Overview
Quantum SEDONet (Spectral-Embedded Deep Operator Network) represents an advancement in neural-operator inference by integrating spectrally-embedded features into a quantum computing framework. This system aims to address limitations in the original Quantum DeepONet, specifically regarding the representation of oscillatory features when processing query coordinates with limited spectral structure. Quantum SEDONet introduces a method for assigning a spectral basis to each trunk coordinate based on its boundary condition, such as Fourier features for periodic coordinates and Chebyshev features for bounded, non-periodic coordinates.
The design allows for the integration of both periodic and non-periodic representations within a single problem, with the basis selected on a per-coordinate rather than per-problem basis. This spectral embedding is implemented without incurring additional qubits or increasing circuit depth, provided its dimension remains within the network width. Furthermore, the parameter count experiences only a marginal increase, specified as a few percent. The proposed system demonstrates reductions in mean relative L2 error across various benchmarks, while maintaining consistency with classical evaluation paths.
Research Context
The foundational technology for Quantum SEDONet is Quantum DeepONet. This predecessor system accelerates neural-operator inference through the evaluation of an orthogonally parameterized network on a quantum computer. Quantum DeepONet has been observed to reproduce the accuracy of its classical counterpart in ideal simulation, while achieving asymptotically lower inference cost. A identified characteristic of Quantum DeepONet is that its trunk network processes query coordinates with limited spectral structure. This characteristic necessitates the network to learn oscillatory features predominantly through its intrinsic nonlinearities.
Approach
Quantum SEDONet's methodological core involves assigning spectral bases to trunk coordinates. The selection of the spectral basis is contingent upon the coordinate's boundary condition:
- Fourier features are applied to periodic coordinates.
- Chebyshev features are applied to bounded, non-periodic coordinates.
This assignment strategy is implemented on a per-coordinate basis, allowing for the concurrent use of both representation types within a single problem. For instance, the channel Poisson problem, one of the benchmarks, utilizes Fourier features in its periodic direction and Chebyshev features in its bounded direction simultaneously.
The implementation of this spectral embedding utilizes unary amplitude encoding. This encoding method ensures that the embedding does not introduce additional qubits or increase circuit depth, contingent on the embedding's dimension remaining within the network's width. The integration of these spectral features results in an increase in the parameter count, which is quantified as only a few percent.
Findings
Evaluations of Quantum SEDONet across four distinct benchmarks yielded the following outcomes:
- For the antiderivative problem, the mean relative L2 error was reduced by 54.1%.
- For the advection problem, the mean relative L2 error was reduced by 49.6%.
- For the Burgers problem, the mean relative L2 error was reduced by 36.0%.
- For a mixed-boundary channel Poisson problem, the mean relative L2 error was reduced by 36.2%.
A consistent observation across these evaluations was the agreement between quantum and classical evaluation paths, which held to within $10^{-8}$ throughout. The channel Poisson problem specifically demonstrated the system's ability to implement coordinate-wise boundary-matched spectral embedding without incurring additional quantum-resource cost, by simultaneously employing Fourier features in the periodic direction and Chebyshev features in the bounded direction.
Why This Matters
The development of Quantum SEDONet offers a method to enhance the spectral learning capabilities of quantum neural operators. By improving the representation of oscillatory features through targeted spectral embeddings, the system contributes to the accuracy and efficiency of solving partial differential equations on quantum computing platforms, potentially expanding the range of problems that can be effectively tackled.