.$$

Smooth profile scales can follow the canonical hierarchy near the endpoint and merge to a common fixed-order scale in the tail. They must stay perturbative, respect the factorization region, and avoid artificial kinks. Multiplicative matching is another legitimate prescription when its denominators are well behaved; it differs from additive matching beyond the claimed accuracy. General event-shape resummation and its conditions are developed in Banfi, Salam, and Zanderighi 2005, §§ 2–4.

Assemble the prediction and state its uncertainty

For a finite thrust bin $B=[\tau_a,\tau_b)$, a compact theoretical prediction has the form

$$\sigma_B^{\mathrm{pred}} =\int_{\tau_a}^{\tau_b}\mathrm d\tau\, \frac{\mathrm d\sigma_{\mathrm{match}}}{\mathrm d\tau} +\delta_B^{\mathrm{power}},$$

where $\delta_B^{\mathrm{power}}$ denotes the stated nonperturbative or power-suppressed contribution, not an automatically known number. Near the peak, $Q\tau$ can approach a hadronic scale and a simple fixed power correction may need to be replaced by a nonperturbative function. The parton-level formula also does not include detector response unless an explicit response map is added.

Before presenting a number or curve, specify and test the following:

Ingredient	What to state	Decisive check
Measurement	process, frame, bin edges, cuts, recombination, normalization	analytic and numerical soft/collinear limits
Fixed order	coupling scheme, order, masses, input parameters, $\mu_R$ and any $\mu_F$	pole cancellation, order comparison, benchmark
Factorization and resummation	operator formula, power expansion, logarithmic accuracy, profiles, matching scheme	anomalous-dimension cancellation and fixed-order expansion
Numerical integration	parameterization, mappings, precision, sample count, seed policy	convergence, variance, and independent point or integral
Long-distance input	PDFs, hadronization parameters, or power model and its domain	input covariance and variation across the applicable region
Reproducibility	code/source revision, data-set identifiers, configuration, machine-readable outputs	independent rerun with frozen inputs

Keep uncertainty sources distinct before combining them. Renormalization- and factorization-scale variations, resummation-profile variations, and matching- scheme differences diagnose selected omitted terms; they are not random draws with a universal confidence level. PDF or calibrated input errors may carry a probabilistic covariance, while Monte Carlo integration has a sampling variance conditional on the integrand. Power-correction and factorization models require their own assumptions. Quadrature is justified only by a covariance model, and correlations across bins are often physically important. Empirical limitations of conventional scale variation are quantified in Bagnaschi et al. 2015, §§ 3 and 5.

Limitations and common pitfalls

IRC safe means small corrections. It means unresolved singularities can cancel for the declared measurement. Large endpoint, non-global, small-radius, or veto logarithms can still spoil a fixed-order expansion.

KLN makes an exclusive amplitude finite. Cancellation requires the specified sum over degenerate states with compatible measurement weights. Charged asymptotic states, confinement, and initial-state collinear singularities can require different observables or additional factorization.

A list of hard, collinear, and soft regions proves factorization. Region naming does not define operators, subtract overlaps, establish the measurement convolution, or settle Glauber exchange. Import only a theorem whose process and observable hypotheses match the calculation.

The subtraction term is the real matrix element everywhere. It must match the real term locally in every unresolved limit and be integrable after the chosen map. Away from those limits it is scheme dependent.

One scale-variation band is a statistical interval. Residual scale dependence probes only some higher-order structures. Keep it labeled as a diagnostic unless an explicit, calibrated probabilistic model is supplied.

Parton-level and measured distributions are identical. Hadronization, underlying event, particle decays, reconstruction, and detector migration are additional maps. Their relevance depends on the process and bin.

Exercises

1. Check thrust in limiting configurations

Show that two back-to-back massless particles have $T=1$. Then calculate thrust for three equal-energy momenta separated by $120^\circ$ and verify the soft and exactly collinear limits.

Solution

For two momenta $\boldsymbol p$ and $-\boldsymbol p$, choose the axis along $\boldsymbol p$. The numerator and denominator are both $2\lvert\boldsymbol p\rvert$, so $T=1$ and $\tau=0$.

For the symmetric three-particle event, take the trial axis along one momentum. The absolute projections are $E$, $E/2$, and $E/2$, while the denominator is $3E$. Thus $T=2/3$. An axis between two momenta gives a smaller sum, so this is the maximum and $\tau=1/3$.

Adding a momentum $\lambda\boldsymbol q$ changes both sums by at most $O(\lambda)$, hence the maximized ratio approaches its original value. Replacing $\boldsymbol p$ by $z\boldsymbol p$ and $(1-z)\boldsymbol p$ leaves every absolute projection and the denominator unchanged after summing the daughters. These are the soft and collinear tests.

2. Show that an NLO subtraction residual is finite

Let an unresolved parameter be $\lambda\to0^+$ and suppose $R=a/\lambda+b+O(\lambda)$, $A=a/\lambda+c+O(\lambda)$, and $F_{n+1}=F_n+d\lambda+O(\lambda^2)$. Show that $R F_{n+1}-A F_n$ is finite. What fails if $F_{n+1}-F_n$ approaches a nonzero constant?

Solution

Expanding gives

$$\begin{aligned} R F_{n+1}-A F_n ={}&\left(\frac a\lambda+b\right) (F_n+d\lambda) -\left(\frac a\lambda+c\right)F_n +O(\lambda)\\ ={}&ad+(b-c)F_n+O(\lambda). \end{aligned}$$

The $1/\lambda$ terms cancel, so the residual has a finite limit. If $F_{n+1}-F_n\to\Delta F\neq0$, the residual contains $a\Delta F/\lambda$ and is nonintegrable with the assumed measure. That is the local signature of an IRC-unsafe measurement.

3. Verify additive matching

Take $\sigma_{\mathrm{res}}=1+\alpha_s(L^2+c)+\alpha_s^2R_2+\cdots$ and $\sigma_{\mathrm{FO}}^{[1]}=1+\alpha_s(L^2+c+d)$. Match through first order and expand the result.

Solution

The first-order expansion of the resummed result is $[\sigma_{\mathrm{res}}]_{[1]}=1+\alpha_s(L^2+c)$. Therefore

$$\begin{aligned} \sigma_{\mathrm{match}} ={}&\sigma_{\mathrm{res}} +\sigma_{\mathrm{FO}}^{[1]} -[\sigma_{\mathrm{res}}]_{[1]}\\ ={}&1+\alpha_s(L^2+c+d)+\alpha_s^2R_2+\cdots. \end{aligned}$$

The complete fixed-order coefficient, including the nonsingular $d$, is reproduced, while the higher-order logarithms in $R_2$ and beyond remain resummed. Omitting the final subtraction would double count the constant and logarithmic terms already present at first order.

4. Separate the scales in a hadronic prediction

A hadron-collider observable uses $\alpha_s(\mu_R)$, parton distributions $f_i(x,\mu_F)$, hard and soft evolution scales, and, in one formulation, a rapidity scale $\nu$. State what each scale does and which cancellations a physical prediction must satisfy.

Solution

$\mu_R$ defines ultraviolet-renormalized couplings and operators. $\mu_F$ defines the separation between incoming collinear physics in the parton distributions and the short-distance coefficient; DGLAP evolution of the parton distributions cancels the coefficient’s $\mu_F$ dependence through the calculated order. Hard, jet, and soft scales are chosen near the natural virtualities of factorized functions and are evolved to a common scale; the sum of their anomalous dimensions must cancel in the combined cross section. $\nu$ is required only if separated sectors have a rapidity divergence, and its dependence must also cancel after the sectors are combined.

Residual dependence on any of these auxiliary scales is of higher order within a valid factorization. Varying them can diagnose missing terms but does not prove factorization and does not by itself define a confidence interval.

Choose a specialist continuation

The best next route depends on which part of the prediction you want to make more powerful:

Goal	Continue with
Collider amplitudes, subtraction, resummation, and precision observables	Scattering calculations for phenomenology
QCD, electroweak theory, hadrons, or nuclei	QFT for particle and nuclear physics
Operator bases, matching, running, and multiscale EFT	Renormalization and EFT for working researchers
Monte Carlo integration, lattice methods, and reproducible computation	Computational field theory onboarding
Collective excitations, transport, and response	QFT for quantum matter
Curved spacetime, cosmology, and gravitational EFT	QFT for gravity and cosmology
Structural questions and theorem hypotheses	Mathematical foundations for physicists

For a deeper version of this lesson, continue directly to measurement functions, real–virtual subtraction, hard–jet–soft factorization, and fixed-order matching.

References

• Bagnaschi, Emanuele, Matteo Cacciari, Alberto Guffanti, and Laura Jenniches. “An Extensive Survey of the Estimation of Uncertainties from Missing Higher Orders in Perturbative Calculations.” Journal of High Energy Physics 02 (2015): 133. DOI.
• Banfi, Andrea, Gavin P. Salam, and Giulia Zanderighi. “Principles of General Final-State Resummation and Automated Implementation.” Journal of High Energy Physics 03 (2005): 073. DOI.
• Becher, Thomas, Alessandro Broggio, and Andrea Ferroglia. Introduction to Soft-Collinear Effective Theory. Lecture Notes in Physics 896. Cham: Springer, 2015. DOI.
• Catani, Stefano, and Michael H. Seymour. “A General Algorithm for Calculating Jet Cross Sections in NLO QCD.” Nuclear Physics B 485 (1997): 291–419; erratum 510 (1998): 503–504. DOI.
• Collins, John C., Davison E. Soper, and George Sterman. “Factorization of Hard Processes in QCD.” In Perturbative Quantum Chromodynamics, edited by A. H. Mueller, 1–91. Singapore: World Scientific, 1989; updated author manuscript, 2004. DOI.
• Kinoshita, Toichiro. “Mass Singularities of Feynman Amplitudes.” Journal of Mathematical Physics 3 (1962): 650–677. DOI.
• Lee, T. D., and Michael Nauenberg. “Degenerate Systems and Mass Singularities.” Physical Review 133 (1964): B1549–B1562. DOI.