Abstract:We present a novel framework for learning cost-efficient latent representations in problems with high-dimensional state spaces through nonlinear dimension reduction. By enriching linear state approximations with low-order polynomial terms we account for key nonlinear interactions existing in the data thereby reducing the problem's intrinsic dimensionality. Two methods are introduced for learning the representation of such low-dimensional, polynomial manifolds for embedding the data. The manifold parametrization coefficients can be obtained by regression via either a proper orthogonal decomposition or an alternating minimization based approach. Our numerical results focus on the one-dimensional Korteweg-de Vries equation where accounting for nonlinear correlations in the data was found to lower the representation error by up to two orders of magnitude compared to linear dimension reduction techniques.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to learn cost - effective latent representations in high - dimensional state spaces through nonlinear dimensionality reduction techniques. Specifically, the paper introduces a new framework that enriches linear state approximations with low - order polynomial terms to account for key nonlinear interactions in the data, thereby reducing the intrinsic dimension of the problem. Two methods are proposed in the paper to learn the data - embedding representations of such low - dimensional polynomial manifolds. These methods can obtain the manifold parameterization coefficients through orthogonal decomposition or regression methods based on alternating minimization. The numerical results focus on the one - dimensional Korteweg - de Vries equation. It is found that, compared with linear dimensionality reduction techniques, considering the nonlinear correlations in the data can reduce the representation error by as much as two orders of magnitude.
### Main contributions:
1. **Propose a new dimensionality reduction framework**: By embedding low - order polynomial terms in the modal basis expansion, the nonlinear correlations in the data are considered, thereby achieving more effective dimensionality reduction.
2. **Two learning methods**: The polynomial manifold method based on POD and the polynomial manifold method based on alternating minimization are introduced. Both methods can effectively reduce the dimension of high - dimensional data.
3. **Numerical verification**: Through numerical experiments on the one - dimensional Korteweg - de Vries equation, the effectiveness of the proposed methods is demonstrated, especially in reducing the representation error.
### Key technical points:
- **Polynomial manifold**: By introducing low - order polynomial terms in the modal basis expansion, the nonlinear correlations in the data are captured.
- **Alternating minimization**: By solving a Procrustes - type problem, a basis matrix with orthogonal columns is found, thereby achieving better data compression performance.
- **Least - squares method**: The entire optimization process is completely data - driven and can be calculated using a standard least - squares solver.
### Numerical experiment results:
- **Korteweg - de Vries equation**: The experimental results show that, compared with the traditional POD method, the proposed polynomial manifold method has a significant improvement in the representation error and can be reduced by as much as two orders of magnitude.
### Conclusion:
The paper proposes a new framework that effectively captures the nonlinear correlations in high - dimensional state spaces by embedding low - order polynomial terms in the modal basis expansion, thereby achieving more efficient dimensionality reduction. Both learning methods (the method based on POD and the method based on alternating minimization) perform well in numerical experiments, especially when dealing with complex data. These methods are not only innovative in theory but also show great potential in practical applications.