Sylvester-Gallai type theorems for approximate collinearity

We study questions in incidence geometry where the precise position of points is `blurry’ (e.g. due to noise, inaccuracy or error). Thus lines are replaced by narrow tubes, and more generally affine subspaces are replaced by their small neighborhood. We show that the presence of a sufficiently large number of approximately collinear triples in a set of points in d dimensional complex space implies that the points are close to a low dimensional affine subspace. This can be viewed as a stable variant of the Sylvester-Gallai theorem and its extensions. Building on the recently found connection between Sylvester-Gallai type theorems and complex Locally Correctable Codes (LCCs), we define the new notion of stable LCCs, in which the (local) correction procedure can also handle small perturbations in the euclidean metric. We prove that such stable codes with constant query complexity do not exist. No impossibility results were known in any such local setting for more than 2 queries.

💡 Research Summary

The paper addresses a fundamental robustness gap in incidence geometry and its applications to coding theory. Classical Sylvester‑Gallai (SG) theorems assert that if a finite set of points in a vector space has the property that every line determined by two points contains a third point, then the whole set lies on a single line (or, more generally, on a low‑dimensional affine subspace). This statement is exact: it assumes perfect knowledge of point locations and exact collinearity. In practice, however, data are noisy, measurements are imprecise, and points rarely sit exactly on a line. The authors therefore replace lines by narrow tubes (the ε‑neighbourhood of a line) and affine subspaces by their ε‑neighbourhoods, and they ask what geometric structure can be inferred when a large fraction of triples are “approximately collinear”.

Main geometric result (stable SG theorem).
Let X be a set of n points in ℂ^d. Suppose that at least a δ‑fraction of all possible triples (i,j,k) satisfy the following: the line through x_i and x_j, thickened to a tube of radius ε, contains x_k. Under this hypothesis the authors prove that X is close to a low‑dimensional affine subspace A ⊂ ℂ^d. More precisely, there exists a subspace of dimension O(1/δ) such that every point of X lies within distance O(ε·poly(1/δ)) of A. The proof proceeds by constructing an “approximate dependency graph” whose edges encode the existence of an ε‑tube containing a third point. Using combinatorial properties of this graph (high average degree, expansion) they build a matrix whose rows correspond to approximate linear relations among the points. By a careful spectral analysis they lower‑bound the smallest singular value of this matrix, which forces the points to be nearly contained in a low‑rank subspace. The argument is a robust analogue of the classic SG proof: it replaces exact linear dependencies with ε‑approximate dependencies and shows that the error does not accumulate catastrophically.

Stable locally correctable codes (stable LCCs).
Locally correctable codes (LCCs) are error‑correcting codes that allow the recovery of any symbol by querying only a few other symbols. Traditional LCC definitions assume that the recovery is exact: a queried symbol is expressed as an exact linear combination of the queried symbols. The authors introduce stable LCCs, where the recovery rule is allowed to be approximate in the Euclidean norm. Formally, a code C ⊂ ℂ^n is a (q,ε,δ)‑stable LCC if for each coordinate i there are at least δ·n^{q‑1} q‑tuples of other coordinates such that i can be expressed as an ε‑approximate linear combination of the q symbols in the tuple.

The paper shows that the stable SG theorem directly yields impossibility results for such codes. If a stable LCC had constant query complexity q (independent of n) and a fixed approximation parameter ε, then the set of codewords would generate a point set satisfying the hypothesis of the stable SG theorem with δ bounded away from zero. Consequently the entire code would be forced into a low‑dimensional subspace, contradicting the requirement that a good code have large dimension (i.e., high rate). The authors prove that no constant‑query stable LCCs exist over the complex numbers, extending the known non‑existence of 2‑query LCCs to all constant‑query regimes. This is the first impossibility result for locally correctable codes that tolerates Euclidean perturbations beyond the exact algebraic setting.

Technical contributions and methods.

1. ε‑approximate collinearity model: The authors formalize the notion of an ε‑tube around a line and define the fraction of approximately collinear triples. This model captures both additive noise and small multiplicative perturbations.
2. Graph‑matrix framework: They construct a dependency graph G whose vertices are points and whose hyperedges correspond to ε‑collinear triples. From G they derive a matrix M whose rows encode the approximate linear relations. The key analytic step is to prove that M has a large spectral gap despite the presence of noise.
3. Spectral lower bound: By adapting tools from random matrix theory and perturbation analysis, they show that the smallest singular value of M is bounded below by a function of δ and ε. This forces the rank of the point configuration to be O(1/δ).
4. Application to coding theory: The same matrix M can be interpreted as the generator matrix of a code. The spectral bound translates into a dimension bound for any code that admits many stable local correction rules.
5. Generalization to higher‑dimensional affine subspaces: While the paper focuses on lines (1‑dimensional subspaces), the techniques extend to approximate dependencies on higher‑dimensional flats, yielding analogous low‑dimensional containment results.

Implications and future directions.
The stable SG theorem provides a rigorous justification for dimensionality‑reduction techniques in noisy settings: if a data set exhibits many near‑collinear relationships, it must be close to a low‑dimensional manifold. This could be leveraged in robust PCA, subspace clustering, and manifold learning, where one often assumes that data lie near a union of low‑dimensional subspaces. In coding theory, the non‑existence of stable constant‑query LCCs suggests that any locally correctable scheme tolerant to Euclidean noise must either increase the number of queries or sacrifice the locality property altogether. An open question is whether similar impossibility results hold over finite fields or for non‑linear codes. Another promising direction is to quantify the trade‑off between ε, δ, and the achievable dimension, potentially leading to explicit bounds for practical algorithms that aim to recover low‑dimensional structure from noisy incidence data.

In summary, the paper delivers a robust geometric theorem that bridges approximate incidence geometry with the theory of locally correctable codes, establishing both positive structural guarantees for noisy point configurations and strong negative results for stable LCCs with constant query complexity.