Local Denoising Circulation Overview

• Local denoising circulation is a framework where denoising is performed using localized, structurally constrained evidence to preserve edges and textures.
• Methods include edge detection, adaptive smoothing with clustering or kernel regression, and circulant patch-group structures like Haar-tSVD to enhance robustness.
• In diffusion language models, it is formalized as a discrete curl that diagnoses order dependence, guiding circulation-aware regularization for improved inference.

Local denoising circulation denotes a family of local-information-flow viewpoints on denoising in which each estimate is formed from nearby, structurally constrained evidence rather than from a single global basis. In image denoising, it refers to edge-aware, patch-aware, or diffusion-based schemes in which denoising influence circulates within homogeneous regions, is redirected by local geometry, and is restricted across discontinuities; in diffusion LLMs, it is formalized as a discrete curl, namely the log-ratio between two order-induced pseudo-joints obtained by swapping a pair of unresolved positions (Basak et al., 2024, Kim, 10 May 2026).

1. Scope and terminological status

The phrase is not used uniformly across the literature. In several denoising papers, the underlying mechanism is said to “naturally support this interpretation,” even when the paper “does not use the word ‘circulation’” explicitly; in those cases, the term names a local, anisotropic propagation of denoising influence through neighborhoods, patches, or multiscale basis supports (Basak et al., 2024, Vasilyeva et al., 2024). By contrast, in diffusion LLMs the notion is explicit and formal: local denoising circulation is the elementary order-swap quantity that diagnoses whether arbitrary-order denoising is path independent (Kim, 10 May 2026).

Across these usages, two invariants recur. First, denoising is local in the operational sense: estimation is carried out from a local window, a local patch group, a local domain of basis support, or a local conditional over unresolved coordinates. Second, “circulation” refers either to a literal circulant representation, as in patch-group denoising with Haar-tSVD, or to a flow-like or curl-like quantity that measures how denoising influence propagates and whether it is conservative (Kong et al., 29 Dec 2025).

This breadth also marks an important distinction. In imaging, local denoising circulation is primarily a structural interpretation of how edge maps, local clusters, and local basis functions steer smoothing. In diffusion LLMs, it is a mathematically defined obstruction to compatibility: if circulation is zero on all reachable elementary squares, pseudo-joints are order invariant; if it is nonzero, decoding is intrinsically path dependent (Kim, 10 May 2026).

2. Edge-aware local circulation in adaptive image denoising

A canonical image-denoising formulation appears in the gray-scale model

$$z_{ij} = f(x_i,y_j) + \epsilon_{ij},\quad (x_i,y_j) = (i/n, j/n),\quad i,j=1,\dots,n,$$

with i.i.d. noise $\epsilon_{ij}$ of mean $0$, variance $\sigma^2$, and an image $f$ that is smooth except along jump location curves (JLCs). Locality enters in two distinct stages: edge detection and smoothing (Basak et al., 2024).

For edge detection, each pixel $(i,j)$ uses a local window

$$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$

to fit a local plane

$$\hat z_{ij}(x,y) = \hat\beta_0^{(i,j)} + \hat\beta_1^{(i,j)}(x-x_i) + \hat\beta_2^{(i,j)}(y-y_j),$$

and the gradient vector $$\vec v_{ij}=(\hat\beta_1^{(i,j)},\hat\beta_2^{(i,j)})$$ is compared with those of two neighboring pixels along the gradient direction. For smoothing, the method switches according to edge proximity. Far from edges, it performs local polynomial kernel regression inside an ellipse $E$ chosen so that no edge pixels lie inside $\epsilon_{ij}$0. Near edges, it uses a local clustering neighborhood

$\epsilon_{ij}$1

splits intensities into two clusters by maximizing

$\epsilon_{ij}$2

and estimates $\epsilon_{ij}$3 by a weighted average over pixels in the same cluster as $\epsilon_{ij}$4: $\epsilon_{ij}$5 The weights $\epsilon_{ij}$6 are patch-based similarities in a local feature space (Basak et al., 2024).

This yields the paper’s own interpretive description of local denoising circulation: noise is reduced by accumulating and averaging information from nearby, structurally similar pixels, with information flowing only locally in space and respecting edge boundaries. The mechanism is described as a “two-stage local diffusion process”: first diffusion of structural information through local gradient estimation and edge detection, then diffusion of intensity information within the discovered structure. The method is “not strictly multi-scale in the wavelet sense, but it has a multi-resolution flavor,” because edge detection uses a relatively larger window, smoothing uses smaller adapted neighborhoods, and clustering uses patch-based distances at a slightly larger scale (Basak et al., 2024).

The theoretical statements are local as well. Under the stated smoothness and Gaussian-noise assumptions, the estimated JLCs are Hausdorff-consistent with rate

$\epsilon_{ij}$7

for example with $\epsilon_{ij}$8 and $\epsilon_{ij}$9, $0$0. The clustering statistic $0$1 separates jump from no-jump regions almost surely under the stated conditions, the clustering-based estimator satisfies

$0$2

and the integrated estimator satisfies

$0$3

at non-singular points away from singularities (Basak et al., 2024).

Experimentally, the same local-circulation picture is tied to improved denoising under heavy noise. For the synthetic square-circle image at $0$4, the reported RMSEs are 9.81 for the ellipse method, 9.9 for clustering, and 11.18 for steering; for Peppers at $0$5, the ellipse method gives 10.27 versus 14.45 for steering and 10.53 for pure clustering. The reported qualitative pattern is that steering kernels are competitive at low noise, whereas the integrated local method is more robust at higher noise and better preserves clearly defined edges (Basak et al., 2024).

3. Circulant local representations and Haar-tSVD

A second major usage makes “circulation” literal through circulant patch-group structure. In Haar-tSVD, denoising follows the standard patch-based nonlocal self-similarity framework, but replaces per-group PCA learning with a fixed transform justified by circulant representation. For a group $0$6 of $0$7 similar patches, each vectorized as $0$8, the local group representation is

$0$9

whose Gram matrix $\sigma^2$0 is also circulant (Kong et al., 29 Dec 2025).

The crucial PCA–Haar link is explicit. If $\sigma^2$1 is a power of $\sigma^2$2, the first row of the Haar matrix $\sigma^2$3 is

$\sigma^2$4

which coincides with the dominant eigenvector

$\sigma^2$5

of the group covariance. This makes Haar a data-agnostic, closed-form orthogonal transform that approximates PCA for circulation-structured groups. Local denoising is therefore realized by grouping similar patches, modeling their local circulant structure, and applying a fixed Haar transform along the group dimension instead of learning a local PCA basis for every group (Kong et al., 29 Dec 2025).

The resulting forward transform combines a global t-SVD projection with local Haar: $\sigma^2$6 followed by hard thresholding

$\sigma^2$7

and inverse transform

$\sigma^2$8

The method is therefore a one-step local collaborative filter: transform, threshold, inverse, aggregate (Kong et al., 29 Dec 2025).

The adaptive variant uses local eigenvalue information from the circulant Gram structure. With the alternating eigenvector

$\sigma^2$9

the rank position $f$0 of its eigenvalue $f$1 is used to adjust the CNN-estimated noise level: $f$2 The reported ablation is that removing the eigenvalue-based adjustment reduces PSNR by $f$3–$f$4 dB on several datasets. The paper further reports that the method is parallelizable and achieves $f$5 speedup over a naive MATLAB serial implementation (Kong et al., 29 Dec 2025).

4. Related local and multiscale formulations

Local denoising circulation also appears in nonlinear diffusion and in analyses of locality itself. In a multiscale Perona–Malik framework, the noised image is used as an initial condition for

$f$6

and the multiscale algorithm has two main ingredients: “performing local image denoising in each local domain of basis support” and “constructing multiscale basis functions to construct a coarse resolution representation by a Galerkin coupling.” In this interpretation, denoising first occurs locally on coarse neighborhoods $f$7, then propagates globally through overlapping basis supports and projected coarse dynamics (Vasilyeva et al., 2024).

A separate line of work sharpens the meaning of locality by showing that Non-Local Means is local in practice. The estimator

$f$8

was historically treated as genuinely non-local, but the empirical study reports that, on average on natural images, the bias of the NLM estimator is an increasing function of the radius of the similarity searching zone, because noise disrupts the order of similarity between patches. The mean squared error has “an absolute minimum for a disk of radius 3 to 4 pixels,” so the practically optimal regime is local rather than global (Postec et al., 2013).

A one-dimensional spectral analogue uses a partial circulant matrix

$f$9

followed by SVD and component selection by the total variations of left singular vectors. Signal components are then reconstructed from a low-rank approximation using only the signal-related singular directions. Here too, denoising is local in the sense of overlapping neighborhoods, yet circulant at the representation level (Chen et al., 2020).

These formulations suggest a common pattern: locality may be spatial, spectral, or multiscale, while circulation may describe either propagative smoothing or an explicit circulant algebraic structure.

5. Local denoising circulation as discrete curl in diffusion LLMs

In diffusion LLMs, local denoising circulation is formal rather than metaphorical. At reverse time $(i,j)$0, with unresolved coordinates $(i,j)$1, the denoiser induces local conditionals

$(i,j)$2

For a block $(i,j)$3 of unresolved positions and an order $(i,j)$4, these local conditionals compose into an order-induced pseudo-joint

$(i,j)$5

which depends on order unless the conditionals are compatible with a single global joint (Kim, 10 May 2026).

Fix a context $(i,j)$6, two unresolved positions $(i,j)$7, and candidate tokens $(i,j)$8. The two two-step pseudo-joints are

$(i,j)$9

and

$$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$0

The local denoising circulation is then

$$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$1

equivalently,

$$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$2

This is the paper’s discrete analogue of curl: zero circulation means the local denoising field is curl-free on that elementary square (Kim, 10 May 2026).

The global consequence is exact. For any two permutations $$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$3 of a block $$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$4, the global order gap

$$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$5

decomposes into a sum of local circulations along any adjacent-swap path turning $$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$6 into $$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$7. Thus local denoising circulation is the fundamental local obstruction to order invariance. The paper also separates this incompatibility-driven path dependence from two other effects: conditional-dependence error in parallel updates, quantified by conditional total correlation $$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$8, and order-specific estimation error, quantified by order-specific local KL terms (Kim, 10 May 2026).

The framework is explicitly inference-only. It proposes aggregate diagnostics such as absolute expected circulation and a normalized variant $$N(x_i,y_j) = \{(x_{i+s},y_{j+t}) : s,t=-l,\dots,l\}$$9, and notes that

$$\hat z_{ij}(x,y) = \hat\beta_0^{(i,j)} + \hat\beta_1^{(i,j)}(x-x_i) + \hat\beta_2^{(i,j)}(y-y_j),$$0

A model is “genuinely order-free” on a block only to the extent that measured circulation is near zero and order-induced pseudo-joint differences are negligible (Kim, 10 May 2026).

6. Distinctions, misconceptions, and future directions

Several misconceptions are corrected by the literature itself. First, local denoising circulation is not identical to classical multiscale analysis: one image-denoising paper states that its method is “not strictly multi-scale in the wavelet sense, but it has a multi-resolution flavor” (Basak et al., 2024). Second, non-locality is not guaranteed by a large search zone: for NLM, the empirical optimum is local because noisy patch distances are unreliable at long range (Postec et al., 2013). Third, in diffusion LLMs, zero circulation is not the whole story: even under perfect compatibility, one-shot parallel decoding can still incur a penalty equal to the conditional total correlation among block tokens (Kim, 10 May 2026).

The practical implications are domain specific. In adaptive image denoising, parameter sensitivity and computational load remain central trade-offs; the same paper points to 3D images, color images, improved edge detection, and more complex noise models as potential extensions, while also noting that optimization and parallelization are important future work (Basak et al., 2024). In Haar-tSVD, the local circulant formulation already yields a one-step, highly parallelizable denoiser, but the authors still note limitations when patch grouping is unreliable under severe noise and suggest further multi-scale or downsample-then-super-resolve strategies (Kong et al., 29 Dec 2025). In diffusion LLMs, the formal circulation viewpoint leads to concrete design directions: circulation-aware regularization

$$\hat z_{ij}(x,y) = \hat\beta_0^{(i,j)} + \hat\beta_1^{(i,j)}(x-x_i) + \hat\beta_2^{(i,j)}(y-y_j),$$1

commutator-aware scheduling, parallel trajectory supervision, and potential-based DLMs (Kim, 10 May 2026).

Taken together, these works indicate that local denoising circulation is best understood not as a single algorithm but as a technical motif. In imaging, it is a description of how denoising influence is confined, redirected, or aggregated by local geometry, local clusters, and local basis supports. In circulant patch methods, it is encoded directly in the algebra of patch groups. In diffusion LLMs, it becomes a precise compatibility diagnostic. The common thread is that denoising is locally organized, and the question is whether the induced local field is structure preserving, computationally tractable, and, in the formal probabilistic setting, conservative.