Rank-One Iterative Source Steering (ISS)

• Rank-One ISS is a matrix update algorithm for multichannel blind source separation and dereverberation that avoids explicit matrix inversion.
• It employs structured rank-one corrections within a majorization-minimization framework to reduce computational complexity and enhance numerical stability.
• The method has been applied in various frameworks like CTF-MNMF and IVA, achieving substantial runtime reductions while maintaining separation performance.

Rank-One Iterative Source Steering (ISS) is a class of matrix update algorithms used for efficient multichannel blind source separation (BSS), dereverberation, and robust parameter estimation in highly reverberant and underdetermined environments. ISS is motivated by the need to update demixing or diagonalization matrices without explicit matrix inversion, replacing Iterative Projection (IP) steps with a sequence of structured rank-one corrections. This approach yields substantial computational savings and improved numerical stability while maintaining state-of-the-art separation performance across a range of BSS and spatial modeling frameworks (Xie et al., 30 Sep 2025, Nakashima et al., 2021, Nakashima et al., 2022, Scheibler et al., 2020, Ikeshita et al., 2022, Bando et al., 2023).

1. Mathematical Foundations of Rank-One ISS

ISS is formulated in the context of surrogate majorization-minimization (MM) approaches for BSS and spatial covariance models. The key objective is typically a function of the demixing matrix $W$ (or diagonalizer $Q$) and the separated signals, such as

$$\mathcal{L}(W) = \sum_{i=1}^m w_i^H V_i w_i - \log|\det W|^2,$$

where $V_i$ are positive-definite surrogate matrices derived from the data, source model statistics, or neural surrogates. Traditionally, optimizing $\mathcal{L}$ involves solving linear systems or inverting $W$ or $V_i$. Rank-One ISS bypasses explicit inversion by parameterizing each MM update as a rank-one correction: $W \leftarrow W - v_{k} w_{k}^H,$ where $w_{k}^H$ is the $k$th row of $Q$0, and $Q$1 is a vector of steering coefficients. These coefficients are derived via closed-form stationary conditions or by locally minimizing the auxiliary cost function for a single row or column direction, frequently leveraging the matrix determinant lemma and quadratic forms (Xie et al., 30 Sep 2025, Ikeshita et al., 2022, Nakashima et al., 2021, Bando et al., 2023). For each update, only quadratic forms and inner products are required:

• For non-diagonal terms: $Q$2.
• Diagonal normalization: $Q$3.

This strategy generalizes across MM-IVA, CTF-MNMF, FastFCA, and other related frameworks.

2. Algorithmic Structure and Pseudocode

A full ISS iteration over all sources or components is realized as either a sequence of row- or column-wise rank-one updates, integrated into BSS/EM-style iterative schemes. The canonical pseudocode for a frequency bin is given below (Xie et al., 30 Sep 2025, Ikeshita et al., 2022, Nakashima et al., 2021, Bando et al., 2023):

$$\mathcal{L}(W) = \sum_{i=1}^m w_i^H V_i w_i - \log|\det W|^2,$$2

For block coordinate ISS or the ISS2 extension, two columns (or blocks) may be updated per sweep (Ikeshita et al., 2022). ISS also admits efficient sequential (per-row/column) and batch sweep implementations (Xie et al., 30 Sep 2025, Nakashima et al., 2021). When integrated into factorization-based models (e.g. CTF-MNMF, FastFCA), ISS steps alternate with updates to spectral or NMF parameters.

3. Computational Complexity and Efficiency

Rank-One ISS achieves a significant reduction in computational complexity relative to IP-based algorithms:

• ISS complexity: $Q$4 per iteration, where $Q$5 is the number of frequencies, $Q$6 the number of sensors/sources, $Q$7 the number of frames (Ikeshita et al., 2022, Xie et al., 30 Sep 2025).
• IP complexity: $Q$8 due to matrix inversions or linear system solutions per source.

This difference arises because ISS involves only inner products, quadratic forms, and rank-one matrix-vector operations, with no explicit inversion or determinant computation. Performance profiling in CTF-MNMF shows typical CPU savings of 33–51%, with greater robustness to ill-conditioned scenarios or reverberant impulse responses (Xie et al., 30 Sep 2025).

4. Applications Across Source Separation Frameworks

ISS has been implemented and analyzed in a range of BSS and spatial modeling schemes:

• Convolutive Transfer Function Multichannel NMF (CTF-MNMF): Replaces IP in the E-step, steering the demixing matrix without matrix inversion for large microphone arrays or long CTFs (Xie et al., 30 Sep 2025).
• Independent Vector Analysis (IVA), AuxIVA: ISS and its extensions update demixing matrices or mixing matrices using surrogate functions in MM, offering efficiency in determined and overdetermined separation (Ikeshita et al., 2022, Scheibler et al., 2020, Nakashima et al., 2022).
• Spatio-temporal Filter Models (ILRMA-T): ISS updates both dereverberation and separation filters, supporting integration with methods such as Weighted Prediction Error (WPE) (Nakashima et al., 2021).
• Neural Variational BSS (Neural FastFCA): ISS provides the analytic, gradient-friendly multichannel parameter update within AVI/end-to-end neural separation architectures (Bando et al., 2023).

In online IVA, ISS enables adaptive or selective updates (e.g. only for moving sources), offering further reduction in update cost under dynamic conditions (Nakashima et al., 2022).

5. Theoretical Properties and Convergence

Each ISS update is guaranteed to monotonically decrease the surrogate MM cost function due to its derivation by gradient-zeroing along a row or column (block) direction. This property holds regardless of the presence of spectral or spatial parameter learning, and is maintained within end-to-end differentiable learning (e.g. ISS-unrolled neural systems) (Xie et al., 30 Sep 2025, Scheibler et al., 2020, Ikeshita et al., 2022). However, global optimality is not guaranteed for $Q$9 sources, as with IP and related BCD methods (Ikeshita et al., 2022). Empirically, ISS requires more outer iterations than the fastest block-IP schemes for fixed accuracy, but total runtime is often lower due to reduced per-iteration cost.

6. Empirical Performance and Use Cases

Experimental evaluations across BSS tasks confirm that ISS-based updates (including fully inversion-free sequential ISS) yield source separation performance statistically equivalent to IP-based solvers, as measured by SI-SDR, SDR, and word error rate (WER):

Method	N=2	N=3	N=4	Runtime Reduction
ILRMA-T-IP	14.0	12.3	10.8	–
ILRMA-T-ISS-JOINT	13.8	12.1	10.6	~30–50%
Neural ISS (2-spk)	10.7	7.7	5.6	~2x–6x
Laplace AuxIVA	8.1	3.2	0.7	–

Separation results (SDR, SI-SDR) are maintained within 1 dB of IP baselines, while total runtimes are reduced by 30–50% or more depending on the problem dimension (Xie et al., 30 Sep 2025, Nakashima et al., 2021, Scheibler et al., 2020). For neural ISS and surrogate learning, word error rates may be reduced by more than 30% relative to fixed prior methods (Scheibler et al., 2020). ISS is also more stable under highly reverberant or temporally dynamic mixtures.

7. Extensions and Ongoing Research

The ISS method has been extended to block-rank updates (ISS2), online/adaptive scenarios, and learned majorization with neural surrogates (Ikeshita et al., 2022, Nakashima et al., 2022, Scheibler et al., 2020, Bando et al., 2023). ISS2 matches the sweep-count convergence of block-IP ($$\mathcal{L}(W) = \sum_{i=1}^m w_i^H V_i w_i - \log|\det W|^2,$$0) but retains the lower $$\mathcal{L}(W) = \sum_{i=1}^m w_i^H V_i w_i - \log|\det W|^2,$$1 cost. Ongoing work includes robust moving-source detection for selective ISS, end-to-end neural separation architectures where ISS acts as a differentiable layer, and further integration into low-latency or streaming systems.

ISS is now considered a robust, computationally favorable update rule in advanced BSS and spatial modeling pipelines, particularly in regimes where large matrix inversions are prohibitive (Xie et al., 30 Sep 2025, Nakashima et al., 2021, Nakashima et al., 2022, Scheibler et al., 2020, Ikeshita et al., 2022, Bando et al., 2023).