It is one axis or dimension of a tensor, such as product or time.

What does the core do?

It represents interactions among the reduced components from the different modes.

Is Tucker decomposition always approximate?

It can represent a tensor exactly with sufficient dimensions, while reduced ranks usually create an approximation.

AI Glossary / AI and machine learning / Advanced

Tucker decomposition

Definition

Tucker decomposition represents a multiway array, or tensor, using a smaller core tensor and one factor matrix for each mode. It generalizes matrix-style dimensionality reduction by allowing a separate reduced dimension along each direction of the array.

What it means

A dataset may naturally have several axes, such as product, store and month. Flattening it into a matrix can hide that organization. Tucker decomposition keeps the modes explicit while representing their interactions through a core and factor matrices.

The factors describe components along each mode, and the core describes how those components combine. A reduced-size core can provide compression or a useful representation, provided the approximation retains the information needed for the task.

Tucker decomposition differs from CP decomposition, which expresses a tensor as a sum of rank-one components. Tucker permits a more general core. Its factors are not generally unique without additional constraints, so a component should not automatically be interpreted as a single real-world cause or category.

How it works

For a three-way tensor, choose reduced dimensions R1, R2 and R3. The factor matrices map those reduced coordinates to the original dimensions, while the R1-by-R2-by-R3 core combines them through mode-wise products.

A higher-order singular value decomposition can supply factors from matrix unfoldings of the tensor. Truncating these factors gives an approximation, but a truncated HOSVD is not generally the exact best approximation for every chosen multilinear rank. Iterative procedures such as higher-order orthogonal iteration can refine a solution.

Rank choices determine the balance between compression and reconstruction error. Scaling, missing entries and nonnegativity requirements can change the appropriate formulation. Evaluation should examine both global error and the parts of the tensor important to the application. A small average error can conceal poor reconstruction of rare but consequential patterns.

Three practical examples

1. A retail analyst represents sales across products, stores and months while retaining the three-way organization.

2. A scientific-computing team compresses a large multidimensional simulation output using a low multilinear-rank approximation.

3. A signal-processing researcher explores components across sensors, frequencies and time, treating component interpretation as a separate analytical task.

Think of it this way

Think of describing a complex three-way table through smaller dictionaries for each axis and a compact table of how their entries interact. The limit is that the mathematical factors may be rotated or transformed without changing the reconstruction, so they are not necessarily unique meaningful categories.

Worked example

Suppose a tensor has dimensions 20 × 30 × 40, containing 24,000 values. A Tucker representation with ranks 3, 4 and 5 has a core containing 3 × 4 × 5 = 60 values.

The factor matrices contain 20 × 3 + 30 × 4 + 40 × 5 = 380 values. Together, core and factors contain 440 values, excluding metadata and implementation overhead. That is about 1.83% of the original entry count.

This is a storage-count illustration, not proof of acceptable compression. The chosen ranks may produce substantial reconstruction error, and the value of the representation depends on what information must be preserved.

Illustrative case study

A fictional retail-planning team compresses a product-store-month tensor. Overall reconstruction error is small, but seasonal items with short sales bursts are poorly represented.

The team evaluates error by product type and time period, then adjusts the rank choices and compares with a simpler baseline. It avoids describing factor columns as definitive customer or store types without supporting evidence.

The case illustrates two separate questions: whether the tensor can be represented compactly and whether the compressed representation preserves the patterns needed for planning. One favorable average statistic cannot answer both.

Common misconceptions and mistakes

Tucker rank is one number identical to matrix rank. It is commonly a tuple of mode-specific ranks.
A truncated HOSVD always gives the exact best approximation. It provides a useful construction, but further optimization can matter.
Factor components have unique real-world meanings. Nonuniqueness and modeling choices complicate interpretation.

Business use

Tucker decomposition can be useful when data have meaningful multiple axes and low-dimensional structure along those axes. It can support compression, exploration and specialized predictive pipelines.

Select ranks using a clear quality criterion and evaluate important slices separately. Document preprocessing and missing-data treatment. Compare storage and computation costs including the core and all factors. If components are used in a business explanation, validate their interpretation with domain evidence rather than assigning labels solely from a visually appealing factor plot.

Check your understanding

Why is counting only the core's entries an incomplete estimate of Tucker storage requirements?

Answer: Reconstruction also requires the factor matrices for every mode, along with any necessary metadata.

Related terms

Dimensionality reduction
tensor
multilinear rank
higher-order SVD
CP decomposition
factor matrix
compression

Learn next

Read tensors and matrix singular value decomposition. Count the core and factor parameters for a small Tucker model, then compare compression with reconstruction error on important data slices.

Singular value decomposition

Sources

Tamara G. Kolda and Brett W. Bader, Tensor Decompositions and Applications (2009): https://www.cs.cornell.edu/courses/cs6241/2020sp/readings/Kolda-Bader-2009-survey.pdf . Independent review read section 4, including the core/factor construction, mode ranks and truncated HOSVD limitations. Tensor dimensions and the fictional retail case are original illustrations, not measured compression results.

Previous: Tsetlin machine Next: Tukey - Duckworth test