Low-Rank Tucker Representation (LRTuckerRep)

• Low-Rank Tucker Representation is a tensor decomposition method that expresses data as a multilinear product of a small core tensor and orthonormal factor matrices under bounded Tucker ranks.
• It underpins various applications by enabling statistical, probabilistic, and nonnegative interpretations, useful in array analysis, completion, and inverse problems.
• Recent advances include recovery algorithms and scalable randomized, streaming constructions that ensure efficient compression and robust performance in high-dimensional settings.

Searching arXiv for the cited Tucker-rank papers to ground the article in current arXiv records. Searching arXiv for "Stochastic Iterative Hard Thresholding for Low-Tucker-Rank Tensor Recovery". I’m going to look up the key arXiv entries relevant to Low-Rank Tucker Representation. Low-Rank Tucker Representation (LRTuckerRep) denotes the representation of a tensor, or of a tensorized operator block, by a small core tensor coupled to mode-wise factor matrices under a bounded Tucker multilinear-rank constraint. In its standard form, a $d$-th order tensor is modeled as a multilinear product of a core and mode subspaces, and the representation is used not only for compression but also as a model class for probabilistic array analysis, completion, inverse problems, randomized sketching, streaming approximation, and time-dependent low-rank evolution (Grotheer et al., 2019, Hoff, 2013, Haselby et al., 2023).

1. Formal representation and multilinear rank

The canonical Tucker representation writes a tensor as

$$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$

with core $\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$ and factor matrices $U_k\in\mathbb R^{n_k\times r_k}$. In vectorized form this becomes

$m=(U_K\otimes\cdots\otimes U_1)s,$

where $m=\operatorname{vec}(M)$ and $s=\operatorname{vec}(S)$. The Tucker rank is the tuple of unfolding ranks,

$r_k=\operatorname{rank}(X_{(k)}),$

so low rank is inherently mode-wise and multilinear rather than a single scalar condition. In many formulations the factors are orthonormal, $U_k^TU_k=I_{r_k}$, which aligns the model with HOSVD-style subspace extraction and simplifies both inference and optimization (Grotheer et al., 2019, Hoff, 2013).

This representation generalizes beyond ordinary data tensors. In hierarchical Tucker low-rank matrices, an admissible interaction block between tensor-product row and column clusters is represented as a $2d$-order Tucker tensor,

$$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$0

or equivalently

$$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$1

For equal mode sizes $$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$2 and equal Tucker ranks $$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$3, Tucker storage scales as $$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$4, versus $$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$5 for the full tensor, making LRTuckerRep a concrete compression model rather than only an abstract rank condition (Li et al., 8 Aug 2025, Minster et al., 2019).

2. Statistical, probabilistic, and nonnegative interpretations

One major interpretation of LRTuckerRep is statistical rather than purely algebraic. In the Tucker decomposition model for array data, the observed array is not decomposed directly; instead, the low-rank Tucker structure is placed on a mean array $$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$6 for Gaussian data or on a latent Gaussian array $$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$7 for discrete, ordinal, or non-normal data. The Gaussian model takes the form

$$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$8

while the scale-free semiparametric extension uses

$$\mathcal X = \mathcal G \times_1 U_1 \times_2 \cdots \times_d U_d,$$9

with inference based on a rank likelihood invariant to strictly increasing transformations. Orthogonal equivariance is enforced through uniform priors on Stiefel manifolds and a $\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$0 invariant prior, while shrinkage on the core is induced by

$\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$1

This makes LRTuckerRep a probabilistic model for mean or latent structure, not merely a least-squares approximation of raw entries (Hoff, 2013).

A different line of work treats nonnegative tensors through alternating projections between the nonnegative orthant and the Tucker-rank-bounded set. There the target approximation tensor must be nonnegative, but the core and factors themselves are not required to be nonnegative. The practical iteration alternates

$\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$2

with approximate Tucker projection by STHOSVD or randomized STHOSVD, yielding a low-rank Tucker representation whose reconstruction becomes increasingly close to entrywise nonnegative over iterations. The empirical outcome is that the final approximation error remains close to the initial STHOSVD error while the negative part decays substantially (Sultonov et al., 2022).

An information-geometric interpretation sharpens the relation between low Tucker rank and probabilistic factorization for strictly positive tensors. In that framework, a rank-$\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$3 tensor is exactly the KL-optimal mean-field approximation, and repeated rank-1 KL projections on selected subtensors enforce “bingo” conditions in canonical parameters that reduce Tucker ranks mode by mode. This yields a non-gradient method for nonnegative Tucker rank reduction based on closed-form marginal computations rather than alternating optimization over a core and factor matrices. This suggests that LRTuckerRep can also be understood as a structured independence model on the index set of a positive tensor (Ghalamkari et al., 2021).

3. Geometry, identifiability, and constrained varieties

The bounded-rank feasible set for LRTuckerRep is the Tucker tensor variety

$\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$4

which is the closure of the smooth fixed-rank manifold $\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$5. At rank-deficient points, tangent spaces are no longer sufficient; the correct first-order object is the Bouligand tangent cone $\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$6. An explicit parametrization of this cone shows that any feasible first-order direction can be decomposed into a core/subspace-enlargement part and mode-specific variation parts,

$\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$7

with orthogonality relations across components. This geometry supports gradient-related line-search methods on $\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$8, HOSVD-based retraction, retraction-free updates, and a Tucker rank-adaptive method, TRAM, that alternates fixed-rank optimization with rank-decreasing and rank-increasing steps (Gao et al., 2023).

A complementary theory addresses when a low-Tucker-rank representation is identifiable from incomplete observations. For a partially observed tensor $\mathcal G\in\mathbb R^{r_1\times\cdots\times r_d}$9, finite or unique completion is characterized in terms of algebraic independence of polynomial constraints induced by the sampling pattern after quotienting out Tucker nonuniqueness through a canonical core. The dimension of the fixed-rank Tucker manifold is

$U_k\in\mathbb R^{n_k\times r_k}$0

and the sampling-pattern theory provides deterministic necessary-and-sufficient conditions for finite completability when a nonempty proper subset of Tucker rank components is specified, as well as probabilistic lower bounds on entrywise sampling probabilities that guarantee finite or unique completion with high probability. A notable implication is that joint Tucker structure can identify a tensor even when each unfolding, considered separately as a low-rank matrix completion problem, remains ambiguous (Ashraphijuo et al., 2016).

These two viewpoints are closely related. The variety perspective explains how optimization can move safely on a closed bounded-rank set, while the completion theory clarifies when the underlying LRTuckerRep is algebraically determined by data. Together they show that LRTuckerRep is not only a compact parametrization but also a geometric and identifiability structure.

4. Recovery, inverse problems, and dynamical optimization

In compressive recovery, LRTuckerRep appears as the constraint set in

$U_k\in\mathbb R^{n_k\times r_k}$1

with

$U_k\in\mathbb R^{n_k\times r_k}$2

A stochastic iterative hard-thresholding method, StoTIHT, alternates a batchwise stochastic gradient step with Tucker truncation,

$U_k\in\mathbb R^{n_k\times r_k}$3

where $U_k\in\mathbb R^{n_k\times r_k}$4 is an HOSVD-based approximate projection onto the Tucker-rank-$U_k\in\mathbb R^{n_k\times r_k}$5 model class. Under a tensor restricted isometry property for the full sensing map and each batch operator, and under an $U_k\in\mathbb R^{n_k\times r_k}$6-approximate projection condition, the method satisfies an expected linear-convergence bound

$U_k\in\mathbb R^{n_k\times r_k}$7

Here hard thresholding is the mechanism that repeatedly re-compresses unrestricted iterates into a Tucker representation after each stochastic update (Grotheer et al., 2019).

In magnitude-only inverse problems, Tucker-Structured Phase Retrieval replaces the matrix model $U_k\in\mathbb R^{n_k\times r_k}$8 by a third-order tensor model

$U_k\in\mathbb R^{n_k\times r_k}$9

with each frame satisfying

$m=(U_K\otimes\cdots\otimes U_1)s,$0

The algorithm uses spectral initialization followed by alternating minimization: each $m=(U_K\otimes\cdots\otimes U_1)s,$1 is updated by a low-dimensional phase retrieval subproblem, while $m=(U_K\otimes\cdots\otimes U_1)s,$2, $m=(U_K\otimes\cdots\otimes U_1)s,$3, and $m=(U_K\otimes\cdots\otimes U_1)s,$4 are updated by CGLS on the phase-corrected least-squares objective

$m=(U_K\otimes\cdots\otimes U_1)s,$5

The practical motivation is parameter reduction relative to vectorized low-rank matrix models, especially in under-sampled regimes; however, the method is explicit about lacking a recovery theorem or sample-complexity analysis (Kwon et al., 2022).

For time-dependent tensors and tensor differential equations, LRTuckerRep becomes a dynamical state constrained to the Tucker manifold,

$m=(U_K\otimes\cdots\otimes U_1)s,$6

A nested Tucker projector-splitting integrator recursively applies matrix projector splitting to tensor unfoldings while treating the L-step inexactly through a lower-dimensional Tucker problem. The method exactly reconstructs time-dependent Tucker tensors of the prescribed rank under a mild invertibility condition and is robust to small singular values in the unfoldings, with error bound

$m=(U_K\otimes\cdots\otimes U_1)s,$7

This turns LRTuckerRep into a stable time-evolving reduced model rather than a static decomposition (Lubich et al., 2017).

5. Randomized, streaming, sampled, and hybrid constructions

Randomization provides one route to scalable construction of LRTuckerRep from large tensors. Randomized HOSVD and randomized STHOSVD replace full SVDs of unfoldings by randomized range finders, yielding Tucker approximations

$m=(U_K\otimes\cdots\otimes U_1)s,$8

with expectation bounds of the form

$m=(U_K\otimes\cdots\otimes U_1)s,$9

where $m=\operatorname{vec}(M)$0 is the discarded spectral tail of the $m=\operatorname{vec}(M)$1-th unfolding. Adaptive variants determine ranks from a target tolerance, and a structure-preserving variant, SP-STHOSVD, keeps the core as sampled entries from the original sparse tensor instead of forming dense intermediate projections (Minster et al., 2019).

When the full tensor is unavailable in memory, compressive sketches can still recover an explicit Tucker representation. One-pass recovery from streamed measurements uses leave-one-out sketches of each unfolding and a core sketch to produce

$m=\operatorname{vec}(M)$2

with core recovered by

$m=\operatorname{vec}(M)$3

The recovered tensor satisfies a near-best Tucker approximation guarantee, up to explicit multiplicative constants, relative to the optimal rank-$m=\operatorname{vec}(M)$4 Tucker approximation. The same paper analyzes both Kronecker- and Khatri–Rao-structured measurement ensembles, emphasizing low memory, one-pass acquisition, and recovery runtime sublinear in the size of the ambient tensor after sketch acquisition (Haselby et al., 2023).

A closely related streaming line is the sequential multilinear Nyström method, which assumes access only to sketches of a tensor given as a sum of streamed components,

$m=\operatorname{vec}(M)$5

It recursively accumulates modewise sketches $m=\operatorname{vec}(M)$6, $m=\operatorname{vec}(M)$7, and a compressed core sketch $m=\operatorname{vec}(M)$8, and recovers a Tucker approximation

$m=\operatorname{vec}(M)$9

from QR factorizations and small least-squares systems. The deterministic error analysis is framed in terms of residuals of randomized range finding applied to sequentially compressed unfoldings, and the numerical experiments show substantial speedups over non-sequential multilinear Nyström in streaming settings (Bucci et al., 2024).

Sparse sampling can replace sketching altogether. DEIM fiber sampling constructs a Tucker tensor cross approximation by evaluating only selected fibers. For a rank-$s=\operatorname{vec}(S)$0 approximation of an $s=\operatorname{vec}(S)$1 tensor, the method requires access to only $s=\operatorname{vec}(S)$2 tensor entries. Modewise DEIM selects index sets $s=\operatorname{vec}(S)$3, sampled fiber bundles are assembled into matrices $s=\operatorname{vec}(S)$4, factor matrices are extracted by truncated SVD, and the core is recovered from the sampled intersection tensor by

$s=\operatorname{vec}(S)$5

An iterative black-box variant removes the need for singular vectors of the full unfoldings and is used to compress full-rank right-hand sides in dynamical low-rank approximation (Ghahremani et al., 2024).

Finally, Tucker can also act as an outer shell around a more compressed internal model. In tensor completion with a Tucker wrapper, the tensor is represented as

$s=\operatorname{vec}(S)$6

where the outermost factor matrices are orthonormal and the core itself is stored as a tensor network. This hybrid preserves Tucker-style mode subspaces while reducing the parameter count of the core, and the associated two-level ALS method enjoys linear-rate exact-recovery guarantees under stated assumptions (Cai et al., 2020).

6. Applications, trade-offs, and limitations

The application range of LRTuckerRep is broad. In array statistics it provides scale-free latent structure for discrete multivariate networks, as in the GDELT example with tensor size $s=\operatorname{vec}(S)$7, where the first two singular vectors in the country modes capture broad activity heterogeneity and additional centered factors reveal geographic and non-geographic patterns (Hoff, 2013). In signal recovery it reconstructs synthetic and real tensors from compressed measurements, including a candle video of size $s=\operatorname{vec}(S)$8 with assumed Tucker rank $s=\operatorname{vec}(S)$9 (Grotheer et al., 2019). In phase retrieval it models video tensors such as Mouse $r_k=\operatorname{rank}(X_{(k)}),$0 and Plane $r_k=\operatorname{rank}(X_{(k)}),$1 and empirically outperforms matrix low-rank baselines in under-determined and some over-determined regimes when the Tucker ranks are chosen appropriately (Kwon et al., 2022).

Compression-oriented uses are equally diverse. A multi-exposure stereo stack is encoded through a Tucker-ALS approximation of $r_k=\operatorname{rank}(X_{(k)}),$2, then passed to 3D-HEVC; the method treats bitrate as controllable by Tucker rank and quantization, and reports substantial savings relative to JPEG-XT and direct 3D-HEVC coding on the reported scenes (Sharma et al., 2021). In hyperspectral imaging, alternating projection to a nonnegative Tucker approximation on a $r_k=\operatorname{rank}(X_{(k)}),$3 tensor is reported with compression factor $r_k=\operatorname{rank}(X_{(k)}),$4 while keeping approximation quality close to the initial STHOSVD solution (Sultonov et al., 2022). In scientific computing, HTLR matrices replace classical low-rank $r_k=\operatorname{rank}(X_{(k)}),$5-matrix blocks by Tucker blocks, reducing the factor-storage term from $r_k=\operatorname{rank}(X_{(k)}),$6 to $r_k=\operatorname{rank}(X_{(k)}),$7 and yielding overall memory reductions of roughly $r_k=\operatorname{rank}(X_{(k)}),$8–$r_k=\operatorname{rank}(X_{(k)}),$9 and runtime speedups $U_k^TU_k=I_{r_k}$0–$U_k^TU_k=I_{r_k}$1 in the reported experiments (Li et al., 8 Aug 2025).

Several recurring limitations cut across the literature. Many algorithms assume the multilinear rank is known a priori, including StoTIHT, TSPR, and the DEIM-based constructions (Grotheer et al., 2019, Kwon et al., 2022, Ghahremani et al., 2024). HOSVD-based truncation is usually only an approximate projection onto the Tucker-rank set, not the exact best Tucker approximation, and this approximation factor enters both theory and practice (Grotheer et al., 2019, Sultonov et al., 2022). Some guarantees are strong but model-specific: TRIP assumptions in stochastic recovery, genericity and noiselessness in completion theory, strict positivity in the information-geometric rank-reduction method, or local regularity assumptions in Tucker-variety optimization (Grotheer et al., 2019, Ashraphijuo et al., 2016, Ghalamkari et al., 2021, Gao et al., 2023). Other methods are effective empirically but explicitly lack formal recovery theory, most notably TSPR (Kwon et al., 2022).

Several common misconceptions are corrected by this literature. Low Tucker rank is not a scalar-rank analogue; it is a tuple of mode-wise rank constraints. A Tucker representation need not be learned from full data, because it can be recovered from linear sketches, sparse fibers, streamed components, or magnitude-only observations. Conversely, Tucker factors are not automatically interpretable, nonnegative, or unique: identifiability requires additional structure, and nonuniqueness under orthogonal changes of basis persists unless a particular gauge or canonicalization is imposed. Taken together, these results position LRTuckerRep as a unifying multilinear representation framework whose importance lies as much in its recovery geometry and algorithmic tractability as in its compression capability.