Generalized Unitarity Method

• Generalized unitarity is a method that reconstructs loop amplitudes by cutting internal propagators and sewing together on-shell tree amplitudes.
• It employs spanning sets of cuts and algebraic techniques to extract loop-integrand coefficients while preserving physical symmetries in diverse quantum field theories.
• The approach has versatile applications ranging from higher-dimensional QCD computations and gravitational observables to PT-symmetric one-dimensional scattering.

Generalized unitarity is a method for obtaining loop amplitudes from products of on-shell tree amplitudes by systematically cutting internal propagators and summing over the physical states crossing the cuts. In its modern amplitude-theoretic form, it is generally applicable to both supersymmetric and non-supersymmetric amplitudes, including non-planar contributions, and it reconstructs either integrands or fully regulated amplitudes from a spanning set of generalized cuts (Bern et al., 2011). In a distinct one-dimensional scattering usage, “generalized unitarity” denotes exact algebraic relations among reflection and transmission amplitudes in non-Hermitian linear systems, especially for ${\cal PT}$-symmetric potentials (Mostafazadeh, 2014).

1. On-shell reconstruction and spanning cuts

The basic cut object is a product of tree amplitudes,

$$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$

with the sum running over all states crossing the cut lines. Generalized unitarity differs from ordinary two-particle unitarity by allowing multiple internal lines to be put on shell. In this way, all or part of the loop integrations are localized, and the amplitude can be matched to an integral representation

$$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$

A spanning set of cuts is a collection of cuts sufficient to determine all integral coefficients in the loop amplitude, and “maximal cuts”, “next-to-maximal cuts”, and related refinements provide a systematic hierarchy for resolving numerator structures (Bern et al., 2011).

Two reconstruction strategies are standard in this framework. The forward strategy merges cut results into an unconstrained integrand while subtracting overlaps; the reverse strategy starts from an ansatz for the integrand with undetermined coefficients and constrains them by matching unitarity cuts. The review literature emphasizes the reverse strategy as favored in practice, while also stressing that any property of on-shell tree amplitudes preserved under sewing can be carried to loop integrands, sometimes in modified form. Explicit examples include dual conformal symmetry and color-kinematics duality (Bern et al., 2011).

2. Dimensional regularization, supersums, and higher-dimensional formulations

A central technical issue is regularization. Four-dimensional cuts capture many amplitudes efficiently, but rational terms and other contributions missed in strictly four dimensions require $D$-dimensional information. One formulation combines generalized unitarity with six-dimensional helicity. In that approach, external states remain four-dimensional, internal loop momenta are six-dimensional, and one-loop dimensionally regularized QCD four-point amplitudes can be constructed while retaining helicity methods; the same framework was also used to confirm that the four-loop four-point amplitude of $\mathcal{N}=4$ super-Yang-Mills theory, including nonplanar contributions, is valid for dimensions less than or equal to six (Bern et al., 2010).

For one-loop QCD and Higgs-to-partons processes, six-dimensional helicity together with $D$-dimensional generalized unitarity produces both the cut-constructible and rational pieces in a single formalism. The amplitude is organized as

$$A_n^{(1)}=\frac{\mu^{2\epsilon}}{(4\pi)^{2-\epsilon}} \Bigg( \sum_{\rm boxes}[C_4^{[0]}I_4+C_4^{[4]}I_4[\mu^4]] +\sum_{\rm triangles}[C_3^{[0]}I_3+C_3^{[2]}I_3[\mu^2]] +\sum_{\rm bubbles}[C_2^{[0]}I_2+C_2^{[2]}I_2[\mu^2]] \Bigg)+O(\epsilon),$$

and the extra polarization states of six-dimensional gluons are removed in the FDH scheme by the state-sum reduction

$C^{\mathrm{FDH}}=C^{(6D)}-2\,C^{\rm scalar}.$

This formalism was illustrated for four- and five-point one-loop amplitudes in QCD, including external fermions, and for the next-to-leading order correction to the Higgs plus three positive-helicity gluons amplitude in the large top-quark mass limit (Davies, 2011).

A different regularization-compatible approach is the Four Dimensional Formulation. There the $d$-dimensional loop momentum is split as

$$\bar \ell=\ell+\tilde \ell,\qquad \bar \ell^2=\ell^2-\mu^2=m^2,$$

internal gauge bosons are treated as 4D massive vector bosons of mass $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$0, the $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$1-dimensional components are mapped to color scalars, and $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$2-dimensional fermions are represented as Dirac fields with mass $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$3. The $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$4-selection rules encode the extra-dimensional sector, many diagrams vanish automatically because of these rules, and both cut-constructible and rational terms are produced simultaneously from the same cuts (Bobadilla et al., 2015).

Supersymmetric state sums admit an analogous organization. In four-dimensional on-shell superspace, intermediate-state sums become Grassmann integrations, and generalized cuts can be written directly in terms of superamplitudes or super form factors. This is the basis for the treatment of MHV and NMHV sectors in $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$5 theories, where quadruple and triple cuts are evaluated as supersymmetric convolutions of tree-level building blocks (Bern et al., 2011).

3. Bases, residues, and direct coefficient extraction

One major line of development concerns the choice of basis for loop integrands. “Prescriptive unitarity” constructs a strictly-diagonal basis of loop integrands in which each coefficient is given by a specifically tailored residue in field theory. In this representation,

$$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$6

and each $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$7 is a single field-theory cut, or on-shell function, attached uniquely to one basis element. The construction proceeds by splitting numerator degrees of freedom into non-contact and contact terms, assigning defining cuts to the non-contact terms, and using contact terms of higher integrals to make all other basis elements vanish on those cuts. For planar, maximally supersymmetric Yang-Mills theory, closed-form representations of all $$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$8-point N$$C=\sum_{\text{states}}A_{\text{tree}}^{(1)}A_{\text{tree}}^{(2)}\cdots A_{\text{tree}}^{(m)},$$9MHV amplitudes through three loops were given in this form (Bourjaily et al., 2017).

A related graph-based development is the construction of a global loop-integrand basis for planar and nonplanar amplitudes. In that framework, one begins from a non-redundant set of representative graphs, augments each graph with irreducible scalar products, removes automorphism redundancy, and obtains a complete basis of integrands. Generalized unitarity cuts, expressed as products of tree amplitudes, are then mapped directly to basis elements, so the coefficients can be read off without ansatze or solving linear equations. The same organization lifts cut data to loop-level integrands and permits the use of the tree-level double copy to generate complete gravitational integrands at any loop order, while bypassing the need to identify higher-loop gauge-theory integrands that obey color-kinematics duality (Bern et al., 2024).

An algebraic route to minimal bases is provided by the duals of Feynman integrals and their intersection numbers. In this formulation, the cut integrand is paired with compactly supported dual forms localized on cut surfaces. If $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$0 is a master basis and $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$1 are duals satisfying $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$2, then

$$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$3

This pairing is invariant under total derivatives, so it bypasses the generation of integration-by-parts identities. The paper introducing this method gave two algorithms for computing multivariate intersection numbers and applied them to 4- and 5-point gluon amplitudes in generic spacetime dimension, including prescriptions for extracting rational terms in the four-dimensional limit (Caron-Huot et al., 2021).

Generalized unitarity also exposes constraints beyond coefficient extraction. In purely on-shell constructions of chiral gauge theories, unitarity produces a unitary S-matrix, but rational terms required by locality can generate inconsistent factorization channels. In four dimensions, the absence of such inconsistencies implies the vanishing of the cubic Casimir of the gauge group; in six dimensions, a non-vanishing symmetric trace of four generators leads to a factorization channel identifying the two-form of the Green-Schwarz mechanism. In this sense, the obstruction traditionally described as gauge anomaly appears as a failure to impose locality consistently on a unitary S-matrix (Huang et al., 2013).

4. Algebraic geometry of cuts, topology of solution spaces, and higher propagator powers

At high loop order, generalized unitarity cuts define algebraic varieties. For a four-dimensional $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$4-loop diagram with $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$5 cuts, the on-shell conditions give a system of $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$6 quadratic polynomial equations in $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$7 loop-momentum components, and the generic solution is a complex curve. Its topology is classified by the genus. Computational algebraic geometry—especially Gröbner bases, primary decomposition, and Macaulay2—was used to reduce these cut systems to plane curves and compute their arithmetic and geometric genera,

$$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$8

The same work used the Riemann–Hurwitz formula when a curve is presented as a ramified covering (Huang et al., 2013).

Concrete examples show how nontrivial topology enters multiloop unitarity. The one-loop triangle cut gives a genus-0 conic. At two loops, the planar double-box with all massive legs gives genus $$\mathcal{A}_n^{(L)}=g^{2L+n-2}\sum_i \frac{1}{S_i}\,c_i\,I_i,\qquad I_i=\int \prod_{l=1}^{L}\frac{d^D l_l}{(2\pi)^D}\frac{n_i}{\prod_{j\in\alpha_i}p_j^2}.$$9, while the non-planar crossed-box gives genus $D$0. At three loops, the planar box-pentagon-box has genus $D$1, the non-planar box-crossed-pentagon genus $D$2, the non-planar crossed-crossed-pentagon genus $D$3, and the Mercedes-logo diagram genus $D$4. Under degenerate kinematics, these curves split into lower-genus branches connected at discrete points. A genus-0 curve admits a rational parameterization, whereas genus $D$5 obstructs rational parameterization in any choice of coordinates (Huang et al., 2013).

Repeated propagators require further extensions because ordinary cut Jacobians become degenerate. For loop integrals with doubled or tripled propagators, the Jacobian on the cut vanishes and the residue is ill-defined in the standard multivariate Cauchy sense. One solution treats the cut as a degenerate multivariate residue and evaluates it using computational algebraic geometry: Gröbner bases, a transformation law for residues, and a Mathematica package calling Macaulay2. This gave direct extraction of integral coefficients for one- and two-loop integrals with doubled or tripled propagators (Sogaard et al., 2014).

A second solution uses differentiation over masses. If

$D$6

then repeated derivatives with respect to the masses translate the problem to integrals with all propagator powers equal to one. The same differentiation applies to the imaginary part, so standard unitarity cuts determine the reduction coefficients of all basis integrals except the tadpole. Bubble, triangle, box, and pentagon reductions with general propagator powers were worked out explicitly in this manner (Feng et al., 2021).

Two-loop integral reduction furnishes another application. A case study of the four-point double-box, double-triangle, and triangle-box generalized the unitarity method to two loops by cutting two propagators depending on $D$7 and two depending on $D$8. For the latter two topologies, complete analytical results were obtained in general $D$9-dimension. Under renormalizable constraints, the analysis found 20 scalar dimensionally shifted bases for the double-triangle and 48 bases for the triangle-box, illustrating that the integral basis can be substantially smaller than the algebraic integrand basis (Feng et al., 2014).

5. Form factors, correlators, integrated amplitudes, resonances, and gravitational observables

Generalized unitarity extends naturally from amplitudes to form factors. For the $\mathcal{N}=4$0 SYM stress-tensor current supermultiplet $\mathcal{N}=4$1, a supersymmetric generalized unitarity cut method was used to compute MHV and NMHV one-loop form factors. The one-loop form factor is expanded in scalar boxes and triangles with supersymmetric polynomial coefficients, quadruple cuts determine box coefficients, triple cuts determine triangle coefficients, and explicit answers were obtained for 3- and 4-point NMHV form factors. The same work also discussed the relation between the form factor with super momentum equal to zero and the logarithmic derivative of the superamplitude with respect to the coupling constant (Bork, 2012).

The same on-shell logic applies to correlation functions of local composite operators. In momentum space, sources for gauge-invariant operators are treated as non-dynamical external fields, and cuts are built from amplitudes, form factors, and multi-operator form factors. This framework was illustrated in $\mathcal{N}=4$2 super-Yang-Mills theory for BPS and non-BPS operators, including correlators related to energy flow in scattering processes and the effective action of a background gravitational field (Engelund et al., 2012).

Unstable particles require a modification of the cutting rules because only stable asymptotic states appear in unitarity sums. In theories with unstable particles, cuts are therefore not to be taken through unstable particles. For one-loop amplitudes containing resonances, generalized unitarity remains reliable under general physical conditions: near resonance, in the narrow-width regime, or in the complex-mass scheme, the cut propagator becomes a nascent delta function and reproduces the leading cut structure of the stable case. Off resonance, the cut through an unstable propagator yields only higher-order corrections, and reconstruction may instead require a formulation using only stable degrees of freedom (Menezes, 2021).

An integrated version of the method reconstructs not just integrands but full amplitudes. “Integrated Unitarity” exploits the relation between cuts and discontinuities of the amplitude in dimensional regularization and writes the amplitude dispersively in terms of lower-loop on-shell amplitudes: $\mathcal{N}=4$3 This method reproduced the four-gluon amplitude in two-loop massless QCD and produced a new result for the four-loop four-point massless planar ladder Feynman integral, expressed in terms of Harmonic Polylogarithms with letters 0 and 1 (Bargiela, 2024).

Gravitational applications have become increasingly explicit. One framework applied multi-loop and generalized-unitarity methods to the effective field theory of a binary composite-particle in order to derive causal effective actions of the dynamical multipoles, energy spectra, and fluxes for tail effects. It treated radiation-reaction, tail, tail-of-tail, and tail-of-tail-of-tail contributions through the four-loop level and seventh order in post-Newtonian gravity (Edison et al., 2022). A further extension introduced a generalized unitarity method for worldline field theory, with factorization applied to bulk graviton modes and worldline fluctuations. By complexifying worldline energies, the method fixed both double and single poles, and it reproduced the gravitational waveform for the scattering of two point masses at next-to-leading order, or $\mathcal{N}=4$4 (He et al., 1 Oct 2025).

6. Generalized unitarity relations in one-dimensional linear scattering

Outside perturbative quantum field theory, “generalized unitarity” also denotes exact relations among scattering amplitudes in one-dimensional linear systems. For an arbitrary linear scattering system that may violate unitarity, time-reversal invariance, $\mathcal{N}=4$5-symmetry, and transmission reciprocity, the transfer-matrix analysis yields

$\mathcal{N}=4$6

This relation involves amplitudes at both $\mathcal{N}=4$7 and $\mathcal{N}=4$8, reduces to $\mathcal{N}=4$9 in the ordinary Hermitian reciprocal case, and can fail at coherent perfect absorption points where the relevant determinant vanishes (Mostafazadeh, 2017).

For $D$0-symmetric scattering potentials satisfying $D$1, transfer-matrix transformations under $D$2, $D$3, and $D$4 imply

$D$5

together with

$D$6

These identities hold for both real and complex $D$7-symmetric potentials. In special cases, such as parity-even or real-valued potentials, they reduce to standard reciprocity and unitarity relations. The same framework accommodates spectral singularities associated in the paper with lasing and coherent perfect absorption (Mostafazadeh, 2014).

The coexistence of these two usages—on-shell reconstruction in quantum field theory and generalized amplitude identities in non-Hermitian one-dimensional scattering—reflects a common structural theme: analytic control of scattering data through on-shell conditions, factorization, and the algebraic organization of singularities. This suggests that generalized unitarity is best regarded not as a single algorithm but as a family of methods centered on the same principle: scattering information becomes tractable when constrained directly on its physical singular loci.