Scattering Amplitudes And The Feynman Rules

Getting From a Diagram to a Number

You draw the lines, you assign the propagators, you plug in the vertices from your Lagrangian, and then you fight with Dirac algebra for six hours because you forgot a minus sign from an anticommutation. That is the rough arc of working with scattering amplitudes and the Feynman rules. It is not glamorous. It is also not as broken as people make it sound, provided you know where the trapdoors are. The rules themselves are mechanically straightforward. For a given QFT with a specified Lagrangian, each internal line becomes a propagator, each vertex contributes a coupling factor and momentum-conserving delta function, and you integrate over all undetermined loop momenta. External legs are amputated and replaced by polarization spinors or wavefunctions. The textbook derivation takes up maybe forty pages in Peskin & Schroeder or Srednicki, and most people never actually derive it from scratch because it is tedious and you already have the result.

What Scattering Amplitudes And The Feynman Rules Actually Mean in Practice

In practice what matters is not the formalism but the bookkeeping. A single tree-level diagram for a four-point process might give you a couple of pages of spinor chains. Add one loop and you are suddenly dealing with tensor integrals that reduce to scalar master integrals through Integration By Parts identities. The Feynman rules tell you what to write down. They do not tell you how to survive writing it down. I spent a week last year debugging a two-loop amplitude in scalar phi-theory with a mass insertion, because my reduction software was silently dropping a boundary term in the IBP reduction. The answer was off by a finite piece that only showed up when I cross-checked against a direct numerical integration over the Feynman parameters. It cost me three days. The fix was to use a different sector decomposition parameterization and verify the residue structure separately. The moral is that automated reduction is useful until it is not, and you need at least one independent check at every level of complexity. Here is a more concrete starting point. Say you want the tree-level amplitude for electron-muon scattering in QED, e minus mu minus to e minus mu minus. The Feynman rules give you a single t-channel photon exchange diagram. You write the amplitude as a product of two currents connected by the photon propagator:

iM = [ubar(pe) gamma^mu u(p)] * (-i g_{mu nu} / q^2) * [ubar(pm) gamma^nu u(p'm)] That is it for the amplitude. Squaring it, summing over spins with trace technology, and averaging over initial spins gives you the standard result that reduces to the Mott cross section in the appropriate limit. The physics is clean. The algebra is where people stumble, usually on sign conventions or the difference between covariant and physical polarizations. One counter-intuitive point that beginners consistently miss: the Feynman gauge propagator contains unphysical longitudinal and timelike polarization states, but they cancel exactly in gauge-invariant observables only after you include all diagrams at a given order. If you truncate a calculation mid-way through, you will get gauge-dependent results that look perfectly reasonable until you change the gauge parameter and watch the answer shift. This is not a bug in perturbation theory; it is a warning to check gauge invariance at every step, not just at the end.

Another thing that is not obvious from the textbooks: on-shell methods have made the Feynman diagram approach look archaic for many processes, and for good reason. The spinor-helicity formalism, BCFW recursion, and the modern amplituhedron program can compute certain tree-level amplitudes in seconds that would take hours by direct diagrammatics. But these methods have hard limits. They work beautifully for massless theories with good UV behavior. They struggle with massive external states, they do not easily handle rational terms in dimensional regularization, and they are not a replacement for Feynman diagrams in loop calculations where unitarity cuts alone leave ambiguities that require additional input. The real bottleneck in most calculations is not finding the amplitude. It is getting from the amplitude to a phenomenologically usable result. That means squaring, integrating over phase space, handling infrared divergences with a subtraction scheme like Catani-Seymour or FKS, and matching to a parton shower if you are doing something at a collider. Each of those steps has its own failure modes. Phase-space integrators can silently converge to the wrong answer if the integrand has sharp peaks that your sampling misses. Infrared subtraction can leave behind spurious logarithms if the resolution parameter is not chosen carefully relative to your jet algorithm. I will be blunt about the limitations. The Feynman rule approach breaks down or becomes impractical when you have many legs at tree level with massive particles, when you need high-multiplicity loop amplitudes beyond two loops in non-abelian gauge theories, and when you are working in regimes where the coupling is too large for perturbation theory to converge. Lattice methods, effective field theory expansions, and numerical bootstrap techniques fill some of those gaps, but they come with their own constraints. Perturbative Feynman diagram calculations are still the default for most collider physics because the alternative is usually worse.

For practical work, I would recommend building a workflow rather than trying to do everything by hand. Use symbolic packages like FormCalc, FeynCalc, or the Mathematica-based FeynArts together with Form for trace evaluation. Feed the resulting expressions into a reduction tool like LiteRed, Reduze, or FIRE for IBP reduction, then use an integral library or a numerical integrator for the master integrals. At tree level, MadGraph5_aMC@NLO handles multi-leg processes efficiently and outputs matrix elements in a form you can feed directly into Monte Carlo generators. The learning curve is real but the time savings are significant. A process that takes me about two hours of manual algebra at tree level with four external legs usually runs in under fifteen minutes once the output is wired into the right chain. The key is not to trust any single component blindly. Verify your Feynman rules against a known result in a simple channel. Check gauge invariance by varying the gauge parameter. Run the same process through two different tools when possible. The field moves fast enough that relying on a single pipeline without independent verification is how you publish incorrect cross sections and waste other people's time.