on multi-source multi-sink hyperedge networks: enumeration ... · on multi-source multi-sink...
TRANSCRIPT
On Multi-source Multi-sink Hyperedge Networks:Enumeration, Rate Region Computation, & Hierarchy
—Ph. D. Dissertation Defense
Congduan Li
ASPITRG & MANLDrexel University
May 29, 2015
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 1 / 73
Motivation: information flow in networks
Coding (mix the input) outperforms routing (receive-forward) in manycircumstances
Efficient information flow in networks: admissible source rate v.schannel capacities
?Data Sources
Relay Nodes/ Routers
Network
Users
?
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 2 / 73
Motivation: information storage in data centers
Coding (mix sources) saves resources, compared with simplereplication
Efficient information storage in data centers: admissible source filesize v.s hard disk capacities
Source Files
Hard Drives
Users
?
?
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 3 / 73
Generalized model
A multisource multisink hyperedge network (hyperedge MSNC)
Paths
In(i) i
e 2 EU
1
s
S T
K |T |
t
1
Ys �(t)
.
.
.
.
.
.
.
.
.
.
.
.
Paths
In(i) i
ReUe
Out(i)
1
s
K |T |
t
1
�(t)
.
.
.
.
.
.
.
.
.
.
.
.
Sources Network
...
Sinks
A[R1, R2, . . .]T � B[H(Y1), H(Y2), . . .]
T
Rate/capacity region: compute
Notation: (K , |EU |) means Ksources, |EU | intermediatehyperedges
We made three major contributions
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 4 / 73
Contribution 1: Revolution, use computer for rate regioncalculation
Conventional: 1 (special) network, info. ineq., manually, 1 paper
Computational: 103, 106, 1012 (arbitrary) networks, computer, # ofpapers?
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
. . .. . .
Conventional way Computational way
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 5 / 73
Contribution 2: enumeration of all networks
Conventional: almost impossible to list all networks due to the largenumber of instances
Computational: enumerate ≥ 109, 1012 general networks
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
In(i) i
eReUe
Out(i)
1
s
S T
|S| |T |
t
1
i
1
s
S T
|S| |T |
t
1
. . .. . .
Conventional way Computational way
?
Trillions of Networks
Enumeration Algorithm
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 6 / 73
Contribution 3: build hierarchy between networks
Conventional: what to do with so many rate regions?
Computational: define embedding and combination operators, build ahierarchy, analyze the rate regions and use them to solve even morenetworks in scale
Conventional way Computational way
?
Trillions of Rate Regions
Hierarchy Operators
. . .. . .
... ...
...
...
... ...
Build hierarchy between themand solve more networks in scale
... ...
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 7 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 8 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 9 / 73
Network Coding: Sources
We assume UOut(s) = Ys
Paths
In(i) i
e
1
s
S T
|S| |T |
t
1
Ys �(t)
.
.
.
.
.
.
.
.
.
.
.
.
Paths
In(i) i
ReUe
Out(i)
1
s
|T |
t
1
�(t)
.
.
.
.
.
.
.
.
.
.
.
.
Sources Network
...
Sinks
K
H(Ys) ≥ ωs
H(YS) =�
s∈SH(Ys)
H(UOut(s)|Ys) = 0
Source Rate
Source Independence
Source Encoding
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 10 / 73
Network Coding: Edges
Paths
In(i) i
e
1
s
S T
|S| |T |
t
1
Ys �(t)
.
.
.
.
.
.
.
.
.
.
.
.
Paths
In(i) i
ReUe
Out(i)
1
s
|T |
t
1
�(t)
.
.
.
.
.
.
.
.
.
.
.
.
Sources Network
...
Sinks
K
Coding Rate Re ≥ H(Ue), e ∈ E
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 11 / 73
Network Coding: Nodes
Paths
In(i) i
e
1
s
S T
|S| |T |
t
1
Ys �(t)
.
.
.
.
.
.
.
.
.
.
.
.
Paths
In(i) i
ReUe
Out(i)
1
s
|T |
t
1
�(t)
.
.
.
.
.
.
.
.
.
.
.
.
Sources Network
...
Sinks
K
Coding Constraints H(UOut(i)|UIn(i)) = 0, i ∈ V\(S ∪ T )
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 12 / 73
Network Coding: Sinks
Paths
In(i) i
e
1
s
S T
|S| |T |
t
1
Ys �(t)
.
.
.
.
.
.
.
.
.
.
.
.
Paths
In(i) i
ReUe
Out(i)
1
s
|T |
t
1
�(t)
.
.
.
.
.
.
.
.
.
.
.
.
Sources Network
...
Sinks
K
Decoding Constraints H(Yβ(t)|UIn(t)) = 0, t ∈ T
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 13 / 73
Special Case: Independent Distributed Source Coding(IDSC)
s1
s2
g1
t1
t2U2
U1
U3
Y1
Y2 t3
Y1
Y2
Y1Y2
E1
E2
D1
D2
D3E3
Y1
Y2
Y1Y2
U2
U1
U3
Y1
Y2
Sources available to all encoders
Decoders demand various subsets of of sources
When sources are prioritized, it becomes MDCS
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 14 / 73
Special Case: Index Coding
Only one intermediate edge that transmits all information of sources
Sources may be directly available at sinks as side information
(K , 1) hyperedge MSNC
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 15 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 16 / 73
Source minimality example
Source minimality example: Source s3 redundant due to not demanded byany sink.
s1
s2
g1
g2
U1
U2
U3
Y1
Y2t3
Y2
t2
t1
Y2
s3Y3
s1
s2
g1
g2
U1
U2
U3
Y1
Y2t3
Y2
t2
t1
Y2
Y1 Y1
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 17 / 73
Node minimality example
Node minimality example: g1, g2 can be merged due to same input
s1
s2
g1
g2
U1
U2
U3
Y1
Y2t3
Y2
t2
t1
Y2
s1
s2
g2
U1
U2
U3
Y1
Y2t3
Y2
t2
t1
Y2
Y1 Y1
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 18 / 73
Edge minimality example
Edge minimality example: U2,U3 are parallel and can be merged
s1
s2
g1
g2
U1
U2
U3
Y1
Y2 t2
t1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2 t2
t1
Y2
Y1 Y1
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 19 / 73
Sink minimality example
Sink minimality example: decoding ability of Y2 at t2 implied by t1,equivalent to let t2 demand Y1 only
s1
s2
g1
g2
U1
U2
U3
Y1
Y2t3
Y2
t2
t1
Y2
Y1Y2
s1
s2
g1
g2
U1
U2
U3
Y1
Y2t3
Y2
t2
t1
Y2
Y1
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 20 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 21 / 73
Equivalent Networks
s1
s2
g1
g2
U1
U2
Y1
Y2
Y2
t2
t1
Y1
Y2
s1
s2
g1
g2
U1
U2
t2
t1
s1
s2 g1
g2
U1
U2Y1
Y2 Y2
t2
t1
Y1
Y1
Y1
Y2
Y1 $ Y2 & U1 $ U2
⇥ Y1 6$ Y2
I.)
II.)
III.)
I , II are equivalent: permutesources and edges
I , III are not equivalent: sourcesare only permuted at source side
II , III are not equivalent:sources are only permuted atsink side
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 22 / 73
Representing a Network
Ordered pair (Q,W), sources 1 . . .K , edges K + 1, . . . ,K + |EU |;Edge definitions Q ⊆{(i ,A)|i ∈ {K + 1, . . . ,K + |EU |}, A ⊆ {1, . . . ,K + |EU |} \ {i}};Sink definitionsW ⊆ {(i ,A)|i ∈ {1, . . . ,K}, A ⊆ {1, . . . ,K + |EU |} \ {i}};Same i allowed to appear in W but not in Q
12
34
12 2
(3, {1, 2}) (4, {1, 3})
(1, {3, 4})(2, {3}) (2, {4})
Q
W
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 23 / 73
Equivalence Under Group Action
Symmetry group G := S{1,2,...,K} × S{K+1,...,K+|EU |};
π ∈ G, then π(Q) 7→ {π((i ,A))|(i ,A) ∈ Q};π((Q,W)) = (π(Q), π(W));
Isomorphic or Equivalent: ∃π ∈ G such thatπ((Q1,W1)) = (Q2,W2).
12
34
12 2
(3, {1, 2}) (4, {1, 3})
(1, {3, 4})(2, {3}) (2, {4})
Q
W
12
34
21 1
(3, {1, 2}) (4, {2, 3})
(2, {3, 4})(1, {3}) (1, {4})
Q0
W 0
⇡ : permute 1, 2
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 24 / 73
Canonical Network: minimal representative in each orbit
Orbit: O(Q,W) := {(π(Q), π(W))|π ∈ G}Networks in an orbit are isomorphic or equivalent to each other
Transversal: one representative for each orbit, the canonical one
Lexicographically order the pairs (i ,A) according to (i ,A) > (j ,A′) ifj < i or i = j ,A′ < A under the lexicographic ordering
Canonical: apply order to (Q,W) and get the minimal one
# 1 2 3 4` 2 1 2 1
# 1 2 3 4` 3 3 4 4
# 1 2 3 4` 2 1 2 1
# 1 2 3 4` 2 1 2 1
# 1 2 3 4` 1 2 1 2
# 1 2 3 4` 4 4 3 3
# 1 2 3 4` 1 2 1 2
# Q W1 {(3, {1, 2}), (4, {1, 3})} {(1, {3, 4}), (2, {3}), (2, {4})}2 {(3, {1, 2}), (4, {2, 3})} {(1, {3}), (1, {4}), (2, {3, 4})}3 {(3, {1, 4}), (4, {1, 2})} {(1, {3, 4}), (2, {3}), (2, {4})}4 {(3, {2, 4}), (4, {1, 2})} {(1, {3}), (1, {4}), (2, {3, 4})}
# 1 is the canonical representative
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 25 / 73
Why this representation
Some minimality constraints easily to take care, e.g., repeated edges,redundant nodes, etc
Smaller orbits, compared with node representation
Orbit-stabilizer theorem: |O(Q,W)| = |G ||Stab((Q,W))|
# 1 2` 2 1
# 1 2` 2 1
# 1 2` 2 1
# 1 2` 1 2
# 1 2` 4 4
# 1 2` 1 2
# Q W1 {(3, {1, 2}), (4, {1, 2})} {(1, {3, 4}), (2, {3}), (2, {4})}2 {(3, {1, 2}), (4, {1, 2})} {(1, {3}), (1, {4}), (2, {3, 4})}
c
d e f
a b
a ab
Subgroup of S{1,2} ⇥ S{3,4} stabilizing (G, T ) = S{3,4}
Subgroup of S{a,b} ⇥ S{c,d,e,f}stabilizing (V, E) = S{d,f}
# a b c d e f1 2 1 3 4 5 62 1 2 3 4 5 63 2 1 4 5 6 3......
24 1 2 3 4 6 5
Isomorphs in Edge and Sink Definition Representation (2)
Isomorphs in Node Representation (24)
# 1 2` 3 3
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 26 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 27 / 73
Leiterspiel Algorithm [BettenBraunFripertinger, 2006]
Computes transversal of orbits on size j subsets Pj(X ) of set X ,incrementally in j ; also gives symmetry group (stabilizer)Lists directly the canonical representatives satisfying some testfunction f , as long as this test has the inherited property, i.e., if asuperset satisfies, its subsets also satisfyInput: set X , group G , inherited test function fOutput: transversal of subsets of different sizes until stop, either fixedj or other stop conditions, symmetry group (stabilizer)
Orbits on Pj�1(X )
Orbits on Pj(X )
Extensionobeying f
Transversal: canonical representatives
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 28 / 73
Enumeration based on Leiterspiel Algorithm
Target: (K , |EU |) non-isomorphic networks
Recall representation of networks: (Q,W), list pool for Q first
Leiterspiel defines edges incrementally from 1 to |EU |
X := {(i, A) |i 2 {K + 1, . . . , K + |EU |}, A ✓ {1, . . . , K + |EU |} \ {i}}
Acting group G := S{1,...,K} ⇥ S{K+1,...,K+|EU |}
Call Leiterspiel T|EU | = Leiterspiel(G, Pf|EU |(X ))
With some minimality constraints inherited
Stop condition: reach number of edges
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 29 / 73
Enumeration based on Leiterspiel Algorithm
Now for each Q ∈ T|EU |, list possible WLeiterspiel incrementally adds sinks from 1 to no possible new sink
Acting group G := S{1,...,K} ⇥ S{K+1,...,K+|EU |}
With some minimality constraints inherited
Y := {(i, A) |9 a directed path in Q from i to at least one edge in A}
Call Leiterspiel T|T | = Leiterspiel(G, Pf|T |(Y))
Stop condition: cannot increase j obeying f
For each W ∈ {T1, . . . ,T|T |}, test all the other minimality conditions on(Q,W), if it passes, we obtain a non-isomorphic network instance
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 30 / 73
Enumeration Results
|M|: number of non-isomorphic networks, listed;|M|: number of networks with edge isomorphism, counted;|Mn|: number of networks with node isomorphism, counted;
(K , |EU |) (1,2) (1,3) (2,1) (2,2) (2,3) (3,1) (3,2)
|M| 4 132 1 333 485 890 9 239 187
|M| 7 749 1 1 270 5 787 074 31 2 829 932
|Mn| 39 18,401 6 ≥ 105 ≥ 1012 582 ≥ ×1011
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 31 / 73
All (2, 2) networks: no direct access btw sources & sinks
s1
s2
g1
g2
U1
U2
Y1
Y2
Y1Y2t1s1
s2
g1
g2
U1
U2
Y1
Y2
t1
Y1Y2t2
Y1 s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2 Y1Y2
t1 Y1s1
s2
g1
g2
U1
U2
Y1
Y2
Y1Y2t1
t2 t2 Y1
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t2
Y1
Y2
t3
t1
Y1
s1
s2
g1
g2
U1
U2
Y1
Y2
Y1Y2t1s1
s2
g1
g2
U1
U2
Y1
Y2
Y1Y2t1
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y1Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1 Y1
t2 Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y1
Y2 s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t2
Y1
Y2
t3
t1
Y1
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
Y1Y2t2
Y2 s1
s2
g1
g2
U1
U2
Y1
Y2
t1 Y1Y2
t2 Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y2
Y1s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y2
Y1s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y2
Y1
s1
s2
g1
g2
U1
U2
Y1
Y2
t2
Y1
Y2
t3
t1
Y1
s1
s2
g1
g2
U1
U2
Y1
Y2
Y1Y2t1
s1
s2
g1
g2
U1
U2
Y1
Y2
Y1Y2t1s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1Y2
Y1Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y1Y2
Y2 s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1Y2
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y1Y2
Y1s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y1
Y1Y2 s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y1
Y2 s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y1
Y2s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y2
Y1
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t2 Y1
Y2
t3
t1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t2
Y1
Y2
t3
t1
Y1
s1
s2
Y1
Y2
g1
t1
t2
Y2U1
U2 Y1Y2
s1
s2
Y1
Y2
g1
t1
t2 Y2
U1
U2
Y1s1
s2
Y1
Y2
g1
U1
U2
t1
t2 Y2
Y1 s1
s2
Y1
Y2
g1
U1
U2
t2 Y1
Y2
t3
t1
Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y2
Y1Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y2
Y1Y2
s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y2
Y1s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2 Y2
Y1s1
s2
g1
g2
U1
U2
Y1
Y2
t1
t2
Y2
Y1
g1
g2
U1
U2
t2 Y1
Y2
t3
t1s1
s2
Y1
Y2
Y2
Instance #1 Instance #2 Instance #3 Instance #4 Instance #6 Instance #7
Instance #8 Instance #9 Instance #10 Instance #11 Instance #13 Instance #14
Instance #15 Instance #16 Instance #17 Instance #18 Instance #19 Instance #20 Instance #21
Instance #22 Instance #23 Instance #24 Instance #25 Instance #26 Instance #27 Instance #28
Instance #29 Instance #30 Instance #31 Instance #32 Instance #33 Instance #34 Instance #35
Instance #36 Instance #38 Instance #39 Instance #40Instance #41
Instance #42
Instance #43 Instance #44 Instance #45 Instance #46
Instance #5
Instance #12
Instance #37
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 32 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 33 / 73
Rate region
Rate region: all possible rate and source entropy vectors satisfying allnetwork constraints.
Collect the N network random variables and their joint entropies.
Define Γ∗N : 2N − 1-dim. cone, region of valid entropy vectors. (revisitlater)
Constraints from network A:L1 = {h ∈ Γ∗N : hYS = Σs∈ShYs} (1)
L2 = {h ∈ Γ∗N : hXOut(k)|Ys= 0} (2)
L3 = {h ∈ Γ∗N : hXOut(i)|XIn(i)= 0} (3)
L4 = {(hT ,RT )T ∈ R2N−1+|E|+ : Re ≥ hUe , e ∈ E} (4)
L5 = {h ∈ Γ∗N : hY β(t)|UIn(t)= 0}. (5)
Rate region (cone) in terms of edge rates and source entropies(derived from [Yan, Yeung, Zhang TranIT 2012]):
R∗(A) = projRE ,H(YS)(con(Γ∗N ∩ L123) ∩ L45) (6)
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 34 / 73
Rate Region Example
A (3, 3) network and its rate region R∗(A)
Rate region: a cone with dimensions of all variables in the network
s1
s2g1
g2
U1
U2
U3Y1
Y2
t3
t2
t1
s3Y3 Y3
Y1Y2g3 R1 � H(Y3)
R2 � H(Y1)
R1 + R2 � H(Y1) + H(Y2) + H(Y3)
R2 + R3 � H(Y1) + H(Y3)
R1 + R2 + 2R3 � H(Y1) + H(Y2) + 2H(Y3)
Y3
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 35 / 73
Rate region
Rate region: a cone in terms of edge rates and source entropies:
R∗(A) = projRE ,H(YS)(con(Γ∗N ∩ L123) ∩ L45) (7)
Involves Γ∗N : 2N − 1-dim. cone, region of valid entropy vectors.
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 36 / 73
Region of Entropic Vectors
Γ∗N :
Open set in general
Γ∗N not fully characterized for N ≥ 4: convex but containsnon-polyhedral part
�⇤N
Inner bounds
Outer bounds
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 37 / 73
Sandwich Bounds
Γ∗N → ΓOutN : Rout(A) = projRE ,H(YS)(ΓOut
N ∩ L12345)
Γ∗N → ΓInN : Rin(A) = projRE ,H(YS)(ΓIn
N ∩ L12345)
R∗(A) = Rout(A) = Rin(A), if Rout(A) = Rin(A)
It becomes: Initial polyhedra → ∩ constraints → projections
Our work following this idea: Li, et. al, Allerton 2012, NetCod 2013,submission TransIT 2014.
Constraints Projection
�N
�InN
�⇤N
Constraints �N
�InN
�⇤N
Constraints
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 38 / 73
Notion of sufficiency
Outer bound typically used is Shannon outer bound→ Ro(A); innerbounds from representable matroids: scalar and vector bounds.
Scalar sufficiency: R∗(A) = Rs,q(A)
Vector sufficiency: R∗(A) = RN′q (A)
Sufficient
Insufficient
Scalar Rs,q(A) Vector RN 0q (A)
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 39 / 73
Rate Region Computation Results
|M|: number of all network instances;The other numbers represent the numbers of instances we can close thegap using various bounds, and hence exact rate region can be obtained
(K , |E|) |M| Rs,2(A) RN+12 (A) RN+2
2 (A) RN+42 (A)
(1, 2) 4 4 4 4 4
(1, 3) 132 122 132 132 132
(2, 1) 1 1 1 1 1
(2, 2) 333 301 319 323 333
(2, 3) 485890 341406 403883 432872 –
(3, 1) 9 4 4 9 9
(3, 2) 239187 118133 168761 202130 –
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 40 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 41 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 42 / 73
Motivated from graph and matroid theory
Inheritance property regarding sufficiency of class of codes
Minor-closed graphs: finite number of forbidden minors[RobertsonSeymour1983-2004]
Rota’s conjecture in matroid theory: finite number of forbiddenminors for Fq representability [Oxley2011]
N � 2
N � 1
N
. . .
. . .
. . . . . .
. . .
. . .
ForbiddenNot Fobidden Size
New Minimal Forbidden Minors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 43 / 73
Similar characterization for networks?
If networks have similar characterization? Possible list of forbiddenembedded networks for sufficiency of linear codes over a field.
Network operations to obtain such embedded networks preservinginsufficiency, & region relationships
Three operations
Sources Network Destinations
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 44 / 73
Source deletion
A′ = A \ kSource Yk deleted, source kstops sending information to thenetwork, H(Yk) = 0
Sinks requiring Yk will no longerdemand it.
kYk
k
...
...
H(Yk0) = 0
k
...
...
t �(t) \ k
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 45 / 73
Edge deletion
A′ = A \ eEdge e deleted, nothing on Ue ,Re = H(Ue) = 0. Re0 = 0
ee
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 46 / 73
Edge contraction
A′ = A/e
Edge e contracted, input to tailof e available for head of e,Re =∞, H(Ue) free.
ee
Re0 = 1
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 47 / 73
Source deletion: A′ = A \ k
For each i ∈ {∗, q, (s, q), o},
Ri (A′) = ProjY\k ,RE′ ({R ∈ Ri (A) |H(Yk) = 0}) (8)
Sufficiency preserved from large to small network as the equation shows.Equivalently, insufficiency preserved from small to large network.
R(A)
R(A0)
H(Yk) = 0Projection
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 48 / 73
Example: Forbidden embedded networks
Goal: minimal forbidden networks for sufficiency
Scalar binary codes considered
k = 1, 2, 3; |E| = 2, 3, 4, 7360 non-isomorphic MDCS
1922 sufficient, 5438 insufficient
12 minimal forbidden minors (Li, et. al submission TransIT 2014)
5438 / 7360 insu�cient
1922 / 7360su�cient
12 forbiddenembedded networks
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 49 / 73
Summary for embedding operators
Three embedding operators: source deletion, edge deletion, edgecontraction
Rate region of smaller network derivable from the associated largernetwork
Sufficiency of linear codes preserved from larger to smaller networksunder embedding operations
Equivalently, insufficiency of linear codes preserved from smallernetworks to larger one
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 50 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 51 / 73
Source merge
After merge: Mergedsources serve as commonsources to the twonetworks
...
...
......
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 52 / 73
Sink merge
After merge: Union theinput and requests ofsinks being merged,respectively
...
...
...
...
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 53 / 73
Intermediate node merge
After merge: union inputand output of the twonodes
...
...
...
......
...
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 54 / 73
Edge merge
Edge merge: create one extranode and four associated edgesto replace the two original edges
First two edges are ordinaryedges connecting with the extranode, the other two edgesconnect the extra node with allthe head nodes of the originaltwo edges, respectively.
Equivalent: create a relay nodeon the two edges, respectively,and then merge the two relaynodes.
...
...
......
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 55 / 73
Source merge
A is obtained by merging A1.S = A2.π(S), then for eachi ∈ {∗, q, (s, q), o}Ri (A) = Proj\π(S)((Ri (A1)×Ri (A2)) ∩ L0),
L0 ={H(Ys) = H(Yπ(s)),∀s ∈ S
}
Remark: essentially replace variables Yπ(s) with the Ys for each s ∈ S.
R(A1) R(A2)8s 2 S, H(Ys) = H(Y⇡(s))
+
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 56 / 73
Obtain Rate Region for Larger Networks
1 2 3 4 5 6
1
2
43
67
8
910
12
05
11
1
{2,3}
2
{1,2,4}
3
{4}
4
{5}
5
{6}
6
{6}
7
{1,3}
8
{2,3}
9
{5} 11
{6}
10
{4,5}
X1 X2 X3 X4 X5 X6
R1
R2
R3R4
R5R6
R7 R9
R10R11
R12
R40
R70
R100
A :
1 2
1
{2}
2
{1,2}
12
R1 � H(X2)R2 � H(X2)
R1 + R2 � H(X1) + H(X2)
X1X2
R1 R2
A1 :5 6
8 76
6
{6}
5
{6}
4
{5}R6 � H(X6)
R7 + R8 � H(X6)R6 + R7 � H(X5) + H(X6)
R6 + R7 + 2R8 � H(X5) + 2H(X6)
X5 X6
R8 R6 R7
A3 :
9 10
1' 2'3'
7
{1',3'}
8
{2',3'}
R9 � H(X30)R10 � H(X10) + H(X20) + H(X30)
X10 X20X30
R9 R10
A4 :
Ri � H(X2), i = 1, 2R1 + R2 � H(X1) + H(X2)
Ri � H(X4), i = 3, 4, 40
R3 + Ri � H(X3) + H(X4), i = 4, 40
R3 + Ri + R5 � H(X3) + 2H(X4), i = 4, 40
R6 � H(X6)Ri + R8 � H(X6), i = 7, 70
R6 + Ri � H(X5) + H(X6), i = 7, 70
R6 + Ri + 2R8 � H(X5) + 2H(X6), i = 7, 70
R9 � H(X3)
R10 � P3i=1 H(Xi)
R11 � H(X5)R12 � H(X4) + H(X5)
R100 + R11 + R12 � H(X4) + 2H(X5) + H(X6)
5 4 3
3
{4}
2'
{4}
1'
{3}
3 4X3 X4
R3R4R5
A2 :4'5'6'
10' 11 12
9
{5'}
11
{6'}
10
{4'5'}
X40X50X60
R11R12R100
A5 :
Ri � H(X4), i = 3, 4R3 + R4 � H(X3) + H(X4)
R3 + R4 + 2R5 � H(X3) + 2H(X4)
R11 � H(X50)R12 � H(X40) + H(X50)
R100 + R11 + R12 � H(X40) + 2H(X50) + H(X60)
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 57 / 73
Summary for combination operators
Four combination operators: source, sink, node, and edge merge
Rate region of combined network derivable from regions of smallernetworks in the combination
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 58 / 73
Outline
1 Enumeration
Network Model
Minimality
Network Equivalence Class
Enumeration algorithm
2 Rate Region Computation
3 Network Operations
Embedding Operations
Combination Operations
Play with both
Enumerate networks
New networks
Software to calculaterate regions or bounds
Combinationoperators
Embeddingoperators
ForbiddenMinors
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 59 / 73
Combination operations suffice?
Answer is NO.
Need cap to limit the network size in the combinations.
NetworkCap: nobeyond
Combinations reach a portion of
large networks
Embeddings enlarge thereachable
portionSmall size
Tiny size
Moderate size
Large size
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 60 / 73
Partial closure of networks
Worst case partial closure ofnetworks: cap the predicted sizeof networks involved in theprocess
Let the pool produce newnetworks until no new networkcan be generated
Seed list of networks
Combinations of pairs
Minors from embedding
New network?
Add to seed list
If within network cap
Yes
StopNO
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 61 / 73
New networks found from a tiny seed list
Start with the single (1, 1), single (2, 1), and the four (1, 2) networks;
!"#"$ !"#%$ !%#%$ !"#"$ !"#%$ !%#%$!&#"$ % % % % % %!&#%$ ' &' &' ' &' &'!(#($ " " " ) &* &+!(#"$ &" &+ &+ "' &"& &**!(#%$ ' ,- &'& ' *&+ +%)!"#($ ( " ( % &' &&!"#"$ (% (% (% %( "*" )""!"#%$ ' &"* &"* ' ("+& *%)&!%#($ ' ' " ' ' "!%#"$ ' ' &- ' ' %%!%#%$ ' ' (*" ' ' %%"'.// %+ (,( *+) )) "%'' &&+"*
012345.64157189:.61:;715/< 9239==45>7.5=7012345.6415;;[email protected]
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 62 / 73
New networks found from a tiny seed list
6 tiny networks can generate new 11635 networks with small cap!
!"#"$ !"#%$ !%#%$ !"#"$ !"#%$ !%#%$!&#"$ % % % % % %!&#%$ ' &' &' ' &' &'!(#($ " " " ) &* &+!(#"$ &" &+ &+ "' &"& &**!(#%$ ' ,- &'& ' *&+ +%)!"#($ ( " ( % &' &&!"#"$ (% (% (% %( "*" )""!"#%$ ' &"* &"* ' ("+& *%)&!%#($ ' ' " ' ' "!%#"$ ' ' &- ' ' %%!%#%$ ' ' (*" ' ' %%"'.// %+ (,( *+) )) "%'' &&+"*
012345.64157189:.61:;715/< 9239==45>7.5=7012345.6415;;[email protected]
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 63 / 73
New networks found from a tiny seed list
With the increase of cap size, number of new networks increases!
!"#"$ !"#%$ !%#%$ !"#"$ !"#%$ !%#%$!&#"$ % % % % % %!&#%$ ' &' &' ' &' &'!(#($ " " " ) &* &+!(#"$ &" &+ &+ "' &"& &**!(#%$ ' ,- &'& ' *&+ +%)!"#($ ( " ( % &' &&!"#"$ (% (% (% %( "*" )""!"#%$ ' &"* &"* ' ("+& *%)&!%#($ ' ' " ' ' "!%#"$ ' ' &- ' ' %%!%#%$ ' ' (*" ' ' %%"'.// %+ (,( *+) )) "%'' &&+"*
012345.64157189:.61:;715/< 9239==45>7.5=7012345.6415;;[email protected]
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 64 / 73
New networks found from a tiny seed list
Embedding operations are important in the process!
!"#"$ !"#%$ !%#%$ !"#"$ !"#%$ !%#%$!&#"$ % % % % % %!&#%$ ' &' &' ' &' &'!(#($ " " " ) &* &+!(#"$ &" &+ &+ "' &"& &**!(#%$ ' ,- &'& ' *&+ +%)!"#($ ( " ( % &' &&!"#"$ (% (% (% %( "*" )""!"#%$ ' &"* &"* ' ("+& *%)&!%#($ ' ' " ' ' "!%#"$ ' ' &- ' ' %%!%#%$ ' ' (*" ' ' %%"'.// %+ (,( *+) )) "%'' &&+"*
012345.64157189:.61:;715/< 9239==45>7.5=7012345.6415;;[email protected]
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 65 / 73
Example to see why integrating embedding is important
Start with the single (1, 1), single (2, 1), and the four (1, 2) networksOnly combination with cap (3, 4), get only 3 networks with size (2, 2)
1
2
3 4
1 12
1
1 1
3 4
1
2
3
1
1
2
3
11
2
3
11
1 2
2 1
3
(2,1)
(1,2)
1
1
3
2
2
4
1
2
1
1
2
1
nodemerge
=
2 2
2
1 1
1
3 4
=
1 2
2 1
3
1 2
2 1
3
sources merge
+
1
2
1
1 2
2 1
3
sources merge
+
=
1
2
1
(1,1)
ALL (1,1), (2,1), (1,2)MINIMAL NETWORKCODING PROBLEMS
(there are 6)
ALLPOSSIBLECOMBINATIONOPERATORS
All (2,2) Minimal Network Coding Problems That are direct Combinations of (1,1),(2,1),and (1,2) Problems.
(There are only 3 out of 333 total (2,2) problems)
(2,2)
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 66 / 73
Example to see why integrating embedding is important
Start with the single (1, 1), single (2, 1), and the four (1, 2) networks
Consider both combination and embedding with same cap, found(2, 2) networks unreachable by combination only
2
2
1
1
34
2
2
1
1
3 4
Contracting edge 3
5
12
2
1
1
34
5
1
Contracting edge 46
2
2
2
1
1
34
2
2
1
1
34
Source merge [1,2] with [2,1]
2
1,2
1
34
5
1
2
1,2
1
34
5
1,2
6
Contracting edge 3
Contracting edge 4
1
2
1
1
3
1
2
1
1
3
2
21
34
5
1
Contracting edge 3
2
1
1
34 5
2
6
Deleting edge 3
Merge two sinks
1
1Edge merge
2 & 22+ +
+
1
2
1
1
3
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 67 / 73
Summary of work thus far
Enumeration, Rate Region Computation, Forbidden Minors, New Networks
Enumerate > 7 ⇥ 105 non-isomorphic networks
Combination operations Embedding operations
Forbidden Minors for su�ciencyNew solvable networks in scale
Calculate their exact rate regions or bounds
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 68 / 73
Dream
Though we have onlinerepository
Want a user-friendly interface toeasily get answer
Draw or input a network
In Database?
Minimality & canonical conversion
Directly Computable?
Reachable by construction?
Get the answer
Not solvable right now
Yes
No
Yes
No
Yes
No
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 69 / 73
Future work
When Shannon outer bound is tight? Any common structure?
Is Shannon outer bound tight for all MDCS, or IDSC?
Is the number of forbidden minors regarding the sufficiency of a classof linear codes finite?
Coverage of the operators in all problems
More operations: node & edge merge, source & sink merge
A notion of forbidden minor which can harness both combination andembedding operators
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 70 / 73
Publication List
1 Steven Weber, Congduan Li, and John Walsh, Rate Region for a class of delay mitigatting codes and P2P networks,46th Annual Conference on Information Sciences and Systems, Princeton University, March, 2012, Invited Paper.
2 Congduan Li, John Walsh, and Steven Weber, ”A Computational Approach for Determining Rate Regions and Codesusing Entropic Vector Bounds”, 50th Allerton Conference, UIUC, Oct 2012.
3 Congduan Li, Jayant Apte, John M. Walsh, and Steven Weber, ”A new computational approach for determining rateregions and optimal codes for coded networks”, The 2013 IEEE International Symposium on Network Coding (NetCod2013), Calgary, Canada, Jun 7-9, 2013.
4 Congduan Li, John M. Walsh, and Steven Weber, ”Matroid bounds on region of entropic vectors”, 51th Allertonconference on Communication, Control and Computation, UIUC, IL, Oct 2013.
5 Jayant Apte, Congduan Li, John MacLaren Walsh, and Steven Weber, Exact pepair problems with multiple sources, inThe 48th annual IEEE Conference on Information Sciences and Systems(CISS), Mar. 2014.
6 Jayant Apte, Congduan Li, and John MacLaren Walsh, Algorithms for Computing Network Coding Rate Regions viaSingle Element Extensions of Matroids, IEEE International Symposium on Information Theory (ISIT) 2014, Honolulu,Hawaii.
7 Congduan Li, Steven Weber, and John M. Walsh, ”Network Embedding Operations Preserving the Insufficiency ofLinear Network Codes”, 52th Allerton conference on Communication, Control and Computation, UIUC, IL, Oct 2014.
8 Congduan Li, Steven Weber, and John M. Walsh, ”Multilevel Diversity Coding Systems: Rate Regions, Codes,Computation, & Forbidden Minors”, IEEE Transactions on Information Theory, 2014, submitted.
9 Congduan Li, Steven Weber, and John M. Walsh, ”Computer Aided Proofs for Independent Distributed Source CodingProblems”, The 2015 IEEE International Symposium on Network Coding (NetCod 2015), Jun, 2015.
10 Congduan Li, Steven Weber, and John M. Walsh, ”Network Combination Operations Preserving the Sufficiency ofLinear Network Codes”, The 2015 IEEE Information Theory Workshop (ITW 2015), submitted.
11 Congduan Li, Steven Weber, and John M. Walsh, ”On Multi-source Networks: Enumeration, Rate Region Computation,and Hierarchy”, IEEE Transactions on Information Theory, 2015, ready to submit.
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 71 / 73
Selected References
X. Yan, R.W. Yeung, and Z. Zhang (2012)
An Implicit Characterization of the Achievable Rate Region for Acyclic MultisourceMultisink Network Coding
IEEE Transactions on Information Theory 58(9), 5625-5639.
A. Betten, M. Braun, H. Fripertinger (2006)
Error-Correcting Linear Codes: Classification by Isometry and Applications
Springer Berlin Heidelberg.
N. Robertson, P. Seymour (2004)
Graph minors. xx. wagner’s conjecture
Journal of Combinatorial Theory, Series B, 92(2), 325-357.
J.G. Oxley (2011)
Matroid Theory
Oxford University, 2011.
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 72 / 73
Q & A
Thank you!
C. Li (ASPITRG & MANL) Ph. D. Dissertation Defense May 29, 2015 73 / 73