Checking five-loop form-factor symbol coefficients
Compare a sample of exact coefficients in two compressed representations of the published five-loop symbol.
Read the same word two ways
The five-loop three-gluon form-factor symbol has words of length ten in a six-letter alphabet. The reference file Esymb.txt stores coefficients in blocks indexed by their final letters. The four-position and eight-position encodings overlap. This lets a reader compare the coefficient of the same word in two tables.
For example, split aabbcbdddd after six letters to read the four-letter block, or after two to read the eight-letter block. The two lookups both return -1/4 in the reference normalization. A missing monomial in a block contributes zero.
These reference tables come from the Dixon, Gürdoğan, McLeod and Wilhelm bootstrap calculation, in the discussion of ancillary files at the end of §4. The learning paper uses a different suffix representation for its prediction task.
Compare two suffix blocks
Fix each of the two eight-letter suffixes, bbcbdddd and dddddddd. Enumerate all 36 ordered two-letter prefixes for each, including prefixes whose coefficients vanish. Read the raw rational coefficients in both representations and compare them exactly.
| Word comparison | Count |
|---|---|
| Tested | 72 |
| Nonzero in both | 15 |
| Zero in both | 57 |
| Different coefficients | 0 |
Reversing the sign of the example coefficient in just the eight-position table introduces one disagreement, at aabbcbdddd. As a separate check, the alphabet reflection interchanging b with c and e with f maps bddddddddd to cddddddddd. The former predicts a coefficient of -26880 for the latter, matching its stored value.
Re-extract the inputs
The small input file contains the two selected octuple blocks and the corresponding slices of the quadruple block dddd. To check it against the source, obtain the versioned Esymb.txt from the reference paper’s ancillary materials and run python3 -I check.py /path/to/Esymb.txt. The script checks the source file’s SHA-256 digest and compares its extracted values with the bundled inputs.
This calculation tests two representations on 72 words and one reflection on a pair of words. It does not run a transformer, reproduce its search, measure prediction accuracy, or establish the full five-loop symbol.
Repeat the calculation
Download the reference inputs (JSON) alongside the script and recorded output.
Save the Python script and recorded output (JSON) in the same directory. The site ran the script with CPython 3.12.14 and its standard library. The calculation uses exact arithmetic and completes in under a second.
python3 -I check.py > rerun.json
cmp result.json rerun.jsonA successful comparison exits with status zero. The files above document the inputs and individual checks.
Sources and interpretation
Learning Scattering Amplitudes with Transformer Reinforcement Learning. arXiv:2609.32096v2. §2 and Eq. (1), observable and symbol alphabet. §§3–4, compact suffix basis, propagation, transformer policy and tree search. §5 and Fig. 2, five-loop reconstruction and six-loop baseline.
Read the research entry for the role of AI in the original work and the range of its conclusions.