Performance highlights

TenCirPauli is built to keep structured quantum workloads compact from the Python call to the native execution kernel. The research examples below collect representative measured wins against TensorCircuit/JAX, QuSpin, PauliPropagation.jl, OpenFermion, and independent Python baselines.

Headline results

The cards show the strongest headline results. The tables below add the resource and symbolic-algebra comparisons without competing with the main message.

Measurement grouping

The measurement workflow has two distinct timed stages. Grouping builds the compatible measurement partitions and bases; reconstruction consumes already sampled rotated bitstrings and computes the corresponding Pauli eigenvalues. The reconstruction number does not include circuit execution, sampling, or Qiskit's grouping runtime.

Workload	Comparison	Measured result	Research example
64-qubit, 1,024-term local QWC workload; support width at most 4	Qiskit SparsePauliOp.group_commuting(qubit_wise=True)	Grouping: 37.9× faster; both produce 26 groups. Reconstruction is not compared because this runner does not implement that post-processing path.	Measurement-grouping example
Same workload	Independent NumPy/Python largest-first baseline	Grouping: 20.0× faster. Reconstruction: 3.0× faster for all 26 groups and 1,024 synthetic shots; outputs agree exactly.	Measurement-grouping example

Memory wins

Workload	Comparison	Resource result	Research example
6-site eager Holstein sparse execution	QuSpin explicit sparse path	About 6× lower peak RSS; at four threads, the native MVP is also about 1.8× faster	Holstein example
7-site lazy Holstein MVP; 20,070,400-dimensional sector	QuSpin matrix-free execution	1.13× faster with about 2× lower peak RSS	Holstein example
4×4 half-filled Fermi–Hubbard MVP	QuSpin matrix-free execution	About 2.16× lower peak RSS with the same 2.650 GB input/output buffers	Fermi–Hubbard example

Native baseline comparisons

These studies compare the native implementation with small independent Python dictionary prototypes, making the benefit of moving repeated symbolic work into the Rust core visible without presenting a prototype as a competing production library.

Workload	Baseline	Measured result	Research example
Fourth-order Pauli BCH; 8 qubits and 44 terms per generator	Python dictionary recurrence	About 4.7× faster native end-to-end	BCH convergence example
Bounded 64-qubit Pauli Lie closure; dimension cap 512	Python dictionary recurrence	About 7.7× faster native closure	Lie-closure example

Classical-shadow workloads

The default classical-shadow research example uses 20 qubits, 8,192 settings, one repeat, and the JAX CPU backend in batches of 64. The release-mode record below includes the public call and synchronized JAX conversion; it is a single Apple arm64/macOS 15.3.1 run with Python 3.11.15, TensorCircuit-NG 1.9.1, JAX 0.10.2, TenCirPauli 0.5.0, and RAYON_NUM_THREADS=4.

Workload	TenCirPauli	TensorCircuit/JAX	Result
Local-Pauli sampling, dense non-Clifford state, 20q/8,192 shots	19.73 s, 2.409 ms/shot	77.56 s, 9.468 ms/shot	Native 3.93× lower total time
Pauli estimator on 8,192 imported snapshots	67.2 μs	1.855 ms	Native 27.3× lower; same-data path
Global-Clifford sampling, dense non-Clifford state, 20q/8,192 shots	106.78 s, 13.034 ms/shot	—	Standalone native workload; no equivalent baseline claimed
Dense-target fidelity post-processing on the same Clifford snapshots	28.54 s	—	Standalone native workload; dense target has 2**20 amplitudes

The run also sampled a 64-qubit stabilizer input with 128 settings in 19.7 ms. These values are informational rather than universal speed guarantees; the exact command, seed, estimates, and cold JAX setup are recorded in the example README. The complete benchmark suite additionally covers RDM, Rényi-2, energy variance, batched estimators, and both fidelity paths.

All numbers are representative results from the linked studies, not universal maxima. Each example contains the workload definition, correctness checks, and the command used to reproduce the comparison.