arXiv:2610.01824
QPI-DeepONet-MAC: A Scalable and Stable Hybrid Classical-Quantum Architecture for Physics-Informed Deep Operator Networks
Said Lantigua, José Valencia, Gilson Giraldi, Renato Portugal, Jonas Maziero
Abstract
General operator learning for parametric partial differential equations (PDEs) is a fundamental challenge at the intersection of artificial intelligence and physics-based modeling. Physics-informed Deep Operator Networks (PI-DeepONets) incorporate governing equations into learning, but face optimization difficulties in high dimensions. Quantum extensions offer additional representational capacity, yet their trainability can be hindered by barren plateaus. Here, we introduce QPI-DeepONet-MAC, a hybrid quantum-classical architecture that integrates parameterized quantum circuits into PI-DeepONets through a multiplicative-and-additive coupling (MAC) of classical representations and quantum expectation values. We prove analytically that the proposed architecture retains universal operator approximation capabilities. Furthermore, we derive bounds on quantum parameter gradients, establishing explicit conditions to avoid barren plateaus alongside a scaling criterion for trainability. Additionally, we derive a lower bound on the number of sensors required for a target accuracy, directly linking sampling requirements to input function regularity, domain dimension, and PDE dynamical growth. Finally, we propose an informational regularization scheme based on quantum coherence and state fidelity dynamics to prevent coherence loss and preserve quantum structure during training. Together, these results establish a theoretical foundation for quantum-enhanced physics-informed operator learning and position QPI-DeepONet-MAC as a scalable, stable framework for solving parametric PDEs.
AI Summary
The paper studies whether parameterized quantum circuits can be incorporated into physics-informed DeepONets while retaining operator approximation capability and controllable quantum-parameter gradients. Its QPI-DeepONet-MAC architecture multiplies shifted classical branch and trunk outputs by quantum expectation values and adds a quantum-dependent bias. Training combines PDE residuals, initial and boundary conditions, optional solution data, and proposed coherence–fidelity regularization. The analytical results comprise a universal operator-approximation theorem, gradient bounds, and a sensor-count criterion. Approximation is developed using classical DeepONet approximation results and a Stone–Weierstrass argument. Writing $a=N^{\mathcal B}\mathcal N^{\mathcal B}$ and $b=N^{\mathcal T}\mathcal N^{\mathcal T}$ for qubit-count–depth products, the stated gradient upper bound scales as $O(b^{-1}a^{-1/2}+a^{-1}b^{-1/2})$. A separate expected squared-gradient lower bound is claimed to be inverse-polynomial under additional shallow-circuit and local-observable conditions. These analyses concern a squared solution-error loss rather than the complete physics-informed training objective. For semilinear evolution equations, the stated sensor criterion is $m\geq[\mathfrak C_L C_{\mathrm{Int}}/(\tilde\epsilon-\epsilon)]^{d/k}e^{\omega td/k}$, connecting tolerance, input regularity $k$, dimension $d$, and dynamical growth. Its derivation combines interpolation or Kolmogorov-width estimates with semigroup stability bounds. The informational regularizer penalizes low coherence and changes between successive quantum states. No numerical benchmarks or hardware experiments are reported. The paper explicitly does not establish an advantage over classical operator learning; practical accuracy, optimization stability, sensor requirements, regularization behavior, and robustness to hardware noise remain subjects for future testing.
Central claim may not hold: The polynomial gradient upper bound (46) cannot establish barren-plateau avoidance or detectable gradients, because zero and exponentially small gradients also satisfy it. The gradient lower-bound argument relies on additional shallow-circuit assumptions and drops a potentially zero or exponentially small trunk expectation-value second moment without establishing a size-controlled positive lower bound. Appendix B incorrectly invokes Stone–Weierstrass to claim uniform approximation of arbitrary L-infinity functions rather than continuous functions. Appendix F derives only a sufficient sensor prescription from an upper error estimate, not the necessary minimum sensor count claimed in the abstract and Theorem Teo-3.2. The informational regularizer evaluates coherence on a diagonalized state, making that penalty constant as written, while its motivation also incorrectly attributes entanglement-induced mixing to a product state and identifies zero entropy with absence of superposition.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 47 pages, 3 figures