MDS Symbol-Pair Codes

• MDS symbol-pair codes are error-correcting codes designed for channels that read overlapping symbol pairs, achieving the Singleton-type bound.
• They are constructed using methods such as classical MDS, repeated-root cyclic, simple-root cyclic, and matrix-product techniques to optimize pair-distance.
• Their design supports efficient decoding and improved error correction in storage and communication systems by addressing unique pair-error challenges.

A maximum distance separable (MDS) symbol-pair code is a class of error-correcting code defined for channels that read overlapping symbol pairs rather than individual symbols. Such codes achieve the largest possible minimum pair-distance for prescribed length and dimension, meeting the Singleton-type bound. MDS symbol-pair codes combine combinatorial, algebraic, and geometric coding-theoretic techniques and generalize classical MDS codes to address the unique challenges of pair-error correction in practical storage and communication systems.

1. Symbol-Pair Channel Model and Metric

A symbol-pair channel outputs for each transmitted word $x=(x_0,\,x_1,\,\ldots,\,x_{n-1})$ the sequence of pairs $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$, using indices modulo $n$ (Chee et al., 2012, Ma et al., 2019). The symbol-pair distance between $x,y\in\mathbb{F}_q^n$ is defined as

$$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$

and the symbol-pair weight of $x$ is $$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$. The minimum symbol-pair distance of a code $C$ is $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$.

Pair-distance satisfies $d_H(x,y)+1\le d_p(x,y)\le 2d_H(x,y)$ for $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$0; if $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$1 has minimum pair-distance $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$2, then $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$3 can correct up to $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$4 pair-errors.

2. Singleton-Type Bound and Definition of MDS Symbol-Pair Codes

The Singleton-type bound for symbol-pair codes [Chee et al., (Chee et al., 2012)]: $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$5 for length $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$6, alphabet size $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$7, and minimum pair-distance $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$8. For linear $$T(x) = \bigl((x_0,x_1),\,(x_1,x_2),\,\ldots,\,(x_{n-1},x_0)\bigr)$$9 codes, this gives $n$0. A symbol-pair code is called MDS if equality holds: $n$1 and $n$2 in the non-linear case. MDS symbol-pair codes thus maximize the error-correction capability for the pair-metric and generalize classical MDS codes (Ma et al., 2019, Chee et al., 2012).

3. Algebraic Constructions and Families of MDS Symbol-Pair Codes

A broad array of algebraic constructions yields MDS symbol-pair codes:

• Classical MDS Codes: Any $n$3 classical MDS code (e.g., Reed–Solomon, extended RS, Goppa codes) is an MDS symbol-pair code, with $n$4 (Chee et al., 2012, Ma et al., 2019).
• Repeated-Root Cyclic Codes: Families of cyclic and constacyclic codes with generator polynomials involving repeated roots (over $n$5 for primes $n$6) give MDS symbol-pair codes, particularly for pair-distances $n$7 (Ma et al., 2020, Ma et al., 2020, Tang et al., 2022, Tang et al., 2022). Explicit classification for degree $n$8 is available in length $n$9.
• Simple-Root Cyclic Codes: Parameter-optimized generator polynomials over $x,y\in\mathbb{F}_q^n$0 provide infinite families with maximal attainable $x,y\in\mathbb{F}_q^n$1 for given $x,y\in\mathbb{F}_q^n$2; e.g., $x,y\in\mathbb{F}_q^n$3 with $x,y\in\mathbb{F}_q^n$4 for $x,y\in\mathbb{F}_q^n$5, or $x,y\in\mathbb{F}_q^n$6 with $x,y\in\mathbb{F}_q^n$7 for $x,y\in\mathbb{F}_q^n$8 odd (Qiu et al., 26 Mar 2025).
• Matrix-Product Codes: Permuted matrix-product codes with underlying nonsingular-by-columns matrices yield families with high $x,y\in\mathbb{F}_q^n$9, such as $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$0 for lengths $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$1 over $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$2 respectively (Zheng et al., 2024, Xu et al., 2023).
• Projective and Geometric Constructions: Ovoids in projective spaces and functional AG codes from elliptic curves furnish $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$3-ary MDS symbol-pair codes of length up to $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$4 ($$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$5) or $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$6 ($$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$7) (Ding et al., 2016).
• Constacyclic Codes and Chain Rings: Specific $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$8-constacyclic codes of length $$d_p(x,y) = |\{\,i : (x_i,x_{i+1}) \neq (y_i, y_{i+1})\,\}|,$$9 over finite fields and chain rings admit a complete characterization of which generator polynomials achieve the MDS bound for the pair metric (Tang et al., 2021).

A selection of families and their parameters is given below.

Family/type	Length $x$0	Pair-distance $x$1	Field(s)
Reed–Solomon/RS (classical)	$x$2	$x$3	Any $x$4
Repeated-root cyclic	$x$5	$x$6	$x$7 odd
Simple-root cyclic	$x$8	$x$9	$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$0
Matrix-product (with permutation)	$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$1	$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$2	$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$3
AG from elliptic curves	$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$4	$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$5	$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$6 any

4. Combinatorial and Structural Properties

The core property is that for every MDS symbol-pair code, the symbol-pair weight of a nonzero codeword $$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$7 satisfies

$$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$8

with equality at the extremal cases. The symbol-pair weight distribution for $$w_p(x) = d_p(x,0) = |\{\,i : (x_i,x_{i+1})\ne(0,0)\}|$$9 MDS codes is computable in closed form—for $C$0: $C$1 where $C$2, and similar for the second sum (Ma et al., 2019, Zhu et al., 2021).

The error-correction capability is $C$3 pair-errors (Ma et al., 2019). Full knowledge of the weight distribution enables estimation of error detection probabilities and informs the design of decoding algorithms tailored to the pair-metric.

5. Matrix-Theoretic and Geometric Criteria

Necessary and sufficient conditions for a linear $C$4 code to be MDS in the symbol-pair metric can be phrased in terms of generator and parity-check matrices:

• For symbol-pair ($C$5), $C$6 is MDS if and only if for every set $C$7 with $C$8, the submatrix formed by all columns in the $C$9-neighborhood of $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$0 has full rank $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$1 (Liu et al., 2021).
• The parity-check matrix characterization involves the column rank of certain submatrices corresponding to sets of coordinates.

For some cyclic and constacyclic codes, characterization of the MDS property requires explicit exclusion of codewords with consecutive nonvanishing support or analysis of the roots of generator polynomials with respect to configured coordinate neighborhoods (Tang et al., 2021, Tang et al., 2022, Qiu et al., 26 Mar 2025).

6. Decoding, Extensions, and Open Problems

Error correction in the symbol-pair metric generalizes classical bounded-distance decoding, permitting adaptation of syndrome and list decoding variants. For Reed–Solomon and closely related codes, syndrome decoders can be modified to take into account the overlap in pair-read vectors (Ma et al., 2019). Matrix-product and cyclic code frameworks suggest further development of efficient decoding algorithms leveraging structural properties of the underlying algebra or geometry.

Current limitations include the upper bounds on achievable code lengths for fixed dimension and pair-distance, and the tightness of these bounds. For example, for pair-distance $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$2, the simple-root cyclic constructions in (Qiu et al., 26 Mar 2025) yield the longest known $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$3-ary MDS symbol-pair codes for many $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$4 that are not prime. For higher $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$5, extending the geometric and cyclic algebraic methods is an open research direction (Ding et al., 2016).

The study of MDS symbol-pair codes also generalizes to $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$6-symbol MDS codes for $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$7: analogous Singleton-type bounds and algebraic characterizations hold, with structural properties reflecting generalizations of the symbol-pair metric (Liu et al., 2021, Xu et al., 2023).

7. Impact and Applications

MDS symbol-pair codes are central in the design of error-correcting systems for next-generation data storage and communication platforms where physical constraints or device limitations result in overlapping or low-resolution reads. Their algebraic and combinatorial richness—in particular, the possibility of attaining much longer lengths than classical MDS codes due to the symbol-pair metric—has driven both theoretical advances and motivates ongoing research into code families, decoding theory, and applications to practical nonvolatile storage architectures (Chee et al., 2012, Ding et al., 2016, Qiu et al., 26 Mar 2025).

• (Chee et al., 2012) Chee et al., Maximum Distance Separable Codes for Symbol-Pair Read Channels
• (Ma et al., 2019) Ma, Luo, Symbol-pair Weight Distributions of Some Linear Codes
• (Ding et al., 2016) Ding, Ge, Zhang, New Constructions of MDS Symbol-Pair Codes
• (Ma et al., 2020, Ma et al., 2020, Tang et al., 2022, Tang et al., 2022) (repeated-root cyclic code constructions)
• (Qiu et al., 26 Mar 2025) New constructions of MDS symbol-pair codes via simple-root cyclic codes
• (Zheng et al., 2024, Xu et al., 2023) (matrix-product code constructions)
• (Zhu et al., 2021) Zhu, Liao, The $d_p(C) = \min_{x\neq y\in C} d_p(x,y)$8-weight distribution for MDS codes
• (Tang et al., 2021) A Characterization of MDS Symbol-pair Codes over Two Types of Alphabets