Theory of Speciation Transitions in Diffusion Models with General Class Structure

TL;DR

The paper develops a general theory of speciation transitions in diffusion-based generative models for targets with arbitrary class structure by introducing Bayes attribution and a free-entropy criterion. A rigorous speciation time $t_{rs}$ is defined via the balance between the mean free-entropy difference and its fluctuations, yielding large-$N$ Scalings: $t_{rs}\sim \tfrac{1}{2}\log N$ when first moments separate and $t_{rs}\sim \tfrac{1}{4}\log N$ when they do not. The authors apply the framework to two Gaussian-mix scenarios and to multi-class 1D Ising mixtures, deriving analytical expressions (via replica methods for Ising) and validating predictions with numerical U-turn experiments, revealing hierarchical speciation times. The results provide a unified description of speciation transitions in diffusion models, enabling principled prediction of when and how trajectories commit to data classes across high-dimensional settings with diverse class definitions. The approach has broad applicability to diffusion-based generation beyond Gaussian mixtures, including models with complex, higher-order, or collective class features.

Abstract

Diffusion Models generate data by reversing a stochastic diffusion process, progressively transforming noise into structured samples drawn from a target distribution. Recent theoretical work has shown that this backward dynamics can undergo sharp qualitative transitions, known as speciation transitions, during which trajectories become dynamically committed to data classes. Existing theoretical analyses, however, are limited to settings where classes are identifiable through first moments, such as mixtures of Gaussians with well-separated means. In this work, we develop a general theory of speciation in diffusion models that applies to arbitrary target distributions admitting well-defined classes. We formalize the notion of class structure through Bayes classification and characterize speciation times in terms of free-entropy difference between classes. This criterion recovers known results in previously studied Gaussian-mixture models, while extending to situations in which classes are not distinguishable by first moments and may instead differ through higher-order or collective features. Our framework also accommodates multiple classes and predicts the existence of successive speciation times associated with increasingly fine-grained class commitment. We illustrate the theory on two analytically tractable examples: mixtures of one-dimensional Ising models at different temperatures and mixtures of zero-mean Gaussians with distinct covariance structures. In the Ising case, we obtain explicit expressions for speciation times by mapping the problem onto a random-field Ising model and solving it via the replica method. Our results provide a unified and broadly applicable description of speciation transitions in diffusion-based generative models.

Figures (6)

• Figure 1: Points scattered on two concentric sphere offer an example of a Pure Density Decomposition, where the two components cannot be distinguished on the basis of their mean.
• Figure 2: Potential of the reverse SDE as a function of $r$ for different times for a mixture of two Gaussians with zero means and different isotropic variances $\sigma^2_{1,2} = 1 \pm \delta$. It is clearly noticeable a change in shape, from a single well for large $t$ to a double well for small $t$. We identify the speciation time in correspondence with the change in curvature in $r=1$.
• Figure 3: Analysis of a mixture of two 1D Ising models, at inverse temperatures $\beta_1=0.5$, $\beta_1=1$. Solid lines show the average free entropy difference for different number of spins $N$ as a function of forward diffusion time. The shading represents the $3\sigma$ confidence interval. Dashed lines show the misattribution fraction during the forward process. Misattribution starts to rise when zero enters the confidence interval.
• Figure 4: Attribution matrices at the end of the backward process for increasing U-turn times computed numerically with the transfer matrix method. Merging times predicted by our criterion are $t_1 \simeq 0.42$ for inverse temperatures $\beta_1$ and $\beta_2$, and $t_2 = 1.39$, for $\beta_2$ and $\beta_3$. The target distribution is a 3-Ising mixture with $\beta_1=0.2$, $\beta_2=0.3$, $\beta_3=1.0$; $N=1600$.
• Figure 5: Attribution matrices at the end of the backward process for increasing U-turn times, computed numerically with the transfer matrix method (top row) compared with the analytical ones (bottom row). The target distribution is a 8-Ising mixture with hierarchically generated temperatures (beta values displayed on the left). System size is $N=1600$.