---
title: Algebraic Local Recoverability
url: https://www.emergentmind.com/topics/algebraic-local-recoverability
type: topic
---

# Algebraic Local Recoverability

Searching arXiv for recent and foundational papers on algebraic local recoverability.
Algebraic local recoverability is the use of algebraic or geometric structure to build codes in which an erased symbol can be reconstructed from a small set of other coordinates. In the literature, this structure appears through evaluation codes built from polynomials, rational functions, function fields, algebraic curves, surfaces, projective bundles, fiber products, and lifted families. The common mechanism is that global functions restrict to low-degree functions on small blocks, fibers, lines, or orbits, so local interpolation, a local parity equation, or a low-weight dual relation determines the missing coordinate [2312.16314, 1603.08876]. A complementary viewpoint treats a fixed dependency pattern as a code topology and asks for the strongest erasure-correction guarantee compatible with that topology; this is the setting of maximal recoverability with locality [1307.4150].

## 1. Core notions and algebraic mechanisms

A linear code \(C \subseteq \mathbb{F}_q^n\) has locality \(r\) if every coordinate can be recovered from at most \(r\) other coordinates. In the evaluation-code language, one writes
\[
C(D,V)=\{(f(P_1),\dots,f(P_n)) : f\in V\},
\]
where \(D=(P_1,\dots,P_n)\) is a tuple of evaluation points and \(V\) is a function space. Local recovery arises when the coordinates are organized into repair groups on which the restriction of every \(f\in V\) has controlled degree or controlled dimension [2312.16314]. The standard Singleton-type bound for locality is
\[
d \le n-k-\left\lceil \frac{k}{r}\right\rceil + 2,
\]
and equality defines optimal locally recoverable codes in several of the constructions surveyed here [1910.13472, 2409.04201].

The dual-code formulation gives a concise algebraic criterion. A coordinate \(i\) is recoverable from a set \(R\) if and only if there exists \(w\in C^\perp\) such that
\[
i\in \operatorname{supp}(w)\subseteq R\cup\{i\}.
\]
Equivalently, the \(i\)-th coordinate has locality \(r\) if \(C^\perp\) contains a word of Hamming weight at most \(r+1\) involving coordinate \(i\) [1307.4150, 1907.05316]. If \(w\in C^\perp\) and \(i\in \operatorname{supp}(w)\), then every codeword \(x\in C\) satisfies
\[
x_i = -w_i^{-1}(w\cdot x),
\]
so the recovery set is \(\operatorname{supp}(w)\setminus\{i\}\) [1907.05316].

This algebraic criterion is compatible with the geometric one. In local-parity constructions, the relation
\[
\sum_{s=1}^{r+1} x_{i,s}=0
\]
is exactly a low-weight dual word; in evaluation constructions, Vandermonde or interpolation matrices express the same dependence in coordinates adapted to a fiber or block [1307.4150, 2409.04201].

## 2. Polynomial, rational-function, and topology-driven models

The polynomial model originates in the use of a degree-\((r+1)\) polynomial \(g(x)\) that is constant on blocks \(A_1,\dots,A_l\) of size \(r+1\). If
\[
\mathcal L(g)=\left\langle g(x)^j x^i: 0\le j\le \frac{k}{r}-1,\; 0\le i\le r-1 \right\rangle,
\]
then on each block \(A_\ell\), the restriction of a codeword becomes a polynomial of degree at most \(r-1\), so one missing symbol is recovered from the other \(r\) values by interpolation. This mechanism underlies the Tamo–Barg family and its reformulations over \(\mathbb P^1\) [1501.04904, 2312.16314].

A recent extension replaces good polynomials by good rational functions. A rational function
\[
h(x)=\frac{f(x)}{g(x)}\in \mathbb F_q(x)
\]
is \((r,l)\)-good if \(\deg(h)=r+1\) and there exist \(l\) disjoint subsets \(A_1,\dots,A_l\subseteq \mathbb P^1(\mathbb F_q)\), each of size \(r+1\), such that \(h\) is constant on each \(A_i\). In the function-field extension \(\mathbb F_q(x)/\mathbb F_q(h(x))\), the condition \(\#h^{-1}(b)=\deg(h)\) is equivalent to total splitting of the rational place \(P_b\), so the number of repair groups is the number of totally split rational places. This Galois-theoretic viewpoint yields explicit families of optimal LRCs and explicit cyclic \(\mathrm{PGL}(2,q)\)-based constructions that can outperform good polynomials of the same degree [2605.11465].

Another rational-function-field approach uses automorphism groups of \(\mathbb F_q(x)\). Since \(\operatorname{Aut}(\mathbb F_q(x)/\mathbb F_q)\cong \mathrm{PGL}_2(q)\), a subgroup \(G\) of size \(r+1\) partitions the \(q+1\) rational places into orbits of size \(r+1\). These orbits become repair groups, and the invariant field \(F^G\) supplies the global coefficients. This produces explicit optimal \(q\)-ary LRCs, including codes of length \(q+1\) via cyclic groups and further families via dihedral groups [1710.09638].

A distinct but closely related algebraic line is maximal recoverability. In data-local \((k,r,h)\)-codes and local \((k,r,h)\)-codes, some parity symbols are local and others are heavy. A code is maximally recoverable if it corrects all erasure patterns that are information theoretically recoverable given the code topology. For data-local codes, maximal recoverability is equivalent to the condition that, for every choice \(E\) of one coordinate from each local group, puncturing those coordinates leaves a \([k+h,k]\) MDS code. In characteristic \(2\), the parity-check framework uses Frobenius powers of coefficients \(\alpha_i\), and the central criterion after puncturing one symbol per group is that
\[
T(S,\mathbf e)=\{\alpha_{i,s}+\alpha_{i,\mathbf e(i)}\}
\]
be \(h\)-wise independent [1307.4150]. This isolates a central distinction: locality specifies allowable dependencies, whereas maximal recoverability specifies the strongest coefficient choice compatible with those dependencies.

## 3. Curves and function fields

The general curve-theoretic framework starts with a separable map of smooth projective curves
\[
g:X\to Y
\]
of degree \(r+1\). If \(S=\{P_1,\dots,P_s\}\subset Y(\mathbb F_q)\) consists of rational points that split completely, then
\[
A=g^{-1}(S)=\{P_{ij}: i=0,\dots,r,\ j=1,\dots,s\}
\]
is partitioned into fibers \(A_j\) of size \(r+1\). If \(L(D)\) is a Riemann–Roch space on \(Y\) with basis \(f_1,\dots,f_m\), and \(x\in \mathbb F_q(X)\) is a primitive element for \(\mathbb F_q(X)/\mathbb F_q(Y)\), then the code space
\[
V=\operatorname{span}\{f_j x^i:\ i=0,\dots,r-1,\ j=1,\dots,m\}
\]
produces an evaluation code of length \(n=(r+1)s\). On each fiber, the \(f_j\) are constant, so the restriction is a polynomial in \(x\) of degree at most \(r-1\), and interpolation on the other \(r\) points recovers the erased symbol [1501.04904, 1603.08876].

This curve formalism extends the Tamo–Barg picture from \(\mathbb P^1\) to higher-genus curves and produces long codes over relatively small fields. Explicit families arise from Hermitian curves and Garcia–Stichtenoth towers. For the Hermitian curve
\[
x^{q_0}+x=y^{q_0+1},
\]
projection to \(y\) yields locality \(r=q_0-1\), projection to \(x\) yields locality \(r=q_0\), and the same curve can also support two disjoint recovering sets for each coordinate [1501.04904]. For Garcia–Stichtenoth towers, the constructions are asymptotically good and the derived rate-distance tradeoffs improve a GV-type benchmark in suitable regimes [1603.08876, 1501.04904].

Curves with separated variables supply a particularly explicit algebraic model. For
\[
\mathcal X:\quad A(Y)=B(X),
\]
with a unique common pole \(Q\), one specializes to \(\phi_1=y\) and \(\phi_2=x\). If the fiber \(y=\beta\) splits completely into \(b\) rational points \(P_\beta=\{P_{\beta,1},\dots,P_{\beta,b}\}\), then the locality is \(r=b-1\), and every erased coordinate in that fiber is recovered from the other \(b-1\) coordinates by Lagrange interpolation in \(x\) [1806.02681]. In special cases, recovery simplifies further: if certain power sums vanish and either \(\operatorname{char}(\mathbb F_q)\mid b\) or \(\ell_0=0\), then
\[
\sum_{i=1}^b f(P_{\beta,i})=0,
\]
so one erasure is repaired by one addition [1806.02681].

The curve perspective also clarifies the geometry-to-coding dictionary. Fibers of a morphism give repair groups, Riemann–Roch spaces control dimension, and divisor or intersection calculations control distance. In this sense, algebraic local recoverability on curves is not an auxiliary decoding trick but a property encoded directly in the function-field architecture of the code [1603.08876, 2312.16314].

## 4. Surfaces, projective bundles, and higher-dimensional geometry

Surface constructions generalize the fiber idea by replacing fibers of curve maps with fibers of a surface fibration. For a surface \(\pi:S\to B\) over a curve \(B\), evaluation points are chosen fiber-by-fiber so that each selected fiber contributes \(r+1\) rational points, giving
\[
n=b(r+1).
\]
Local recovery comes from the geometry of one fiber, while dimension and distance are controlled by divisors and intersection numbers on the ambient surface [1910.13472].

In ruled-surface constructions, the points of a fiber are arranged so that no \(r\) of them lie on a hyperplane after embedding into projective space; in the affine model this gives a Vandermonde matrix. If a function has the form
\[
o = a_0(t)+a_1(t)x+\cdots + a_{r-1}(t)x^{r-1},
\]
then any \(r\) values on a fixed fiber determine the coefficients and therefore the missing value [1910.13472]. Elliptic surfaces use a different mechanism: if \(r+1\) points in a fiber sum to \(0\) in the elliptic-curve group law and none are \(2\)-torsion, then a function in \(\mathcal L(rO)\) is uniquely determined by its values at any \(r\) of those points, with uniqueness proved using Abel’s theorem [1910.13472].

Projective bundles refine this picture and incorporate availability. On
\[
X=\mathbb P^1_x\times \mathbb P^1_y,
\]
one evaluates the space
\[
V=\left\{\sum_{\ell=0}^{r-1} a_\ell(x)y^\ell : a_\ell(x)\in \mathbb F_q[x],\ \deg a_\ell(x)\le b-2\right\}
\]
on \(b\) fibers, each with \(r+1\) points. The evaluation map is injective, the dimension is \(k=(b-1)r\), locality comes from an \(r\times r\) Vandermonde subsystem, and for \(r=1,2,3\) the resulting plane codes satisfy
\[
[n,k,d]_q = [\,b(r+1),\ (b-1)r,\ r+3\,]_q
\]
and are optimal [2409.04201]. For \(r\ge 4\), the paper shows that nonoptimal codes can occur, but a uniformly random choice of the evaluation points yields an optimal code with probability approaching \(1\) as the alphabet size grows [2409.04201].

Surface-based LRCs also appear in explicit cyclic covers of \(\mathbb P^2\). For surfaces
\[
X:\quad w^{r+1}=f_{r+1}(x,y,z),
\]
projection to \([x,y,z]\) has fibers of size \(r+1\) away from the branch locus, and locality follows from a Vandermonde-type recovery matrix on each fiber. This yields explicit optimal examples such as \((18,11,3,2)\) over \(\mathbb F_4\), \((24,17,3,3)\) over \(\mathbb F_5\), and \((110,87,3,4)\) over \(\mathbb F_{11}\) [1701.05212]. The higher-dimensional viewpoint therefore shows that local recoverability is not confined to curve-based AG codes.

## 5. Availability, multiple erasures, and hierarchical recovery

Availability augments locality by requiring several disjoint recovery sets for each coordinate. Fiber products of curves provide a systematic source of such structures. If smooth curves \(\mathcal Y_1,\dots,\mathcal Y_t\) map separably to a common curve \(\mathcal Y\), then their fiber product
\[
\mathcal X=\mathcal Y_1\times_{\mathcal Y}\cdots\times_{\mathcal Y}\mathcal Y_t
\]
comes with projections \(g_j:\mathcal X\to \mathcal Y_j\). Functions built from the primitive elements of the intermediate extensions yield an LRC\((t)\) whose locality is
\[
(d_{h_1}-1,\dots,d_{h_t}-1),
\]
and the recovery sets are disjoint because they come from different projection directions [1612.03841]. This framework produces families from generalized Giulietti–Korchmáros curves, Suzuki curves, van der Geer–van der Vlugt Artin–Schreier fiber products, and Hermitian curves [1612.03841].

Multi-erasure locality is formalized by locality \((r,\delta)\). In the \(J\)-affine variety setting, a coordinate lies in a set \(R\) with \(\#R\le r+\delta-1\) such that the punctured code \(C[R]\) has minimum distance at least \(\delta\); equivalently, any \(\delta-1\) erasures in \(R\) are locally correctable. Subfield-subcodes of \(J\)-affine variety codes realize this through orbit sets
\[
R_P=\{\eta^n\cdot_L P : 0\le n\le q-2\},
\]
whose evaluations satisfy a Vandermonde-type linear system. In the full-index case, the local punctured code on an orbit has parameters \([q-1,r,q-r]\) and is MDS, yielding locality \((r,q-r)\) [1911.07485]. These constructions are notable because some families have lengths \(n\gg q\) and are \((\delta-1)\)-optimal [1911.07485].

A further extension is hierarchical locality. In Reed–Muller codes, nested affine subspaces
\[
\text{point} \subset \text{line} \subset \text{plane} \subset \cdots \subset \mathbb F_q^m
\]
supply a sequence of nested repair sets. In fiber-product codes, the hierarchy comes from nested fibers obtained by successively forgetting coordinates. The resulting \(H\)-level structures have parameters
\[
n_j=\prod_{k=j}^H d_{h_k},\qquad s_j=\prod_{k=j}^H(d_{h_k}-\rho_k+1),\qquad \delta_j=\prod_{k=j}^H \rho_k,
\]
and availability at each level [2310.20533]. This places ordinary locality, availability, and multi-erasure locality inside one nested geometric framework.

Recent work on maximal curves emphasizes that the subgroup-orbit picture can yield much larger availability than earlier examples suggested. Using separable morphisms and automorphism groups of maximal curves, one obtains codes with multiple disjoint recovery sets and, in explicit cases, availability as high as \(q+1\). The same paper also corrects inaccuracies in prior literature concerning fixed fields and parameter formulas for some maximal-curve constructions [2509.15163]. This is one of the places where the algebraic details of the recovery structure materially affect the claimed parameters.

## 6. Optimality, field size, algorithms, and structural issues

The optimality problem splits into several distinct questions. For ordinary LRCs, many polynomial, curve, surface, and projective-bundle constructions meet the Singleton-type bound exactly [1501.04904, 1701.05212, 2409.04201]. For maximal recoverability, the issue is different: the code must correct every erasure pattern not ruled out by the topology itself. In this setting there is a field-size cost. For MR local \((k,r,h)\)-codes with \(h\ge 2\), one has the lower bound \(q\ge k+1\), while the explicit constructions of [1307.4150] give alphabet sizes \(O(k^h)\), then
\[
O\!\left(k^{\left\lceil (h-1)\left(1-\frac{1}{2^r}\right)\right\rceil}\right),
\]
and special improvements \(O(k^{3/2})\) for \(h=3\) and \(O(k^{7/3})\) for \(h=4\) [1307.4150]. The same paper shows that random heavy-parity coefficients typically do not give maximal recoverability unless the field size is roughly \(\Omega(k^{h-1})\), so locality alone does not imply the strongest possible erasure behavior [1307.4150].

The algebraic framework also extends beyond fields. Over a finite chain ring \(R\), an \(R\)-linear code has rank \(K\), and the locality bound becomes
\[
d \le n - K - \left\lceil \frac{K}{r} \right\rceil + 2.
\]
Using well-conditioned sets and good polynomials over \(R[x]\), the Tamo–Barg mechanism extends to free optimal LRCs over chain rings, and similarly to \((r,\rho)\)-locality [2401.05286]. This shows that the essential ingredients are not restricted to vector-space linearity over fields; they can be reformulated in terms of rank, modular dependence, and interpolation over ring-theoretic evaluation sets.

At the structural level, sharp recovery can be computed directly from a parity-check matrix. The Gröbner-test-set algorithm of [1907.05316] constructs, for each coordinate, a minimum-weight dual codeword containing that coordinate, thereby producing a sharp recovery structure. It also yields
\[
\operatorname{loc}(\mathcal C)\ge d(\mathcal C^\perp)-1
\]
and computes the dual distance from the same data [1907.05316]. This algorithmic viewpoint reinforces the basic algebraic fact that recovery sets are encoded in the support structure of the dual code.

Several recurring misconceptions are explicitly resolved in the literature. Locality is not the same as maximal recoverability, because the latter is a topology-relative optimality condition on erasure patterns [1307.4150]. Availability is not the same as global correction, because local recovery consults recovery sets, whereas global correction uses essentially all remaining symbols [1612.03841]. And explicit recent work notes that some earlier maximal-curve constructions required correction at the level of fixed fields and parameter formulas [2509.15163]. Taken together, these clarifications show that algebraic local recoverability is not a single construction technique but a family of algebraic design principles whose precise guarantees depend on topology, geometry, coefficient choice, and ambient algebraic category.

Source: https://www.emergentmind.com/topics/algebraic-local-recoverability