New tensor PCA theory guarantees faster convergence with warm starts

Researchers Yanjin Xiang and Zhihua Zhang have published a significant advance in tensor PCA (principal component analysis). Their paper, 'Finite-Iteration Local Dynamics and Warm Starts for Alternating Power Iteration in Spiked Tensor PCA,' addresses a long-standing challenge: ensuring efficient convergence of alternating power iteration regardless of how the algorithm is initialized. The key innovation is a finite-iteration local theory that works once iterates enter a sufficiently small neighborhood of the true signal. The error then decomposes neatly into a geometrically decaying transient term and an intrinsic noise floor caused by fixed orthogonal noise contractions. This provides a clear theoretical framework for predicting convergence behavior in practice.

The authors also separate the warm-start mechanism from specific spectral constructions, introducing a generic one-sweep principle. If an initializer has correlation γ_N and a first-sweep noise level a_N such that a_N/(γ_N^{d-1} ω_{N,d}) → 0, then one can choose an expanding radius that guarantees entry into the local basin. After entry, local affine contraction yields convergence to the unique informative fixed point. The paper validates this approach for centered-Gram initialization under i.i.d. finite-fourth-moment noise, using a novel pressed-back estimate. This work makes tensor PCA more reliable for applications like signal processing and machine learning where initializations are often arbitrary.