Qinara News · 21 October 2026

Qinara releases ACTIVE, an open-source observable-conditioned quantum circuit reduction tool using NVIDIA CUDA-Q

Independent evaluation with NVIDIA CUDA-Q and results from an IBM Heron r2 processor accompany the open-source release.

Qinara
Quantum software
Open source

We are releasing ACTIVE, an open-source observable-conditioned quantum circuit reduction tool from Qinara.

ACTIVE propagates the target observable backwards through the circuit and removes operations that cannot contribute to the requested expectation value. Because the observable is tracked algebraically, its support can contract during propagation, exposing additional upstream operations for elimination while retaining the original target observable.

Observable-conditioned reduction

ACTIVE propagates the target observable backwards through the circuit and removes operations that cannot contribute to the requested expectation value. The observable is tracked algebraically, so its support can contract during propagation and expose additional upstream operations for elimination while the original target observable is retained.

40-case benchmark snapshot

Original vs ACTIVE-reduced circuits

63.0%

mean reduction in total gates

64.3%

mean reduction in two-qubit gates

52.2%

mean reduction in circuit depth

8.88×10⁻¹⁶

maximum independent original/reduced expectation difference

40/40

independent target-observable checks agreed to numerical precision

40/40

cases with Pauli-term count unchanged

Heisenberg-picture observable propagation used for gate relevance

Diagram of backward observable propagation through gates U1, Uj and Um, showing when a gate is removable for the expectation value of O

Independent evaluation with NVIDIA CUDA-Q

Independent execution and observable-validation path

NVIDIA CUDA-Q , the open platform for hybrid quantum-classical supercomputing, was used as an independent execution and validation environment for both the original and ACTIVE-reduced workloads, outside ACTIVE's own verification stack. The same target observable was evaluated on both circuit variants, providing a separate execution path for checking preservation of the requested expectation value.

CUDA-Q was also used for GPU-accelerated simulation and repeated observable evaluation across the benchmark suite, making it the primary independent execution environment in the validation workflow rather than an internal component of ACTIVE's reduction algorithm.

Matched observable evaluation in NVIDIA CUDA-Q

Flow diagram comparing an original circuit and an ACTIVE-reduced circuit, each evaluated in CUDA-Q against the same observable — CUDA-Q provided an execution path separate from ACTIVE's internal verification stack. Original and ACTIVE-reduced circuits were evaluated against the same target observable; the independent benchmark cross-check agreed to numerical precision in all 40 cases.

40-case benchmark results

Bar chart of mean logical reduction in total gates, two-qubit gates, and depth for eight circuit families — Mean logical reduction by circuit family across the supplied 40-case benchmark suite.

Scatter plot of ACTIVE expectation versus original expectation across 40 cases, lying on the line y equals x — Original versus ACTIVE target-observable expectation values across all 40 benchmark cases.

Selected 8-qubit circuit pairs

Logical gate and depth reductions for eight selected 8-qubit circuits
Circuit	Qubits	Total gates	Δ gates	2Q gates	Δ 2Q	Depth	Δ depth
Support contraction	8	28 → 4	85.7%	4 → 0	100.0%	4 → 1	75.0%
GHZ / local	8	29 → 1	96.6%	7 → 0	100.0%	11 → 1	90.9%
QAOA ring	8	72 → 35	51.4%	32 → 15	53.1%	51 → 15	70.6%
HEA linear	8	69 → 20	71.0%	21 → 5	76.2%	12 → 9	25.0%
QFT-like	8	40 → 2	95.0%	32 → 1	96.9%	16 → 2	87.5%
Ising chain	8	95 → 40	57.9%	42 → 17	59.5%	37 → 21	43.2%
Random scrambler	8	80 → 64	20.0%	16 → 15	6.2%	15 → 13	13.3%
Local Hamiltonian	8	54 → 25	53.7%	14 → 7	50.0%	14 → 10	28.6%

Original 8-qubit support-contraction circuit — Support contractionOriginal

ACTIVE-reduced 8-qubit support-contraction circuit — Support contractionActive

Original 8-qubit QAOA ring circuit — QAOA ringOriginal

ACTIVE-reduced 8-qubit QAOA ring circuit — QAOA ringActive

Results from IBM Quantum

Eight matched 8-qubit original / ACTIVE circuit pairs

IBM Heron r2 processor

Same backend, Qiskit optimisation level, transpiler seed and matched initial physical layouts.

61.9%

mean reduction in total ISA gates

67.2%

mean reduction in ISA two-qubit gates

52.8%

mean reduction in ISA circuit depth

55.9%

mean reduction in estimated single-shot circuit duration

6 of 8 hardware cases had lower point-estimate absolute error against the independent exact statevector reference.

6.25% ISA two-qubit reduction on the random-scrambling negative control.

We executed eight matched 8-qubit original/ACTIVE circuit pairs on an IBM Heron r2 processor, using the same backend, Qiskit optimisation level, transpiler seed and matched initial physical layouts. After transpilation on an IBM processor, the aggregate reductions above were observed.

The random-scrambling negative control showed only 6.25% ISA two-qubit reduction, consistent with ACTIVE exploiting observable/circuit structure rather than producing uniform reduction across arbitrary circuits.

Bar chart of mean ACTIVE reductions on an IBM Heron r2 processor for total ISA gates, ISA two-qubit gates, ISA depth, and estimated single-shot duration — Mean post-transpilation reductions across eight matched 8-qubit original/ACTIVE circuit pairs on an IBM Heron r2 processor.

Fault-tolerant quantum computing

When fault-tolerant quantum computing (FTQC) is achieved, ACTIVE will operate before synthesis and QEC, so operations removed at the logical level will be absent from the downstream FT compilation problem.

Projected downstream effects of logical reduction

T-count

non-Clifford synthesis

T-depth

logical critical path

Magic states

consumption / distillation

QEC cycles

logical depth / scheduling

For non-Clifford reductions, this can reduce T-count, T-depth, synthesis demand and magic-state consumption. Reductions in logical entangling structure and active support can also propagate into lower logical depth, scheduling pressure and QEC cycle requirements.

We are evaluating these effects through the NVIDIA CUDA-Q fault-tolerant workflow, comparing downstream resource estimates before and after ACTIVE under matched FT compilation assumptions.

ACTIVE will be released open source together with the methodology, benchmark circuits, CUDA-Q validation and results on IBM Quantum computers.