QED and QCD:
from Lagrangians to high-precision results.
From a theory's Lagrangian to its Feynman rules and diagrams, to the running of the strong coupling and to one-loop integrals at any precision. What was compared, against what, and where the comparison stops.
From a Lagrangian
to a number at any precision.
A calculation in quantum field theory starts from a theory's Lagrangian and ends with a number that can be compared with an experiment. On the way it needs the theory's Feynman rules, every Feynman diagram with its symmetry factor and sign, the loop integrals the diagrams produce, and the running of the couplings and masses between energy scales. Each step has established programs, built over decades: FORM for diagrams and algebra, LoopTools and QCDLoop for one-loop integrals, CRunDec for the running of QCD.
We built each of these steps in PQCF, exact where the answer is exact and at any precision where it is numerical, and compared every one with those programs, with mpmath, or with the textbook. This page shows what was compared, the results, and where the comparison stops.
| step | in PQCF | compared with | result |
|---|---|---|---|
| Feynman diagrams | pq_diagrams | FORM 5's diagram generator; zero-dimensional field theory | equal in all 24 cases compared; exact |
| Feynman rules | pq_lagrangian | Peskin and Schroeder; the inverse of each quadratic Lagrangian | equal term by term |
| β and γm to five loops | pq_qcd, pq_lie | the formula for any gauge group; an independent transcription in mpmath | exact; 38 digits |
| running of αs and m | pq_qcd | mpmath's ODE solver at 45 digits; CRunDec 3.1.2 | 28 digits; within CRunDec's accuracy |
| the box D0 | pq_oneloop | mpmath's direct integral; LoopTools 2.16; QCDLoop 2.1.0 | 16 digits; about 1e-15 |
| the three-loop β-function | PQCF with FORM and Forcer | the published coefficients | exact |
Tested Reproduced These instruments are on PQCF's main branch since 16–18 November and not yet in a tagged release: release 1.4.0 does not contain them, so reproduce them at the commit their pages name (7570e62).
Every diagram,
with its symmetry factor and sign.
What it does. For given external legs and a number of loops, pq_diagrams generates every Feynman graph of a theory with its weight sign/S: S the symmetry factor, the sign (−1) for each closed fermion loop. It has QGRAF's filters: connected, one-particle irreducible, no tadpoles, no snails.
Against FORM 5's generator (diagrams_, Kaneko's grcc), in 24 cases of QED and QCD: the same number of diagrams and the same multiset of weights in every case.
| case | 1 loop | 2 loops | 3 loops |
|---|---|---|---|
| QED photon self-energy | 1 | 3 | 20 |
| QED electron self-energy | 1 | 3 | 20 |
| QED vertex | 1 | 9 | – |
| QCD gluon self-energy | 4 | 23 | 324 |
| QCD ghost self-energy | 1 | 7 | 97 |
| QCD quark self-energy | 1 | 7 | 97 |
| QCD three-gluon vertex | 8 | 112 | – |
| QCD quark-gluon vertex | 2 | 36 | – |
| QED vacuum graphs (2, 3, 4 loops) | 2, 5, 20 | ||
Against zero-dimensional field theory, exactly: φ⁴ vacuum graphs to six loops, two- and four-point graphs; φ³ vacuum graphs to five loops; a two-field theory; and QED vacuum graphs with N fermion flavours (the fermion-loop signs). In zero dimensions every integral is a number, so the weights of all the graphs must add up to the expansion of the theory's generating function (Cvitanović, Lautrup and Pearson, Phys. Rev. D 18, 1939).
Speed, one core: the three-loop gluon self-energy, 324 diagrams, in 0.1 s. Where it gets slow: vacuum graphs with many identical vertices; φ³ at six loops (388 graphs) takes 100 s.
The scope of the claim: 24 tested cases against FORM and the zero-dimensional checks, not every possible diagram of QED and QCD.
Feynman rules
from a Lagrangian.
pq_lagrangian turns a Lagrangian into Feynman rules: the vertices in momentum space, in a canonical form up to the symmetries of the metric and the structure constants, and the propagators.
- Vertices: φ⁴, φ³, Yukawa, QED, scalar QED and QCD (quark-gluon, three-gluon, four-gluon and ghost-gluon), equal to Peskin and Schroeder's An Introduction to Quantum Field Theory (1995) term by term.
- Propagators: scalar, Dirac and gauge-field (Rξ) propagators, each equal to −Γ2−1 of its own quadratic Lagrangian, evaluated with the Minkowski metric and Dirac matrices at a momentum.
QCD's renormalisation group
to five loops, exactly.
pq_qcd holds the β-function and the quark mass anomalous dimension of QCD to five loops as exact numbers: rationals and ζ-values, polynomials in the number of flavours nf. The sources: β3 hep-ph/9701390; β4 arXiv:1606.08659 and arXiv:1701.01404; γ3 hep-ph/9703278 and hep-ph/9703284; γ4 arXiv:1402.6611.
- β0 to β3 equal, term by term, to van Ritbergen, Vermaseren and Larin's formula for any gauge group, with SU(3)'s group invariants computed by
pq_lie. - β0 to β4 and γ0 to γ4, for nf = 3 to 6, equal to a separate transcription in mpmath to 38 digits.
αs and quark masses
at any precision.
The coupling and the MS-bar quark masses are run between scales by integrating the renormalisation-group equations at the working precision, and the coupling is decoupled at a heavy-quark threshold to three loops (RunDec's conventions, arXiv:1703.03751).
| check | against | agreement |
|---|---|---|
| αs(μ) and m(μ), 10 cases, 1 to 5 loops, nf = 3 to 6 | mpmath's ODE solver at 45 digits | 28 digits |
| the same | CRunDec 3.1.2, AsmMSrunexact | 1e-8 |
| decoupling of αs, down and up, 8 cases | mpmath; CRunDec DecAsDownMS, DecAsUpMS | 28 digits; 1e-14 |
| one loop | the closed form | 55 digits |
About CRunDec. CRunDec (Herren and Steinhauser) is the field's reference program, and its numbers are not wrong. Two of its functions are not meant to be exact: its exact running integrates with adaptive steps at a tolerance of 1e-10 per step, so it is accurate to about 1e-9; and its mass conversion evaluates an expanded formula, not the exact integral. So it is compared to 1e-8, within its own accuracy, and mpmath is the reference for the digits.
Not yet done: decoupling at four loops, and the decoupling of masses.
The one-loop box D0
at any precision.
pq_oneloop computes the one-loop integrals of one to four external legs (A0, the B functions, C0 and D0) at arbitrary precision for real kinematics, in LoopTools' conventions.
- Three Euclidean configurations: PQCF equals mpmath's direct three-dimensional integral to 16 digits.
- Four configurations above threshold: PQCF equals LoopTools 2.16 to about 1e-15.
- QCDLoop 2.1.0, in our build on ARM (aarch64), returns no number for two of the four configurations above threshold, and agrees with the others in the other two. We have not tested an x86 build; the programs may behave differently there.
- Slow above threshold: 45 seconds to 5 minutes at 18 digits on 20 cores. No speed is claimed for D0.
The two- and three-point functions were compared earlier (16 November): B0, B1 and B11 equal to mpmath to 35 digits and to LoopTools to 1e-11 in 60 configurations; C0 in 40 configurations against mpmath's direct integral (28 digits) and contour deformation (1e-12). In our ARM builds QCDLoop and LoopTools disagree with mpmath in several C0 configurations, all where the Källén function is negative (the comparison's page).
How to check it
yourself.
Each comparison has a step-by-step page, tested by us from a fresh clone on 19 November 2026 at commit 7570e62: the reference programs write their answers into a file, PQCF's tests compare with it, and regenerating the file must give it back byte for byte.
| result | how to reproduce | time |
|---|---|---|
| Feynman diagrams and rules | comparison/diagrams/REPRODUCE.md | seconds, once the images are built |
| The running of QCD | comparison/qcd/REPRODUCE.md | seconds |
| One-loop integrals, D0 included | comparison/oneloop/REPRODUCE.md | tests 3.5 minutes; D0's references 2 minutes |
| The three-loop β-function | comparison/beta/REPRODUCE.md | 15 minutes on 8 cores |
The code: github.com/quanterall/pqcf (src/pq_diagrams.erl, pq_lagrangian.erl, pq_qcd.erl, pq_oneloop.erl), and the comparisons' pages: diagrams and rules, QCD's renormalisation group, one-loop integrals.
The three-loop β-function of QCD,
recomputed exactly.
The β-function says how the strong coupling changes with energy. Its first coefficient is the reason quarks behave almost freely at high energies (asymptotic freedom). We computed it to three loops, as a test of a complete physics calculation rather than of single integrals. The result is the known one, first obtained by Tarasov, Vladimirov and Zharkov in 1980 and confirmed by Larin and Vermaseren with FORM (arXiv:hep-ph/9302208).
The result, every coefficient exact (a = αs/4π, β(a) = −β0a2 − β1a3 − β2a4):
- β0 = 11/3 CA − 4/3 TFnf
- β1 = 34/3 CA2 − 20/3 CATFnf − 4 CFTFnf
- β2 = 2857/54 CA3 − 1415/27 CA2TFnf − 205/9 CFCATFnf + 2 CF2TFnf + 158/27 CATF2nf2 + 44/9 CFTF2nf2
For QCD's three colours this is 11 − 2nf/3, 102 − 38nf/3 and 2857/2 − 5033nf/18 + 325nf2/54.
How. The coupling's renormalisation follows from the gluon and ghost self-energies in Landau gauge, where the ghost-gluon vertex needs none:
- The diagrams (338 at three loops) are generated by FORM 5 (arXiv:2601.19982).
- The Feynman rules, colour factors and Dirac traces are applied in FORM, the colour factors with its colour package (arXiv:hep-ph/9802376).
- The renormalisation and the β-function are extracted by PQCF.
Who computed the integrals.
- At one and two loops, PQCF reduced every integral itself, and its β0 and β1 are exact.
- At three loops, the integrals were computed by FORM with Forcer (arXiv:1704.06650): 12 minutes on 8 cores for the gluon, about a minute for the ghost.
- PQCF cannot yet do the three-loop integrals of this calculation. Landau gauge gives them high powers of numerators, and PQCF's choice of equations grows past a million for a single graph, one of which took 10 hours. That is our next piece of work.
A bug we found on the way: FORM's option for dropping “snail” diagrams also removes six non-zero three-loop diagrams. With it, two of β2's six coefficients come out wrong; we now drop only the diagrams that vanish.
The results themselves:
- the renormalisation constants and the β-function, at three loops and, with every integral reduced by PQCF, at two loops;
- the bare self-energies and PQCF's reductions, in comparison/beta/results;
- the method and the programs, in comparison/beta.
How to reproduce the results: a step-by-step page: one script builds the FORM, Forcer and PQCF images and runs the whole calculation, about 15 minutes on 8 cores; the three coefficients must equal the known ones and the results files must come out unchanged. We reran it from a fresh copy of the repository on 15 November.