arXiv:2610.12447
Entanglement entropy and magic of ZX-diagrams
Marcin Szyniszewski, Razin A. Shaikh, Aleks Kissinger
Abstract
Entanglement and non-stabilizerness are complementary resources underlying the complexity of quantum states, yet extracting either quantity from large quantum circuits generally requires exponentially scaling resources. In diagrammatic approaches to quantum computation, such as the ZX-calculus, both quantities likewise typically require expensive tensor contractions that are not native to graphical rewriting. Here we show that both entanglement and magic can be efficiently estimated directly from ZX-diagrams. Flow enables efficient extraction of a normal form separating a potentially extensively entangled graph-state backbone from non-Clifford Pauli gadgets. This structure yields additive upper and lower bounds on bipartite entanglement entropy, which we further tighten through a preprocessing procedure that removes or merges redundant non-Clifford contributions. The same normal form gives an upper bound on logarithmic stabilizer extent. We benchmark these methods in random unitary and monitored circuits, and a circuit combining Trotterized Hamiltonian evolution with Clifford layers, finding that the resulting bounds remain informative even at large qubit numbers, circuit depths, and internal spider counts. Our results provide scalable probes of entanglement and non-stabilizerness of ZX-diagrams and establish a bridge between ZX-calculus and quantum many-body resource characterization, with applications across quantum computing, quantum information, and condensed-matter physics.
AI Summary
The paper asks whether bipartite entanglement and magic can be estimated from ZX-diagrams without contracting the represented state. For state-preparation diagrams satisfying flow conditions, it uses polynomial-time extraction of a normal form $|\psi\rangle=U_N\cdots U_1|G\rangle$, separating a graph-state backbone from Pauli rotations $U_i=e^{-i\varphi_iP_i/2}$. Conditional on the absence of coherent interference between stabilizer branches, the von Neumann entropy satisfies $|S-S_0|\leq\sum_{i\in\mathcal C}h_2(\sin^2(\varphi_i/2))$, where $S_0=\operatorname{rank}_{\mathbb F_2}\Gamma_{AB}$ and $\mathcal C$ contains cut-crossing gadgets. The small-angle correction scales as $\varphi^2\log(1/|\varphi|)$ rather than linearly. The derivation combines branch dephasing with an entropic uncertainty relation; Rényi variants are also given. Independently of branch interference, logarithmic stabilizer extent is bounded above by $2\sum_i\log_2[\sqrt{1-|\sin\varphi_i|}+\sqrt{1-|\cos\varphi_i|}]$, using explicit stabilizer decompositions. Preprocessing fuses redundant gadgets, absorbs Clifford contributions into the backbone, and excludes effectively local rotations from the entropy correction. Benchmarks compare with exact entropy for random circuits on 20 qubits and Clifford-interleaved Trotter evolution on 26 qubits, and with exact magic on eight-qubit unitary and monitored circuits. The entropy estimates remain informative longer than generic entangling-capacity bounds in these tests. The entropy bounds are not certified by the scalable preprocessing: it searches only pairwise fusions, leaving higher-order interference possible. A first violation occurs at step 64 in the structured Trotter benchmark, matching the Clifford-layer periodicity. Applicability requires flow, and the monitored example uses a restricted circuit whose non-Clifford rotations can be moved past measurements.
Central claim may not hold: The abstract presents additive entanglement bounds without stating their essential no-coherent-branch-interference assumption, which the scalable preprocessing does not certify. The Trotter-plus-Clifford benchmark reports an entropy-bound violation at step 64, showing that the proposed preprocessing can yield heuristic rather than certified entropy intervals. The supplied benchmarks do not quantitatively establish the claimed informativeness at large qubit numbers and internal-spider counts, reporting at most 26 qubits and no spider-count scaling.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 6+6 pages, 3+2 figures