MLE Decoder: Optimal Error Correction

• MLE decoding is an algorithm that maximizes the likelihood over error patterns to recover the optimal codeword configuration in both classical and quantum systems.
• It employs diverse strategies such as Taylor expansion, integer programming, and A*-search to efficiently balance performance and computational complexity.
• MLE decoders serve as a gold standard in error correction by significantly reducing error rates despite the NP-hard nature of exact maximum-likelihood decoding.

A Most-Likely-Error (MLE) decoder is a class of decoding algorithm in coding theory and quantum error correction that, given noisy observations and a prior error model, identifies the physical error or codeword configuration with the highest posterior probability, potentially up to code or logical equivalence. The term "MLE decoder" is encountered in both classical and quantum error correction, where it corresponds to the mathematical maximization of the likelihood function over the set of error patterns compatible with the observed syndrome or outputs. Practical MLE decoding often requires approximation or computational optimization due to its inherent NP-hardness for general codes.

1. Mathematical Formulation and Decoder Principle

MLE decoding defines the recovery operation as a global maximization problem:

For a classical code with codebook $\mathcal{C}\subseteq\{0,1\}^n$, received vector $y$, and memoryless channel with transition probabilities $P(y|x)$,

$$\hat{x}_{\mathrm{MLE}} = \arg\max_{x\in\mathcal{C}} P(y|x).$$

In quantum stabilizer codes, noise is modeled as a stochastic Pauli process $E$, the measured syndrome $s$ is extracted, and the MLE decoder seeks the error $E_{\mathrm{MLE}}$ satisfying

$$E_{\mathrm{MLE}} = \operatorname{argmax}_{E\,:~\mathrm{syndrome}(E)=s} P(E),$$

or, modulo stabilizer equivalence, the most-probable logical coset $L$ given the syndrome: $P(L\,|\,s) = \sum_{S\in S}\,P(T_s L S),$ where $S$ runs over the stabilizer group and $T_s$ is a syndrome-determined pure error representative (Iyer et al., 11 Jul 2025).

The same principle governs MLE decoders for block codes over deletion channels, where the objective is to find the codeword $x$ that maximizes the joint probability of generating the observed outputs across all channels (Sabary et al., 2020).

2. Algorithmic Strategies for Efficient MLE Decoding

Due to the intractability of brute-force MLE decoding for large codes, diverse algorithmic paradigms have been developed:

• Taylor-Expansion Rational Map MLE for Linear Codes: ML decoding can be expressed as evaluation of a rational map $f(u)$ at a specialized input ($u^0$) derived from the observed word, with the bit-decision rule determined by the value $f_i(u^0)\geq 1/2$. The rational map can be Taylor-expanded around the non-hyperbolic fixed point $p=(1/2, ..., 1/2)$; terms of the expansion are dictated by the code’s dual distance. Truncating at the dominant term yields an approximate MLE decoder with low complexity and near-ML performance for moderate lengths (Hayashi et al., 2010).
• Error-Building Decoding (EBD): EBD for linear block codes recursively builds "error-building blocks" aligned with the syndrome, leveraging the parity-check matrix alone and minimizing weighted sums of reliabilities via a dynamic program. EBD's complexity is substantially smaller than trellis-ML or exhaustive OSD for many codes (Qiu et al., 5 Jan 2026).
• Integer-Programming-Based MLE Decoding: General stabilizer codes are handled by formulating MLE as an integer optimization problem, minimizing the Hamming weight (or other cost) of the error vector subject to linear syndrome constraints. Logical and stabilizer equivalence is encoded as additional binary variables. While this guarantees ML-optimality, runtime is exponential in general (Harris et al., 2020).
• Minimum-Weight Parity Factor (MWPF) and Hypergraph Techniques: The MLE decoder may be mapped onto a minimum-weight parity factor problem on a hypergraph representing code constraints and error locations. The HyperBlossom algorithm generalizes MWPM and Union-Find methods, offering a primal-dual framework that certifies near-optimality and is practical for large qLDPC codes (Wu et al., 7 Aug 2025).
• A*-Search-Based MLE Decoding (Tesseract): MLE decoding is recast as a shortest-path search in an exponentially large, acyclic graph of error subsets. The Tesseract decoder employs the A* search with an admissible cost heuristic and a suite of pruning strategies for tractable exact or near-exact inference (Beni et al., 14 Mar 2025, Grbic et al., 3 Feb 2026).
• Hybrid Neural-Model Decoders for Short Codes: For short block codes, a serial combination of normalized min-sum (NMS) iterative decoding, neural reliability refinement, and reinforced OSD achieves near-ML frame error rates with 3–5× lower complexity than conventional approaches. Sliding-window early stopping and empirical path updates based on real-time observations help regulate computation and maintain near-ML performance (Li et al., 29 Sep 2025).

3. MLE Decoding in Quantum Error Correction

Quantum codes introduce degeneracy and equivalence between error patterns modulo stabilizers; hence, the MLE decoder maximizes likelihood over cosets:

• Surface Codes and Linear-Time MLE (Erasure Channel): For surface codes subject to erasure, ML decoding reduces to finding any subset of erased edges matching the syndrome parity, uniquely determined by a leaf-peeling algorithm on any spanning forest of the erasure subgraph. This runs in linear time and is provably ML-optimal for the erasure channel (Delfosse et al., 2017).
• Circuit-Level Noise and Marginalization: In the presence of circuit-level errors, maximum-likelihood decoding is formulated via the error-equivalence group and mapped to the partition function of an Ising model on the error variables. Marginalization produces a reduced classical LDPC code, in which free-energy minimization solves the equivalent ML decoding problem (Pryadko, 2019).
• Neural Models for Near-MLE Decoding: Learned decoders such as SAQ-Decoder combine transformer architectures and differentiable logical losses with constraint-aware post-processing to achieve logical error thresholds matching or closely approaching the theoretical ML bound for toric and surface codes. This is accomplished at linear computational complexity, outperforming previous neural and classical decoders (Zenati et al., 9 Dec 2025).

4. Approximations, Complexity, and Performance Considerations

Exact MLE decoding is NP-hard in general (for classical block codes, quantum LDPC, and surface codes under general Pauli noise). Tractable ML decoding is feasible only in special cases (erasure channels, short block codes, thin circuits). Consequently, numerous algorithms rely on approximations:

• Truncation or restriction to low-complexity subspaces (Taylor-ML, EBD, OSD hybrid designs) can approximate full ML behavior at a fraction of the cost.
• Primal–dual LP/ILP relaxations (HyperBlossom, IP) enable certifiable approximation; primal–dual gap certifies solution optimality or proximity to ML when direct minimization is computationally prohibitive.
• A* search with admissible and efficiently computed heuristics allows for heralded near-ML solutions, achieving a balance between speed and worst-case optimality (Beni et al., 14 Mar 2025, Grbic et al., 3 Feb 2026).
• In practical settings, a small fraction of error parameter calibration (e.g., Pauli rates via CER) and completion heuristics (e.g., USS) can enable MLE decoders to outperform baseline decoders by an order of magnitude at moderate code sizes (Iyer et al., 11 Jul 2025).

Empirical studies consistently report that MLE or near-ML decoders deliver substantial improvements in logical or bit error rates, often by an order of magnitude, relative to algebraic or heuristic decoders at the same code parameters (Hayashi et al., 2010, Zenati et al., 9 Dec 2025, Wu et al., 7 Aug 2025).

5. Specialized MLE Decoders in Communication and Deletion Channels

In channels with synchronization errors, such as deletion or insertion channels, the MLE decoder operates by maximizing joint embedding probabilities:

• For two independent deletion channels, the ML decoder chooses the codeword $x$ maximizing $\#(x;y^{(1)})\cdot\#(x;y^{(2)})$, where $\#(x;y)$ is the embedding number of $y$ in $x$. The error probability is dominated by specific error patterns (runs and alternations) and characterized by

$P_e \approx \frac{3q-1}{q-1}\,p^2,$

where $p$ is the deletion probability and $q$ the alphabet size (Sabary et al., 2020).

• For codebooks such as Varshamov-Tenengolts (VT) and shifted VT codes, this error probability analysis extends, providing explicit failure rate estimates for these classes under ML decoding. The theoretical results are consistent with Monte Carlo simulation across code parameters.

6. Theoretical Implications and Comparative Summary

MLE decoding provides a "gold standard"—the optimal achievable performance for a decoding scheme given the error model and code constraints. The growing array of algorithmic designs and efficient implementations (Taylor expansion, error block construction, primal–dual LP/IP, A*, neural post-processed models) reflects ongoing efforts to close the performance–complexity gap for practical decoders. The feasibility and gain of approximate MLE decoding are code- and channel-dependent, with significant impact on code design choices, overhead estimation, and hardware–software partitioning in both classical and quantum error correction.

7. Selected Empirical and Complexity Results

Method/Context	Accuracy Relative to ML	Complexity	System/Code Family
Taylor-ML truncation (Hayashi et al., 2010)	Near-ML (within 0.1 dB)	$O(n^l)$ with sparse terms	BCH, Random [n,k] codes
EBD for Hamming (Qiu et al., 5 Jan 2026)	True MLE	$~10\%$ of trellis ML	Extended Hamming $(64,128,256)$
SAQ-Decoder (Zenati et al., 9 Dec 2025)	$\Delta_p <0.3\%$ to ML	Linear in syndrome	Surface/Toric quantum codes
HyperBlossom (Wu et al., 7 Aug 2025)	4.8x to 1.6x gain vs. matching/BPOSD	Near-linear average; worst-case exp	Surface, color, LDPC
Tesseract (Beni et al., 14 Mar 2025, Grbic et al., 3 Feb 2026)	Matches IP ML	$2\!-\!5\times$ faster than IP	Surface, color, bicycle QEC codes
ML in deletion channel (Sabary et al., 2020)	Exact ML	Poly$(n)$ for $n$ moderate	DNA storage, VT/SVT codes

MLE decoding remains central in benchmarking and designing both classical and quantum codes under the operationally relevant regime of finite size, moderate noise, and real-time computing constraints. Its continued optimization and approximation will remain of critical interest for high-performance, scalable information processing systems.