splitenum enumerates every compatible collection of Harnack splits of the degree-d simplex and records the real scheme of the T-curve each collection produces. By Haas' theorem a T-curve is maximal exactly when its signs arise from Harnack's by such a collection, so every node of the search is an M-curve and the output is the exact distribution of nodes over the M-schemes realized this way. All isotopy computations are done by libisotopy.
Zoe Geiselmann, Michael Joswig, Lars Kastner and Marcel Wack. Complete enumeration of maximal T-curves in degrees eight and nine. FPSAC 2027. arXiv:XXXX.XXXXX
Related work:
- 121 Patchworked Curves of Degree Seven, arXiv:2602.06888, data at dmg-lab/CombinatorialPatchworking.
- Fast Isotopy Computation for T-Curves, ICMS 2026, arXiv:2604.09221, the algorithm behind libisotopy.
| Degree | Splits | Nodes | M-schemes | Data |
|---|---|---|---|---|
| 5 | 78 | 7,915 | 1 | |
| 6 | 222 | 1,035,083 | 3 | |
| 7 | 444 | 296,740,751 | 14 | |
| 8 | 873 | 105,949,232,275 | 38 | statistics, witnesses |
| 9 | 1,437 | 57,065,860,838,381 | 225 | statistics, witnesses |
Requires a C++20 compiler with OpenMP. MPI and CUDA are needed only for their targets.
git clone --recurse-submodules <this repository>
cd splitenum
make # libisotopy, then ./splitenum
make test # test suite
make tools # bin/witnesses, bin/pcomcheck and the other tools
make mpi # ./splitenum-mpi, needs mpicxx
make gpu # ./splitenum-gpu, needs nvcc; CUDA_ARCHS="75 86" selects targetsARCH defaults to -march=native; build with ARCH= for binaries that run on
other machines.
MAX_DELTA, default 8, sizes fixed arrays inside libisotopy and is part of its
ABI. The Makefile passes it to both builds; always change it through make MAX_DELTA=9, which degree 9 needs, and never for one side only.
./splitenum --delta 6 --out out/d6
./splitenum --delta 7 --count-only
./splitenum --delta 8 --scheduler jobs --classify gpu --out out/d8
./splitenum -hA run can be stopped at any time; rerunning the same command resumes it.
--count-only counts the tree without any isotopy computation. With
--scheduler jobs a master hands out budgeted jobs to the worker threads,
--classify gpu sends the classification through the CUDA kernel of
libisotopy, and --listen PORT together with --connect DIR spreads one run
over several machines.
Everything is written to the --out directory:
splits.csv: the split table; witnesses refer to its split ids.types.csv: one row per real scheme and slack profile, with the node count and the shallowest, then lexicographically smallest, witness collection.run.json: parameters and statistics of the run.
bin/witnesses RUN_DIR OUT_PREFIX re-verifies every witness of a finished run
and writes it as a .pcom patchwork file.
GPL-3.0, see LICENSE.