Lecture Graphical Networks
(Reading: NIT .–.)
∙ Graphical multicast network
∙ Max-flow min-cut theorem
∙ Network coding theorem
∙ Graphical multimessage network
© Copyright – Abbas El Gamal and Young-Han Kim
Graphical multicast network
̂j
M
j
C
C N
M ̂N
M
C
k
̂k
M
∙ Weighted directed acyclic graph G = (N , E , C)
∙ A (nR , n) code:
A message set [ : nR ]
Source encoder mj (m) ∈ [ : nCj ]
Relay encoder k ∈ [ : N]: mkl (mjk : (j, k) ∈ E) ∈ [ : nCkl ]
Decoder k ∈ D: m̂ k (mjk : (j, k) ∈ E)
∙ Pe(n) , achievability, and C: Defined as before
/
Cutset upper bound on C
S Sc ̂j
M
j
C
C N
M ̂
M
C
k
̂k
M
Theorem .
C ≤ min min C(S),
j∈D S⊂N
∈S, j∈S c
where C(S) = ∑k∈S, l∈S c Ckl
/
Proof
∙ Consider a cut (S , S c ) and a destination node j ∈ S c
∙ For any (nR , n) code, M
̂ j is a function of M(S , S c ) = {Mkl : k ∈ S , l ∈ S c }
∙ By Fano’s inequality,
̂ j ) + nєn
nR ≤ I(M; M
̂ j ) + nєn
≤ H(M
≤ H(M(S , S c )) + nєn
≤ nC(S) + nєn
∙ Since the above holds for every cut and every destination, in the limit
R ≤ min min C(S)
j∈D S⊂N
∈S, j∈S c
/
Capacity of graphical unicast network
∙ Unicast: D = {N}
Theorem . (Max-flow min-cut theorem (Ford–Fulkerson ))
C= min C(S)
S⊂N
∈S,N∈S c
∙ Capacity is achieved error-free using routing with finite block length (NIT .)
∙ Information can be treated as a commodity
∙ Ford–Fulkerson algorithm finds capacity and optimal routing
∙ Continues to hold for networks with cycles and delays
/
Capacity of graphical unicast network
∙ Example:
S Sc
M ̂
M
C=
Minimum cut: S = {, , , }
Achieved by routing bit along → → → and bits along → → →
∙ Cutset bound also achievable for multicast networks:
Not always achieved using only routing
/
Butterfly multicast network
M
̂
M
M M
M ⊕ M
M ⊕ M
M = (M , M )
M ⊕ M
M M
M
̂
M
∙ Cutset bound: C ≤
∙ Routing achieves R = R + R =
∙ Achieving cutset bound requires network coding!
Treating information as a commodity is not optimal in general
/
Capacity of graphical multicast network
Network coding theorem (Ahlswede–Cai–Li–Yeung )
C = min min C(S)
j∈D S⊂N
∈S,j∈S c
∙ Achieved by random binning (Ahlswede–Cai–Li–Yeung )
∙ Also achieved error free using linear coding: Li–Yeung–Cai (),
Koetter–Médard () (NIT ..)
∙ Continues to hold for networks with broadcasting, cycles, and delays
/
Proof of achievability via binning: Unicast case
g g
M ̂M
M g ̂M
M
g g
∙ Random codebook generation:
Source encoder: Generate gj (m) ∼ Unif[ : nCj ] for each m ∈ [ : nR ], e.g.,
g (m) ∼ Unif[ : n ] and g (m) ∼ Unif[ : n ]
Relay encoder: Generate gkl (m∗k ) ∼ Unif[ : nCkl ] for each m∗k = (mjk : (j, k) ∈ E), e.g.,
g (m , m ) ∼ Unif[ : n ]
The mappings gjk , (j, k) ∈ E, induce functions of m at every edge, e.g.,
m (m) = g (g (m), g (g (m)))
∙ Encoding: To send message m, node j transmits mjk (m) over edge (j, k) ∈ E
/
∙ Decoding:
Analysis
of the probability of error
̂ (̂) =
∙ Consider the probability of error averaged over M and (Gjk : (j, k) ∈ E):
̂
P(E) = P{M ̸= M}
̃ for some m
= nR P M∗N (m) = M∗N (m) ̃ ̸= m|M = m
m
̃ for some m
= P M∗N () = M∗N (m) ̃ ̸=
̃
≤ P M∗N () = M∗N (m)
̃ =
m̸
nR
= ( − ) P M∗N () = M∗N ()
∙ Thus, it suffices to bound the collision probability P{M∗N () = M∗N ()}
/
Collision probability
M ̂M
M ̂M
M
∙ We decompose P{M∗N () = M∗N ()} by different cuts (N = ):
P M∗ () = M∗ () = P M∗ () = M∗ (), M∗ () = M∗ (), M∗ () = M∗ ()
+ P M∗ () = M∗ (), M∗ () ̸= M∗ (), M∗ () = M∗ ()
+ P M∗ () = M∗ (), M∗ () = M∗ (), M∗ () ̸= M∗ ()
+ P M∗ () = M∗ (), M∗ () ̸= M∗ (), M∗ () ̸= M∗ ()
∙ No collision at node , collision at nodes , , :
P M∗ () = M∗ (), M∗ () = M∗ (), M∗ () = M∗ ()
/
= P (G (), G ()) = (G (), G ())
Proof of achievability: Multicast case
=
∙ We repeat the unicast argument for each destination j ∈ D
∙ By the union of events bound,
P(E) ≤ N−
j∈D n(minS:∈S,j∈S c C(S)−R)
≤ |D|N− n(min j∈D minS:∈S,j∈S c C(S)−R)
,
which tends to as n → ∞, if R < minj∈D minS:∈S,j∈S c C(S)
/
Graphical multimessage network
̂
M ̂ j , M
(M ̂ j )
j
M
C
C N
M ̂ N
M
C
M
k
̂
M ̂ k , M
(M ̂ k )
∙ Each node j ∈ [ : N − ] wishes to send Mj to a set Dj ⊆ [j + : N]
∙ A (nR , . . . , nRN− , n) code:
Message sets [ : nR ], . . . , [ : nRN− ]
Encoder k ∈ [ : N − ]: mkl (mk , mjk : (j, k) ∈ E)
̂ jl (mkl : (k, l) ∈ E) for j such that l ∈ Dj
Decoder l ∈ ∪j Dj : m
∙ Pe(n) , achievability, and C : defined as before
/
Cutset bound for graphical multimessage network
Theorem .
If (R , . . . , RN− ) is achievable, then
Rj ≤ C(S)
j∈S
for all S ⊂ N such that Dj ∩ S c ̸= for some j ∈ [ : N − ]
∙ Tight for multimessage multicast (Dj = D, j ∈ [ : N − ])
∙ Will be later extended to noisy networks
∙ Not tight in general
/
Multimessage multicast
Theorem . (Ahlswede–Cai–Li–Yeung )
The capacity region is the set of (R , . . . , Rk ) such that
Rj ≤ C(S)
j∈S
for all S ⊂ N with [ : k] ∩ S ̸= and D ∩ S c ̸=
∙ Key idea: augmented single-message multicast network (Read NIT ..)
R
M
R
G
Rk
G k
/
Multiple unicast
∙ Routing is insufficient: Consider the following example (Cjk = )
M ̂
M
M ̂
M
∙ Cutset bound is not tight: Consider the following example (Cjk = )
M ̂
M
M ̂
M
∙ Capacity is not known in general
/
Summary
∙ Cutset bounds on the capacity of graphical networks
∙ Max-flow min-cut theorem for graphical unicast networks
∙ Routing alone does not achieve the capacity of general graphical networks
∙ Network coding theorem for graphical multicast networks
∙ Linear network coding achieves the capacity of graphical multimessage
multicast networks (error-free and with finite block length)
/
References
Ahlswede, R., Cai, N., Li, S.-Y. R., and Yeung, R. W. (). Network information flow. IEEE Trans. Inf. Theory,
(), –.
Ford, L. R., Jr. and Fulkerson, D. R. (). Maximal flow through a network. Canad. J. Math., (), –.
Koetter, R. and Médard, M. (). An algebraic approach to network coding. IEEE/ACM Trans. Netw., (),
–.
Li, S.-Y. R., Yeung, R. W., and Cai, N. (). Linear network coding. IEEE Trans. Inf. Theory, (), –.
/
