Contents lists available atScienceDirect
Digital
Signal
Processing
www.elsevier.com/locate/dsp
Team-optimal
distributed
MMSE
estimation
in
general
and
tree
networks
Muhammed
O. Sayin
a,
∗
, Suleyman
S. Kozat
b, Tamer Ba ¸sar
aaTheDepartmentofElectricalandComputerEngineering,UniversityofIllinoisatUrbana-Champaign,Champaign,IL61801,USA bTheDepartmentofElectricalandElectronicsEngineering,BilkentUniversity,Bilkent,Ankara06800,Turkey
a
r
t
i
c
l
e
i
n
f
o
a
b
s
t
r
a
c
t
Articlehistory:
Availableonline22February2017
Keywords: Distributednetworks DistributedKalmanfilter Teamproblem Finite-horizon
DistributedMMSEestimation
Weconstructteam-optimalestimationalgorithmsoverdistributednetworksforstateestimationinthe finite-horizon mean-square error (MSE)sense. Here, we have a distributed collection of agents with processingandcooperationcapabilities.Theseagentsobservenoisysamplesofadesiredstatethrough alinearmodelandseek tolearnthisstatebyinteractingwitheachother.Althoughthisproblemhas attracted significantattention and been studied extensively in fields includingmachine learningand signal processing, allthe well-knownstrategiesdo not achieveteam-optimal learningperformance in the finite-horizonMSEsense. To thisend, weformulatethe finite-horizondistributedminimum MSE (MMSE)whenthereisnorestriction onthesizeofthedisclosedinformation,i.e.,oracleperformance, overanarbitrarynetworktopology.Subsequently,weshowthatexchangeoflocalestimatesissufficient to achievethe oracle performance onlyovercertainnetwork topologies.By inspecting thesenetwork structures,weproposerecursivealgorithmsachievingtheoracleperformancethroughthedisclosureof localestimates.Forpracticalimplementationswealsoprovideapproachestoreducethecomplexityof thealgorithmsthroughthetime-windowingoftheobservations.Finally,inthenumericalexamples,we demonstratethesuperiorperformanceoftheintroducedalgorithmsinthefinite-horizonMSEsensedue tooptimalestimation.
©2017ElsevierInc.Allrightsreserved.
1. Introduction
Overa distributednetwork ofagents withmeasurement, pro-cessing and communication capabilities, we can have enhanced processing performance, e.g., fast response time, relative to the centralizednetworksbydistributingtheprocessingpoweroverthe networks[1–4]. Mainly, distributedagents observethe true state ofthe system through noisy measurements from different view-points,processtheobservationdatainordertoestimatethestate, andcommunicatewitheachothertoalleviatetheestimation pro-cessinafullydistributedmanner.Notably,theagentscanrespond tostreaming data inan onlinemanner by disclosinginformation among each other at certain instances. This framework is con-veniently used to modelhighly complex structuresfrom defense applicationstosocialandeconomicalnetworks[5–8].Asan exam-ple,saythat wehave radarsystemsdistributedover an areaand seekingto locate hostile missiles, i.e., the location ofthe missile isthe underlyingstate of thesystem. In that respect,distributed
*
Correspondingauthor.E-mailaddresses:sayin2@illinois.edu(M.O. Sayin),kozat@ee.bilkent.edu.tr (S.S. Kozat),basar1@illinois.edu(T. Ba ¸sar).
processing approach has vital importance in terms of detecting the missiles and reacting as fast as possible. In particular, even iftheviewpointsofafew radarsystemsareblockeddueto envi-ronmentalobstacles,throughthecommunicationamongtheradar systems,eachsystemshouldstillbeabletolocatethemissiles. Ad-ditionally,sinceeachradarsystemnotonlycollectsmeasurements butalsoprocessthemtolocatethemissiles,theoverallsystemcan respondtothemissilesfasterthanacentralizedapproachinwhich measurements ofallthe radarsystemsare collected ata central-izedunitandprocessedtogether.
Althoughthereis anextensive literatureon thistopic, e.g., [2, 6–11]andreferencestherein,westill havesignificantandyet un-explored problems for disclosure and utilization of information amongagents.Priorworkhasfocusedonthecomputationally sim-plealgorithmsthataimtominimizecertaincostfunctionsthrough theexchangeoflocalestimates,e.g., diffusionorconsensusbased estimation algorithms [2,9,3,12–14],due to processing power re-latedpracticalconcerns. However,thereis atrade-off intermsof computationalcomplexityandestimationperformance.
Formulatingtheoptimaldistributedestimationalgorithmswith respect to certain performance criteria isa significant and unex-plored challenge. To this end, we consider here the distributed
http://dx.doi.org/10.1016/j.dsp.2017.02.007 1051-2004/©2017ElsevierInc.Allrightsreserved.
estimation problem asa team problem for distributed agents in whichagentstakeactions,e.g.,whichinformationtodiscloseand howtoconstructthelocalestimate.Thisdiffersfromtheexisting approaches in which agents exchange their local estimates. Fur-thermore,weaddresstheoptimalityofexchanginglocalestimates with respect to the team problemover arbitrary network struc-tures.
We examine the optimal usage of the exchanged information based on its content rather than a blind approach in which ex-changed information ishandled irrespective of the content asin thediffusionorconsensus basedapproaches. Insuch approaches, the agents utilize the exchanged information generally through certain static combination rules, e.g., the uniform rule [15], the Laplacianrule[16]ortheMetropolisrule[17].However,ifthe sta-tisticalprofile ofthe measurement data variesover thenetwork, i.e., each agent observes diverse signal-to-noise ratios, by ignor-ingthe variationinnoise, theserulesyield severedegradation in theestimationperformance[2].Insuchcasestheagentscan per-formbetter even without cooperation[2].Therefore, the optimal usageoftheexchangedinformationplaysanessentialrolein per-formanceimprovementintheteamproblem.
Consider distributed networks of agents that observe noisy samplesof an underlying state (possibly multi-dimensional) over a finite horizon. The agentscan exchange information withonly certainother agentsateach timeinstant.Inparticular,agents co-operate with each other as a team according to a certain team cost depending on the agents’ actions. To this end, each agent constructsa localestimateofthe underlyingstate andconstructs messages to disclose to the neighboring agents at each time in-stant.We particularlyconsidera quadraticcostfunction andthat theunderlyingstateandmeasurementnoisesarejointlyGaussian. We note that restrictions on the sent messages, e.g., on the size of the disclosed information,have significant impact on the optimalteamactions.Weintroducetheconceptoftheoracle per-formance,inwhich thereis norestrictionon thedisclosed infor-mation. In that case, a time-stamped information disclosure can beteam-optimal andweintroduce theoptimaldistributedonline learning (ODOL) algorithm using the time-stamped information disclosure. Through a counter example, we show that the oracle performancecannotbeachievedthroughtheexchangeoflocal es-timates in general. Then, we analytically show that over certain networks, e.g., treenetworks, agents can achieve the oracle per-formance through the exchange of local estimates. We propose the optimaland efficient distributedonline learning (OEDOL) al-gorithm, whichis practical forreal life applicationsand achieves theoracle performance over treenetworksthrough theexchange of local estimates. Finally, we introduce the time windowing of the measurements in the team cost andpropose a recursive al-gorithm,sub-optimaldistributedonlinelearning(SDOL)algorithm, combiningthe received messages linearlythrough time-invariant combinationweights.
We can list our main contributions as follows: 1) We intro-duce a team-problem to minimize finite horizon mean square error cost function in a distributed manner. 2) We derive the ODOL algorithm achieving the oracle performance over arbitrary networksthrough time-stampedinformationexchange.3) We ad-dresswhetheragentscanachievetheoracleperformancethrough the disclosure of local estimates. 4) We propose a recursive al-gorithm, the OEDOL algorithm, achieving the oracle performance overcertain networktopologieswithtremendously reduced com-municationload.5)Wealsoformulatesub-optimalversionsofthe algorithmswithreducedcomplexity.6) Weprovidenumerical ex-amplesdemonstratingthesignificantgainsdueto theintroduced algorithms.
Theremainder ofthepaperis organizedasfollows.We intro-duce the team problem fordistributed-MMSE estimation in
Sec-Fig. 1. The neighborhoods of ith agent over the distributed network. tion 2.We studythetreenetworks, exploitthenetwork topology to formulate the OEDOL algorithm that reducesthe communica-tion load andintroduce cell structures, which is relatively more connectedthan treenetworks,in Section3.We proposethe sub-optimalversionsoftheODOLalgorithmforpractical implementa-tions in Section 4. In Section 5,we provide numerical examples demonstratingsignificant gainsduetothe introducedalgorithms. WeconcludethepaperinSection6withseveralremarks.
Notation: We work with real data for notational simplicity.
N(
0,
.)
denotes the multivariate Gaussian distribution with zero mean and designated covariance. For a vector a (or matrix A), a (or A) is its ordinary transpose. We denotethe vector whose terms are all 1s (or all 0s) by 1 (and 0). We denote random variablesbyboldlowercaseletters,e.g.,x.Theoperatorcol{·}
pro-ducesacolumnvectororamatrixinwhichtheargumentsofcol{·}
arestackedone undertheother.Foramatrix A,diag{
A}
operator constructs a diagonalmatrix with the diagonal entries of A. For agivensetN
,diag{
N }
createsadiagonalmatrixwhosediagonal block entriesareelementsoftheset.Theoperator⊗
denotesthe Kroneckerproduct.2. Teamproblemfordistributed-MMSEestimation
Consideradistributednetworkofm agentswithprocessingand communication capabilities. In Fig. 1, we illustrate this network through an undirected graph, where the vertices and the edges correspond totheagentsandthecommunicationlinksacrossthe network,respectively.Foreachagenti,wedenotethesetofagents whose information could be received at least after k hops, i.e.,
k-hopneighbors,by
N
i(k),andπ
(k) i:=
N
(k)
i
is thecardinalityofN
(k)i (see Fig. 1b)1.Weassume that
N
(0)i
= {i}
andN
(k)i
= ∅
fork
<
0.Notethatthesequenceofthesets{
N
i(0),
N
(1)i
,
. . .
}
isa non-decreasingsequencesuchthatN
i(k)⊆
N
i(l)ifk<
l.1 Fornotationalsimplicity,wedefineN
Here,atcertain time instants, theagentsobserve anoisy ver-sionof a time-invariantandunknown state vector x
∈ R
p which isa realizationofaGaussianrandomvariablex with mean¯
x andauto-covariance matrix
x. In particular, at time instant t, each agenti observesanoisyversionofthestateasfollows:
yi,t
=
Hix+
ni,t,
t=
1, . . . ,
T,
whereHi
∈ R
q×p isamatrixcommonlyknownbyallagents,andni,t
∈ R
q is a realization of a zero-mean white Gaussian vector process{
ni,t}
with auto-covarianceni. Correspondingly, the
ob-servation yi,t
∈ R
q is a realization of the random process{
yi,t}
, where yi,t=
Hix+
ni,t almosteverywhere (a.e.).Thenoise ni,t is also independent of the state x and the other noise parametersnj,τ , j
=
i andτ
≤
t.Weassumethatthestatisticalprofilesofthenoiseprocesses are commonknowledge of theagents since they canreadilybeestimatedfromthedata[18].
Theagentshavecommunicationcapabilitiesandatcertaintime instants,i.e.,aftereachmeasurement,they canexchange informa-tionwiththeneighboringagentsasseeninFig. 1a.Letzi,j,t
∈ R
r denotetheinformationdisclosedbyi to j attime t,andr≥
0 is thesizeofthedisclosedinformation.Weassumethatthereexists aperfectchannel betweenthe agentssuch thatthe disclosed in-formationcanbetransmittedwithinfiniteprecision.Therefore,we denotetheinformationavailabletoagent-i attimet byδ
i,t=
yi,t
,
yi,τ,
zj,i,τ,
for j∈
N
i,
τ
=
1, . . . ,
t−
1andlet
σ
i,t denotethesigma-algebrageneratedbytheinformation setδ
i,t.Furthermore,wedefinethesetofallσ
i,t-measurable func-tionsfromR
qt× R
rπi(t−1) toR
r byi,t.Importantly, here,which
information to disclose is not determined a priori in the prob-lemformulation. Let
γ
i,j,t bethe decisionstrategy forzi,j,t,then agent-i choosesγ
i,j,t, j∈
N
i,fromtheseti,t,i.e.,
γ
i,j,t∈
i,t andγ
i,j,t(δ
i,t)
=
zi,j,t,basedonhis/herobjective.In addition to the disclosed information zi,j,t, j
∈
N
i, agent-i takes action ui,t∈ R
q, where the corresponding decision strat-egyη
i,t is chosen from the seti,t, which is the set of all
σ
i,t-measurablefunctionsfromR
qt×R
rπi(t−1)toR
p,i.e.,η
i,t∈
i,t andη
i,t(δ
i,t)
=
ui,t.Here,weconsiderthattheagentshavea com-moncostfunction:T
t=1 m j=1 x−
uj,t2,
where all actions ui,t, i
=
1,
. . . ,
m and t=
1,
. . . ,
T are costly, andagent-i should take actions ui,t and zi,j,t, j=
1,
. . . , π
i andt
=
1,
. . . ,
T , accordingly. Therefore, this corresponds to a team-problem,inwhichagent-i facesthefollowingminimization prob-lem: min γi,j,t∈i,t,ηi,t∈i,t, j∈Ni,t=1,...,T T t=1 m j=1 Ex−
η
j,t(δ
j,t)
2.
(1)Wepointout thatboth
i,t and
i,t areinfinitedimensional,i.e.,
(1)isafunctionaloptimizationproblemandtheoptimalstrategies canbe anonlinear functionoftheavailable information. Further-more,the agentsshould also constructthe disclosedinformation accordinglysince other agents’ decisions uj,t directly depend on thedisclosedinformation.
2.1.Alowerbound
Inordertocomputetheteamoptimalstrategies,wefirst con-structalowerboundontheperformance oftheagentsby remov-ing the limitation on the size of the disclosed information, i.e.,
r
→ ∞
. In that case, the following propositionprovides an opti-malinformationdisclosurestrategy.Fig. 2. Timestampedinformationdisclosureover anetworkof6 agentsattime instantt=4.
Proposition2.1.Whenr
≥
mq,atimestampedinformationdisclosure strategy,inwhichagentstransmitthemostcurrentversionofthe avail-ableinformation(e.g.,seeFig. 2),canleadtotheteam-optimalsolution.Proof. Through the time stamped information disclosure, each agentcanobtainthemeasurements oftheotheragentsseparately ina connectednetwork.However, themeasurements ofthe non-neighboringagentscould only bereceived aftercertain hopsdue to the partially connected structure, i.e., certain agents are not directly connected. As an example, the disclosed information of
j
∈
N
i(2)reachestoi bypassingthroughtwocommunicationlinks asseeninFig. 1b.Inparticular,thiscaseassumesthateachagent has accessto full information fromthe other agents,albeit with certainhops,andcorrespondstothedirectaggregationofall mea-surementsacrossthenetworkateachagent.Correspondingly,attimet,alltheinformationaggregatedatith
agentisgivenby
δ
io,t:=
yi,ττ≤t,
yj,ττj∈N≤t−1(1) i, . . . ,
yj,ττ≤t−κi j∈Ni(κi),
(2)wheretheinformationfromthefurthestagent isreceivedatleast after
κ
i hops. Therefore, a time-stamped information disclosure strategycanleadtotheteam-optimalsolution.2
Let
σ
oi,t denotethesigma-algebrageneratedbytheinformation set
δ
oi,t andoi,t be theset of all
σ
oi,t-measurable functionsfrom
R
qt×R
qπi(t−1)×. . .×R
qπκi
i |t−κi|+ (where
|
t−
κ
i
|
+=
0 ift−
κ
i<
0) toR
p.Then,agent-i facesthefollowingminimizationproblem:min ηi,t∈oi,t, t=1,...,T T
t=1 m j=1 Ex−
η
j,t(δ
oj,t)
2,
(3) whichisequivalentto T t=1 min ηi,t∈oi,t Ex−
η
i,t(δ
oi,t)
2 (4) sinceγ
i,j,t, j∈
N
i,issettothetime-stampedstrategyandη
i,t has impactonlyonthetermEx−
η
i,t(δ
io,t)
2.Letδ
o i,t bedefinedby
δ
oi,t:=
yi,ττ≤t,
yj,ττ≤t−1 j∈Ni(1), . . . ,
yj,ττ≤t−κi j∈Ni(κi)Then,teamoptimalstrategy inthelowerbound,i.e., oracle strat-egy,
η
oi,t andthecorrespondingactionu o
i,t aregivenby
η
oi,t(δ
oi,t)
=
uoi,t=
E[
x|δ
oi,t= δ
oi,t]
(5)andwedefineuoi,t
:=
E[
x|δ
oi,t]
.2.2. ODOLalgorithm
Since the state and the observation noise are jointly Gaus-sian randomparameters, we cancompute (5)through a Kalman-like recursion [19]. Therefore, we provide the following ODOL algorithm. We introduce a difference set
i,t
:= δ
oi,t\δ
oi,t−1=
yi,t,
yj,t−1 j∈Ni(1),
. . . ,
yj,t−κi j∈N(κi ) i and a vector wi,t=
col{
i,t}
.Then,fort≥
1 theiterations oftheODOLalgorithmare givenby uoi,t=
I−
Ki,tH¯
i uoi,t−1+
Ki,twi,t,
Ki,t= ˆ
i,t−1H¯
i¯
Hiˆ
i,t−1H¯
i+ ¯
ni−1
,
ˆ
i,t=
I−
Ki,tH¯
iˆ
i,t−1
,
where2uoi,0= ¯
x,ˆ
i,0=
x,H¯
i:=
Pi⊗
Iq H , H:=
col{
H1, . . . ,
Hm} ,
¯
ni:=
Pi⊗
Ipn Pi
⊗
Ip ,n
:=
diagn1
, . . . ,
nm ,andPi is thecorrespondingpermutationmatrix.We point out that thisis a lower bound on the original cost function(1),i.e., T
t=1 m i=1 Ex−
uoi,t2≤
min γi,j,t∈i,t,ηi,t∈i,t, i=1,...,m; j∈Ni;t=1,...,T T t=1 m j=1 Ex−
η
j,t(δ
j,t)
2,
(6)wherewesubstituteteam-optimalaction(whenr
→ ∞
)uoi,t back into(4)andsumovert=
1,
. . . ,
T andi=
1,
. . . ,
m.However,the lower boundis not necessarilytight depending onr. By Proposi-tion 2.1, time-stamped information disclosure strategy, in which the size of the disclosed information is q×
m, yields the ora-cle solution.This implies that when r≥
qm, the lower bound is tight.Furthermore,teamoptimalsolutions arelinearinthe avail-able information and can be constructed through the recursive algorithm ODOL. However, qm islinear inthe numberof agents,m, andinlarge networksthiscancauseexcessivecommunication loadyetcommunicationloadiscrucialfortheapplicabilityofthe distributed learning algorithms [14,13]. Therefore, in the follow-ing section, we provide a sufficient condition on the size of the disclosed information, which depends on the network structure (ratherthanitssize),inordertoachievethelowerbound(6). 3. Distributed-MMSEestimationwithdisclosureoflocalestimate
Intheconventionaldistributedestimationalgorithms,e.g., con-sensus anddiffusion approaches, agentsdisclose their local esti-mates,whichhavesize p (notethat thisdoesnot dependonthe networksize). The followingexample addresseswhetherthe dis-closureoflocalestimatescanachievethelowerbound(6)ornot.
2 Iftheinversefailstoexist,apseudo-inversecanreplacetheinverse[19].
Fig. 3. A cycle network of 4 agents. 3.1. Acounterexample
Consider a cyclenetworkof 4 agents asseenin Fig. 3,where
p
=
q=
1,Hi=
1,ni
=
σ
2
n,fori
=
1,
. . . ,
4,andx
=
σ
x2.Weaim toshowthatagent-1’soracleactionattimet=
3,i.e.,uo1,3,cannot beconstructedthroughtheexchangeoflocalestimates.At time t
=
2, agent-2 and agent-3 have the following oracle actions: uo2,2=
E[
x|
y2,2=
y2,2,
y2,1=
y2,1,
y1,1=
y1,1,
y4,1=
y4,1]
=
σ
x2 4(
σ
2 x+
σ
n2)
(
y2,2+
y2,1+
y1,1+
y4,1),
(7) uo3,2=
E[
x|
y3,2=
y3,2,
y3,1=
y3,1,
y1,1=
y1,1,
y4,1=
y4,1]
=
σ
x2 4(
σ
2 x+
σ
n2)
(
y3,2+
y3,1+
y1,1+
y4,1).
(8) Note that since there are two hops between agents 2 and 3, att
=
2,agentsdonothaveaccesstoeachother’sanymeasurement yet.Attimet=
3,agent-1’soracleactionisgivenbyuo1,3
=
E[
x|
y1,3=
y1,3,
y1,2=
y1,2,
y1,1=
y1,1,
y2,2=
y2,2,
y2,1=
y2,1,
y3,2=
y3,2,
y3,1=
y3,1,
y4,1=
y4,1],
=
σ
x2 8(
σ
2 x+
σ
n2)
(
y1,3+
y1,2+
y1,1+
y2,2+
y2,1+
y3,2+
y3,1+
y4,1).
Assume that uo1,3 canbe obtainedthrough the exchange oflocal estimates:
ˆ
u1,3
:=
E[
x|
y1,3=
y1,3,
y1,2=
y1,2,
y1,1=
y1,1,
uo2,2=
uo2,2,
uo2,1=
uo2,1,
uo3,2
=
uo3,2,
uo3,1=
uo3,1].
(9)Since all parameters are jointlyGaussian, the local estimates are alsojointlyGaussian,u
ˆ
1,3,islinearinuo2,2 anduo3,2.Furthermore, the measurements y2,2, y3,2, and y4,1 are only included in uo2,2 anduo3,2.Therefore,weobtainˆ
u1,3= · · · +
α
u2o,2+ β
uo3,2= · · · +
α
σ
2 x 4(
σ
2 x+
σ
n2)
(
y2,2+
y4,1+ · · · )
+ β
σ
x2 4(
σ
2 x+
σ
n2)
(
y3,2+
y4,1+ · · · ),
where
· · ·
referstotheotherterms.However,theequalityofuˆ
1,3 anduo1,3 impliesα
= β =
1/
2 due tothecombinationweights ofy2,2 and y3,2,respectively,and
α
+ β =
1/
2 duetothe combina-tion weightof y4,1,whichleads toa contradiction.Hence, which information to discloseover arbitrary networksforteam-optimalFig. 4. Anexampletreenetwork.NoticetheeliminatedlinksfromFig. 1toavoid multi-pathinformationdiffusion.
solutions should be considered elaborately. In the following, we analyticallyshowthatlowerboundcouldbeachievedthroughthe disclosureoflocalestimatesover“treenetworks”.
3.2.Treenetworks
A network has a “tree structure” if its corresponding graph is a tree, i.e., connected andundirected without any cycles [20]. As an example, the conventional star or line networks have tree structures. We remark that for an arbitrary network topol-ogy we can also construct the spanning tree of the network andeliminate thecycles. Inthe literature, thereexists numerous distributed algorithms for minimum spanning tree construction
[21–25].
Importantly, the following theoremshows that over tree net-works we can achieve the performance of the oracle algorithm throughthedisclosureofthelocalestimatesonly.
Theorem3.1.Considertheteam-problemoveratreenetwork,inwhich r
=
p.Then,exchangeoflocalestimatescanleadtotheteam-optimal solution,i.e.,agentscanachievetheoracleperformance.Proof. Initially,agent-i hasaccessto yi,1 onlyandthe oracle ac-tion is uoi,1
=
E[
x|
yi,1=
yi,1]
. At time t=
2, the oracle action is givenby uoi,2=
E x|
yi,τ=
yi,ττ=1,2,
yj,1=
yj,1 j∈Ni,
(10)whichcanbewrittenas uoi,2
=
E x|
yi,τ=
yi,τ τ=1,2,
E[
x|
yj,1=
yj,1]
j∈Ni=
E x|
yi,τ=
yi,τ τ=1,2,
uoj,1=
uoj,1 j∈Ni.
(11)Thisimpliesthatfort
=
1 andt=
2,the oracleperformance can beachievedthroughthedisclosureoflocalestimate.Therefore,we can consider the oracle action (10) even though agents disclose their local estimate instead of time-stamped information disclo-sure.As seen in Fig. 4, over a tree network, fork
∈ {
1,
. . . , κ
i}
we haveN
(k) i=
j∈NiN
(k) i∩
N
(k−1) j.
(12)Notethatthesetsin(12)aredisjointas
N
(k) i∩
N
(k−1) j1∩
N
(k) i∩
N
(k−1) j2= ∅
(13)forall j1
,
j2∈
N
i and j1=
j2. Notably, over a tree network, by(13),we can partitionthe collectionsetofthemeasurements re-ceivedafteratleastk-hopsasfollows
yj,τ j∈N(k) i=
yj,τ j∈N(k) i ∩N (k−1) j1,
. . . ,
yj,τ j∈Ni(k)∩N(jπk−1) i.
(14)In the time-stamped information disclosure, at time t
=
3, agent-i hasaccesstoδ
oi,3,definedin(2).Wedenotethesetofnew measurementsreceivedbyi over j attimet
=
2 byoj,i,2
:=
=yj,2 yk,2 k∈N(1) i ∩N (0) j
,
yk,1 k∈N(2) i ∩N (1) j,
whichcanalsobewrittenas
oj,i,2
= δ
oj,2\
=δoj,1 yj,1,
yi,1 =o i,j,1=δoi,1,
(15)where we exclude the informationsent by i to j at time t
=
1, i.e., yi,1.Then,wecanwritetheaccessedinformationastheunion of new measurement yi,3, new measurements received over the neighboringagentsandtheaccessed informationattimet=
2 as follows:δ
io,3=
yi,3,
oj1,i,2, . . . ,
o jπi,i,2, δ
oi,2.
(16)Notethat thesetsonthe righthandside of(16)aredisjointdue totreestructure.Furthermore,by(15)and(16),thesigma-algebra generatedby
δ
oi,3 isequivalent to thesigma-algebrageneratedby theset
{
yi,τ}τ≤3,
{δ
oj,τ}
τj∈≤N2i .Since uoj,t=
E[
x|δ
oj,t= δ
oj,t]
,we ob-tain uoi,3=
E x|
yi,τ=
yi,τ τ≤3,
uoj,τ=
uoj,ττ≤2 j∈Ni.
(17) By(15),wehaveoj,i,t
= δ
oj,t\
δ
oj,t−1∪
oi,j,t−1,
(18)which implies that for t
≥
2,oj,i,t is constructible from
δ
io,τand
δ
oj,τ forτ
≤
t. Hence, by induction, we conclude that the lower bound can be achieved through the exchange of local es-timates.2
Remark3.1. When theexpectationofthe state isconditioned on infinitenumberofobservationsover evenaconstructedspanning tree,onlyafinitenumberoftheobservationsismissingcompared tothecaseoverafullyconnectednetwork.Hence,evenifwe con-struct the spanning tree of that network, we would still achieve the lower bound over a fullyconnected (or centralized)network asymptotically. As an illustrative example,in Fig. 10, we observe thattheMMSEperformanceoverthefullyconnected,starandline networks are asymptotically the same. Similarly, in [26–28], the authors show that the performance of the diffusion based algo-rithmscould approachthe performance ofafully connected net-workundercertainregularityconditions.
In thesequel, we propose the OEDOL algorithm that achieves thelowerboundovertreenetworksiteratively.
3.3. OEDOLalgorithm
ByTheorem 3.1,overatreenetwork,oracleactioncanbe con-structedby
uoi,t
=
E x|
yi,τ=
yi,τ τ≤t,
uoj,τ=
uoj,ττ≤t−1 j∈Ni (19)through the disclosure of oracle actions, i.e., local estimates. We remarkthat uo
i,t is linear inthe previous actions uoi,τ ,
τ
≤
t−
1.Inordertoextractnewinformation,i.e.,innovationpart,weneed to eliminate the previously received information at each instant ontheneighboringagents.Thisbringsinadditionalcomputational complexity.On thecontrary, agentscan justsend thenew infor-mationcomparedtothepreviouslysentinformation,e.g.,si,t.Note that here agents disclose the same informationto the neighbor-ing agents.Sincewe are conditioningon thelinear combinations oftheconditionedvariableswithouteffectingtheirspannedspace, i.e., si,t is computable from uoi,τ for
τ
≤
t andvice versa, agentscanstillachievetheoracleperformancebyreducedcomputational load,yet.
Attimet,agent-i receiveslocalmeasurement yi,t andsent in-formationfromtheneighboringagents,
ri,t
:=
col{
sj1,t−1, . . . ,
sjπi,t−1}.
Weaimto determinethecontent ofthereceived informationri,t toextract the innovationwithin them andutilizethisinnovation intheupdateoftheoracleaction.
Initially,attimet
=
1,agent-i hasonlyaccesstothelocal mea-surementyi,1.Then,theoracleactionisgivenbyuoi,1
= (
I−
xHi(
HixHi
+
ni)
− 1H i)
¯
x+
xHi(
HixHi
+
ni)
− 1y i,1.
Letuoi,0
= ¯
x andˆ
i,0=
x,andsetBi,1= ˆ
i,0Hi(
Hiˆ
i,0Hi+
ni)
−1 andAi,1
=
I−
Bi,1Hi.Then,weobtainuoi,1
=
Ai,1uoi,0+
Bi,1yi,1,
ˆ
i,1=
Ai,1ˆ
i,0.
Next, instead of sending uo
i,1, agent-i sends to the neighboring agents, j
∈
N
i,si,1
=
uoi,1−
Ai,1uoi,0=
Bi,1yi,1.
Correspondingly, attime t
=
2, agent-i receives yi,2 andri,2. Letri,2 be the corresponding random vector. Then, conditioning the stateandthereceivedinformationonthepreviously available in-formation yi,1
=
yi,1,wehave x yi,2 ri,2yi,1
=
yi,1∼ N
⎛
⎝
⎡
⎣
u o i,1 Hiuoi,1 ¯ Hi,1uoi,1⎤
⎦ ,
⎡
⎣
ˆi,1 ˆi,1H i ˆi,1H¯i,1 Hiˆi,1 Hiˆi,1Hi+ni Hiˆi,1H¯i,1 ¯ Hi,1ˆi,1 H¯i,1ˆi,1Hi H¯i,1ˆi,1H¯i,1+ ¯Gi,1⎤
⎦
⎞
⎠ ,
where H¯
i,1:=
col Bj1,1Hj1, . . . ,
Bjπi,1Hjπi andG¯
i,1=
diag{
Gi,1}
, whereGi,1:=
col Bj1,1nj1Bj1,1
, . . . ,
Bjπi,1njπiB jπi,1 . LetH
˜
i,1:=
col{
Hi,
H¯
i,1}
andG˜
i,1:=
diag{
ni,
G¯
i,1}
andset!
Bi,2 Ci,2"
= ˆ
i,1H˜
i,1˜
Hi,1ˆ
i,1H˜
i,1+ ˜
Gi,1−1
,
Ai,2=
I−
Bi,2Hi−
Ci,2H¯
i,1.
Then,weobtain uoi,2=
Ai,2uoi,1+
Bi,2yi,2+
Ci,2ri,2,
(20)ˆ
i,2=
Ai,2ˆ
i,1 andagent-i sendssi,2
=
uoi,2−
Ai,2uoi,1=
Bi,2yi,2+
j∈Ni Ci(,j2)Bj,1yj,1=sj,1
,
whereC(i,j2)denotesthecorresponding jthblock ofCi,2.Therefore, attimet
=
3,agent-i receivesfrom j∈
N
i:sj,2
=
Bj,2yj,2+
k∈Nj\i
(
C(jk,2)Bj,1yj,1)
+
C(ji,)2Bi,1yi,1.
(21)Since thelast termontherighthandside of(21)isknownby i,
wehave E
[
ri,2|δ
io,2= δ
oi,2] =
= ¯Hi,2⎡
⎢
⎢
⎣
Bj1,2+$k∈Nj1\iC(j1k),2Bk,2..
.
Bjπi,2+$k∈Njπi\iC (k) jπi,2Bk,2⎤
⎥
⎥
⎦
uoi,2+
Di,2si,1,
(22) where Di,2:=
col{
C(ji1),2,
. . . ,
C (i) jπi,2}
,and Gi,2=
⎡
⎢
⎢
⎣
Bj1,2n j1Bj1,2+ $ k∈Nj1\iC( k) j1,2Bk,1nkBk,1(C (k) j1,2)..
.
Bjπi,2n jπ iB jπi,2+ $ k∈Njπi\iC (k) jπi,2Bk,1nkBk,1(C (k) jπi,2)⎤
⎥
⎥
⎦
(23)By(20),(22),and(23),thenextoracleactionuoi,3 isgivenby uoi,3
=
Ai,3uoi,2+
Bi,3yi,3+
Ci,3(
ri,3−
Di,2si,1).
Subsequently, agent-i sends si,3
=
uoi,3−
Ai,3uoi,2 andthe received informationfrom j∈
N
i yieldssj,3
=
Bj,3yj,3+
k∈Nj C(jk,3)⎛
⎝
Bk,2yk,2+
l∈Nk\j Ck(l,)2Bl,1yl,1⎞
⎠
=
Bj,3yj,3+
k∈Nj\i C(jk,3)⎛
⎝
Bk,2yk,2+
l∈Nk\j Ck(l,)2Bl,1yl,1⎞
⎠
+
C(ji,)3 si,2−
Ci(,j2)sj,1.
Then, H¯
i,3 isgivenby¯
Hi,3=
⎡
⎢
⎣
Bj1,3Hj1+Cj1,3H¯j1,2..
.
Bjπi,3Hjπi+Cjπi,3H¯jπi,2⎤
⎥
⎦ −
⎡
⎢
⎢
⎣
C(i) j1,3H¯ (i) j1,2..
.
C(jiπ) i,3H¯ (i) jπi,2⎤
⎥
⎥
⎦ .
(24) Correspondingly,wehave Gi,3=
⎡
⎢
⎣
Bj1,3n j1Bj1,3+Cj1,3G¯j1,2Cj1,3..
.
Bjπi,3n jπiBjπi,3+Cjπi,3G¯jπi,2Cjπi,3⎤
⎥
⎦
−
⎡
⎢
⎢
⎣
C(j1i),3G(j1i),2(C(j1i),3)..
.
C(ji) πi,3G (i) jπi,2(C (i) jπi,3)⎤
⎥
⎥
⎦ .
(25)Therefore,theoracleactioncanbewrittenas
uoi,4
=
Ai,4uoi,3+
Bi,4yi,4+
Ci,4(
ri,4−
Di,3si,2+
Ti,3ri,2),
where Ai,4,
Bi,4,
Ci,4,
Di,3 aredefinedaccordinglyandTi,3
:=
⎡
⎢
⎢
⎣
C(j1i),3C(i,j12). .
.
C(jiπ) i,3C (jπi) i,2⎤
⎥
⎥
⎦ .
Following identical steps, for t
≥
1, the OEDOL algorithm is givenbyuoi,t
=
Ai,tuoi,t−1+
Bi,tyi,t+
Ci,twi,t,
(26)ˆ
i,t=
Ai,tˆ
i,t−1,
(27)wherewi,t istheinnovationtermextractedfromthereceived in-formation,whichevolvesaccordingto
wi,t
=
ri,t−
Di,t−1si,t−2+
Ti,t−1wi,t−2.
(28) Theweightingmatrices Ai,t, Bi,t,Ci,t,Di,t,andTi,t aredefinedby!
Bi,t Ci,t"
= ˆ
i,t−1H˜
i,t−1×
˜
Hi,t−1ˆ
i,t−1H˜
i,t−1+ ˜
Gi,t−1−1
,
(29) Ai,t=
I−
Bi,tHi−
Ci,tH¯
i,t−1,
(30) Di,t=
col C(ji) 1,t, . . . ,
C (i) jπi,t (31) Ti,t=
⎡
⎢
⎢
⎣
C(i) j1,tC (j1) i,t−1 ··· 0..
.
. .
.
..
.
0 ··· C(jiπ) i,tC (jπi) i,t−1⎤
⎥
⎥
⎦ ,
(32)where H
˜
i,t=
col{
Hi,
H¯
i,t}
, G˜
i,t=
diag{
ni,
G¯
i,t}
and G¯
i,t=
diag
{
Gi,t}
.By(24)and(25),theintermediateparameters H¯
i,t andGi,t evolveaccordingto
¯
Hi,t=
⎡
⎢
⎣
Bj1,tHj1+Cj1,tH¯j1,t−1..
.
Bjπi,tHjπi+Cjπi,tH¯jπi,t−1⎤
⎥
⎦ −
⎡
⎢
⎢
⎣
C(i) j1,tH¯ (i) j1,t−1..
.
C(jiπ) i,tH¯ (i) jπi,t−1⎤
⎥
⎥
⎦ ,
(33) Gi,t=
⎡
⎢
⎣
Bj1,tn j 1B j1,t+Cj1,tG¯j1,t−1Cj1,t..
.
Bjπi,tn jπiBjπi,t+Cjπi,tG¯jπi,t−1Cjπi,t⎤
⎥
⎦
−
⎡
⎢
⎢
⎣
C(i) j1,tG (i) j1,t−1 C(i) j1,t..
.
C(jiπ) i,tG (i) jπi,t−1 C(jiπ) i,t⎤
⎥
⎥
⎦
(34)and we initialize the parameters as H
¯
j,τ=
0 and Gj,τ=
0 forτ
<
1.Then,agent-i sends si,t=
uoi,t−
Ai,tuoi,t−1.
ThedetaileddescriptionofthealgorithmisprovidedinTable 1.
3.4.Computationalcomplexity
In (26), the combination matrices Ai,t
,
Bi,t, Ci,t, Di,t−1, and Ti,t−1 are independent of the streaming data although they are time-varying. Hence they can be computed before-hand. In that case,thecomputationalcomplexityoftheiterationsforeachagent is dominated by the term Ci,twi,t. Therefore, the average com-putational complexity is on the order of p2π
2, whereπ
2:=
1/
m$
mi=1π
2i, i.e., O
(
p2
π
2)
. Otherwise, the computational com-plexity of the algorithm is mainly dominated by the matrix in-version in (29), note that G¯
i,t∈ R
pπi×pπi, unless the network is sparsely connected, i.e.,π
im for i=
1,
. . . ,
m. Therefore, overTable 1
ThedescriptionoftheOEDOLalgorithm. Algorithm: The OEDOL Algorithm. Initialization: For i=1 to m do uo i,0= ¯x, ˆi,0= x, ¯ Hi,τ=0, Gi,τ=0, and wi,τ=0 forτ<1 End for Iterations: Do for t≥1 For i=1 to m do Construction of Weights: For j=1 to m do
CalculateH¯j,tand Gj,tby(33)and(34).
Determine combination matrices via(29)–(32). End for
Construct ri,tthrough received sj,t−1for j∈ Ni
Extraction of Innovation: wi,t=ri,t−Di,t−1si,t−2+Ti,t−1wi,t−2 Update: uo i,t=Ai,tuoi,t−1+Bi,tyi,t+Ci,twi,t ˆi,t=Ai,tˆi,t−1
Disclose si,t=uoi,t−Ai,tuoi,t−1to the neighbors.
End for
Table 2
Acomparisonofthecomputationalcomplexitiesoftheproposedalgorithms. Algorithm Without weights Pre-computed weights ODOL O(qm)3 O(qm)2
OEDOL Omp3π3 Op2π2
a non-sparse network, theaverage complexity is on theorder of
mp3
π
3 (π
3 is defined accordingly) since each agent i computes Bj,t andCj,t for j=
1. . . ,
m.Inparticular,thecomplexityisgiven by Omp3π
3,whileitis O(
qm)
3fortheODOLalgorithm.Note that over tree networks, we have m−
1 edges and correspond-inglyaverageneighborhoodsizeissmall.Hence,disclosureoflocal estimatesovertreenetworksalsoreducesthecomputational com-plexitycomparedto thetime-stampeddisclosurestrategy in gen-eral(inadditiontothesubstantialreductionincommunication).InTable 2,we tabulatethe computationalcomplexities ofthe intro-ducedalgorithms.
Wepointoutthatdiffusionorconsensusbasedalgorithmshave relatively low complexity, i.e., on the order of p2 in the least-mean-square based algorithms andon the order of p3 in quasi-Newtonbasedalgorithms,sinceexchangedinformationishandled irrespective ofthe content. Such algorithms also present appeal-ing performance for certain applications in addition to the low computational complexity. However, they do not achieve the or-acleperformance.
In the following, we analyze whether the agents can still achievetheoracleperformancethroughtheexchangeoflocal esti-matesoverthenetworksnotintreestructure.
3.5. Treenetworksinvolvingcellstructures
While constructing thespanning tree, we cancelcertain com-municationlinksin ordertoavoidmulti-pathinformation propa-gation.However,wealsoobservethatinafullyconnectednetwork agents can achieve the oracle performance through the disclo-sureoflocalobservations.Inparticular,sincealloftheagentsare connected,eachagentcanreceivetheobservationsacrossthe net-work directly. Correspondingly, in a fully connectednetwork, we can achieve identicalperformance withthe ODOLalgorithm only throughthe disclosureofthelocalestimatesasstatedinthe fol-lowingcorollaryformally.
Fig. 5. An example tree network involving cell structures.
Corollary3.1.Considertheteam-problemoverafullyconnected net-work,inwhichr
=
p.Then,exchangeoflocalestimatescanleadtothe team-optimalsolution,i.e.,agentscanachievetheoracleperformance.Proof. Overa fullyconnectednetwork,
κ
i=
1 andthe oracle ac-tionisgivenby uoi,t=
E&
x|{
yi,τ=
yi,τ}
τ≤t,
{
yj,τ=
yj,τ}
τj∈N≤t−i1'
(35)andwecanalsoobtain(35)by(11)sincewehave
oj,i,t
= δ
oj,t\ {δ
io,t\ {
yi,t}},
which implies that
oj,i,t is constructible from
δ
io,τ andδ
oj,τ forτ
≤
t.Theproofisconcluded.2
We point out that the team-optimal strategies can fail if a link or node failure occurs. However, once a link failure is de-tected,team-optimalstrategiescan berecomputedbyeliminating thefailed linkin thenewnetwork configuration.Hence, through suchstrategies,wecanincreasetherobustnessoftheteamagainst linkandnodefailures.
We define a “cell structure” as a sub-network in which all agentsare connectedto each other.Intuitively, consideringa cell structureasa“single”agent,thecell(i.e.,alltheagentsinthecell) canbe involvedinthe treesuch thatthe agentscanstill achieve the oracle performance through thedisclosure of local estimates (althoughtheremaybe loopsinthecell). Welist thefeatures of thecellstructures,e.g.,seeninFig. 5,asfollows:
•
Agentsoutofacellcanconnect toatmostoneoftheagents withinthatcell.•
Acellstructureconsistsofatleast2agents.•
Anagentcanbelongtomorethanonecell.•
Twodifferentagentscannot belong to morethan one cell at thesametime.•
Alloftheagentsbelongtoatleastacellina connected net-work.•
Each agent has also the knowledge ofthe cells of the other agents.•
Each agent labels its cells from its own and its first order neighbor’spoint ofview.As anexample,foragent-i,C
i,i1de-notesthe cell involving both i and i1. Notethat ifthe same cellalsoincludesi2,
C
i,i1=
C
i,i2.Thefollowingtheoremshowsthatagentscanachievetheoracle performance over treenetworksinvolvingcell structuresthrough thedisclosureofthelocalestimates.
Theorem3.2.Considertheteamproblemovertreenetworksinvolving cellstructures.Then,exchangeoflocalestimatescanleadtothe team-optimalsolution,i.e.,agentscanachievetheoracleperformance.
Proof. Initially, we have
oj,i,1
= δ
j,1= {
yj,1}
and the oracle action is also given by (11) over this network topology. Note that the information received by j at t=
2 is given byδ
oj,2=
yj,2,
=ok, j,0 yk,1 k∈Nj,
δ
oj,1,
whichyieldsoj,i,2
= δ
oj,2\
⎧
⎨
⎩
δ
oj,1∪
k∈Ci,j yk,1⎫
⎬
⎭
,
= δ
o j,2\
⎧
⎨
⎩
δ
oj,1∪
k∈Ci,jok,j,1
⎫
⎬
⎭
ando
j,j,t
= ∅
bydefinition.Duetothecellstructure,wehave k∈Cj,iko,j,1
=
oi,j,1∪
k∈Ci,j\jok,i,1
,
= δ
o i,1∪
k∈Ci,j\jδ
ko,1.
Correspondingly,fort
>
0 wehaveoj,i,t
= δ
oj,t\
⎧
⎨
⎩
δ
oj,t−1∪
oi,j,t−1∪
k∈Ci,j\jok,i,t−1
⎫
⎬
⎭
(36) ando
j,i,t isconstructible bythesets
δ
io,τ andδ
oj,τ for j∈
N
i andτ
≤
t.Note that overtree networks,C
i,j\
j= ∅
fori=
1,
. . . ,
m,j
∈
N
i,and(36)leadsto(18).Hence,fort>
0 weobtain(19)and theproofisconcluded.2
Notethatthenetworkcanhaveloopswithinthecellstructures and agents can still achieve the oracle performance through the diffusionofthelocalestimates.Thisincreasestherobustnessofthe teamstrategiesagainstthelinkfailures.Inthesequel,weprovide thesub-optimalextensionsofthealgorithmsforpractical applica-tions.
4. Sub-optimalapproaches
Minimization of the cost function (4) optimally requires rel-atively excessive computations. We aim to mitigate the problem sub-optimally yet in a computationally efficient approach while achieving comparable performance withthe optimal case. As an example,we canapproximatethe costmeasure (4)through To
≥
max{
κ
i}
i=1,...,msizetime-windowingasfollowsT
t=1 min ηi,t∈si,t Ex−
η
i,t(δ
is,t)
2,
(37) whereagent-i hastheinformationsetδ
is,t:=
yi,τ t−To<τ≤t
,
y j,τ t−To<τ≤t−1 j∈Ni(1), . . . ,
yj,τ t−To<τ≤t−κi j∈N(κi) i,
(38)if t