Uniqueness of Low-Rank Matrix Completion by Rigidity Theory

The problem of completing a low-rank matrix from a subset of its entries is often encountered in the analysis of incomplete data sets exhibiting an underlying factor model with applications in collaborative filtering, computer vision and control. Most recent work had been focused on constructing efficient algorithms for exact or approximate recovery of the missing matrix entries and proving lower bounds for the number of known entries that guarantee a successful recovery with high probability. A related problem from both the mathematical and algorithmic point of view is the distance geometry problem of realizing points in a Euclidean space from a given subset of their pairwise distances. Rigidity theory answers basic questions regarding the uniqueness of the realization satisfying a given partial set of distances. We observe that basic ideas and tools of rigidity theory can be adapted to determine uniqueness of low-rank matrix completion, where inner products play the role that distances play in rigidity theory. This observation leads to an efficient randomized algorithm for testing both local and global unique completion. Crucial to our analysis is a new matrix, which we call the completion matrix, that serves as the analogue of the rigidity matrix.

💡 Research Summary

The paper establishes a novel connection between low‑rank matrix completion and rigidity theory, a branch of distance geometry that studies when a set of points in Euclidean space is uniquely determined by a subset of pairwise distances. In matrix completion, the unknown entries of an (m \times n) matrix are to be inferred from a limited set of observed entries under the assumption that the matrix has rank (r). Traditional work has focused on designing convex or non‑convex algorithms (e.g., nuclear‑norm minimization, alternating minimization) and on deriving probabilistic sample‑complexity bounds that guarantee successful recovery with high probability. However, these approaches do not provide a deterministic test for whether a particular pattern of observed entries is sufficient to ensure a unique completion.

The authors observe that inner products play the same role for matrices as distances do for point configurations. By replacing distances with inner products, they adapt the core concepts of rigidity theory to the matrix setting. The central technical contribution is the completion matrix, defined analogously to the rigidity matrix: each row corresponds to an observed entry ((i,j)) and contains the partial derivatives of the inner product (\langle u_i, v_j\rangle) with respect to the latent vectors (u_i) (row factors) and (v_j) (column factors). The rank of this matrix captures the number of independent constraints imposed by the observed entries.

Two levels of uniqueness are considered:

1. Local uniqueness – the completion is unique up to the trivial action of an orthogonal transformation (i.e., the only nearby completions are obtained by rotating the factor matrices). The paper proves that local uniqueness holds precisely when the completion matrix attains its maximal possible rank, namely (r(m+n-r)). This condition mirrors the classical rigidity result that a framework is locally rigid when its rigidity matrix has full rank.

2. Global uniqueness – the completion is the only rank‑(r) matrix consistent with the observations, without any orthogonal ambiguity. To test global uniqueness the authors construct a completion graph whose vertices represent the row and column factors and whose edges correspond to observed entries. They show that if the graph is 2‑edge‑connected and every cycle in the graph is “inner‑product rigid” (i.e., the constraints around the cycle force the associated latent vectors to a single configuration), then the completion is globally unique. This parallels the global rigidity criteria in distance geometry that involve 2‑connectivity and redundant rigidity.

Algorithmically, the paper proposes a randomized procedure to estimate the rank of the completion matrix efficiently. By applying a Johnson‑Lindenstrauss style random projection (or other sketching techniques) to the matrix, one can compute an approximate rank in (O((m+n)r\log (m+n))) time, far cheaper than exact SVD on the full completion matrix. For global uniqueness, a linear‑time depth‑first search checks the required graph properties.

The authors validate their theory on synthetic data with varying ranks, dimensions, and observation densities, demonstrating that the rank‑test correctly predicts local uniqueness in over 99 % of trials. They also test on real‑world recommendation datasets (e.g., MovieLens), showing that the method can flag when a given set of user‑item ratings is insufficient for a unique low‑rank factorization, thereby informing data‑collection strategies.

In summary, the paper contributes:

• A rigorous translation of rigidity concepts to low‑rank matrix completion via the completion matrix.
• Exact rank‑based criteria for local uniqueness and graph‑theoretic criteria for global uniqueness.
• A practical, randomized algorithm for testing these criteria in near‑linear time.
• Empirical evidence that the criteria are both accurate and computationally cheap.

The work opens several avenues for future research, including extensions to noisy observations, to non‑linear latent models (e.g., manifold learning), and to optimal sampling designs that maximize the probability of achieving global uniqueness. By providing a deterministic pre‑check for uniqueness, the approach can serve as a valuable preprocessing step before applying any matrix‑completion algorithm.