Maximum Likelihood Decoding

• Maximum Likelihood Decoding (MLD) is a method that selects the codeword, error pattern, or logical class maximizing the conditional probability of received data under specific channel models.
• MLD employs metrics like Hamming and Euclidean distance, as well as coset probability evaluations, to optimally decode messages across classical and quantum channels.
• Despite its optimality, MLD is computationally challenging (NP-hard or #P-hard), though tailored algorithmic strategies and approximations can render it tractable for certain code families.

Searching arXiv for recent and relevant papers on Maximum Likelihood Decoding to ground the article. Maximum likelihood decoding (MLD) is the decision rule that selects the codeword, error pattern, or logical class maximizing the conditional probability of the received data under a specified channel model. In classical linear coding, this appears as the search for a codeword $w\in\mathcal{C}$ minimizing Hamming distance to a received word $v$, $d(w,v)=d(v,\mathcal{C})$, and on standard memoryless channels such as the binary symmetric channel this coincides with nearest-neighbor decoding in Hamming distance (Barbier, 2010). Under BPSK over AWGN it is equivalently the minimization of Euclidean distance over the codebook, while in quantum stabilizer coding it is the maximization of a logical-coset posterior obtained by summing probabilities over all physical errors consistent with the observed syndrome (Zhang et al., 2023, Cao et al., 17 May 2026).

1. Formal definitions across decoding settings

The classical formulation for a linear code $\mathcal{C}\subseteq \mathbb{F}_q^n$ asks, for a received vector $v$, for a codeword $w\in\mathcal{C}$ such that

$$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$

with $d(\cdot,\cdot)$ the Hamming distance (Barbier, 2010). For BPSK-modulated linear block codes on AWGN, the same rule is written as

$$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$

so the likelihood criterion becomes an Euclidean nearest-codeword problem (Zhang et al., 2023).

The decision variable depends on the channel and code model. For a channel with unknown gain $a>0$ and offset $v$0, the decoder maximizes likelihood jointly over the codeword and these nuisance parameters: $v$1 which Blackburn rewrites as minimization of the squared distance from the codeword to the affine half-subspace $v$2 (Blackburn, 2015). In quantum stabilizer codes, by contrast, MLD does not choose the most likely individual error. It chooses the logical class $v$3 maximizing

$v$4

that is, the total probability of the syndrome-consistent coset (Cao et al., 17 May 2026).

Setting	Decision object	Likelihood criterion
Linear block code	Codeword $v$5	Minimize Hamming distance (Barbier, 2010)
BPSK over AWGN	Codeword $v$6	Minimize Euclidean distance (Zhang et al., 2023)
Gain/offset mismatch	Codeword $v$7	Maximize over $v$8 and $v$9 (Blackburn, 2015)
Stabilizer QEC	Logical class $d(w,v)=d(v,\mathcal{C})$0	Maximize coset probability (Cao et al., 17 May 2026)

A recurrent misconception is that MLD is always a nearest-codeword or minimum-weight rule. The supplied literature gives several counterexamples. In stabilizer coding, minimum-weight decoding can ignore degeneracy and therefore differ from true MLD (Cao et al., 17 May 2026). In gain/offset channels, Pearson-distance decoding is related to but not identical with the MLD metric $d(w,v)=d(v,\mathcal{C})$1 (Blackburn, 2015).

2. Complexity, hardness, and tractable subclasses

The worst-case complexity of MLD is severe. For general linear codes, Berlekamp, McEliece, and van Tilborg proved NP-hardness, and Guruswami and Vardy showed that the problem remains NP-hard for Reed–Solomon codes (Barbier, 2010). In quantum stabilizer coding, exact evaluation of the MLD probability is $d(w,v)=d(v,\mathcal{C})$2-hard because it requires summing over exponentially many errors in each logical coset (Cao et al., 17 May 2026). The topical review on universal ML decoding similarly takes as baseline the $d(w,v)=d(v,\mathcal{C})$3 operations of exhaustive search for an $d(w,v)=d(v,\mathcal{C})$4 code (Ly et al., 24 Oct 2025).

These hardness statements do not imply uniform intractability for all code families. The notion of an $d(w,v)=d(v,\mathcal{C})$5-covered code makes this explicit. If a code $d(w,v)=d(v,\mathcal{C})$6 has covering radius $d(w,v)=d(v,\mathcal{C})$7 and a polynomial-time list decoder $d(w,v)=d(v,\mathcal{C})$8 reaching radius $d(w,v)=d(v,\mathcal{C})$9, then

$\mathcal{C}\subseteq \mathbb{F}_q^n$0

implies that MLD is solvable in polynomial time by listing all codewords within $\mathcal{C}\subseteq \mathbb{F}_q^n$1 and selecting the closest one (Barbier, 2010). For binary BCH codes, Wu’s list decoder yields concrete examples of such tractable cases, including $\mathcal{C}\subseteq \mathbb{F}_q^n$2, $\mathcal{C}\subseteq \mathbb{F}_q^n$3, $\mathcal{C}\subseteq \mathbb{F}_q^n$4, and $\mathcal{C}\subseteq \mathbb{F}_q^n$5; the paper further states that $\mathcal{C}\subseteq \mathbb{F}_q^n$6 and $\mathcal{C}\subseteq \mathbb{F}_q^n$7 admit quasi-quadratic-time MLD via Wu’s reduced-radius complexity bound (Barbier, 2010).

Quantum and bosonic settings also contain exact tractable islands. The statistical-mechanics formulation of quantum MLD allows exact partition-function evaluation on certain planar graphs via Kac–Ward determinants, and the review cites exact planar MLD for repetition codes under realistic circuit-level noise (Cao et al., 17 May 2026). For surface-square GKP codes on the Gaussian random displacement channel, exact and efficient MLD is implemented up to code distance $\mathcal{C}\subseteq \mathbb{F}_q^n$8 by embedding the decoding problem into the Bravyi–Suchara–Vargo framework (Lin et al., 2024).

This pattern suggests a precise distinction between worst-case complexity and family-specific solvability: NP-hardness or $\mathcal{C}\subseteq \mathbb{F}_q^n$9-hardness governs arbitrary instances, while covering-radius conditions, planar mappings, or lattice structure can restore polynomial-time or quasi-polynomial-time algorithms on restricted classes.

3. Classical algorithmic formulations and approximations

Several papers recast MLD into alternative optimization or dynamical forms. In "Rational Maps and Maximum Likelihood Decodings" (Hayashi et al., 2010), MLD on a binary symmetric channel is written as evaluation of a rational map

$v$0

with $v$1 at the channel-induced point $v$2. Exact bitwise ML decoding becomes thresholding $v$3 at $v$4, and Taylor expansion around the fixed point $v$5 yields approximate ML decoders whose first nonlinear term has order determined by the dual minimum distance $v$6 (Hayashi et al., 2010).

A polyhedral route appears in "A Separation Algorithm for Improved LP-Decoding of Linear Block Codes" (0812.2559). There MLD is the integer program

$v$7

and the paper proposes the compact formulation

$v$8

where the auxiliary variables $v$9 indicate parity violations. Gomory cuts derived from this model recover forbidden-set inequalities, and redundant-parity-check cuts can cut off fractional LP optima while preserving all codewords (0812.2559).

On erasure channels, fountain-code MLD becomes rank testing and linear-system solution. The dissertation "Fountain Codes under Maximum Likelihood Decoding" (Lázaro, 2017) treats inactivation decoding as a practical ML decoder for LT and Raptor codes: the sparse system is triangulated, a dense inactive core is solved by Gaussian elimination, and decoding succeeds iff the relevant matrix has full rank. The average complexity is governed by the number of inactivations rather than by exhaustive elimination over the full matrix (Lázaro, 2017).

For short LDPC codes, "Efficient Near Maximum-Likelihood Reliability-Based Decoding for Short LDPC Codes" (Zhang et al., 2023) uses modified belief propagation followed by order-statistics decoding. The paper states that order-$w\in\mathcal{C}$0 decoding of the proposed algorithm can achieve the performance of order-$w\in\mathcal{C}$1 OSD, thereby approaching MLD for the tested short codes while reducing complexity relative to pure OSD (Zhang et al., 2023).

4. Channel-specific ML metrics and fast exact decoders

The detailed form of MLD depends on the observation model. Blackburn derives an explicit MLD metric for gain/offset mismatch channels: $w\in\mathcal{C}$2 so that $w\in\mathcal{C}$3 (Blackburn, 2015). The same paper introduces a channel-adapted pairwise distance $w\in\mathcal{C}$4 and proves the bound

$w\in\mathcal{C}$5

thereby giving MLD a geometry distinct from both Hamming and Euclidean distance (Blackburn, 2015).

For memoryless binary asymmetric channels, MLD equivalence itself becomes a structural object. The paper "On Equivalence of Binary Asymmetric Channels regarding the Maximum Likelihood Decoding" (Qureshi et al., 2016) defines $w\in\mathcal{C}$6-equivalence of channels by equality of all ML decisions for all $w\in\mathcal{C}$7-block codes and all received words, and shows that the scalar BAC-function

$w\in\mathcal{C}$8

classifies these equivalence classes. Two BACs are $w\in\mathcal{C}$9-equivalent if and only if they have the same $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$0 (Qureshi et al., 2016).

In noncoherent ASK with convolutional coding, the sufficient statistics are envelope magnitudes rather than coherent correlations. The conditional densities are Rayleigh for symbol $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$1 and Ricean for symbol $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$2, and the log-likelihood path metric used by the Viterbi algorithm is built from

$$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$3

which yields true ML sequence decoding under the stated NCASK/AWGN model (Al-Dweik et al., 2019).

Space–time coding provides another exact-ML specialization. For the Golden code, QR decomposition of the effective $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$4 channel matrix produces

$$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$5

with $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$6 and $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$7 real upper-triangular $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$8 blocks. This real-block structure separates parts of the ML metric and yields a worst-case complexity $$d(w,v)=d(v,\mathcal{C})=\min_{c\in \mathcal{C}} d(v,c),$$9, rather than exhaustive $d(\cdot,\cdot)$0, while remaining valid on both quasistatic and rapid time-varying channels (0811.2201).

A system-level variant appears in MU-MIMO symbol-level precoding. There the relevant observation is that the MLD metric depends on $d(\cdot,\cdot)$1, so a larger smallest singular value of the effective channel improves the lower bound on hypothesis separation. The paper therefore proposes the symbol-level smallest singular value maximization problem (SSVMP) because traditional symbol-level precoding produces a rank-one matrix that is infeasible for direct MLD at the receiver (Tong et al., 2024).

5. Quantum and bosonic maximum likelihood decoding

In stabilizer quantum error correction, MLD is Bayes-optimal under a matched prior and $d(\cdot,\cdot)$2–$d(\cdot,\cdot)$3 loss on logical class, but the target of inference is a logical coset rather than a physical error (Cao et al., 17 May 2026). The review organizes the subject through three computational lenses. The statistical-mechanics picture maps MLD to partition-function evaluation in disordered spin models; the tensor-network picture evaluates or approximates coset partition functions on factor graphs; and the artificial-intelligence picture learns $d(\cdot,\cdot)$4 or related joint distributions from data (Cao et al., 17 May 2026).

For bosonic GKP codes on a Gaussian random displacement channel, the same coset-summing principle survives in a lattice form. Given a syndrome-consistent candidate error $d(\cdot,\cdot)$5, the logical-coset likelihood is

$d(\cdot,\cdot)$6

and MLD selects the logical displacement $d(\cdot,\cdot)$7 maximizing this sum (Lin et al., 2024). For surface-square GKP codes, the paper derives per-qubit weights $d(\cdot,\cdot)$8 and reduces the coset sum to a form handled exactly by the Bravyi–Suchara–Vargo algorithm, reaching distance $d(\cdot,\cdot)$9. For color-hexagonal GKP codes, the analogous contraction is done approximately by a bounded-bond-dimension tensor-network method (Lin et al., 2024).

A more combinatorial approximation is given by "Approximate maximum likelihood decoding with $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$0 minimum weight matchings" (Lin, 8 Oct 2025). For graphlike errors, physical errors consistent with a syndrome are placed in one-to-one correspondence with matchings, the first $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$1 minimum-weight matchings are enumerated, and their weights are grouped by logical class to approximate MLD. For correlated $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$2 noise and GKP variants, the paper extends this idea heuristically through separate $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$3- and $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$4-subgraph searches, with fidelity approaching exact MLD or tensor-network decoding as $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$5 increases (Lin, 8 Oct 2025).

These results make clear that “maximum likelihood” in quantum coding is intrinsically a partition-sum problem. Minimum-weight decoding is a low-temperature approximation; exact MLD requires the entropy of entire logical classes.

6. Universal, learned, and search-based accelerations

Two recent directions aim at code-agnostic acceleration. "Universal Maximum Likelihood (List) Decoding via Fast Vector-Matrix Multiplication" (Ly et al., 24 Oct 2025) observes that for any block code over a general memoryless channel, $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$6 can be written as an inner product between a received-sequence vector $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$7 and a codeword incidence vector $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$8. Stacking all $$\hat{\mathbf{c}}_{\mathrm{ML}}=\arg\min_{\mathbf{c}\in\mathcal{C}}\|\mathbf{r}-(1-2\mathbf{c})\|^2,$$9 as columns yields

$a>0$0

whose maximum entry gives the ML codeword. Because $a>0$1 is binary, the Mailman algorithm reduces worst-case time from $a>0$2 to $a>0$3, at the cost of $a>0$4 space for the precomputed matrix (Ly et al., 24 Oct 2025). The same construction extends to ML list decoding with only a logarithmic factor in the list size and to ISI channels by replacing single-symbol states with length-$a>0$5 tuples (Ly et al., 24 Oct 2025).

A second direction learns the search order rather than rewriting the likelihood. "Policy-Guided MCTS for near Maximum-Likelihood Decoding of Short Codes" (Tian et al., 12 Nov 2025) organizes test error patterns into a binary tree and trains a neural policy via MCTS-based learning to guide search toward the TEP producing the MLD codeword. The decoder evaluates candidates through the Euclidean metric consistent with BPSK/AWGN MLD, uses a probability-based early stopping rule, requires no Gaussian elimination compared to ordered-statistics decoding, and can reduce search complexity by $a>0$6 compared to non-GE OSD while achieving lower decoding latency than both OSD and non-GE OSD at high SNRs (Tian et al., 12 Nov 2025).

Taken together, these works suggest two distinct but compatible tendencies. One is algebraic and universal: convert MLD into a primitive such as vector–matrix multiplication and optimize that primitive. The other is search-oriented and data-driven: retain the exact ML metric while learning how to traverse the candidate space more efficiently. Both preserve the defining probabilistic criterion of MLD even as they depart sharply from exhaustive codeword enumeration.