Eigen-Domain Clustering

• Eigen-Domain Clustering is a family of methods that embed data in eigenvector or eigenvalue spaces to transform discrete partitioning into a continuous spectral problem.
• It encompasses classical spectral clustering on undirected graphs, deep clustering through eigenbasis refinement, and non-Hermitian clustering for directed networks.
• Practical implementations focus on proper eigen-gap estimation, choice of affinity matrices, and efficient clustering rules to balance computational cost and clustering accuracy.

Eigen-Domain Clustering (EDC) denotes clustering methods that operate after an eigen-structured transformation of the data. In the classical spectral sense, EDC is “clustering in an eigenvector space”: objects are embedded using leading eigenvectors or singular vectors of a matrix that encodes affinities, associations, or covariance structure, and a conventional clustering rule is then applied in that derived domain (Mirzal et al., 2010). Across later work, the same phrase is also used for variants built on within-class scatter eigen-bases, covariance eigenvalues, or covariance eigenspaces (Guo et al., 2021, Deng et al., 15 Jul 2025, Ali et al., 18 Jun 2026). This suggests that EDC is best understood as a family of eigen-structured clustering constructions rather than a single canonical algorithm.

1. Spectral foundations and canonical formulation

The foundational spectral formulation casts clustering as a relaxed trace maximization problem over a symmetric matrix. For an undirected graph with affinity matrix $W$ and weight matrix $\Phi$, the general weighted-association objective can be written as

$$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$

with $\tilde{Z}^T \tilde{Z} = I_K$. After relaxing the binary indicator constraints, Ky Fan’s theorem implies that the continuous optimum is attained by the top $K$ eigenvectors of

$H = \Phi^{-1/2} W \Phi^{-1/2}.$

Vertices are then represented by the rows of the eigenvector matrix and clustered in that $K$-dimensional eigen-domain (Mirzal et al., 2010).

The same paper extends this viewpoint to bipartite and directed data. For bipartite graphs or data matrices $A$, the relaxed $K$-way co-clustering solution is obtained from the first $K$ left and right eigenvectors, equivalently the singular vectors, of the normalized matrix $\Phi$0. The row and column problems reduce to spectral clustering on $\Phi$1 and $\Phi$2, yielding the stated equivalence that column clustering implies row clustering and vice versa. For directed graphs, one formulation constructs the symmetric surrogate

$\Phi$3

and again performs clustering in the corresponding eigenvector space.

Setting	Eigen-domain object	Clustered entities
Unipartite graph	Top $\Phi$4 eigenvectors of $\Phi$5	Vertices
Bipartite graph / matrix	First $\Phi$6 left and right singular vectors of $\Phi$7	Rows and columns
Directed graph	Top eigenvectors of $\Phi$8	Vertices

This classical formulation gives EDC its most general mathematical meaning: a discrete partitioning problem is relaxed into an eigenvalue or singular-value problem, then returned to a hard partition by clustering in the resulting embedding. The eigen-domain is therefore not merely a visualization space; it is the relaxed solution space of the original combinatorial objective.

2. Eigen-domain construction in deep clustering

A distinct but closely related use of EDC appears in deep clustering through "Deep Embedded K-Means" (DEKM), where the eigen-domain is constructed from the within-class scatter matrix of K-means rather than from a graph affinity matrix. DEKM begins with an autoencoder embedding $\Phi$9, pre-trained by reconstruction loss, and then discards the decoder. Clustering thereafter is driven entirely by K-means and an eigen-domain refinement step. The within-class scatter matrix is

$$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$0

and its eigendecomposition

$$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$1

defines an orthonormal transformation $$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$2 whose eigenvectors are sorted in ascending order of eigenvalues. The transformed representation is

$$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$3

Small eigenvalues are interpreted as directions with high cluster-structure information, while large eigenvalues correspond to directions with high within-cluster spread (Guo et al., 2021).

DEKM does not truncate to a lower-dimensional subspace. Instead, it works in the full eigen-domain and greedily optimizes the worst direction: for each point, it modifies only the last eigen-coordinate so that it moves toward the centroid along the eigen-direction associated with the largest eigenvalue. The resulting loss is

$$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$4

where $$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$5 differs from $$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$6 only in that last coordinate. This procedure alternates with hard K-means assignments and replaces the common reconstruction-plus-clustering compromise by a purely clustering-driven refinement.

The method is explicitly positioned against DEC, IDEC, DCEC, DCN, and DKM. Its novelty lies in the reparameterization of the embedding into eigen-directions ordered by within-cluster variance, followed by greedy reduction of variance along the least favorable axis. Empirically, the paper reports strong gains on several datasets: on MNIST, AE+K-means reaches ACC $$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$7 and NMI $$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$8, whereas DEKM reaches ACC $$\max J_u = K \, \mathrm{tr}\big(\tilde{Z}^T \Phi^{-1/2} W \Phi^{-1/2} \tilde{Z}\big),$$9 and NMI $\tilde{Z}^T \tilde{Z} = I_K$0; on RCV1-10K, DEKM reports ACC $\tilde{Z}^T \tilde{Z} = I_K$1 and NMI $\tilde{Z}^T \tilde{Z} = I_K$2, the best among the listed baselines. In this formulation, EDC is not graph-based spectral clustering but an eigen-basis refinement of a learned embedding.

3. Directed-network EDC and non-Hermitian embeddings

Another major line of work argues that, for directed graphs, the eigen-domain should be built directly from the non-Hermitian adjacency matrix rather than from symmetrizations, Laplacians, or SVD-based surrogates. In this setting, the adjacency matrix $\tilde{Z}^T \tilde{Z} = I_K$3 is analyzed under a weighted, inhomogeneous directed Erdős–Rényi model with low-rank expectation $\tilde{Z}^T \tilde{Z} = I_K$4 and second-moment matrix $\tilde{Z}^T \tilde{Z} = I_K$5. The key spectral threshold is

$\tilde{Z}^T \tilde{Z} = I_K$6

Eigenvalues of $\tilde{Z}^T \tilde{Z} = I_K$7 with modulus larger than $\tilde{Z}^T \tilde{Z} = I_K$8 generate isolated outlier eigenvalues of $\tilde{Z}^T \tilde{Z} = I_K$9; these carry the cluster signal (Coste et al., 2021).

The method uses both left and right eigenvectors of $K$0. If $K$1 and $K$2 are the informative left and right eigenvectors associated with the $K$3 largest outlier eigenvalues, the node embedding is

$K$4

A Gaussian Mixture Model is then fitted in this embedding space. The paper states that the use of both left and right eigenvectors is mandatory in sparse regimes and gives numerical evidence for the superiority of Gaussian Mixture clustering over the widely used k-means algorithm.

This formulation differs sharply from symmetric spectral relaxations. It does not regularize, trim, or normalize the data, and it treats directionality as intrinsic rather than as a nuisance to be symmetrized away. The associated Master Theorem gives asymptotics for isolated eigenvalues and explicit overlaps between sample and population eigenvectors. In sparse directed stochastic block models, this yields a low-dimensional eigen-domain in which communities appear as mixture components. The same work also emphasizes that in the constant-degree sparse regime strong or weak consistency is impossible because a constant fraction of nodes are isolated; the meaningful objective is detection or partial consistency rather than exact recovery.

4. Eigenvalue-domain and eigenspace-domain variants

Later work broadens the meaning of eigen-domain beyond eigenvector embeddings of affinity graphs. In a heterogeneous hybrid analog–digital massive MIMO receiver, improved EDC is used for sensing the number of sources by clustering features derived from covariance eigenvalues rather than eigenvectors. For each group $K$5, a sample covariance matrix is eigendecomposed to obtain eigenvalues $K$6, which are standardized by

$K$7

then mapped to a two-dimensional feature space,

$K$8

Aggregating these points over all groups gives $K$9, on which DBSCAN identifies a dense noise cluster near the origin and a complementary signal set $H = \Phi^{-1/2} W \Phi^{-1/2}.$0. The number of targets is estimated by

$H = \Phi^{-1/2} W \Phi^{-1/2}.$1

Here the eigen-domain is the spectrum of the covariance matrix, geometrically re-expressed as clustered points in a nonlinear 2D space. The paper reports that the three proposed number-sensing methods achieve $H = \Phi^{-1/2} W \Phi^{-1/2}.$2 sensing at moderate-to-high SNRs, while the improved 1D-CNN is superior under extremely-low SNR conditions (Deng et al., 15 Jul 2025).

A second extension appears in personalized system identification, where systems are clustered through the alignment of covariance eigenspaces. Each system $H = \Phi^{-1/2} W \Phi^{-1/2}.$3 estimates

$H = \Phi^{-1/2} W \Phi^{-1/2}.$4

extracts its top-$H = \Phi^{-1/2} W \Phi^{-1/2}.$5 eigenvectors $H = \Phi^{-1/2} W \Phi^{-1/2}.$6, and compares them across systems through directional similarity scores. For each eigen-direction,

$H = \Phi^{-1/2} W \Phi^{-1/2}.$7

and the aggregate similarity is

$H = \Phi^{-1/2} W \Phi^{-1/2}.$8

Spectral clustering on $H = \Phi^{-1/2} W \Phi^{-1/2}.$9 then recovers groups of systems with shared dynamics. The paper supplies covariance concentration bounds, Davis–Kahan eigenspace perturbation bounds, pairwise false-merge probabilities, and a global clustering success guarantee, and reports lower personalized model-estimation error than training-based clustering and non-clustered baselines (Ali et al., 18 Jun 2026).

Taken together, these variants show that EDC can refer not only to clustering in a space spanned by eigenvectors, but also to clustering in an eigenvalue-derived feature space or in a space of leading covariance eigenspaces. The shared principle is that spectral structure is treated as the primary representation on which clustering is performed.

5. Assumptions, limitations, and recurring design choices

EDC methods are united by a reliance on spectral separation, but the form of that separation differs across domains. In classical spectral clustering, the main caveats concern the choice of similarity matrix, the relaxation from discrete indicators to real-valued orthonormal embeddings, graph connectivity, degree normalization, and the need to specify $K$0. The same framework also inherits computational costs from eigenvector or SVD computation, especially on large graphs or matrices (Mirzal et al., 2010).

In DEKM, the critical assumptions are those of K-means and within-cluster compactness. The method is tailored to K-means, assumes clusters are roughly isotropic Gaussians, requires the number of clusters $K$1, and introduces an eigendecomposition of $K$2 at every iteration. Its greedy update acts only on the last eigen-direction, which the paper notes may converge more slowly than optimizing multiple directions at once, even though it performs best empirically among the tested strategies (Guo et al., 2021).

For direct adjacency-eigen EDC on directed networks, the theory assumes a low-rank expectation matrix $K$3, independent entries, diagonalizability, and delocalized eigenvectors. Informative structure is detectable only when the corresponding eigenvalues lie outside the non-Hermitian bulk defined by the threshold $K$4. The method also requires selecting the number $K$5 of informative eigenpairs, typically by counting eigenvalues outside the bulk circle (Coste et al., 2021).

In the H$K$6AD MIMO setting, improved EDC is unsupervised and interpretable but becomes ambiguous when signal eigenvalues approach noise eigenvalues at very low SNR; the paper explicitly states that it is inferior to FCNN at $K$7 dB, and that the improved 1D-CNN is superior under extremely-low SNR conditions (Deng et al., 15 Jul 2025). In personalized system identification, success depends on full state observation, Gaussian assumptions, shared dynamics within each cluster, known $K$8, and nonzero eigengaps. The finite-sample analysis also shows that unstable dynamics increase $K$9 and therefore the required number of rollouts (Ali et al., 18 Jun 2026).

A plausible implication is that EDC is most effective when the spectral object being used—graph operator, within-class scatter matrix, covariance spectrum, or covariance eigenspace—has a clear gap between informative and non-informative directions. Where that gap collapses, clustering degrades regardless of whether the eigen-domain is constructed from eigenvectors, eigenvalues, or subspaces.

6. Terminological ambiguity and acronym collisions

The acronym “EDC” is not stable across the literature, and this has generated a recurrent source of confusion. In "Efficient Dynamic Clustering-Based Document Compression for Retrieval-Augmented-Generation," EDC refers to Efficient Dynamic Clustering-based document Compression, not Eigen-Domain Clustering. The paper explicitly states that it never mentions eigenvalues, eigenvectors, spectral clustering, eigen-decomposition, or an eigen-domain representation; its clustering is a query-specific, root-based partitioning in embedding space, followed by LLM-based compression (Li et al., 4 Apr 2025).

A second collision occurs in "EDC: Equation Discovery for Classification," where EDC stands for Equation Discovery for Classification. That method is a grammar-based symbolic classifier for binary classification and contains no eigenvalue, eigenvector, or clustering component. Its “EDC” is therefore unrelated to spectral or eigen-domain clustering (Toussaint et al., 28 Oct 2025).

This ambiguity matters because the phrase “Eigen-Domain Clustering” accurately describes some lines of work—spectral clustering in eigenvector space, deep eigen-domain refinement, covariance-eigenvalue clustering, and covariance-eigenspace clustering—but not every method abbreviated as EDC. The term should therefore be read in context: in some papers it denotes a mathematically spectral clustering paradigm, while in others it is only an acronymic coincidence.