ON FREE STOCHASTIC DIFFERENTIAL EQUATIONS V.KARGIN 1 1 Abstract 0 2 The paper derives an equation for the Cauchy transform of the solution of a n freestochasticdifferentialequation(SDE).Thisnewequationisusedtosolve a J severalparticular examplesoffreeSDEs. 4 1 ] R 1. INTRODUCTION P h. Free stochastic differential equations generalize classical stochastic differential equations at tothesetting offreeprobability. Hereisanexampleofsuchanequation: m [ Xt = a(Xt)dt+b∗(Xt)(dZt)b(Xt) 1 Inthisequation, X isaself-adjoint operator, a(X )andb(X )areoperator-valued functions v t t t 7 ofX ,andthedrivingnoiseZ isanoperatorprocesswithfreeincrements. Thatis,theincre- t t 9 ments Z Z ,s > t,areassumed tobe freefrom past realizations ofZ .Theprocess Z is 6 s t t t − 2 usuallythefreeBrownianprocess,inwhichcasetheincrementshavesemicircledistributions; . 1 however,otherchoicesarepossible. 0 Informally, thereadermaythinkabout X asverylargerandom matricesandZ asmatri- 1 t t 1 ceswithindependent Gaussianrandomvariablesasentries. Theseentriesfollowindependent : v Brownian motions andweareinterested inthelawofthe eigenvalues ofX .Thefreeproba- t i X bilitytheoryisaconvenient abstraction whichintendstomodelthesituation whenthesizeof r thematricesisverylarge. a Thestudy offreestochastic differential equations (“freestochastic calculus”) ismoredif- ficult than in the classical case because of non-commutativity of coefficients and noise. This papercontributes bydeveloping anewtoolfortheanalysis oftheseequations. The idea of free stochastic calculus was first suggested in [S]. It was later developed and formalized in[KS],[B],[BSa],and[A],whichintroduced stochastic integration withrespect tofreeBrownianmotionasarigorous basisforfreestochastic calculus. Theyalsoderivedan analogoftheItoˆ formula,whichallowsustoobtainidentities likethefollowing: a a [W2(dW )+W (dW )W +(dW )W2] = W3 2 W dt, t t t t t t t a − t Z0 Z0 Date:November2010. DepartmentofMathematics,StanfordUniversity,PaloAlto,CA94305,USA.e-mail:[email protected]. 1 2 V.KARGIN whereW denotes thefreeBrownianmotion(theWignerprocess). Ananalogous formulain t theclassical situation is a a 3B2(dB ) = B3 3 B dt, t t a − t Z0 Z0 where B is the standard Brownian motion. Note the different coefficient before the integral t ontheright-hand side. The classical Itoˆ formula is very helpful in the study of stochastic differential equations. Unfortunately, the range of applicability of the free Itoˆ formula is smaller. This difficulty callsforadifferent methodapplicable tothosefreeSDEs,whicharenotsolvable withtheItoˆ formula. Onepossibility istoseekanequationfortheevolution ofthespectraldistribution of thesolution. Suchanequation wasderivedbyBianeandSpeicherin[BSb]fortheequation X = a(X )dt+dW . (1) t t t They showed that the density of the spectral probability measure of X , which we denote p t t andwhichgeneralizes theeigenvalue distribution ofamatrix,satisfiesthefreeFokker-Planck equation: ∂p ∂ t = [p (Hp +a)]. (2) t t ∂t −∂x HereH denotes amultipleoftheHilberttransform: u(y) Hu(x) := p.v. dy. (3) x y Z − Still, the approach through the free Fokker-Planck equation has its own disadvantages. First,itisapplicableonlytoequationsthathavethespecialform(1),thatis,onlytoequations with the constant diffusion coefficient. Second, the free Fokker-Planck equation (2) is not a bonafidepartialdifferential equationsinceitincludestheHilberttransformoperator. Forthis reason, itissomewhatdifficulttosolvethisequation. Thepurpose ofthispaperistoapproach thefreestochastic equations byderiving adiffer- entialequation fortheCauchytransform ofthesolution. Recall that the resolvent of operator X is defined as the operator-valued function of a t complex parameter Gt(z) := (Xt z)−1. The Cauchy transform of Xt is defined as the − expectation of the resolvent: gt(z) := E[(Xt z)−1]. It is useful because the knowledge − of the Cauchy transform is sufficient to recover all properties of the spectral probability distribution ofX . ItturnsoutthatifX solves t t dX = a(X )dt+b(X )(dW )c(X ), t t t t t theng (z)satisfiesthefollowingequation: t dg t = E(a G2)+E(b c G )E(b c G2), (4) dt − t t t t t t t t where we use a , b , and c to denote a(X ), b(X ), and c(X ), respectively. This is the t t t t t t statementofTheorem3.2below. Equation (4)isnotausual differential equation since itinvolves expectations. Ingeneral, these expectations are difficult to compute because the coefficients a , b , and c , and the t t t resolvent G are not free from each other. However, if the coefficients are polynomials, it is t ONFREESTOCHASTICDIFFERENTIALEQUATIONS 3 possibletoperformfurtherreductiontoadifferentialequationaswewillshowinProposition 3.4. Aswejustsaid,theknowledgeoftheCauchytransformcanbeusedtorecoverthespectral probabilitydistribution. Inparticular,thefreeFokker-Planckequation(2)canbederivedfrom (4)aswillbeshowninCorollary3.6. IncertaincasesitisnotpossibletocomputetheCauchytransformexplicitly,butitispossi- bletodetectthebehaviorofitssingularities. Thisknowledgecanprovideuswithinformation aboutthesupportofthespectraldistribution. Inparticular,itcanshowushowthenormofthe solution grows. For a simple example of this approach, let us consider the well-known case of the free Ornstein-Uhlenbeck equation: dX = θX dt+σdW . (5) t t t − For this equation, it is easy to compute the Cauchy transform using equation (4) and re- cover the known result that for positive θ, the spectral probability distribution of X con- t verges to a stationary solution, which is a semicircle distribution supported on the interval [ σ 2/θ ,σ 2/θ ]. − | | | | Asamoredifficultexample,consider theequation p p 1/2 1/2 dX = θX dt+X (dW )X , t t t t t whichcanbethoughtofasafreeanalogoftheequationforthe“geometricBrownianmotion”, dx = θx dt+x dB . t t t t LetX = I.Equation(4)leadstothefollowingdifferential equation: 0 ∂g ∂g +z(θ 1 zg) = g(θ 1 zg), ∂t − − ∂z − − − with the initial condition g(0,z) = (1 z)−1. The method of characteristics gives us a − functional equation fortheCauchytransform: z+g 1 = e(θ 1 zg)t. (6) − − − θ=1/2 1.4 1.2 t=0.1 t=0.2 1 t=0.3 0.8 t=0.4 0.6 0.4 0.2 00 0.5 1 1.5 2 2.5 3 3.5 4 FIGURE 1. θ = 1/2 Whileitisdifficulttoextractanexplicitanalyticalformulaforthesolutionofthisequation, wecaninvestigate howthesupport ofthedistribution changes withtime. Itturns outthatfor 4 V.KARGIN θ=2 1.4 1.2 t=0.1 1 t=0.2 t=0.3 0.8 t=0.4 0.6 0.4 0.2 00 1 2 3 4 5 6 7 FIGURE 2. θ = 2 θ=−1 1.6 t=0.1 1.4 t=0.2 1.2 t=0.3 1 t=0.4 0.8 0.6 0.4 0.2 00 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 FIGURE 3. θ = 1 − θ=−1, larger times 4.5 4 t=1 3.5 3 2.5 t=0.75 2 t=0.5 1.5 t=0.25 1 0.5 00 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 FIGURE 4. θ = 1,Largetimes − ONFREESTOCHASTICDIFFERENTIALEQUATIONS 5 θ < 0,the support of thedistribution shrinks tozero. If θ is between 0and 1, then the lower boundaryofthesupportdecreasestozeroandtheupperboundarygrowsexponentially fastto infinity. Ifθ > 1, then both thelowerand upper boundary ofthe support grow exponentially fasttoinfinity. Wecan solve equation (6)numerically and recover thedensity ofthe spectral distribution by using Stieltjes formula. Figures 1-4 show the evolution of the density for various values of parameters and illustrate the complexity of the behavior of the spectral distribution. For example, Figure 2 shows that even if θ > 1, the spectral distribution does not approach infinityimmediately. Thereisatransition periodinwhichasignificant portionofthespectral distribution remains below λ = 1. Similarly, Figure 3 shows that for θ < 0, the distribution does not collapse to zero immediately. Onlywhen time increases, the distribution begins the rapidapproach tozero,asshowninFigure4. Letuscomparethisresultwiththeclassical analog. ByusingtheItoˆ formula, itiseasyto showthatthesolution oftheclassical equation forthegeometric Brownianmotionis 1 x = exp (θ )t+B . t t { − 2 } Hence,withprobability1,theclassicalsolutionwilldecreaseexponentiallytozeroifθ < 1/2, and will grow exponentially to infinity if θ > 1/2. However, the support of the solution distribution is(0, )forallt. Thisisquiteunlikethebehavior ofthefreeSDEsolution. ∞ Note that the equation for the geometric Brownian motion can be generalized to the free probability settinginadifferentway: dX = θX dt+X dW +(dW )X . t t t t t t Thebehavior ofthe solution ofthis equations is quite different. In particular, the ratio ofthe standarddeviationtotheexpectationis 2(e2t 1).Thisratiogrowsexponentiallyfastwith − t, quite unlike the previous example, where this ratio equals √t. Unfortunately, the partial p differential equation associated with equation is more difficult to solve and it is not clear whetherthesolution becomesunbounded infinitetime. Finally,letusconsider thefollowingequation: dX = kX (dW )X , t t t t and let the initial condition be X = aI. For this equation it is possible to write an explicit 0 formula for thespectral distribution ofthesolution. Aninteresting feature ofthis equation is that the solution blows up in finite time τ = (ak) 2, by which we means that the operator − normofthesolution becomesinfiniteastapproaches τ. Another interesting feature is that as time t approaches τ, the spectral distribution con- vergestoafixeddistribution. Ifk = a 1,thenthedensityofthisdistribution is − √4ξ 1 f(ξ) = − , 2πξ3 6 V.KARGIN 2 1.8 1.6 t=0.4 1.4 1.2 t=0.3 1 0.8 t=0.2 0.6 t=0.1 0.4 0.2 00 0.5 1 1.5 2 2.5 3 3.5 4 FIGURE 5. Densityfunctions fort = 0.1,0.2,0.3,0.4. 2.5 2 1.5 1 0.5 0 0 0.5 1 1.5 2 2.5 3 3.5 4 FIGURE 6. Densityfunctions fort =0.5,0.6,0.7,0.8 which is supported on the interval [1/4, ). Otherwise, it is a scaled version of this distri- ∞ bution. The behavior of the solution density for various times is illustrated in Figures 5 and 6. SeveralspecificclassesoffreeSDEhavealreadybeeninvestigated intheliterature. Biane andSpeicherin[BSb]andGaoin[G]studiedthefreeOrnstein-Uhlenbeckequation(5). Biane and Speicher proved that its solution converges to a stationary process with a semicircle dis- tribution. Gaoconsidered free Ornstein-Uhlenbeck processes withafree Levydriving noise, andshowedthateveryself-decomposable probability measureonthereallinecanberealized asadistribution ofsuchaprocess. Capitaine and Donati-Matin in[CD-M]defined the freeWishart process and found that it satisfiesthefreeSDEoftheform: dX = λdt+ X dZ +dZ X , t t t t∗ t p p ONFREESTOCHASTICDIFFERENTIALEQUATIONS 7 where Z is the complex Wigner process. Demni in [D] studied the so-called free Jacobi t processes whichsatisfyequations similartothefollowing: dX = (θI X)dt+ I X dZ X + X dZ I X . t − − t t t t t∗ − t p p p p With exception of the free Ornstein-Uhlenbeck process, we study a different set of free SDEs,andweapproachtheseequationswithadifferentpointofviewbasedonthedifferential equations fortheCauchytransform. ForthefreeOrnstein-Uhlebeck process, ourresultsagreewithresultsin[BSb]. The rest of the paper is organized as follows. Section 2 provides preliminary information about free stochastic integration and Itoˆ formulas. Section 3describes main results. . In par- ticular, Section3.1isdevoted toalocalexistence anduniqueness result. Section3.2presents generalresultsabouttheCauchytransformofthesolutionandSection3.3providesexamples. 2. FREE STOCHASTIC INTEGRATION 2.1. ThefreeBrownianmotion. Forthebasicsoffreeprobabilitytheorywereferto[VDN] and[NS]. Alloperators thatweconsider belong toanon-commutative W -probability space ∗ ( ,E), that is, to a von Neumann operator algebra with a faithful normal trace E. We A A denotetheusualoperatornormby X ,andtheL2-normby X := E(X X). 2 ∗ k k k k The spectral probability distribution of a self-adjoint operator X is a probability p∈ A measureµonRsuchthat E Xk = xkµ(dx). R (cid:16) (cid:17) Z ItsCauchytransform isthefunction µ(dx) g (z) = . X R x z Z − ItcanbedefineddirectlyintermsofoperatorX astheexpectationoftheresolvent: g (z) = X E[GX (z)],whereGX (z) := (X z)−1.Theprobability measureµcanberecovered from − itsCauchytransform bytheStieltjesinversion formula: 1 µ(B)= lim Img(x+iε)dx, (7) π ε↓0 ZB provided thatB isBorelandµ(∂B) = 0. This fact is the starting point of our approach, since we will study the evolution of the Cauchytransform asatooltoinvestigate theevolution ofthecorresponding probability mea- sure. The most important concept in free probability theory is that of free independence. Let A denote an arbitrary element of algebra . The sub-algebras ,..., of algebra i i 1 n A A A A (and operators that generate them) are said to be freely independent or free, if the following condition holds: E(A ...A ) = 0, i(1) i(m) 8 V.KARGIN provided thatE(A ) = 0and i(s+1) = i(s)foreverys. Twoparticular consequences of i(s) 6 thisdefinitionisthat(i)E(AB) = E(A)E(B)ifAandB arefree,and(ii) E(AX AX ) = E(A2)E(X )E(X ), (8) 1 2 1 2 ifAisfreefromX andX andE(A) =0. 1 2 ThefreeBrownianmotion,ortheWignerprocess,isafamilyofoperatorsW ,wheret 0, t ≥ that satisfies the following properties: (1) W = 0; (2) the increments of W are free in the 0 t sense of Voiculescu, i.e., if t > s, then W W is free from the subalgebra which is t s s − W generated by all W with τ s , and (3) the spectral distribution of W W is semicircle τ t s ≤ − withzeroexpectation andvariancet s. − The choice of W is not unique, and in the rest of the paper we assume that a particular t realization ofW isfixed. t 2.2. Free stochastic integral. Itoˆ-style free stochastic integration with respect to the free Brownian motion was defined and studied in [KS] and [BSa]. Their results show that under certainassumptions ontheoperator coefficientsa andb ,itispossibletodefinetheintegral t t 1 I = a (dW )b , t t t Z0 whereW isthefreeBrownianmotion. t Let us briefly recall the construction of the integral. For details of the construction, the readerisadvisedtoseeDefinition2.2.1andSection3in[BSa]andSection3andTheorem14 in[A]. Suppose thata andb arefunctions ofW ,τ t.Thatis,leta andb belong tothe t t τ t t ≤ sub-algebra . Assume also that max a , b C for all t [0,1] and that t a t t t t W {k k k k} ≤ ∈ → and t b are continuous mappings in the operator norm. Let t ,...,t and τ ,...,τ be t 0 n 1 n → realnumberssuchthat 0 = t t ... t = 1, 0 1 n ≤ ≤ ≤ and 0 τ t . k k 1 ≤ ≤ − Wedenotethesetoft ,...,t andτ ,...,τ as∆. Let 0 n 1 n d(∆) = max (t τ ). k k 1 k n − ≤ ≤ Considerthesum n I(∆) = a (W W )b τi ti − ti−1 τi i=1 X It turns out that as d(∆) 0, the sums I(∆) converge in operator norm and the limit → does not depend on the choice of t and τ . The limit is called the free stochastic integral i i 1 and denoted as a (dW )b . An important point in the proof of convergence is that the 0 t t t convergence of sums in the operator norm depends on a free analogue of the Burkholder- R Gundymartingaleinequalities. ONFREESTOCHASTICDIFFERENTIALEQUATIONS 9 Averyusefultoolinthestudyofstochastic integralsistheItoˆ formula. Afreeprobability analogueoftheItoˆ formulawasdevelopedin[BSa]. Intermsofformalrules,itcanbewritten asfollows: a dt b dt = a dt b dW c = a dW b c dt = 0, t t t t t t t t t t · · · a dW b c dW d = E(b c )a d dt. (9) t t t t t t t t t t · Notethattheruleinthesecondlineissignificantly different fromtheclassical case. Interms offreestochastic integrals, thesecondrulecanbewrittenasfollows: 1 1 1 t a dW b c dW d = ( a dW b )c dW d t t t t t t τ τ τ t t t · Z0 Z0 Z0 Z0 1 t + a dW b ( c dW d ) t t t τ τ τ Z0 Z0 1 + E(b c )a d dt. t t t t Z0 (Compare Theorem 4.1.2. in [BSa] or Proposition 8 and Corollary 10 in [A].) Here is an illustration (aparticular caseofProposition 4.3.2in[BSa]): LetW bethefreeBrownianmotionanddefine t ∂(Wn) := Wn 1dW +Wn 2(dW )W +...+(dW )Wn 1. t t − t t − t t t t − Then, α (W )n = d(Wn) a t Z0 α α = ∂(Wn)+ Wn k l 2(dW )Wk(dW )Wl, t t − −− t t t t Z0 Z0 0 k+l n 2 ≤ X≤ − anditfollowsthat a ⌊n/2⌋−1 a ∂(Wn) = (W )n (n 2k 1)C Wn 2k 2tkdt, (10) t a − − − k t − − Z0 k=0 Z0 X whereC aretheCatalannumbers, k 1 2k C := = E[(W )2k]. k 1 k+1 k (cid:18) (cid:19) AnanalogousformulafortheclassicalItoˆ integralwithrespecttotheBrownianmotionB t isquitedifferent: a n(n 1) a nBtn−1(dBt) = Ban− 2− Btn−2dt. Z0 Z0 Below,wewillusethefreeItoˆ formulainordertocomputethemomentsofvariableX . t 10 V.KARGIN 3. FREE STOCHASTIC DIFFERENTIAL EQUATIONS 3.1. Existenceanduniqueness. Afreestochastic differential equation (freeSDE) dX = a(X )dt+b(X )(dW )c(X ) (11) t t t t t isaconvenient shortcut notation forthefollowingintegralequation: t t X = X + a(X )dτ + b(X )dW c(X ). (12) t 0 τ τ τ τ Z0 Z0 Weconsideronlyequationswiththecoefficientsthatdonotdependexplicitlyontime,andwe willalwaysassumethata(X ),b(X ),andc(X )arelocallyoperator Lipschitzfunctions. (A t t t functionf : R Ciscalledlocallyoperator-Lipschitz, ifitisalocallybounded,measurable → function, andifforallA > 0,thereisaconstant K > 0,suchthat A f(X) f(Y) K X Y A k − k ≤ k − k forallself-adjointoperatorsX andY withthenormlessthanA.Forexample,allpolynomials arelocallyoperator-Lipschitz.) Equations(11)and(12)areparticular casesofthefollowingmoregeneralequations: m dX = a(X )dt+ b (X )(dW )c (X ) (13) t t i t t i t i=1 X and t m t X = X + a(X )dτ + b (X )dW c (X ). (14) t 0 τ i τ τ i τ Z0 i=1Z0 X OurresultsfortheCauchytransformofX canbeextendedtothismoregeneralsettingat t theexpenseofmorecumbersomenotation. Theexistenceofthesolutionofequation(14)mayfailforlargetifthenormofthesolution approachesinfinityinfinitetime. However,forsufficientlysmallt > 0wehavethefollowing local existence result. (See Theorem 3.1 in [BSb] for a sufficient condition of the global existence in a simpler class of free SDEs, and Theorem 5.2.1 in [O] for an existence and uniqueness resultinthecaseofclassical SDEs). Theorem 3.1. Suppose that a , b , and c are locally operator Lipshitz functions and X is i i i bounded inoperator norm. Then, there exist t > 0andafamilyofoperators X defined for 0 t allt [0,t )andbounded inoperatornorm,suchthatX = X,andX isauniquesolution 0 0 t ∈ of(14)fort < t . 0 Proof: TheproofproceedsbyPicard’smethodofsuccessiveapproximations. Wewillgive (0) theproveforthecasem = 1.Thegeneralcaseissimilar. DefineX =X,and t t t X(N+1) = X + a(X(N))dτ + b(X(N))dW c(X(N)). (15) t τ τ τ τ Z0 Z0 Weaimtoshowthatthisprocessconvergesforallsufficientlysmallt. Forthis,itisenoughto showthatforasufficiently smallt > 0 andallt < t andN 1,thefollowing twoclaims 0 0 ≥