Chapter 18: Langevin Dynamics and Sampling. In the chapter mind map this icon labels Role of Noise in Sampling & Exploration. The discussion below is excerpted and lightly edited from § The Sampling vs Optimization Distinction in Mathematics for AI and Machine Learning.

This distinction between sampling and optimization is fundamental in machine learning:

• Optimization (gradient ascent/descent): Seeks to find a single point (or small set of points) that optimizes an objective function. The goal is convergence to an optimum.
• Sampling (Langevin dynamics, MCMC): Seeks to generate a sequence of points whose empirical distribution matches a target distribution. The goal is exploration and representation of the full distribution.

__In generative modeling, we need sampling, not optimization.__ We want to generate diverse samples that collectively represent the data distribution, not just find its modes.

the book figure demonstrates the fundamental difference between optimization and sampling for a 2D mixture of two Gaussians. The background shows the probability density as contour lines, with two distinct modes. The black dashed line shows a deterministic gradient ascent trajectory starting from a low-density region. This path follows the score field directly and converges to a single mode, demonstrating mode collapse: the deterministic process gets trapped in one mode and cannot explore the other. In contrast, the red solid line shows a Langevin dynamics trajectory with the same starting point. The injected noise allows the process to explore both modes, visiting regions around both peaks. This visualization makes concrete why noise is essential: it enables exploration of the entire distribution, diversity in generated samples, and correctness of the stationary distribution. Without noise, we would only generate samples from a single mode, failing to represent the true multi-modal distribution.

This need for sampling in generative modeling directly motivates the introduction of Langevin dynamics, which provides a principled way to balance gradient guidance with exploratory noise.