Diffusion-based Sparse-Grid generative models for density estimation

In density estimation, generative models are usually categorized under unsupervised learning due to the lack of labeled data. These models apply various indirect loss functions to refine neural network training, yet face specific challenges. Issues like mode collapse and instability in generative adversarial networks are notable, while normalizing flows are constrained by the need to calculate the Jacobian matrix's determinant, limiting network design. While neural networks are well-suited for handling very high-dimensional data, they can be overly complex for moderately high dimensions. Here, traditional sparse polynomial approximation offers advantages by avoiding complex training requirements. This research employs a score-based diffusion model combined with sparse grid interpolation to estimate the unknown density function. The method involves generating labeled data pairs linking samples from the standard Gaussian distribution to the target distribution using the diffusion model. This model transports the Gaussian distribution to the target density through a backward stochastic differential equation, where a Monte Carlo method approximates the score function at any point, facilitating function interpolation for the transport model. A sparse grid interpolant can be built based on the labeled data. We leverage the Tasmanian library [32] for building this sparse-grid-based generative model. We demonstrate the performance of our method using a set of multi-dimensional benchmark distributions.