symmetries in entropy space - ece.drexel.eduece.drexel.edu/walsh/walsh_itw_2016.pdf ·...

18
Symmetries in Entropy Space Jayant Apte, Qi Chen, John MacLaren Walsh Department of Electrical and Computer Engineering Drexel University Philadelphia, PA [email protected] Thanks to NSF CCF-1421828 1

Upload: hoangliem

Post on 18-Jun-2018

214 views

Category:

Documents


0 download

TRANSCRIPT

Symmetries in Entropy Space

Jayant Apte, Qi Chen, John MacLaren Walsh

Department of Electrical and Computer Engineering

Drexel University

Philadelphia, PA

[email protected]

Thanks to NSF CCF-1421828

1

Co-Authors & Software Advertisement

Jayant Apte (Ph.D)Defended 10-Aug-2016

Dissertation presents symmetry exploiting algorithms for proving network coding capacity regions.

Try our software: ITAP & ITCP

Qi Chen, Ph.D. Ph.D. CUHK w/ R. W. Yeung

Postdoc @ Drexel Fall, 2015

Postdoc @ INC, CUHK, 2016

On the Job Market!

NEW IT&NC SOFTWARE TOOL RELEASES!!!! (ASPITRG homepage & GitHub)Information Theoretic Achievability Prover (ITAP):For a given network, determine entire rate region associated w/ a class of linear codes!Also, can determine achievability of a specified rate point, bounds for linear secret sharing,and test where a specified polymatroid is linear over a specified field.Information Theoretic Converse Prover (ITCP):For a given network coding problem, computes polyhedral cones outer bounding its capacityregion using a custom new symmetry exploiting polyhedral projection algorithm!Both Shannon & non-Shannon outer bounds are supported. Also works with secret sharing.

http://www.ece.drexel.edu/walsh/aspitrg/software.html

Outline

1. Entropy Region ¯

⇤N

& Shannon Outer Bound �N

2. Portion Symmetric Under Group G, ¯

⇤G

and G

3. Power-Set Orbit Structures & Hierarchical Results

4. ¯

⇤G

?=

G

for N = 4, 5 (exhaustive) and N � 6.

Entropy Region ¯

⇤N

& Shannon Outer Bound �N

1. X = (X1, . . . , XN

) N discrete RVs

2. every subset XA = (Xi

, i 2 A) A ✓{1, . . . , N} ⌘ [N ] has joint entropy

h(XA).

3. h = (h(XA) |A ✓ [N ] ) 2 R2N �1en-

tropic vector

• Example: for N = 3, h =

(h1, h2, h3, h12, h13, h23, h123).

4. a ho

2 R2N �1 is entropic if 9 joint PMF

pX s.t. h(pX) = ho

.

5. Region of entropic vectors = �

⇤N

6. Closure ¯

⇤N

is a convex cone

h2

h1

h12

Shannon bound �N

: h 2 R2N �1 s.t.

hA + hB � hA[B + hA\B (1)

hA[B � hA (2)

i.e. submodular & non.-dec.

for N 2 {2, 3}, �n

=

¯

⇤N

,

but for N � 4, ¯

⇤N

( �N

.

¯

⇤N

is an unknown non-polyhedral convex cone for N � 4.

determining capacity regions of all networks under network coding , complete

characterization of ¯

⇤N

Entropy Region ¯

⇤N

– Why Care?

• All fundamental laws of information theory

• Network Coding Capacity Region (Streaming, Distributed Storage, & Caching)

• Implications among conditional independences (Graphical Models & Machine

Learning)

• Fundamental Limits for Secret Sharing Schemes

• More (inequalities for subgroup sizes, matrix rank inequalities, Kolmogorov

Complexities, etc.)

Group Action & Symmetries

h2

h1

h12

(1,0,1)

(0,1,1)

⇡ = (12)

Any permutation ⇡ 2 Sn

, ⇡ : {1, . . . , n} ! {1, . . . , n} a bijection, is a symmetry of ¯

⇤n

.

Under ⇡, the RVs (X1, . . . , Xn

) 7! (X⇡(1), . . . , X⇡(n)),

so hA 7! h⇡(A), with ⇡(A) := {⇡(i)|i 2 A}.

Example: ⇡ = (12),

⇡([h1, h2, h12, h3, h13, h23, h123]T

) = [h2, h1, h12, h3, h23, h13, h123]T .

Clearly, if h 2 ¯

⇤n

or �n

, ⇡(h).

(definition is insensitive to ordering of RVs)

Key Question: ¯

⇤G

?=

G

S2 =

⇤S2

Fix

S2

h1

h2

h12

What about those h 2 ¯

⇤n

fixed under a group of problem symmetries G ✓ Sn

?

Fix

G

(H) := {h 2 H|⇡(h) = h 8⇡ 2 G} (3)

Define G

= Fix

G

(�

n

) and ⇤G

= Fix

G

(�

⇤n

).

For what types of problem symmetries can we expect Shannon-type inequalities to be

su�cient? i.e., for which G does ¯

⇤G

=

G

?

Orbits in the Power Set

N = {1, . . . , n}

Fix

G

, and hence ¯

⇤G

and G

on depend on G through 2

N //G – the orbits in the

power set.

• ⇡(h) = h 8⇡ 2 G () hA = h⇡(A)8⇡ 2 G, 8A ✓ N .

• OG

(A) := {⇡(A)|⇡ 2 G}. hA = hB 8B 2 OG

(A)

• OG

=

�O

G

(A)|A 2 2

N , the power set orbits. (Partitions 2

N )

Note: Multiple groups can yield the same power set orbits.

i.e. can have OG

= OG

0 for G 6= G0.

Power Set Orbits, N = 4

{;}

O(1)

O(12)

O(123)

{N}

(a) OS4 = OA4

{;}

O(1)

O(12)

O(123)

{N}

O(13)

(b) OD4 = OC4

{;}

O(1)

O(123)

{N}

O(13)

O(14) O(12)

(c) OV =h(12)(34),(13)(24)i

{;}

O(1)

O(123)

{N}

O(3)

O(134)

O(13)O(14)

O(12)O(34)

(d) OS24=h(12)(34)i

Figure 1: Indecomposable power set orbits on N = {1, 2, 3, 4}

Power Set Orbits under di↵erent groups are also ordered by refinement of associated

partition of the power set.

Power Set Orbits, N = 5

{;}

O(1)

O(12)

O(123)

O(1234)

{N}

(a) OS5 = OA5 = OGA(1,5)

{;}

O(1)

O(12)

O(123)

O(1234)

{N}

O(13)

O(124)

(b) OD5 = OC5

Figure 2: Indecomposable orbit structures on N = {1, 2, 3, 4, 5}

Observe that while a poset, need not be a lattice.

¯

⇤G

?=

G

– Some Implications

• If G G0, then OG

OG

0 , i.e. partition OG

refines OG

0

– (extra group elements can force more equivalences)

Thm. 2: If OG

OG

0 , and ⇤G

=

G

, then ⇤G

0 =

G

0 .

• Equiv., if OG

OG

0 , and ⇤G

0 ( G

0 , then ⇤G

( G

.

Chen & Yeung – Partition Symmetrical Entropy Functions:¯

⇤G

?=

G

for G = Sn1 ⇥ S

n2 ⇥ · · · ⇥ Snk

Thm. 1 (Qi Chen & R. W. Yeung): Let p = {N1, N2, . . . , Nt

} be a t-partition of N ,

and Gp

= SN1 ⇥ SN2 ⇥ · · · ⇥ SNt . For |N | � 4,

⇤Gp

=

Gp if and only if p = {N} or

{{i}, N \ {i}} for some i 2 N .

What about general ¯

⇤G

?=

G

? Remaining cases?

• For any G, let p = N//G = {{⇡(i)|⇡ 2 G}|i 2 N}, then G Gp

– Thm. 1: ¯

⇤G

( G

unless N//G = {N} or N//G = {i, N \ {i}}, i 2 N .

• N//G = {N}: G is transitive.

• N//G = {i, N\}: G = S1 ⇥ G0,

– G fixes some i, then G0 on remainder is transitive on N \ {i}.

• Priority (Pessimistic) – Maximal Transitive Subgroups

• Priority (Optimistic) – Minimal Transitive Subgroups

¯

⇤G

?=

G

– Another Implication

Thm. 3: Let G ⇥ S1 act on N [ {n + 1} (fix n + 1, G acts on N ). Then

⇤S1⇥G

=

S1⇥G

=)

⇤G

=

G

.

Proof:

• orbits of OG⇥S1 = each orbit in O

G

repeated twice: once w/o n + 1, and once w/

n + 1.

• projN G⇥S1 =

G

– ✓: First 2

N � 1 coordinates are same subsets obeying same inequalities, must

also have the remaining coordinates exist obeying more inequalities.

– ◆: Take h 2 G

, extend to h0 on N [ {n + 1} via h0A = hA\N , then

h0 2 G⇥S1 .

• projN ⇤G⇥S1

=

⇤G

– ✓: Take X1, . . . , Xn

from X1, . . . , Xn+1 realizing h0 2 ⇤G⇥S1

. Realizes a

h 2 ⇤G

.

– ◆: Extend as previous, Xn+1 = 0, deterministic.

¯

⇤G

?=

G

– Complete Answer for N = 4

Symmetric & Alternating:

S4 : �(1234), (12)�A4 : �(123), (12)(34)�

Dihedral & Cyclic:

D4 : �(1234), (13)�C4 : �(1234)�

Normal Klein 4-group:

V : �(12)(34), (13)(24)�

Double Transp.:

S24 : �(12)(34)�

(1, 3)-Partition:

S1 ⇥ S3 : �(234), (23)�S1 ⇥ A3 : �(234), (243)�

(2, 2)-Partition:

S2 ⇥ S2 : �(12), (34)�

(1, 1, 2)-Partition:

S1 ⇥ S1 ⇥ S2 : �(34)�

Trivial: �()�

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

=

⇤G

G

=

⇤G

G

=

⇤G

G

=

⇤G

Power Set Orbits, N = 4

{;}

O(1)

O(12)

O(123)

{N}

(a) OS4 = OA4

{;}

O(1)

O(12)

O(123)

{N}

O(13)

(b) OD4 = OC4

{;}

O(1)

O(123)

{N}

O(13)

O(14) O(12)

(c) OV =h(12)(34),(13)(24)i

{;}

O(1)

O(123)

{N}

O(3)

O(134)

O(13)O(14)

O(12)O(34)

(d) OS24=h(12)(34)i

Figure 1: Indecomposable power set orbits on N = {1, 2, 3, 4}

Power Set Orbits under di↵erent groups are also ordered by refinement of associated

partition of the power set.

Power Set Orbits, N = 4

{;}

O(1)

O(12)

O(123)

{N}

(a) OS4 = OA4

{;}

O(1)

O(12)

O(123)

{N}

O(13)

(b) OD4 = OC4

{;}

O(1)

O(123)

{N}

O(13)

O(14) O(12)

(c) OV =h(12)(34),(13)(24)i

{;}

O(1)

O(123)

{N}

O(3)

O(134)

O(13)O(14)

O(12)O(34)

(d) OS24=h(12)(34)i

Figure 1: Indecomposable power set orbits on N = {1, 2, 3, 4}

Power Set Orbits under di↵erent groups are also ordered by refinement of associated

partition of the power set.

Power Set Orbits, N = 4

{;}

O(1)

O(12)

O(123)

{N}

(a) OS4 = OA4

{;}

O(1)

O(12)

O(123)

{N}

O(13)

(b) OD4 = OC4

{;}

O(1)

O(123)

{N}

O(13)

O(14) O(12)

(c) OV =h(12)(34),(13)(24)i

{;}

O(1)

O(123)

{N}

O(3)

O(134)

O(13)O(14)

O(12)O(34)

(d) OS24=h(12)(34)i

Figure 1: Indecomposable power set orbits on N = {1, 2, 3, 4}

Power Set Orbits under di↵erent groups are also ordered by refinement of associated

partition of the power set.

Power Set Orbits, N = 4

{;}

O(1)

O(12)

O(123)

{N}

(a) OS4 = OA4

{;}

O(1)

O(12)

O(123)

{N}

O(13)

(b) OD4 = OC4

{;}

O(1)

O(123)

{N}

O(13)

O(14) O(12)

(c) OV =h(12)(34),(13)(24)i

{;}

O(1)

O(123)

{N}

O(3)

O(134)

O(13)O(14)

O(12)O(34)

(d) OS24=h(12)(34)i

Figure 1: Indecomposable power set orbits on N = {1, 2, 3, 4}

Power Set Orbits under di↵erent groups are also ordered by refinement of associated

partition of the power set.

¯

⇤G

?=

G

– Complete Answer for N = 5

Symmetric, Alternating, & Gen. A�ne:

S5 : �(12345), (12)�, A5 : �(12345), (123)�GA(1, 5) : �(12345), (2345)�

Cyclic & Dihedral:

C5 : �(12345)�D5 : �(12345), (25)(34)�

(2, 3)-Partition:

S2 ⇥ S3 : �(12), (345), (34)�S2 ⇥ A3 : �(12), (345)�S3

5 : �(12)(45), (345)�

(1, 4)-Partition:

S1 ⇥ S4 : �(2345), (23)�S1 ⇥ A4 : �(2345), (23)(45)�

Fix 1, Rest Cyclic or Dihedral:

S1 ⇥ C4 : �(2345)�S1 ⇥ D4 : �(2345), (24)�

Fix 1, Rest Klein 4-group:

S1 ⇥ V : �(23)(45), (24)(35)�

(1, 1, 3)-partition:

S1 ⇥ S1 ⇥ S3 : �(345), (34)�S1 ⇥ S1 ⇥ A3 : �(345)�

(1, 2, 2)-partition:

S1 ⇥ S2 ⇥ S2 : �(23), (45)�

(1, 1, 1, 2)-partition:

S1 ⇥ S1 ⇥ S1 ⇥ S2 : �(45)�

trivial: �()�

S1 ⇥ S24 : �(23)(45)�

G

=

⇤G

G

=

⇤G

G

=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

Power Set Orbits, N = 5

{;}

O(1)

O(12)

O(123)

O(1234)

{N}

(a) OS5 = OA5 = OGA(1,5)

{;}

O(1)

O(12)

O(123)

O(1234)

{N}

O(13)

O(124)

(b) OD5 = OC5

Figure 2: Indecomposable orbit structures on N = {1, 2, 3, 4, 5}

Observe that while a poset, need not be a lattice.

¯

⇤G

?=

G

– Some Answers for N � 6

Thm. 5: For n � 6,

• ⇤Cn

( Cn ,

• ⇤Dn

( Dn ,

• ⇤S1⇥Cn�1

( S1⇥Cn�1 , &

• ⇤S1⇥Dn�1

( S1⇥Dn�1 .

Proof shows that the ray with coordinates matching Vamos projection lies in Cn but

is cut o↵ by Zhang-Yeung non-Shannon inequality.

Summary

Symmetric, Alternating, & Gen. A�ne:

S5 : �(12345), (12)�, A5 : �(12345), (123)�GA(1, 5) : �(12345), (2345)�

Cyclic & Dihedral:

C5 : �(12345)�D5 : �(12345), (25)(34)�

(2, 3)-Partition:

S2 ⇥ S3 : �(12), (345), (34)�S2 ⇥ A3 : �(12), (345)�S3

5 : �(12)(45), (345)�

(1, 4)-Partition:

S1 ⇥ S4 : �(2345), (23)�S1 ⇥ A4 : �(2345), (23)(45)�

Fix 1, Rest Cyclic or Dihedral:

S1 ⇥ C4 : �(2345)�S1 ⇥ D4 : �(2345), (24)�

Fix 1, Rest Klein 4-group:

S1 ⇥ V : �(23)(45), (24)(35)�

(1, 1, 3)-partition:

S1 ⇥ S1 ⇥ S3 : �(345), (34)�S1 ⇥ S1 ⇥ A3 : �(345)�

(1, 2, 2)-partition:

S1 ⇥ S2 ⇥ S2 : �(23), (45)�

(1, 1, 1, 2)-partition:

S1 ⇥ S1 ⇥ S1 ⇥ S2 : �(45)�

trivial: �()�

S1 ⇥ S24 : �(23)(45)�

G

=

⇤G

G

=

⇤G

G

=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

Symmetric & Alternating:

S4 : �(1234), (12)�A4 : �(123), (12)(34)�

Dihedral & Cyclic:

D4 : �(1234), (13)�C4 : �(1234)�

Normal Klein 4-group:

V : �(12)(34), (13)(24)�

Double Transp.:

S24 : �(12)(34)�

(1, 3)-Partition:

S1 ⇥ S3 : �(234), (23)�S1 ⇥ A3 : �(234), (243)�

(2, 2)-Partition:

S2 ⇥ S2 : �(12), (34)�

(1, 1, 2)-Partition:

S1 ⇥ S1 ⇥ S2 : �(34)�

Trivial: �()�

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

6=

⇤G

G

=

⇤G

G

=

⇤G

G

=

⇤G

G

=

⇤G

n = 4 :

n = 5 :

n � 6 :

Cn 6=

⇤Cn

,

Dn 6=

⇤Dn

{1, . . . , n}//G not a (1, n � 1) or (n) partition =)

G

6=

⇤G

S1⇥Cn�1 6=

⇤S1⇥Cn�1

,

S1⇥Dn�1 6=

⇤S1⇥Dn�1

Fig. 1(a)

Fig. 1(b)

Fig. 1(c)

Fig. 1(d)

Fig. 2(a)

Fig. 2(b)

Fig. 3

Sn =

⇤Sn

, S1⇥Sn =

⇤S1⇥Sn

The Way Forward

1. For n � 6, answer ¯

⇤G

?=

G

for other maximal transitive groups.

2. Permutations of the ground set form just one set of interesting symmetries for �N

,

others include

• Combinatorial Symmetry Group: permutations of the extreme rays which

leave the face lattice intact. (Huge – Subgroup of SM

, w/ M= # of extreme

rays)

• A�ne Symmetry Group: Those Combinatorial symmetries whose ray

permutation can be generated by multiplying by a 2

N � 1 ⇥ 2

N � 1 invertible

matrix. (Also a large group)

• Restricted A�ne Symmetry Group: Those a�ne symmetries associated with

vectors representing the rays of fixed length. (can be computed w/ sympol)

3. Which of the latter are also symmetries of

¯

⇤N

?

4. What sort of dimensionality reduction can be achieved by exploiting the a�ne

symmetries when calculating rate regions for network coding, storage repair

tradeo↵s, caching regions, etc?