A Broad Dynamical Model for Pattern Formation by Lateral Inhibition Murat Arcak∗ January 9, 2012 2 1 0 2 n a 1 Introduction J 6 Spatial patterns ofgene expression arecentral to thedevelopment ofmulti-cellular organisms. Most math- ] S ematical studies of pattern formation investigate diffusion-driven instability, which is a mechanism that D amplifies spatial inhomogeneities in a class of reaction-diffusion systems (see, e.g., [1]). However, many . patterning events in multi-cellular organisms rely on cell-to-cell contact signaling, such as the Notch path- h t way [2], and do not involve diffusible proteins for intercellular communication. A particularly interesting a m phenomenon inthisformofcommunication islateral inhibition wherebyacellthatadoptsaparticular fate [ inhibits its immediate neighbors from doing likewise [3], thus leading to ‘fine-grained’ patterns. There is 1 increasing interest inunderstanding theNotchsignaling circuitry inmammaliancellsthatleadstosuchlat- v eral inhibition [4, 5]. Recent studies showed thatalateral inhibition pathway also functions inE.Coli, and 7 7 enablesthebacteria toinhibitthegrowthofotherE.Colistrains indirectcontact[6]. 4 Dynamical models areofgreatinterest forunderstanding the circuit topologies involved inlateral inhi- 1 . bition and for predicting the associated patterns. Several simplified models have been employed for Notch 1 0 signallingpathwaysin[3]and[5]. Theobjectiveofthispaperistopresentanabstractdynamicalmodelthat 2 captures the essential features of lateral inhibition and to demonstrate with dynamical systems techniques 1 : thatthesefeaturesindeedleadtopatterning. AlthoughthismodelisnotmeantspecificallyforNotchsignal- v i ing,itencompasses asspecialcasesthelateralinhibition modelin[3]aswellasaslightly modifiedversion X oftheonein[5]. r a Ourmodeltreatstheevolutionofconcentrations ineachcellasaninput-outputsystem,wheretheinputs represent theinfluenceofadjacentcellsandtheoutputscorrespond totheconcentrations ofthespecies that interact with adjacent cells. The input-output models for the cells are then interconnected according to an undirected graph where the nodes represent the cells, and the presence of a link between twonodes means that the corresponding cells are in contact. The main assumption on the input-output model is that each constantinputyieldsauniqueandgloballyasymptoticallystablesteady-state,andthatthevalueoftheoutput at this steady-state is a decreasing function of the input. This decreasing property captures the inhibition of the cell function by its neighbors. The model allows multiple inputs and outputs, and is restricted by a monotonicity assumption, following the definition of monotonicity for dynamical systems with inputs and outputs[7]. ∗Department of Electrical Engineering and Computer Sciences, University of California, Berkeley. Email: ar- [email protected]. 1 Using this model, wefirstgive aninstability condition for the homogeneous steady-state, applicable to arbitrary contact graphs. We then focus our attention on bipartite graphs, and demonstrate the emergence of a “checkerboard” pattern, exhibiting alternating high and low values of concentrations in adjacent cells. Next, we establish a strong monotonicity property of the interconnected model for bipartite graphs, which implies that almost every bounded solution (except for a measure-zero set of initial conditions) converges to a steady-state [8, 9]. A graph is bipartite if and only if it contains no odd-length cycles, and Cartesian products of bipartite graphs are also bipartite [10]. Thus, the results of this section are applicable, among others, to grid graphs (one dimensional path graphs and their Cartesian products in higher dimensions) whichareappropriate forrepresenting arraysofcells. 2 Lateral Inhibition Model and Preliminaries Welet G be anundirected, connected graph where the nodes represent the cells, and the presence of a link betweentwonodesmeansthatthecorrespondingcellsareincontact. Inpreparationforthedynamicalmodel studiedbelow,weletN denotethenumberofcellsanddefinethematrix P∈RN×N: d−1 ifnodesiand jareadjacent, p = i (1) ij 0 otherwise, whered denotesthedegreeofnodei. Itfollowsthat Pisanonnegative row-stochastic matrix,thatis: i P1=1 (2) where 1 denotes the vector of ones. The matrix P is identical to the probability transition matrix for a random walk on the graph G. Theproperties summarized below therefore follow from standard results for randomwalks(see,e.g.,[11]): Lemma 1. P possesses real eigenvalues λ ≤···≤ λ all of which lie in the interval [−1,1], and corre- N 1 sponding real, linearly independent eigenvectors v, i = 1,···,N. In particular, λ = 1, and v = 1 is a i 1 1 corresponding eigenvector. IfGisbipartite,thenλ =−1,andaneigenvectorv issuchthattheentriesare N N either1or−1,andtwoentries corresponding toadjacent nodeshaveopposite signs. Leti=1,···,N denotethecells,andconsider thedynamicalmodel: x˙i= f(xi,ui) yi=h(xi) (3) where xi∈X ⊂Rnisavectordescribingthestateofreagentconcentrations incelli,ui∈U ⊂Rm describes the ‘input’ from adjacent cells, and yi ∈Y ⊂Rm describes the ‘output’ that serves as an input to adjacent cells. Inparticular, U =(P⊗I )Y (4) m wherePisasdefinedin(1),U :=[u1T···uNT]T andY :=[y1T···yNT]T. Iffollowsfrom(1)thattheinputui istheaverage oftheoutputs yk overallneighbors kofcelli. Thus, wehenceforth taketheinput andoutput spacestobeidentical: U =Y . Weassumethat f(·,·)andh(·)arecontinuously differentiable andfurthersatisfythefollowingproperty: Assumption 1. For each constant input u∗, system (3) has a globally asymptotically stable steady-state x∗:=S(u∗)withtheadditional property that: ∂f(x,u) det ,0. (5) ∂x (cid:12)(cid:12)(x,u)=(x∗,u∗)! (cid:12) (cid:12) (cid:12) 2 ThemapS :U →X and,therefore, themapT :U →U defined by: T(·):=h(S(·)), (6) arecontinuously differentiable. Following the terminology in [7], we will refer to S(·) as the input-state characteristic, and to T(·) as theinput-output characteristic. Ournextassumption isthat(3)isamonotone system inthesense of[7],as definedbelow. Accordingtotheclassicaldefinition forsystemswithoutinputsandoutputs [9],amonotone system is one that preserves a partial ordering of the initial conditions as the solutions evolve. The partial orderingisdefinedwithrespecttoapositivitycone K intheEucledeanspacethatisclosed,convex,pointed (K∩(−K)={0}), and has nonempty interior. Givensuch acone, x(cid:22) xˆ means xˆ−x∈K, x≺ xˆ means x(cid:22) xˆ and x, xˆ, and x≪ xˆ means that xˆ−x is in the interior of K. The system x˙ = f(x) is then defined to be monotone iftwosolutions x(t)and xˆ(t)starting withtheorder x(0)(cid:22) xˆ(0)maintain x(t)(cid:22) xˆ(t)forall1 t≥0. Themorerestrictivenotionofstrongmonotonicitystipulatesthat x(0)≺xˆ(0)implies x(t)≪xˆ(t)forallt>0. Themonotonicity conceptwasextended tosystemswithinputsandoutputs in[7]: Definition 1. Given positivity cones KU,KY,KX for the input, output, and state spaces, the system x˙ = f(x,u), y=h(x) is said to be monotone if x(0)(cid:22) xˆ(0) and u(t)(cid:22)uˆ(t) for all t ≥0 imply that the resulting solutions satisfy x(t)(cid:22) xˆ(t)forallt≥0,andtheoutputmapissuchthat x(cid:22) xˆimpliesh(x)(cid:22)h(xˆ). Assumption2. Thesystem (3)ismonotonewithrespecttoKU =Rm , KY =−KU,and KX =K,where K is ≥0 somepositivity coneinRn. Asobserved in[7,RemarkV.2], monotonicity implies thattheinput-state andinput-output characteris- ticsarenondecreasingwithrespecttothesameordering;thatis,u(cid:22)uˆwithrespecttoKU impliesS(u)(cid:22)S(uˆ) withrespecttoKX andT(u)(cid:22)T(uˆ)withrespecttoKY. SinceKY =−KU inAssumption2,weconcludethat T(·)isnonincreasing withrespect tothestandard orderinduced by KU =Rm . Thisnonincreasing property ≥0 means that, if two cells are in contact, an increase in the output value of one has the opposite effect on the other, which is why (3)-(4) is referred to as a “lateral inhibition” model. We note from the nonincreasing property ofT(·)that: ∂T(u) T′(u):= (7) ∂u isanonpositive matrixinRm×m,anddenoteitsspectral radiusas: ρ(T′(u)). (8) Weconcludethissectionbyquotinglemmasthatwillbeusedinthesequel. Lemmas2and3arefrom[12]: Lemma 2. Given the system x˙ = f(x,u), y=h(x) with continuously differentiable f(·,·) and h(·), the lin- earization x˙=Ax+Bu, y=Cx about a point (x∗,u∗) satisfying f(x∗,u∗)=0 is also monotone with respect tothesamepositivity cones. Lemma3. Thelinearsystem x˙=Ax+Bu,y=Cxismonotoneifandonlyif: 1) x∈KX implies Ax∈KX, 2)u∈KU implies Bu∈KX, 3) x∈KX impliesCx∈KY. 1Here,“forallt”isunderstoodas“foralltimestthatbelongtothecommondomainofexistenceofthetwosolutions.” 3 Thefollowinglemma,provenin[12]forsingle-input, single-output systemsandextendedin[13]tothe multivariablecase,determinesstabilityofapositivefeedbacksystembasedonthe‘dcgain’oftheopen-loop system: Lemma 4. Suppose the linear system x˙= Ax+Bu, y=Cx is monotone with respect to cones KU,KY,KX suchthat KU =KY andAisHurwitz. If−(I+CA−1B)isHurwitz,thensoisA+BC. If−(I+CA−1B)hasan eigenvalue withapositiverealpart,thensodoes A+BC. Inthespecialcaseofsingle-input, single-output systems,thestabilitycondition aboveamountstochecking whether the dc gain −CA−1B is greater or smaller than one. In the multi-input, multi-output case, this condition is equivalent to inspecting whether the spectral radius of the dc gain matrix is greater or smaller thanone. Thefollowingtestfrom[7,14]isusefulforcertifying monotonicity withrespecttoorthantcones: Lemma 5. Consider the system x˙ = f(x,u), y = h(x), x ∈ X ⊂ Rn, u ∈ U ⊂ Rm, y ∈ Y ⊂ Rp, where the interiors of X and U are convex, and f(·,·) and h(·) are continuously differentiable. If there exist ǫ ,···,ǫ ,δ ,···,δ ,µ ,···,µ ∈{0,1}suchthat: 1 n 1 m 1 p ∂f (−1)ǫj+ǫk j(x,u)≥0 ∀x∈X,∀u∈U,∀j,k (9) ∂x k ∂f (−1)ǫj+δk j(x,u)≥0 ∀x∈X,∀u∈U,∀j,k (10) ∂u k ∂h (−1)ǫj+µk k(x,u)≥0 ∀x∈X,∀j,k, (11) ∂x j then the system is monotone with respect to the positivity cones KU ={u∈Rm | (−1)δjuj ≥0}, KX ={x∈ Rn |(−1)ǫjxj≥0},KY ={y∈Rp |(−1)µjyj≥0}. 3 Instability of the Homogeneous Steady-State Notethatsystem(3)-(4)admitsspatiallyhomogeneous solutionsoftheform xi(t)=x(t),i=1,···,N,where x(t)satisfies: x˙ = f(x,h(x)). (12) Inparticular, ifthemapT(·)hasafixedpoint: u∗=T(u∗), (13) then(12)admitsthesteady-state: x∗=S(u∗). (14) Forsingle-input,single-outputsystemswithU =R ,thenonincreasingpropertyofthemapT :R →R ≥0 ≥0 ≥0 indeedguarantees auniquefixedpointu∗ in(13). The “lumped model” (12) describes the dynamics of the Nn-dimensional system (3) reduced to the n- dimensional invariant subspace wherethesolutions arespatially homogeneous. Thus,thesteady-state x∗ of thelumpedmodeldefinesthehomogeneoussteady-state xi=x∗,i=1,···,N,forthefullsystem(3)-(4). Asa startingpointfortheanalysisofpatternformation,wenowgiveaninstabilitycriterionforthehomogeneous steady-state: 4 Theorem1. Considerthesystem(3)-(4)andsupposeAssumptions1and2hold. Letλ denotethesmallest N eigenvalue ofPasinLemma1,andletu∗,x∗ beasin(13),(14). If: λ ρ T′(u∗) <−1, (15) N (cid:0) (cid:1) thenthehomogeneous steady-state xi=x∗,i=1,···,N,isunstable. Proof: LetX:=[x1T···xNT]T,andnotethatthelinearization of(3)-(4)aboutthehomogeneous steady-state [x∗T,···,x∗T]T givestheJacobian matrix: I ⊗A+P⊗(BC) (16) N where: ∂f(x,u) ∂f(x,u) ∂h(x) A:= , B:= , C:= . (17) ∂x (cid:12)(cid:12)(x,u)=(x∗,u∗) ∂u (cid:12)(cid:12)(x,u)=(x∗,u∗) ∂x (cid:12)(cid:12)x=x∗ (cid:12) (cid:12) (cid:12) WerecallfromLemma1that (cid:12)(cid:12) (cid:12)(cid:12) (cid:12)(cid:12) λ 1 V−1PV =Λ:= ... , (18) whereV =[v ···v ],andapplythefollowingsimilaritytransformλNation to(16): 1 N (V−1⊗I )[I ⊗A+P⊗(BC)](V⊗I )=I ⊗A+Λ⊗(BC). (19) n N n N Thismatrixisblock-diagonal, withthekthdiagonal blockgivenby: A+λ BC. (20) k Claim: If λ ρ T′(u∗) <−1, (21) k (cid:0) (cid:1) then(20)hasapositiveeigenvalue. The theorem follows from this claim because, if (15) holds, then (20) has a positive eigenvalue for k = N, which implies instability. To prove the claim, we note from Assumption 2 and Lemma 2 that the linearsystem x˙=Ax+Bu,y=CxismonotonewithrespecttoKU =Rm ,KY =−KU,andKX =K. Wewrite ≥0 A+λ BC=A+BC whereC :=λ Candnotethat(21)impliesλ <0. Thus,thesystem x˙=Ax+Bu,y=C x k k k k k k is monotone with KU =KY. In addition, Assumptions 1 and 2 imply that A is Hurwitz, as can be deduced from [12,Lemma6.5]. Thus,itfollowsfrom thesecond statement ofLemma4thatif−(I+C A−1B)hasa k positiveeigenvalue, thensodoes(20). Theremainingtaskisthustoprovethat −(I+C A−1B)=−I−λ CA−1B (22) k k hasapositiveeigenvalue. Tothisend,wefirstshowthat T′(u∗)=−CA−1B. (23) Since f(S(u),u)≡0, (24) differentiation gives: ∂f(x,u) ∂S(u) ∂f(x,u) + =0. (25) ∂x (cid:12)(cid:12)x=S(u) ∂u ∂u (cid:12)(cid:12)x=S(u) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) 5 Next,itfollowsfromthedefinition(6)that ∂h(x) ∂S(u) T′(u)= . (26) ∂x (cid:12)(cid:12)x=S(u) ∂u (cid:12) (cid:12) Combining(25)and(26),andsubstituting (17),weve(cid:12)rify(23). Substituting (23),wethenrewrite(22)as −I+λ T′(u∗), (27) k andconclude thatitindeed hasapositive eigenvalue, because λ <0impliesthatλ T′(u∗)isanonnegative k k matrix and (21) implies that its spectral radius exceeds one. Since the spectral radius is an eigenvalue for nonnegative matrices(see,e.g.,[15]),theconclusion follows. (cid:3) The eigenvectors v of P used in the similarity transformation (19) may be interpreted as the spatial k modes of the system. Thus, the stability properties of the matrix (20) for each k determines whether the corresponding mode decays orgrowsintime. Since thespectral radius isnonnegative andλ ,k=1,···,N, k areindecreasingorder,whenevertheinstabilitycriterion(21)holdsforaparticularmodek,italsoholdsfor highervaluesofk. Becauselargerwavenumberskimplyhigherspatialfrequencycontentinv ,weconclude k thattheinstability condition abovesetsthestagefortheformationofhigh-frequency spatialpatterns. 4 Patterning in Bipartite Graphs 4.1 Emergence ofCheckerboard Patterns Forbipartitegraphs, whereλ =−1asstatedinLemma1,theinstability condition inTheorem1is: N ρ(T′(u∗))>1. (28) Thiscondition indicates thegrowthofthehighest spatial-frequency modev whichexhibits opposite signs N foradjacentnodes. Thus,concentrations inadjacentnodesmoveinoppositedirectionsinthevicinityofthe homogeneous steady-state. Wenowshowthat,ifthemap T2(·):=T(T(·)) (29) hastwofixedpointsu ,u otherthanu∗,satisfying: 1 2 u =T(u ), u =T(u ), (30) 1 2 2 1 then the system (3)-(4) has an inhomogeneous steady-state with two sets of concentrations, each assigned to one of two adjacent cells. We will refer to this steady-state as a “checkerboard” pattern, since adjacent cells adopt distinct states. Although this term may be associated with cells arranged as agrid graph in two dimensional space, wewilluseitbroadlyforanyspatial arrangement thatformsabipartitegraph. Proposition1. LetGbeabipartitegraphandletthesetsI⊂{1,···,N}andI′={1,···,N}−Ibesuchthat notwonodesinthesamesetareadjacent. Ifthereexistu ∈U andu ∈U,u ,u ,satisfying (30),then 1 2 1 2 xi=S(u ), i∈I, xi=S(u ), i∈I′, (31) 1 2 and xi=S(u ), i∈I, xi=S(u ), i∈I′, (32) 2 1 aresteady-states forsystem(3)-(4). 6 Proof: Toshowthat(31)isasteady-state, wenotethat,ifi∈I,thenyi=T(u )and,ifi∈I′,thenyi=T(u ). 1 2 From (4), the input ui to a node in I is T(u ) because all neighbors of this node belong to I′. Likewise, 2 the input ui to a node in I′ is T(u ) because all neighbors of this node belong to I. Since T(u )=u and 1 2 1 T(u )=u ,weconclude that(31)isindeedasteady-state, andidentical arguments applyto(32). (cid:3) 1 2 Theorem2. Considerthesystem(3)-(4)andsupposeAssumptions1and2,andthehypotheses ofProposi- tion1hold. If,inaddition, ρ(T′(u )T′(u ))<1, (33) 1 2 thenthesteady-states (31)and(32)areasymptotically stable. Beforegivingtheproof,wenotethat(30)corresponds toaperiod-twoorbitofthediscrete-timesystem: u(t+1)=T(u(t)), (34) and (33)implies theasymptotic stability ofthis orbit. Likewise, (28)indicates instability ofthefixedpoint u∗forthisdiscrete-timesystem. Thus,aninterestingdualityexistsbetween(34)andthespatially-distributed system (3)-(4) defined on a bipartite graph: A bifurcation from a stable fixed point to a stable period-two orbitin(34)correspondstotheemergenceofstablecheckerboardpatternsfromahomogeneoussteady-state in(3)-(4). Inthesingle-input, single-output casewithU =R ,whereT :R →R isanonincreasing function ≥0 ≥0 ≥0 byAssumption2,condition(28)indeedimpliestheexistenceofaperiod-twoorbit(30). Toseethis,assume to the contrary that u∗ is the unique fixed point of T2(·). Since T(·) is continuous and nonincreasing, this uniquenesspropertywouldimplythatu∗isaglobalattractorforallsolutionsofthedifferenceequation(34) starting in R [16, Lemma1.6.5]. This, however, contradicts (28), which implies instability of u∗ for this ≥0 scalardifferenceequation. The argument above does not suggest the uniqueness of the pair (u ,u ), and multiple pairs satisfying 1 2 (30)mayexist. However,weclaimthatatleastonepairsatisfies: dT2(u) dT2(u) = =T′(u )T′(u )<1, (35) du (cid:12) du (cid:12) 1 2 (cid:12)(cid:12)u=u1 (cid:12)(cid:12)u=u2 (cid:12) (cid:12) (cid:12) (cid:12) whichisthescalarequivalent of(33),(cid:12)sinceT′(u )T′(cid:12)(u )isnonnegative. Toseethis,notefrom(28)that: 1 2 dT2(u) =T′(u∗)T′(u∗)>1 (36) du (cid:12) (cid:12)(cid:12)u=u∗ (cid:12) (cid:12) andsuppose,incontrastto(35),thatthederiva(cid:12)tiveofT2(·)isgreaterthanorequaltooneateachofitsfixed points. ThisimpliesthatT2(u)≥uforallu≥u∗,because T2(u)−uhasnonnegativeslopeatzero-crossings and, thus, remains nonnegative for u ≥u∗. The inequality T2(u) ≥ u implies unbounded growth of T2(·) whichisacontradiction becauseT(·)iscontinuous andnonincreasing, thus,bounded. ProofofTheorem2: LetNI and NI′ :=N−NI denotethecardinalities ofthesetsIandI′,andindexthe cellssuchthati=1,···,N belongtoI,andi=N +1,···,N belongtoI′. ThenthematrixPhastheform: I I 0 P P= 12 (37) P 0 21 7 where P12∈RNI×NI′,P21∈RNI′×NI. LetX:=[x1T···xNT]T,andnotethatthelinearization of(3)-(4)about (31)givestheJacobian matrix: I ⊗A P ⊗(B C ) NI 1 12 1 2 (38) P ⊗(B C ) I ⊗A where 21 2 1 NI′ 2 ∂f(x,u) ∂f(x,u) ∂h(x) A := , B := , C := , j=1,2. (39) j j j ∂x (cid:12)(cid:12)(x,u)=(S(uj),uj) ∂u (cid:12)(cid:12)(x,u)=(S(uj),uj) ∂x (cid:12)(cid:12)x=S(uj) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) From the definition(cid:12)(1), the matrix DP, where D (cid:12)is a diagonal matrix of the n(cid:12)ode degrees, is symmetric. Since D−1/2(DP)D−1/2=D1/2PD−1/2 isalsosymmetric,wewrite: 0 R D1/2PD−1/2= (40) RT 0 whereR∈RNI×NI′ isappropriately defined. Then,weapplythefollowingsimilaritytransformation to(38): I ⊗A P ⊗(B C ) I ⊗A R⊗(B C ) (D1/2⊗I ) NI 1 12 1 2 (D−1/2⊗I )= NI 1 1 2 . (41) n P ⊗(B C ) I ⊗A n RT⊗(B C ) I ⊗A 21 2 1 NI′ 2 2 1 NI′ 1 Thestructure of(40)issuchthatitcandiagonalized withanorthonormal matrixoftheform: Q Q Q 0 Q= 1 1 3 (42) Q −Q 0 Q 2 2 4 whichresultsin: Λ + 0 R −Λ+ where Λ is a diagonal matrix ofthReTstr0ictlyQp=osQitive eigenvalues0of0P, the columns of Q and Q s(p4a3n) + 3 4 the null spaces of RT and R, respectively, and the dimensions of the zero diagonal blocks in (43) are con- sistent with the dimensions of these null spaces (which we denote as n and n , respectively). From the 3 4 orthonormality of Q,wegettheidentities: 1 QTQ =QTQ = I (44) 1 1 2 2 2 r QTQ =I QTQ =I (45) 4 4 n4 3 3 n3 QTQ =0 QTQ =0, (46) 1 3 2 4 whereristhedimension ofΛ . Likewise,equation(43)implies: + RQ =Q Λ RTQ =Q Λ (47) 2 1 + 1 2 + RQ =0 RTQ =0. (48) 4 3 WenowreturntotheJacobian matrix(41)andfurtherapplythefollowingsimilarity transformation: 2QT ⊗I 0 1 n QT300⊗In 2QQT4T20⊗⊗IInn RTIN⊗I(⊗BA2C11) RI⊗NI(′B⊗1CA12) Q10⊗In Q20⊗In Q30⊗In Q40⊗In (49) 8 where the leftmost matrix is the inverse of the rightmost matrix from (44)-(46). Likewise, using (44)-(48), itisnotdifficulttoshowthattheproduct (49)equals: I ⊗A Λ ⊗(B C ) r 1 + 1 2 Λ+⊗(B2C1) Ir⊗A2 In3⊗A1 In4⊗A2 . (50) SinceAssumptions1and2implythatA andA areHurwitz[12,Lemma6.5],stabilityof(50)isdetermined 1 2 bytheupperleftblockswhich,uponasimilaritytransformationwithanappropriatepermutationmatrix,are block-diagonalized intorblocksoftheform: A λB C 1 i 1 2 (51) λB C A i 2 1 2 i=1,···,r. Wewillnowshowthat(51)isHurwitzforanyλ ∈[−1,1]. Sincealleigenvalues ofPlieinthisinterval i byLemma1,thiswillconcludetheproof. Wedonotprovideaseparateprooffortheasymptoticstabilityof (32), as identical arguments apply when the indices 1 and 2 are swapped in (51). If λ =0, (51)is Hurwitz i because A and A areHurwitz. Ifλ ,0,thenweapplythesimilaritytransformation: 1 2 i I 0 A λ B C I 0 A λ2B C 1 i 1 2 = 1 i 1 2 (52) 0 λ−1I λ B C A 0 λI B C A i i 2 1 2 i 2 1 2 andrewritetheresultas: A+BC (53) where A λ2B C 0 A:= 1 i 1 2 , B= , C=[C 0]. (54) 1 0 A B 2 2 Weclaimthatthelinearsystemdefinedbythetriplet(C,A ,B)ismonotonewithrespecttoKU =KY =Rm , ≥0 and KX =−K×K where K isasinAssumption2. Toseethis,firstnotefromLemma2that(C ,A ,B )and 1 1 1 (C ,A ,B ) are monotone with respect to the cones specified in Assumption 2. By Lemma 3, this means 2 2 2 that: x∈K ⇒ A x∈K, u∈Rm ⇒ B u∈K, x∈K ⇒ C x∈Rm , j=1,2. (55) j ≥0 j j ≤0 Wenowshowthattheconditions ofLemma3holdfor(C,A,B)withKU =KY =Rm ,KX =−K×K: ≥0 1)Suppose x=[xTxT]T ∈−K×K,thatis x ∈−K, x ∈K. Then, 1 2 1 2 A x +λ2B C x Ax= 1 1 i 1 2 2 ∈−K×K (56) A x 2 2 because, from(55), A x ∈−K,A x ∈K,C x ∈Rm and,hence, B C x ∈−K. 1 1 2 2 2 2 ≤0 1 2 2 2)Wewanttoshowthatu∈Rm impliesBu∈−K×K. FromthedefinitionofBin(54),Bu∈−K×K means ≥0 B u∈K. Itfollowsfromthesecondimplication in(55)thatu∈Rm indeedimplies B u∈K. 2 ≥0 2 3)Toprovemonotonicity withKY =Rm ,weneedtoshowthat x ∈−K and x ∈K implyC[xTxT]T ∈Rm . ≥0 1 2 1 2 ≥0 Thisisindeedtrue,sinceC[xTxT]T =C x and,from(55), x ∈−K impliesC x ∈Rm . 1 2 1 1 1 1 1 ≥0 9 Having verified the conditions of Lemma 3, we conclude that (C,A,B) is monotone with respect to KU =KY =Rm . In addition, the matrix A in (54) is Hurwitz, as A and A are Hurwitz. Thus, it follows ≥0 1 2 fromthefirststatement inLemma4that,if−(I+CA−1B)isHurwitz,thensois(53). Notethat A−1 −λ2A−1B C A−1 0 CA−1B=[C 0] 1 i 1 1 2 2 =−λ2C A−1B C A−1B (57) 1 0 A−1 B i 1 1 1 2 2 2 2 2 and,fromaderivation similartotheonefor(23),T′(u )=−C A−1B , j=1,2.Thus,(57)gives: j j j j −(I+CA−1B)=−I+λ2T′(u )T′(u ), (58) i 1 2 and(33)andλ ∈[−1,1]implythat−(I+CA−1B)isindeed Hurwitz. FromLemma4,this meansthat (53) i and,thus,(51)isHurwitzi=1,···,r,concluding theproof. (cid:3) 4.2 Generic Convergence to Steady-States Thusfarwehave studied local asymptotic stability properties ofthe steady-states. Strongly monotone sys- tems(as defined intheparagraph above Definition 1)have been shown topossess a“generic convergence” property [8, 9] which means that almost every bounded solution (except for a measure-zero set of initial conditions) converges to the set of steady-states. Below wefirst prove monotonicity of (3)-(4) in Theorem 3 and, next establish strong monotonicity in Theorem 4, thereby concluding generic convergence for this system. Theorem3. IfGisbipartite andAssumption2holds,thenthesystem(3)-(4)ismonotone. Proof: LetI⊂{1,···,N}andI′={1,···,N}−IbedefinedasinProposition 1,andsuppose thatin(3),the cells are indexed such that i=1,···,N belong to I, and i= N +1,···,N belong to I′ as in the proof of I I Theorem 2, where NI is the cardinality of set I. Let XI :=[x1T···xNIT]T, XI′ :=[xNI+1T···xNT]T, and defineUI,UI′,YI,YI′ similarly. Then,theinterconnection condition (4)becomes: UI = (P ⊗I )YI′ (59) 12 m UI′ = (P ⊗I )YI (60) 21 m where P and P areasin(37). Ablockdiagram illustrating thisinterconnection isdepicted inFigure1. 12 21 x1 U:=UI x2 YI UI′ YI′ Y P ⊗I P ⊗I ... 21 ... 12 xNI xN Figure1: Ablockdiagramforthesystem(3)-(4)whenthecontactgraphisbipartiteandthecorrespondingintercon- nectionmatrixPisdecomposedasin(37). To prove the monotonicity of this feedback system, we establish the monotonicity of the feedforward systemwithinputU:=UI andoutputY:=(P ⊗I )YI′: 12 m 10