Answers are grounded in this paper and cite their sources.
Sequential alternating least squares (ALS) fits a canonical polyadic (CP) tensor by updating one factor matrix at a time. For tensors of order d ≥ 3 represented by r CP components with full-column factor matrices and mode dimensions at least r, we establish the sharp fitting-width threshold r^(d−1) for generic finite exact termination. Below this threshold, almost every joint target-factor and initialization tuple fails after every finite sequence of single-mode updates, including adaptive schedules. At or above it, one cyclic sweep reconstructs the target generically in every mode order, even when block minimizers are nonunique. At the threshold, this conclusion holds for every fixed full-column target; at every larger width, it holds for every fixed orthogonally decomposable target with nonzero signed weights. The proofs use explicit sequential spanning constructions and Fourier witnesses for nonvanishing residual polynomials. Further results establish pointwise tensor stability across rank changes, Gaussian perturbation bounds, a finite-bit threshold algorithm at fixed order for supplied sufficiently accurate rounded data, and a companion CP-ALS threshold for general Tucker-core targets. The finite-step obstruction is compatible with asymptotic convergence below the threshold.
No takes yet. Share an insight, caveat, or question.
Tan et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: