Atiyo Ghosh,Antonio A. Gentile,Mario Dagrada,Chul Lee,Seong-Hyok Kim,Hyukgeun Cha,Yunjun Choi,Brad Kim,Jeong-Il Kye,Vincent E. Elfving
Abstract:Harmonic functions are abundant in nature, appearing in limiting cases of Maxwell's, Navier-Stokes equations, the heat and the wave equation. Consequently, there are many applications of harmonic functions from industrial process optimisation to robotic path planning and the calculation of first exit times of random walks. Despite their ubiquity and relevance, there have been few attempts to incorporate inductive biases towards harmonic functions in machine learning contexts. In this work, we demonstrate effective means of representing harmonic functions in neural networks and extend such results also to quantum neural networks to demonstrate the generality of our approach. We benchmark our approaches against (quantum) physics-informed neural networks, where we show favourable performance.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is: how to introduce the inductive bias of harmonic functions into machine - learning models, especially for neural networks and quantum neural networks. Specifically, the goal of the paper is to develop effective methods to represent harmonic functions and generalize these methods to quantum neural networks to demonstrate their universality. In addition, the paper also verifies the effectiveness of the proposed methods through multiple practical application scenarios (such as heat distribution, electrostatic imaging, and robot navigation).
### Analysis of the Core Problems in the Paper
1. **The Importance of Harmonic Functions**
- Harmonic functions exist widely in nature and appear in the limit cases of Maxwell's equations, Navier - Stokes equations, heat conduction equations, and wave equations.
- These functions have important uses in practical applications such as industrial process optimization, robot path planning, and the calculation of the first - exit time of random walks.
2. **Limitations of Existing Methods**
- Although harmonic functions have a wide range of applications, in the field of machine learning, few studies have attempted to introduce inductive biases for harmonic functions.
- Although the traditional finite element method (FEM) performs excellently in solving related equations, it lacks the ability of data - driven modeling, and it is difficult to handle noisy data and boundary conditions, as well as solve inverse problems.
3. **Solutions in the Paper**
- **Classical Neural Networks**: The paper shows how to represent harmonic functions in neural networks, especially achieving exact harmonicity using complex - valued neural networks in two - dimensional simply - connected domains.
- **Quantum Neural Networks**: It is further extended to quantum neural networks, and by using the automatic differentiation technique in quantum circuits, the representation of harmonic functions is achieved.
- **Multi - Connected Domain Processing**: For multi - connected domains, the paper proposes a method based on domain decomposition, combined with a loss function to ensure the continuity and harmonicity between different sub - domains.
4. **Experimental Verification**
- The paper verifies the effectiveness of the proposed methods through multiple practical application scenarios (such as heat distribution, electrostatic imaging, and robot navigation), and conducts benchmark tests with the existing physics - informed neural networks (PINNs), showing superior performance.
### Formula Summary
- **Definition of Harmonic Function**:
\[
\nabla^2 \phi = 0
\]
- **Representation by Complex - valued Neural Network**:
\[
\phi_H(x, y) = \Re \circ NN_h \circ c
\]
where \( c(x, y)=x + iy \), \( \Re(x + iy)=x \).
- **Output of Quantum Neural Network**:
\[
\phi_QH(x, y)=\Re\{\langle \emptyset | \Psi_{\theta,x,y} \rangle\}
\]
where \( |\Psi_{\theta,x,y}\rangle = \hat{U}_{\theta_2} QFM(x, y) \hat{U}_{\theta_1} |\emptyset\rangle \).
- **Loss Function for Multi - Connected Domains**:
\[
L_{\partial \omega}=\sum_{i,j} \mathbb{E}_{x \sim \partial \omega_{ij}} \left[ \left( \phi_H^{(i)}(x)-\phi_H^{(j)}(x) \right)^2+\left\| \nabla \phi_H^{(i)}(x)-\nabla \phi_H^{(j)}(x) \right\|^2_2 \right]
\]
Through the above methods, the paper successfully introduces the inductive bias of harmonic functions into classical and quantum neural networks and demonstrates its effectiveness in multiple practical applications.