Abstract:We present a sequential version of the multilinear Nyström algorithm which is suitable for the low-rank Tucker approximation of tensors given in a streaming format. Accessing the tensor $\mathcal{A}$ exclusively through random sketches of the original data, the algorithm effectively leverages structures in $\mathcal{A}$, such as low-rankness, and linear combinations. We present a deterministic analysis of the algorithm and demonstrate its superior speed and efficiency in numerical experiments including an application in video processing.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **How to efficiently perform low - rank Tucker approximation of tensors in a streaming data environment**. Specifically, the paper proposes a method called **Sequential Multilinear Nyström (SMLN) algorithm** for handling the low - rank Tucker decomposition of tensors given in a streaming format. This algorithm accesses and processes the original data through random sampling techniques and effectively utilizes the low - rank structure and linear combination characteristics in tensors. Compared with traditional non - streaming methods, the SMLN algorithm has a significant improvement in computational speed and efficiency, especially in application scenarios such as video processing. ### Main problem description: 1. **Streaming data processing**: How to perform low - rank approximation of tensors when data arrives one by one and cannot be all stored. 2. **Computational efficiency**: How to reduce computational complexity and avoid operations that become computational bottlenecks, especially when dealing with large - scale tensors. 3. **Accuracy**: How to maintain a relatively high approximation accuracy while ensuring computational efficiency. ### Solutions: - **Random Sketching**: Through random sampling techniques, only part of the tensor information is used to perform low - rank approximation, reducing the direct operation on the entire tensor. - **Sequential Processing**: Compress each mode of the tensor step by step, avoiding processing the entire tensor at once, thereby reducing computational complexity. - **Multilinear Nyström extension**: Extend the classical Nyström method to the multilinear case, which is suitable for processing high - order tensors. ### Application scenarios: - **Video processing**: Through the streaming processing of video frames, efficient video compression and analysis are achieved. ### Formula summary: - The Tucker decomposition of a tensor is represented as: \[ \mathcal{A}=\mathcal{C}\times_1 F_1\times_2 F_2\cdots\times_d F_d \] where $\mathcal{C}\in\mathbb{R}^{r_1\times r_2\times\cdots\times r_d}$ is the core tensor and $F_k\in\mathbb{R}^{n_k\times r_k}$ is the factor matrix. - The core steps of the SMLN algorithm include: - Randomly generate sampling matrices $X_k$ and $Y_k$. - Calculate small tensors $\mathcal{B}_k = \mathcal{A}\times_k Y_k^T$ and matrices $\Omega_k=\mathcal{A}_{-k}X_k$. - Use QR decomposition and projection matrices to construct low - rank approximations. Through these techniques, the SMLN algorithm can efficiently perform low - rank approximation of tensors in a streaming data environment and has broad application prospects.

A sequential multilinear Nyström algorithm for streaming low-rank approximation of tensors in Tucker format

A multilinear Nyström algorithm for low-rank approximation of tensors in Tucker format

Randomized algorithms for streaming low-rank approximation in tree tensor network format

Low-Rank Tucker Approximation of a Tensor from Streaming Data

Efficient Alternating Least Squares Algorithms for Low Multilinear Rank Approximation of Tensors

Randomized Algorithms for the Low Multilinear Rank Approximations of Tensors

Fast and Low-Memory Compressive Sensing Algorithms for Low Tucker-Rank Tensor Approximation from Streamed Measurements

Tensor-Based Sketching Method for the Low-Rank Approximation of Data Streams

Efficient randomized tensor-based algorithms for function approximation and low-rank kernel interactions

Efficient Algorithms for Low Tubal Rank Tensor Approximation with Applications to Image Compression, Super-Resolution and Deep Learning

Fast and Accurate Randomized Algorithms for Low-rank Tensor Decompositions

The Computation of Low Multilinear Rank Approximations of Tensors Via Power Scheme and Random Projection

Parallel Randomized Tucker Decomposition Algorithms

Effective Streaming Low-Tubal-Rank Tensor Approximation via Frequent Directions

Multilinear Low-Rank Tensors on Graphs & Applications

Practical Sketching Algorithms for Low-Rank Tucker Approximation of Large Tensors

An Iterative Reweighted Method for Tucker Decomposition of Incomplete Multiway Tensors

Efficient algorithms for Tucker decomposition via approximate matrix multiplication

Randomized Algorithms for the Approximations of Tucker and the Tensor Train Decompositions.

Online Robust Low-Rank Tensor Modeling for Streaming Data Analysis