Overcoming critical slowing down in frustrated spin systems by learned multiscale sampling

Imagine you're working on a simulation of a complex physical system, like a spin model, where interactions can lead to frustrating configurations that slow down your computations. Traditional methods, like cluster algorithms, excel in many scenarios but struggle when faced with even slight frustration, leading to what's known as critical slowing down. This means that as you approach a critical point, the algorithms take longer and longer to converge, making them inefficient for practical use. This is a significant hurdle when trying to model real-world systems accurately and efficiently.

To tackle this issue, the authors propose a new approach that learns the relevant clusters instead of constructing them. They introduce the wavelet conditional renormalization group (WCRG) method, which samples the probability distribution of fluctuations in a frustrated two-dimensional model. By recursively generating configurations from coarse to fine scales, WCRG maintains efficiency, achieving a sampling complexity of O(log2 L), which is a marked improvement over standard local MCMC methods. This method not only reproduces key statistical properties of the system but also clarifies the tradeoff between sampling speed and the model's expressiveness, providing a more robust solution for simulating frustrated systems.

The approach introduces a novel method for sampling in frustrated systems, addressing a significant limitation of existing algorithms.

The method shows solid performance across different phases, though it relies on the expressiveness of the underlying model.