Learning to simplify a gluon amplitude
Cheung, Dersy and Schwartz trained networks to find groups of spinor-helicity terms that simplify together. A four-gluon example reduces 17 terms to one. Its value can be checked against the known Parke–Taylor amplitude using exact arithmetic.
Choosing which terms to simplify
A sum of tree diagrams can conceal a short answer. In a spinor-helicity representation, momentum conservation and the Schouten identity let visibly different products of brackets represent the same amplitude. The difficulty for a simplifier is deciding which terms to combine before applying those identities.
Cheung, Dersy and Schwartz generated training pairs by expanding short expressions with known identities. They trained an encoder-decoder transformer to reverse the process on short inputs. For longer inputs, a separate model trained with contrastive learning groups terms likely to simplify together. The sequential algorithm sends those groups through the transformer. The authors tested the method on amplitudes computed with established physics techniques.
A model output with an analytic benchmark
Consider the four-dimensional massless gluon tree amplitude with helicities (--++), cyclic color order (1,2,3,4), and all momenta outgoing. Appendix G gives a 17-term input for a particular choice of reference spinors. The model returns the expression on the left below. The right side is the known Parke–Taylor formula with negative helicities on legs 1 and 2.
Angle and square brackets are antisymmetric spinor contractions. Keep the cyclic denominator in order, especially the final angle bracket from leg 4 back to leg 1. Reversing that bracket changes the sign. Both forms use the same stripped normalization and apply at on-shell, momentum-conserving points where the displayed denominators are nonzero.
A check with exact momenta
We chose undotted and dotted spinors with rational components and solved momentum conservation before evaluating either formula. The spinors may vary independently at complexified kinematics. At the three points in the calculation file, both formulas gave, in order, 1/2, 3/70 and 49/40.
A second check changes the spinors of one leg while leaving its momentum fixed. Multiplying its undotted spinor by 3/2 and its dotted spinor by 2/3 multiplies the amplitude by 9/4 for a negative-helicity gluon and by 4/9 for a positive-helicity gluon. Both formulas gave the expected weight for each leg at all three points. Flipping the overall sign broke equality, and the script rejected an input that violated momentum conservation.
What the benchmark tells us
The four-gluon example tests whether a learned simplification recovers the value and helicity weights of a known amplitude. The paper also reports a reduction from 298 terms to two for three massless scalars and two positive-helicity gravitons in a theory with minimal gravitational coupling and cubic scalar self-interaction, Eq. (4.7). That example and the input-length dependence in Fig. 9 are separate parts of the authors’ benchmark.
For another proposed compact amplitude, start by checking that all terms have the same little-group weights. Then test the expression against an independent representation at nonsingular, momentum-conserving points. A mismatch at one admissible point disproves an identity. Agreement at finitely many points still leaves a symbolic identity to establish.
Limits of this result
- The comparison tests one tree-level model output. It supplies neither a general proof of equivalence nor a guarantee that learned simplifications are minimal.
- The equation locators and tested expression follow arXiv v2 of 19 November 2024.
Checks performed
Exact rational evaluation and helicity scaling
Compared the published compact output with Parke–Taylor at three rational points and twelve individual leg rescalings. The site calculation did not run model training or inference, transcribe the full 17-term input, or evaluate the five-point example.
Sources & citations
- Learning the Simplicity of Scattering AmplitudesClifford Cheung, Aurélien Dersy and Matthew D. Schwartz · 2024-11-19
arXiv:2408.04720v2. §1, Eq. (1.1), Parke–Taylor benchmark. §§4.2–4.3, grouping and sequential simplification. §4.4, Fig. 9 and Eq. (4.7), physical examples. Appendix G, four-gluon input in Eq. (G.2) and the unnumbered output immediately following.
Index terms spinor-helicity · MHV · maximally helicity violating · Parke–Taylor · Parke-Taylor · gluon · transformer · contrastive learning · Schouten · arXiv:2408.04720