Efficient implementation of linear programming decoding

While linear programming (LP) decoding provides more flexibility for finite-length performance analysis than iterative message-passing (IMP) decoding, it is computationally more complex to implement in its original form, due to both the large size of the relaxed LP problem, and the inefficiency of using general-purpose LP solvers. This paper explores ideas for fast LP decoding of low-density parity-check (LDPC) codes. We first prove, by modifying the previously reported Adaptive LP decoding scheme to allow removal of unnecessary constraints, that LP decoding can be performed by solving a number of LP problems that contain at most one linear constraint derived from each of the parity-check constraints. By exploiting this property, we study a sparse interior-point implementation for solving this sequence of linear programs. Since the most complex part of each iteration of the interior-point algorithm is the solution of a (usually ill-conditioned) system of linear equations for finding the step direction, we propose a preconditioning algorithm to facilitate iterative solution of such systems. The proposed preconditioning algorithm is similar to the encoding procedure of LDPC codes, and we demonstrate its effectiveness via both analytical methods and computer simulation results.

💡 Research Summary

This paper addresses the computational challenges of linear programming (LP) decoding for low‑density parity‑check (LDPC) codes and proposes a fast, scalable implementation that narrows the gap with iterative message‑passing (IMP) decoders. The authors begin by revisiting the original LP formulation of Feldman et al., which relaxes each parity‑check constraint into an exponential number of linear inequalities. Because the number of constraints grows as 2^{d_j‑1} for a check node of degree d_j, direct use of generic LP solvers quickly becomes infeasible for practical code lengths.

To overcome this, the authors first modify the Adaptive LP (ALP) decoding scheme introduced in earlier work. In classic ALP, an initial LP containing only box constraints is solved, and at each iteration all violated parity‑inequalities are added. The new “constraint‑removal” step discards constraints that have become redundant after new ones are introduced. The authors prove that, after each iteration, at most one linear constraint per parity‑check remains active. They call this the single‑constraint property. Consequently, the size of every intermediate LP is bounded by the number of checks m, independent of the maximum check degree. This dramatically reduces memory usage and the number of non‑zero entries in the constraint matrix, making the LP problems structurally similar to the original parity‑check matrix.

Having obtained a sequence of small, sparse LPs, the second contribution focuses on solving each LP efficiently using a sparse interior‑point method. The interior‑point algorithm requires a Newton step at each iteration, which involves solving a linear system A Δx = b. As the algorithm approaches the optimal solution, A becomes increasingly ill‑conditioned, causing standard Conjugate Gradient (CG) iterations to stall. The authors therefore design a preconditioning strategy that exploits the LDPC structure. They construct a preconditioner M by selecting a tree‑like subgraph of the Tanner graph and forming a matrix that mimics the encoding (Gaussian elimination) process of LDPC codes. M is easy to apply (only matrix‑vector products) and captures the dominant spectral characteristics of A, thereby reducing the condition number of M^{-1}A. When used within a Preconditioned Conjugate Gradient (PCG) solver, the number of CG iterations drops substantially, especially in the later interior‑point stages.

The paper provides both theoretical analysis and extensive simulations. Theoretical results show that the preconditioned system’s eigenvalue spread shrinks by roughly a factor of √κ, where κ is the original condition number, and that the greedy subgraph selection algorithm yields a near‑optimal preconditioner with linear‑time complexity in the number of non‑zeros. Simulations on regular (3,6) and (4,8) LDPC codes demonstrate that the combined adaptive‑LP with single‑constraint property and the proposed preconditioner reduces total decoding time by 30–50 % compared with the original ALP and with an unpreconditioned interior‑point implementation, while preserving the error‑rate performance. Moreover, the ML‑certificate property of LP decoding remains intact: whenever the solution is integral it is guaranteed to be the maximum‑likelihood codeword, and non‑integral solutions (pseudocodewords) are correctly identified as decoding failures.

Finally, the authors discuss future work, including automated tuning of the preconditioner, extension to irregular LDPC ensembles, and hardware‑friendly parallel implementations. In summary, the paper delivers a two‑pronged approach—constraint reduction and structure‑aware preconditioning—that makes LP decoding of LDPC codes far more practical without sacrificing its theoretical advantages.