Supported Gates¶

Clifft parses Stim circuit format and supports nearly all Stim gates, plus non-Clifford extensions.

Pauli Gates¶

All Pauli gates are single-qubit Cliffords absorbed at compile time, so they do not become active-state actions unless a later dynamic dependency requires one.

Single-Qubit Clifford Gates¶

All single-qubit Cliffords are absorbed AOT — they update the Clifford frame \(U_C\) at compile time and have zero cost at runtime.

Pauli-Product Clifford Gates¶

These gates accept the same product syntax as MPP, including multiple whitespace-separated products in one instruction. Prefixing a Pauli term with ! negates the product and reverses the corresponding phase gate. A qubit may appear only once in each product, which is stricter than Stim for some syntactically valid Hermitian products.

The frontend absorbs SPP and SPP_DAG into the Clifford frame during tracing, so they have no runtime cost. SPP Z0 and S 0 represent the same projective Clifford operation.

Non-Clifford Extensions¶

Clifft extends Stim with discrete and arbitrary-angle non-Clifford gates. These operations can add active stabilizer coordinates and expand the active state vector.

T Gates¶

Pauli-Product Phase Gates¶

These gates accept the same product syntax as MPP, including multiple whitespace-separated products in one instruction. Prefixing a Pauli term with ! negates the product and reverses the corresponding phase gate. A qubit may appear only once in each product, which is stricter than Stim for some syntactically valid Hermitian products.

TPP and TPP_DAG emit one generalized T operation per product. TPP Z0 and T 0 represent the same projective operation.

Rewrite Gates¶

These gate names are accepted by the parser as fixed rewrite rules into Clifft-native gates. They are rewritten during parsing and do not appear as distinct frontend, planner, or executor gate types.

Continuous Rotations¶

Clifft extends the Stim gate set with arbitrary-angle rotation gates. All angle parameters are in half-turns (multiply by pi to get radians).

Single-Qubit Rotations¶

Finite angles within an absolute tolerance of 1e-12 half-turns of a multiple of 0.5, such as 0.5, -1, and 1.5, are canonicalized to that Clifford rotation during tracing. This policy intentionally absorbs small numerical errors introduced by circuit generation or serialization; a deliberate overrotation smaller than the tolerance is therefore treated as the canonical Clifford. Deviations outside the tolerance remain continuous rotations. Each U3 component receives the same treatment.

Two-Qubit Pauli Rotations¶

Duplicate target qubits (e.g. R_XX(0.5) 3 3) are rejected at parse time.

Multi-Qubit Pauli Rotation¶

The target list uses Stim's Pauli product syntax (e.g. X0*Y1*Z2). Maximum target count is 64 qubits per instruction.

For two- and multi-qubit Pauli rotations, finite angles within the shared 1e-12 half-turn tolerance of a Clifford angle are absorbed into the Clifford frame during tracing. This includes square-root Pauli rotations at half-integer angles and projective Pauli gates at odd-integer angles. The peephole pass applies the same rule to HIR rotations produced directly or by fusion.

Two-Qubit Clifford Gates¶

Two-qubit Cliffords are also absorbed at compile time.

Measurements and Resets¶

Multi-Qubit Measurements¶

Noise Channels¶

READOUT_NOISE flips already-recorded bits rather than acting on a qubit, so its targets are measurement-record references. With one argument the flip is symmetric; with two, a recorded 0 flips with probability p01 and a recorded 1 with probability p10. Record targets do not take the ! inversion marker — swap the two probabilities instead. Noisy measurements (e.g., M(0.01) 0) are parser shorthand for a clean measurement followed by READOUT_NOISE(0.01) rec[-1].

DEPOLARIZE3(p) a b c applies one of the 63 non-identity Pauli products on a,b,c with probability p/63 each, and identity with probability 1-p. PAULI_CHANNEL_3 uses the same lexicographic Pauli order as PAULI_CHANNEL_2, extended to three qubits: IIX, IIY, IIZ, IXI, IXX, ..., ZZZ.

CORRELATED_ERROR(p) X0 Z1 applies the listed Pauli product with probability p. Pauli terms may be whitespace-separated or combined with *; all Pauli targets on the instruction form one product. Repeated terms on the same qubit multiply modulo Pauli phase, so E(1) X0 Z0 is equivalent to a Y error and E(1) X0 X0 is an identity event.

ELSE_CORRELATED_ERROR(p) must immediately follow CORRELATED_ERROR or another ELSE_CORRELATED_ERROR. Its p is conditional on no earlier link in the chain firing. Clifft lowers each contiguous chain to one noise site with absolute channel probabilities.

Leakage and Loss Annotations¶

All three are recognized only by the leakage/loss sampler — clifft.compile() rejects them and points to clifft.noncomp.sample(). See the Leakage and Loss guide.

Identity Gates¶

These are accepted for compatibility with Stim circuits but have no effect.

Annotations and Control Flow¶

OBSERVABLE_INCLUDE currently supports rec[-k] measurement-record targets. Stim also permits Pauli-term targets on OBSERVABLE_INCLUDE, which Clifft does not currently parse.

MPAD(p) is accepted; the optional probability noisily flips the padded measurement-record bits.

Expectation Value Probes¶

EXP_VAL evaluates the expectation value of one or more Pauli products at the exact point in the circuit where it appears. It uses the same Pauli product syntax as MPP — multiple whitespace-separated products per instruction are supported, each producing one float64 result in [-1, 1].

H 0
EXP_VAL X0          # single Pauli: <X> on qubit 0
EXP_VAL X0*Y1*Z2    # multi-qubit product
EXP_VAL X0*X1 Z0*Z1 # two products in one instruction

Results are available via SampleResult.exp_vals (shape (shots, num_exp_vals)).

Not Yet Supported¶