A new preprint on arXiv presents a closed-form law relating depth and width to neural collapse in deep linear networks. The authors note that how depth and width shape learned representations is typically studied only after training. In their tractable feature-learning setting, they show this dependence can instead be obtained directly in closed form.
The result offers a rare analytical handle on a phenomenon usually observed empirically. Because the setting is limited to deep linear networks, the law may not directly transfer to nonlinear architectures, but it provides a concrete starting point for understanding how architectural choices influence representation geometry.
As a single-source preprint, the findings have not yet been peer-reviewed. The abstract does not give the explicit form of the law, but the claim of a closed-form expression is itself notable for a field that often relies on simulations and post-hoc analysis.