Given an $N$-th order tensor $\mathcal{A} \in \mathbb{R}^{I_1\times I_2\times\ldots\times I_N}$ we can find an approximation of $\mathcal{A}$, $\hat{\mathcal{A}}$ by means of Higher Order Singular value Decomposition (HOSVD). That is,

$$\hat{\mathcal{A}}=\mathcal{Y}\times_1\mathbf{U}_1\times_2 \mathbf{U}_2 \ldots \times_N \mathbf{U}_N$$

where $\mathcal{Y}$ is a core tensor, $\times_i$ stands for mode-$i$ product (Tensor matrix multiplication) and matrices $\mathbf{U}_d|_{d=1}^N$ are mentioned as orthogonal projection matrices.

Q1: Matrices $\mathbf{U}_d|_{d=1}^N$ are generally rectangular, not square. However, an orthogonal matrix is generally square. How can this contradiction be explained?

Q2: Similarly a projection matrix is by definition square. How $\mathbf{U}_d|_{d=1}^N$ are projection matrices even though they are rectangular?

Q3: What matrices $\mathbf{U}_d|_{d=1}^N$ project and where? What is the physical interpretation of the term projection in HOSVD?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.