v0.1.4

Fixed

• MR, MRX, and MRY no longer double-count their measurement flip probability as both a pre-measurement Pauli error and a measurement-result flip. (#124)
• Fixed a bug where M(p) instructions incorrectly flipped the qubit, not just the measurement record. This would affect circuits where qubits were measured and not immediately reset. (#138)
• Fixed a bug where MR !q instructions (with measurement record inversion) produced wrong measurement results. (#138)
• Out-of-order OBSERVABLE_INCLUDE indices now produce the correct sampler column order and output shape. Missing indices below the maximum mentioned id appear as deterministic-zero columns, and columns are emitted in sorted logical-index order. (#127)
• matmul_gf2 no longer silently corrupts parity for inner products with more than 255 set bits. The float32→uint8 cast in JAX saturates at 255, which previously made % 2 always return 1 once a row-sum reached 256. The modulo is now applied on float32 before the uint8 cast. (#130)
• CompiledDetectorSampler.sample now raises ValueError when separate_observables=True is combined with prepend_observables=True or append_observables=True (matching Stim). Previously these combinations silently dropped observable columns. The prepend_observables=True + append_observables=True combination is now supported and returns columns in [obs, det, obs] order, matching Stim. (#134, #140)
• OBSERVABLE_INCLUDE with Pauli targets (e.g. OBSERVABLE_INCLUDE(0) X1) now raises ValueError instead of silently producing wrong observable bits or crashing with IndexError. tsim only supports measurement-record targets (rec[-k]). (#140)
• Empty DETECTOR and OBSERVABLE_INCLUDE annotations (without targets) no longer crash the parser; they now produce zero detector/observable bits, matching Stim semantics. (#127)
• CompiledDetectorSampler.sample with use_detector_reference_sample=True or use_observable_reference_sample=True no longer returns fewer rows than shots when called with an explicit batch_size that exactly divides shots. (#131)
• Incorrect visualization of CORRELATED_ERROR and ELSE_CORRELATED_ERROR instructions in the pyzx diagram renderer. Previously, error vertices were rendered as classical spiders instead of bold quantum spiders. (#133)
• Sweep-bit targets (e.g. CX sweep[0] 1) were silently parsed as unconditional gates. The parser now raises NotImplementedError, since sweep parameters are not supported by Tsim. (#136)

Added

• TPP and TPP_DAG instructions — applies exp(-i pi/8 P) or exp(+i pi/8 P) (up to global phase) for a Pauli product P, i.e., phases the -1 eigenspace of P by exp(i pi/4) or exp(-i pi/4). (#121)
• Circuit.is_clifford now supports REPEAT blocks. (#128)

v0.1.3

Fixed

• DEPOLARIZE2 channel was missing the p_ZZ probability term, which was always set to 0. This lead to incorrect noise models that were missing ZZ errors. (#103)
• Samplers now gracefully handle circuits with no measurements or no detectors, returning empty (shots, 0) arrays matching stim's behavior instead of raising an error (#106)
• MPP, MXX, MYY, MZZ instructions now support a bit flip probability argument (#118)

Added

• Zoomable timeline and timeslice diagrams. Circuit.diagram now accepts a zoomable option, enabled by default, to support pan and zoom in notebooks for the timeline-svg and timeslice-svg diagram types (#116)
• HERALDED_PAULI_CHANNEL_1 and HERALDED_ERASE noise channel instructions with herald bit indicating whether the noise event occurred (#107)
• CXSWAP, CZSWAP, SWAPCX, SWAPCZ two-qubit gate instructions (#105)
• C_NXYZ, C_XNYZ, C_XYNZ, C_NZYX, C_ZNYX, C_ZYNX axis-cycling gate variants with negated axes (#105)
• H_NXY, H_NXZ, H_NYZ Hadamard-like gate variants with negated axes (#105)
• II two-qubit identity instruction that acts trivially (#105)

Changed

• I_ERROR, II_ERROR, and QUBIT_COORDS instructions now allocate qubit lanes instead of being silently skipped (#105)

v0.1.2

Fixed

• Exact scalar reduction during sum/product operations to prevent underflows/overflows of int32 on large diagrams. Unfortunately, this change comes with a 2x performance overhead, but results in more stable numerical results (#93)
• Normalization issues for circuits with arbitrary rotation gates now raise a warning instead of an error (#91)
• Parsing errors for invalid Stim circuits now raise useful exceptions (#91)

Added

• SPP and SPP_DAG instructions — generalized S gate that phases the -1 eigenspace of Pauli product observables by i or -i. Supports multi-qubit Pauli products and inverted targets (#97)
• MXX, MYY, MZZ two-qubit parity measurement instructions, delegating to existing MPP infrastructure. Also adds II_ERROR support (#96)
• MPAD instruction for padding the measurement record with fixed bit values (#95)

v0.1.1 🚀 🌕

This release adds new stabilizer decomposition strategies, faster noise sampling, and scientifically accurate pyzx visualization. A detailed description of Tsim, including some benchmarks, is now available in this publication.

Added

• Improved stabilizer decomposition strategies. When compiling a sampler, you can now choose between three different strategies: "cat5", "bss", and "cutting". The default is "cat5" and applies to T and arbitrary rotations; see arxiv.org/abs/2106.07740 (#77)
• Sparse geometric channel sampler for noise modeling based on this repo. This significantly improves performance when the stabilizer rank is low. (#64)
• Circuit.append method for programmatic circuit construction (#65) by @jaideepkathiresan
• Circuit.is_clifford property and automatic replacement of U3 gates with Clifford equivalents for pi/2 rotations (#69)
• Improved pyzx visualization. Now doubled ZX notation is used when using the "pyzx" argument in Circuit.diagram, which is a technically accurate depiction of the quantum circuit (#86)
• Automatic batch size selection based on available memory (#84)

Changed

• Tsim now uses pyzx-param==0.9.3 which fixes a bug where diagrams were not fully reduced in the absence of noise
• Tsim will now make sure that marginal probabilities are normalized and raise an error if they are not. Wrong normalization can be the result of rare underflow errors that will be addressed in a future release (#87)
• Use BLAS matmul kernel for tensor contractions (#63)
• Circuit flattening deferred to ZX graph construction time (#71)
• White background for SVG plots, which are now readable in dark mode (#85)

New Contributors

• @jaideepkathiresan made their first contribution in #65

v0.1.0

Tsim is a GPU-accelerated quantum circuit sampler based on ZX-calculus stabilizer rank decomposition. It feels just like Stim, but supports non-Clifford gates.

Tsim is described in arXiv:2604.01059.

Highlights

• Clifford+T circuit simulation via stabilizer rank decomposition, following Sutcliffe and Kissinger (2024)
• Full Stim compatibility -- supports all Stim instructions, plus T, T_DAG, R_Z, R_X, R_Y, and U3
• Noise channels via parametric rewrites
• Arbitrary rotation gates via magic cat state decomposition from Qassim et al. (2021)
• GPU acceleration via JAX
• Documentation and tutorials at queracomputing.github.io/tsim

Installation

pip install bloqade-tsim

For GPU support:

pip install "bloqade-tsim[cuda13]"