Network Coding Solutions for

the Combination Network and its Subgraphs

Han Cai∗ , Tuvi Etzion† , Moshe Schwartz‡ , and Antonia Wachter-Zeh§
∗ Electrical & Computer Engineering, Ben-Gurion University of the Negev, Israel, [email protected]
† Computer Science, Technion – Israel Institute of Technology, Israel, [email protected]
‡ Electrical & Computer Engineering, Ben-Gurion University of the Negev, Israel, [email protected]
§ Institute for Communications Engineering, Technical University of Munich, Germany, [email protected]
arXiv:1901.01058v1 [cs.IT] 4 Jan 2019

Abstract—The combination network is one of the simplest and the starting node of the edge. When we consider the function
insightful networks in coding theory. The vector network coding on the edge as a linear combination of the h messages, the
solutions for this network and some of its sub-networks are vector of coding coefficients (for the h messages) is called the
examined. For a fixed alphabet size of a vector network coding
solution, an upper bound on the number of nodes in the network global coding vector. To recover the h messages, a receiver R
is obtained. This bound is an MDS bound for subspaces over a should obtain h global coding vectors whose linear span has
finite field. A family of sub-networks of combination networks dimension h. In other words, the h × h matrix formed by these
is defined. It is proved that for this family of networks, which h global coding vectors should be invertible. This h × h matrix
are minimal multicast networks, there is a gap in the minimum is called a transfer matrix of R. The previous description
alphabet size between vector network coding solutions and scalar
network coding solutions. This gap is obtained for any number constitutes the framework for scalar linear network coding.
of messages and is based on coloring of the q-Kneser graph and The framework for vector network coding was presented in [9].
a new hypergraph generalization for it. Each message and each packet is a vector of length t and
the coding coefficients are t × t matrices. The global coding
I. I NTRODUCTION vectors, on the edges, consist of h matrices of size t × t, which
Network coding has been attracting increased attention together form t × (ht) matrices. W.l.o.g., we assume that each
for almost two decades since the seminal papers [1], [16]. t × (ht) matrix of a global coding vector is a generator matrix
Multicast networks have received most of this attention. An of a t-subspace of F htq . To recover the h messages, a receiver
recent survey on the foundation of multicast network coding R should have on its ℓ incoming edges, ℓ > h, h such global
can be found in [13]. The multicast network-coding problem coding vectors which form together an (ℓt) × (ht) transfer
can be formulated as follows: given a network with one source matrix of rank ht.
which has h messages, for each edge find a function of the The field size of the solution is an important parameter that
packets received at the starting node of the edge, such that each directly influences the complexity of the calculations at the
receiver can recover all the messages from its received packets. network nodes. It is known that any field size q > N suffices
Such an assignment of a function to each edge is called a for a solution. However, it is conjectured that the smallest
solution for the network. Therefore, the received packets on an field size allowing a solution is much smaller [12], [13]. An
edge can be expressed as functions of the source messages. If efficient algorithm to find such a field size and the related
these functions are linear, we obtain a linear network coding network code was given in [15]. It is conjectured that the
solution, otherwise we have a nonlinear solution. In linear minimum alphabet size is much smaller, but this was proved
network coding, each linear function on an edge consists of only for two messages [12]. For this purpose we distinguish
coding coefficients for each incoming packet. If the coding between the smallest alphabet size required for each one of the
coefficients and the packets are scalars, it is called a scalar three types of network coding solutions. Given a network N ,
network coding solution. If the messages and the packets are we define qs (N ) to be the smallest field size q for which N
vectors and the coding coefficients are matrices then it is has a scalar linear solution. Similarly, qn (N ) is the smallest
called a vector network coding solution. A network which alphabet size q (q not necessarily a prime power) for which
has a solution is called a solvable network. It is well-known N has a scalar nonlinear solution, and qv (N ) is the smallest
that a multicast network with one source, h messages, and value qt , q a prime power, such that N has a vector solution
N receivers, is solvable if and only if the min-cut between the over F tq . By definition, qs (N ) > qv (N ) > qn (N ), and we
source and each receiver is at least h [13]. define the vector gap by
The functions on the edges of the network form the network
gapv (N ) , qs (N ) − qv (N ).
code. The coding coefficients form the network coding vectors
on the edges. The vector of coding coefficients is called Two other gaps (qs (N ) − qn (N ) and qv (N ) − qn (N )) are
the local coding vector when the function on the edge is defined similarly, but this paper will be mostly devoted to the
considered as a linear combination of the packets received at vector gap.
 One of the most celebrated families of networks is the networks called minimal multicast networks, for any number of
family of combination networks [22], which were used for messages. The gap is increasing with the number of messages.
various topics in network coding. The Nh,r,s combination This also proves the existence of a gap for two messages which
network, where s > h, is shown in Fig. 1. The network has was left open in [11]. The networks which will be used for this
three layers: the first layer consists of a single source with h purpose are sub-networks of the combination networks. The
messages. The source transmits r packets to the r nodes of the proof will be based on the chromatic number of the q-Kneser
middle layer. Any s nodes in the middle layer are connected graph and a generalized version of it, the q-Kneser hypergraph,
to a receiver, and each one of the (rs) receivers demands all which was not defined before. The coloring problem raises
the h messages. It was proved in [22] that a solution for such an intriguing combinatorial problem which has independent
a network exists if and only if a related error-correcting code intellectual merit. Several more related problems will be
exists. This network was also generalized to compare scalar presented in Section IV and will be considered in the full
and vector network coding [11]. Its sub-networks were used version of this paper. The same is true for some proofs of
to prove that finding the minimum required field size of a claims in the paper.
(linear or nonlinear) scalar network code for a certain multicast
II. V ECTOR S OLUTION AND B OUND FOR MDS C ODES
network is NP-complete [21].
In this section, we first describe the three types of solutions
x1 , . . . , x h for the Nh,r,s combination network. The key result is the
following theorem proved in [22]. Let (r, qh , r − s + 1)q
denote a code over F q of length r with qh codewords and
minimum Hamming distance r − s + 1. If this code is linear,
...... r nodes it is denoted by [r, h, r − s + 1]q .
s edges Theorem 1. ( [22]) The Nh,r,s combination network is solvable
over F q if and only if there exists an (r, qh , r − s + 1)q code.
(rs) receivers
In view of Theorem 1, what are the functions on the edges of
the Nh,r,s combination network in the three types of solutions?
Figure 1. The N h,r,s combination network: it has an edge from the source to
each of the r nodes in the middle layer. Each of the (rs) receivers is connected 1) For the scalar nonlinear solution, an (r, qh , r − s + 1)q
to a unique set of s middle-layer nodes, and demands all of the h source code, each coordinate in a codeword is a function of
messages.
h information symbols which are represented by the
The goal of this work is to consider two problems which are h messages. The function for the ith symbol of a
related to vector coding solutions for combination networks codeword is the function on the link from the source
and their sub-networks. In Section II, we describe network to the ith node in the middle layer.
coding solutions (vector, scalar, linear and nonlinear) for the 2) For the scalar linear solution, an [r, h, r − s + 1]q code
Nh,r,s combination network. In particular, we consider the is required. It has an r × h generator matrix and the
Nh,r,h combination network and the maximum number of h entries of its ith column are the coding coefficients of
nodes in the middle layer for such a network. This number is the linear function on the link from the source to the ith
related to the largest length of certain MDS codes. While there node in the middle layer.
exists a proof on the upper bound of such length for linear and In both cases, the nodes of the middle layer transmit their
nonlinear codes, we are not aware on any proof based on the information to the related receivers. Each receiver obtains s
properties of the subspaces. These codes are also MDS array symbols from the middle-layer nodes, each one has the same
codes which were considered in the past for storage [5] and global coding vector on its incoming and outgoing edges.
are very popular today as distributed-storage codes, e.g., see Since the minimum Hamming distance of the code is r − s + 1,
[8], [24] and references therein. In Section III, the vector gap it follows that for each two different sets of h messages, each
is considered. Such vector gaps, which are very large, were receiver obtains a different s-tuple of symbols from the middle
considered in [11] for any number of messages h > 2. The layer nodes. Hence, it can recover the h messages.
networks which were used for the proof are generalizations of For the vector network coding solution, the h matrices of
the combination networks in which for each receiver there are size t × t on the edges from the source to the middle-layer
some redundant edges on the paths between the source and the nodes form together a t × (ht) matrix which has dimension t,
receiver. The extra edges were used to distribute the (ht)-space i.e., it represents a t-subspace of F ht
q . Now, to have a solution
formed by the h vector messages of length t on more than h for the Nh,r,s combination network, each s subspaces, related
edges. This enables some edges to transmit only a fraction of to the edges between the source and the middle-layer nodes,
a one-dimensional space. However, a similar idea cannot be span the (ht)-space defined by the messages of the source.
used for scalar linear network coding. The question whether A fundamental combinatorial structure that underpins some
such gaps can be obtained if there are no such redundant edges of the generalized combination networks is a structure we call
remained open. In Section III, we give a positive answer to a (t; h, α)q -independent configuration. We use [Vt ] to denote
this question and prove that there exists a vector gap in such all the t-dimensional subspaces of a vector space V, and [ba]
to denote the Gaussian coefficient (where the field size q is that receiver has a full rank (ht) × (ht) transfer matrix from
understood from context). which it can recover the h messages.
Definition 2. Let q be a prime power, t, h, α be positive integers, Lemmas 3 and 4 form a generalization for an upper bound
α 6 h, and denote V = F ht on the length of MDS code (use α = h in Lemma 3). The
q . A ( t; h, α ) q -independent configu-
related results for (scalar) linear codes are given in [18].
ration (IC) is a set C = {V1 , V2 , . . . , Vm } ⊆ [Vt ], such that for
Corollary 7 [18, p. 321] asserts that for an [n, k, n − k + 1]q
all 1 6 i1 < i2 < · · · < iα 6 m,
MDS code, we have that n 6 q + k − 1. This result is
dim(Vi1 + Vi2 + · · · + Viα ) = αt. strengthened in Theorem 11 [18, p. 326] by using a more
complicated proof based on projective geometry. The theorem
We say |C| = m is the size of the IC. asserts that if k > 3 and q is odd then n 6 q + k − 2. A more
Lemma 3. Let C be a (t; h, α)q -IC. If α > 2 then complicated proof for the same result is given for nonlinear
codes in [20, pp. 12-13].
q ( h − α +2 ) t − 1 Lemmas 3 and 4 can be generalized for a family of net-
|C| 6 + α − 2.
qt − 1 works which generalize the combination network [11]. Some
Proof: If α = 2 the claim is immediate by considering interesting consequences implied by this generalization will
the size of a t-spread [10]. be discussed in the full version of this paper.
Assume now α > 2, and denote V , F ht q . Let us write
We can use Lemma 4 to upper bound the vector gap
C = {V1 , V2 , . . . , Vm }, and define in the Nh,r,h combination networks. For this we will use
Bertrand’s postulate (e.g., see [2]) that the interval [n, 2n]
W1 , V1 + V2 + · · · + Vα−2 , contains a prime power for any integer n; and that the interval
where dim(W1 ) = (α − 2)t. By the definition of an IC, F ht [ x, x + x21/40 ] contains a prime for all large enough x [3].
q =
This implies the following result.
W1 + W2 , where W2 ∈ [( h−αV+2) t]. It follows that any vector
v ∈ Vj , α − 1 6 j 6 m, may be written uniquely as v = Theorem 5. For all positive integers h and r, let N denote the
v1 + v2 , where v1 ∈ W1 and v2 ∈ W2 . We now define Nh,r,h combination network. Then gapv (N ) 6 r + h − 3, and
for all large enough r, gapv (N ) 6 (r − 1)21/40 + h − 2.
Vj′ , v2 : v1 + v2 ∈ Vj , v1 ∈ W1 , v2 ∈ W2 ,


III. M INIMAL M ULTICAST N ETWORKS

for all α − 1 6 j 6 m. It is easily seen that dim(Vj′ ) = t.
Furthermore, for any α − 1 6 j1 < j2 6 m, In this section we will prove that for each number of
messages h > 2, there exists a minimal multicast network
dim(W1 + Vj′1 + Vj′2 ) = αt ⇒ dim(Vj′1 + Vj′2 ) = 2t. for which vector network coding outperforms scalar network
Thus, the set Vi′ α−16i6m contains |C| − α + 2 pairwise
 coding. A minimal multicast network can deliver h messages
disjoint t-subspaces of W2 . Thus, from the source to the receivers, but if any edge is removed,
it can deliver at most h − 1 messages to at least one of
[(h−α1+2)t] the receivers. From a practical point of view, considering
|C| − α + 2 6 . such minimal networks is interesting as it minimizes the used
[1t ]
network resources. From a theoretical point of view, minimal
networks can be regarded as a fair setting for a comparison
We now make the connection between ICs and a certain between the three types of network coding solutions.
family of combination networks.
Definition 6. A multicast network N is said to be minimal if
Lemma 4. The Nh,r,h combination network has a vector solu-
every edge crosses a cut of size h.
tion over F q with messages of length t if and only if there exists
a (t; h, h)q -IC of size r. Thus, in a minimal network, the removal of any edge makes
at least one cut have size strictly less than h, and therefore the
Proof: In the first direction assume that a vector solution
new network is incapable of a solution.
over F q with messages of length t exists. We note that by
To achieve the goal of this section, a sub-network of the
construction, any node i in the middle layer has a subspace ∗ , will be used.
Nh,r,h combination network, denoted by Nh,r,h
Vi ⊆ V , F ht q , with dim(Vi ) 6 t. If the terminal R j gets
from the middle layer the subspaces Vj1 , . . . , Vjh , then The network Nh,r,h∗ has one source in the first layer, and [htt]
nodes in the middle layer, each node represents a different t-
dim(Vj1 + · · · + Vjh ) = ht, subspace of F ht
q . From each h nodes in the middle layer which
which implies that dim(Vi ) = t. Thus, {Vi }16i6r is a represent the t-subspaces Vi1 , . . . , Vih for which
(t; h, h)q -IC. Vi1 + · · · + Vih = F ht
q , (1)
In the other direction, assume C = {V1 , . . . , Vr } is a
(t; h, h)q -IC. We can easily construct a vector network coding there are links to a unique receiver.
solution to the Nh,r,h combination network. Simply send Vi to For the remainder of this work, let G ( X ) be any t × (ht)
the ith middle layer node. Since C is a (t; h, h)q -IC it follows generator matrix for a t-subspace X of F ht
q . Also, the splitting
of a t × (ht) matrix G is the h matrices of size t × t obtained Thus, for these cases
by taking the first t columns of G, then the next t columns,
qs (N ) > qt + qt−1 − 1,
and so on.
It is obvious from the definition of Nh,r,h ∗ that for vector and therefore we have a gap of at least
network coding, the minimum alphabet size for which it is
solvable is qt . The coding coefficients on the edge from gapv (N ) = qs (N ) − qv (N ) > qt−1 − 1.
the source to the node represented by the t-subspace X are
formed by splitting of G ( X ) into h matrices of size t × t. The The scope of Theorem 7 is somewhat limited due to the
global coding vector from a node u of the middle layer to a restrictions on the value of t. We can remove these restrictions,
receiver R j is the same one as the global coding vector (which but severely reduce the guaranteed vector gap to merely 1.
coincides with the local coding vector) from the source to u.
It implies by (1) that the (ht) × (ht) transfer matrix of each Theorem 8. For a prime power q and any integer t > 2, there
receiver is of full rank. It is not difficult to prove that a smaller exists a minimal network N with two messages for which
alphabet size is impossible. gapv (N ) > 1.
For the scalar solution we form a new hypergraph Proof: We will prove that χ(qK2t:t ) > qt + 1. Recall
G = (V, E), where V is the set of middle-layer vertices of that the vertex set of qK2t:t is [Vt ], where V = F2t q . Assume a
∗ . Each set of h vertices from the middle layer from which
Nh,r,h V
coloring of qK2t:t with c colors. Let Ui ⊂ [ t ], 1 6 i 6 c, be
there are links to a joint receiver R j of the third layer (i.e (1) is the set of vertices colored with color i. Then each Ui is a 1-
satisfied), are connected in G by a hyperedge. When h = 2 this intersecting family in the language of [14], and an anticode of
hypergraph is the well-known q-Kneser graph qK2t:t . Hence, diameter t − 1 in the language of [23]. Also, the set {Ui }16i6c
we will denote the general hypergraph by qKht:t h and call it forms a tiling (partition) of [Vt ].
the q-Kneser hypergraph. (This is not to be confused with the
By [14], for all 1 6 i 6 c, |Ui | 6 [2tt−−11], and Ui is either
q-Kneser graph qKht:t )
A coloring of a graph G = (V, E) is an assignment of a
       
V V
set of colors to the set of vertices V such that for each edge U∈ : V1 ⊆ U or U ∈ : U ⊆ V2t−1 ,
t t
{u, v} ∈ E, the vertices u and v are assigned different colors.
where V1 , V2t−1 are subspaces of V of dimensions 1 and
The chromatic number of a graph G, denoted χ( G ), is the
2t − 1, respectively. However, by [23], there is no tiling of V
minimum number of colors in which we can color G. Before
by Ui of these shapes. Thus,
we discuss the q-Kneser hypergraph we will concentrate on
∗ , i.e. a
the q-Kneser graph [6], [7] which is related to N2,r,2 [2tt]
sub-network of a combination network with two messages. χ(qK2t:t ) > = qt + 1,
∗ [2tt−−11]
The network N2,r,2 has two messages and in a scalar
network coding solution on the link between the source and since such a tiling of V will have size qt + 1. It follows that
each node in the middle layer there is a global coding vector
from F2q . The two vectors on two distinct such edges, which gapv (N ) = qs (N ) − qv (N ) > (qt + 2 − 1) − qt = 1.
transmit information to two middle layer nodes (that represent
two disjoint t-subspaces of F2t q ), must be linearly independent. For h messages, the vector network code for the Nh,r,h ∗
The set of such pairs of nodes is exactly the pairs of vertices ∗
network is exactly as in the N2,r,2 network. The coding
which define edges in qK2t:t . Hence, each color of a vertex coefficients on the edge from the source to the middle-layer
in qK2t:t will be associated with a vector of F2q , such that node represented by the t-subspace X is formed by splitting
two different colors will be associated with two linearly G ( X ) to h matrices of size t × t. For the scalar linear
independent vectors. Since the largest set of vectors in F2q network code we consider the q-Kneser hypergraph qKht:t h . Our
which are pairwise linearly independent is q + 1, it follows generalization is different from other generalizations, e.g. [19]
that if the chromatic number of qK2t:t is c then the alphabet and references therein. As for the coloring of the graph, we
size for the linear scalar solution is the smallest prime power generalize the definition of coloring to hypergraphs as follows.
greater than or equal to c − 1. The vertices are colored by a set of colors in a way that each
This coloring method and the reduction from the network vertex in an hyperedge has a different color. The chromatic
to the q-Kneser graph are similar to some ideas in [12], [21]. number of such a hypergraph G, denoted χ( G ), is the smallest
Our method is unique by using the q-Kneser graph and also number of colors required to color G.
its new generalization.
Theorem 9. For any prime power q and integers t > 2, h > 3,
Theorem 7. For a prime power q and an integer t > 2, t ∈
h −1
{2, 3} or t < q log q − q, there exists a minimal network N h qht − 1
with two messages for which gapv (N ) > qt−1 − 1.
χ(qKht:t )>
qt − 1
= ∑ qit .
i =0
Proof: By [6], [7], the chromatic number of qK2t:t is
Proof: Let U be a set of (qht − 1)/(qt − 1) pairwise-
t t −1
χ(qK2t:t ) = q + q for t = 2, 3 or t < q log q − q. disjoint subspaces of F ht
q . Such a set is called a spread and
 h
it exists for all h and t [10]. The vertices of qKht:t related to 3) Can vector gaps for multicast networks with two mes-
the t-subspaces that are in U should be colored in a different sages be larger than the one obtained for minimal
colors, which implies the claim of the theorem. multicast networks?
Corollary 10. For any prime power q and integers t > 2, h > 3, 4) What is the largest possible vector gap as a function of
there exists a minimal network N with h messages for which h and t for a multicast network with h messages?
gapv (N ) > ∑hi=−21 qit . ACKNOWLEDGMENT
A. Wachter-Zeh and M. Schwartz were supported in part by
In [11] it was proved that for even h > 4 there exists a
a German Israeli Project Cooperation (DIP) grant under grant
multicast network (not minimal) N for which gapv (N ) =
2 no. PE2398/1-1 and KR3517/9-1.
q( h−2) t /h+o ( t), and for odd h > 5 there exists a mul-
R EFERENCES
ticast network (not minimal) N for which gapv (N ) =
2 [1] R. Ahlswede, N. Cai, S.-Y. R. Li, and R. W. Yeung, “Network informa-
q( h−3) t /( h−1)+o ( t). Corollary 10 implies that if t is fixed a tion flow”, IEEE Trans. Inform. Theory, vol. 46, pp. 1204–1216, 2000.
vector gap larger than q f ( t) for any function f (t) can be [2] T. M. Apostol, Introduction to Analytic Number Theory. Springer-Verlag,
obtained. It is well-known [9], [11] that a scalar linear network NY, 1976.
[3] R. C. Baker, G. Harman, and J. Pintz, “The difference between con-
coding solution can be translated to a vector coding solution secutive primes, II”, Proceedings of the London Mathematical Society,
with vectors of length t over F q . Corollary 10 and the vector vol. 83, pp. 532–562, 2001.
gaps proved in [11] imply that a translation from a vector [4] A. Beutelspacher and J. Ueberberg, “A characteristic property of ge-
ometric t-spreads in finite projective spaces”, Eur. J. Comb., vol. 12,
coding solution with vectors of length t over F q to a scalar pp. 277–281, 1991.
linear solution will require an alphabet of size q f ( h,t), with an [5] M. Blaum, J. Bruck, and A. Vardy, “MDS array codes with independent
parity symbols”, IEEE Trans. Inform. Theory, vol. 42, pp. 529–542,
interesting trade-off between h and t in f (h, t). 1996.
Finally, Theorem 9 can be improved and as a consequence [6] A. Blokhuis, A. E. Brouwer, and T. Szőnyi, “On the chromatic number
also Corollary 10. A lower bound on the chromatic number of q-Kneser graphs”, Designs, Codes and Crypt., vol. 65, pp. 187–197,
h 2012.
of qKht:t is obtained by using a normal spread (also called a [7] A. Chowdhury, C. Godsil, and G. Royle, “Colouring lines in projective
geometric spread) [4], [17] and the chromatic number of qK2t:t space”, J. Combin. Theory Ser. A, vol. 113, pp. 39–52, 2006.
as given in [6], [7]. For example we have: [8] A. Dimakis, P. B. Godfrey, Y. Wu, M. J. Wainwright, and K. Ramchan-
dran, “Network coding for distributed storage systems”, IEEE Trans.
Theorem 11. For a prime power q and an integer t > 2, Inform. Theory, vol. 56, pp. 4539–4551, 2010.
t ∈ {2, 3} or t < q log q − q, there exists a minimal net- [9] J. B. Ebrahimi and C. Fragouli, “Algebraic algorithms for vector network
coding”, IEEE Trans. Inform. Theory, vol. 57, pp. 996–1007, 2011.
work N with h messages for which gapv (N ) > q(h−1)t + [10] T. Etzion and A. Vardy, “Error-correcting codes in projective space”,
∑hi=−11 qit−1 − 1. IEEE Trans. on Inform. Theory, vol. 57, pp. 1165–1173, 2011.
[11] T. Etzion and A. Wachter-Zeh, “Vector network coding based on
In the full version of the paper we will also prove that the subspace codes outperforms scalar linear network coding”, IEEE Trans.
vector-gap problem, for minimal multicast networks with two on Inform. Theory, vol. 64, pp. 2460–2473, 2018.
[12] C. Fragouli and E. Soljanin, “Information flow decomposition for
messages can be reduced to sub-networks of the combination network coding”, IEEE Trans. Inform. Theory, vol. 52, pp. 829–848,
network. The proof is based on a few reductions, where the 2006.
first one is similar to the one in [12]. [13] C. Fragouli and E. Soljanin, “(Secure) Linear network coding multicast”,
Designs, Codes, and Crypt., vol. 78, pp. 269–310, 2016.
IV. C ONCLUSION [14] P. Frankl and R. M. Wilson, “The Erdős-Ko-Rado theorem for vector
spaces”, J. Combin. Theory Ser. A, vol. 43, pp. 228–236, 1986.
The family of combination networks and their sub-networks [15] S. Jaggi, P. Sanders, P. A. Chou, M. Effros, S. Egner, K. Jain,
L. Tolhuizen, “Polynomial time algorithms for multicast network code
to prove two results. The first one is an upper bound on the construction”, IEEE Trans. Inform. Theory, vol. 51, pp. 1973–1982,
number of nodes in the middle layer for a vector network 2005.
coding solution. The second one is that for any number of [16] S.-Y. R. Li, R. W. Yeung, and N. Cai, “Linear network coding”, IEEE
Trans. Inform. Theory, vol. 49, pp. 371–381, 2003.
messages vector network coding outperforms scalar network [17] G. Lunardon, “Normal spreads”, Geometriae Dedicata, vol. 75, pp. 245–
coding for minimal multicast networks with respect to the field 261, 1999.
size. The first result is an MDS bound for vector spaces and the [18] F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-Correcting
Codes, North-Holland, 1978.
proof is based on vector spaces and is simpler than the one for [19] F. Meunier, “Colorful subhypergraphs in Kneser hypergraph”, The
nonlinear MDS codes. The second result induces an interesting Electronic Journal of Combinatorics, vol. 21, #P1.8, 2014.
question on the chromatic number of q-Kneser hypergraphs. [20] D. Raghavarao, Constructions and Combinatorial Problems in the
Design of Experiments, John Wiley, 1971.
There are a few more problems which are induced directly [21] A. Rasala Lehman and E. Lehman, “Complexity classification of net-
from our discussion. work information flow problems”, Proc. 15th Annu. ACM/SIAM Symp.
1) Can the vector gap in minimal multicast networks with Disc. Alg. (SODA), pp. 142–150, New Orleans, LA, January 2004.
[22] S. Riis and R. Ahlswede, “Problems in network coding and error
more than two messages be reduced to subgraphs of the correcting code”, Lecture Notes in Computer Science, vol. 4123, pp. 861–
combination networks? 897, 2006.
2) What is the maximum vector gap for minimal multicast [23] M. Schwartz and T. Etzion, “Codes and anticodes in the Grassman
graph”, J. Combin. Theory Ser. A, vol. 97, pp. 27–42, 2002.
networks, with h messages and vectors of length t? [24] N. Silberstein, T. Etzion, and M. Schwartz, “Locality and availability of
Is it the one obtained by using the chromatic number array codes constructed from subspaces”, IEEE Trans. Inform. Theory,
h )?
χ(qKht:t to appear.