SDCut: Clustering & Meshing Contexts

• SDCut is a context-dependent label with two distinct uses: one in deep subspace clustering and another in neural meshing.
• In deep subspace clustering, SDCut constructs full n×n affinity matrices and performs complete spectral decomposition, resulting in O(n³) complexity.
• In neural meshing, SDCut acts as a non-official alias for Self-Supervised Dual Contouring, employing self-supervised losses to extract meshes without ground-truth data.

SDCut is a context-dependent label rather than a single stable method name. In the deep subspace clustering literature represented by "Scalable Deep Subspace Clustering Network" (Mrabah et al., 24 Dec 2025), SDCut is presented as a baseline that still constructs full $n \times n$ affinity or self-expression matrices and performs spectral decomposition in sample space, thereby retaining overall $\mathcal{O}(n^3)$ complexity. In a separate neural meshing context, "Self-Supervised Dual Contouring" explicitly states that "SDCut" is not its official abbreviation, but that whenever SDCut appears elsewhere it refers to the same SDC method (Sundararaman et al., 2024). Any technical account of SDCut therefore requires domain disambiguation before methodological interpretation.

1. Terminological scope

The available arXiv sources support two distinct uses of the label.

Context	Use of “SDCut”	Source
Deep subspace clustering	Comparator that still uses full $n \times n$ affinity or self-expression and full spectral decomposition	(Mrabah et al., 24 Dec 2025)
Neural meshing	Non-official alias used elsewhere for Self-Supervised Dual Contouring (SDC)	(Sundararaman et al., 2024)

This split is not a minor naming detail. In the clustering setting, SDCut denotes a method situated within the classical full-affinity spectral pipeline. In the meshing setting, SDCut does not define a separate method at all; it is instead described as a naming variant for SDC. A common misconception is to assume that the same acronym necessarily identifies the same algorithmic family across papers. The sources here do not support that assumption.

2. SDCut in the deep subspace clustering pipeline

Within the SDSNet comparison, SDCut belongs to the class of subspace clustering methods built around self-expressiveness. The underlying setting is the standard one in which each data point is expressed as a linear combination of other points, leading to an affinity or self-expression matrix $\mathbf{C} \in \mathbb{R}^{n \times n}$. The comparison is framed against the conventional least-squares self-expression objective

$$\min_{\mathbf{C}} \frac{1}{2}\|\mathbf{X} - \mathbf{X}\mathbf{C} \|_F^2 + \lambda \|\mathbf{C}\|_F^2,$$

whose closed-form solution is itself $n \times n$ and "not scalable" (Mrabah et al., 24 Dec 2025).

In that framing, SDCut "still constructs full $n \times n$ affinity/self-expression." Its "spectral decomposition and final clustering are performed on this large matrix," and its "complexity remains $\mathcal{O}(n^3)$ overall" (Mrabah et al., 24 Dec 2025). The affinity construction is further characterized as "dense and global," with "heavy memory use." This places SDCut on the side of methods that preserve the classical spectral-clustering bottleneck even when paired with learned feature extraction.

The same comparative account also states that SDCut "may optimize the self-expression only after network convergence, or as a post-hoc step" (Mrabah et al., 24 Dec 2025). In that respect, SDCut is not merely a computational baseline; it also functions as a representative of a decoupled optimization strategy in which representation learning and affinity learning are not necessarily updated simultaneously.

3. Computational profile and scalability limits

The key technical property attributed to SDCut is its retention of the full-matrix regime. In the general background given for deep subspace clustering, constructing the full affinity matrix requires $\mathcal{O}(n^2)$ memory and time, while spectral decomposition requires $\mathcal{O}(n^3)$ time (Mrabah et al., 24 Dec 2025). Because SDCut remains in this regime, the dominant cost is not removed by replacing shallow features with deep ones.

This characterization matters because the subspace clustering bottleneck is structural rather than cosmetic. If affinity construction is exhaustive and eigendecomposition is carried out in sample space, the principal asymptotic barrier remains. The SDSNet discussion makes this explicit by contrasting SDCut with factorized alternatives that never form full pairwise affinities. A plausible implication is that SDCut is suitable only while dataset size remains compatible with dense $\mathcal{O}(n^3)$0 storage and cubic-time eigendecomposition, whereas large-scale deployment requires reformulating the affinity representation itself rather than only improving the encoder (Mrabah et al., 24 Dec 2025).

The comparison also clarifies that the scaling problem is inseparable from the geometry of the spectral step. SDCut performs "full eigendecomposition in sample space," not in a reduced anchor space (Mrabah et al., 24 Dec 2025). That distinction is the core reason it is treated as a non-scalable reference point.

4. Contrast with landmark-based and jointly optimized formulations

SDCut’s role in the SDSNet paper is sharpened by direct contrast. SDSNet avoids full affinity construction by factorizing the self-expression as

$\mathcal{O}(n^3)$1

where $\mathcal{O}(n^3)$2 and $\mathcal{O}(n^3)$3, and by solving spectral clustering in the reduced anchor space rather than the original sample space (Mrabah et al., 24 Dec 2025). Its joint latent-space objective combines auto-encoder reconstruction with landmark-based self-expression: $\mathcal{O}(n^3)$4 subject to $\mathcal{O}(n^3)$5 (Mrabah et al., 24 Dec 2025).

The comparison is explicit on several points. Whereas SDCut uses a "dense and global" affinity with cubic complexity, SDSNet uses a "factorized, anchor-based" affinity with linear scaling in $\mathcal{O}(n^3)$6 for fixed $\mathcal{O}(n^3)$7. Whereas SDCut performs eigendecomposition in sample space, SDSNet reduces the eigenproblem to the anchor space. Whereas SDCut may optimize self-expression after network convergence or as a post-hoc step, SDSNet "JOINTLY minimizes the reconstruction loss ... and the self-expression term" (Mrabah et al., 24 Dec 2025).

These contrasts do not merely describe SDSNet; they define the operational meaning of SDCut in that source. SDCut is the method class that SDSNet claims to supersede computationally by replacing full-matrix self-expression and sample-space spectral decomposition with factorized affinities and joint optimization.

5. SDCut as a naming variant for Self-Supervised Dual Contouring

A different use of SDCut appears in the literature on neural isosurface extraction. "Self-Supervised Dual Contouring" states unambiguously that "SDC stands for 'Self-Supervised Dual Contouring.'" It further states that the paper "does not use the term 'SDCut' as the official abbreviation for the method," and that "Whenever SDCut appears elsewhere, it refers to the same SDC method" (Sundararaman et al., 2024).

In that usage, SDCut denotes a self-supervised extension of Neural Dual Contouring. The method predicts mesh vertices from an input SDF grid, infers faces from sign changes, and eliminates dependence on ground-truth or precomputed proxy meshes. Training is driven by two self-supervised losses. The distance loss

$\mathcal{O}(n^3)$8

aligns the point-to-mesh distance with the magnitude of the SDF, while the normal consistency loss

$\mathcal{O}(n^3)$9

aligns mesh face normals with SDF-derived normals at edge intersections (Sundararaman et al., 2024). The overall mesh objective is

$n \times n$0

The same source characterizes the method as removing "the need for ground truth meshes during training," avoiding "QEF or its limitations/ill-definition," and enabling end-to-end optimization with upstream Deep Implicit Networks (Sundararaman et al., 2024). Accordingly, when SDCut is used in this meshing context, it should be read not as a distinct algorithm but as a non-official label for SDC.

6. Disambiguation, citation practice, and recurrent confusion

The principal scholarly issue surrounding SDCut is disambiguation. In one source, it is a deep subspace clustering comparator defined by full affinity construction, full spectral decomposition, and cubic complexity (Mrabah et al., 24 Dec 2025). In another, it is expressly not the official name of the method at all, but an alternative label referring to Self-Supervised Dual Contouring (Sundararaman et al., 2024).

A plausible implication is that acronym-only retrieval is unreliable across domains. The broader literature represented here already includes another unrelated use of the closely neighboring acronym SDC, namely "shifted delta coefficients" in user-defined keyword spotting (V et al., 2024). The result is that SDC, SDCut, and related strings can point to distinct research objects in subspace clustering, neural meshing, and speech processing. For technical writing, the stable unit of reference is therefore not the acronym by itself but the combination of domain, method description, and citation context.

For researchers encountering the term in recent arXiv usage, the safest interpretation is contextual. If the surrounding discussion concerns self-expression matrices, spectral clustering, and $n \times n$1 cost, SDCut denotes the full-affinity subspace clustering baseline described in the SDSNet comparison. If the discussion concerns SDFs, dual contouring, and self-supervised mesh losses, SDCut denotes the same method as Self-Supervised Dual Contouring, whose official abbreviation is SDC (Mrabah et al., 24 Dec 2025, Sundararaman et al., 2024).