Guruswami–Sudan Algorithm

• Guruswami–Sudan algorithm is an interpolation-based list decoding method that decodes Reed–Solomon codes beyond the unique decoding limit using multiplicity constraints.
• It constructs a nonzero bivariate polynomial with weighted-degree and multiplicity conditions to extract all low-degree message polynomials via factor extraction.
• Recent advancements optimize its interpolation step, leading to faster algorithms and extensions to generalized and algebraic-geometry codes.

The Guruswami–Sudan algorithm is an interpolation-based list-decoding algorithm for Reed–Solomon codes that decodes beyond the classical half-minimum-distance barrier and, in the standard adversarial model, has long served as the state of the art up to the Johnson bound (Brakensiek et al., 14 Apr 2025). In its classical form, the method constructs a nonzero bivariate polynomial subject to multiplicity constraints at the received points and then extracts all low-degree message polynomials that occur as linear factors in the second variable (Mohamed et al., 2016). Subsequent work has turned this paradigm into a broad framework encompassing generalized Reed–Solomon, algebraic-geometry, Hermitian, and complex Reed–Solomon codes, while also producing faster interpolation methods, deterministic root-finding procedures for the Reed–Solomon case, and several closely related alternatives that match the Guruswami–Sudan decoding radius without reproducing its full list-decoding guarantees (Beelen et al., 2022, Chatterjee et al., 7 Nov 2025).

1. Algebraic decoding setting and decoding radius

For $k \le n \le q$, a Reed–Solomon code evaluates a polynomial $f \in \mathbb F_q[x]$ of degree at most $k-1$ at distinct points $\alpha_1,\dots,\alpha_n$, and a generalized Reed–Solomon code adds nonzero column multipliers $v_1,\dots,v_n$; these are $[n,k,d]_q$ MDS codes with $d=n-k+1$ (Riasat et al., 4 Feb 2025). Classical unique decoding corrects only up to half the minimum distance, whereas list decoding seeks all codewords within a larger radius.

Modern summaries in the supplied literature describe Guruswami–Sudan as efficiently list decoding Reed–Solomon codes up to the Johnson radius $1-\sqrt{R}$, where $R=k/n$, equivalently up to $n-\sqrt{nk}$ errors in the paper’s discussion (Brakensiek et al., 14 Apr 2025). In finite-parameter form, with interpolation multiplicity $f \in \mathbb F_q[x]$0 and $f \in \mathbb F_q[x]$1-degree parameter $f \in \mathbb F_q[x]$2, the decoding radius is stated as

$f \in \mathbb F_q[x]$3

and by choosing $f \in \mathbb F_q[x]$4 and $f \in \mathbb F_q[x]$5 large enough one approaches $f \in \mathbb F_q[x]$6 (Mohamed et al., 2016). This places the algorithm strictly beyond the unique-decoding regime while preserving polynomial-time complexity.

A recurrent theme in later work is that this radius is a benchmark rather than a universal optimum across all noise models. In the fully adversarial setting, Guruswami–Sudan remains the baseline efficient list decoder for Reed–Solomon codes (Brakensiek et al., 14 Apr 2025). In semi-adversarial or random-error models, however, other interpolation-based unique decoders can exceed Johnson-type thresholds because they exploit additional stochastic structure rather than improve the worst-case adversarial bound (Brakensiek et al., 14 Apr 2025).

2. Core mechanism: multiplicity interpolation and factor extraction

The classical Guruswami–Sudan architecture has two stages. First, it constructs

$f \in \mathbb F_q[x]$7

subject to three conditions: $f \in \mathbb F_q[x]$8 is nonzero, it vanishes at the received points with multiplicity $f \in \mathbb F_q[x]$9, and its coefficient polynomials satisfy

$k-1$0

(Mohamed et al., 2016). Second, it recovers all polynomials $k-1$1 with $k-1$2 such that $k-1$3, equivalently all factors $k-1$4 of $k-1$5 (Mohamed et al., 2016).

The multiplicity condition is expressed through Hasse derivatives. For $k-1$6, the $k-1$7-th Hasse derivative at $k-1$8 is defined as the coefficient of $k-1$9 in $\alpha_1,\dots,\alpha_n$0, and multiplicity at least $\alpha_1,\dots,\alpha_n$1 means that every derivative indexed by

$\alpha_1,\dots,\alpha_n$2

vanishes at the interpolation point (Nielsen, 2014). The weighted-degree constraint is the $\alpha_1,\dots,\alpha_n$3-weighted degree: for a monomial $\alpha_1,\dots,\alpha_n$4,

$\alpha_1,\dots,\alpha_n$5

(Nielsen, 2014).

The fundamental correctness statement is that if the actual message polynomial $\alpha_1,\dots,\alpha_n$6 is close enough to the received word and the interpolation constraints above hold, then

$\alpha_1,\dots,\alpha_n$7

is a factor of $\alpha_1,\dots,\alpha_n$8 (Mohamed et al., 2016). The root-finding stage is therefore not an auxiliary optimization but the second algebraic half of the decoder. Over finite fields, the Roth–Ruckenstein procedure is a standard method for extracting candidate roots coefficient-by-coefficient; the complex-field adaptation discussed later shows how sensitive this step is to the underlying arithmetic model (Mohamed et al., 2016).

A common misconception is that Guruswami–Sudan is “just interpolation.” In the supplied literature, interpolation is consistently presented as only the first stage; the practical decoder also needs a factorization or root-finding procedure to recover the candidate message polynomials (Mohamed et al., 2016).

3. Interpolation as bottleneck and the development of faster algorithms

A large fraction of the Guruswami–Sudan literature is devoted to the interpolation step because it is typically the dominant cost. One line of work reformulates interpolation with multiplicities as simultaneous polynomial approximation. In the multivariate framework, the usual Reed–Solomon Guruswami–Sudan problem is the special case $\alpha_1,\dots,\alpha_n$9 of a broader interpolation problem with multiplicities and weighted-degree constraints, and this reformulation yields complexity

$v_1,\dots,v_n$0

for classical Reed–Solomon interpolation, or

$v_1,\dots,v_n$1

with re-encoding; the paper explicitly states a speedup by a factor $v_1,\dots,v_n$2 over the previous fastest known algorithm in that setting (Chowdhury et al., 2014).

Another line accelerates the Kötter–Nielsen–Høholdt interpolation routine. The divide-and-conquer version preserves the same algebraic problem—minimal weighted-degree interpolation in $v_1,\dots,v_n$3 with multiplicity constraints—but reduces the complexity from

$v_1,\dots,v_n$4

for classical KNH to

$v_1,\dots,v_n$5

which is quasi-linear in $v_1,\dots,v_n$6 up to logarithmic factors (Nielsen, 2014). The key observation is that, for multiplicity conditions at one point, only the residue modulo $v_1,\dots,v_n$7 matters, so local Hasse-derivative information can be propagated through transformation matrices instead of repeated full polynomial updates (Nielsen, 2014).

The interpolation bottleneck has also been attacked through ideal-theoretic and prefactor-based methods. A binary-exponentiation construction of the interpolation ideal $v_1,\dots,v_n$8, based on

$v_1,\dots,v_n$9

replaces direct multiplicity-$[n,k,d]_q$0 construction by repeated ideal squaring and randomized ideal multiplication; the paper reports both asymptotic and practical gains over iterative interpolation, with further reduction under re-encoding (0812.4937). A related reduction observes that certain univariate constituents $[n,k,d]_q$1 of the interpolation polynomial have known polynomial divisors before any linear system is solved. Re-encoding yields one class of such divisors, and the paper introduces Sierpinski prefactors, derived from binomial coefficients that vanish modulo the field characteristic, to reduce the number of interpolation unknowns while leaving the decoding radius unchanged (Senger, 2013).

For generalized Reed–Solomon codes, a different optimization strategy is iterative refinement of the interpolation module. Instead of solving the full $[n,k,d]_q$2 problem at once, one starts from $[n,k,d]_q$3, enlarges the module via micro-steps $[n,k,d]_q$4 or $[n,k,d]_q$5, and optionally performs root-finding at intermediate radii. The paper’s conclusion is that this preserves worst-case asymptotic complexity while improving average-case behavior because the decoder may terminate early when the actual error pattern is easier than the final target radius (Nielsen et al., 2014).

4. Extensions to broader code families and arithmetic settings

The Guruswami–Sudan paradigm extends beyond classical Reed–Solomon codes, but the ambient algebra changes substantially. For general algebraic-geometry codes, the interpolation object is

$[n,k,d]_q$6

with $[n,k,d]_q$7 lying in appropriate function-field modules rather than in $[n,k,d]_q$8. A fast AG-code realization represents the relevant spaces as free $[n,k,d]_q$9-modules of rank $d=n-k+1$0, where $d=n-k+1$1 is the smallest positive element in the Weierstrass semigroup at a chosen place, and obtains complexity

$d=n-k+1$2

for a full Guruswami–Sudan-style decoder (Beelen et al., 2022). A later refinement improves the interpolation stage to

$d=n-k+1$3

while keeping the same GS success criterion and root-finding architecture (Beelen et al., 2023).

For one-point Hermitian codes, the same high-level structure survives, but the coefficient ring becomes the coordinate ring of the Hermitian curve and ordinary degree is replaced by the pole-order function $d=n-k+1$4. A sub-quadratic realization of Guruswami–Sudan for these codes uses polynomial-ring matrix minimization over $d=n-k+1$5, reaching interpolation complexity

$d=n-k+1$6

and root-finding complexity

$d=n-k+1$7

while retaining the Johnson-range decoding guarantee for one-point Hermitian codes (Nielsen et al., 2014).

A very different extension treats complex Reed–Solomon codes. There the algebraic framework of Guruswami–Sudan remains conceptually intact, but the dominant difficulty is numerical stability rather than algebraic existence. The interpolation system is solved over $d=n-k+1$8 by singular value decomposition, the Roth–Ruckenstein root-finding stage is modified by thresholding small coefficients, and candidate roots are refined by Newton’s method; the paper further embeds the decoder in a generalized-minimum-distance loop driven by intrinsic soft information (Mohamed et al., 2016). This usefully separates what is intrinsic to Guruswami–Sudan from what is specific to exact finite-field arithmetic.

5. Variants, alternatives, and the role of soft or erasure information

Several decoding methods in the supplied literature are best understood relative to Guruswami–Sudan rather than as direct modifications of it. A prominent example is multiplicity-enhanced power decoding for Reed–Solomon codes. By extending the classical key-equation approach with multiplicities, one obtains a one-pass shift-register or Hermite–Padé decoder whose decoding radius

$d=n-k+1$9

differs from the finite-parameter Guruswami–Sudan radius only by $1-\sqrt{R}$0, and therefore approaches the Johnson radius asymptotically (Rosenkilde, 2015). The essential distinction is that this algorithm is a partial bounded-distance decoder, not a list decoder: it may fail for some error patterns within its nominal radius, whereas Guruswami–Sudan is guaranteed to output all nearby codewords (Rosenkilde, 2015).

An analogous phenomenon appears for one-point Hermitian codes. Improved power decoding can reach the same radius as the Hermitian Guruswami–Sudan decoder using powered key equations and a Padé approximation problem, but it avoids the expensive root-finding stage and remains a partial decoder that can fail inside its nominal radius (Puchinger et al., 2017). Likewise, Wu’s rational-interpolation decoder is presented as the rational analogue of Guruswami–Sudan: for generalized Reed–Solomon codes, it has the same decoding range and the same list-size parameter after the change of variables $1-\sqrt{R}$1, but it interpolates a rational relation attached to error positions rather than a polynomial graph attached to correct positions (Beelen et al., 2012).

The Guruswami–Sudan framework also interacts naturally with erasures and soft information. In threshold-based multi-trial error/erasure decoding for concatenated systems, the outer Reed–Solomon decoder is used repeatedly with different reliability thresholds. The relevant parameter is the error/erasure tradeoff factor $1-\sqrt{R}$2; for bounded-minimum-distance decoding $1-\sqrt{R}$3, while Guruswami–Sudan can realize effective tradeoff factors throughout the interval $1-\sqrt{R}$4 (Senger et al., 2011). Because the true Guruswami–Sudan tradeoff depends on the number of erasures, the paper approximates it by tangent decoders with fixed $1-\sqrt{R}$5, derives optimal threshold sets, and concludes that although GS is stronger per trial, multiple lower-complexity BMD trials can match or even outperform fewer GS trials in the concatenated architecture studied (Senger et al., 2011).

Soft-decision list decoding enters through the Koetter–Vardy variant. In Construction D lattice decoding, Euclidean perturbations are converted into reliability vectors and fed to a Koetter–Vardy decoder for Reed–Solomon subfield subcodes; this supplies the code-level engine for polynomial-time lattice list decoding to radius approaching $1-\sqrt{R}$6 of the minimum distance (Mook et al., 2020). This is not a change to Guruswami–Sudan’s core algebra, but an important illustration of how multiplicity-based interpolation extends from hard-decision Hamming decoding to weighted soft-input settings.

6. Contemporary status, applications, and common points of confusion

Recent work continues to treat Guruswami–Sudan as the canonical adversarial Reed–Solomon list decoder. A 2025 paper on semi-adversarial errors explicitly frames GS as the long-standing state of the art for list decoding Reed–Solomon codes up to the Johnson bound against adversarial errors, while emphasizing that its own gains occur only because the noise model is weaker than fully adversarial corruption (Brakensiek et al., 14 Apr 2025). This distinction is important: improvements beyond Johnson in structured noise models do not constitute adversarial improvements over Guruswami–Sudan.

Another contemporary development is deterministic root-finding for the Reed–Solomon case. A 2025 result shows that Reed–Solomon codes can be deterministically list decoded from agreement $1-\sqrt{R}$7 in time $1-\sqrt{R}$8 over every finite field (Chatterjee et al., 7 Nov 2025). The paper does not alter the interpolation paradigm; instead, it derandomizes the special bivariate factorization instances arising in Sudan and Guruswami–Sudan decoding. In that sense, it strengthens the algorithmic status of Guruswami–Sudan rather than replacing it (Chatterjee et al., 7 Nov 2025).

The algorithm also appears as a black-box subroutine in problems that are not themselves new list-decoding results. In an efficient covering algorithm for generalized Reed–Solomon codes, Guruswami–Sudan is used simply