Why do physics-informed constraints like Landau theory improve generalization better than pure neural approaches?

This is a sharp question, but the archive provided doesn't directly address why physics-informed constraints improve generalization in neural networks compared to pure learning approaches.

However, I can point toward the underlying intuition in the papers you've been given. The key principle appears in work on stability and structural fidelity: when you embed physical laws or conservation principles into a model's training or architecture, you're reducing the hypothesis space the network has to search. This acts like an implicit regularizer.

For instance, [AG-2025.01-1027] shows that physics-inspired optimization (using Hamiltonian dynamics) stabilizes neural networks by "reducing dependence on initial conditions while stabilizing the result near the minimum of the loss function." The physical structure—energy conservation in parameter space—constrains the learning trajectory itself. Similarly, [AG-2024.12-1755] demonstrates that when ML-based inverse modeling is constrained to respect hydrodynamic conservation laws, the resulting parameters can regenerate consistent physical fields from radiographic data.

The broader pattern: constraints like Landau theory (which codifies symmetry-breaking behavior in phase transitions) function as inductive biases. They tell the network *which features matter* before training begins, so it doesn't waste capacity learning spurious patterns. In contrast, pure neural approaches must discover these patterns from data alone, which typically requires more examples and is prone to overfitting on finite datasets.

The preprints don't contain a direct empirical comparison of Landau-constrained networks versus unconstrained ones on generalization benchmarks.