Decentralized Coded Caching Insights

• Decentralized coded caching is a two-phase scheme where users independently cache segments of files, creating overlap that enables efficient coded-multicast delivery.
• It achieves near-optimal memory–rate tradeoffs with a performance gap to centralized methods bounded between 1 and 1.5, which narrows as the number of users increases.
• Variants extend the canonical model to heterogeneous caches, finite file sizes, coded prefetching, and secure or network-specific delivery, addressing modern practical challenges.

Decentralized coded caching (DCC) is a two-phase caching paradigm in which each cache independently and randomly stores a subset of every file during an offline placement phase, and the server later exploits the resulting overlap structure to transmit coded multicast messages that simultaneously satisfy multiple demands over a shared link (Maddah-Ali et al., 2013, Karat et al., 2019). In the canonical Maddah-Ali–Niesen formulation, placement is uncoordinated and demand-agnostic, yet still creates coded-multicasting opportunities and attains an order-optimal memory–rate tradeoff (Maddah-Ali et al., 2013). Subsequent work established exact worst-case optimality for the Ali–Niesen decentralized scheme under decentralized uncoded placement when $N \ge K$ (Karat et al., 2019), bounded its multiplicative gap to centralized coded caching between $1$ and $3/2$ (Yan et al., 2016), and extended the framework to heterogeneous caches, finite file size, shared caches, multi-access networks, wireless and fog architectures, secure D2D delivery, and asynchronous online settings (Amiri et al., 2016, Jin et al., 2016, Peter et al., 2021, Dutta et al., 2022, Girgis et al., 2017, Zewail et al., 2018, Jiang et al., 2019).

1. Historical development and conceptual position

DCC emerged as the decentralized counterpart to centralized coded caching. The central issue addressed in the original formulation is whether coded-multicasting gain can still be created when no central coordinating server is available during placement (Maddah-Ali et al., 2013). The Maddah-Ali–Niesen decentralized scheme answers this affirmatively by replacing coordinated combinatorial placement with independent random placement, while retaining multicast coding in delivery (Maddah-Ali et al., 2013, Karat et al., 2019).

A defining historical result is that the decentralized scheme achieves a rate close to the optimal centralized scheme (Maddah-Ali et al., 2013). The original order-optimality statement was later sharpened in two directions. First, the multiplicative gap between decentralized and centralized coded caching was shown to lie between $1$ and $3/2$, with both bounds achievable in some cases, and the gap can be arbitrarily close to $1$ if the number of users is large enough (Yan et al., 2016). Second, for $N \ge K$, the Ali–Niesen decentralized delivery was shown to be optimal in the worst case under decentralized uncoded placement by mapping delivery to an index-coding instance and proving $\alpha=\kappa$ for the induced problem (Karat et al., 2019).

This historical trajectory places DCC in a distinct position. It is not merely a heuristic relaxation of centralized coded caching; within the uncoded decentralized-placement model and worst-case distinct demands for $N \ge K$, it is rate-optimal (Karat et al., 2019). At the same time, its architectural rationale is robustness to unknown or changing user sets, asynchronous arrivals, and lack of placement coordination (Maddah-Ali et al., 2013, Miraftab et al., 2 Aug 2025).

2. Canonical model, placement, and coded delivery

The canonical DCC system consists of one server storing $N$ files $1$0, each of size $1$1 bits, and $1$2 users connected through a single error-free shared broadcast link, where user $1$3 has a cache of size $1$4 bits (Maddah-Ali et al., 2013, Karat et al., 2019). Placement occurs without knowledge of future demands. In the standard decentralized placement, each user independently caches a subset of $1$5 bits of each file, chosen uniformly at random; equivalently, each bit of any file is cached by any given user with probability $1$6, independently across users and bits (Karat et al., 2019).

This random uncoded placement partitions each file into subfiles indexed by user subsets. In the notation of (Karat et al., 2019), file $1$7 is partitioned into $1$8 disjoint subfiles $1$9, where $3/2$0 consists of the bits cached precisely by the users in $3/2$1. For large $3/2$2,

$3/2$3

The same law-of-large-numbers sizing appears throughout later decentralized variants, including asynchronous fog architectures and shared-cache models (Jiang et al., 2019, Peter et al., 2021).

The canonical delivery phase is MAN-style XOR multicasting. For each subset $3/2$4 with $3/2$5, the server sends

$3/2$6

where $3/2$7 denotes the portion of the file demanded by user $3/2$8 that is stored exclusively in the caches of users in $3/2$9 (Karat et al., 2019). Because each user in $1$0 already stores all XOR terms except its own, the transmission is simultaneously useful to all users in the subset. The original decentralized formulation also includes an alternative delivery by enough random linear combinations of each requested file, and the overall scheme uses the better of the two procedures (Maddah-Ali et al., 2013).

One common misconception is that decentralized placement destroys the coded-multicasting structure that makes coded caching effective. The opposite is true in the original theory: even without coordination, the random overlap structure induces enough subfiles of the right exclusivity pattern to support the same XOR logic as centralized MAN delivery (Maddah-Ali et al., 2013, Karat et al., 2019).

3. Rate expressions, optimality theorems, and converse structure

For the standard decentralized scheme, the achieved worst-case rate for $1$1 is

$1$2

normalized by file size $1$3 (Karat et al., 2019). In the original order-optimal form, the overall achievable decentralized rate is

$1$4

which recovers the above expression in the main regime of interest (Maddah-Ali et al., 2013).

The central converse mechanism is index coding. For $1$5 and worst-case distinct demands, the induced index-coding instance satisfies

$1$6

so the Ali–Niesen decentralized delivery matches the minrank lower bound and is optimal for decentralized uncoded placement (Karat et al., 2019). The same $1$7 mechanism later reappears in decentralized shared-cache networks and in error-correcting overlays (Peter et al., 2021).

The relation to centralized coded caching is also quantitatively sharp. Let $1$8. Then

$1$9

and

$3/2$0

so decentralized coded caching is never better than centralized coded caching, never worse by more than a factor $3/2$1, and becomes asymptotically indistinguishable as $3/2$2 grows (Yan et al., 2016).

These converse results clarify the status of DCC. The phrase “order-optimal” in the original work refers to a universal constant-factor approximation to the optimal memory–rate tradeoff (Maddah-Ali et al., 2013). Later work then established exact optimality for the main uncoded decentralized model with $3/2$3 (Karat et al., 2019), and a tight comparison with centralized MAN delivery (Yan et al., 2016).

4. Placement and delivery variants beyond the canonical scheme

A large branch of the literature modifies either placement or delivery while remaining within the decentralized ethos. The most direct motivation is finite file size. The paper "New Order-Optimal Decentralized Coded Caching Schemes with Good Performance in the Finite File Size Regime" proposes a decentralized random coded caching scheme and a partially decentralized sequential coded caching scheme, shows that the sequential scheme outperforms the random scheme in the finite file size regime, and proves that both attain the same memory–load tradeoff as Maddah-Ali–Niesen as file size goes to infinity (Jin et al., 2016). In a different finite-subpacketization direction, "Decentralized Coded Caching Without File Splitting" replaces coded subfile caching by coded file caching with online clique cover or matching delivery, derives expected-rate approximations by the differential equations method, and reports that coded file caching is significantly more effective than uncoded caching in reducing the delivery rate (Saberali et al., 2017).

Another major line introduces coded prefetching. "Novel Decentralized Coded Caching through Coded Prefetching" stores random portions of an MDS-coded version of each file in user caches and uses the reconstruction property of MDS codes to reduce transmissions that are useful only for a small subset of users (Wei et al., 2018). In the illustrative case $3/2$4, $3/2$5, $3/2$6, the uncoded decentralized rate $3/2$7 is reduced to $3/2$8, $3/2$9, and $1$0 for $1$1, $1$2, and $1$3 MDS codes, respectively (Wei et al., 2018). More recently, "On Coded Caching Systems with Decentralized Linear Coding Placement" studies decentralized random linear coding placement, derives achievable and converse bounds, and shows that the bounds meet under certain conditions; for $1$4, the scheme is exactly optimal in its placement class for all $1$5 (Ma et al., 29 Apr 2026).

A different strategy is to reduce the subpacketization cost by translating centralized constructions. "From Centralized to Decentralized Coded Caching" gives a generic translation from any centralized constant-rate coded caching scheme to a decentralized Type B scheme with target coding gain $1$6, rate

$1$7

with high probability, and subpacketization subexponential in $1$8 (Chen et al., 2018). This result does not preserve the original i.i.d. uncoded placement of Ali–Niesen; instead, it uses power-of-two choices over a set of virtual caches. A plausible implication is that the term “decentralized” in later work denotes a family of placement models rather than a single algorithmic template.

The following table organizes representative variants.

Variant	Core modification	Representative papers
Finite-file-size DCC	Random and sequential decentralized coded caching	(Jin et al., 2016)
Coded prefetching	MDS-coded file expansion before random placement	(Wei et al., 2018)
No file splitting	Whole-file placement with clique-cover or matching delivery	(Saberali et al., 2017)
Translated Type B placement	Virtual caches and power-of-two choices	(Chen et al., 2018)
Linear-coded placement	Random linear coding symbols per file	(Ma et al., 29 Apr 2026)

These variants do not overturn the canonical theory; they modify its implementation constraints. Finite-$1$9 work targets variance and subpacketization (Jin et al., 2016, Saberali et al., 2017), coded-prefetching work changes the algebra of side information (Wei et al., 2018, Ma et al., 29 Apr 2026), and translation-based work aims at target coding gain with improved subpacketization scaling (Chen et al., 2018).

5. Heterogeneity, shared caches, and generalized access structures

The canonical DCC model assumes equal caches, equal file sizes, and direct one-cache-per-user access. A substantial later literature relaxes each of these assumptions. For heterogeneous user memories, "Decentralized Coded Caching with Distinct Cache Capacities" and "Decentralized Caching and Coded Delivery with Distinct Cache Capacities" study decentralized placement with user-specific cache fractions $N \ge K$0, define subfiles $N \ge K$1 by exclusive cache ownership, and design group-aware deliveries that improve the required rate when $N \ge K$2 (Amiri et al., 2016, Amiri et al., 2016). In the regime $N \ge K$3, the achievable rate is

$N \ge K$4

with strict improvement over the prior heterogeneous decentralized baseline (Amiri et al., 2016). The gain increases as cache capacities become more skewed (Amiri et al., 2016, Amiri et al., 2016).

For arbitrary file sizes, arbitrary cache sizes, and arbitrary popularity, "Optimization-based Decentralized Coded Caching for Files and Caches with Arbitrary Size" introduces a general caching parameter $N \ge K$5, exact nondifferentiable load formulas, nonconvex worst-case and average-load minimization problems, Complementary GP–based iterative algorithms for stationary points, low-complexity soft-max approximations, and information-theoretic converse bounds (Wang et al., 2019). This paper is significant because it moves DCC from closed-form symmetric models to optimization over heterogeneous instances.

Shared-cache and multi-access models generalize the access structure itself. "Decentralized and Online Coded Caching with Shared Caches" considers $N \ge K$6 helper caches, each serving multiple users, derives the optimal worst-case delivery time for any association profile under uncoded decentralized placement, improves delivery for redundant demands, and provides optimal linear error-correcting delivery (Peter et al., 2021). "An Optimal Decentralized Multi-access Coded Caching System" further allows each user to access $N \ge K$7 caches. It derives a closed-form per-user delivery rate, proves linear-optimality for $N \ge K$8 by index-coding arguments, and recovers decentralized shared caching and conventional decentralized caching as special cases (Dutta et al., 2022).

These generalizations show that DCC is not tied to the one-user–one-cache abstraction. The same design logic—uncoded decentralized placement followed by index-coded XOR delivery—extends to heterogeneous memories (Amiri et al., 2016, Amiri et al., 2016), arbitrary file-size and popularity profiles (Wang et al., 2019), shared caches (Peter et al., 2021), and multi-access cache connectivity (Dutta et al., 2022).

6. Wireless, asynchronous, secure, and system-level formulations

DCC has been adapted to network settings in which the relevant performance metric is not only load but also delay, fronthaul cost, throughput, or confidentiality. In fog radio access networks, "Decentralized coded caching in wireless networks: trade-off between storage and latency" studies decentralized placement at both edge nodes and users, and a coded delivery scheme combining multicast XORs, zero-forcing, interference alignment, and fronthaul-assisted soft transfer, with performance measured by normalized delivery time (NDT) (Girgis et al., 2017). For the special case with caches only at the edge nodes, the decentralized scheme is approximately optimal, and the pipelined NDT is optimal for several parameter regimes (Girgis et al., 2017).

A different F-RAN problem is asynchrony. "Decentralized Asynchronous Coded Caching Design and Performance Analysis in Fog Radio Access Networks" studies online, time-slotted arrivals with deadlines, introduces the encoding set collapsing rule and encoding set partition method, and proves that the asynchronous load satisfies

$N \ge K$9

hence

$\alpha=\kappa$0

The scheme recovers synchronous decentralized MAN delivery when $\alpha=\kappa$1 (Jiang et al., 2019). This directly formalizes the load–delay tradeoff in asynchronous DCC.

In user-cooperative broadcast networks, "Coded Caching for Broadcast Networks with User Cooperation" proposes decentralized caching with parallel server and user transmissions, defines cooperation gain and parallel gain, and proves order-optimality when each user's cache size is larger than the threshold

$\alpha=\kappa$2

which approaches $\alpha=\kappa$3 as $\alpha=\kappa$4 (Chen et al., 2020). The paper also shows that always letting more users parallelly send information could cause high transmission delay (Chen et al., 2020).

Security and D2D delivery add another axis. "Device-to-Device Secure Coded Caching" develops a decentralized secure D2D scheme based on non-perfect secret sharing and one-time pad keying, requiring only a lower bound $\alpha=\kappa$5 on the active user count during placement, and guaranteeing both secure caching and secure delivery (Zewail et al., 2018). In a distinct D2D/cellular model, "Throughput Analysis of Decentralized Coded Content Caching in Cellular Networks" studies decentralized coded content placement by random GF(2) linear combinations and proves a throughput gain of order $\alpha=\kappa$6 under Zipf requests relative to decentralized uncoded caching (Kiskani et al., 2016).

Recent system-oriented work carries DCC into content-centric architectures. "Improving performance of content-centric networks via decentralized coded caching for multi-level popularity and access" integrates DCC with Content-Centric Networking, uses multi-level popularity and differentiated access, introduces color-based FIFO queues aligned with access privileges, and enables in-network recoding of uncoded data (Miraftab et al., 2 Aug 2025). The stated effect is elimination of queue-search overhead and improved throughput, delay, and cache-hit behavior relative to conventional CCN implementations (Miraftab et al., 2 Aug 2025).

Across these settings, the mathematical core of DCC remains recognizable: decentralized placement generates overlap; delivery exploits it through coded multicast. What changes is the surrounding optimization objective—NDT in wireless fog networks (Girgis et al., 2017), fronthaul load under deadlines (Jiang et al., 2019), throughput in D2D cellular systems (Kiskani et al., 2016), secure delivery in D2D (Zewail et al., 2018), or CCN queueing and recoding behavior (Miraftab et al., 2 Aug 2025).

7. Limitations, misconceptions, and open directions

Several limitations recur across the literature. The canonical model assumes homogeneous cache sizes, uniform file popularity, worst-case one-file requests, and large file size so that subfile sizes concentrate (Maddah-Ali et al., 2013, Karat et al., 2019). Many extensions relax one of these assumptions but not all at once. Subpacketization remains a central difficulty in both classical decentralized MAN and many of its descendants (Karat et al., 2019, Jin et al., 2016, Chen et al., 2018). Finite-file-size work was motivated precisely because the asymptotic gains of decentralized random placement can degrade substantially when $\alpha=\kappa$7 is limited (Jin et al., 2016).

A second persistent issue is analytical tractability under heterogeneity and online randomness. In asynchronous F-RANs, the exact fronthaul load for random $\alpha=\kappa$8 is difficult, and the paper provides bounds rather than exact closed forms (Jiang et al., 2019). In optimization-based generalized DCC, the exact objectives are nonconvex and nondifferentiable, which necessitates Complementary GP approximations and soft-max relaxations (Wang et al., 2019). In secure and shared-cache settings, exact optimality is usually proved under specific structural assumptions such as uncoded placement, linear delivery, or worst-case distinct demands (Zewail et al., 2018, Peter et al., 2021, Dutta et al., 2022).

A third misconception is that decentralized means purely i.i.d. uncoded placement and nothing else. The literature now includes uncoded random placement (Maddah-Ali et al., 2013, Karat et al., 2019), MDS-coded prefetching (Wei et al., 2018), random linear coding placement (Ma et al., 29 Apr 2026), grouping-based or Type B decentralized constructions (Chen et al., 2018), and group-based secure D2D placement with pre-placed keys (Zewail et al., 2018). This suggests that “decentralized” denotes the absence of global placement coordination rather than a unique stochastic law.

Open directions explicitly identified in the cited work include nonuniform popularity, heterogeneous cache and file sizes, multi-request users, improved subpacketization tradeoffs, adaptive partitioning for stochastic arrivals (Jiang et al., 2019), broader optimality proofs via generalized independence numbers (Karat et al., 2019), tighter lower bounds on subpacketization versus coding gain (Chen et al., 2018), extensions to coded placement under heterogeneity (Amiri et al., 2016), multi-access regimes beyond the currently solved cases (Dutta et al., 2022), and demand privacy or security overlays beyond the one-time-pad constructions already studied (Zewail et al., 2018, Miraftab et al., 2 Aug 2025).

In that sense, DCC has evolved from a single decentralized counterpart of MAN caching into a broad theory of uncoordinated cache placement plus coded delivery. Its most stable facts are the canonical uncoded-placement rate formula (Maddah-Ali et al., 2013, Karat et al., 2019), the exact optimality result for $\alpha=\kappa$9 under decentralized uncoded placement (Karat et al., 2019), and the sharp comparison to centralized coded caching (Yan et al., 2016). Its most active frontier is the systematic redesign of placement and delivery under finite-$N \ge K$0, heterogeneity, security, online operation, and network-specific objectives.