Conditional Confinement Criteria and Empirical Diagnostics for Quadratically Regularized Transformer Objectives
Abstract Coupled $L^2$ regularization makes neural-network objectives with non-negative data-fitting losses coercive and their Gibbs densities normalizable, but these properties do not guarantee differential confinement of the parameter-space derivatives. This study establishes this distinction for Transformer objectives. A bilinear-logit counterexample shows that bounded inputs and positive quadratic regularization are insufficient to ensure $\Psi_{s,\lambda}(\theta) =-\Delta\mathcal{F}_{\lambda}(\theta) +s^{-1}\|\nabla\mathcal{F}_{\lambda}(\theta)\|^2 \rightarrow+\infty$ in every parameter-space direction. We derive a sufficient criterion that balances the asymptotic growth of the data-loss gradient and Hessian trace against the regularization strength and temperature. We then develop a scalable finite-directional diagnostic based on Hessian--vector products and stochastic trace estimation. Experiments with GPT-Neo-125M on Penn Treebank and WikiText-103 reveal substantial variation across directions, checkpoints, temperatures, and regularization implementations. Within the investigated near-reference regime, the negative data-loss Hessian-trace contribution and the constant regularizer trace dominate the positive radial term, and the measured diagnostic does not exhibit uniform quadratic growth across the sampled directions. The results separate global mathematical confinement from finite directional evidence and provide a reproducible framework for investigating asymptotic loss geometry in large Transformer models.
Authors
- Sayantan Dutta (ORCID: https://orcid.org/0000-0002-8187-4193)
- Abhijit Das
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11379021/v1
- Primary Topic
- Stochastic Gradient Optimization Techniques
- Type
- preprint