Title: Utilizing Non-Identical Diffusion Models for MIMO-OFDM Channel Synthesis

Abstract:

This study introduces a novel approach to wireless orthogonal frequency division multiplexing (OFDM) channel generation through a proposed method known as the non-identical diffusion model. In contrast to traditional diffusion models, which rely on a scalar-valued time index to denote a global noise level, our approach extends this concept by employing an element-wise time indicator. This modification allows for a more precise capture of local error variations. By enabling the characterization of reliability for individual elements—such as subcarriers in OFDM systems—within the noisy input, the non-identical diffusion framework yields enhanced generation outcomes, particularly when the initial state is biased.

Our primary focus is the reconstruction of wireless multi-input multi-output (MIMO) OFDM channel matrices. In this context, initial channel estimates display significant disparities in reliability across different elements, a phenomenon driven by the pilot scheme. Standard time embeddings, predicated on the assumption of uniform noise progression, are unable to account for such variability across different pilot configurations and noise intensities. To address this limitation, we introduce a matrix, sized to match the input, which governs the element-wise progression of noise.

We demonstrate both the theoretical validity and numerical effectiveness of this proposed non-identical diffusion scheme, adhering to a diffusion process similar to that of existing methods. Furthermore, for MIMO-OFDM channel generation, we propose a dimension-wise time embedding strategy. The paper concludes with the development and evaluation of various training and generation techniques, with their performance compared through comprehensive numerical experiments.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC