Hierarchical Locally Recoverable Codes

• Hierarchical LRCs are erasure-correcting codes with a multilevel locality structure that employs nested repair groups for step-wise recovery of erasures.
• They optimize trade-offs by using explicit constructions such as pyramid codes and algebraic-geometric methods to balance local repair efficiency with global minimum distance.
• Applications in distributed storage and geo-distributed systems leverage these codes to minimize repair bandwidth and latency while ensuring robust data availability.

Locally recoverable codes with hierarchy—hierarchical locally recoverable codes (H-LRCs)—form a broad class of erasure-correcting codes equipped with a multilevel structure of locality. Each code symbol is protected not just by a single small local group, but by a sequence of nested repair groups of increasing size and repair capability. This design enables efficient, progressive local recovery for small numbers of erasures and robust global recovery for larger erasure patterns. The theory underpins modern distributed storage architectures, where multi-tiered repair granularity is critical for reliability, bandwidth minimization, and latency reduction.

1. Foundational Definitions: Hierarchical Locality

Consider an $[n,k,d]$ linear code $\mathcal{C}$ over a finite field $\mathbb{F}_q$:

• Single-level locality $(r,\delta)$: Every symbol $c_i$ lies in a subset $S_i$ such that $\dim(\mathcal{C}|_{S_i})\le r$ and $\mathcal{C}|_{S_i}$ has minimum distance at least $\delta$.
• Two-level hierarchical locality: For each $i$, there is a “middle” repair code $\mathcal{C}$0 of length $\mathcal{C}$1, dimension $\mathcal{C}$2, minimum distance $\mathcal{C}$3, such that $\mathcal{C}$4 itself has $\mathcal{C}$5-locality.
• $\mathcal{C}$6-level hierarchical locality: Every symbol is protected by a chain of nested repair codes, each with parameters $\mathcal{C}$7, $\mathcal{C}$8, such that each level refines the nesting of the previous.

This structure induces, for each symbol, a chain of increasingly larger recovery sets: the smallest enables correction of few erasures, with higher sets invoked as more erasures occur. Definitions admit both information-symbol-restricted and all-symbol locality variants (Sasidharan et al., 2015).

2. Minimum Distance Bounds and Recovery Capabilities

The nested locality structure fundamentally impacts the code’s minimum distance. For a code with two-level hierarchical locality $\mathcal{C}$9, $\mathbb{F}_q$0:

$\mathbb{F}_q$1

[(Sasidharan et al., 2015), Theorem 3.1].

In the general $\mathbb{F}_q$2-level setting:

$\mathbb{F}_q$3

[(Sasidharan et al., 2015), Theorem 5.2].

These bounds precisely express the redundancy trade-off induced by introducing multiple locality layers: smaller repair groups enable efficient local recovery but necessarily reduce the minimum distance for a fixed code length and dimension.

3. Explicit and Optimal Constructions

Two principal paradigms enable explicit H-LRC constructions:

• Pyramid code generalizations: Starting from a systematic $\mathbb{F}_q$4 MDS code, the generator matrix is extended via block splitting to create nested local and mid-level parity structures, achieving codes that saturate the distance bounds under divisibility conditions $\mathbb{F}_q$5 (Sasidharan et al., 2015).
• Multilevel evaluation/affine group/coset constructions: Codes are constructed as evaluations of tailored polynomials over algebraically selected group chains or subspace partitions (e.g., via multiplicative subgroup lattice or nested F-adic expansions) (Sasidharan et al., 2015, Dukes et al., 2022, Haymaker et al., 2023). This approach generalizes to arbitrary levels—each layer of hierarchy is encoded by coset trees and idempotents or by algebraic geometric stratification (Reed–Muller, fiber product, surface codes, etc.).

Recent advances employ algebraic geometry (e.g., surface fibrations, fiber products) and combinatorial techniques (e.g., sum-rank codes, integrated interleaving) to achieve optimal H-LRCs over broad parameter regimes and often with minimized field size requirements (Dukes et al., 2022, Araujo et al., 1 Feb 2026, Haymaker et al., 2023, Blaum, 2020).

4. Availability, Hierarchical Recovery, and Maximal Recoverability

A major refinement is addition of availability: for each symbol, multiple disjoint repair sets exist at each tier, enabling high parallelism and multiple local repair options (Carvalho et al., 2021, Haymaker et al., 2023). Examples include Cartesian-product evaluation (fibers in distinct coordinate directions) and multidimensional cyclic/affine schemes (Chen et al., 2020).

Parallel to this, maximally recoverable codes with hierarchical locality (MRCs) correct all theoretically correctable erasure patterns under the locality constraints imposed by the hierarchy. These codes achieve tight minimum distance (for two-level, $\mathbb{F}_q$6) and are characterized by explicit combinatorial-independence conditions on the underlying parity-check arrays. General constructions leverage layered Vandermonde and Gabidulin blocks, BCH code design for high-wise independence, and flexible partitioning of local and global parities (Nair et al., 2019, Shivakrishna et al., 2021, Martínez-Peñas et al., 2018).

5. Algebraic, Geometric, and Cyclic H-LRCs

Algebraic-Geometric Methods

Algebraic geometry offers a powerful toolkit for constructing H-LRCs, leveraging covers of curves, fiber product structures, and (more recently) fibrations of algebraic surfaces. Codes arising from rational, elliptic, Kummer, Hermitian, and Artin–Schreier curve towers provide explicit two- and multilevel hierarchy with high rates and optimal locality parameters (Ballentine et al., 2018, Haymaker et al., 2023, Araujo et al., 1 Feb 2026).

Cyclic and Multidimensional Codes

Cyclic/LRC codes with hierarchical locality exploit the properties of roots in the code's generator polynomials to embed multi-tiered locality. By tailoring the zero set, one engineers nested repair structures efficiently representable in quasicyclic or tail-biting convolutional form, enabling further extensions to integrated interleaving and multidimensional cyclicity with availability (Chen et al., 2020, Blaum, 2020).

6. Parameter Trade-offs, Field Size, and Complexity

The principal trade-off in hierarchical locality is between local recovery efficiency (minimizing the number of symbols accessed for typical small erasure events) and global code parameters—dimension, rate, and minimum distance. Adding hierarchy (smaller $\mathbb{F}_q$7, more levels) reduces the local repair burden at the expense of overall dimension (rate loss), as precisely captured in the form of the distance bound. Most optimal constructions require field sizes linear or quasi-polynomial in the code length; number-theoretic and combinatorial tools (e.g., Chebotarev density, explicit BCH-independence) are essential in ensuring parameters are attainable without field-size barriers (Sasidharan et al., 2015, Dukes et al., 2022, Shivakrishna et al., 2021).

Complexity of encoding and local repair is typically $\mathbb{F}_q$8 for evaluation codes and quadratic in group size for small repairs. Hierarchical codes enable lowest-latency repairs for common single-symbol erasures, with escalation to higher tiers only for more severe local failures (Dukes et al., 2022, Carvalho et al., 2021).

7. Applications and Open Directions

Hierarchical LRCs are foundational in the architecture of geo-distributed storage and multi-tiered cache networks, where distinct failure domains and repair costs motivate variable, nested recovery granularity. Substantial research focuses on expanding the available parameter regimes (via algebraic geometry, sum-rank codes, and F-adic constructions), reducing field sizes, and integrating availability with hierarchical locality. Extension to more than two levels with all parameters unconstrained, explicit small-field constructions, and optimizing bandwidth/latency trade-offs remain active open problems (Dukes et al., 2022, Haymaker et al., 2023). Asymptotically good H-LRC sequences have been constructed from function field towers, showing deep connections between code theory and arithmetic geometry (Ballentine et al., 2018).

---

Key references: (Sasidharan et al., 2015, Carvalho et al., 2021, Nair et al., 2019, Shivakrishna et al., 2021, Dukes et al., 2022, Chen et al., 2020, Araujo et al., 1 Feb 2026, Haymaker et al., 2023, Blaum, 2020, Ballentine et al., 2018, Martínez-Peñas et al., 2018).