Quantum-Native Transformer Architecture: Interference, Partition Functions, and Nonlinear Schrodinger Dynamics

Abstract

Timo Aukusti Laine

This paper introduces the Quantum Semantic Circuit (QSC), a quantum-native transformer architecture designed to bridge the gap between classical transformers and quantum mechanics. We map classical transformer components to their quantum mechanical counterparts, classifying these correspondences as exact identities, derivational approximations, representational equivalences, or functional analogies. Key findings include the exact correspondence between L2-normalized embeddings and quantum states, softmax attention weights and Boltzmann distributions, quantum interference as a central process, and the derivation of the cubic nonlinear Schrödinger equation (NLSE) as the unique minimal norm-preserving nonlinearity suitable for quantum-native activation functions. We demonstrate that classical sinusoidal position encoding and residual connections are leading-order approximations of exact quantum phase rotations and unitary evolution, respectively. We propose that the dimensional expansion within the classical feed-forward network (FFN) is a real-valued shadow of quantum symmetry breaking and restoration, potentially linked to the Higgs mechanism in large models. A quantum simulator pilot on a lexical disambiguation task validates the architectural self-consistency of the QSC and confirms that the NLSE nonlinearity performs measurable computational work beyond linear evolution. This work provides a theoretical foundation for exploring quantum- enhanced language models and identifies key hardware requirements for their physical realization.

PDF