Tucker decomposition

Short description: Tensor decomposition

In mathematics, Tucker decomposition decomposes a tensor into a set of matrices and one small core tensor. It is named after Ledyard R. Tucker although it goes back to Hitchcock in 1927. Initially described as a three-mode extension of factor analysis and principal component analysis it may actually be generalized to higher mode analysis, which is also called higher-order singular value decomposition (HOSVD).

It may be regarded as a more flexible PARAFAC (parallel factor analysis) model. In PARAFAC the core tensor is restricted to be "diagonal".

In practice, Tucker decomposition is used as a modelling tool. For instance, it is used to model three-way (or higher way) data by means of relatively small numbers of components for each of the three or more modes, and the components are linked to each other by a three- (or higher-) way core array. The model parameters are estimated in such a way that, given fixed numbers of components, the modelled data optimally resemble the actual data in the least squares sense. The model gives a summary of the information in the data, in the same way as principal components analysis does for two-way data.

For a 3rd-order tensor $T\in F^{n_1\times n_2\times n_3}$, where $F$ is either $\mathbb{ℝ}$ or $\mathbb{ℂ}$, Tucker Decomposition can be denoted as follows, $$T={𝒯}{\times }_1{U}^{(1)}{\times }_2{U}^{(2)}{\times }_3{U}^{(3)}$$ where ${𝒯}\in F^{d_1\times d_2\times d_3}$ is the core tensor, a 3rd-order tensor that contains the 1-mode, 2-mode and 3-mode singular values of $T$, which are defined as the Frobenius norm of the 1-mode, 2-mode and 3-mode slices of tensor ${𝒯}$ respectively. $U^{(1)},U^{(2)},U^{(3)}$ are unitary matrices in $F^{d_1\times n_1},F^{d_2\times n_2},F^{d_3\times n_3}$ respectively. The j-mode product (j = 1, 2, 3) of ${𝒯}$ by $U^{(j)}$ is denoted as ${𝒯}\times U^{(j)}$ with entries as

$\begin{array}{rl}({𝒯}{\times }_1{U}^{(1)})(n_1,d_2,d_3)& =\sum _{i_1=1}^{d_1}{𝒯}(i_1,d_2,d_3)U^{(1)}(i_1,n_1)\\ ({𝒯}{\times }_2{U}^{(2)})(d_1,n_2,d_3)& =\sum _{i_2=1}^{d_2}{𝒯}(d_1,i_2,d_3)U^{(2)}(i_2,n_2)\\ ({𝒯}{\times }_3{U}^{(3)})(d_1,d_2,n_3)& =\sum _{i_3=1}^{d_3}{𝒯}(d_1,d_2,i_3)U^{(3)}(i_3,n_3)\end{array}$

Taking $d_i=n_i$ for all $i$ is always sufficient to represent $T$ exactly, but often $T$ can be compressed or efficiently approximately by choosing $d_i<n_i$. A common choice is $d_1=d_2=d_3=\min (n_1,n_2,n_3)$, which can be effective when the difference in dimension sizes is large.

There are two special cases of Tucker decomposition:

Tucker1: if $U^{(2)}$ and $U^{(3)}$ are identity, then $T={𝒯}{\times }_1{U}^{(1)}$

Tucker2: if $U^{(3)}$ is identity, then $T={𝒯}{\times }_1{U}^{(1)}{\times }_2{U}^{(2)}$ .

RESCAL decomposition can be seen as a special case of Tucker where $U^{(3)}$ is identity and $U^{(1)}$ is equal to $U^{(2)}$ .

See also

• Higher-order singular value decomposition
• Multilinear principal component analysis