Collinear Factorization Violation in QCD

• Collinear factorization violation is the breakdown of universal, process-independent splitting behavior in QCD amplitudes, notably occurring in space-like limits due to color and kinematic correlations.
• Generalized splitting amplitudes that depend on non-collinear partons reveal how Coulomb-Glauber exchanges and soft-collinear entanglement drive the violation beyond tree level.
• The phenomenon impacts practical predictions in hard scattering processes, influencing observables in deep-inelastic and exclusive channels where standard collinear factorization fails.

Collinear factorization violation denotes the failure of the strict, process-independent factorization of QCD amplitudes or cross sections in a collinear limit. In the strict form, the singular collinear factor depends only on the momenta and quantum numbers of the collinear partons. A substantial body of work shows that this statement is exact at tree level and remains valid in time-like collinear limits, but it fails beyond tree level in space-like configurations and in other Glauber-sensitive settings, where the singular factor can acquire explicit dependence on non-collinear partons through color and kinematic correlations (Catani et al., 2012). At the same time, several apparent violations have been shown to be artifacts of incomplete proofs or inappropriate factorization procedures rather than genuine breakdowns (Bodwin et al., 2010).

1. Strict versus generalized factorization

The canonical amplitude-level statement of strict collinear factorization is that, when a set of partons becomes collinear, the full amplitude factorizes into a universal splitting amplitude multiplying a reduced amplitude in which the collinear set is replaced by a parent parton. In the notation used in recent generalizations, strict factorization has the schematic form

$\ket{M(c,h)} \simeq Sp(c;)\,\ket{M(;h)},$

where $c$ denotes the collinear set and $h$ the remaining hard partons. “Strict” means that $Sp(c;)$ depends only on the collinear sector (Cieri et al., 2024).

The generalized form replaces this by

$\ket{M(c,h)} \simeq Sp(c;;h)\,\ket{M(;h)},$

so that the splitting amplitude retains dependence on non-collinear external partons. This generalized structure was formulated for multiparton collinear limits to all orders, with the singular behavior encoded by all-order splitting matrices or generalized splitting amplitudes (Catani et al., 2012). In this sense, “collinear factorization violation” does not usually mean that singular behavior ceases to factorize at all; it means that the simplest universal and process-independent collinear factor ceases to exist.

The same refinement extends to more elaborate infrared limits. For several collinear directions, the naive product

$Sp(c_A;_A)\,Sp(c_B;_B)$

is replaced in general kinematics by a single non-factorizable object

$Sp(c_A,c_B;_A,_B;h),$

and in simultaneous soft-collinear limits naive multiplicative factorization into a soft current times independent splitting amplitudes also fails beyond tree level (Cieri et al., 2024).

2. Kinematic domains and the role of space-like limits

A central distinction is between time-like and space-like collinear configurations. In the time-like region, all collinear partons are either all final-state or all initial-state, so the collinear splitting is purely final-state or purely initial-state. In the space-like region, at least one collinear parton is initial-state and at least one is final-state. The strict factorization formula holds at tree level in both regions, but beyond tree level it remains strictly valid only in the time-like region (Catani et al., 2011).

The violation in the space-like region is not merely an analytic-continuation subtlety. In one-loop and higher-loop amplitudes, the space-like splitting matrix depends on color operators and invariants involving non-collinear partons. This dependence encodes the fact that causality distinguishes initial-state and final-state interactions, so the collinear subsystem does not behave as a self-contained object in the same way as in time-like kinematics (Catani et al., 2012).

The effective-field-theory analysis sharpens this statement. The factorization-violating situation is specifically associated with an incoming colored particle becoming collinear to an outgoing colored particle. When no incoming direction is collinear to an outgoing one, the Glauber contribution is entirely contained in the soft contribution, and ordinary hard/soft/collinear factorization is preserved. The same analysis further states that there is no violation for purely final-state collinear splittings and no violation if there is only one incoming colored particle (Schwartz et al., 2017).

Deep-inelastic scattering is a special case. Because the non-collinear colored partons are all in the final state, the initial-state/final-state Coulomb-Glauber interactions responsible for the breakdown are absent, so the space-like splitting matrices effectively retain a strictly factorized form (Catani et al., 2012).

3. Dynamical origin: Coulomb-Glauber exchange and soft-collinear entanglement

The standard mechanism behind strict-factorization breaking is the exchange of Coulomb or Glauber gluons. At one loop, the factorization-breaking operator is anti-Hermitian and has the form

$$\Delta^{(1)}_{mC}(\mu^2) = \frac{\alpha_s(\mu^2)}{2\pi}\, i\pi \sum_{\substack{i\in C\ j\in NC}} \mathbf{T}_i\cdot\mathbf{T}_j\, \Theta(-z_i)\,\mathrm{sign}(s_{ij}),$$

where $C$ is the collinear set and $NC$ the non-collinear partons. The explicit dependence on non-collinear partons is the direct signal of the breakdown of strict universality (Catani et al., 2012).

In effective field theory, this physics is reproduced by explicit Glauber operators. The relevant momentum region has Glauber scaling with transverse momentum parametrically larger than the light-cone components, and the contribution is non-analytic in external kinematics because of the rapidity regulator. That non-analyticity is precisely what allows the same operator to vanish when there is no Glauber pinch and to survive when an initial-state and final-state direction become collinear (Rothstein et al., 2016). The one-loop Glauber exchange reproduces the imaginary factorization-violating term; double Glauber exchange reproduces the leading real two-loop non-factorizing structure (Schwartz et al., 2017).

A distinct but related subtlety arises in factorization proofs themselves. Standard leading-region analyses can miss configurations in which low-energy collinear gluons couple to soft gluons. This does not, by itself, imply a failure of collinear factorization. In Feynman gauge, the loophole is closed by collinear approximations, soft approximations, graphical Ward-Takahashi identities, cancellation of soft eikonals in color-singlet hadronic states, and absorption of low-energy collinear gluons into jet functions (Bodwin et al., 2010). The significance is methodological: soft-collinear communication must be controlled explicitly, because naively excluding it leaves the proof incomplete.

4. Perturbative structure and what survives in observables

The perturbative pattern is highly structured. At one loop, strict factorization fails at amplitude level in the space-like region, but the violating term is anti-Hermitian, so it contributes only a phase. For the simplest squared amplitudes this phase cancels, and the one-loop squared splitting matrix can remain strictly factorized even though the amplitude-level splitting matrix does not (Catani et al., 2011).

At two loops, new non-abelian color correlations appear. The generalized factorization formula contains structures involving one collinear and two non-collinear partons, with both Hermitian and anti-Hermitian pieces. These are no longer reducible to the square of the one-loop phase, and they provide the first genuinely richer factorization-breaking structures (Catani et al., 2012).

For physical cross sections, the cancellation pattern depends on the observable and the hard process. For pure QCD non-inclusive observables in hadron-hadron scattering with two incoming colored partons, the one-loop effect cancels and the two-loop effect still cancels in physical cross sections, while the first non-zero surviving contribution appears at three loops. The physical mechanism is the non-commutation of Coulomb/Glauber exchanges with ordinary soft eikonal exchanges, the same color structure that underlies super-leading logarithms (Forshaw et al., 2012). Beyond pure QCD, factorization violation can already contribute at two loops if the hard subprocess contains matrix-element contributions with phase differences between different color topologies (Forshaw et al., 2012).

These surviving effects have implications beyond the splitting amplitudes themselves. The literature connects them to the non-abelian structure of logarithmically enhanced terms, to super-leading logarithms, to factorization issues of mass singularities, and to TMD factorization questions (Catani et al., 2012). A recurrent theme is that partial cancellation in squared amplitudes is real but incomplete; it does not restore strict universality at sufficiently high perturbative order.

5. Concrete manifestations and diagnostic examples

A direct leading-twist breakdown has been exhibited for the exclusive process

$c$0

in the gluon-GPD channel. The obstruction is a trapped Glauber gluon exchanged in the $c$1-channel at leading power. Both $c$2 and $c$3 are pinched, so the contour cannot be deformed into a purely collinear or purely soft region. Under naive collinear factorization, the resulting GPD$c$4DA convolution develops an endpoint-like divergence in the combined limit $c$5, $c$6. The paper concludes that the standard leading-twist collinear description fails for this channel, while quark-GPD channels do not share the problem (Nabeebaccus et al., 2024).

Vector angularities in Drell-Yan were introduced as a controlled probe of such effects. The observable

$c$7

reduces at $c$8 to the standard boson transverse momentum, where factorization is established. The baseline SCET factorization formula presented for small $c$9 explicitly omits Glauber contributions. In Pythia studies, multi-parton interactions were used as a proxy for factorization-violating physics: the effect was negligible at $h$0 but became visible for $h$1, with hadronization effects comparatively small (Bijl et al., 2023). This does not prove a QCD factorization theorem with violation included; it provides an observable family designed to interpolate between a factorization-protected limit and potentially Glauber-sensitive regimes.

Outside perturbative QCD proper, the same language appears as a consistency test. For the off-shell Sudakov form factor on the Coulomb branch of $h$2 super Yang-Mills, a three-loop method-of-regions analysis found that the hard region factorizes cleanly, while mixed ultrasoft-collinear regions do not. The obstruction is that in propagators such as

$h$3

the ultrasoft-collinear coupling remains of leading parametric order in the relevant region. At three loops, distinct mismatch factors are required for different mixed sectors, so a simple multiplicative $h$4 ansatz cannot be maintained (Belitsky et al., 28 May 2025).

A further extension appears in on-shell studies of gauge anomalies. One-loop anomalous 4-point amplitudes with graviton exchange can be embedded into 5-point amplitudes whose collinear limits exhibit singularities such as $h$5 or $h$6 that cannot be generated by universal one-loop splitting amplitudes. In this framework, anomaly-cancellation conditions are recast as conditions for the preservation of collinear factorization in one-loop 5-point amplitudes (Alviani et al., 3 Sep 2025).

6. Apparent violations, non-violations, and the boundary of applicability

Not every failure of a naive factorized formula is a genuine violation of collinear factorization. A prominent example is the exclusive process $h$7 at the two-parton twist-3 level. Earlier literature reported residual collinear divergences that seemed to violate universality. The later reanalysis showed that the issue came from inserting the Fierz identity too early, at the level of hard kernels rather than the full amplitude. Once the infrared analysis is done first and the valence-quark equations of motion are used for power counting, the offending term is identified as twist-4 and factorization is restored (Cheng et al., 2017).

Similarly, the all-orders proof of hard-soft-collinear factorization for well-separated final-state jets establishes universal collinear splitting functions and soft currents, with soft-collinear overlap removed by Wilson-line vacuum denominators. That proof explicitly excludes configurations involving initial-state/final-state collinearity and Glauber complications. Its scope therefore supports the robustness of collinear factorization within its stated kinematic domain rather than contradicting the later violation results (Feige et al., 2014).

Multi-collinear refinements also do not constitute violations. The decomposition of triple-collinear splitting functions into strongly ordered $h$8 terms plus a remainder $h$9 that is finite whenever any pair of collinear particles becomes collinear makes the unresolved structure more transparent, but it leaves standard collinear factorization intact (Braun-White et al., 2022). The revisited leading-order triple-collinear splitting functions likewise separate iterated single-unresolved contributions from genuinely correlated double-unresolved content without implying any breakdown of factorization (Braun-White, 2022). In celestial CFT, recursive OPEs and descendant contributions reproduce known multi-collinear gluon splitting behavior rather than violating it (Ebert et al., 2020).

Other uses of the phrase concern the factorization properties of observables rather than the universal splitting amplitudes themselves. In jet substructure, a necessary condition for factorization is that the phase-space constraints on soft and collinear dynamics be independent. Generic declustering and filtering fail this “soft-collinear observable phase space” criterion because the measurement operator depends on recombination history or subjet counting in a way that entangles soft and collinear sectors. By contrast, pruning, trimming, and $Sp(c;)$0-subjettiness can factorize if their parameters scale appropriately (Walsh et al., 2011).

Finally, hard-scale ambiguity in collinear factorization is conceptually distinct from factorization violation. The comparison with $Sp(c;)$1-factorization shows that the conventional choice $Sp(c;)$2 arises dynamically as the collinear limit of a more general transverse-momentum-dependent description, but the precise choice of $Sp(c;)$3 remains intrinsically ambiguous and this ambiguity is not removed by higher orders. The paper explicitly states that collinear factorization is not invalid; rather, it carries an additional theoretical uncertainty absent in $Sp(c;)$4-factorization (Guiot, 2018).

In this broader landscape, “collinear factorization violation” therefore names several related but distinct phenomena: failure of strict process independence in space-like limits, observable-level sensitivity to Glauber exchange, entanglement of soft and collinear sectors in specific exclusive or off-shell problems, and algorithm-induced failures of soft-collinear phase-space separation. The common thread is that universality is more conditional than the simplest textbook formula suggests.