Shock-wave Theory and Symmetry-reduced SGD, Miyagawa 2026

Taiki Miyagawa links shock-wave theory to symmetry-quotiented stochastic gradient descent in a mathematical framework presented on arXiv as arXiv:2606.18303, submitted 16 Jun 2026 and accepted to the 35th International Conference on Artificial Neural Networks (ICANN) 2026. The paper derives effective training dynamics on a quotient manifold and shows that, after symmetry reduction and local-entropy coarse-graining, the dynamics satisfy a viscous Hamilton-Jacobi equation, and under a gradient-field assumption the coarse-grained loss gradient obeys a Burgers-type equation.

What does the paper prove?

The paper proves that symmetry-quotiented and locally coarse-grained SGD dynamics satisfy specific partial differential equations: a viscous Hamilton-Jacobi equation for the effective dynamics and a Burgers-type equation for the gradient of the coarse-grained loss, enabling the rigorous establishment of shock formation. Miyagawa first quotients parameter symmetries, applies local-entropy coarse-graining, and then shows the resulting effective dynamics live on the quotient manifold and follow those PDEs. The author states that, given the assumption that raw parameter dynamics reduce to a gradient field on the quotiented space, "shock formation can be established rigorously."

How does the link apply to common architectures?

The analysis covers multilayer perceptrons, convolutional neural networks, Transformers, and mean-field networks, each shown to obey either the Hamilton-Jacobi or Burgers-type equations after symmetry reduction and coarse-graining. Miyagawa develops the connection in general mathematical terms using differential geometry, Lie group theory, and fluid mechanics, then applies the framework across those architectures to demonstrate that the same classes of PDEs appear in their symmetry-reduced dynamics.

Why it matters

The paper reframes training-phase transitions in terms of classical PDE phenomena such as shock formation, offering a principled mathematical basis for observables that are invariant under parameter symmetries. Miyagawa notes that in architectures such as Transformers, raw parameter norms can be distorted by symmetry redundancy and may therefore be misleading, while symmetry-corrected quotient observables give a principled basis for monitoring, forecasting, and controlling training-phase transitions. If the theoretical link translates to practical diagnostics, it could change how researchers interpret training signals and how they design monitoring tools for deep models.

What to watch

Look for the ICANN 2026 presentation and any follow-up work that tests the conjectured diagnostics on real training runs. The paper includes a conjecture that the framework "yields practical diagnostics for deep learning," so empirical validation that symmetry-corrected quotient observables improve monitoring or forecasting of training-phase transitions will be the key next milestone.

Additional factual details: the manuscript is available on arXiv under identifier arXiv:2606.18303 and the submission date is 16 Jun 2026; the paper was accepted to the 35th International Conference on Artificial Neural Networks (ICANN) 2026. The author frames the results using tools from differential geometry, Lie group theory, and fluid mechanics, and explicitly connects local-entropy coarse-graining and symmetry quotienting to viscous Hamilton-Jacobi and Burgers-type dynamics across multiple neural architectures.