Interpolation synthesis for quadratic polynomial inequalities and combination with EUF TingGan1,LiyunDai1,BicanXia1,NaijunZhan2,DeepakKapur3,andMingshuaiChen2 1 LMAM&SchoolofMathematicalSciences,PekingUniversity {gant,dailiyun,xbc}@pku.edu.cn, 6 2 StateKeyLab.ofComputerScience,InstituteofSoftware,CAS 1 [email protected] 0 2 3 DepartmentofComputerScience,UniversityofNewMexico [email protected] v o N Abstract. Analgorithmforgeneratinginterpolantsforformulaswhicharecon- 0 junctionsofquadraticpolynomialinequalities(bothstrictandnonstrict)ispro- 1 posed. The algorithm is based on a key observation that quadratic polynomial inequalitiescanbelinearizediftheyareconcave.AgeneralizationofMotzkin’s ] O transpositiontheoremisproved,whichisusedtogenerateaninterpolantbetween twomutuallycontradictoryconjunctionsofpolynomialinequalities,usingsemi- L definiteprogrammingintimecomplexityO(n3+nm))withagiventhreshold, . s wherenisthenumberofvariablesandmisthenumberofinequalities.Using c the framework proposed by [21] for combining interpolants for a combination [ ofquantifier-freetheorieswhichhavetheirowninterpolationalgorithms,acom- 3 binationalgorithmisgivenforthecombinedtheoryofconcavequadraticpoly- v nomialinequalitiesandtheequalitytheoryoveruninterpretedfunctionssymbols 2 (EUF).Theproposedapproachisapplicabletoallexistingabstractdomainslike 0 octagon,polyhedra,ellipsoidandsoon,thereforeitcanbeusedtoimprovethe 8 scalabilityofexistingverificationtechniquesforprogramsandhybridsystems.In 4 addition,wealsodiscusshowtoextendourapproachtoformulasbeyondconcave 0 quadraticpolynomialsusingGro¨bnerbasis. . 1 0 6 Keywords: Programverification,Interpolant,Concavequadraticpolynomials,Motzin’s 1 theorem,Semi-definiteprogramming. : v 1 Introduction i X InterpolantshavebeenpopularizedbyMcMillan[15]forautomaticallygenerating r invariantsofprograms.Sincethen,developingefficientalgorithmsforgeneratinginter- a polantsforvarioustheorieshasbecomeanactiveareaofresearch;inparticular,methods havebeendevelopedforgeneratinginterpolantsforPresburgerarithmetic(bothforin- tegers as well as for rationals/reals), theory of equality over uninterpreted symbols as wellastheircombination.Mostofthesemethodsassumetheavailabilityofarefutation proofofα∧βtogeneratea“reverse”interpolantof(α,β);calculihavebeenproposed to label an inference node in a refutational proof depending upon whether symbols of formulas on which the inference is applied are purely from α or β. For proposi- tional calculus, there already existed methods for generating interpolants from reso- lution proofs [11,16] prior to McMillan’s work, which generate different interpolants fromthosedonebyMcMillan’smethod.ThisledD’Silvaetal[6]tostudystrengthsof variousinterpolants. In Kapur, Majumdar and Zarba [10], an intimate connection between interpolants andquantifiereliminationwasestablished.Usingthisconnection,existenceofquantifier- free as well as interpolants with quantifiers were shown for a variety of theories over container data structures. A CEGAR based approach was generalized for verification of programs over container data structures using interpolants. Using this connection betweeninterpolantgenerationandquantifierelimination,Kapur[9]hasshownthatin- terpolantsformalatticeorderedusingimplication,withtheinterpolantgeneratedfrom αbeingthebottomofsuchalatticeandtheinterpolantgeneratedfromβ beingthetop ofthelattice. Nonlinearpolynomialsinequalitieshavebeenfoundusefultoexpressinvariantsfor softwareinvolvingsophisticatednumbertheoreticfunctionsaswellashybridsystems; an interested reader may see [27,28] where different controllers involving nonlinear polynomialinequalitiesarediscussedforsomeindustrialapplications. Weproposeanalgorithmtogenerateinterpolantsforquadraticpolynomialinequal- ities (including strict inequalities). Based on the insight that for analyzing the solu- tionspaceofconcavequadraticpolynomial(strict)inequalities,itsufficestolinearize them. We prove a generalization of Motzkin’s transposition theorem to be applicable for quadratic polynomial inequalities (including strict as well as nonstrict). Based on thisresult,weprovetheexistenceofinterpolantsfortwomutuallycontradictorycon- junctionsα,β ofconcavequadraticpolynomialinequalitiesandgiveanalgorithmfor computing an interpolant using semi-definite programming. The algorithm is recur- sivewiththebasisstepofthealgorithmrelyingonanadditionalconditiononconcave quadraticpolynomialsappearinginnonstrictinequalitiesthatanynonpositiveconstant combinationofthesepolynomialsisneveranonzerosumofsquarepolynomial(called NSOSC). In this case, an interpolant output by the algorithm is either a strict in- equalityoranonstrictinequalitymuchlikeinthelinearcase.Incase,thisconditionis notsatisfiedbythenonstrictinequalities,i.e.,thereisanonpositiveconstantcombina- tions of polynomials appearing as nonstrict inequalities that is a negative of a sum of squares,thennewmutuallycontradictoryconjunctionsofconcavequadraticpolynomi- als infewer variablesare derivedfrom the inputaugmented withthe equality relation deduced, and the algorithm is recursively invoked on the smaller problem. The out- putofthisalgorithmisingeneralaninterpolantthatisadisjunctionofconjunctionof polynomial nonstrict or strict inequalities. The NSOSC condition can be checked in polynomialtimeusingsemi-definiteprogramming. Wealsoshowhowseparatingtermst−,t+ canbeconstructedusingcommonsym- bols in α,β such that α ⇒ t− ≤ x ≤ t+ and β ⇒ t+ ≤ y ≤ t−, whenever (α∧β) ⇒ x = y. Similar to the construction for interpolants, this construction has thesamerecursivestructurewithconcavequadraticpolynomialssatisfyingNSOSCas thebasisstep.Thisresultenablestheuseoftheframeworkproposedin[17]basedon hierarchical theories and a combination method for generating interpolants by Yorsh andMusuvathi,fromcombiningequalityinterpolatingquantifier-freetheoriesforgen- erating interpolants for the combined theory of quadratic polynomial inequalities and theoryofuninterpretedsymbols. Obviously,ourresultsaresignificantinprogramverificationasallwell-knownab- stractdomains,e.g.octagon,polyhedra,ellipsoid andsoon,whicharewidelyusedin the verification of programs and hybrid systems, are quadratic and concave. In addi- tion,wealsodiscussthepossibilitytoextendourresultstogeneralpolynomialformu- lasbyallowingpolynomialequalitieswhosepolynomialsmaybeneitherconcavenor quadraticusingGro¨bnerbasis. Wedevelopacombinationalgorithmforgeneratinginterpolantsforthecombination ofconcavequadraticpolynomialinequalitiesanduninterpretedfunctionsymbols. In [5], Dai et al. gave an algorithm for generating interpolants for conjunctions of mutually contradictory nonlinear polynomial inequalities based on the existence of a witness guaranteed by Stengle’s Positivstellensatz [22] that can be computed using semi-definiteprogramming.Theiralgorithmisincompleteingeneralbutifeveryvari- ablesrangesoveraboundedinterval(calledArchimedeancondition),thentheiralgo- rithm is complete. A major limitation of their work is that formulas α,β cannot have uncommonvariables4.However,theydonotgiveanycombinationalgorithmforgener- atinginterpolantsinthepresenceofuninterpretedfunctionsymbolsappearinginα,β. Thepaperisorganizedasfollows.Afterdiscussingsomepreliminariesinthenext section, Section 3 defines concave quadratic polynomials, their matrix representation and their linearization. Section 4 presents the main contribution of the paper. A gen- eralizationofMotzkin’stranspositiontheoremforquadraticpolynomialinequalitiesis presented. Using this result, we prove the existence of interpolants for two mutually contradictoryconjunctionsα,β ofconcavequadraticpolynomialinequalitiesandgive analgorithm(Algorithm2)forcomputinganinterpolantusingsemi-definiteprogram- ming. Section 5 extends this algorithm to the combined theory of concave quadratic inequalitiesandEUFusingtheframeworkusedin[21,17].Implementationandexper- imentalresultsusingtheproposedalgorithmsarebrieflyreviewedinSection6,andwe concludeanddiscusfutureworkinSection7. 2 Preliminaries Let N, Q and R be the set of natural, rational and real numbers, respectively. Let R[x]bethepolynomialringoverRwithvariablesx=(x ,··· ,x ).Anatomicpoly- 1 n nomial formula ϕ is of the form p(x)(cid:5)0, where p(x) ∈ R[x], and (cid:5) can be any of =,>,≥,(cid:54)=;withoutanylossofgenerality,wecanassume(cid:5)tobeanyof>,≥.Anar- bitrarypolynomialformulaisconstructedfromatomiconeswithBooleanconnectives andquantificationsoverrealnumbers.LetPT(R)beafirst-ordertheoryofpolynomi- als with real coefficient, In this paper, we are focusing on quantifier-free fragment of PT(R). Laterwediscussquantifier-freetheoryofequalityoftermsoveruninterpretedfunc- tion symbols and its combination with the quantifier-free fragment of PT(R). Let Σ beasetof(new)functionsymbols.LetPT(R)Σ betheextensionofthequantifier-free theorywithuninterpretedfunctionsymbolsinΣ. Forconvenience,weuse⊥tostandforfalseand(cid:62)fortrueinwhatfollows. 4Seehoweveranexpandedversionoftheirpaperunderpreparationwheretheyproposeheuris- ticsusingprogramanalysisforeliminatinguncommonvariables. Definition1. AmodelM = (M,f )ofPT(R)Σ consistsofamodelM ofPT(R) M andafunctionf :Rn →Rforeachf ∈Σ witharityn. M Definition2. LetφandψbeformulasofaconsideredtheoryT,then – φisvalidw.r.t.T,writtenas|= φ,iffφistrueinallmodelsofT; T – φentailsψ w.r.t.T,writtenasφ |= ψ,iffforanymodelofT,ifψ istrueinthe T model,soisφ; – φ is satisfiable w.r.t. T, iff there exists a model of T such that in which φ is true; otherwiseunsatisfiable. Notethatφisunsatisfiableiffφ|= ⊥. T Craig showed that given two formulas φ and ψ in a first-order theory T such that φ |= ψ, there always exists an interpolant I over the common symbols of φ and ψ suchthatφ|=I,I |=ψ.Intheverificationliterature,thisterminologyhasbeenabused following[15],whereanreverseinterpolantI overthecommonsymbolsofφandψis definedforφ∧ψ |=⊥as:φ|=I andI∧ψ |=⊥. Definition3. Let φ and ψ be two formulas in a theory T such that φ∧ψ |= ⊥. A T formulaI saidtobea(reverse)interpolantofφandψifthefollowingconditionshold: i φ|= I; T ii I∧ψ |= ⊥;and T iii I onlycontainscommonsymbolsandfreevariablessharedbyφandψ. Ifψ isclosed,thenφ |= ψ iffφ∧¬ψ |= ⊥.Thus,I isaninterpolantofφand T T ψ iff I is a reverse interpolant of φ and ¬ψ. In this paper, we just deal with reveres interpolant,andfromnowon,weabuseinterpolantandreverseinterpolant. 2.1 Motzkin’stranspositiontheorem Motzkin’stranspositiontheorem[18]isoneofthefundamentalresultsaboutlinear inequalities; it also served as a basis of the interpolant generation algorithm for the quantifier-freetheoryoflinearinequalitiesin[17].Thetheoremhasseveralvariantsas well.Belowwegivetwoofthem. Theorem1 (Motzkin’s transposition theorem [18]). Let A and B be matrices and letαandβbecolumnvectors.ThenthereexistsavectorxwithAx≥αandBx>β, iff forallrowvectorsy,z≥0: (i)if yA+zB =0thenyα+zβ ≤0; (ii)if yA+zB =0andz(cid:54)=0thenyα+zβ <0. Corollary1. LetA ∈ Rr×n andB ∈ Rs×n bematricesandα ∈ Rr andβ ∈ Rs be columnvectors.DenotebyA ,i = 1,...,r theithrowofAandbyB ,j = 1,...,s i j thejthrowofB.ThentheredoesnotexistavectorxwithAx ≥ αandBx > β,iff thereexistrealnumbersλ ,...,λ ≥0andη ,η ,...,η ≥0suchthat 1 r 0 1 s r s (cid:88) (cid:88) λ (A x−α )+ η (B x−β )+η ≡0, (1) i i i j j j 0 i=1 j=1 s (cid:88) η >0. (2) j j=0 Proof. The“if”partisobvious.Belowweprovethe“onlyif”part. ByTheorem1,ifAx≥αandBx>βhavenocommonsolution,thenthereexist tworowvectorsy∈Rr andz∈Rswithy≥0andz≥0suchthat (yA+zB =0∧yα+zβ >0)∨(yA+zB =0∧z(cid:54)=0∧yα+zβ ≥0). Letλ =y ,i=1,...,r,η =z ,j =1,...,sandη =yα+zβ.Thenitiseasyto i i j j 0 checkthatEqs.(1)and(2)hold. (cid:116)(cid:117) 3 Concavequadraticpolynomialsandtheirlinearization Definition4 (ConcaveQuadratic).Apolynomialf ∈R[x]iscalledconcavequadratic (CQ),ifthefollowingtwoconditionshold: (i) f hastotaldegreeatmost2,i.e.,ithastheformf =xTAx+2αTx+a,whereA isarealsymmetricmatrix,αisacolumnvectoranda∈Risaconstant; (ii) thematrixAisnegativesemi-definite,writtenasA(cid:22)0.5 Example1. Letg =−x2+2x −x2+2x −y2,thenitcanbeexpressedas 1 1 1 2 2 T T x −1 0 0 x 1 x 1 1 1 g1 =x2 0 −1 0 x2+21 x2. y 0 0 −1 y 0 y −1 0 0 Thedegreeofg1is2,andthecorrespondingA= 0 −1 0 (cid:22)0.Thus,g1isCQ. 0 0 −1 Itiseasytoseethatiff ∈ R[x]islinear,thenf isCQbecauseitstotaldegreeis1 andthecorrespondingAis0whichisofcoursenegativesemi-definite. Aquadraticpolynomialcanalsoberepresentedasaninnerproductofmatrices(cf. (cid:28) (cid:18)1 xT (cid:19)(cid:29) [13]),i.e.,f(x)= P, . xxxT 5A being negative semi-definite has many equivalent characterizations: for every vector x, xTAx ≤ 0;everykthminorofA≤ 0ifk isoddand≥ 0otherwise;aHermitianmatrix whoseeigenvaluesarenonpositive. 3.1 Linearization Considerquadraticpolynomialsf andg (i=1,...,r,j =1,...,s), i j f =xTA x+2αTx+a , i i i i g =xTB x+2βTx+b , j j j j whereA ,B aresymmetricn×nmatrices,α ,β ∈ Rn,anda ,b ∈ R;letP := i j i j i j i (cid:18)a αT(cid:19) (cid:18)b βT(cid:19) i i , Q := j j be(n+1)×(n+1)matrices,then α A j β B i i j j (cid:28) (cid:18)1 xT (cid:19)(cid:29) (cid:28) (cid:18)1 xT (cid:19)(cid:29) f (x)= P , , g (x)= Q , . i i xxxT j j xxxT ForCQpolynomialsf sandg sinwhicheachA (cid:22)0,B (cid:22)0,define i j i j K ={x∈Rn |f (x)≥0,...,f (x)≥0,g (x)>0,...,g (x)>0}. (3) 1 r 1 s (cid:28) (cid:18)1 xT (cid:19)(cid:29) Givenaquadraticpolynomialf(x)= P, ,itslinearizationisdefined xxxT (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:18)1 xT(cid:19) asf(x)= P, ,where (cid:23)0. x X x X Let X =(X ,X ,X ,...,X ,...,X ,...,X ,...,X ) (1,1) (2,1) (2,2) (k,1) (k,k) (n,1) (n,n) be the vector variable with n(n+1) dimensions corresponding to the matrix X. Since 2 (cid:28) (cid:18)1 xT(cid:19)(cid:29) X isasymmetricmatrix, P, isalinearexpressioninx,X. x X Now,let (cid:18)1 xT(cid:19) (cid:28) (cid:18)1 xT(cid:19)(cid:29) K ={x| (cid:23)0, ∧r P , ≥0, 1 x X i=1 i x X (cid:28) (cid:18)1 xT(cid:19)(cid:29) ∧s Q , >0, forsomeX}, (4) j=1 j x X whichisthesetofallx∈Rnonlinearizationsoftheabovef sandg s. i j In[7,13],whenK andK aredefinedonlywithf withoutg ,i.e.,onlywithnon- 1 i j strictinequalities,itisprovedthatK =K . BythefollowingTheorem2,weshowthat 1 K =K alsoholdseveninthepresenceofstrictinequalitieswhenf andg areCQ.So, 1 i j whenf andg areCQ,theCQpolynomialinequalitiescanbetransformedequivalently i j toasetoflinearinequalityconstraintsandapositivesemi-definiteconstraint. Theorem2. Let f ,...,f and g ,...,g be CQ polynomials, K and K as above, 1 r 1 s 1 thenK =K . 1 Proof. Foranyx ∈ K,letX = xxT.Thenitiseasytoseethatx,X satisfy(4).So x∈K ,thatisK ⊆K . 1 1 Next,weproveK ⊆ K.Letx ∈ K ,thenthereexistsasymmetricn×nmatrix 1 1 (cid:18)1 xT(cid:19) X satisfying(4).Because (cid:23)0,wehaveX−xxT (cid:23)0.Thenbythelasttwo x X conditionsin(4),wehave (cid:28) (cid:18)1 xT (cid:19)(cid:29) (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:28) (cid:18)0 0 (cid:19)(cid:29) f (x)= P , = P , + P , i i xxxT i x X i 0xxT −X =(cid:28)P ,(cid:18)1 xT(cid:19)(cid:29)+(cid:10)A ,xxT −X(cid:11)≥(cid:10)A ,xxT −X(cid:11), i x X i i (cid:28) (cid:18)1 xT (cid:19)(cid:29) (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:28) (cid:18)0 0 (cid:19)(cid:29) g (x)= Q , = Q , + Q , j j xxxT j x X j 0xxT −X =(cid:28)Q ,(cid:18)1 xT(cid:19)(cid:29)+(cid:10)B ,xxT −X(cid:11)>(cid:10)B ,xxT −X(cid:11). j x X j j Since f and g are all CQ, A (cid:22) 0 and B (cid:22) 0. Moreover, X − xxT (cid:23) 0, i.e., i j i j xxT −X (cid:22)0.Thus,(cid:10)A ,xxT −X(cid:11)≥0and(cid:10)B ,xxT −X(cid:11)≥0.Hence,wehave i j f (x)≥0andg (x)>0,sox∈K,thatisK ⊆K. (cid:116)(cid:117) i j 1 3.2 Motzkin’stheoreminMatrixForm (cid:28) (cid:18)1 xT(cid:19)(cid:29) If P, is seen as a linear expression in x,X, then Corollary 1 can be x X reformulatedas: Corollary2. Let x be a column vector variable of dimension n and X be a n×n symmetricmatrixvariable.SupposeP ,P ,...,P andQ ,...,Q are(n+1)×(n+1) 0 1 r 1 s symmetricmatrices.Let (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:28) (cid:18)1 xT(cid:19)(cid:29) W=ˆ{(x,X)|∧r P , ≥0,∧s Q , >0}, i=1 i x X i=1 j x X thenW =∅iffthereexistλ ,λ ,...,λ ≥0andη ,η ,...,η ≥0suchthat 0 1 r 0 1 s (cid:88)r (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:88)s (cid:28) (cid:18)1 xT(cid:19)(cid:29) λ P , + η Q , +η ≡0, and i i x X j j x X 0 i=0 j=1 η +η +...+η >0. 0 1 s 4 AlgorithmforgeneratinginterpolantsforConcaveQuadraticPoly- nomialinequalities Problem1. Giventwoformulasφandψonnvariableswithφ∧ψ |=⊥,where φ=f ≥0∧...∧f ≥0∧g >0∧...∧g >0, 1 r1 1 s1 ψ =f ≥0∧...∧f ≥0∧g >0∧...∧g >0, r1+1 r s1+1 s inwhichf ,...,f ,g ,...,g areallCQ,developanalgorithmtogeneratea(reverse) 1 r 1 s CraiginterpolantI forφandψ,onthecommonvariablesofφandψ,suchthatφ|=I andI∧ψ |=⊥.Forconvenience,wepartitionthevariablesappearinginthepolynomials above into three disjoint subsets x = (x ,...,x ) to stand for the common variables 1 d appearinginbothφandψ,y=(y ,...,y )tostandforthevariablesappearingonlyin 1 u φandz=(z ,...,z )tostandforthevariablesappearingonlyinψ,whered+u+v = 1 v n. Sincelinearinequalitiesaretriviallyconcavequadraticpolynomials,ouralgorithm (Algorithm IGFQC in Section 4.4) can deal with the linear case too. In fact, it is a generalizationofthealgorithmforlinearinequalities. Theproposedalgorithmisrecursive:thebasecaseiswhennosumofsquares(SOS) polynomial can be generated by a nonpositive constant combination of nonstrict in- equalitiesinφ∧ψ.Whenthisconditionisnotsatisfied,i.e.,anSOSpolynomialcanbe generatedbyanonpositiveconstantcombinationofnonstrictinequalitiesinφ∧ψ,then itispossibletoidentifyvariableswhichcanbeeliminatedbyreplacingthembylinear expressionsintermsofothervariablesandthusgenerateequisatisfiableproblemwith fewervariablesonwhichthealgorithmcanberecursivelyinvoked. (cid:28) (cid:18)1 xT(cid:19)(cid:29) Lemma1. LetU ∈R(n+1)×(n+1) beamatrix.If U, ≤0foranyx∈Rn x X (cid:18)1 xT(cid:19) andsymmetricmatrixX ∈Rn×nwith (cid:23)0,thenU (cid:22)0. x X Proof. AssumethatU (cid:54)(cid:22)0.Thenthereexistsacolumnvectory=(y ,y ,...,y )T ∈ 0 1 n Rn+1suchthatc:=yTUy=(cid:10)U,yyT(cid:11)>0.DenoteM =yyT,thenM (cid:23)0. (cid:18)1 xT(cid:19) (cid:18)1 xT (cid:19) Ify (cid:54)=0,thenletx=(y1,...,yn)T,andX =xxT.Thus, = = 0 y0 y0 x X xxxT (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:68) (cid:69) (cid:28) (cid:18)1 xT(cid:19)(cid:29) 1 M (cid:23),and U, = U, 1 M = c >0,whichcontradictswith U, ≤ y2 x X y2 y2 x X 0 0 0 0. If y = 0, then M = 0. Let M(cid:48) = |U(1,1)|+1M, then M(cid:48) (cid:23) 0. Further, let 0 (1,1) c 10···0 00···0 (cid:18)1 xT(cid:19) M(cid:48)(cid:48) =M(cid:48)+... ... ... ....ThenM(cid:48)(cid:48) (cid:23)0andM(cid:48)(cid:48)(1,1) =1.Let x X =M(cid:48)(cid:48),then 00···0 10···0 (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:42) 00···0(cid:43) U, x X =(cid:104)U,M(cid:48)(cid:48)(cid:105)= U,M(cid:48)+... ... ... ... 00···0 10···0 (cid:42) (cid:43) |U |+1 00···0 = U, (1,1c) M +... ... ... ... 00···0 |U |+1 (1,1) = (cid:104)U,M(cid:105)+U c (1,1) =|U |+1+U >0, (1,1) (1,1) (cid:28) (cid:18)1 xT(cid:19)(cid:29) whichalsocontradictswith U, ≤ 0.Thus,theassumptiondoesnothold, x X thatisU (cid:22)0. (cid:116)(cid:117) Lemma2. LetA={y∈Rm |A y−α ≥0,B y−β >0, fori=1,...,r,j = i i j j 1,...,}beanonemptysetandB ⊆Rmbeannonemptyconvexclosedset.IfA∩B =∅ andtheredoesnotexistalinearformL(y)suchthat ∀y∈A,L(y)>0, and ∀y∈B,L(y)≤0, (5) thenthereisalinearformL (y)(cid:54)≡0andδ ,...,δ ≥0suchthat 0 1 r r (cid:88) L (y)= δ (A y−α )and ∀y∈B,L (y)≤0. (6) 0 i i i 0 i=1 Proof. Since A is defined by a set of linear inequalities, A is a convex set. Using the separation theorem on disjoint convex sets, cf. e.g. [1], there exists a linear form L (y)(cid:54)≡0suchthat 0 ∀y∈A,L (y)≥0, and ∀y∈B,L (y)≤0. (7) 0 0 From(5)wehavethat ∃y ∈A, L (y )=0. (8) 0 0 0 Since ∀y∈A,L (y)≥0, (9) 0 then A y−α ≥0∧...∧A y−α ≥0∧ 1 1 r r B y−β >0∧...∧B y−β >0∧−L (y)>0 1 1 s s 0 hasnosolutionw.r.t.y.UsingCorollary1,thereexistλ ,...,λ ≥ 0,η ,...,η ≥ 0 1 r 0 s andη ≥0suchthat r s (cid:88) (cid:88) λ (A y−α )+ η (B y−β )+η(−L (y))+η ≡0, (10) i i i j j j 0 0 i=1 j=1 s (cid:88) η +η >0. (11) j j=0 Applyingy in(8)to(10)and(11),itfollows 0 η =η =...=η =0, η >0. 0 1 s Fori=1,...,r,letδi = ληi ≥0,then r (cid:88) L (y)= δ (A y−α )and ∀y∈B,L (y)≤0. (cid:116)(cid:117) 0 i i i 0 i=1 The lemma below asserts the existence of a strict linear inequality separating A and B defined above, for the case when any nonnegative constant combination of the linearizationoff sispositive. i Lemma3. LetA={y∈Rm |A y−α ≥0,B y−β >0, fori=1,...,r,j = i i j j 1,...,}beanonemptysetandB ⊆Rmbeannonemptyconvexclosedset,A∩B =∅. ThereexistsalinearformL(x,X)suchthat ∀(x,X)∈A,L(x,X)>0, and ∀(x,X)∈B,L(x,X)≤0, whenevertheredoesnotexistλ ≥0,s.t.,(cid:80)r λ P (cid:22)0. i i=1 i i Proof. Proof is by contradiction. Given that A is defined by a set of linear inequal- ities and B is a closed convex nonempty set, by Lemma 2, there exist a linear form L (x,X)(cid:54)≡0andδ ,...,δ ≥0suchthat 0 1 r (cid:88)r (cid:28) (cid:18)1 xT(cid:19)(cid:29) L (x,X)= δ P , and ∀(x,X)∈B,L (x,X)≤0. 0 i i x X 0 i=1 I.e.thereexistsansymmetricalmatrixL(cid:54)≡0suchthat (cid:28) (cid:18)1 xT(cid:19)(cid:29) (cid:88)r (cid:28) (cid:18)1 xT(cid:19)(cid:29) L, ≡ δ P , , (12) x X i i x X i=1 (cid:28) (cid:18)1 xT(cid:19)(cid:29) ∀(x,X)∈B, L, ≤0. (13) x X