Combination Networks: Theory and Applications

• Combination networks are defined as two-hop relay multicast topologies where a server connects to H relays that serve binomial(H, r) users via orthogonal, error-free links.
• They enable rigorous analysis of multicast capacity, minimality, and the distinctions between scalar-linear and vector-linear solutions, with implications for network coding.
• In coded caching, combination networks drive topology-aware placement strategies and low-subpacketization designs that exploit relay-user incidence for efficient message delivery.

Combination networks are a class of two-hop relay multicast topologies studied in network coding and coded caching. In the caching formulation, a server with a library of $N$ files is connected to $H$ relays, and $K=\binom{H}{r}$ users are each connected to a distinct subset of $r$ relays through orthogonal, error-free links; in the network-coding formulation, a generalized combination network $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$ has $h$ source messages, $r$ middle-layer nodes, $\ell$ parallel links on source–middle and middle–receiver edges, $\alpha$ middle nodes per receiver, and $\varepsilon$ direct links from the source to each receiver (Wan et al., 2017, Liu et al., 2020). Across these formulations, the topology is used to study multicast capacity, minimality, scalar-linear versus vector-linear solvability, memory–load tradeoffs, secrecy, interference management, and low-subpacketization code design (Bidokhti et al., 2014).

1. Canonical models and parameterizations

A standard $H$0 combination network consists of a single server, $H$1 relays, and $H$2 users, where each user is connected to a unique set of $H$3 relays, and the links are orthogonal, non-interfering, and error-free (Yan et al., 2018). In several coded-caching papers, the relays have no caches and each user has a cache of size $H$4 files; in other variants, both relays and users have caches, or the relay–user hop is replaced by a partially connected wireless interference network (Zewail et al., 2017, Roushdy et al., 2018).

In network coding, the combination network is a highly structured multicast network. A multicast network is minimal if deleting any edge destroys solvability, and the combination network $H$5 is minimal only for $H$6, in which case it is the minimal combination network (Cai et al., 2019). Generalized combination networks extend this structure by allowing $H$7 parallel links and $H$8 direct source–receiver links; the nontrivial regime emphasized in the scalar/vector gap literature is $H$9 (Liu et al., 2020).

A recurrent structural distinction is between ordinary and resolvable combination networks. In the resolvable case, $K=\binom{H}{r}$0, the users can be partitioned into parallel classes, which is useful for placement delivery array constructions and transformations from shared-link coded-caching schemes (Yan et al., 2018).

2. Multicast coding, minimality, and capacity regions

Minimal multicast networks are extremal objects in the sense that every edge is critical for solvability, and minimal combination networks provide a particularly clean test case for linear network coding (Cai et al., 2019). For minimal multicast networks, every non-source node must have in-degree at least $K=\binom{H}{r}$1, and if a minimal multicast network admits a $K=\binom{H}{r}$2-vector solution, then a corresponding $K=\binom{H}{r}$3 solution exists for the $K=\binom{H}{r}$4-parallelization of the network (Cai et al., 2019).

A separate line of work studies multicasting nested message sets over combination networks. The source transmits a common message $K=\binom{H}{r}$5 and a private message $K=\binom{H}{r}$6 to public and private receivers, respectively. Standard linear superposition coding is optimal for networks with two public receivers and any number of private receivers, but it stops being optimal when the number of public receivers increases. Two refinements—pre-encoding at the source and block Markov encoding—yield inner bounds that are capacity for combination networks with three or fewer public receivers and any number of private receivers; the block Markov scheme may strictly include the pre-encoding/linear-superposition region (Bidokhti et al., 2014).

The converse methodology in this setting relies on sub-modularity of the entropy function. The same paper introduces an equivalent graphical representation and a lemma described as being of independent interest, and it extends the block Markov idea to broadcast channels with two nested messages (Bidokhti et al., 2014).

3. Scalar-linear and vector-linear solvability

The scalar/vector comparison is one of the most technically developed aspects of combination networks. In the minimal-network literature, $K=\binom{H}{r}$7 denotes the smallest field size for a scalar-linear solution and $K=\binom{H}{r}$8 the smallest alphabet size $K=\binom{H}{r}$9 for a vector-linear solution; in the generalized-network literature, the gap is often written as

$r$0

These are two different conventions for quantifying the same qualitative question: how much alphabet-size reduction vector coding can provide (Cai et al., 2019, Liu et al., 2020).

For minimal networks with two source messages, the largest possible scalar/vector gap is attained by sub-networks of the combination network called Kneser networks. For $r$1,

$r$2

and for the full minimal combination network $r$3,

$r$4

The same work shows that, for $r$5, a $r$6-linear solution exists if and only if $r$7, and that

$r$8

for $r$9 or $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$0 (Cai et al., 2019). This directly contradicts the common intuition that the full minimal combination network should exhibit the largest vector advantage; the largest known gaps arise in carefully chosen sub-networks rather than in the full minimal structure.

For generalized combination networks, new upper and lower bounds are stated in terms of the maximum number $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$1 of middle-layer nodes that admit a $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$2-linear solution. If $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$3, then

$$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$4

and for $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$5 a stronger upper bound is obtained (Liu et al., 2020). Lower bounds come from two different techniques: a probabilistic construction based on the Lovász Local Lemma, and an algebraic construction via covering Grassmannian codes (Liu et al., 2020).

These bounds lead to asymptotic statements on the alphabet-size gap. With all network parameters fixed except $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$6, the gap satisfies $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$7 as $$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$8 (Liu et al., 2020). A later version states that the upper and lower bounds together give

$$(\varepsilon,\ell)-\mathcal{N}_{h,r,\alpha\ell+\varepsilon}$$9

showing logarithmic growth in the number of middle-layer nodes (Liu et al., 2020).

4. Combination networks as coded-caching topologies

The coded-caching literature treats the combination network as a practically motivated relay architecture. Under uncoded placement, once the cache contents and the user demands are known, the delivery problem reduces to a general index coding problem. For combination networks with end-user caches, the acyclic index coding converse bound is not tight; a tighter converse that leverages the network topology is derived, together with an inequality that generalizes sub-modularity of entropy. The same work proposes several achievable schemes—DIS, IES, CICS, and ICICS—and proves order optimality or exact optimality in some $h$0 regimes under uncoded cache placement (Wan et al., 2017).

Several later schemes explicitly move beyond topology-agnostic delivery. The Separate Relay Decoding delivery Scheme (SRDS) directly leverages the combination-network topology when generating multicast messages, rather than first generating shared-link MAN messages and only then routing them through the relays. SRDS is extended to decentralized combination networks, more general relay networks, and combination networks with cache-aided relays and users, and is reported to reduce the normalized download time relative to previously available schemes (Wan et al., 2017).

Another delivery-side development keeps MAN placement and multicast-message generation independent of topology but introduces a two-phase delivery that creates additional side information in a first phase and exploits it in a second phase. In that setting, the download time is shown to be proportional to $h$1, and the scheme is order optimal under uncoded placement for some parameter regimes (Wan et al., 2018).

A distinct line of work argues that the placement phase itself should depend on topology. In $h$2 combination networks with end-user caches, asymmetric coded placement tailored to relay–user incidence improves over symmetric, topology-independent schemes. For a coded caching gain $h$3, the proposed scheme attains memory–load points $h$4 and matches the cut-set lower bound

$h$5

when

$h$6

thereby proving information-theoretic optimality for large cache sizes (Wan et al., 2018). Related asymmetric constructions based on relay coordination likewise show that restricting subfile placement to user groups that can actually be served together reduces redundancy relative to earlier symmetric coded placement (Wan et al., 2018).

5. Relay caches, secrecy constraints, interference, and random topology

Combination networks with caches at both relays and end users add another layer of side information. A centralized coded-caching scheme based on $h$7 MDS coding jointly optimizes placement and delivery and decomposes the network into virtual multicast sub-networks. If the aggregate memory visible to a user through its own cache and its $h$8 attached relays satisfies

$h$9

then

$r$0

so the server can be completely disengaged in the delivery phase (Zewail et al., 2017). The same model is extended to secure delivery, secure caching, and the joint secure-delivery/secure-caching setting, using one-time pads for secure delivery and non-perfect secret sharing for secure caching; the second hop is shown to meet the cut-set lower bound (Zewail et al., 2017).

A related formulation further exploits relay caches as side information during server transmission rather than treating the server load as independent of the cached contents of relays. The claimed effect is a larger coded-caching gain on the server-to-relay hop and a strict improvement over earlier state-of-the-art schemes (Wan et al., 2018). This reinforces a broader point in the literature: relay caches are not merely local storage; they can alter the structure of the coded multicasts themselves.

In radio-access combination networks, the relay layer becomes a partially connected interference network with fronthaul constraints. The performance metric is the normalized delivery time (NDT), and three schemes are analyzed: MDS-IA, soft-transfer, and zero-forcing. MDS-IA is emphasized for low fronthaul capacity, soft-transfer becomes favorable as the fronthaul capacity increases, and ZF is feasible when $r$1, in which case the cloud is silent during delivery (Roushdy et al., 2018).

The strict $r$2 incidence pattern has also been relaxed. For two-hop relay networks in which each user connects to a random subset of relays, MAN multicast packets can be routed through the network according to a linear program that minimizes worst-case delivery time, and a dynamic algorithm is proposed to reduce computational complexity while approximating the LP solution (Bayat et al., 2019). This suggests that combination-network methods function not only as exact symmetric constructions but also as templates for more general combination-type relay topologies.

6. Low-subpacketization constructions and extended incidence models

A major practical limitation of coded caching over combination networks is subpacketization. Placement delivery arrays (PDAs), introduced for the shared-link model by Yan et al., were generalized to combinational PDAs (C-PDAs), whose ordinary symbols are constrained so that the intended users share a common relay. Given a $r$3 C-PDA, the induced scheme has

$r$4

and three low-subpacketization schemes are proposed (Yan et al., 2018). Two apply to resolvable networks via a transformation from ordinary PDAs; the third applies to arbitrary $r$5 and uses

$r$6

for

$r$7

thereby achieving the cut-set lower bound for sufficiently large cache sizes (Yan et al., 2018).

Direct CPDA constructions remove some of the restrictions of earlier grouping methods. A new algorithm realizes a coded-caching scheme from a CPDA with subpacketization at most $r$8 and does not require $r$9. Two classes of CPDAs are then constructed for arbitrary positive integers $\ell$0 and $\ell$1 with $\ell$2, yielding more flexible memory ratios and substantially smaller subpacketization than previously known constructions based on grouping (Cheng et al., 2019).

The incidence pattern can also be generalized so that each $\ell$3-relay set serves $\ell$4 users rather than one. In an $\ell$5 multiaccess combination network, a trivial repetition of an $\ell$6 scheme multiplies the per-relay load by $\ell$7. An extension of the Zewail–Yener method reduces relay load but has subpacketization exponential in the number of users, which motivates a direct construction based on combinational design theory and a hybrid construction for arbitrary $\ell$8. These newer schemes trade a moderate increase in relay load for much lower subpacketization (Huang et al., 2021).

Across these strands, a stable pattern emerges. The most effective schemes are those that exploit the relay–user incidence structure directly: in multicast coding through Kneser-type sub-networks and $\ell$9-Kneser objects, in coded caching through topology-aware delivery and asymmetric placement, and in implementation-oriented designs through PDAs, CPDAs, and multiaccess generalizations.