オオイシ　クニオ   大石　邦夫    所属   コンピュータサイエンス学部 コンピュータサイエンス学科    職種   教授
言語種別	日本語
発行・発表の年月	2026/07
形態種別	学術論文
査読	査読あり
標題	Approximate Joint Diagonalization for Blind Separation of Mixtures in the Rectangular Mixing Case
執筆形態	共著
掲載誌名	IEEE Access
掲載区分	国外
出版社・発行元	IEEE
巻・号・頁	14,103929-103943頁
総ページ数	15
担当区分	最終著者
著者・共著者	Sinya Saito and Kunio Oishi
概要	Blind Source Separation (BSS) using rectangular mixing models—where observations outnumber sources—offers superior separation performance, but its high computational complexity hinders real-world application in Approximate Joint Diagonalization (AJD) via Alternating Least Squares (ALS). To overcome this limitation, this paper presents a canonical representation for the AJD problem under rectangular mixing matrix conditions. By integrating Principal Component Analysis (PCA) for dimensionality reduction, target matrices are projected onto a signal subspace, transforming conventional ALS into computationally efficient algorithms. During ALS iterations, a residual-square mixing matrix minimizes the squared error between reduced-dimensional target matrices and their replicas, followed by multiplicative updates to achieve successive matrix diagonalization. Experimental results confirm that the proposed subspace-based ALS approach significantly reduces computational requirements compared to standard methods while fully preserving source separation performance.
DOI	10.1109/ACCESS.2026.3709317
論文（査読付）ファイル	DOWNLOAD