Linear Time Approximation Schemes for the Gale-Berlekamp Game and Related Minimization Problems

We design a linear time approximation scheme for the Gale-Berlekamp Switching Game and generalize it to a wider class of dense fragile minimization problems including the Nearest Codeword Problem (NCP) and Unique Games Problem. Further applications include, among other things, finding a constrained form of matrix rigidity and maximum likelihood decoding of an error correcting code. As another application of our method we give the first linear time approximation schemes for correlation clustering with a fixed number of clusters and its hierarchical generalization. Our results depend on a new technique for dealing with small objective function values of optimization problems and could be of independent interest.

💡 Research Summary

The paper tackles the Gale‑Berlekamp Switching Game, a combinatorial optimization problem that can be expressed as finding a rank‑1 matrix xyᵀ that minimizes the Hamming distance to a given ±1 matrix M. Recognizing that this game is a special case of a dense fragile MIN‑k CSP (constraint satisfaction problem), the authors develop a linear‑time (in the size of the input) (1+ε)‑approximation scheme that works for the game and a broad class of related problems.

A “fragile” constraint is one that becomes unsatisfied as soon as any variable participating in it is flipped. In dense instances each variable participates in Ω(n^{k‑1}) constraints, guaranteeing that the total contribution of any variable to the objective is large. The Gale‑Berlekamp game corresponds to a fragile dense MIN‑2CSP, and the same structural property holds for several other problems: the k‑ary Nearest Codeword Problem (NCP), the Unique Games Problem with a constant number of colors, correlation clustering with a fixed number of clusters, and its hierarchical generalization.

The algorithm proceeds in three conceptual phases. First, it draws a constant‑size random multiset of (k‑1)‑tuples S₁,…,S_s from the variable set. For each possible assignment to the variables in the union of these samples, it estimates for every variable v and each possible label i an unbiased estimator ˆb(v,i) of the true cost b(v,i) (the number of violated constraints if v were set to i). Using these estimates it performs a greedy assignment x^{(1)} that minimizes ˆb(v,i) for each v. Variables for which the two best labels differ by Θ(n) – called “clear‑cut” variables – are shown to be set correctly in this step.

Second, the algorithm recomputes the exact costs b(x^{(1)},v,i) for all variables, performs another greedy choice x^{(2)} based on these exact values, and identifies the set C of clear‑cut variables (those whose chosen label is strictly better than any alternative by at least a constant fraction of the dense bound). The variables in C are guaranteed to match the optimal solution.

Third, the algorithm fixes the assignments of C and applies a known additive‑error PTAS (e.g., the O(ε)‑additive algorithm for MAX‑k CSP) to the remaining variables V\C. Because |V\C| may be large but the additive error depends only on ε, the total cost of the final solution is at most OPT + ε·|V\C|·poly(1/δ). The overall running time is O(n^{k}) + 2^{O(1/ε²)}: the O(n^{k}) term is unavoidable since any algorithm must read the whole dense input, while the exponential dependence on 1/ε matches the best known bounds for dense CSPs.

The authors instantiate this framework for several concrete problems. For the k‑ary NCP, each equation is an XOR of k variables; the constraints are fragile, so the same linear‑time PTAS applies. For Unique Games with a constant alphabet, the permutation constraints are fragile, yielding a linear‑time PTAS for dense instances. For correlation clustering with a fixed number d of clusters, the problem reduces to a dense fragile MIN‑k CSP when d=2, and to a slightly more general dense problem for d>2; the authors obtain a running time of O(n²·2^{O(d/ε²)}), dramatically improving on the previous O(n⁹·d/ε²·log n) bound. The hierarchical clustering variant receives its first PTAS with the same runtime.

Two technical innovations distinguish this work from prior PTAS for dense CSPs. First, instead of using a large O((log n)/ε²) sample and a single greedy pass, the authors use a constant‑size sample and perform two greedy refinements, which suffices to correctly identify all clear‑cut variables. Second, the additive‑error algorithm is invoked only after the clear‑cut variables have been fixed, rather than at the beginning, reducing the overall computational burden.

In summary, the paper presents the first linear‑time approximation schemes for the Gale‑Berlekamp game and a suite of dense fragile minimization problems, introduces a novel two‑stage greedy refinement combined with additive‑error PTAS, and demonstrates broad applicability to coding theory, unique games, and clustering problems. The results close the gap between theoretical PTAS guarantees and practical, near‑linear runtime performance for a rich class of dense optimization problems.