Tensor Robust Principal Component Analysis

• Tensor Robust Principal Component Analysis is a family of methods that decomposes multi-dimensional tensors into low-rank and sparse components to isolate structured variation from outliers.
• The framework includes diverse formulations such as t-SVD, Tucker, and atomic-norm models, each offering unique trade-offs in recovery guarantees and computational efficiency.
• TRPCA is applied in video background subtraction, image denoising, and hyperspectral imaging, leveraging both convex and nonconvex optimization techniques for robust data recovery.

Tensor robust principal component analysis (TRPCA) is the tensorial extension of robust principal component analysis to multi-dimensional arrays, with the central objective of decomposing an observed tensor into a low-rank component and a sparse component. In its most common form, one observes a tensor $\mathcal{X}$ and seeks $\mathcal{L}$ and $\mathcal{S}$ such that $\mathcal{X}=\mathcal{L}+\mathcal{S}$, where $\mathcal{L}$ captures structured, low-dimensional variation and $\mathcal{S}$ captures gross corruptions, outliers, or moving foregrounds (Lu et al., 2018). Because tensor rank is not unique, TRPCA is not a single model class but a family of formulations built on distinct tensor notions, including t-SVD/tubal-rank models, Tucker-rank models, and atomic-norm or CP-style models (Geng et al., 2022, Driggs et al., 2019).

1. Formal problem and model classes

The canonical TRPCA observation model is

$\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$

with $\mathcal{L}_0$ low-rank and $\mathcal{S}_0$ sparse. In the t-SVD framework, the standard convex program is

$$\min_{\mathcal{L},\mathcal{S}} \ \|\mathcal{L}\|_*+\lambda\|\mathcal{S}\|_1 \quad \text{s.t.}\quad \mathcal{X}=\mathcal{L}+\mathcal{S},$$

where $\mathcal{L}$0 is a tensor nuclear norm induced by the t-product and $\mathcal{L}$1 is the entrywise $\mathcal{L}$2-norm (Lu et al., 2018). This model is the direct tensor analogue of matrix RPCA.

A common simplification is to treat TRPCA as synonymous with the t-SVD/TNN formulation. The literature is broader. One major branch uses Tucker-rank constraints and solves

$\mathcal{L}$3

typically through alternating projections or factorized optimization (Cai et al., 2023, Dong et al., 2022). Another branch uses the tensor atomic norm, defining low-rankness through rank-1 unit-norm atoms $\mathcal{L}$4 and solving

$\mathcal{L}$5

with $\mathcal{L}$6 (Driggs et al., 2019).

This multiplicity of models reflects a basic fact stated explicitly in the literature: tensor rank is not unique, and different TRPCA variants correspond to different tensor notions of low-rankness (Geng et al., 2022). A precise reading of TRPCA therefore requires attention to the underlying tensor algebra and rank surrogate.

2. t-SVD, tensor nuclear norms, and convex TRPCA

The most influential TRPCA line is based on the tensor-tensor product, or t-product, together with tensor singular value decomposition. For a third-order tensor $\mathcal{L}$7, the t-product is defined through a block circulant embedding or, equivalently, by applying the discrete Fourier transform along the third mode so that multiplication decouples slicewise in the Fourier domain (Lu et al., 2018). Under this framework, a tensor admits a t-SVD

$\mathcal{L}$8

where $\mathcal{L}$9 and $\mathcal{S}$0 are orthogonal tensors and $\mathcal{S}$1 is $\mathcal{S}$2-diagonal (Lu et al., 2018).

The associated tubal rank is the number of non-zero singular tubes in $\mathcal{S}$3. Lu et al. define a tensor average rank

$\mathcal{S}$4

and a tensor nuclear norm

$\mathcal{S}$5

A central theoretical result is that this tensor nuclear norm is the convex envelope of tensor average rank on the unit ball of the tensor spectral norm (Lu et al., 2018). This gives the t-SVD-based TRPCA model the same convex-analytic status that the matrix nuclear norm has in matrix RPCA.

Algorithmically, the convex program is usually solved by ADMM. The low-rank update is a proximal step for the tensor nuclear norm and becomes tensor singular value thresholding: FFT along mode 3, SVD on each frontal slice, soft-thresholding of singular values, and inverse FFT. The sparse update is standard entrywise soft-thresholding, followed by a dual update (Lu et al., 2018). When $\mathcal{S}$6, the entire construction reduces exactly to matrix RPCA (Lu et al., 2018).

A transform-based generalization replaces the discrete Fourier transform with a general invertible linear transform $\mathcal{S}$7 satisfying $\mathcal{S}$8. In that setting, the tensor nuclear norm becomes

$\mathcal{S}$9

and the corresponding TRPCA model reduces to matrix RPCA when the tensor degenerates to a matrix (Lu et al., 2019). This makes explicit that t-SVD/TRPCA is not tied uniquely to the Fourier transform.

3. Identifiability and recovery theory

The exact recovery theory of TRPCA closely parallels that of matrix RPCA, but with tensor-specific incoherence and sparsity conditions. In the probabilistic t-SVD framework, exact recovery holds under tensor incoherence conditions on the singular tensor subspaces, random sparse support, and a parameter choice

$\mathcal{X}=\mathcal{L}+\mathcal{S}$0

with the low-rank tensor recoverable when tubal rank and sparsity are sufficiently small relative to the tensor dimensions (Lu et al., 2018). The proof follows the dual-certificate style familiar from matrix RPCA.

A deterministic alternative replaces randomness by a tensor rank-sparsity incoherence principle. In this formulation, one introduces the tangent spaces $\mathcal{X}=\mathcal{L}+\mathcal{S}$1 and $\mathcal{X}=\mathcal{L}+\mathcal{S}$2, together with two quantities: $\mathcal{X}=\mathcal{L}+\mathcal{S}$3 The decomposition is unique if $\mathcal{X}=\mathcal{L}+\mathcal{S}$4, and a sufficient condition is

$\mathcal{X}=\mathcal{L}+\mathcal{S}$5

The main deterministic exact recovery theorem strengthens this to

$\mathcal{X}=\mathcal{L}+\mathcal{S}$6

under which the convex RTPCA program exactly recovers $\mathcal{X}=\mathcal{L}+\mathcal{S}$7 for an explicit interval of regularization parameters $\mathcal{X}=\mathcal{L}+\mathcal{S}$8 (Shen et al., 2020). The same work states the tensor rank-sparsity uncertainty principle

$\mathcal{X}=\mathcal{L}+\mathcal{S}$9

for every nonzero tensor $\mathcal{L}$0 (Shen et al., 2020).

The atomic-norm line establishes a different guarantee. For $\mathcal{L}$1, exact recovery by

$\mathcal{L}$2

holds with $\mathcal{L}$3 under coherence bounds $\mathcal{L}$4, $\mathcal{L}$5, a random support model for $\mathcal{L}$6, and a rank condition on the weighted average Tucker rank $\mathcal{L}$7 (Driggs et al., 2019). That paper states that the resulting guarantees improve on existing tensor-RPCA guarantees, including those for matrix RPCA (Driggs et al., 2019).

More recent nonconvex theory replaces convex surrogates by scale-invariant ratios. In the TNF model,

$\mathcal{L}$8

and the TRPCA objective becomes

$\mathcal{L}$9

Under incoherence and random-support conditions similar in spirit to convex TRPCA, one can prove that the ground-truth pair is a local minimizer when the low-rank tensor is sufficiently low in rank and the sparse tensor is sufficiently sparse (Zheng et al., 7 May 2025). The result is explicitly local rather than global.

4. Algorithmic developments and scalability

The original convex TRPCA literature is dominated by ADMM with tensor singular value thresholding and soft-thresholding (Lu et al., 2018). This remains the reference optimization scheme for t-SVD/TNN models, but it is not the only algorithmic template.

A major shift is toward factorized nonconvex methods. Under Tucker decomposition, one can directly recover the low-dimensional tensor factors through scaled gradient descent coupled with iteration-varying thresholding. The 2022 Tucker-based ScaledGD method starts from a tailored spectral initialization, updates the sparse component by thresholding, and then performs preconditioned gradient steps on the Tucker factors and core. Its main theoretical result is linear convergence to the true low-rank tensor at a constant rate independent of the condition number, provided the corruption level is not too large (Dong et al., 2022).

A t-SVD analogue appears in the 2025 RTPCA-SGD model, which factorizes the low-rank tensor as

$\mathcal{S}$0

and alternates soft-thresholding for the sparse component with scaled gradient updates for $\mathcal{S}$1 and $\mathcal{S}$2. The paper states recovery guarantees under mild assumptions and linear convergence at a constant rate independent of the condition number, while emphasizing lower runtime than TNN-based RTPCA (Feng et al., 8 Jan 2025). The same work introduces a learnable self-supervised deep unfolding model, RTPCA-LSGD, with learnable $\mathcal{S}$3, $\mathcal{S}$4, $\mathcal{S}$5, and $\mathcal{S}$6 (Feng et al., 8 Jan 2025).

For large-scale Tucker-rank problems, Robust Tensor CUR Decompositions (RTCUR) replace full low-rank projections by tensor CUR sketches inside an alternating-projection scheme. RTCUR projects between the low-rank tensor set and the sparse tensor set, and develops four variants—RTCUR-FF, RTCUR-FC, RTCUR-RF, and RTCUR-RC—for different application settings (Cai et al., 2023). In the regime considered in that paper, the per-iteration complexity is $\mathcal{S}$7 for Fiber CUR and $\mathcal{S}$8 for Chidori CUR, as compared with at least $\mathcal{S}$9 for standard HOSVD-based Tucker projections (Cai et al., 2023).

These developments support a useful corrective to a common misconception: TRPCA algorithms do not uniformly rely on repeated tensor SVD of the full tensor. The current algorithmic landscape includes convex proximal methods, factorized scaled-gradient methods, deep unfolding, and CUR-based alternating projections (Dong et al., 2022, Feng et al., 8 Jan 2025, Cai et al., 2023).

5. Weighted, transformed, and nonconvex extensions

A recurrent critique of standard TNN-based TRPCA is that it shrinks all tensor singular values equally. Several papers argue that this is overly biased because large singular values often encode more significant structure than small ones. The tensor adjustable logarithmic norm,

$\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$0

is one response: it yields weights $\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$1, so that larger singular values have smaller weights and are penalized less (Geng et al., 2022). This leads to a nonconvex N-TRPCA model and, after patch grouping based on nonlocal self-similarity, to NN-TRPCA, where each group tensor is recovered separately (Geng et al., 2022).

A related nonconvex line replaces both the tensor nuclear norm and the sparse $\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$2-norm by folded-concave penalties. In one formulation,

$\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$3

where $\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$4 is a SCAD/MCP-type tensor $\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$5-norm and $\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$6 is a SCAD/MCP-type sparsity measure (Li et al., 2019). The corresponding algorithm uses majorization minimization together with generalized tensor singular value thresholding and weighted soft-thresholding (Li et al., 2019).

Weighting can also be imposed directly in the Fourier domain. Frequency-Weighted RTPCA introduces a frequency-filtered tensor nuclear norm

$\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$7

with band-dependent coefficients $\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$8, and derives a frequency-filtered tensor SVT operator for the low-rank subproblem (Wang et al., 2020). Global Weighted TRPCA (GWTRPCA) goes further by weighting both intra-frontal-slice and inter-frontal-slice singular values in the Fourier domain, and proposes an adaptive weight learning strategy by a Modified Cauchy Estimator (Wang et al., 2022).

Transform-based TRPCA generalizes these ideas by allowing any invertible transform $\mathcal{X}=\mathcal{L}_0+\mathcal{S}_0,$9 satisfying $\mathcal{L}_0$0, thereby replacing Fourier-specific t-SVD by a broader class of tensor algebras (Lu et al., 2019). The transform choice becomes application-dependent; the image recovery experiments in that paper explicitly compare DCT- and ROM-based variants (Lu et al., 2019).

6. Applications, benchmarks, and domain-specific interpretations

TRPCA has been applied to background/foreground separation in videos, image denoising, impulse noise removal, moving object tracking, background modeling, hyperspectral image denoising, healthcare, manufacturing processes, and general tensor data in signal processing (Lu et al., 2018, Shen et al., 2020, Wang et al., 2022). These applications reflect the same decomposition semantics: low-rank components capture slowly varying or shared structure, while sparse components isolate anomalies, outliers, or moving objects.

The t-SVD/TRPCA model is especially prominent in image and video settings. Standard use cases include stacking frames into $\mathcal{L}_0$1, recovering $\mathcal{L}_0$2 as the background and $\mathcal{L}_0$3 as moving objects or sparse corruption, and then evaluating background quality, foreground masks, PSNR, SSIM, FSIM, AGE, pEPs, pCEPs, or runtime depending on the task (Geng et al., 2022, Feng et al., 8 Jan 2025). Tucker-rank TRPCA has also been used for color video background subtraction and network clustering on co-authorship tensors (Cai et al., 2023).

A domain-specific and methodologically atypical example is the application to lightning images. That work treats a sequence of lightning frames as a 3-way tensor, decomposes it into a low-rank tensor representing stationary background and a sparse tensor representing lightning and other dynamic features, and then applies an SVD “unfold” technique specifically to the sparse tensor (Yilmaz et al., 2022). The paper reports principal component vectors $\mathcal{L}_0$4, $\mathcal{L}_0$5, and $\mathcal{L}_0$6, 2D energy density contour maps, 3D vector (ether) fields, Lyapunov exponent, approximate entropy, and fractal dimension; it further reports that Frame 78 has a fractal dimension of $\mathcal{L}_0$7, which it compares with the spectrum of the Fibonacci Hamiltonian (Yilmaz et al., 2022). These interpretations are specific to that application and should not be read as general consequences of TRPCA itself.

Across these lines of work, a consistent conclusion emerges. TRPCA is best understood not as a single algorithm, but as a structured family of low-rank-plus-sparse decompositions whose precise meaning depends on the tensor model, rank surrogate, transform, and optimizer. The field spans convex exact-recovery theory, deterministic identifiability, weighted and nonconvex rank surrogates, factorized gradient methods, and application-specific pipelines built on the recovered sparse and low-rank components (Lu et al., 2018, Shen et al., 2020, Feng et al., 8 Jan 2025).