Two Examples of Submodularity in Wireless Communications

Preview:

Citation preview

Two Examples ofSubmodularity in WirelessCommunications

Ni Ding

13 June 2017

www.data61.csiro.au

Outline

IntroductionSubmodular FunctionTarski Fixed Point TheoremSubmodular (Set) Function Minimization

Adaptive Modulation in Network-coded Two-way Relay ChannelSystemTwo-player Game ModelPure Strategy Nash EquilibriumCournot Tatonnement

Communication for OmniscienceSystemMinimum Sum-ratePrincipal Sequence of PartitionsModified Decomposition AlgorithmExtensions: Secret Capacity, Clustering and Data Compression

Conclusion

2 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Submodularity

• a property of functions defined on lattice [1]1 [2]2.

I lattice: fundamental algebraic structure on partial order.

• applications: economics, machine learning, operations research.

Study on Machine Learning in [3]3:

Submodularity imposes a structure which allows much strongermathematical results than we would be able to achieve without it.

• submodularity on

I vector lattice: discrete convexity and comparative staticsI set lattice: combinatorial optimization, e.g., in graph theory

1Topkis 20012Murota 20053Vondrak 2007

3 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Lattice

Poset

For a set L and a binary order , (L,) is a poset (partially ordered set)if either a b or a 6 b,∀a, b ∈ L.

examples: (RN ,≤), (1, . . . , 4N ,≤), (2V ,⊆) and (1, 1, 2, 3,⊆)

Lattice

A poset (L,) is a lattice with notation (L,∨,∧) ifa ∨ b = supa, b ∈ L and a ∧ b = infa, b ∈ L,∀a, b ∈ L with sup andinf w.r.t.

• maximum∨L = supL and minimum

∧L = inf L exist;

examples: (RN ,∨,∧), (1, . . . , 4N ,∨,∧) with r ∨ r′ = (maxri , r ′i : i ∈ 1, . . . ,N)and r ∧ r′ = (minri , r ′i : i ∈ 1, . . . ,N); (2V ,∪,∩) and (1, 1, 2,∪,∩)

4 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Submodular Function

Submodularity

f : (L,∨,∧) 7→ R is submodular if

f (a) + f (b) ≥ f (a ∨ b) + f (a ∧ b), ∀a, b ∈ L.

f is supermodular if −f is submodular.

5 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Tarski Fixed Point TheoremN-player game model Ω = N , Ai , ci (a)i∈N witha ∈ A = ×i∈NAi ⊆ RN :

• pure strategy Nash equilibrium (PSNE): best response functionψ : A 7→ A with ai ∈ Ai , a−i ∈ ×i ′∈N\iAi ′ and

ψi (a−i ) ∈ arg minci (ai , a−i ) : ai ∈ Ai

, ∀i ∈ N .

a∗ is an PSNE if a∗ = ψ(a∗), i.e., a∗ is a fixed point of ψ.

• question: PSNE exists for discrete A? supermodular game withstrategic complements: ψ : (A,∨,∧) 7→ (A,∨,∧) is non-decreasingif ci is submodolar ∀i .

Tarski Fixed Point Theorem [4]4

The fixed points of non-decreasing ψ : (A,∨,∧) 7→ (A,∨,∧) form a(nonempty) lattice.

4Tarski et. al 1955

6 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Submodular (Set) FunctionMinimization

For f : (2V ,∪,∩) 7→ R, consider

minf (X ) : X ⊆ V (1)

combinatorial optimization: NP-complete or NP-hard in general

SFM (submodular function minimization) algorithm

If f is submodular, i.e.,

f (X ) + f (Y ) ≥ f (X ∪ Y ) + f (X ∩ Y ), ∀X ,Y ⊆ V ,

(1) can be solved in polynomial time and the minimizers form a lattice:⋃argminf (X ) : X ⊆ V and

⋂argminf (X ) : X ⊆ V exist [5]5.

5Fujishige 2005

7 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Adaptive Modulation inNetwork-coded Two-way RelayChannel

network-coded two-way relay channel (NC-TWRC): two userscommunicate via a center node, relay ‘R’.

user 1wireless

Rwireless

user 2

physical-layer network coding (PNC): messages x1 and x2 transmittedsimultaneously in phase I, the superposition z broadcast in phase II.

user 1x1

Rx2

user 2

phase I: multiple access (MAC)

user 1z

Rz

user 2

phase II: amplify and forward (AF)

8 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Adaptive Modulation inNC-TWRC

assumption: m-quadrature amplitude modulation (m-QAM) adopted byeach user:

• constellation size mi = 2ai of user i with ai ∈ 0, 1, . . . , thenumber of bits/symbol, determined by user i

Strategic Complements

increasing best response:

• spectral efficiency: one tends to transmit while the other does so

• equal share of the channel: one increases ai while the other−i = 1, 2 \ i does so

proposal: two-player game model parameterized by user-to-user channelsignal-to-noise ratios (SNRs)

9 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Two-player Game ModelΩγ = N , Γ, Ai , ci (γi , a)i∈N :• N = 1, 2;• γ = (γ1, γ2) ∈ Γ = Γ1 × Γ2 with γi ∈ Γi being SNR of user i-to-user−i channel determined by PNC scheme;

• a = (a1, a2) ∈ A = A1 ×A2 = 0, 1, . . . ,Am2 with Am being themaximum number of bits/symbol

• cost function: ci : Γi ×A 7→ R+

ci (γi , a) = ce(γi , ai ) + cr (a)

with the cost associated with transmission error rate

ce(γi , ai ) =− ln(5Pb)(2ai − 1)

1.5γi

and the cost associated with spectral efficiency and fairness

cr (a) =a−i + 1

ai + 1

10 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Pure Strategy Nash Equilibrium(PSNE)

Submodularity

ci : (A,∨,∧) 7→ R+ is submodular, i.e.,

ci (γi , a) + ci (γi , a′) ≥ ci (γi , a ∨ a′) + ci (γi , a ∧ a′),

for all a, a′ ∈ (A,∨,∧), i ∈ N and γ ∈ Γ.

Tarski Fixed Point Theorem =⇒ Existence of PSNE

Pure strategy θ : Γ 7→ A, where θ(γ) = (θ1(γ), θ2(γ)) with θi (γ) ∈ Ai

being the pure strategy of user i when SNRs are γ = (γ1, γ2):

• PSNE θ∗ exists

• The largest PSNE θ∗

and the smallest PSNEs θ∗ exist

11 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Cournot Tatonnementdetermine extremal PSNEs: Cournot tatonnement [6]6

• Let ψ : Γ×A 7→ A and ψ : Γ×A 7→ A be the maximal andminimal best response functions, respectively, with

ψi (γ, a−i ) =∨

arg minai∈Ai

ci (γi , ai , a−i )

ψi(γ, a−i ) =

∧arg min

ai∈Ai

ci (γi , ai , a−i )

• recursions with θ(0)

(γ) =∨A and θ(0)(γ) =

∧A:

θ(γ) := ψ(γ,θ(γ))

θ(γ) := ψ(γ,θ(γ))

Convergence

θ(k)(γ) and θ(k)(γ) converge monotonically downward and upward

to θ∗(γ) and θ∗(γ), respectively, for all γ.

6Vives 1990

12 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Experiment Iperformance of extremal PSNEs:A = 0, 1, . . . , 92 and simulation lasts for 104 symbol durations:

extremal PSNEs θ∗

and θ∗ are compared to the single-agent adaptivemodulation (Single-AM) and 2-QAM scheme.

−6 −4 −2 0 2 4 610−5

10−4

10−3

10−2

10−1

100

γ(dB)

biterrorrate

(BER)

θ∗

θ∗

Single-AM2-QAM

(a) bit error rate (BER)

−6 −4 −2 0 2 4 60

2

4

6

8

10

γ(dB)

bits

per

symbol

duration

θ∗

θ∗

Single-AM2-QAM

(b) spectral efficiency

13 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Experiment IIexample of Cournot tatonnement

A = 0, 1, . . . , 92 and the sequences θ(k)

1 (γ) and θ(k)1 (γ) generated

by Cournot tatonnement for certain γ

1 2 3 40

2

4

6

8

iteration index k

strategy

ofschedu

ler1 θ

(k)1 (γ)

θ∗1(γ)

θ(k)1 (γ)θ∗1(γ)

14 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Communication for Omniscience

indices of users: a finite ground set V with |V | > 1discrete correlated random source: ZV = (Zi : i ∈ V )

• user i observes an i.i.d. n-sequence Zni of Zi in private

communication for omniscience (CO) [7]7:

• users exchange Zi s directly over noiseless broadcast channels

• goal: attain omniscience, the state that each user recovers ZnV

Minimum Sum-rate Problem

how to attain omniscience with RCO(V ), the minimum total number oftransmissions: value of RCO(V ) and an optimal rate vectorr∗V = (r∗i : i ∈ V )

7Csiszar et. al 2004: CO formulated based on the study on secret capacity

15 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Example: Coded CooperativeData Exchange (CCDE)

client 1

Z1 = [Wa,Wb,Wc ,Wd ,We ]ᵀ

client 2

Z2 = [Wa,Wb,Wf ]ᵀ

client 3

Z3 = [Wc ,Wd ,Wf ]ᵀ

3-mobile clients in V = 1, 2, 3; Zi : partial observation of a packet setwith Wj denoting a packet

Solutions to Minimum Sum-rate Problem

RCO(V ) = 72 and r∗V = (r∗1 , r

∗2 , r∗3 ) = ( 5

2 ,12 ,

12 ): by packet-splitting

Wj =⇒W(1)j ,W

(2)j ; transmit (r1, r2, r3) = (5, 1, 1) with ri denote the

number of linear combinations of packet chunks W(k)j .

16 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Omniscience-achievabilityFor X ⊆ V : r(X ) =

∑i∈X ri for rV = (ri : i ∈ V )

H(X ): the amount of randomness in ZX measured by Shannon entropy

Omniscience-achievability [7]8

An omniscience-achievable rV satisfies the Slepian-Wolf (SW) constraint:r(X ) ≥ H(X |V \ X ) = H(V )− H(V \ X ),∀X ( V .

achievable rate vector set:

R(V ) = rV ∈ R|V | : r(X ) ≥ H(X |V \ X ),∀X ( V

minimum sum-rate:

RCO(V ) = minr(V ) : rV ∈ R(V )

constant sum-rate set: Rα(V ) = rV ∈ R(V ) : r(V ) = αoptimal rate vector set: RRCO(V )(V )

8Csiszar et. al 2004

17 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Nonemptiness of BasePolyhedron

For α ∈ R+, let

fα(X ) =

H(X |V \ X ) X ( V

α X = V.

polyhedron: P(fα,≥) = rV ∈ R|V | : r(X ) ≥ fα(X ),∀X ⊆ V base polyhedron: B(fα,≥) = rV ∈ P(fα,≥) : r(V ) = fα(V ) = α

• B(fα,≥) = Rα(V ) 6= ∅ ⇐⇒ ∃ achievable rV with r(V ) = α

dual set function: f #α (X ) = fα(V )− fα(V \ X )

• B(fα,≥) = B(f #α ,≤) [5]9;

why consider B(f #α ,≤)? f #

α is intersecting submodular, i.e.,

f #α (X ) + f #

α (Y ) ≥ f #α (X ∪ Y ) + f #

α (X ∩ Y ), ∀X ,Y : X ∩ Y 6= ∅.9Fujishige 2005

18 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Minimum Sum-rateΠ(V ): the set of all partitions of V and Π′(V ) = Π(V ) \ V achievability of α: B(f #

α ,≤) 6= ∅ iff α = minP∈Π(V )

∑C∈P f #

α (C ) [5]10

Minimum Sum-rate

RCO(V ) = maxP∈Π′(V ) φ(P) with the finest maximizer P∗. Here,

φ(P) =∑C∈P

H(V \ C |C )

|P| − 1.

interpretation: ∀C ∈ P, the cut C ,V \ C imposes SW constraintr(V \ C ) ≥ H(V \ C |C ) so that∑

C∈P

r(V \ C ) = (|P| − 1)r(V ) ≥∑C∈P

H(V \ C |C )

A multi-way cut P ∈ Π′(V ) imposes r(V ) ≥ φ(P).

10Fujishige 2005: minP∈Π(V )

∑C∈P f #

α (C) is called the Dilworth truncation of f #α .

19 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Principal Sequence of Partitions

Principal Sequence of Partitions (PSP) [8]11:

minP∈Π(V )

∑C∈P f

#α (C ) is a piecewise linear increasing curve in α

that is fully characterized by p ≤ |V | − 1 critical points

H(V ) = α0 > α1 > α2 > . . . > αp ≥ 0.

Let Pj be the finest minimizer of minP∈Π(V )

∑C∈P f

#α (C ).

P0 P1 P2 . . . Pp

where P P ′ denotes P ′ is strictly finer than P.

The first critical point determines the solutions to the minimumsum-rate problem: RCO(V ) = α1,P∗ = P1.

11Nagano et. al 2010

20 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Example of PSP

client 1

Z1 = [Wa,Wb,Wc ,Wd ,We ]ᵀ

client 2

Z2 = [Wa,Wb,Wf ]ᵀ

client 3

Z3 = [Wc ,Wd ,Wf ]ᵀ

0 1 2 3 4 5 6

−5

0

5

10

α1 = 72,P1 = 1, 2, 3

α0 = 0,P0 = 1, 2, 3

α

minP∈Π

(V)

∑ C∈P

f# α(C

)

minP∈Π(V )

∑C∈P f

#α (C )

α

PSP results:α0 > α1 and P0 P1:

• No omniscience-achievablerV if α < α1, because α 6=minP∈Π(V )

∑C∈P f #

α (C );

• RCO(V ) = α1 andP∗ = 1, 2, 3 = P1

21 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Properties of φ(P) in PSP

αj and Pj :

• If j = 1, αj = φ(Pj);

• When j > 1, let α = φ(Pj) and Pj′ be the finest minimizer ofminP∈Π(V )

∑C∈P f #

α (C ). Then,

αj < α < α1

j ′ < j =⇒ Pj′ Pj

Suggestion: A Recursive Algorithm

• iteratively updates α and P, the estimation of α1 = RCO(V ) andP∗ = P1;

• terminate when α = φ(P).

22 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Modified DecompositionAlgorithm

recursion in modified decomposition algorithm (MDA):

α := φ(P)

P(n) is the finest minimizer of minP∈Π(V )

∑C∈P f #

α(n) (C ) and

P(0) = i : i ∈ V

Optimality of the MDA algorithm

α(n) and P(n) converge monotonically towards α1 = RCO(V ) and

P1 = P∗, respectively. Also returns rV ∈ B(f #RCO(V ),≤) = RRCO(V )(V )

• minP∈Π(V )

∑C∈P f #

α(n) (C ) reduces to⋂argminf #

α(n) (X )− r(X ) : i ∈ X ⊆ V ,∀i ∈ V , SFM due to the

intersecting submodularity of f #α(n)

• complexity: O(|V |2 · SFM(|V |))12

12SFM(|V |): the complexity of minimizing submodular function f : 2V 7→ R.

23 | Two Examples of Submodularity in Wireless Communications | Ni Ding

ExperimentV = 1, . . . , 5: Wm is an independent uniformly distributed random bit:

Z1 = (Wb,Wc ,Wd ,Wh,Wi ),

Z2 = (We ,Wf ,Wh,Wi ),

Z3 = (Wb,Wc ,We ,Wj ),

Z4 = (Wa,Wb,Wc ,Wd ,Wf ,Wg ,Wi ,Wj ),

Z5 = (Wa,Wb,Wc ,Wf ,Wi ,Wj ),

0 1 2 3 4 5 6 7 8 9 10

−20

−10

0

10

P3 = 1, 2, 3, 4, 5

P2 = 4, 5, 1, 2, 3

P1 = 1, 4, 5, 2, 3 = P∗P0 = 1, 2, 3, 4, 5

α

minP∈Π

(V)

∑ C∈P

f# α(C

)

minP∈Π(V )

∑C∈P f

#α (C )

α

(c) PSP

0 1 2 3

5.8

6

6.2

6.4

6.6

iteration index

α

α(n)RCO(V )

(d) α(n) by MDA algorithm

24 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Extensions of CO: SecretCapacity

secret capacity CS(V ): the maximum rate at which the secret key can begenerated by the users in V with results in [7]13:

• dual relationship: RCO(V ) = H(V )− CS(V )

• mutual dependence upper bound on CS(V ):

CS(V ) ≤ I (V ) = minP∈Π′(V )

∑C∈P H(C )− H(V )

|P| − 1︸ ︷︷ ︸mutual dependence in ZV

tightness [9]14: CS(V ) = I (V ) = H(V )− RCO(V )question: how to achieve CS(V )? with interactive communication rater(V ) = RCO(V )? silly! CS(V ) can be attained with r(V ) ≤ RCO(V ) [7]

13Csiszar et. al 200414Chan et. al 2015

25 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Extensions of CO: Clustering

Inspired by the name ‘mutual dependence’

I (V ) = minP∈Π′(V )

∑C∈P H(C )− H(V )

|P| − 1

is proposed in [9]15 as a generalization of Shannon’s mutual informationto multivariate case: I (V ) = H(1) + H(2)− H(1, 2) whenV = 1, 2.

• realization: I (V ) is the similarity measure of more than two rvs.

I limitation in existing clustering algorithms: pairwisesimilarity/dissimilarity measure

I agglomerative clustering result given by PSP determined inO(|V |2 · SFM(|V |)) time

question: can PSP clustering frame work provide more objective overviewof the dataset?

15Chan et. al 2015

26 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Extensions of CO (Digiscape):Source Coding with SideInformation

1

Zn1

2

Zn2

. . . . . . |V |

Zn|V |

T

sensor nodes i ∈ V = 1, . . . , |V | reveal all information to sink T .

• for lossless data compression/aggregation, SW constraints:

r(X ) ≥ H(X |V \ X ),∀X ( V , r(V ) = H(V ) =⇒ RH(V )(V )

• an extreme, one of the unfairest, rV ∈ RH(V )(V ) can be determinedin O(|V |) time

• question: how to find a fair rate allocation in RH(V )(V ) efficiently?27 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Conclusion

two examples of submodularity in wireless communications:

• vector lattice: the existence of PSNEs in a game modeledadaptive modulation problem in NC-TWRC

• set lattice: polynomial time algorithm for solving CO problem

future:

• vector lattice: more applications of discrete convexity, e.g.,the energy-delay trade-off in data aggregation tree inDigiscape, and monotone comparative statics

• set lattice: improving efficiency for determining PSPI less call of SFM algorithmI improving complexity SFM(|V |): SFM belongs to worst

polynomial algorithm category, e.g., |V |5 to |V |8

28 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Bibliography I

D. M. Topkis, Supermodularity and complementarity. Princeton: PrincetonUniversity Press, 2001.

K. Murota, “Note on multimodularity and l-convexity,” Math. Oper. Res.,vol. 30, no. 3, pp. 658–661, Aug. 2005.

J. Vondrak, “Submodularity in combinatorial optimization,” Ph.D. dissertation,Dept. Appl. Math., Charles Univ., Prague, 2007.

A. Tarski et al., “A lattice-theoretical fixpoint theorem and its applications,”Pacific J. Math., vol. 5, no. 2, pp. 285–309, 1955.

S. Fujishige, Submodular functions and optimization, 2nd ed. Amsterdam, TheNetherlands: Elsevier, 2005.

X. Vives, “Nash equilibrium with strategic complementarities,” J. Math. Econ.,vol. 19, no. 3, pp. 305 – 321, 1990.

I. Csiszar and P. Narayan, “Secrecy capacities for multiple terminals,” IEEETrans. Inf. Theory, vol. 50, no. 12, pp. 3047–3061, Dec. 2004.

29 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Bibliography II

K. Nagano, Y. Kawahara, and S. Iwata, “Minimum average cost clustering,” inProc. Advances in Neural Inf. Process. Syst., Vancouver, Candada, 2010, pp.1759–1767.

C. Chan, A. Al-Bashabsheh, J. Ebrahimi, T. Kaced, and T. Liu, “Multivariatemutual information inspired by secret-key agreement,” Proc. IEEE, vol. 103,no. 10, pp. 1883–1913, Oct. 2015.

30 | Two Examples of Submodularity in Wireless Communications | Ni Ding

Recommended