Multi-Parameter Noise Geometry in Variational Quantum Circuits: A Layer-Resolved Extrapolation Framework Beyond Scalar ZNE
Kylymdar Omorov
Abstract
Standard Zero-Noise Extrapolation (ZNE) in Variational Quantum Eigensolvers (VQE) assumes a single global noise scaling parameter, failing to capture heterogeneous and non-linear noise propagation across layered quantum circuits. We simulate VQE parameter optimisation at a challenging L=6 hardware-efficient depth, extracting the exact expectation values at each iteration. When the optimal parameters theta_star are found, they are fixed, and the multi-parameter noise surface is mapped. We propose a multi-parameter noise framework, modeling expectation values as a function of layer-wise noise scaling factors E(lambda_1, lambda_2, ..., lambda_L). Using structured unitary folding U_l = U_l (U_l^dagger U_l)^((lambda_l - 1) / 2), we independently scale the noise contribution of each ansatz layer and empirically map the resulting multi-dimensional noise surface. We analyze the separability of inter-layer noise interactions, identify layers with disproportionate error contributions via the gradient dE/dlambda_l, and compare scalar ZNE with multivariate polynomial regression and Gaussian Process (GP) extrapolation across H_2, LiH, and BeH_2 under depolarizing, amplitude damping, and coherent over-rotation noise. Our results demonstrate that noise is non-uniform and partially coupled across layers, and that layer-resolved ZNE with higher-order extrapolation can improve accuracy and variance under realistic NISQ noise models. These findings provide geometric insight into noise propagation in variational circuits and motivate targeted, layer-specific error mitigation strategies.