Error-and-Erasure Decoding

• Error-and-erasure decoding is a method addressing both unflagged errors and declared erasures by leveraging specific tradeoffs, such as 2ε + τ < dmin, to ensure successful decoding.
• It utilizes classical algebraic approaches and advanced decoder capability functions to map the balance between error correction and erasure handling.
• Modern implementations incorporate reliability-based erasure selection and iterative, soft-aided techniques to optimize performance across various coding scenarios.

Error-and-erasure decoding is the family of decoding methods in which the receiver corrects both errors, where a received symbol is wrong but not flagged, and erasures, where a symbol position is declared unknown. In its classical form for a linear code of minimum distance $d_{\min}$, the decoder trades one erasure against roughly half an error-correction step; in modern formulations, that tradeoff is expressed by decoder-specific capability functions, probabilistic reliability models, or channel-dependent exponent criteria. The topic spans algebraic decoding of Reed–Solomon and BCH codes, generalized remainder and affine-variety codes, sum-rank-metric codes, iterative decoders for product and staircase codes, and information-theoretic erasure/list decoders for discrete memoryless channels (Senger et al., 2011, Hörmann et al., 2022, 0801.4544).

1. Classical decoding radius and generalized tradeoff laws

For a linear block code $[n,k,d_{\min}]$, the classical unique-decoding statement is that a decoder can correct up to $t$ errors if $2t \le d_{\min}-1$, up to $e$ erasures if $e \le d_{\min}-1$, or any mixture $(t,e)$ satisfying

$2t + e \le d_{\min} - 1.$

For Reed–Solomon codes $(q;n,k,d_{\min})$ with $d_{\min}=n-k+1$, standard bounded minimum distance decoding with erasure handling succeeds whenever

$[n,k,d_{\min}]$0

equivalently

$[n,k,d_{\min}]$1

errors remain among the non-erased symbols (Senger et al., 2011).

A broader algebraic viewpoint replaces the fixed coefficient $[n,k,d_{\min}]$2 by a decoder-specific tradeoff parameter or capability function. For binary codes, a $[n,k,d_{\min}]$3-extended bounded-distance decoder corrects any pattern satisfying

$[n,k,d_{\min}]$4

so smaller $[n,k,d_{\min}]$5 makes errors “cheaper” relative to erasures (Senger et al., 2010). For Reed–Solomon codes, the same idea is abstracted by a decoder capability function $[n,k,d_{\min}]$6, where guaranteed decoding is characterized by

$[n,k,d_{\min}]$7

The maximal correctable number of errors for a fixed erasure count $[n,k,d_{\min}]$8 is then

$[n,k,d_{\min}]$9

which defines an error-correction boundary in the $t$0-plane (Senger et al., 2011).

This generalization is structurally important because it separates the geometry of the decoder’s success region from the strategy used to choose erasures. The classical inequality $t$1 is therefore one special case inside a larger family of error-and-erasure tradeoff laws (Senger et al., 2011).

2. Reed–Solomon decoding and decoder capability functions

For Reed–Solomon codes, the decoder capability function unifies several algebraic decoders. For BMD decoding, including Berlekamp–Massey or Sugiyama with erasures,

$t$2

so the success condition is exactly $t$3, and

$t$4

For an IRS-based decoder for $t$5-punctured Reed–Solomon codes,

$t$6

and

$t$7

For the Guruswami–Sudan list decoder in the high-multiplicity limit,

$t$8

with success region

$t$9

and boundary

$2t \le d_{\min}-1$0

These examples show that error-and-erasure decoding regions may be linear or nonlinear, but the adaptive machinery only needs the mapping $2t \le d_{\min}-1$1 (Senger et al., 2011).

The same subject also has a multi-trial form. Forney’s generalized minimum distance decoding performs several trials with increasing erasure counts and approaches ML performance when the number of trials is approximately $2t \le d_{\min}-1$2, but with a complexity increase by a factor $2t \le d_{\min}-1$3. A later treatment generalized threshold-based multi-trial decoding to decoders with tradeoff factor $2t \le d_{\min}-1$4, including BMD and Guruswami–Sudan decoding. For a constant-$2t \le d_{\min}-1$5 decoder, the guaranteed decoding condition is essentially

$2t \le d_{\min}-1$6

and the optimal thresholds for $2t \le d_{\min}-1$7 trials are

$2t \le d_{\min}-1$8

with

$2t \le d_{\min}-1$9

For classical BMD, $e$0 and the thresholds reduce to

$e$1

In the studied RS$e$2 setting, BMD decoders with $e$3 trials can yield lower residual codeword error probability than GS decoders with $e$4 trials if $e$5 is only slightly larger than $e$6 (Senger et al., 2011).

A recurrent misconception is that stronger list decoders always benefit more from erasures. For the Guruswami–Sudan decoder, the error-correction boundary decreases concavely with $e$7, and the cited simulations show that adaptive erasures can be practically negligible for the studied code, unlike the BMD case (Senger et al., 2011).

3. Reliability-based erasing strategies and residual error optimization

Modern error-and-erasure decoding often begins with a soft-output channel but ends with a hard/erasure algebraic decoder. The central design problem is then to decide which symbols to erase and how many erasures to introduce.

For Reed–Solomon codes over a symbol-by-symbol channel, one computes a symbol unreliability

$e$8

or, under equiprobable signaling,

$e$9

with a nearest-neighbor approximation in practice. The received symbols are sorted by decreasing unreliability,

$e \le d_{\min}-1$0

and for a given $e \le d_{\min}-1$1, the $e \le d_{\min}-1$2 most unreliable symbols are erased before applying the algebraic decoder. If $e \le d_{\min}-1$3 denotes the number of errors among the non-erased symbols, then decoding succeeds iff $e \le d_{\min}-1$4, so the residual codeword error probability is

$e \le d_{\min}-1$5

and the optimal erasure count is

$e \le d_{\min}-1$6

The exact computation uses the probability generating function

$e \le d_{\min}-1$7

but direct evaluation for all $e \le d_{\min}-1$8 is about $e \le d_{\min}-1$9. Two approximations were proposed: a Hoeffding-based windowed summation of order $(t,e)$0, and an empirical “$(t,e)$1-boundary” approximation of order $(t,e)$2 that uses only one or two PGF coefficients (Senger et al., 2011).

An analogous per-vector optimization was developed for binary codes with $(t,e)$3. There the hard decisions are derived from antipodal signaling, the symbol unreliability on AWGN is

$(t,e)$4

and the optimal number of erased least-reliable positions is

$(t,e)$5

with

$(t,e)$6

The exact algorithm has complexity $(t,e)$7, while a Hoeffding-based approximation reduces this to $(t,e)$8 (Senger et al., 2010).

For the Reed–Solomon code $(t,e)$9 on AWGN with 256-QAM and Gray labeling, adaptive single-trial error/erasure decoding yields roughly $2t + e \le d_{\min} - 1.$0 dB gain in $2t + e \le d_{\min} - 1.$1 at practical residual codeword error probabilities compared to errors-only BMD decoding, and captures most of the gain of Forney GMD while using only one decoding run. By contrast, the potential gain from adaptive erasures is negligible for the Guruswami–Sudan list decoder in the same setup (Senger et al., 2011).

A separate line of work analyzed the decoding error probability of binary BCH codes under BDD, EaE decoding, and anchor-bit variants. It derived closed-form expressions for the decoding transition probabilities and showed that anchor bits reject candidate codewords that disagree with highly reliable positions, significantly reducing miscorrections while retaining most of the extra decoding capability of EaE decoding (Miao et al., 29 Sep 2025).

4. Beyond Hamming-metric symbol erasures

Error-and-erasure decoding is not confined to Reed–Solomon or binary Hamming-metric settings. Several algebraic generalizations replace simple symbol counts by weighted degrees, row/column spaces, or Gröbner-basis support sets.

For irreducible polynomial remainder codes, one starts from pairwise coprime monic irreducible moduli $2t + e \le d_{\min} - 1.$2 and the CRT map

$2t + e \le d_{\min} - 1.$3

The error pattern is split into known positions $2t + e \le d_{\min} - 1.$4 and unknown positions $2t + e \le d_{\min} - 1.$5, with locator polynomials

$2t + e \le d_{\min} - 1.$6

and the joint decoding condition becomes

$2t + e \le d_{\min} - 1.$7

Two fixed-transform approaches incorporate erasures without recomputing the CRT basis: one multiplies by the erasure locator on the polynomial side, the other divides it out of $2t + e \le d_{\min} - 1.$8. Both lead to partial extended-Euclid algorithms that recover a scalar multiple of $2t + e \le d_{\min} - 1.$9, and both reduce to the known error-only decoders when $(q;n,k,d_{\min})$0. Approach II appears to be new even for Reed–Solomon codes (Yu, 2012).

For affine variety codes, the central structural statement is that the vector space indexed by codeword positions and the vector space indexed by the support of the Gröbner basis of the vanishing ideal are canonically isomorphic. The isomorphism is realized by extension through linear feedback shift registers induced by the Gröbner basis and generalized discrete Fourier transforms. This lemma yields a unified system for encoding and decoding erasures and errors in a class of affine variety codes, and systematic encoding becomes a special erasure-only decoding problem (Matsui, 2012).

For linearized Reed–Solomon codes in the sum-rank metric, the decoder handles three classes of corruption: $(q;n,k,d_{\min})$1 where $(q;n,k,d_{\min})$2 are full errors, $(q;n,k,d_{\min})$3 row erasures, and $(q;n,k,d_{\min})$4 column erasures. The unique-decoding condition is

$(q;n,k,d_{\min})$5

and the syndrome-based Berlekamp–Massey-like decoder operates over a skew polynomial ring with complexity at most $(q;n,k,d_{\min})$6 operations in $(q;n,k,d_{\min})$7. This extends the classical RS and Gabidulin error-and-erasure inequality to the sum-rank setting and enables decoding in multishot network coding via the sum-subspace reduction (Hörmann et al., 2022).

These frameworks show that the essential object in error-and-erasure decoding is not always the number of bad positions. Depending on the code family, the relevant “erasure budget” may be a weighted degree, a rank quantity, or a root-space dimension. This suggests that the classical Hamming-metric inequality is best viewed as one instance of a more general algebraic balance law.

5. Iterative, ternary, and soft-aided decoding architectures

In high-rate product and staircase codes, error-and-erasure decoding is used as a low-complexity way to inject limited soft information into iterative bounded-distance decoding. For a BI-AWGN channel, the receiver quantizes each sample to a ternary alphabet $(q;n,k,d_{\min})$8 using a threshold $(q;n,k,d_{\min})$9: $d_{\min}=n-k+1$0 Component decoders then operate on ternary vectors. A standard EaE decoder for a BCH component code corrects all patterns satisfying

$d_{\min}=n-k+1$1

where $d_{\min}=n-k+1$2 is the number of errors on non-erased positions and $d_{\min}=n-k+1$3 the number of erasures. Density-evolution analysis for product and staircase codes with ternary message passing showed that additional coding gains of up to $d_{\min}=n-k+1$4 dB are possible by ternary decoding, at only a small additional increase in complexity compared to traditional binary message passing. The same analysis also showed that even-weight BCH subcodes can improve decoding performance at high rates (Rapp et al., 2021).

The extrinsic message-passing decoder required by density evolution is, however, much more complex than intrinsic message passing. A later simplification derived EMP outputs from IMP decoding plus additional logical operations based on the component-code structure and the EaE rule. Simulation results showed that the number of BDD steps is reduced to being comparable with IMP, and a heuristic modification yielded up to $d_{\min}=n-k+1$5 dB improvement compared to standard EMP decoding (Miao et al., 2022).

A related soft-aided architecture for product codes uses dynamic reliability scores. Initial reliabilities are derived from the channel magnitudes, erasures are generated by a threshold on $d_{\min}=n-k+1$6, and a 5-bit dynamic reliability score per bit determines whether a bit becomes an “anchor” that is not allowed to flip. When a component decode would flip an anchor bit, the decision is discarded and the reliability score is reduced. This decoder, based on dynamic reliability scores and error-and-erasure decoding, achieved coding gains of up to $d_{\min}=n-k+1$7 dB compared to conventional hard-decision decoders (Miao et al., 2021).

Error-and-erasure decoding has also been integrated into universal noise-guessing decoders. In GRAND-EDGE, jammed positions are treated as erasures, and the parity-check matrix is split into erased and non-erased columns: $d_{\min}=n-k+1$8 The erased bits are then reconstructed by solving

$d_{\min}=n-k+1$9

via Gaussian elimination. For a random linear code of rate $[n,k,d_{\min}]$00, GRAND-EDGE and ORBGRAND-EDGE lowered both BLER and computational complexity by up to five order of magnitude compared to the original GRAND and ORBGRAND algorithms, and ORBGRAND-EDGE improved BLER by up to three orders of magnitude compared to Ordered Statistics Decoding in the reported simulations (Ercan et al., 2023).

A common misunderstanding is that ternary or erasure-aware decoders are merely heuristic variants of hard-decision decoding. In the cited product, staircase, BCH, and GRAND settings, the improvement is tied to explicit decoding radii, density-evolution equations, or closed-form error-probability calculations rather than to purely empirical thresholding (Rapp et al., 2021, Miao et al., 29 Sep 2025, Ercan et al., 2023).

6. Information-theoretic formulations, universality, and asymptotic behavior

In information theory, erasure decoding is often formulated as a Neyman–Pearson tradeoff between undetected errors and erasures, or between list-error probability and expected list size. For constant-composition random codes over a compound class of DMCs, the $[n,k,d_{\min}]$01-MMI decoder accepts a message $[n,k,d_{\min}]$02 if

$[n,k,d_{\min}]$03

and erases otherwise. If $[n,k,d_{\min}]$04, there is at most one accepted message; if not, the rule becomes a list decoder. The incorrect-message exponent is

$[n,k,d_{\min}]$05

while the erasure exponent is

$[n,k,d_{\min}]$06

and the optimal weighting function for list decoding is

$[n,k,d_{\min}]$07

In the relevant regime, the exponents take the Forney form

$[n,k,d_{\min}]$08

showing that universal decoding can match Forney’s classical exponents for symmetric channels in the small-erasure-exponent regime (0801.4544).

A related universal treatment exists for channels with non-causal side information at the transmitter. There the decoder uses the metric

$[n,k,d_{\min}]$09

and parameters $[n,k,d_{\min}]$10 define both erasure decoding ($[n,k,d_{\min}]$11, $[n,k,d_{\min}]$12) and list decoding ($[n,k,d_{\min}]$13, $[n,k,d_{\min}]$14). The same parametric rule yields universally achievable exponents for total error, undetected error, list error, and average number of incorrect messages on the list, extending the Forney and Csiszár–Körner perspective to the Gel’fand–Pinsker setting (0903.2203).

On the pure erasure channel, exact ensemble-average formulas can be written for unambiguous decoding, list decoding, and ML decoding of random matrix ensembles. For a fixed erasure set $[n,k,d_{\min}]$15, the number of codewords consistent with the unerased symbols is

$[n,k,d_{\min}]$16

Unambiguous decoding succeeds iff $[n,k,d_{\min}]$17; list decoding with list size $[n,k,d_{\min}]$18 succeeds when $[n,k,d_{\min}]$19; ML decoding chooses uniformly among all consistent codewords. The ensemble-average decoding error probabilities and their error exponents were computed explicitly, and for unambiguous decoding the variance exponent implies strong concentration: roughly speaking, the ratio between a random code’s decoding error probability and the ensemble average converges to $[n,k,d_{\min}]$20 with high probability as blocklength grows (Chan et al., 2021).

Finally, typical random-channel performance can differ sharply from worst-case guarantees. For binary locally recoverable codes with locality $[n,k,d_{\min}]$21 and availability $[n,k,d_{\min}]$22, majority-logic decoding over the BSC has bit-failure probability bounded by

$[n,k,d_{\min}]$23

and over the BEC by

$[n,k,d_{\min}]$24

If $[n,k,d_{\min}]$25, the block decoding failure probability vanishes asymptotically for fixed $[n,k,d_{\min}]$26, and the analysis shows that majority-logic decoding can successfully correct virtually all error and erasure patterns of weight linear in the blocklength (Ly et al., 13 Jan 2026). This makes explicit a broader point: worst-case error-and-erasure guarantees and stochastic-channel behavior can be quantitatively far apart.

Error-and-erasure decoding is therefore not a single algorithmic technique but a layered concept. At one extreme it is an algebraic radius condition such as $[n,k,d_{\min}]$27; at another it is a universal hypothesis-testing rule with sphere-packing exponents; in modern iterative systems it is a ternary-message architecture with explicit complexity-performance tradeoffs. Across these settings, the defining principle remains the same: decoder performance improves when uncertainty is separated into unknown values at known locations and unknown values at unknown locations, and the decoder is designed to exploit that distinction.