Tucker 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[1] although it goes back to Hitchcock in 1927.[2] 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 ${\displaystyle T\in F^{n_{1}\times n_{2}\times n_{3}}}$, where ${\displaystyle F}$ is either ${\displaystyle \mathbb {R} }$ or ${\displaystyle \mathbb {C} }$, Tucker Decomposition can be denoted as follows,

$${\displaystyle T={\mathcal {T}}\times _{1}U^{(1)}\times _{2}U^{(2)}\times _{3}U^{(3)}}$$

where ${\displaystyle {\mathcal {T}}\in F^{d_{1}\times d_{2}\times d_{3}}}$ is the core tensor, a 3rd-order tensor that contains the 1-mode, 2-mode and 3-mode singular values of ${\displaystyle T}$, which are defined as the Frobenius norm of the 1-mode, 2-mode and 3-mode slices of tensor ${\displaystyle {\mathcal {T}}}$ respectively. ${\displaystyle U^{(1)},U^{(2)},U^{(3)}}$ are unitary matrices in ${\displaystyle F^{d_{1}\times n_{1}},F^{d_{2}\times n_{2}},F^{d_{3}\times n_{3}}}$ respectively. The j-mode product (j = 1, 2, 3) of ${\displaystyle {\mathcal {T}}}$ by ${\displaystyle U^{(j)}}$ is denoted as ${\displaystyle {\mathcal {T}}\times U^{(j)}}$ with entries as

${\displaystyle {\begin{aligned}({\mathcal {T}}\times _{1}U^{(1)})(n_{1},d_{2},d_{3})&=\sum _{i_{1}=1}^{d_{1}}{\mathcal {T}}(i_{1},d_{2},d_{3})U^{(1)}(i_{1},n_{1})\\({\mathcal {T}}\times _{2}U^{(2)})(d_{1},n_{2},d_{3})&=\sum _{i_{2}=1}^{d_{2}}{\mathcal {T}}(d_{1},i_{2},d_{3})U^{(2)}(i_{2},n_{2})\\({\mathcal {T}}\times _{3}U^{(3)})(d_{1},d_{2},n_{3})&=\sum _{i_{3}=1}^{d_{3}}{\mathcal {T}}(d_{1},d_{2},i_{3})U^{(3)}(i_{3},n_{3})\end{aligned}}}$

Taking ${\displaystyle d_{i}=n_{i}}$ for all ${\displaystyle i}$ is always sufficient to represent ${\displaystyle T}$ exactly, but often ${\displaystyle T}$ can be compressed or efficiently approximately by choosing ${\displaystyle d_{i}<n_{i}}$. A common choice is ${\displaystyle d_{1}=d_{2}=d_{3}=\min(n_{1},n_{2},n_{3})}$, which can be effective when the difference in dimension sizes is large.

There are two special cases of Tucker decomposition:

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

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

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