k-Lines Clustering of Points

• k-Lines clustering is a problem that partitions points in Euclidean space into k groups, each fitted by a line that minimizes the squared distances from points.
• It generalizes k-means by replacing point centers with one-dimensional subspaces, enabling the capture of linear structural trends in data.
• Heuristic, spectral, and coreset-based methods address its computational challenges, supporting applications in motion segmentation, imaging, and trajectory analysis.

k-Lines clustering is the problem of partitioning a set of points in Euclidean space into $k$ groups, each associated with an optimal line such that the sum of squared distances from points to their closest line is minimized. This optimization generalizes classical $k$-means to the case where centers are replaced by one-dimensional linear subspaces. The $k$-lines problem arises naturally in subspace clustering, computational geometry, and large-scale data analysis, and has motivated the development of novel algorithmic and complexity-theoretic frameworks.

1. Formal Problem Definition and Context

Given a set $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$ and an integer $k \geq 1$, the $k$-lines clustering task is to find $k$ lines $L_1, ..., L_k$ in $\mathbb{R}^d$ minimizing

$$\sum_{i = 1}^n \min_{1 \leq j \leq k} \text{dist}^2(p_i, L_j)$$

where $k$0 is the Euclidean distance from point $k$1 to line $k$2 (Bentert et al., 19 Dec 2025, Statman et al., 2020, Ding et al., 2012). In decision form, one asks whether there exist $k$3 such that the sum falls below a given budget $k$4. In applications, the problem supports both unsupervised segmentation of high-dimensional data and geometric inference of underlying linear structure.

2. Computational Complexity and Exact Algorithms

The $k$5-lines clustering problem is W[1]-hard parameterized by $k$6 and, under the Exponential Time Hypothesis (ETH), admits no algorithm running in $k$7 time. This lower bound is established by a reduction from the Regular Multicolored Independent Set problem, introducing a polynomial gap between YES and NO instances, thus demonstrating strong computational hardness even in the plane ($k$8) (Bentert et al., 19 Dec 2025).

Despite this hardness, it is possible to obtain an exact algorithm with running time $k$9. This is achieved via the algebraic-geometry / sign-condition technique, which samples cells in the parameter space of all possible $k$0-tuples of lines. Each cell is defined by the set of sign conditions on polynomials encoding point-line proximity, and the collection of such cells covers all possible “nearest-line” clusterings of $k$1. For each cell, one computes the optimal line-fitting within each cluster (using PCA-type routines for $k$2 subproblems), ultimately finding the globally optimal solution (Bentert et al., 19 Dec 2025).

The table summarizes the key complexity results:

Parameterization	Complexity	Source
$k$3 (W[1]-hardness)	$k$4 XP	(Bentert et al., 19 Dec 2025)
ETH lower bound	No $k$5	(Bentert et al., 19 Dec 2025)

This approach is primarily of theoretical interest, as the algebraic-cell enumeration is infeasible for large $k$6 or $k$7.

3. Approximation Schemes and Sampling Algorithms

A polynomial-time approximation scheme (PTAS) for $k$8-lines clustering (with an $k$9 metric) is provided via uniform random sampling, $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$0-rotation, symmetric sampling, and recursive projection (Ding et al., 2012). The central procedure estimates candidate line directions using reflective symmetries in small random subsamples of the centered data. Each candidate is formed by computing means of all subsets of a small sample and defining a line through these means.

For fixed $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$1, by removing at most a $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$2 fraction of outliers, $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$3 lines can be computed to within $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$4 times the optimal sum of squared distances with high probability. When $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$5 is constant, the approach achieves linear time in $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$6 for fixed parameters.

Key ingredients include:

• Slab Partition: Removal of outliers via thresholding on projections
• $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$7-Rotation: Guarantees that sufficiently distinct sampled directions yield good approximations to the underlying structure
• Symmetric Sampling: Reflection of small samples to ensure coverage of the line arrangement space.

Coreset-based approaches and $P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$8-approximation are also realized through a k-means++-like seeding (k-Line-Means++), where D$P = \{p_1, ..., p_n\} \subseteq \mathbb{R}^d$9-sampling is adapted to directional data, with theoretical guarantees based on a proxy metric $k \geq 1$0; the resulting cost is within $k \geq 1$1 of optimal, in time $k \geq 1$2 (Statman et al., 2020).

4. Ensemble and Spectral Methods for $k \geq 1$3-Lines Clustering

Ensemble geometric algorithms extend $k \geq 1$4-subspaces methods by running randomized KSS (K-subspaces) algorithms multiple times with distinct initializations, then aggregating cluster assignments. The EKSS method collects a co-association matrix counting the frequency with which each pair of points is co-clustered across multiple runs; this matrix quantifies geometric proximity in “evidence accumulation” terms (Lipor et al., 2017).

The clustering pipeline involves:

1. Multiple KSS runs initialized by sampling $k \geq 1$5 unit-norm directions independently.
2. Alternating Lloyd-style assignment and line re-estimation by leading singular vector computation.
3. Formation of a co-association matrix $k \geq 1$6.
4. Sparsification of $k \geq 1$7 via thresholding, then spectral clustering (using the Laplacian of $k \geq 1$8) followed by $k \geq 1$9-means in the embedded space.

A prototypical recovery result states that, given sufficient separation between underlying lines (expressed as a lower bound on pairwise principal angles), sufficiently large sample sizes, and sufficiently many random restarts ($k$0, with $k$1 an angular separation parameter), EKSS recovers true clusters with high probability. This theoretical guarantee extends to noisy settings and provides strong empirical performance in the presence of intersecting or close subspaces (Lipor et al., 2017).

Computational and statistical features:

• Each KSS iteration is $k$2.
• The ensemble is highly parallelizable.
• Thresholding parameter $k$3 balances cluster connectivity and false association.

5. Coreset Construction and Scalable Approaches

For very large data, coresets (small, weighted subsets achieving provably small error for all solutions) enable scalable $k$4-lines clustering. Given a sufficiently large number $k$5 (related to the optimal k-line solution), one constructs an $k$6-coreset of size $k$7 in $k$8 time (Statman et al., 2020). The k-Line-Means++ seeding method sharply reduces candidate coreset size while preserving clustering accuracy.

The relationship between the proxy metric $k$9 and the squared distance metric is characterized, allowing the approximation algorithm to transfer probabilistic guarantees from the clustering++ literature to the $k$0-lines setting. Handling of outliers is realized via iterative exclusion of farthest points in the sampling phase; the guarantees extend to approximate optimal post-outlier removal.

Empirical studies on high-dimensional and large-scale datasets (UCI, MNIST, Wikipedia) demonstrate that this two-stage method (seeding + spectral coreset) achieves substantial runtime and memory improvements over prior spectral approaches with minimal loss in accuracy (Statman et al., 2020).

6. Parameterized Complexity and Open Problems

The $k$1-lines clustering problem, in contrast to line covering, is W[1]-hard, precluding fixed-parameter tractable (FPT) algorithms for $k$2 unless W[1]=FPT (Bentert et al., 19 Dec 2025). The best-known algorithm has slice-wise polynomial (XP) complexity: $k$3. For finite $k$4, the general projective clustering algorithm for $k$5-dimensional flats of dimension $k$6 yields a computational complexity of $k$7, which is tight under ETH.

No nontrivial FPT-approximation is known: the existence of a constant-factor FPT-approximation for $k$8-lines remains open. A plausible implication is that beyond $k$9 norms, this entire framework fails, as the $L_1, ..., L_k$0 subproblem becomes NP-hard for $L_1, ..., L_k$1 and inapproximable for $L_1, ..., L_k$2.

The algebraic-geometric approach provides a complete theoretical understanding of the decision problem, but is not practical for large $L_1, ..., L_k$3 or $L_1, ..., L_k$4. Heuristic or approximate methods—spectral, randomized, or coreset-based—are used in applied contexts.

7. Applications, Implementational Insights, and Practical Considerations

$L_1, ..., L_k$5-lines clustering is fundamental in unsupervised learning for data lying near multiple one-dimensional structures, such as motion segmentation, molecular imaging, and trajectory analysis. When used in high dimensions, dimension reduction (e.g., via Johnson–Lindenstrauss projection) is recommended before clustering (Ding et al., 2012). Careful implementation of sampling routines, early pruning, and batch distance computation are necessary for efficiency. The ensemble and coreset paradigms allow scaling to massive datasets.

Practical guidance distilled from the literature includes:

• For moderate $L_1, ..., L_k$6, PTAS or $L_1, ..., L_k$7-approximate methods achieve near-linear runtime in $L_1, ..., L_k$8 with removal of an $L_1, ..., L_k$9 fraction of outliers.
• Thresholds for ensemble aggregation should be set logarithmically in the sample size to optimize the trade-off between noise robustness and cluster connectivity.
• Direct implementations of XP or algebraic geometry methods are not feasible at scale; heuristic spectral and sampling approaches dominate in real-world tasks.

Research continues on narrowing the approximation gap, derandomizing seeding steps, robustifying against adversarial outliers, and extending to broader models, such as affine lines and higher-dimensional flats. Theoretical advances in parameterized and approximation complexity for $\mathbb{R}^d$0-lines clustering remain a significant area of interest.