Abstract:We introduce novel estimators for computing the curvature, tangent spaces, and dimension of data from manifolds, using tools from diffusion geometry. Although classical Riemannian geometry is a rich source of inspiration for geometric data analysis and machine learning, it has historically been hard to implement these methods in a way that performs well statistically. Diffusion geometry lets us develop Riemannian geometry methods that are accurate and, crucially, also extremely robust to noise and low-density data. The methods we introduce here are comparable to the existing state-of-the-art on ideal dense, noise-free data, but significantly outperform them in the presence of noise or sparsity. In particular, our dimension estimate improves on the existing methods on a challenging benchmark test when even a small amount of noise is added. Our tangent space and scalar curvature estimates do not require parameter selection and substantially improve on existing techniques.
What problem does this paper attempt to address?
This paper aims to solve the problem of estimating geometric properties on data manifolds, especially the estimation of curvature, tangent space and dimension. Traditionally, although Riemannian geometry provides a rich theoretical basis for geometric data analysis and machine learning, in practical applications, these methods often have difficulty achieving ideal results in statistical performance, especially when dealing with noisy data or low - density data. By introducing the method of diffusion geometry, the author proposes a series of new estimators that can accurately and robustly calculate these geometric properties. These new methods not only perform equivalently to the existing state - of - the - art methods under ideal conditions (i.e., dense, noise - free data), but also significantly outperform the existing methods in the presence of noise or sparse data.
### Core problems of the paper
1. **Curvature estimation**: How to estimate scalar curvature, Ricci curvature and Riemann curvature from data manifolds.
2. **Tangent space estimation**: How to estimate the tangent space at each point from data manifolds.
3. **Dimension estimation**: How to estimate local and global dimensions from data manifolds.
### Solutions
The paper utilizes the diffusion geometry theory to estimate the Laplace operator \(\Delta\) on the manifold through the heat kernel (heat kernel), and then constructs important objects in Riemannian geometry, such as tangent space, dimension and curvature. The specific steps are as follows:
1. **Diffusion graph and diffusion geometry**:
- Use the diffusion graph method to construct a normalized heat kernel matrix \(K_\epsilon(x_i, x_j)=\exp\left(-\frac{\|x_i - x_j\|^2}{\epsilon}\right)\), where \(\epsilon\) is the bandwidth parameter.
- Estimate the approximate value \(\hat{\Delta}_\epsilon\) of the Laplace operator \(\Delta\) through the heat kernel matrix.
2. **carré du champ formula**:
- Use the carré du champ formula \(g(\nabla f,\nabla h)=\frac{1}{2}(f\Delta(h)+h\Delta(f)-\Delta(fh))\) to estimate the Riemannian metric.
- Calculate the estimated value \(\hat{\Gamma}(\hat{f},\hat{h})\) of carré du champ through \(\hat{\Delta}_\epsilon\).
3. **Dimension estimation**:
- Construct a Gram matrix \(G(p)\) to estimate the tangent space at each point.
- Estimate the dimension \(d\) through the eigenvalues of the Gram matrix.
4. **Tangent space and curvature estimation**:
- Estimate the tangent space through the first \(d\) eigenvectors of the Gram matrix.
- Use the Hessian matrix and the carré du champ formula to estimate the curvature.
### Experimental results
The paper verifies the effectiveness of the proposed method through a series of experiments, especially the performance on noisy data and low - density data. The experimental results show that the new method significantly outperforms the existing state - of - the - art methods in these cases. Specifically:
- **Dimension estimation**: On noisy data and low - density data, the accuracy of the new method is significantly higher than that of other methods.
- **Tangent space estimation**: Even in the presence of a large amount of noise, the new method can robustly recover the tangent space.
- **Curvature estimation**: The new method performs well in estimating scalar curvature, Ricci curvature and Riemann curvature, especially for high - dimensional manifolds.
### Conclusion
The new methods proposed in the paper successfully solve the problem of estimating geometric properties on data manifolds through the diffusion geometry theory, especially when dealing with noise and low - density data. These methods are not only of great theoretical significance, but also provide powerful tools for practical applications.