Maximum-Likelihood Decoding (MLD)

• MLD is a decoding method that identifies the most likely transmitted codeword by maximizing channel output likelihood, achieving minimal error rates.
• It formulates decoding as an optimization problem—often through integer or linear programming—while being NP-hard for general codes.
• Practical approximations, such as LP relaxations, Monte Carlo methods, and specialized algorithms, enable near-optimal performance in communications and quantum error correction.

Maximum-likelihood decoding (MLD) is the process of selecting, from a codebook, the codeword most likely to have been transmitted given the observed noisy channel output. MLD achieves the minimum possible error probability under the assumed channel model and therefore remains the gold standard for optimal decoding in coding theory, communications, and quantum error correction. However, MLD is computationally intractable (NP-hard) for general codes and channel models, leading to significant research on specialized classes, fast approximations, and algorithmic relaxations.

1. Mathematical Formulation and Optimality Criteria

Under a memoryless channel, the MLD rule selects the codeword $\hat c$ maximizing the conditional probability $P(y|c)$ of observing received vector $y$: $\hat c = \arg\max_{c \in \mathcal{C}} P(y|c)\,.$ For the binary-input additive white Gaussian noise (AWGN) channel with BPSK modulation, this reduces to minimizing the Euclidean distance between $y$ and the codeword’s bipolar embedding $s$: $$\hat c = \arg\min_{c \in \mathcal{C}} \|y - s(c)\|_2^2\,.$$ MLD is optimal in minimizing word error rate (WER), achieving the theoretical best possible reliability for any decoder.

2. Complexity and Fundamental Hardness

Berlekamp, McEliece, and van Tilborg proved that exact MLD for general linear block codes is NP-hard, and this result extends to Reed-Solomon codes and many non-binary families [0702147]. The worst-case time complexity is $O(q^k n)$ for a $[n,k]_q$ code—exponential in code dimension, regardless of the structure of the code.

However, several tractable regimes exist:

• For certain asymptotically good expander and low-density parity-check (LDPC) codes, MLD can be accomplished in expected polynomial time over some channels and rate/noise regimes, aided by the existence of ML certificate decoders (e.g., LP decoders of Feldman) and ML-verifiable decoding [0702147].
• For A-covered codes (codes whose covering radius does not exceed the decoding radius of a polynomial-time list decoder, e.g., Wu's algorithm for binary BCH codes), MLD is polynomial-time solvable (Barbier, 2010).
• For the 2D surface code with independent bit/phase-flip noise, exact MLD is $O(n^2)$ using matchgate simulation techniques (Bravyi et al., 2014).

3. Algorithmic Approaches and Relaxations

3.1 Linear and Integer Programming Relaxations

Feldman's relaxation of MLD for linear codes casts decoding as an integer program (IP), then relaxes it to a linear program (LP). Solutions are exact when the LP yields an integral optimum; otherwise, fractional pseudocodewords can prevent exactness. LP-based separation algorithms (incorporating forbidden-set or redundant parity-check cuts) iteratively tighten the LP until the true ML codeword is found, with competitive empirical performance for moderate block lengths (0812.2559).

3.2 Ordered Statistics and Reliability-Based Decoding

Ordered statistics decoding (OSD) systematically tests low-weight error patterns in the most reliable positions. Combining belief propagation (BP) with OSD in a multistage approach, as in modified BP + OSD schemes, achieves near-MLD performance for short LDPC codes at moderate complexity—the BP phase improves reliability, allowing lower order OSD to achieve the performance of higher-order pure OSD (Zhang et al., 2023). Policy-guided Monte Carlo Tree Search (MCTS) frameworks further optimize the enumeration of candidate error patterns using neural network policies to guide search, approaching near-MLD with significantly reduced complexity for short codes (Tian et al., 12 Nov 2025).

3.3 Matrix-Vector and Universal Frameworks

A universal vector-matrix framework for MLD reduces the per-block complexity for arbitrary block codes from $P(y|c)$0 to $P(y|c)$1 via codebook pre-indexing and the Mailman algorithm for fast vector-matrix multiplication, at the expense of $P(y|c)$2 space (Ly et al., 24 Oct 2025).

3.4 Quantum Decoding Algorithms

Quantum annealing machines (QAMs) can encode MLD as a quadratic unconstrained binary optimization (QUBO) problem mapped to an Ising Hamiltonian, solved physically on superconducting qubit arrays. Embedding complexity for parity-check-based QUBO constructions scales more favorably with code sparsity, allowing practical ML decoding up to blocklengths $P(y|c)$3 for LDPC codes on current hardware, though with performance degradation relative to classical BP as $P(y|c)$4 grows (Ide et al., 2020). MLD for quantum codes is implemented exactly via dynamic programming (GKP surface codes) or approximated via tensor network contraction (tensor network approximate MLD), enabling the determination of quantum capacity thresholds in bosonic Gaussian channels (Lin et al., 2024, Bravyi et al., 2014).

4. Channel Models and MLD Metric Variants

MLD must be tailored to the statistical model of the channel:

• For the memoryless binary asymmetric channel (BAC), MLD partitions the $P(y|c)$5 parameter space into equivalence classes of channels yielding identical decoding regions, with the boundary functions and region sizes governed by the BAC-function $P(y|c)$6 (Qureshi et al., 2016).
• In channels with gain and offset mismatch, the MLD metric becomes a function of codeword variance and Pearson correlation, reducing to minimization of a correlation-based distance (Blackburn, 2015).
• For partial response (ISI) channels, graph-theoretic approaches based on linear programming and message passing can yield exact ML detectors under weak nonnegativity conditions, circumventing the exponential complexity of classical trellis-based Viterbi decoding for some memory/channel classes (0707.1241).
• Hidden Markov models (e.g., convolutional codes as Markov paths) require ML sequence detection, where NLL-based tests can minimize complexity per symbol, yielding $P(y|c)$7 decoding with tight optimality in high SNR limits (0711.3077).

5. Practical Approximations and Hybrid Strategies

Approximate MLD methods seek favorable trade-offs between complexity and error performance:

• $P(y|c)$8-minimum-weight-matching decoders enumerate multiple lightest matchings in quantum code decoding, systematically approaching MLD fidelity as $P(y|c)$9 grows, interpolating between polynomial-complexity MWM and intractable MLD (Lin, 8 Oct 2025).
• Taylor expansions of rational map representations of the MLD rule (in the likelihood domain) yield low-order approximations with accuracy governed by the dual code’s minimum distance; for many practical codes, a third-order expansion virtually matches true ML error rates while dramatically reducing computational burden (Hayashi et al., 2010).
• Tensor network and matrix-product-state (MPS) contraction methods provide scalable, systematic approximation to MLD probabilities for large surface codes and other topological structures; approximation accuracy is adjustable via bond dimension (Bravyi et al., 2014, Lin et al., 2024).

6. Special Cases, Islands of Tractability, and Limitations

Certain code/channel pairs admit efficient MLD, including codes with large decoding radii relative to their covering radius (A-covered codes), expander and LDPC codes in certain noise/rate regimes, and surface and GKP codes under specific noise models [0702147], (Barbier, 2010, Bravyi et al., 2014, Lin et al., 2024). However, in general, the complexity of exact MLD remains prohibitive, and algorithms with ML guarantee are rare outside these islands. Observed performance gaps between practical methods and MLD, even under strong approximations, highlight the continued challenge: tight pipelines for error-rate–complexity optimization, especially as code length and noise models increase in complexity.

7. Impact and Applications

MLD continues to serve as the fundamental decoding benchmark, underpinning the analysis and development of suboptimal decoders in wireless communications, storage, quantum computing, and related disciplines. Advances in near-MLD methods, complexity reductions, and rigorous tractability boundaries provide critical guidance for system design and theoretical performance limits. In quantum error correction, high-accuracy MLD enables meaningful capacity lower bounds for quantum channels and sharp threshold analyses for topological codes (Lin et al., 2024, Bravyi et al., 2014). While general MLD is likely to remain intractable for large-scale block codes, progress in code class–specific methods, scalable approximations, and hybrid classical–quantum decoding continues to narrow the practical gap between MLD and implementable decoding.