---
title: Symbol-Pair Metric Overview
url: https://www.emergentmind.com/topics/pair-symbol-metric
type: topic
---

# Symbol-Pair Metric Overview

The **pair-symbol metric**, more commonly called the **symbol-pair metric**, is a distance on words over a finite alphabet induced by a read model in which a channel outputs **overlapping adjacent pairs** rather than individual symbols. For a length-\(n\) word over \(\mathbb F_q\), the relevant observation is the cyclic pair-read vector \(\big((x_0,x_1),(x_1,x_2),\dots,(x_{n-1},x_0)\big)\), and distance is measured by the Hamming distance between such pair-read vectors. The metric was proposed for **symbol-pair read channels** motivated by high-density storage systems, where isolated symbol reads may be physically impractical and pair-errors are more natural than ordinary symbol errors. In the subsequent literature, the pair-symbol metric became a central object for the construction of **MDS** and **AMDS** codes, for exact distance computation in cyclic and constacyclic families, and for extensions to the general \(b\)-symbol metric [1605.08859] [2206.12668].

## 1. Formal definition and elementary structure

Let \(x=(x_0,x_1,\dots,x_{n-1})\in \mathbb F_q^n\), with indices taken modulo \(n\). The symbol-pair read vector is
\[
\pi(x)=\big((x_0,x_1),(x_1,x_2),\dots,(x_{n-1},x_0)\big).
\]
For \(x,y\in \mathbb F_q^n\), the symbol-pair distance is
\[
d_p(x,y)=\left|\left\{i\in \mathbb Z_n:(x_i,x_{i+1})\neq (y_i,y_{i+1})\right\}\right|,
\]
equivalently the Hamming distance between \(\pi(x)\) and \(\pi(y)\). The associated symbol-pair weight is
\[
w_p(x)=\left|\left\{i\in \mathbb Z_n:(x_i,x_{i+1})\neq (0,0)\right\}\right|,
\]
so that for linear codes \(d_p(x,y)=w_p(x-y)\), and the minimum symbol-pair distance of a linear code is the minimum pair weight of its nonzero codewords [2111.12946] [2010.04329].

The pair-symbol metric differs from the Hamming metric because it counts disagreements of **adjacent ordered pairs** rather than coordinatewise disagreements. A standard comparison used throughout the literature is
\[
d_H(x,y)+1\le d_p(x,y)\le 2\,d_H(x,y),
\qquad 0<d_H(x,y)<n,
\]
and similarly for linear codes,
\[
d_H(C)+1\le d_p(C)\le 2\,d_H(C).
\]
A sharper structural statement is that if the support of the Hamming difference between \(x\) and \(y\) splits into \(L\) consecutive blocks modulo \(n\), then
\[
d_p(x,y)=d_H(x,y)+L.
\]
This formula makes explicit that the pair-symbol metric is sensitive not only to the number of symbol disagreements but also to their **local clustering** [2010.04329] [2111.12946].

Notation varies across the literature. The same distance appears as \(d_p\), \(d_{sp}\), \(D_p\), or \(d_2\), and the corresponding weight as \(w_p\), \(w_{sp}\), or \(w_2\). These are different notational realizations of the same pair-symbol metric [1911.02184].

## 2. Channel model, decoding meaning, and order sensitivity

The symbol-pair metric is tied to the **symbol-pair read channel**, where the decoder does not receive \(x\) itself but the cyclic list of adjacent pairs. This model was introduced to guard against **pair-errors** in high-density storage systems, where the read process may produce overlapping symbol pairs instead of isolated symbols. In that setting, the natural performance parameter is the minimum symbol-pair distance \(d_p(C)\), not the ordinary Hamming distance [1605.08859] [2206.10806].

As in classical coding theory, distance controls correction capability. The literature repeatedly uses the fact that a code with minimum pair distance \(d_p\) can correct up to
\[
\left\lfloor \frac{d_p-1}{2}\right\rfloor
\]
symbol-pair errors. This is the pair-metric analogue of the standard Hamming-metric decoding radius and explains why code constructions aim to maximize \(d_p\) rather than only \(d_H\) [2010.04329] [2206.10806].

A common misconception is that pair distance is determined solely by Hamming weight. The literature shows instead that **coordinate order matters**. In the study of simplex and variation simplex codes, two permutations of the same simplex codeword were shown to have different symbol-pair weights: one had pair weight \(24\), another had pair weight \(21\). Thus rearranging coordinates can preserve Hamming weight while changing symbol-pair weight. This order sensitivity is one of the defining features of the pair-symbol metric and underlies several later construction methods based on coordinate permutations [1911.02184].

## 3. Bounds, optimality notions, and asymptotic length formulas

The basic upper bound is the **Singleton-type bound** for symbol-pair codes:
\[
|C|\le q^{\,n-d_p+2},
\]
for a length-\(n\) \(q\)-ary symbol-pair code with minimum pair distance \(d_p\). A code meeting this bound with equality is called an **MDS symbol-pair code**. For a linear \([n,k]\) code, equality yields
\[
d_p=n-k+2.
\]
Several papers also use the term **AMDS symbol-pair code** for the one-step-short case \(d_p=n-k+1\), or equivalently for linear codes of size \(q^{n-d_p+1}\) [2012.05050] [2206.00462].

The same literature develops further extremal bounds in the generalized \(b\)-symbol setting. For
\[
[b]_q=\frac{q^b-1}{q-1},
\]
a Griesmer-type lower bound for linear \(b\)-symbol codes is
\[
n \ge \left\lceil \frac{g_q\!\left(k,q^{b-1}d\right)}{[b]_q} \right\rceil,
\]
equivalently
\[
\frac{q^b-1}{q-1}\,n \ge g_q\!\left(k,q^{b-1}d\right).
\]
For fixed \(k,q,b\), the optimal length function satisfies
\[
n_q^b(k,d)=\left\lceil \frac{g_q\!\left(k,q^{b-1}d\right)}{[b]_q}\right\rceil
\]
for all sufficiently large \(d\). In particular, for \(b=2\), this yields an eventual exact length formula for linear codes in the pair-symbol metric [2507.07728].

Upper-bound phenomena also appear through covering theory. One paper proves that if a symbol-pair code has minimum pair distance \(D\), then for
\[
D>17 \quad\text{and}\quad n > \frac{(D-1)(q^2-1)}{4(q-1)},
\]
the code is not MDS. This is presented as a general upper bound on the lengths of MDS symbol-pair codes [2206.12668].

## 4. Exact distance computation and algebraic determination

A major strand of the theory concerns exact determination of pair distances for algebraically structured codes, especially repeated-root cyclic and constacyclic codes.

For cyclic codes of length \(p^e\) over \(\mathbb F_{p^m}\), every code has the form
\[
C=\langle (x-1)^i\rangle,\qquad 0\le i\le p^e.
\]
The pair distances of **all** such codes were determined exactly. When \(e=1\),
\[
d_p(C)=
\begin{cases}
2, & i=p,\\[1mm]
i+2, & 0\le i\le p-2,\\[1mm]
p, & i=p-1.
\end{cases}
\]
For \(e\ge 2\), the complete classification is piecewise, with values such as \(3\), \(4\), \(3p^k\), \(4p^k\), \(2(\beta+2)p^k\), and \(p^e\) on explicitly described intervals of \(i\). This theorem gives the exact minimum symbol-pair distance for every repeated-root cyclic code of length \(p^e\) [1607.01887].

For repeated-root cyclic codes over
\[
\mathfrak R=\mathbb F_{p^m}[u]/\langle u^4\rangle
\]
of length \(n=p^k\), the decisive invariant is the **third torsional degree** \(t_3\), defined via
\[
Tori_3(C)=\big((x-1)^{t_3}\big).
\]
The paper shows that both the minimum symbol-pair weight and the minimum RT weight are determined from \(t_3\). In particular,
\[
wt_{\mathrm{RT}}(C)=t_3+1,
\]
while the minimum symbol-pair weight is a piecewise function of \(t_3\) indexed by the \(p\)-adic position of \(t_3\). This reduces a pair-metric distance problem over a non-chain ring to a torsion-degree computation [2010.02471].

For reducible cyclic codes \(C(q,m,h,e)\), cyclotomic numbers, generalized cyclotomic numbers, and Gaussian periods are used to determine possible symbol-pair weights through the count
\[
w_p(c(a,b))=n-T(a,b).
\]
Under the stated \(e=2\) conditions, the minimum pair distance satisfies
\[
d_p(C(q,m,h,2))=2d_H(C(q,m,h,2))=hq^m-hq^{m-1}.
\]
The same work also identifies several **three symbol-pair weight** families and determines their symbol-pair weight distributions [2309.04068].

| Framework | Structural mechanism | Pair-metric consequence |
|---|---|---|
| Cyclic codes of length \(p^e\) | Generator \(\langle (x-1)^i\rangle\) | Exact classification of \(d_p\) for every \(i\) |
| Repeated-root cyclic codes over \(\mathfrak R\) | Third torsional degree \(t_3\) | Piecewise formula for minimum pair weight; \(wt_{\mathrm{RT}}=t_3+1\) |
| Reducible cyclic codes \(C(q,m,h,e)\) | Cyclotomic numbers and Gaussian periods | Explicit pair-weight sets; in \(e=2\) cases \(d_p=2d_H\) |

These results show that the pair-symbol metric is not merely an auxiliary reformulation of Hamming distance. In the algebraic code families where exact analysis is possible, pair distance exhibits its own piecewise regimes, its own extremal constructions, and its own governing invariants.

## 5. Explicit construction of MDS and AMDS symbol-pair codes

The construction theory of optimal pair-symbol codes is extensive and uses projective geometry, elliptic curves, repeated-root cyclic codes, constacyclic codes, matrix-product codes, and simple-root cyclic codes.

A foundational geometric construction proves that a linear MDS \((n,5)_q\) symbol-pair code exists **if and only if**
\[
5\le n\le q^2+q+1.
\]
For pair distance \(6\), linear MDS \((n,6)_q\) symbol-pair codes exist for all
\[
6\le n\le q^2+1,
\]
using ovoids in \(\mathrm{PG}(3,q)\). With elliptic curves, linear MDS \((n,d+2)_q\) symbol-pair codes are constructed for
\[
7\le d+2\le n\le q+\lfloor 2\sqrt q\rfloor+\delta(q)-3,
\]
where \(\delta(q)=0\) or \(1\) [1605.08859].

Repeated-root cyclic constructions produce high pair distances. For prime \(p\) with \(5\mid (p-1)\), MDS \((5p,7)_p\) and \((5p,8)_p\) symbol-pair codes are constructed, and for every odd prime \(p\) there is an MDS \((4p,7)_p\) symbol-pair code [2010.04329]. Other repeated-root constructions over \(\mathbb F_p\) with \(3\mid (p-1)\) give MDS \((3p,10)_p\) and \((3p,12)_p\) symbol-pair codes [2012.05050]. A later paper constructs six AMDS families with
\[
d_p\in\{4,6,7,8,12,13\},
\]
including one class of unbounded lengths and one class with minimum symbol-pair distance \(13\) [2206.10806].

Constacyclic classification results identify exactly which codes in certain repeated-root constacyclic classes are MDS. Over \(\mathbb F_{p^m}\), the listed MDS families include \(\langle x-\alpha_0\rangle\), \(\langle (x-\alpha_0)^2\rangle\), \(\langle (x-\alpha_0)^4\rangle\) in the special case \(p=3,s=2\), \(\langle (x-\alpha_0)^k\rangle\) for \(s=1\), \(\langle (x-\alpha_0)^{p^s-2}\rangle\), \(\langle x^2-\alpha_0\rangle\), \(\langle (x^2-\alpha_0)^k\rangle\) for \(s=1\), and \(\langle (x^2-\alpha_0)^{p^s-1}\rangle\). The same work proves there are **no other nontrivial MDS symbol-pair constacyclic codes** in the studied classes over finite fields and the ring \(\mathbb F_{p^m}+u\mathbb F_{p^m}\) with \(u^2=0\) [2111.12946].

Matrix-product methods with permutation equivalence produce six further MDS classes. Under the stated congruence conditions, the constructions yield MDS \((3n,7)_q\), \((3n,8)_q\), \((3n,10)_q\), \((4n,7)_q\), \((4n,8)_q\), and \((4n,10)_q\) symbol-pair codes over prime power fields. The key mechanism is that permutation equivalence preserves Hamming parameters but can enlarge symbol-pair distance by changing adjacency patterns [2401.01619].

Simple-root cyclic constructions extend the known range of long optimal codes. Three infinite families are obtained:
\[
(4q+4,7)_q \quad\text{for } q\equiv 1\pmod 4,
\]
\[
(4q-4,8)_q \quad\text{for } q\equiv 3\pmod 4,
\]
and
\[
(2q+2,9)_q \quad\text{for odd prime powers } q.
\]
The same paper states that for pair distance \(7\) or \(8\), these \(q\)-ary MDS symbol-pair codes achieve the longest known code length when \(q\) is not a prime, and for pair distance \(9\), they attain the longest known code length regardless of whether \(q\) is prime [2503.20137].

## 6. Generalizations, distributions, covering theory, and quantum extensions

The pair-symbol metric is the case \(b=2\) of the **\(b\)-symbol metric**. For
\[
\pi_b(x)=\big((x_0,\dots,x_{b-1}),\dots,(x_{n-1},x_0,\dots,x_{b-2})\big),
\]
the \(b\)-symbol weight and distance are
\[
w_b(x)=\#\{i:(x_i,\dots,x_{i+b-1})\neq 0\},
\qquad
d_b(x,y)=d(\pi_b(x),\pi_b(y)).
\]
The standard comparison becomes
\[
\min\{d_H(x,y)+b-1,n\}\le d_b(x,y)\le \min\{b\,d_H(x,y),n\}.
\]
This generalization is central in covering theory, list decoding, and asymptotic length questions [2206.12668] [2507.07728].

Weight distribution theory reveals further metric-specific behavior. For \([n,k,d]_q\) MDS codes, the symbol-pair weight distribution \(B_w\) satisfies
\[
B_0=1,\qquad B_w=0\ \text{for }1\le w\le d,
\]
and explicit formulas are derived for \(d<w\le n\). For cyclic simplex codes, all nonzero codewords have the same symbol-\(b\)-weight:
\[
w_b(\mathbf c_a)=
\begin{cases}
q-p^{m-b}, & b<m,\\[1mm]
q-1, & m\le b\le q-1.
\end{cases}
\]
In particular, in the pair-symbol case \(b=2\), every nonzero cyclic simplex codeword has
\[
w_2=q-p^{m-2}.
\]
Variation simplex codes exhibit a different uniform value,
\[
w_2(\mathbf c_a)=q-p^{m-1}+p^{m-2},
\]
while also illustrating that coordinate rearrangement can change pair weight [1911.02184].

Covering theory in the pair and \(b\)-symbol metrics diverges from its Hamming counterpart. The literature proves that there is **no perfect linear symbol-pair code with minimum pair distance \(7\)** and no perfect \(b\)-symbol metric code if
\[
b>\frac{n+1}{2}.
\]
It also determines the covering radius of Reed–Solomon codes:
\[
R_b(RS(n,k))=\min\{n-k+b-1,n\}.
\]
These results feed into generalized Singleton-type bounds for list-decodable \(b\)-symbol codes [2206.12668].

A recent quantum extension relates symbol-pair weight to symplectic weight for CSS decoding. For a binary vector \(c\) and its cyclic left shift \(L(c)\),
\[
\mathrm{wt}_{sp}(c)=\mathrm{wt}_{symp}(c\mid L(c)).
\]
Using this identity, a symbol-pair syndrome decoder for classical cyclic codes is lifted to a decoder for CSS codes, yielding improved correction capability for certain cyclic CSS constructions [2409.10979].

Taken together, these developments show that the pair-symbol metric has evolved from a storage-motivated distance on overlapping pairs into a mature metric framework with its own extremal theory, exact algebraic classifications, covering and list-decoding bounds, metric-specific weight distributions, and even quantum-decoding applications.

Source: https://www.emergentmind.com/topics/pair-symbol-metric