# Standardized quantum transistor block enables differentiable learning on gait dynamics

## TL;DR

This paper introduces a quantum transistor block for differentiable learning on gait dynamics, focusing on standardization and compatibility with classical systems.

## Contribution

The QT introduces a standardized, analyzable quantum-layer primitive with fixed port contract and closed-form gain/saturation.

## Key findings

- The QT network achieved a mean test accuracy of 0.960 and mean F1 of 0.931 on gait classification.
- Classical baselines outperformed the QT with F1 scores in the 0.962–0.964 range.
- The QT's design enables portable compilation and predictable shot/latency budgeting for quantum co-processor integration.

## Abstract

We introduce the Quantum Transistor (QT), a standardized variational quantum building block inspired by the operating-point and gain semantics of classical transistors. The QT is specified as a two-qubit template (gate g, channel t), but the experiments reported here instantiate the non-entangling special case in which g is deterministically prepared in \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$|1\rangle$$\end{document} and is not reused; consequently, the controlled-\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$R_y(\phi )$$\end{document} bias reduces to an unconditional single-qubit \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$R_y(\phi )$$\end{document} on the channel qubit t. In this instantiation the QT implements an analytically tractable, bounded scalar nonlinearity \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$s\mapsto \langle Z_t\rangle$$\end{document} with closed-form gain and saturation, and the full QT stack is efficiently classically simulable (a shallow classical contraction composed with bounded trigonometric scalar nonlinearities). We evaluate this standardized, single-qubit-effective QT stack on subject-aware gait classification with HyperBand and grouped cross-validation. In a locked-out 3-fold grouped protocol, the QT network reaches the mean test accuracy 0.960 and the mean F1 0.931. Strong classical baselines (CNN and Transformer encoders operating directly on the same spectrogram windows from which the QT’s 8-D inputs are derived) achieve F1 in the 0.962–0.964 range. Although the present QT stack (in its non-entangling, efficiently classically simulable instantiation) is outperformed by parameter-matched classical baselines, the objective of this paper is to validate a standardized, analyzable quantum-layer primitive rather than to claim superiority on this dataset. Accordingly, we position the evaluated system as a quantum-inspired, hybrid structured model study and we do not claim quantum advantage in this manuscript. The QT’s fixed port contract, closed-form gain/saturation, and constant-depth instruction template enable portable compilation, lightweight conformance tests (midpoint/slope/noise contraction), and predictable shot/latency budgeting. These properties are directly relevant in settings where a quantum co-processor is available as part of the system (e.g., integration with quantum sensing or other quantum-data pipelines). We therefore contextualize accuracy with strong classical baselines while outlining the block-level extensions (trainable biasing, pooling, richer encodings, and genuinely data-dependent gating) needed to close the remaining performance gap.