Global Low Rank and Local Smoothness for Tensor Recovery

This direction is merged into another direction: hybrid local-nonlocal regularization. Its last version stays readable and citable here.

Version 1 of 122 papers30 conceptsfrozen 2026-10-04

When we try to reconstruct a corrupted video, a grainy medical scan, or a sparse hyperspectral image, we are solving a "tensor completion" problem. High-dimensional data—tensors—rarely consist of random noise; they possess deep internal structures. However, traditional mathematical tools often force a trade-off: you can either capture the broad, global patterns of the data or the sharp, local details, but rarely both at once.

A powerful research direction has emerged to bridge this gap by combining global low-rank constraints with local smoothness priors. By treating these two properties as complementary forces, researchers can reconstruct complex signals even when the vast majority of the data is missing.

The Global Skeleton: Low-Rank Redundancy

The first pillar of this approach is the observation that multi-dimensional data is highly redundant. In a video, adjacent frames are nearly identical; in a medical MRI, different slices share the same anatomical geometry. Mathematically, this means the data can be represented as a "low-rank" tensor.

Low-rank tensor completion (LRTC) acts like a global skeleton. It looks across all dimensions simultaneously to find the most efficient representation of the data. As Haijin Zeng and colleagues demonstrate in Tensor completion using enhanced multiple modes low-rank prior and total variation, exploiting these low-rank properties across all modes allows for the recovery of data from as little as 2.5% of the original pixels.

However, a purely low-rank approach has a weakness: it tends to over-smooth. Because it prioritizes the most dominant global patterns, it often discards high-frequency information, leaving the reconstructed image looking blurry or "blocky."

The Local Skin: Smoothness and Texture

To fix the blurring caused by global models, researchers introduce a second pillar: local regularization. The most common tool here is Total Variation (TV), which penalizes sharp, noisy jumps between adjacent pixels while allowing for "piecewise smooth" regions.

If the low-rank constraint is the skeleton, the local prior is the skin. Xutao Li and Yunming Ye showed in Low-Rank Tensor Completion with Total Variation for Visual Data Inpainting that this combination prevents reconstruction artifacts like ringing. The low-rank prior ensures the broad structural consistency, while the TV term preserves the sharp edges of objects.

This synergy is particularly effective for simultaneous tasks. Tatsuya Yokota and Hidekata Hontani leveraged this in Simultaneous Visual Data Completion and Denoising, proving that the same dual-prior framework can fill in missing pixels and strip away electronic noise at the same time.

Beyond Smoothness: From Gradients to Deep Learning

While Total Variation is excellent for preserving edges, it struggles with complex textures like the weave of a fabric or the grain of wood. Recent work has evolved the "local" component of the framework to be more expressive:

  1. Convolutional Sparse Coding (CSC): Instead of just looking at pixel gradients, CSC learns a dictionary of local patterns. Chuan Chen et al. in Tensor Completion via Few-shot Convolutional Sparse Coding used this to recover high-frequency textures that standard low-rank methods typically blur.
  2. Plug-and-Play (PnP) Denoisers: Rather than using a fixed mathematical formula for local smoothness, researchers like Wen-Hao Xu and Michael Ng "plug" pre-trained deep learning denoisers into the low-rank framework. In Deep plug-and-play prior for low-rank tensor completion, they show this allows the model to learn what "natural" local textures look like from data, rather than relying on human-defined rules.
  3. Generative Diffusion Priors: The most recent frontier involves using diffusion models as the local prior. Zerui Tao and Qibin Zhao introduced DiffBCP in Bayesian Tensor Decomposition with Diffusion Model Prior, which uses the generative power of diffusion to "hallucinate" missing high-frequency details while the low-rank component keeps the hallucination anchored to the global structure of the original data.

Why the Hybrid Approach Wins

The success of this direction lies in its balance of constraints and freedom. A purely global model is too constrained (everything must fit a simple low-rank pattern), leading to blur. A purely local model has too much freedom (it doesn't know how a pixel in the top-left corner relates to one in the bottom-right), leading to structural incoherence.

By tying them together—as seen in the work of Hongbing Zhang on Fast and Accurate Low-Rank Tensor Completion—the global prior provides the "where" (the coordinates of the structure) and the local prior provides the "what" (the sharp detail of the texture). This dual-scale processing is now the gold standard for high-fidelity reconstruction in fields ranging from satellite hyperspectral imaging to sparse-view CT scans.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_global_low_rank_and_local_smoothness_for_963e7d25,
  title = {Global Low Rank and Local Smoothness for Tensor Recovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-low-rank-and-local-smoothness-for-tensor-recovery/rdr_963e7d25f22f427ea58549cec1b324a0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_global_low_rank_and_local_smoothness_for_963e7d25_v1,
  title = {Global Low Rank and Local Smoothness for Tensor Recovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-low-rank-and-local-smoothness-for-tensor-recovery/rdr_963e7d25f22f427ea58549cec1b324a0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}