1 Fairness in Communication for Omniscience Ni Ding∗, Chung Chan‡, Qiaoqiao Zhou‡, Rodney A. Kennedy∗ and Parastoo Sadeghi∗ Abstract—Weconsidertheproblemofhow tofairlydistribute battery usage of some users. Fairness has been considered in the minimum sum-rate among the users in communication for [8]–[10]forCCDEandin[11]forCO.However,themethods omniscience (CO). We formulate a problem of minimizing a proposed in [8], [9] can only determine an integer valued weightedquadraticfunctionoverasubmodularbasepolyhedron fairest minimum sum-rate strategy in terms of Jain’s fairness which contains all achievable rate vectors, or transmission strategies, for CO that have the same sum-rate. By solving index [12], which may not be applied to general CO systems 6 it, we can determine the rate vector that optimizes the Jain’s where the transmission rates could be fractional; the method 1 fairness measure, a more commonly used fairness index than in [10] gives a fractional optimizer but relies on building a 0 the Shapley value in communications engineering. We show multi-layer hypergraph model which can be quite complex 2 that the optimizer is a lexicographically optimal (lex-optimal) for large scale systems; the Shapley value proposed in [11] base and can be determined by a decomposition algorithm n (DA) that is based on submodular function minimization (SFM) incurs exponentialcomplexity and does not coincide with the a algorithm and completes in strongly polynomial time. We prove Jain’sfairestsolution,the commonlyusedfairnessmeasurein J that the lex-optimal minimum sum-rate strategy for CO can be communications engineering. We will give example to show 7 determined by finding the lex-optimal base in each user subset the difference between Jain’s fairest solution and Shapley 2 in thefundamentalpartition and thecomplexity can bereduced value in this paper. accordingly. ] The main purpose of this paper is to determine the fairest T solution in the minimum sum-rate strategy set for CO. We I I. INTRODUCTION . consider the problem of minimizing a weighted quadratic s Communication for omniscience (CO) is a problem pro- function on a constraint set. The constraint set contains all c [ posed in [1]. It is assumed that there is a group of users achievable transmission strategies that have the same sum- in the system and each of them observes a component of a rate which is greater than or equal to the minimal one for 1 discrete memorylessmultiple source in private. The users can CO. This problem is equivalent to determining the Jain’s v 5 exchangetheir informationin order to attain omniscience, the fairest transmission strategy when all users are assigned the 8 state that each user obtains the total information in the entire same weight. We show that the constraint set of this problem 2 multiple source in the system. A typicalexampleis the coded is a submodular base polyhedron and the minimizer is the 7 cooperativedataexchange(CCDE)problem[2]whereagroup lexicographicallyoptimal(lex-optimal)base.Theproblemcan 0 ofgeographicallycloseuserscommunicatewitheachothervia be reduced to an SFM problem, and the lex-optimal base . 1 error-free broadcast channels in order to recover a packet set. can be determined by a decomposition algorithm (DA) in 0 The fundamental problem in CO and CCDE is how to O(|V|2·SFM(|V|)) time, where |V| is the cardinality of the 6 achieve omniscience with minimum total transmission rate, 1 user set and SFM is the complexityof minimizing a submod- or minimum sum-rate. The studies in [3], [4] show that the : ularfunctionwhichisstronglypolynomial.Wealsoshowthat v minimumsum-ratecanbeobtainedbysolvinganoptimization thelex-optimalminimumsum-ratestrategycanbedetermined i X problem over the partitions of user set, while the authors in by obtaining the lex-optimal base for each user subset in the [5],[6]proposepolynomialtimealgorithmsthatdeterminethe r fundamentalpartition [3], the optimalpartition corresponding a minimumsum-rateandacorrespondingstrategybasedonsub- to the minimum sum-rate. The task of determining the lex- modular function minimization (SFM) techniques. However, optimal base in a user subset is less complex than in the the minimum sum-rate strategy is not unique in general, and ground/entire user set and can be completed in parallel in thealgorithmsin[5],[6]utilizingEdmond’sgreedyalgorithm distributed systems. [7]necessarilyreturnanextremalpoint,oranunfairminimum sum-rate strategy. On the other hand, fairness is an important factor in CO. II. SYSTEMMODEL For example, in CCDE, the users are considered as peers in Let V = {1,2,...} be a finite set. We assume that there wirelesscommunications.Afairtransmissionstrategyencour- are |V| > 1 users in the system. Let Z = (Z : i ∈ V) V i ages them to take part in CO and helps prevent driving the be a vector of discrete random variables indexed by V. For ∗Ni Ding, Rodney A. Kennedy and Parastoo Sadeghi are with the each i∈V, user i can privately observe an n-sequenceZni of Research School of Engineering, College of Engineering and Com- the random source Zi that is i.i.d. generated according to the puter Science, the Australian National University (email: {ni.ding, rod- joint distribution PZ . We allow users exchange their sources ney.kennedy, parastoo.sadeghi}@anu.edu.au). The work of Ni Ding (email: V directly so as to let all of them recover the source sequence [email protected]) has been done when she was a junior research assistantwiththeInstituteofNetworkCodingatChineseUniversityofHong ZnV. Let rV = (ri : i ∈ V) be a rate vector indexed by V. KongfromNov23,2015toFeb6,2016. We callr anachievableratevector,ortransmissionstrategy, ‡ChungChan(email:[email protected]) andQiaoqiaoZhou(email: V if omniscience is possible by letting users communicate with [email protected]) arewiththeInstitute ofNetwork CodingatChinese University ofHongKong. the rates designated by rV. Let r be the function associated 2 withrV suchthatr(X)= i∈Xri,∀X ⊆V.ForX,Y ⊆V, B(fˆ4#,≤) let H(ZX) be the amount of randomnessin ZX measured by 2 P(fˆ#,≤) Shannon entropy and H(ZPX|ZY) be the conditional entropy Jain’4sfairestrate of Z given Z . It is shown in [1] that an achievable rate Shapleyvalueˆr X Y V vector must satisfy the Slepian-Wolf constraints: r(X) ≥ r3 1 (λ,λ,λ) H(Z |Z ),∀X ⊂V. Let α∈R and X V\X + H(Z |Z ) X ⊂V f (X)= X V\X . 0 α 2 (α X =V 1 3 The polyhedron and base polyhedron of f are respectively 2 α r 2 1 r P(f ,≥)={r |r(X)≥f (X),∀X ⊆V}, 0 0 1 α V α B(f ,≥)={r |r ∈P(f ,≥),r(V)=α}, Fig. 1: In the CO system in Example II.1, the Shapley value α V V α ˆr =(8,2,2) is the gravity center of B(fˆ#,≤). The Jain’s where B(fα,≥) contains all achievable rate vectors that have faVirest r3ate3 v3ector is r∗ = (2,1,1) 6= ˆr 4. Instead, r∗ = sum-rate equal α. B(fα,≥) = ∅ means that there is no (2,1,1)=argmax{r(VV)|r ∈P(fˆ#,≤)V,r ≤(2,2,2V)}. achievable rate vector that has sum-rate α. Let f#(X) = V 4 V α f (V) − f (V \ X),∀X ⊆ V be the dual set function of α α fα. We have B(fα#,≤) = B(fα,≥) [13].1 Denote Π(V) the It can be shown that B(fˆ#,≤)=B(f#,≤)=B(f ,≥) is a set of all partitions of V. It is shown in [3], [4], [11] that f# 4 4 4 α convex region, as shown in Fig. 1, instead of a singleton. is intersecting submodular and B(f ,≥) 6= ∅ if α is no less α than the minimum sum-rate In this paper, we assume that α ≥ α∗(V). We will show H(Z |Z ) thatthis assumptionholdsfor all elementsin the fundamental α∗(V)= max V\C C . (1) partitionP∗ inSectionIV-B.SinceB(fˆ#,≤)=B(f#,≤)= P∈Π(V):|P|>1 |P|−1 α α CX∈P B(fα,≥) when α≥α∗(V), we mainly present the results in We call the maximizer of (1) the fundamental partition and terms of fˆ#. In Section VI, we will show how to solve the α denote by P∗. If α≥α∗(V), B(fˆ#,≤)=B(f#,≤) where fairness problem by the oracle calls of f#.2 α α α fˆ#(X)= min f#(C), ∀X ⊆V α α P∈Π(X)C∈P III. MOTIVATION:JAIN’S FAIRNESS VS. SHAPLEYVALUE X is the Dilworth truncation of fα# and fˆα# is submodular. It is shown in [11] that a fair rate vector in B(fˆα#,≤) can Usually, B(fα,≥) is not a singleton when it is nonempty. be determinedby the Shapley valueˆrV [14]. Each tuple rˆi of So, a natural question that follows is to find a rate vector in ˆrV quantifies the average marginal contribution of user i in B(fα,≥) that distributes the transmission load as evenly as CO, and ˆrV is the gravity center of B(fˆα#,≤) [14]. But, the possible among the users. Jain’s fairest3 rate vector in B(fˆα#,≤) is the minimizer of Example II.1. Consider the user set V = {1,2,3} and let min{ r2 |r ∈B(fˆ#,≤)} (2) i V α W be an independent uniformly distributed random bit. The i i∈V X three users observe respectively that does not usually coincide with the Shapley value. Z =(W ,W ,W ,W ,W ), 1 a b c d e Example III.1. In the system in Example II.1, in B(fˆ#,≤), Z2 =(Wa,Wb,Wf), the Shapley value is rˆV = (83,23,23), while the Jain’s4fairest Z =(W ,W ,W ). rate vector is (2,1,1). See Fig. 1. 3 c d f Inthissystem,itcanbeshownthatα∗(V)=3.5andB(f#,≤ It can be easily seen based on the definition of Shapley α ) = B(f ,≥) = ∅ if α < 3.5. For α = 3.5, B(f# ,≤) = valuein[14]thatˆr isusuallybiasedsothattheuserhasmore α 3.5 V B(f ,≥)={(2.5,0.5,0.5)} is singleton. However, for α= informationonZ transmitsmore.Forexample,inthesystem 3.5 V 4, we have f# as in Example II.1, we have H(Z ) greater than H(Z ) and 4 1 2 f#(∅)=0,f#({1})=3,f#({2})=1,f#({3})=1, H(Z3).ThethreevertexpointsofB(fˆ4#,≤),(2,1,1),(3,0,1) 4 4 4 4 and (3,1,0), all have r greater than the other entries, and so 1 f4#({1,2})=4,f4#({1,3})=4,f4#({2,3})=3, does the gravity center ˆrV. However, in communications or f#({1,2,3})=4, networkengineering,we wantto distributethe ratesas evenly 4 as possible without considering the users’ prior knowledge and its Dilworth truncation as about the sources. So, the ideal rate vector is in the form of fˆ4#(∅)=0,fˆ4#({1})=3,fˆ4#({2})=1,fˆ4#({3})=1, λ·χV, where χX is the characteristic or incidence vector of fˆ4#({1,2})=4,fˆ4#({1,3})=4,fˆ4#({2,3})=2, X ⊆V.IfsucharatevectordoesnotbelongtoB(fˆα#,≤),we fˆ#({1,2,3})=4. 4 2Theoracletakes X ⊆V asaninputandoutputs fα#(X). of1thTehedumaaljsoertitfyunstcutidoinesfoα#nCanOdaitrsebbaasseedpoolnyhtheder(oinnteBrs(efcα#tin,g≤))s,ueb.mg.,od[3u]l–a[r6it]y. to31Ja(ifna’isrefsat)irn[1e2ss].index of rV is |(VP|iP∈Vi∈rVi)r2i2 ranges from |V1| (unfairest) 3 need to at least find one that is as close as possible to it, e.g. whichdep(r∗ ,1)={1,2,3},dep(r∗,2)={2},dep(r∗ ,3)= V V V the rate vector (2,1,1) in Fig. 1. In this sense, the minimizer {1,2,3}andLemmaIV.1holds.Forexample,fori=1,j =2, of (2) provides a better solution than Shapley value, and it is we have j ∈ dep(r∗,i)\{i}, ri∗ = 0.6 ≥ rj∗ = 0.5. Also, worth discussing on how to solve problem (2) efficiently. TwV(r∗V) = (0.5,0.V6,0.6)≥lex Twwi V(rV) forwajny other rV ∈ B(fˆ#,≤). For example, for r = (2.2,1,0.8) ∈ B(fˆ#,≤), 4 V 4 IV. MAIN RESULTS we have TwV(rV) = (0.5,0.55,0.8) and TwV(r∗V) ≥lex LetwV bea positiveweightvector.We consideraproblem TwV(rV) because the 1st entries of TwV(rV) and TwV(r∗V) that is more general than (2): are equal while the 2nd entry of TwV(r∗V) is greater than r2 that of TwV(rV). If we change wV to (1,1,1), we have min{ i |r ∈B(fˆ#,≤)}. (3) r∗ = (2,1,1). It also can be shown that Lemma IV.1 holds w V α V i∈V i and (2,1,1) is the lex-optimal base in B(fˆ#,≤). X 4 The objectivefunctioncan be consideredas a weightedJain’s fairness measure in that (3) reduces to (2) when wV = B. Lex-optimal minimum sum-rate strategy χV = (1,...,1). The main purpose of this section is to Let the fundamental partition be P∗ = {C ,...,C } 1 |P∗| show that the fairest minimum sum-rate strategy for CO can where ∅6=C ⊂V,∀m∈{1,...,|P∗|}. We have [3] be determined based on the fundamental partition P∗. We m start this section by presenting some existing results on the α∗(V)= f# (C )= fˆ# (C ). α∗(V) m α∗(V) m minimization problem (3) in Section IV-A, where we show CmX∈P∗ CmX∈P∗ that the minimizer of (3) is a lexicographically optimal (lex- Also, for each C ∈P∗, we have [8] optimal)baseinB(fˆ#,≤).Basedontheseresults,wepresent m our main result in Sαection IV-B: The lex-optimal minimum fα#∗(V)(Cm)=fˆα#∗(V)(Cm)≥α∗(Cm). sum-ratestrategyisadirectmergeofthethelex-optimalbases Alternatively speaking, in order to achieve omniscience with of the user subsets in the fundamental partition P∗. the minimum sum-rate α∗(V), each user subset C in the m fundamental partition P∗ must transmit exactly fˆ# (C ) α∗(V) m A. Lex-optimal Base timeswhichis greaterthanor equalto theminimumsum-rate A set X is called rV-tight if r(X) = fˆα#(X).4 Let for achieving the omniscience in Cm itself. Let hCm(X) = dep(rV,i) be the smallest rV-tight set that includes i. For fˆα#∗(V)(X),∀X ⊆ Cm. For Cm,Cm′ ∈ P∗ such that r ∈B(fˆ#,≤), dep(r ,i) can be expressed as [13] C 6= C , denote r ⊕ r = (r : i ∈ C ⊔ C ) V α V thme directm′merging ofCmr aCnmd′ r .i For exammple, mfo′r dep(rV,i)={j |∃ξ >0: rV +ξ(χi−χj)∈B(fˆα#,≤)}, r = (r ,r ,r ) = (2,3C,m4) and rCm′ = (r ,r ) = (4,1), where ξ(χi − χj) is called elementary transformation. It rCCmm ⊕ rC1m′ 4= 5(r1,...,r5) = (2,4C,m1′,3,4).2Th3e following is shown in [13] that the local optimality with respect to theoremstatesthatthelex-optimalminimumsum-ratestrategy the direction χi − χj implies the global optimality: r∗V is w.r.t.wV isadirectmergeofthelex-optimalbaseineachuser the minimizer of (3) if i∈V (rwi∗i)2 ≤ i∈V wri2i,∀rV = subset Cm w.r.t. wCm in the fundamental partition P∗. r∗V + ξ(χi − χj) ∈ B(fˆα#P,≤). The condiPtion is equivalent Theorem IV.3. The lex-optimal base in B(fˆα#∗(V)) w.r.t. wV to (4) in the lemma below. is r∗ = r∗ ⊕ ... ⊕ r∗ , where P∗ is the fundamental Lemma IV.1 ([13,Theorem8.1]). r∗V istheminimizerof (3) wpartVit.ion anCd1r∗Cm is theCle|Px∗-o|ptimalbase in B(hCm,≤) w.r.t. iff Cm wri∗i ≥ wrj∗j, ∀j ∈dep(r∗V,i)\{i}. (4) fˆ# Pr(oCof:) F=orhany(CrV),∀C∈ ∈B(Pfˆα∗#∗,(Vi.e).,,≤C), ris(Crm)-tig=ht α∗(V) m Cm m m m V Tor(d4L,e2er,ti1n)g(T(w2oV,f3(u,rs1Ve)r))in==dic(0e(s.wr5σsσ,(u(111c),)h1,..t5.h).a,twwrwrhσσσσ(e(|(|(Vr1V1e))||))σ≤)(1.w).h.=e≤re1wr,σσσ(σ((||ViV()|2|))),is=e.ga3n., Bfro′V(rh=aCllmr∗Cu,≤s1e⊕)r w.su..r.b.⊕ts.erwt∗CCC|Pmm∗.|Tiwnhhetehnre,erfr′Vu∗Cnm∈daiBms(etfhnˆαe#ta∗ll(eVxp)-a,or≤tpitt)iimonneaclPebs∗as.saerLioleyft and σ(3) = 2. It is shown in [15] that, for r∗V and any other [13,Lemma3.1].Also,foralli∈Cm ∈P∗,dep(r′V,i)⊆Cm rdwrVoσσm((ii∈′′i))nBaft(oefsrˆα#Ts,ow≤mV)e(,rwVi′e). ahIntadviesdewcrnσ∗aσo(l(ilit)e)edd=bTywrwσσTV((iiw)()rV,∗V∀(ri)∗V>l)exi≥′iclaeonxgdTrawwrpVσ∗σh(((iiic′r′))aVl>l)y. acbnoadnsediwrwti′ii.orn.≥t.(4wwr)j′Vji,n.∀Lje∈mmdeapI(Vr.∗V1.,iT)h\er{eif}o,rew,rh′Vichisistheeqluexiv-aolpetnimt taol So, r∗ is called the lex-optimal base in B(fˆ#,≤) w.r.t. w . Example IV.4. Let the user set be V ={1,2,3,4}. The four V α V users observe respectively Example IV.2. In the system in Example II.1, for w = V (4,2,1), we have lex-optimal base r∗ = (2.4,1,0.6), for Z =(W ,W ,W ,W ,W ), V 1 c d f g h Z =(W ,W ,W ,W ), du4eAtontrhVe-ctiognhsttrsaeitntXr(mXe)an≤srfaˆαt#e(sXatu),raftoiornξi>ne0n,trrieVs+riξfo·rχaill∈/i∈P(Xfˆα#,i,.e≤., Z32 =(Wca,Wdd,Weg,Wfh,Wg,Wh), ),∀i ∈X. The example in Section V shows how tight set is related to the Z =(W ,W ,W ). minimizerr∗ of(3). 4 a b f V 4 pathtor∗ pathtor∗ V V directionλ χ directionλ w 1 V 1 V directionλ χ directionλ w r∗ 2 V\S1 r∗ 2 V\S1 V V 1 B(fˆ#,≤) 1 B(fˆ#,≤) 4 4 r3 P(fˆ#,≤) r3 P(fˆ#,≤) 4 4 0 0 1 1 2 2 r2 0 0 1 r1 r2 0 0 1 r1 (a)path(0,0,0)→(0,1,1)→(2,1,1)w.r.t.wV =(1,1,1) (b)path(0,0,0)→(0,1,0)→(2.4,1,0.6)w.r.t.wV =(4,2,1) Fig. 2: The path towards the lex-optimal r∗ w.r.t. w in the system in Example II.1. V V Inthissystem, we haveminimumsum-rateforCO α∗(V)=6 Example V.1. We explain the path (0,0,0) → (0,1,1) → and the fundamental partition P∗ = {{1,2,3},{4}}. For (2,1,1) towards the lex-optimal base r∗ = (2,1,1) w.r.t. V C = {1,2,3}, we have h (X) = fˆ#(X),∀X ⊆ C and w =(1,1,1)inFig.2(a)byusing(5)asfollows.Weinitiate 1 C1 6 1 V B(h ,≤) 6= ∅ because h (C ) = 5 > α∗(C ) = 4. r∗ =(0,0,0) and the direction λ·w =λ·χ =(λ,λ,λ). C1 C1 1 1 V V V Consider determining the lex-optimal base in B(fˆ#,≤) w.r.t. We increase λ from 0 to move along the direction χ until 6 V w =χ =(1,1,1,1)byTheoremIV.3.Thelex-optimalbase at λ = 1 we reach a boundary of P(fˆ#,≤) that corre- V V 4 w.r.t. w = (1,1,1) in B(h ,≤) is r∗ = (r∗,r∗,r∗) = sponds to the inequality r({2,3}) ≤ fˆ#({2,3}) = 2. Here, (5,5,5C).1ForC ={4},r∗ =C1r∗ =1.TCh1edirect1me2rgi3ngof the boundary point (1,1,1) is necessa4rily the maximizer of r3 3an3dr is r2∗ ⊕r∗ =C2(5,54,5,1). Itcan be shown that max{r(V) | r ∈ P(fˆ#,≤),r ≤ λ ·w = (1,1,1)}. We rC∗1 ⊕r∗ C2is theCl1ex-opCt2imal 3bas3e3w.r.t. w = (1,1,1,1) in cannot increasVe λ any4more aloVng the1direVction λ·χ since BC(1fˆ#,≤C)2, which is also the Jain’s fairest mVinimum sum-rate otherwise we will be out of P(fˆ#,≤) so that the reVsulting 6 4 strategy. vector does not belong to B(fˆ#,≤). For the set {2,3} that 4 is associated with the boundary, we know that we can at V. PATH TOWARDSr∗ least allocate fˆ4#({2,3}) = 2 rates evenly to users 2 and V 3. Therefore, by letting λ = 1 and S = {2,3}, we assign 1 1 TheproblemunsolvedinTheoremIV.3ishowtodetermine rates r∗ =λ w =1 and r∗ =λ w =1 so that the rate r∗ 2 1 2 3 1 3 V thelex-optimalbaseineachusersubset.Inthissection,weuse is updated to (0,1,1). By knowing that we cannot increase theexistingresultsin[15]toshowthatthereexistsapiecewise any further along any dimension in S ,5 we set the direction 1 linear path towards the lex-optimal base, which can be found λ·χ =(λ,0,0) and continue to increase the value of λ by a decomposition algorithm in the existing literature. fromVr\∗S1=(0,1,1).The purposeis to see ifwe candistribute Let the distinct values of wri∗i of the lex-optimal base r∗V the remVaining rates fˆ4#(V)−fˆ4#(S1) = 2 evenly among the Sbe0 =λ1,∅.a.n.,dλSNns=uc{hit|hwrai∗ti 0≤=λnλ,0i∈<Vλ}1.<We..h.a<veλalNl.SDnefnoormte iussesrestibnyVth\eSc1o.nAsttrλai=nt2r,(w{1e,r2e,a3c}h)th≤efsˆe4#co({n1d,b2o,3u}n)da=ry4thiant a chain [15] P(fˆ#,≤). We set λ =2 and S = {1,2,3}. For S \S = 4 2 2 2 1 {3}, we assign r∗ = λ w = 2. r∗ is updated to (2,1,1). C :∅=S ⊂S ⊂...⊂S =V 3 2 3 V S 0 1 N Here, the second boundary point (2,1,1) is the maximizer of suchthatri∗ =λnwi,foralli∈Sn\Sn−1.Sincedep(r∗V,i)⊆ max{r(V) | rV ∈ P(fˆ4#,≤),rV ≤ λ2 · wV = (2,2,2)}. Sn,∀i ∈ Sn, r∗(Sn) = fˆα#(Sn). It means that we cannot Now, we have V \ S2 = ∅, which means all fˆ4#(V) = 4 increase the value of r∗ for all i∈S because otherwise the rates have been allocated. We get r∗ = (2,1,1) as the lex- i n V condition r∗(S ) ≤ fˆ#(S ) in the polyhedron P(fˆ#,≤) is optimal rate vector w.r.t. w = (1,1,1). In the same way, we n α n α breached.Alternativelyspeaking,r∗ ∧λ isthemaximizerof canshowthepath(0,0,0)→(0,1,0)→(2.4,1,0.6)towards V n max{r(V)|rV ∈P(fˆα#,≤),rV ≤λn·wV}[15],wherer∗V∧ r∗V =(2.4,1,0.6) w.r.t. wV =(4,2,1) in Fig. 2(b). Based on λn·wV = (min{ri∗,λn·wi}: i ∈ V) is the componentwise this path, we have λ1 = 0.5 corresponds to S1 = {2} and minimization of r∗V and λn·wV. It is clear that a piecewise λ2 =0.6 corresponds to S2 ={1,2,3}. linear path in P(fˆ#,≤) towards the lex-optimal base r∗ can α V Due to the min-max theorem [7] be determined by considering the problem max{r(V)|rV ∈P(fˆα#,≤),rV ≤λ·wV}, (5) max{r(V)|rV ∈P(fˆα#,≤),rV ≤λ·wV}= min{fˆ#(X)+λw(V \X)|X ⊆V}, (6) where λ ∈ [0,+∞). This path also determines the values of α all S and λ . n n 5ItmeansthatS1 isr∗V-tight, i.e.,r∗(S1)=fˆ4#(S1). 5 Algorithm 1: decomposition algorithm (DA) [16] user4inC ={4}todeterminethelex-optimalbaser∗ and r∗ , respec2tively. C1 input : S,S in CS C2 output: the set that contains S, S and all Sn in CS such that S ⊂Sn ⊂S VII. CONCLUSION 1 begin We considered the problem of minimizing a weighted 2 λ= fˆα#(S)−fˆα#(S); quadratic function over the set of all achievable rate vectors w(S\S) with the same sum-rate for CO. The objective function was 3 get S as the maximal minimizer of a generalization of the Jain’s fairness index. We showed that min{fˆα#(X)−λw(X):X ⊆V}; if the constraint set was nonempty the optimizer was a lex- optimal base that could be searched by a DA algorithm in if S =S then return (S,S) else return DA(S,S)∪DA(S,S) stronglypolynomialtime.Wealsoshowedthatthelex-optimal 4 end minimum sum-rate strategy for CO could be determined by directly merging the lex-optimal bases for all user subset in the fundamental partition, which could be completed in a problem (5) reduces to min{fˆ#(X) − λw(X) | X ⊆ V}, decentralized manner with lower complexity. α which is an SFM problem and can be efficiently solved by REFERENCES the DA algorithm in Algorithm 1 that is proposed in [16]. [1] I. Csisza´r and P. Narayan, “Secrecy capacities for multiple terminals,” Example V.2. In B(fˆ#,≤) in Example II.1, consider the IEEETrans.Inf.Theory,vol.50,no.12,pp.3047–3061, Dec.2004. problem of finding the4 lex-optimal base r∗ w.r.t. w = [2] S.ElRouayheb,A.Sprintson,andP.Sadeghi,“Oncodingforcoopera- V V tivedataexchange,” inProc.IEEEInf.TheoryWorkshop,Cairo,2010, (4,2,1) by the DA algorithm. We call DA(∅,V), i.e., S = ∅ pp.1–5. and S = V. We get λ = 4 and S = {2}. Since S 6= [3] C.Chan, A.Al-Bashabsheh, J.Ebrahimi, T.Kaced, andT.Liu, “Mul- 7 tivariate mutual information inspired by secret-key agreement,” Proc. S = V, DA(∅,{2}) is called where we get λ = 0.5 and IEEE,vol.103,no.10,pp.1883–1913, Oct.2015. S = {2}= S. It returns {∅,{2}}. On the other hand, in the [4] N. Ding, R. A. Kennedy, and P. Sadeghi, “Estimating minimum callofDA({2},V),wegetλ=0.6andS =V =S.Itreturns sum-rate for cooperative data exchange,” in Proc. IEEE Int. Symp. Inf. Theory, Hong Kong, China, 2015. [Online]. Available: {{2},V}.WethengetthechainC : ∅⊂{2}⊂{1,2,3}=V S http://arxiv.org/abs/1502.03518 corresponding to λ0 = 0, λ1 = 0.5 and λ2 = 0.6. By [5] T.CourtadeandR.Wesel,“Codedcooperative dataexchangeinmulti- calling DA algorithm to determine the lex-optimal base r∗ hopnetworks,”IEEETrans.Inf.Theory,vol.60,no.2,pp.1136–1158, w.r.t. w = (1,1,1), we get the chain C : ∅ ⊂ {2,3} ⊂V Feb.2014. V S [6] N.Milosavljevic, S.Pawar, S.E.Rouayheb, M.Gastpar, andK.Ram- {1,2,3}= V corresponding to λ0 = 0, λ1 = 1 and λ2 = 2. chandran, “Efficient algorithms for the data exchange problem,” arXiv The results are consistent with those in Example V.1. preprintarXiv:1502.06149, 2015. [7] J. Edmonds, “Submodular functions, matroids, and certain polyhedra,” inCombinatorialOptimization—Eureka,YouShrink! Berlin:Springer, 2003,pp.11–26. VI. COMPLEXITY [8] N.Milosavljevic, S.Pawar, S.ElRouayheb, M.Gastpar, andK.Ram- chandran, “Deterministic algorithm for the cooperative data exchange In step 2 in Algorithm 1, the value of the Dilworth trun- problem,” in Proc. IEEE Int. Symp. Inf. Theory, St. Petersburg, 2011, cation fˆα# at certain X ⊆ V can be obtained by oracle pp.410–414. calls of f# in O(|V|· SFM(|V|)) time, where SFM is the [9] N. Ding, R. A. Kennedy, and P. Sadeghi, “Fairest constant sum-rate α transmissionforcooperativedataexchange:AnM-convexminimization complexity of minimizing a submodular function. In step 3, approach,” in Proc. 22nd Int. Conf. Telecommun., Sydney, Australia, the maximal minimizer of min{fˆα#(X)−λw(X) | X ⊆ V} 2015.[Online].Available: http://arxiv.org/abs/1502.03517 canbeobtainedbyconsideringtheproblemofmin{f#(X)− [10] S.Tajbakhsh,P.Sadeghi,andR.Shams,“Ageneralized modelforcost α andfairnessanalysisincodedcooperative dataexchange,”inProc.Int. λw(X) | X ⊆ V : i ∈ V}. For each i ∈ V, it completes in Symp.NetworkCoding,Beijing, 2011,pp.1–6. O(|V|·SFM(|V|))timebyoraclecallsoffα#.Todetermineall [11] N. Ding, C. Chan, T. Liu, R. A. Kennedy, and P. Sadeghi, “A game- S inchainC ,theDAalgorithmiscalledfornogreaterthan theoreticperspectiveoncommunicationforomniscience,”arXivpreprint n S arXiv:1510.08953, 2015. 2|V|−1times[17]. Therefore,the complexityof determining [12] R. Jain, D.-M. Chiu, and W. R. Hawe, A quantitative measure of r∗V by oracle calls of fα# is O(|V|2 · SFM(|V|)), which is fairness anddiscrimination forresourceallocation insharedcomputer stronglypolynomialascomparedtotheexponentiallygrowing system. Hudson,MA:EasternResearchLaboratory,DigitalEquipment complexity in obtaining the Shapley value ˆr [14]. Corporation, 1984. V [13] S.Fujishige,Submodularfunctionsandoptimization,2nded. Amster- Consider the complexity of determining the lex-optimal dam,TheNetherlands: Elsevier, 2005. minimumsum-rate strategy by TheoremIV.3. The fundamen- [14] L. S. Shapley, “Cores of convex games,” Int. J. Game Theory, vol. 1, no.1,pp.11–26,Dec.1971. tal partition P∗ can be determined in O(|V|2 · SFM(|V|)) [15] S. Fujishige, “Theory of principal partitions revisited,” in Research time [18]. After obtaining P∗, we can determine the lex- TrendsinCombinatorialOptimization,W.Cook,L.Lova´sz,andJ.Vy- optimal base r∗ for each C ∈ P∗ and get the lex- gen,Eds. Berlin: SpringerBerlinHeidelberg, 2009,pp.127–162. optimal minimuCmmsum-rate stratmegy r∗. Since |C | < |V|, [16] K. Nagano and Y. Kawahara, “Structured convex optimization under V m submodularconstraints,” arXivpreprintarXiv:1309.6850,2013. mOu(|cChmlo|2w·erSFthMan(|Cthmat|)o)f, fithnedincgomrp∗l.exInityadodfitifionnd,inr∗g r∗Cfomr,ailsl [17] Klem.NsaogvaenrobaasnedpKol.yAtoipheasr,a”,J“aEpqaunivJa.lIenndc.eaonfdcAopnpvle.xMmatihn.i,mviozla.ti2o9n,nporo.b3-, V Cm pp.519–534, 2012. C ∈P∗ can be found in parallel in distributed systems. For m [18] K.Nagano,Y.Kawahara,andS.Iwata,“Minimumaveragecostcluster- example, in Example IV.4 where P∗ = {{1,2,3},{4}}, we ing,”inProc.AdvancesinNeuralInf.Process.Syst.,Vancouver,2010, can allow a user delegate in C = {1,2,3}, say, user 1, and pp.1759–1767. 1