Multilevel Diversity Coding (MDC)

• Multilevel Diversity Coding (MDC) is an information-theoretic framework that mandates ordered reconstruction, where decoders recover progressively larger subsets of independent messages based on their access to encoder outputs.
• It leverages techniques such as superposition coding, multilayer Slepian–Wolf strategies, and entropy inequalities to precisely characterize symmetric rate regions and ensure efficient distributed storage and exact repair.
• MDC further extends to secure and sliding secret-sharing variants, integrating methods from network coding and combinatorial optimization to address secrecy constraints and multilevel data protection.

Multilevel Diversity Coding (MDC) is an information-theoretic framework for coding multiple independent messages or source components with ordered decoding requirements: a decoder that accesses more descriptions, encoders, or storage nodes must reconstruct a larger prefix of the information or a higher diversity level of the stored data. In the classical symmetric formulation, there are $L$ encoders and $L$ independent sources, and any subset $U\subseteq\{1,\dots,L\}$ with $|U|=\alpha$ must reconstruct $(S_1^n,\dots,S_\alpha^n)$; in storage-oriented formulations, the requirement is often stated as “message $M_k$ is recoverable from any $k$ nodes.” This common multilevel structure has generated a large literature spanning symmetric source coding, distributed source coding, distributed storage with exact repair, and secrecy-constrained variants (Jiang et al., 2011, Tian et al., 2015, Xiao et al., 2015).

1. Core formulations and modeling conventions

The classical symmetrical multilevel diversity coding model has $L$ independent discrete memoryless sources $(S_1,\dots,S_L)$, $L$ encoders, and a decoder that may observe an arbitrary nonempty subset $L$0 of encoder outputs. Symmetry means that the reconstruction requirement depends only on $L$1: if $L$2, the decoder must reconstruct the first $L$3 sources. A rate tuple $L$4 is admissible if there exist blocklength-$L$5 encoders with $L$6 and vanishing reconstruction error for every nonempty $L$7 (Jiang et al., 2011).

A storage-oriented convention, used in multilevel diversity coding with regeneration, places $L$8 independent messages $L$9 across $U\subseteq\{1,\dots,L\}$0 storage nodes and requires that either the first $U\subseteq\{1,\dots,L\}$1 messages be reconstructible from any $U\subseteq\{1,\dots,L\}$2 nodes, or, equivalently in a different indexing convention, that message $U\subseteq\{1,\dots,L\}$3 itself be recoverable from any $U\subseteq\{1,\dots,L\}$4 nodes. Formally, one sees statements such as

$U\subseteq\{1,\dots,L\}$5

or

$U\subseteq\{1,\dots,L\}$6

depending on the paper’s notation (Tian et al., 2015, Shao et al., 2015).

A further generalization is distributed multilevel diversity coding, where there are $U\subseteq\{1,\dots,L\}$7 vector sources $U\subseteq\{1,\dots,L\}$8, one at each encoder, with arbitrary correlation within each layer and independence across layers. For any nonempty subset $U\subseteq\{1,\dots,L\}$9, the decoder using encoder outputs from $|U|=\alpha$0 must reconstruct the first $|U|=\alpha$1 components of each corresponding source: $|U|=\alpha$2 This extends MDC from centralized coding of a common collection of sources to a distributed Slepian–Wolf-type setting (Xiao et al., 2015).

2. Symmetric rate regions, superposition coding, and entropy inequalities

For classical SMDC, the canonical achievability method is superposition coding: each source $|U|=\alpha$3 is encoded separately, typically by an $|U|=\alpha$4 MDS-like code, and the encoder output is the superposition of the per-source codewords. Yeung and Zhang showed that superposition coding achieves the entire admissible rate region of symmetric MDC; Roche, Yeung, and Hau had already established minimum-sum-rate optimality. In particular, the minimum sum rate is

$|U|=\alpha$5

and the full region can be written through supporting hyperplanes indexed by $|U|=\alpha$6 (Jiang et al., 2011).

A standard implicit characterization is

$|U|=\alpha$7

where each $|U|=\alpha$8 is the optimum of a linear program over $|U|=\alpha$9-resolutions. Guo and Yeung made this characterization explicit: for ordered $(S_1^n,\dots,S_\alpha^n)$0, they showed

$(S_1^n,\dots,S_\alpha^n)$1

identified a finite subset of inequalities sufficient for the coding rate region, and proved that only a much smaller subset needs to be checked when determining achievability of an ordered rate tuple, although the cardinality of that smaller set grows at least exponentially fast with $(S_1^n,\dots,S_\alpha^n)$2 (Guo et al., 2018).

The converse theory for SMDC is built around subset entropy inequalities. Han’s subset entropy inequality underlies the minimum-sum-rate converse, while later work introduced a sliding-window subset entropy inequality and used the Madiman–Tetali subset entropy inequality to clarify the structure of the Yeung–Zhang converse. These tools also support extensions such as SMDC with an all-access encoder and secure symmetrical multilevel diversity coding (Jiang et al., 2011).

3. Distributed and storage-centric extensions

Distributed multilevel diversity coding replaces the centralized encoder view by $(S_1^n,\dots,S_\alpha^n)$3 distributed encoders observing correlated vector sources. The paper on distributed MDC proposes a multilayer Slepian–Wolf architecture: for each layer $(S_1^n,\dots,S_\alpha^n)$4, encoder $(S_1^n,\dots,S_\alpha^n)$5 sends a bin index of rate $(S_1^n,\dots,S_\alpha^n)$6, with intralayer constraints

$(S_1^n,\dots,S_\alpha^n)$7

for appropriate $(S_1^n,\dots,S_\alpha^n)$8, and total rates

$(S_1^n,\dots,S_\alpha^n)$9

This multilayer Slepian–Wolf scheme is optimal for $M_k$0, and for general $M_k$1 under the symmetrical entropy-wise condition $M_k$2 whenever $M_k$3 (Xiao et al., 2015).

In distributed storage, the corresponding model is multilevel diversity coding with regeneration, also called MLDR or MLD-R. Here there are $M_k$4 storage nodes, messages $M_k$5, per-node storage $M_k$6, and exact repair from any $M_k$7 helpers each sending at most $M_k$8. The constraints are

$M_k$9

and

$k$0

The normalized tradeoff between storage and repair bandwidth is the analogue of the regenerating-code tradeoff, but now aggregated across multiple reconstruction levels (Shao et al., 2015).

For MLDR, two outer bounds were derived: $k$1 where $k$2 and $k$3. Their intersection gives the MBR point

$k$4

and separate coding of the messages with per-level regenerating codes achieves this point (Shao et al., 2015).

The full tradeoff region can nevertheless exhibit genuine multilevel coupling. For the $k$5 case, Tian and Liu characterized the complete tradeoff region and showed that mixed coding can strictly improve on separate coding in the interior. For $k$6, they exhibited a mixed code achieving

$k$7

which is not achievable by any separate-coding scheme for that message profile (Tian et al., 2015).

4. Secrecy-constrained MDC

Secrecy has been incorporated into both source-coding and storage versions of MDC. In secure symmetrical multilevel diversity coding, there are $k$8 independent sources, $k$9 encoders, and an eavesdropper that may access any subset $L$0 with $L$1. The legitimate receiver must reconstruct $L$2 from any $L$3 with $L$4, while perfect secrecy requires

$L$5

For the corresponding secure single-level problem, the admissible region is characterized exactly, and for general secure SMDC, superposition coding remains optimal for the minimum sum rate (Balasubramanian et al., 2012).

Weakly secure SMDC imposes per-message security constraints. Each message $L$6 must satisfy

$L$7

with $L$8. The complete characterization of when superposition coding is sum-rate optimal is unusually sharp: for every pair $L$9 with $(S_1,\dots,S_L)$0 and $(S_1,\dots,S_L)$1, one must have either $(S_1,\dots,S_L)$2, or $(S_1,\dots,S_L)$3. When this condition fails, two joint coding strategies yield rate savings, using some coding components for one message as the encryption key for another (Guo et al., 2020).

Sliding secure SMDC imposes a “security priority” indexed by a gap parameter $(S_1,\dots,S_L)$4. Classical reconstruction remains unchanged, but for each source $(S_1,\dots,S_L)$5 with $(S_1,\dots,S_L)$6, perfect secrecy is required whenever no more than $(S_1,\dots,S_L)$7 encoders are accessible: $(S_1,\dots,S_L)$8 When the first $(S_1,\dots,S_L)$9 sources are constants, the model becomes $L$0 multilevel secret sharing. For $L$1, superposition coding is optimal. For the $L$2 problems, the rate regions are characterized, and superposition coding is suboptimal for both the sliding secure SMDC and multilevel secret sharing cases. The main mechanism behind rate reduction is that the previous source $L$3 can be used as the secret key of $L$4 (Guo et al., 2024).

In distributed storage, secrecy enters through repair observations. MDC with secure regeneration (MDC-SR) combines multilevel reconstruction, exact repair, and information-theoretic secrecy against an eavesdropper that observes repair data. In the baseline model, for any $L$5-subset $L$6 of nodes,

$L$7

Two outer bounds show that separate coding of the messages using secure regenerating codes achieves the MBR point of the normalized storage-capacity/repair-bandwidth tradeoff region (Shao et al., 2017). A later generalization distinguishes type 1 compromised nodes, where the eavesdropper sees only stored contents $L$8, from type 2 nodes, where it sees full repair data $L$9. The MBR point remains unchanged as long as $L$00, so the MBR point depends only on the total number $L$01 of compromised nodes in that regime (Shao et al., 2018).

5. Optimality landscape: superposition, separate coding, mixed coding, and network coding

The coding-theoretic picture across MDC is highly structured. In classical SMDC, superposition coding is optimal for both the minimum sum rate and the entire admissible rate region; this is one of the rare multilevel source-coding settings where full source separation is exactly optimal (Jiang et al., 2011). In distributed multilevel diversity coding with correlated sources, multilayer Slepian–Wolf coding is optimal for $L$02 and under a symmetry condition for general $L$03, again giving a layered optimal architecture (Xiao et al., 2015).

In storage with exact repair, separate coding is optimal at the extremes but not generally in the interior. Tian and Liu showed that separate coding achieves the minimum-storage point for general MDCR and that it is fully optimal for $L$04; they also proved, for $L$05, that mixed coding can strictly enlarge the tradeoff region when $L$06 (Tian et al., 2015). Shao, Liu, and Tian later established that separate coding also achieves the MBR point for general MLDR, resolving the question left open by Tian and Liu (Shao et al., 2015).

The secure variants preserve this “extremes versus interior” theme but with different mechanisms. For weakly secure SMDC, joint coding can strictly outperform superposition when the per-message secrecy thresholds permit one message to serve as a key for another; the paper gives two such pairwise strategies (Guo et al., 2020). For sliding secure SMDC, the previous source $L$07 can serve as the secret key for $L$08, and the minimum sum rate of the general $L$09 multilevel secret sharing problem is achieved by such a joint scheme; for the general sliding secure SMDC problem, superposition coding of the $L$10 sets $L$11, $L$12, $L$13, $L$14, $L$15 achieves the minimum sum rate (Guo et al., 2024).

For MDC-SR, the principal extremal result is again at MBR. The normalized MBR point is

$L$16

and separate coding with per-level secure regenerating codes achieves it (Shao et al., 2017). In the generalized compromised-node model, the same MBR point persists when $L$17, despite the weaker eavesdropper (Shao et al., 2018).

A recent application-oriented development studies two-level priority coding over three edge-disjoint paths under arbitrary blockage patterns. There, the complete capacity region is characterized for two priority levels, superposition coding achieves the region in general, and network coding is required only in a specific corner case. The analysis is unified by showing equivalence to designing encoding schemes over combination networks (Dogan et al., 23 Aug 2025).

6. Computational, combinatorial, and application perspectives

Beyond the classical symmetric models, multilevel diversity coding systems (MDCS) form a broader class of multi-source multi-sink networks with prioritized source prefixes as decoder demands. For MDCS, rate regions can be written in entropy-space form using source independence, encoder determinism, decoder determinism, and rate constraints, and then approximated by the Shannon outer bound together with inner bounds from scalar and vector linear codes and from superposition coding (Li et al., 2014).

A large computational study enumerated all non-isomorphic MDCS instances up to several parameter sizes, computed Shannon outer bounds and linear-code inner bounds, and proved exact rate regions for thousands of instances when the bounds matched. These computations produced several structural conclusions: scalar binary codes suffice for many instances, vector binary codes close many remaining gaps, and superposition coding is sufficient in some but not all asymmetric instances. The same work also described how to generate computer-aided human-readable converse proofs and explicit achievability codes (Li et al., 2014).

The minor-theoretic viewpoint is especially notable. Using source deletion, encoder deletion, encoder contraction, and encoder unification, the paper identifies embedding operations that preserve the sufficiency properties of scalar and vector codes. From thousands of instances for which binary scalar codes are insufficient, it extracts 12 forbidden smallest embedded MDCS instances, giving a combinatorial explanation for when scalar binary linear codes fail (Li et al., 2014).

On the explicitly computable side of symmetric MDC, the finite characterization of the SMDC rate region has practical algorithmic significance. The original description involved uncountably many inequalities and constant terms defined through linear optimization problems; the explicit formulation reduces achievability testing to a finite family of inequalities, although the smaller symmetry-reduced set still grows at least exponentially fast with $L$18 (Guo et al., 2018).

At the application level, recent two-level priority coding for resilience to arbitrary blockage patterns uses MDC to control the received information and offer distinct reliability guarantees based on stream priority while keeping design and operational complexity low as the number of network paths increases. The problem of constructing high-performing MDC schemes is shown there to be equivalent to designing encoding schemes over combination networks, linking MDC directly to modern network-coding analyses for ultra-reliable low-latency communication (Dogan et al., 23 Aug 2025).