TenCirPauli moves Pauli-heavy quantum workflows from Python object graphs into compact native data and Rust execution. It is the TensorCircuit-NG-facing layer for Hamiltonians, measurement grouping, symmetry reduction, restricted sectors, structured operators, classical-shadow snapshots, and observable propagation.
| Capability | Main entry points |
|---|---|
| Pauli algebra and Hamiltonians | PauliWord, PauliOperator, products, commutators, dense/COO/CSR targets, matrix-free MVP |
| Measurement planning | group_commuting(mode="qubit_wise"), QWCGroupingResult, basis-aware bitstring reconstruction |
| Classical-shadow snapshots | Snapshots.sample(), Pauli/global-Clifford protocols, native estimators, RDM, Rényi-2 entropy, and fidelity |
| Static stabilizer-code analysis | StabilizerCode, ordered syndromes, logical representatives, error classification, and decoder-correction verification; distinct from StabilizerState |
| Symmetry reduction | find_z2_symmetries(), taper_z2(), restrict_charge() with U1Sector or ChargeSector |
| Native circuit execution | U1Circuit, PropagationCircuit, SPPSCircuit, value/gradient and expectation terminals |
| Structured quantum operators | FermionOperator, BosonOperator, QuditWeylOperator, HybridOperator, MajoranaOperator |
| Fermion and chemistry workflows | Jordan–Wigner, parity, Bravyi–Kitaev, optional PySCF ingestion, TensorCircuit-NG integration |
- Scalable symbolic work stays in Rust over packed Pauli, occupation, charge, and structured representations.
- Public results are lazy native-backed handles by default; Python materializes terms or arrays only when an API explicitly asks for them.
- Matrix-free and restricted-sector plans avoid dense matrices and full Hilbert spaces when the workload has useful structure.
- TensorCircuit-NG and JAX can remain at the backend boundary through reusable
backend_mvp_plan()plans, while native CPU paths use coarse-grained Rust execution and default parallelism.
These are representative runs from the linked research studies, not universal maxima. The benchmark page collects the workload definitions and runnable examples.
| Workload | Comparison | Result |
|---|---|---|
| 60-qubit, two-particle U(1) VQE | TensorCircuit-NG Python/JAX path | 688× faster on the first compiled value-and-gradient call; 2.6× faster when steady |
| 28-qubit Pauli-propagation VQE | PauliPropagation.jl | 7.8× faster on the first call; 1.43× faster when steady |
| 12-mode fermion mapping | OpenFermion | 22× faster for Jordan–Wigner; 12.6× faster for Bravyi–Kitaev |
| 64-qubit, 1,024-term QWC grouping | Qiskit | 38× faster grouping on the linked workload, with native reconstruction included in the study |
| 4×4 Fermi–Hubbard restricted MVP | QuSpin | 1.73× faster and 2.16× lower peak memory |
| If you need to... | Start with... |
|---|---|
| Build or transform Pauli operators | PauliOperator.from_terms() and PauliWord.from_string() |
| Group terms for measurements | operator.group_commuting(); use QWC mode for measurement-ready bases |
| Estimate observables from randomized measurements | Snapshots.sample() followed by expectation(), estimate_many(), or energy() |
| Analyze a static qubit stabilizer code | StabilizerCode.from_css() / from_generators(); the code space has dimension 2**nlogical, while StabilizerState is a pure-state tableau |
| Apply a large Hamiltonian without a matrix | operator.native_mvp_plan() or operator.mvp() |
| Keep TensorCircuit-NG/JAX active | operator.backend_mvp_plan() and tencirpauli.backend_mvp() |
| Reduce a symmetry sector | operator.find_z2_symmetries() / operator.taper_z2() |
| Work at fixed particle number or additive charge | U1Sector, ChargeSector, and operator.restrict_charge() |
| Run a native circuit objective | U1Circuit, PropagationCircuit, or SPPSCircuit |
| Map or compile structured operators | FermionQubitMapping, MajoranaOperator, and the structured operator classes |
mode="general" is available for algebraic commuting groups, but its result is intentionally measurement_ready=False; use QWC grouping when the output must directly describe product-basis measurements.
TenCirPauli is designed as a TensorCircuit-NG companion, not a replacement for its circuit frontend. TensorCircuit-NG circuits can be converted with U1Circuit.from_circuit() or PropagationCircuit.from_circuit(), and Pauli backend plans can be called through TensorCircuit-NG's NumPy or JAX backend. The Rust core itself has no Python or TensorCircuit-NG dependency.
The research index links to reproducible, manual studies covering Fermi–Hubbard, Holstein, SYK Majorana, BCH convergence, Lie closure, fermion mapping, U(1) VQE, Pauli propagation, measurement grouping, and classical shadows.
python -m pip install tencirpauliReleased wheels cover common CPython 3.10+ platforms. Source builds require Rust, Cargo, and maturin; chemistry interop is optional via python -m pip install 'tencirpauli[chemistry]'.
Read the documentation for concepts and API details, performance notes for benchmark context, and CONTRIBUTING.md for local development. The local quality gate is python scripts/check.py --benchmark smoke.
Apache License 2.0.