arXiv:2609.25947
Bridge of 's: Quantum Circuit Optimization with Schrödinger Bridges
Lino S. Hofstetter, Lia Yeh, Prakash Murali
Abstract
Quantum circuit optimization replaces a circuit with an equivalent one of fewer gates and lower depth, reducing execution cost and error rate. We ask whether a generative model can learn this transformation directly from examples, rather than selecting from a fixed rewrite library or rigid algebraic routines. We present Bridge of $Ψ$'s (BOPS), a generative model based on Schrödinger bridges, using a custom denoiser architecture, that learns a transformation from a source circuit into an equivalent optimized circuit. We train it on data constructed to be hard for existing optimizers, by applying rewrite rules backwards so that each input has a known lower-cost target. On held-out 8 qubits $\times$ 64 depth Clifford+$T$ circuits, BOPS reduces gate count by $2.46\times$ and depth by $2.45\times$ in geometric mean, outperforming all nine baseline optimizers. This constitutes the first generative model bridging quantum circuits and frontier machine learning methods, opening up the quantum compilation stack to learned optimization along multiple axes.
AI Summary
Can a generative model learn to replace a quantum circuit with an equivalent, lower-cost circuit rather than select predefined rewrites? Bridge of $\Psi$’s (BOPS) formulates Clifford+$T$ optimization as a source-conditioned Schrödinger bridge. A 98.8-million-parameter U-shaped diffusion transformer transports circuit-grid embeddings toward optimized equivalents, using time-only downsampling and grouped attention. Training uses roughly 1.95 million pairs, generated primarily by reversing verified one- to three-qubit rewrites, including nested expansions. Equivalence checks are excluded from training; inference samples candidates and accepts only verified improvements, otherwise retaining the input. On 2,686 held-out rule-generated circuits fitting an $8\times64$ grid, whole-circuit inference achieves geometric-mean gate-count and depth reductions of $2.28\times$ and $2.26\times$. Adding windowed optimization raises these to $2.46\times$ and $2.45\times$, versus $2.11\times$ for both metrics from the strongest tested baseline, QUESO. Windowed BOPS reaches or beats the known target gate count on 77.2% of inputs, versus 26.1% for QUESO. Finetuning to $16\times192$ yields verified improvements on 98.4% of inputs with windowing. The strongest results concern the synthetic training distribution. On 68 external benchmarks, using ten times the main inference budget, BOPS achieves $1.13\times$ gate reduction versus QUESO’s $1.26\times$, while matching the best aggregate depth reduction at $1.08\times$. Whole-circuit performance deteriorates sharply with length; windowing misses cross-window reductions. Verification can remain undecided at 16 qubits. Inference is also costlier: $14.4$ seconds per windowed $8\times64$ input on a GH200, versus roughly $0.03$–$0.04$ seconds for pattern-based CPU baselines.
Error found: The claim of outperforming all nine baseline configurations includes a GUOQ run with infrastructure failures on 21.6% of inputs, so superiority over a fully functioning GUOQ run is not established. The abstract's broad first-generative-model claim conflicts with its acknowledgement of prior diffusion-based quantum-circuit synthesis, whereas the narrower source-conditioned Schrödinger-bridge novelty claim is not contradicted by the cited discussion.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 25 pages, 11 figures, 12 tables