Mathematical physics,
exact and at any precision.
PQCF began as a photonic simulator. Its symbolic engine, its reconstruction over finite fields and its arbitrary-precision numerics now serve five areas of mathematical physics. Every result is checked against an established program, a published result or an exact formula.
From operator algebra
to five-loop QCD.
Where this stands. Feynman-integral reduction is part of PQCF's latest release, 1.4.0 (8 November). The other four areas are on PQCF's main branch since 14–18 November and not yet in a release: tested and published with their comparisons, reproducible at the commits their pages name, but their interfaces may still change. To reproduce them, check out those commits: release 1.4.0 does not contain them.
- Tested
- checked by our own tests against an independent reference
- Benchmarked
- timed against an established tool on the same machine
- Reproduced
- rerun by us from a fresh clone, with a step-by-step page
- Prototype
- working, but not yet in its final place or form
- In development
- implementation underway; not yet complete
- Planned
- intended work that has not started
- Independently reviewed
- evaluated by a reviewer outside the development team: nothing yet
Quantum optics
Operator algebra for bosonic modes and atoms, master equations and quantum trajectories, exact steady states and spectra (also as rational functions of the parameters), cumulant expansions, collective atoms and input-output networks, at any precision.
- Four problems with closed-form solutions (the Rabi oscillation, collapse and revival, a decaying cavity, the steady state of resonance fluorescence): PQCF agrees with the closed forms to 36 digits, QuTiP 5.3.1 to 5–16 digits. The settings: PQCF's master equations summed by their Taylor series at 128 bits (about 38 digits); QuTiP's
sesolve,mesolveandsteadystateat their default tolerances and at the tightest its integrator accepts on these problems (atol 1e-14, rtol 1e-12), in double precision. This is a comparison on these problems and settings, not a general limit of QuTiP. - QuTiP is faster at time evolution, 12 to 280 times on the driven Jaynes–Cummings model; PQCF is faster on the steady state of resonance fluorescence.
- The driven Kerr cavity and the Kerr resonator with two-photon drive and loss equal their exact published solutions (Drummond and Walls; Bartolo et al.) to 30 digits.
- Cumulant equations agree with QuantumCumulants.jl to 1e-13; the collective-atom solver with QuTiP's PIQS within PIQS's single precision.
Quantum field theory
Feynman diagrams with their symmetry factors and fermion signs, Feynman rules from a Lagrangian, Dirac traces in d dimensions, one-loop integrals, angular-momentum algebra.
- Feynman diagrams of QED and QCD equal to FORM's generator in all 24 cases compared.
- Feynman rules: the QED and QCD vertices equal Peskin and Schroeder's term by term.
- Dirac traces equal to FORM on 60 traces of up to 12 matrices.
- One-loop integrals at any precision: two-point functions equal to mpmath to 35 digits and LoopTools to 1e-11; C0 in 40 configurations; the box D0.
QCD
Colour factors of SU(N) with N symbolic, the β-function and the quark mass anomalous dimension to five loops, the running and decoupling of αs and quark masses at any precision.
- β0…β3 equal, term by term, to the formula for any gauge group; all five loops to an independent transcription to 38 digits.
- αs and quark masses against mpmath to 28 digits, and against CRunDec within its own accuracy.
- The three-loop β-function recomputed from its diagrams, every coefficient exact; at three loops the integrals were Forcer's.
Feynman integrals
Integration-by-parts reduction to master integrals, solved modulo primes and reconstructed exactly, and the masters' differential equations.
- 3 to 7 times faster than FIRE on the tested problem, FIRE's own two-loop double box, on one core, with the same answers. Larger families are not yet compared.
- Four-loop integrals of FORM's own Forcer checks, every answer exact; Forcer, built for them, is 17 to 25 times faster.
- Next: larger families, spread across machines.
High-precision mathematics
The symbolic engine and the numerics under everything above: exact algebra, series, special functions, arbitrary precision, differential equations and geometry.
- Symbolic algebra measured against FORM 5 every week; gcd level with FORM, factorisation 4 to 8 times slower.
- Arbitrary precision against mpmath to 48 digits; series against SymPy on 32 expansions.
- Spectra to 30–38 digits (the quartic oscillator, hydrogen); perturbation series to high order against Bender and Wu.
- Holographic QCD: seven computations from the action to the result, checked against their papers (below).
Large-scale symbolic computation,
measured against FORM.
PQCF computes with symbols as well as numbers. Example: the exact transfer matrix of an interferometer whose every angle is a parameter, and the exact gradients that optimise it. Our goal is to support large-scale symbolic computation, with expressions of millions of terms, not only the circuit-sized ones the framework needs today.
We measure ourselves against FORM, the system of J.A.M. Vermaseren and colleagues for very large symbolic calculations in particle physics (version 5, January 2026). We run the same reference tasks in both systems every week until the end of the project. Each run checks that the answers agree term for term, and the results are public: the latest run, the history and the method.
Where we stand: 14 November 2026. Two kinds of work matter for a symbolic engine: expanding large expressions, and compiling them into programs that evaluate them fast under many parameter values. We measured both against FORM 5. All times are single-threaded, on a shared development machine; times vary from run to run by up to about a third, FORM's too.
1. Expansion: the weekly reference tasks.
- Dense polynomial products (Fateman's benchmark and a dense product): 3 to 5 times faster than FORM. Fateman's f·(f + 1) with n = 15, 46,376 terms: 0.65 seconds against FORM's 3.22. On 5 November it took 33 seconds.
- Symbolic permanents, the building block of boson sampling: the 7 × 7 in a third of FORM's time, the 8 × 8 (40,320 terms) in the same time as FORM, 0.19 seconds; the 6 × 6, a few milliseconds either way, 2.8 times FORM's time. On 7 November they took 0.6 to 2.6 times FORM's time.
- Our own workload, the symbolic transfer matrix of an interferometer with 4 to 8 modes (up to 276,800 terms): 1.5 to 2.4 times FORM's time in this run, 1.1 to 1.5 on 7 November; measured old against new under the same load, the 8-mode case got 1.5 times faster, so the higher ratio here is the machine's load.
- Sparse products (Monagan and Pearce, up to 591,235 terms): 1.5 to 1.7 times FORM's time, from 2.2 to 2.7 on 7 November and 25 to 32 on 5 November.
- Polynomial powers: 1.7 to 2.8 times FORM's time, from 3.6 to 4.3 on 7 November and 10 to 25 on 5 November.
How: the polynomial arithmetic moved from Erlang into Rust, and large polynomials stay in Rust between operations, so only the final result is built as Erlang data. Every answer is still checked to be identical, term for term, to FORM's and to the earlier Erlang code.
What changed on 14 November. Profiling the reference tasks showed that, for powers and sparse products, the arithmetic in Rust was the smaller part; most of the time went into bringing the result back into Erlang.
- The result comes back as a handle, not copied out of the protected worker; its size is counted in Rust before it is built, so the size limit holds exactly as before; and it is built faster.
- Powers of small polynomials multiply by merging sorted terms instead of hashing.
- The symbolic permanent runs in Rust.
Old and new, run one after the other on the same core under the same load: powers 1.6 to 2.2 times faster, sparse products 1.6 to 1.9, permanents 2.9 to 5.3, the 8-mode interferometer 1.5. The same answers, exactly; PQCF's 408 tests pass.
The ratios above are from one run, with FORM and PQCF measured together on core 8. The shared machine was loaded during it (load average about 7), which made both slower than in the morning's run; the next weekly run will show the ratios under normal load.
2. Compiled programs: FORM's own benchmarks. FORM's authors measure expression simplification on resultants: polynomials of 2,562 to 43,166 terms that, written out, take 29,163 to 587,880 operations to evaluate. FORM 5 has four levels of optimisation, O1 to O4. On 5 November PQCF's compiled programs needed 4.2 to 5.8 times the operations of FORM's best level, O4. With PQCF 1.3.0 they need fewer operations than O4 on all three:
- res(7,4): 2,741 operations against O4's 2,936.
- res(7,5): 10,243 against 10,722.
- res(7,6): 34,799 against 35,419.
How we got there, step by step, each measured against FORM:
- Horner schemes, as FORM uses: the polynomial is nested by one variable after another, so each power is computed once.
- Shared subexpressions: every part that appears more than once, also with another sign or numeric factor, is computed once. With this, PQCF already passes FORM's O2 at the default order.
- A search over variable orders. The order in which variables are nested decides most of the cost. PQCF searches it with a Monte Carlo tree search, as FORM's O3 does, running the candidates in parallel across cores, or across machines, and giving the same answer on any number of them. Each candidate is scored by counting its operations in Rust, without building the program.
One test settled where the remaining gap was. FORM can print the order its O3 chose; with that order, PQCF's scheme needed fewer operations than FORM's own result. So the work went into the search, not the scheme.
Where FORM is still ahead: time. FORM's O4 reaches its result in 2.6 to 57 seconds on one core; PQCF's search takes 22 to 335 seconds on 20. Programs are compiled once and evaluated many times, which is what PQCF needs them for, but the gap is real.
Not compared here: FORM 5's own paper tests Feynman-integral reductions through FORM's packages. PQCF now reduces Feynman integrals in its own way, measured against FIRE, a dedicated program for them: see Reducing Feynman integrals.
Polynomial gcd and factorisation, added on 14 November. Until then these were outside our engine's scope. They are now measured against FORM 5 with FLINT, on the problems of T. Ueda's benchmark suite (polybench, 5 variables), and the answers are identical term for term:
- Greatest common divisor: level with FORM. 10 problems with a common factor take 0.083 s against FORM's 0.075 s; 10 with only a monomial in common, 0.052 s against 0.044 s.
- Factorisation: slower than FORM. Products of two factors take about 8 times FORM's time (1.02 s against 0.13 s for 5 problems); polynomials that do not factor, about 4 times (0.077 s against 0.019 s).
Both are now weekly reference tasks of the comparison. They are the foundation of the work on holographic QCD.
Two routes. PQCF's own engine is exact and bounded, built for safety inside a running service. For sizes beyond it, PQCF can hand polynomial work to FORM itself: FORM runs as a separate program, and its results are checked to be identical to PQCF's own. FORM's parallel version and its expressions stored on disk reach far beyond what we measure here.
The method, the latest run and the full history: comparison/form; the release notes: PQCF 1.3.0, and the change of 14 November in the changelog.
How to reproduce the results: a step-by-step page from a fresh copy of the repository: one script builds FORM at the version we measured and runs both programs on the same core, about 4 minutes the first time, 80 seconds after. We reran it from a fresh copy of the public repository on 15 November; every check passed.
Holographic QCD in PQCF,
from the action to the spectrum.
FORM is built for one kind of computation: very large polynomials in perturbative quantum field theory. Other parts of theoretical physics need other tools: curvature of a metric, equations of motion from an action, series with logarithms, special functions, eigenvalue problems solved to many digits. We added these to PQCF's symbolic engine and tested them on holographic QCD, models in which the mesons and glueballs of the strong force are computed from a five-dimensional gravity theory.
Every result below is computed by PQCF from the model's action and checked against the published paper, an exact formula, or an independent program (SymPy, mpmath, SciPy). The code, the results and the method are public: comparison/holographic.
What was added to the engine.
- Algebra: multivariate polynomial gcd, factorisation and rational functions in normal form, measured against FORM (see Large-scale symbolic computation).
- Analysis: series with fractional powers and logarithms, with Bessel, Laguerre, hypergeometric and Gamma functions, checked against SymPy on 32 expansions; and arithmetic to any number of digits, checked against mpmath to 48 digits.
- Geometry: curvature of a metric in any dimension, and canonical forms of tensor expressions, checked against SymPy's Butler–Portugal algorithm.
- Equations: equations of motion from an action, their reduction to Schrödinger form, and eigenvalue solvers to 25–30 digits; and a high-precision integrator for nonlinear differential equations.
The results.
| model | what PQCF computes | against | agreement |
|---|---|---|---|
| Soft wall, arXiv:hep-ph/0602229 | meson masses m² = 4, 8, 12, 16 from the action | the exact m² = 4(n + 1) | 30 digits, by two independent solvers |
| Hard wall, arXiv:hep-ph/0501128 | meson masses; the a₁, decay constants, f_π, m_π for the paper's two fits | zeros of the Bessel function J₀; the paper's Table II | 30 digits; Table II to its rounding, except f_π (1.6–2.1% lower: we use zero quark mass, explained) and one a₁ mass (0.35% lower, equal to an independent SciPy solution) |
| Sakai–Sugimoto, arXiv:hep-th/0412141 | vector and axial mesons λ = 0.66931, 1.56877, 2.87432, 4.54610 | the paper's 0.67, 1.6, 2.9, 4.5 | agrees; to 25 digits, an independent SciPy solution agrees to its accuracy |
| Black-brane free energy, arXiv:hep-th/0002230 | holographic renormalisation from the metric up: the divergences cancel | F = −π²N²T⁴/8 | exact |
| Weyl anomaly, arXiv:hep-th/9806087 | the near-boundary expansion on two curved boundaries | a = c = N²/4 | exact |
| Gubser–Nellore, arXiv:0804.0434 | speed of sound c_s² against temperature for the paper's potential | the paper's Δ = 3.93, its limits 1/3 and 0.15, its figures | agrees; the dip at T_c (c_s² = 0.045) as in the figures |
| Improved holographic QCD, arXiv:0707.1349 | scalar and tensor glueball mass ratios | the paper's ratios | agrees: the first four scalar and four tensor glueballs within 0.6%, at 12 digits (corrected on 15 November, below) |
A difference we found and corrected. For improved holographic QCD (Gürsoy, Kiritsis and Nitti, arXiv:0707.1349; the framework in arXiv:0707.1324), we first published glueball mass ratios 6–15% lower than the paper's, and said we had not found the reason. The error was ours.
- The cause: one parameter of the model, a. The paper uses the symbol in two ways. In its large-coupling limit, a = 3/16 gives linear confinement; but its interpolating formulas, the ones that are integrated, need a = 3/8 for the same behaviour. Our script used 5/16, which matches neither.
- With a = 3/8: the first excited scalar over the lowest is 1.875 against the paper's 1.866; the lowest tensor over the lowest scalar is 1.394 against 1.393. All ratios of the first four scalar and four tensor glueballs agree within 0.6%.
- What we checked: the background confines linearly, as the paper requires; the ratios are identical, to four decimals, for four starting values, including the paper's own; a longer integration changes no mass; the first 12 scalar masses squared rise in nearly equal steps, as in the paper's figure; fitting the model's parameter b₀ to the paper's target gives 4.15 against its 4.2.
- What remains open: the last 0.6%, which is more than the rounding of the paper's numbers. This computation runs at 12 digits, so it is a numerical result at that precision, unlike the 25–30-digit checks above. A study at higher precision with an independent solver is next.
- Details: IHQCD.md, which keeps the first, wrong run as a record.
How to reproduce the results: Docker and the repository are enough. Each computation is one command, from seconds to four minutes; improved holographic QCD takes 40 minutes on 16 cores, or 4 minutes for a quick check of its two main ratios. We ran every one from a fresh copy of the public repository, and each gave the same numbers.
Where it can be checked: soft and hard walls, hard-wall mesons, Sakai–Sugimoto, black-brane free energy, Weyl anomaly, Gubser–Nellore, improved holographic QCD. The plan behind it is decision record SY2.