Real–Virtual Cancellation and Subtraction

At next-to-leading order, real-emission and virtual contributions are usually divergent before they are combined. Subtraction makes the cancellation computationally usable by adding and subtracting a local approximation that has the real matrix element’s unresolved limits, is simple enough to integrate analytically in the unresolved variables, and is tied to an exact map between $(n+1)$- and $n$-particle phase space.

Required background. Infrared and Collinear Safety supplies the measurement-function limits; their approach must also make the remainder integrable against the singular measure. Soft and Collinear Singularities supplies the universal unresolved factors reproduced by the counterterm.

The add–subtract identity

In dimensional regularization with $d=4-2\epsilon$, first consider a process with only final-state unresolved singularities. The NLO correction, excluding the Born term, is schematically

$$\delta\sigma^{\mathrm{NLO}} =\int_{n+1} \mathrm d\sigma^R_{\epsilon}\,F_{n+1} +\int_n \mathrm d\sigma^V_{\epsilon}\,F_n.$$

For massless incoming partons, a scheme-dependent mass-factorization contribution must also be added on the lower-multiplicity phase space; it is suppressed in the formulas below but included in the pole checks.

The real term is singular when one emitted parton is soft or two partons are collinear. The virtual term contains explicit $1/\epsilon^k$ poles. Since the terms live on different phase spaces, directly adding sampled values is meaningless.

Choose a local counterterm $\mathrm d\sigma^A$ and an exact map

$$\{p\}_{n+1}\longmapsto \{\widetilde p\}_{n},\Phi_{\mathrm{unres}}$$

such that

$$\mathrm d\sigma^A F_n(\{\widetilde p\}_n) \longrightarrow \mathrm d\sigma^R F_{n+1}(\{p\}_{n+1})$$

in every singly unresolved limit. Add zero in the form of the counterterm minus its integral. For compactness, set $\widetilde F_n=F_n(\{\widetilde p\}_n)$ in the master formula:

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

The first line is finite when the matched remainder, including the measurement and phase-space measure, is absolutely integrable; equality of unresolved limits alone does not establish this. Setting $\epsilon=0$ under an integral also requires a justified interchange of limit and integration, for example an integrable bound uniform in the regulator, as in the model below. The integrated counterterm exposes poles that cancel the virtual poles in the second line. This master identity and its exact dipole phase-space implementation are developed in Catani and Seymour 1997, § 2.1, pp. 297–298; §§ 7.1–7.2, pp. 343–346.

What the counterterm must reproduce

A valid counterterm is more than the leading power of one limit. It must satisfy all of the following within the claimed process class:

• reproduce every soft, collinear, and soft-collinear overlap of $\mathrm d\sigma^R$;
• preserve on-shell conditions and total momentum under the map;
• evaluate the lower-multiplicity measurement on the mapped momenta;
• retain the correct color and spin correlations when the factorization formula requires them;
• avoid introducing nonintegrable singularities away from the physical unresolved regions;
• admit an analytic or otherwise independently controlled integral over $\Phi_{\mathrm{unres}}$.

The counterterm is not unique. Different choices can differ by integrable functions and give very different Monte Carlo variances while producing the same physical result.

A one-dimensional model

Let $f$ be a fixed measurable function with finite $f(0)$ and $\int_0^1 \lvert f(x)-f(0)\rvert\,\mathrm dx/x<\infty$. A sufficient condition is $\lvert f(x)-f(0)\rvert\le Cx^\alpha$ near zero for constants $C<\infty$ and $\alpha>0$, together with integrability away from zero: the local weighted remainder is then bounded by the integrable function $Cx^{\alpha-1}$. The core cancellation appears in

$$I(\epsilon) =\int_0^1 \mathrm dx\,x^{-1-\epsilon}f(x) +\frac{f(0)}{\epsilon},$$

where the second term models a virtual pole. Add and subtract $f(0)$:

$$\begin{aligned} I(\epsilon) ={}&\int_0^1 \mathrm dx\,x^{-1-\epsilon}[f(x)-f(0)] \\ &+f(0)\int_0^1 \mathrm dx\,x^{-1-\epsilon} +\frac{f(0)}{\epsilon}. \end{aligned}$$

For real $\epsilon<0$ the elementary integral is $-1/\epsilon$, so the last two terms cancel. On $0<x<1$, $x^{-\epsilon}\le1$, so the absolute value of the remaining integrand is bounded by $\lvert f(x)-f(0)\rvert/x$, independently of $\epsilon<0$. This bound is integrable by assumption. Dominated convergence therefore permits $\epsilon\to0^-$ inside the remainder integral and gives

$$I(0)=\int_0^1 \mathrm dx\,\frac{f(x)-f(0)}{x}.$$

Continuity alone is insufficient: $f(0)=0$ and $f(x)=1/\ln(e/x)$ for $x>0$ give a continuous function whose weighted integral diverges, since $t=\ln(e/x)$ turns it into $\int_1^\infty \mathrm dt/t$. The weighted integrability condition is the relevant toy analogue of the unresolved cancellation requirement. Numerically, the subtraction must still be evaluated stably: when $x$ is tiny, subtracting nearly equal floating-point values can lose precision even though the subtracted integral is finite. Its integrand need not be bounded at zero.

Maps and overlap organization

An unresolved map typically identifies an emitter, an unresolved momentum, and possibly a spectator that absorbs recoil. Its Jacobian must satisfy an exact factorization

$$\begin{aligned} \mathrm d\Phi_{n+1}(P;\{p\}_{n+1}) ={}&\mathrm d\Phi_n(P;\{\widetilde p\}_n)\\ &\times[\mathrm d\Phi_{\mathrm{unres}}] \,J. \end{aligned}$$

Dipole subtraction partitions the singular approximation among emitter–spectator pairs and interpolates covariantly between soft and collinear limits. FKS subtraction instead partitions phase space into sectors with at most one soft and one collinear direction, then uses plus distributions in suitable variables Frixione, Kunszt, and Signer 1996, §§ 2–4, pp. 404–429. Antenna subtraction groups radiation between color-connected hard radiators. These are method families, not interchangeable formulas; their mappings, overlap rules, and integrated terms must be kept as coherent packages.

At NNLO, double-unresolved limits overlap in more ways and the simple NLO identity expands into several multiplicities and iterated counterterms. The same principles remain, but this page does not compare complete NNLO schemes.

Cancellation and numerical checks

A reliable implementation exposes the following tests separately:

1. Pointwise limits. For trajectories approaching each unresolved region, verify $R/A\to1$ or $(R-A)/R\to0$ at the predicted rate, including spin/color-correlated cases.
2. Pole cancellation. Sum the coefficients of every $1/\epsilon^k$ pole from the renormalized virtual term, integrated subtraction, and factorization counterterms before numerical integration.
3. Map checks. Verify on-shellness, momentum conservation, phase-space Jacobians, and invertibility away from boundaries.
4. Technical-parameter independence. If a slicing or restriction parameter is introduced, show a stable plateau with residual power corrections smaller than the quoted numerical uncertainty.
5. Counterterm deformation. For an integrable deformation compatible with the map, replace $A$ by $A+\delta A$ and recompute its consistently mapped integrated image. The first line then changes by $-\int_{n+1}\delta A$, while the second changes by $+\int_n\!\int_1\delta A$; the physical answer must remain unchanged within integration error.
6. Independent benchmark. Reproduce a lower-point analytic result or a separately implemented phase-space point.

Pole cancellation is necessary but not sufficient. A wrong finite term, symmetry factor, measurement map, or phase-space Jacobian can pass the pole test and still shift the answer.

Common pitfalls

Subtracting the matrix element but not the measurement. The local approximation must be multiplied by $F_n$ evaluated on mapped momenta. Using $F_{n+1}$ or unmapped momenta can spoil integrability.

Taking four dimensions too early. Keep $\epsilon$ until analytic poles from the integrated counterterm and virtual contribution have canceled. Only the explicitly finite brackets may be evaluated at $\epsilon=0$.

Using a soft approximation in a soft-collinear overlap without a partition. Double counting or missing an overlap leaves residual singular behavior. Follow the chosen scheme’s overlap construction exactly.

Diagnosing correctness from a stable integral alone. A biased integrand can converge beautifully. Perform local limits, pole checks, and independent normalization tests before trusting Monte Carlo stability.

Check your understanding

For $f(x)=1+x$, evaluate the one-dimensional model both before and after subtraction for real $\epsilon<0$. Confirm that the pole cancels and that $I(0)=1$. Then take $f(0)=0$ and $f(x)=1$ for $0<x\le1$. Find $I(\epsilon)$ and explain why no finite $\epsilon\to0^-$ limit exists.

Solution

For $f(x)=1+x$, the regulated real integral is $-1/\epsilon+1/(1-\epsilon)$. Since $f(0)=1$, the virtual term $1/\epsilon$ cancels the pole, leaving $I(\epsilon)=1/(1-\epsilon)\to1$. After subtraction, $f(x)-f(0)=x$ gives the same result directly from $\int_0^1 x^{-\epsilon}\,\mathrm dx=1/(1-\epsilon)$.

For the specified jump, $f(0)=0$, so the virtual term vanishes and $I(\epsilon)=\int_0^1 x^{-1-\epsilon}\,\mathrm dx=-1/\epsilon\to+\infty$ as $\epsilon\to0^-$. The putative limiting remainder is $\int_0^1 \mathrm dx/x$, which diverges. This function violates the weighted integrability condition: its assigned endpoint value does not match its nonzero limit from the right.

No runnable subtraction laboratory is currently available. The pointwise limits, pole-cancellation test, mapping checks, and one-dimensional model above provide the static reproducibility path.

Where to continue

• Phase-Space Integration and Monte Carlo Estimators explains how the finite brackets are sampled efficiently.
• Fixed-Order Organization and Scale Dependence states how the NLO coefficient enters a reported prediction.
• Validation and Theory Uncertainties separates pole, mapping, numerical, and benchmark checks.

References

• 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. Open preprint.
• Frixione, Stefano, Zoltan Kunszt, and Adrian Signer. “Three-Jet Cross Sections to Next-to-Leading Order.” Nuclear Physics B 467 (1996): 399–442. DOI. Open preprint.