Can NDA estimate coefficients in theories where multiple mass scales compete?

# Can NDA Estimate Coefficients in Multi-Scale Theories?

The short answer: yes, but you need the right computational tools. The archive contains two complementary approaches to this problem.

The Gradient-Based Approach

The most direct answer comes from modern effective field theory work. Researchers have built differentiable, multi-scale EFT likelihoods that can simultaneously handle renormalization-group evolution, matching conditions, and experimental constraints [AG-2026.03-1506]. Think of this as teaching a neural network to "understand" how coupling constants change as you move between energy scales—the gradient-based optimization then finds the best-fit coefficients across all scales at once. This has made 374-parameter global EFT fits practical, which would have been nearly impossible with older methods.

Why this matters: you can now estimate coefficients in theories with competing mass scales without having to decouple scales artificially or solve the running equations by hand.

Where Multiple Scales Create Structure

However, competing mass scales don't just pose a computational challenge—they can *generate* physical hierarchies. One mechanism involves disorder in "theory space" [AG-2025.01-782]: if you arrange quantum fields along a chain with disordered mass terms, they localize like electrons in a dirty conductor. Fields at opposite ends of the chain develop exponentially suppressed couplings, naturally producing hierarchies. This shows that multi-scale competition can itself be a feature, not a bug, that your fitting procedure should exploit.

For neutrino physics specifically, certain combinations of mass parameters turn out to be scale-independent—they don't run with energy [AG-2025.11-1162]. This means your fitting procedure can anchor on these invariant relations to constrain the landscape more tightly.

---

The provided preprints don't discuss NDA (non-derivative analysis or network-driven anomaly detection) by name, so I cannot comment on that specific technique's performance in multi-scale settings.