Two-Dimensional Smoothing Techniques

• Two-Dimensional Smoothing (TDS) is a family of techniques that impose explicit smoothness criteria on 2D data like images, meshes, and surfaces while preserving structural details.
• Global variational approaches and local directional methods illustrate TDS's flexibility, with applications ranging from image processing to statistical surface estimation and mesh optimization.
• The methodology spans spline-based, operator smoothing, and multigrid frameworks, balancing noise reduction with edge and quality preservation in complex computational tasks.

Searching arXiv for recent and representative papers on “Two-Dimensional Smoothing” and closely related usages. Two-Dimensional Smoothing (TDS) denotes a class of procedures that modify a two-dimensional object under explicit smoothness criteria while retaining application-dependent structure. In the literature, the object being smoothed may be a two-dimensional sequence, an image, a planar triangular mesh, a reduced potential-energy surface, a regression surface, or a dynamical state measured in a stronger function space. The operational meaning of “smoothing” therefore varies: it can mean suppressing fluctuation by penalizing second differences, selecting directionally compatible local averages, relocating mesh vertices only when local quality improves, constructing a continuous low-energy surface through an extra dimension, or proving that a two-dimensional flow map becomes regular from $H$ into $H^2$ after positive pullback time (Chen et al., 21 Jul 2025, Mastriani et al., 2016, Lau et al., 2021, Cui et al., 26 Aug 2025).

1. Scope and principal meanings

Across the cited work, TDS is not a single algorithm but a family of two-dimensional regularization ideas. In one line of work, the data are arranged as $Z=(z_{i,j})\in\mathbb{R}^{m\times n}$ and the trend $G=(g_{i,j})$ is defined as the minimizer of a global loss combining fidelity and row/column second-difference penalties (Chen et al., 21 Jul 2025). In SAR despeckling, directional smoothing computes several directional local averages in a moving $2\text{D}$ window and chooses the direction whose average is closest to the center pixel, thereby smoothing along local structure rather than across edges (Mastriani et al., 2016). In mesh optimization, two-dimensional Smart Laplacian smoothing relocates vertices of planar triangular meshes while preserving connectivity and accepts a move only if the local mesh quality improves (Zhao et al., 2015). In nuclear-fission calculations, two-dimensional smoothing refers to replacing a discontinuous $2\text{D}$ potential-energy surface by a continuous low-energy surface $z(x,y)$ inside a local $3\text{D}$ search region (Lau et al., 2021). In statistics, it includes tensor-product P-splines, thin-plate splines, smoothing thin plate splines with $\mathcal H$-matrix compression, and triangulation splines on polygonal domains (Bach et al., 2022, Cavieres et al., 2024, Pyeon et al., 10 Jun 2026). In PDE and multigrid analysis, smoothing denotes either a regularity lift, such as $(H,H^2)$-smoothing for stochastic $H^2$0 Navier–Stokes, or damping of high-frequency error components by a multigrid relaxation scheme (Cui et al., 26 Aug 2025, Zhu et al., 2023).

Setting	Smoothed object	Governing idea
2D sequences and images	$H^2$1, image intensity field	second-difference penalization or directional local averaging
Planar triangular meshes	vertex coordinates	move accepted only if local quality improves
Potential-energy surfaces	$H^2$2 in a 3D search box	lowest-energy continuous surface under continuity constraints
Statistical surface estimation	$H^2$3	spline basis plus roughness penalty
PDE and multigrid	solution differences or error modes	regularity gain or high-frequency damping

This range suggests that TDS is best understood as a methodological category unified by two-dimensional structure and explicit smoothness control, rather than by a unique algebraic template.

2. Global and local formulations on two-dimensional sequences and images

The most explicit algorithm named TDS in the cited corpus is the global second-difference formulation for a general two-dimensional sequence $H^2$4, $H^2$5. It posits a decomposition

$H^2$6

where $H^2$7 is the trend sequence and $H^2$8 is the fluctuation sequence, and defines $H^2$9 as the minimizer of

$Z=(z_{i,j})\in\mathbb{R}^{m\times n}$0

with first differences $Z=(z_{i,j})\in\mathbb{R}^{m\times n}$1, $Z=(z_{i,j})\in\mathbb{R}^{m\times n}$2, second differences

$Z=(z_{i,j})\in\mathbb{R}^{m\times n}$3

and global smoothing parameter $Z=(z_{i,j})\in\mathbb{R}^{m\times n}$4 (Chen et al., 21 Jul 2025). The normal equation is

$Z=(z_{i,j})\in\mathbb{R}^{m\times n}$5

where $Z=(z_{i,j})\in\mathbb{R}^{m\times n}$6 and $Z=(z_{i,j})\in\mathbb{R}^{m\times n}$7 are sparse real symmetric positive semidefinite penalty matrices. The same work proves existence and uniqueness for fixed $Z=(z_{i,j})\in\mathbb{R}^{m\times n}$8, gives the vectorized solution

$Z=(z_{i,j})\in\mathbb{R}^{m\times n}$9

and defines the fluctuation sequence as

$G=(g_{i,j})$0

(Chen et al., 21 Jul 2025). It further introduces TDS-I, TDS-II, and TDS-III, which replace the single $G=(g_{i,j})$1 by directional global parameters $G=(g_{i,j})$2 or row-/column-specific diagonal parameter matrices $G=(g_{i,j})$3.

A distinct image-oriented formulation appears in directional smoothing for SAR despeckling. There the observed image obeys the multiplicative model

$G=(g_{i,j})$4

and smoothing is performed in a local $G=(g_{i,j})$5 window by computing directional averages

$G=(g_{i,j})$6

$G=(g_{i,j})$7

then selecting

$G=(g_{i,j})$8

(Mastriani et al., 2016). This is a nonlinear adaptive directional smoother rather than a global optimizer. The same paper applies the method homomorphically in the log domain and evaluates it using NV, MSD, ENL, DR, and Pratt figure of merit (Mastriani et al., 2016).

These two strands illustrate a basic division inside image and array smoothing. One class uses a global regularized objective with a single solution $G=(g_{i,j})$9; the other uses local directional candidates and a nonlinear selection rule. This suggests that, even within classical data smoothing, TDS spans both variational and local-compatibility paradigms.

3. Geometry-aware smoothing of meshes and physical surfaces

In planar mesh optimization, two-dimensional Smart Laplacian smoothing operates on a triangular mesh without changing connectivity. Ordinary Laplacian smoothing moves a vertex to the geometric center of its neighboring vertices, but it may degrade local quality and even create invalid or inverted elements. Smart Laplacian smoothing adds a quality-based acceptance test: evaluate the quality of the incident triangles before relocation, compute the Laplacian candidate position, re-evaluate the incident-triangle qualities at that position, and accept the move only if the local mesh quality improves; the explicit example criterion is that the minimum quality among all incident triangles increases (Zhao et al., 2015). The GPU implementation described there organizes the procedure into initiating, neighbor finding, constraint determination, and iterative smoothing, compares AoS and SoA layouts, compares Form A

$2\text{D}$0

with Form B

$2\text{D}$1

and reports that AoS is always faster than SoA, Form B is faster than Form A, and the implementation reaches speedups up to $2\text{D}$2 on a GeForce GT640 (Zhao et al., 2015).

A different geometric meaning of TDS appears in nuclear-fission potential-energy surfaces. There, the problem is a discontinuous reduced $2\text{D}$3 PES in coordinates $2\text{D}$4, caused by an omitted coordinate $2\text{D}$5. The proposed Frontier DPM seeks a continuous surface $2\text{D}$6 inside a local $2\text{D}$7 search space, minimizing total energy while enforcing a maximum gradient constraint $2\text{D}$8 and an overlap threshold $2\text{D}$9 between adjacent HFB states (Lau et al., 2021). The method represents partial surfaces by their frontiers, expands them one planar site at a time, and prunes states with identical frontier configurations by keeping only the lowest-energy representative. The brute-force search space scales as $2\text{D}$0, whereas Frontier DPM reduces dependence to the frontier width, with an estimated cost on the order of $2\text{D}$1 or the analogous expression with $2\text{D}$2 and $2\text{D}$3 interchanged (Lau et al., 2021). In the $2\text{D}$4 example, it removes a diagonal discontinuity in $2\text{D}$5, eliminates zero-overlap adjacencies, and reduces density distances (Lau et al., 2021).

These uses show that geometric TDS is not primarily about denoising. In mesh processing it is a quality-monotonic relocation rule on a fixed connectivity graph; in nuclear PES repair it is a constrained search for a physically continuous low-energy surface embedded in a higher-dimensional space.

4. Statistical surface smoothing: tensor products, thin plates, and triangulations

A central statistical formulation of TDS is the tensor-product spline surface

$2\text{D}$6

with anisotropic penalty

$2\text{D}$7

(Bach et al., 2022). In the Bayesian formulation, the spline coefficients have partially improper Gaussian prior

$2\text{D}$8

and the technical bottleneck is the pseudo-determinant in the smoothing-parameter posterior. The paper resolves this by deriving closed-form expressions for $2\text{D}$9 and its derivatives using Kronecker-structured eigendecompositions, enabling adaptive Metropolis-Hastings updates for $z(x,y)$0 in arbitrary dimension and making the $z(x,y)$1 case computationally straightforward (Bach et al., 2022).

Thin-plate smoothing supplies a second major line. Robust thin-plate splines replace quadratic data fit by a convex M-estimation loss,

$z(x,y)$2

where, in the $z(x,y)$3 case,

$z(x,y)$4

(Kalogridis, 2022). The representer form remains

$z(x,y)$5

which in the classical two-dimensional thin-plate case corresponds to an affine polynomial plus the $z(x,y)$6 kernel. The same paper proves optimal convergence rates under mild assumptions and emphasizes invariance of the thin-plate penalty under rigid transformations (Kalogridis, 2022).

Classical smoothing thin plate splines become computationally demanding because the kernel matrix is dense. An efficient realization compresses the dense TPS matrix by an $z(x,y)$7-matrix and solves the resulting Schur-complement SPD system by conjugate gradients (Cavieres et al., 2024). In that work the smoothing objective is

$z(x,y)$8

with $z(x,y)$9 the two-dimensional bending-energy penalty, and the dense kernel matrix $3\text{D}$0 is approximated blockwise by low-rank admissible far-field blocks using Adaptive Cross Approximation (Cavieres et al., 2024). The reported interpolation RMSE on an $3\text{D}$1 training grid and $3\text{D}$2 prediction grid is $3\text{D}$3 for the exact method, the Schur-complement CG method, and the $3\text{D}$4-matrix method, while the latter exhibits practically linear growth in the reported timing plots and strong sensitivity to the ACA tolerance $3\text{D}$5 (Cavieres et al., 2024).

For irregular polygonal domains, Bayesian triangulation splines replace tensor products and global radial kernels by spline spaces on constrained Delaunay triangulations. The domain $3\text{D}$6 is triangulated, the spline space is

$3\text{D}$7

and the roughness penalty is the thin-plate functional

$3\text{D}$8

(Pyeon et al., 10 Jun 2026). The prior is placed jointly on triangulation complexity and coefficient vectors, which allows adaptation to unknown global smoothness and to local spatial inhomogeneity; the paper proves optimal posterior contraction under a global Sobolev assumption and an oracle-type spatial adaptation guarantee over shape-regular triangulations (Pyeon et al., 10 Jun 2026).

Taken together, these statistical formulations show that TDS can be global and Euclidean, robust to contamination, computationally compressed, or boundary-conforming and locally adaptive. The choice among them is governed less by the word “smoothing” than by domain geometry, robustness requirements, and computational regime.

5. Smoothing as regularity gain and multigrid damping in two dimensions

In PDE and dynamical-systems analysis, smoothing has a different meaning from denoising or surface fitting. For stochastic Navier–Stokes on the two-dimensional torus $3\text{D}$9, the relevant result is an $\mathcal H$0-smoothing theorem: for tempered initial sets in the energy space $\mathcal H$1, there exist random variables $\mathcal H$2 and $\mathcal H$3 such that

$\mathcal H$4

This is established under $\mathcal H$5, $\mathcal H$6, and

$\mathcal H$7

and is then used to promote the random attractor from $\mathcal H$8 to a tempered $\mathcal H$9-random attractor of finite fractal dimension in $(H,H^2)$0 (Cui et al., 26 Aug 2025). Here smoothing means that the cocycle becomes locally Lipschitz from a weaker phase space into a stronger one after positive pullback time.

In multigrid analysis for $(H,H^2)$1 Stokes flow, smoothing refers to damping high-frequency error. The stabilized non-staggered discretization

$(H,H^2)$2

is transformed by a distributive operator so that the smoothing problem reduces to the scalar blocks $(H,H^2)$3 and $(H,H^2)$4 (Zhu et al., 2023). The LFA smoothing factor is

$(H,H^2)$5

the Poisson block yields

$(H,H^2)$6

and the stabilized pressure block yields a $(H,H^2)$7-dependent optimal factor with an effective convergence zone approximately

$(H,H^2)$8

(Zhu et al., 2023). The paper emphasizes that this smoothing property is independent of mesh size but depends on the pressure-stabilization parameter.

These meanings broaden the concept of TDS beyond direct manipulation of images or surfaces. In such settings, smoothing is an operator-theoretic property of evolution or iteration, not an explicit fitted surface $(H,H^2)$9. A plausible implication is that “smoothing” in two dimensions is best treated as a structural effect whose concrete realization depends on the governing operator.

6. Evaluation, limitations, and acronym ambiguity

Comparisons across the cited literature show that TDS methods are usually assessed by criteria native to their application rather than by a universal measure. The global second-difference TDS algorithm reports MSE, PSNR, and SSIM and, in its numerical table with $H^2$00, attains the best listed values across AWGN, complex noise, MWGN, and SPN, with core code time reported as $H^2$01 s (Chen et al., 21 Jul 2025). The same paper, however, explicitly states that TDS is not suitable for filtering out salt-and-pepper noise in general image settings and suggests combining TDS with median filtering; in its image experiments, median filtering ranks first for SPN while TDS-I ranks second (Chen et al., 21 Jul 2025). For SAR despeckling, EDS is presented as superior in edge preservation, but the paper itself notes a tension between the scalar metrics in Table I and the narrative, since some classical filters have higher ENL and lower NV than EDS; the claimed advantage is therefore the compromise between despeckling and edge retention rather than uniformly stronger scalar scores (Mastriani et al., 2016).

In computational geometry and numerical linear algebra, performance claims are likewise application-specific. The Smart Laplacian GPU study reports speedups up to $H^2$02, consistent superiority of AoS over SoA, faster convergence/runtime for Form B than Form A, and a slight advantage for CUDA Dynamic Parallelism (Zhao et al., 2015). The efficient smoothing thin plate spline study reports that the $H^2$03-matrix method preserves interpolation RMSE while reducing computational cost, but also shows extreme degradation when the ACA tolerance is loosened to $H^2$04 (Cavieres et al., 2024). Bayesian triangulation splines are reported to outperform existing approaches in estimation accuracy while maintaining low model complexity in simulations on irregular domains (Pyeon et al., 10 Jun 2026).

A persistent source of confusion is terminological rather than methodological. In wireless communications, “TDS” in “TDS-OFDM” denotes time-domain synchronous, not two-dimensional smoothing (Gao et al., 2015, Liu et al., 2012). Those papers are nevertheless relevant to smoothing in a narrower sense because they use one-dimensional and two-dimensional moving average or Wiener filtering to refine channel estimates over neighboring subcarriers and OFDM symbols, but that use of “2-D smoothing” is a channel-estimation subroutine inside TDS-OFDM rather than a definition of TDS itself (Liu et al., 2012). The acronym collision is therefore substantive: not every arXiv paper labeled “TDS” is about Two-Dimensional Smoothing.

The literature as a whole indicates three common misconceptions. First, TDS is not synonymous with windowed image filtering; global variational, spline-based, and search-based formulations are equally central. Second, stronger smoothing is not automatically better; several papers make explicit trade-offs between edge preservation, local quality monotonicity, physical continuity, or posterior adaptation. Third, two-dimensionality does not imply a rectangular Euclidean domain; triangulated polygonal domains, planar meshes, and reduced physical manifolds all support legitimate TDS formulations (Lau et al., 2021, Pyeon et al., 10 Jun 2026).