New analysis of iterated linear programming for the binary perceptron
Researchers analyzed how repeated linear programming steps can find solutions for a specific constraint satisfaction problem.
The binary perceptron is a random constraint satisfaction problem. It asks for a sign vector that satisfies a system of random linear inequalities.
The solution space has an extreme geometry. At every positive constraint density, all but a vanishing fraction of the solutions are isolated.
The work provides a precise analysis of iterated linear programming for the binary perceptron. This includes a randomized polynomial-time algorithm with a final rounding step.
For Gaussian matrices and any fixed number of rounds, the iterated linear programs are characterized through a deterministic state-evolution recursion. This characterization also holds for matrices with i.i.d. entries drawn from any centered, unit-variance subgaussian law.
Numerical evaluations of the recursion suggest the algorithm succeeds at constraint density with zero margin. This is beyond the range guaranteed by single-LP analysis and the guarantee from Jo and Lee, 2026.
The proof uses the Lindeberg principle to compare Gaussian and subgaussian matrices. The researchers used a regularized version of each linear program.
Key facts
- The algorithm succeeds at constraint density with zero margin.
- The characterization holds for matrices with i.i.d. entries from any centered, unit-variance subgaussian law.
- The analysis uses a deterministic state-evolution recursion for Gaussian matrices.