arXiv:2611.24196v1 Announce Type: new Abstract: Trace-ratio optimization problems from high-dimensional tensor data appear across numerous research domains including machine learning, signal processing, as well as pattern recognition. Up to now, analytical closed-form expressions remain unavailable for tensor trace-ratio tasks; prevailing techniques mostly adopt inner-outer iterative schemes to tackle such optimization tasks. However, when dealing with high-dimensional tensor data sets, existing multilinear discriminant analysis methods often encounter limitations such as heavy computational costs, insufficient memory space, and interference from redundant features. This paper yields two core contributions.First, we present a closed-form solution under the full row rank condition, and construct a randomized solver targeting large-scale tensor trace-ratio optimization. In contrast to prior approaches, the devised algorithm achieves substantial computational speedup.Its core mechanism is to reformulate the original large-scale tensor trace-ratio optimization into a tensor least-squares formulation, instead of the much more complicated (generalized) tensor eigenproblems. We derive theoretical error-bound results which verify the reliability of the newly developed algorithm. Second, we build a sparse multilinear discriminant analysis model leveraging the above closed-form result combined with tensor \ell_{2,0}-norm regularization, and propose a sparse multilinear discriminant method for tensor feature selection. The sparse method can identify key discriminative features from high-dimensional data and suppress redundant information. Comprehensive numerical evaluations are conducted over multiple publicly available real-world benchmarks.The proposed methods demonstrate significant advantages in computational efficiency and scalability for large-scale problems, while achieving competitive classification accuracy on some datasets.