Bellman Systems with M ean Field Dep endent Dynamics∗
1 Introduction
In recent literature,mainly scalar Bellman equations which are coupled with a Fokker-Planck equation are studied and we refer to starting paper[9]or to a survey[6],see also[7].These equations model a Nash game with a large number of players behaving similarly,so that the decision can be approximate by a single decision make(a representative agent).So we have a scalar Bellman equation.The present paper considers a model suggested by Bensoussan,accompanied by co-authors(see[3]),where the decision of finite number of N players of large population of Nash-game-playersare approximated by N representative agents.So the Bellman system is a system of backward parabolic equations coupled with a forward mean field equation.
其次,从性质上来看,学校是公益性质的,学校只能成立非营利的体育活动组织,并通过组织向学生提供课外体育服务,这类组织的法人一般由学校领导或教师担任,人员的性质也决定了不能以营利为目的。而社会上的体育俱乐部主要是以营利为目的,更具有功利性,其经营和管理人员在利益驱动下更有动力。相比而言,学校建设并运转体育俱乐部的积极性并不高。离开了体育俱乐部这样的组织,学生课外体育锻炼也就失去了很好的平台。
(1)电磁吸附工装的应用 目前电磁吸附工装在工件搬运及机械加工行业有着广泛的应用,采用磁力吸盘对铁磁性材料进行吸合固定,如机械加工车削、铣削、磨削、镗削、锻件模具加工的紧固工作。
In principle,the dependence of the coefficients on the data in the nonlinearities of the equation may be a functional one.But in order to have the first insight in the difficulties and in order to simplify the presentation,we confine to a point-wise dependence of Hamiltonians with respect to the mean field variable here.Generally,obtaining the existence of solution of the Bellman mean field dependent system with growth with respect to the mean field variable in nonlinearities is the critical subject.Without additional assumptions only a poor growth behaviour is permitted.For the case of scalar Bellman equations,for obtaining the global solvability,[10–11]gave a quite exhaustive analysis of the growth conditions for the Hamiltonians concerning the dependence on∇u(u is a generalized value function)and the mean field m,which appears in the pay off functional.
In comparison,in the present paper,we restrict ourselves to Hamiltonians which grow quadratically with respect to∇u,which is from the point of view of PDE analysis the most interesting case.In a related paper[1],a mean field dependence of the pay off functional is assumed,but no such a dependence of m in the dynamics of the system is considered.Also in[1],thegrowth propertiesof the data with respect to m,werecrucial to obtain global solvability of the problem.
In this paper we go much beyond the scalar theory and the theory developed in[1]and obtain the existence result for much broader class of problems.As a key tool for the existence of a solution we use the method of sub and super solution used in the context of Bellman systems in[2].
2 Derivation of the System
In this paper we study a system of partial differential equations which arises as a necessary condition of a Nash-Point-problem for Vlasov-McKean-functionals
where(0,T)is a given time interval,Ω ⊂ R d is a cube(0,1)d and Q:=(0,T)×Ωis a space-time cylinder.The function v:=(v1,···,v N)with N ∈ N is the vector of the control functions,i.e.,
is the control of the i-th player,where i=1,···,N and M ∈ N is given.For every i=1,···,N the function
is the so-called pay off function of the i-th player.Finally,the function
is the so-called mean field variable.This means that for a given v,it is a weak nonnegative solution of the following parabolic equation(“mean field equation”):
Then using also the assumption(2.21),we see that it satisfies
基于激光雷达技术的京津冀区域气溶胶特性分 析 ………… 温玉海,倪晓昌,董 昊,程飞帆,张 琦(18)
对照组采用常规的方法拔出尿管:在拔管前采用间歇性夹管方式放尿,夹闭导尿管,每2~3小时开放1次,训练膀胱功能。观察组运用循证护理,具体如下:
is a given mapping and we postpone the discussion about its structure to the end of the section.
For a bounded v,under natural assumptions on the mapping g,the problem(2.2)–(2.3)has a unique solution
which allows us to define(2.1).Since the mean field variable depends on the choice of v,we will frequently also write m=m(v),which should not be understood as an algebraic relation,whenever the couple(m,v)solves(2.2).
In(2.4)we use the standard notation for Bochner,Sobolev and Lebesgue spaces,and the subscript“per” indicates the periodicity with respect to Ω and this notation will be kept through the whole paper.In what follows,we also omit writing the dependence of function on(t,x)explicitly to shorten all formulae,i.e.,we use the following abbreviation f(m,v)for f(t,x,m(t,x),v(t,x))or for f(·,m,v)in what follows,where f:=(f 1,···,f N).
Having the mean field variable m and the control function v,we can define a “pre-version”of the so-called Bellman system for a further function via the backward parabolic system
Similarly,wereplace v in themean field equation(2.2)and obtain thebackward-forward system
We call this system the “pre-Bellman equation” since v has not been replaced by a feed back formula
yet and we introduce the meaning of(2.6)in the next subsection.Moreover,L=(L 1,···,L N)are the so-called modified Lagrangians1 We use here the word modified since they diff er from standard Lagrangians in Bellman systems,which is however here caused by the fact that f and g depend on m..
It is evident,that(2.2)and(2.5)do not form a closed problem and one needs to connect v with m and∇u via some relationship.It will be shown in the next subsection that the necessary condition for classical Nash–Point problem may serve as such constraint.In addition assuming further certain qualitative properties of f,we will be able to give a good meaning to the feed back formula(2.6)and thus to avoid the presence of the control variable v in the analysis.The main goal of the paper is to introduce certain structural assumptions on f and g such that they describe very general mean field dependent Bellman system on one hand,and for which we can establish the existence of a weak solution on the other hand.
另外,科研成果具有时效性,若科研成果未在时效期内转换成实际生产力,就失去了价值。当前高校和企业之间更多采用传统的合作模式,技术信息的传递不能保证实时性,一部分科研成果只是以理论成果的形式存放在学校科研管理部门,在科研成果转化推广的过程中,学校、企业、个人之间缺少信息实时流通的平台,在时间和空间上存在节点距离[2]。
2.1 Derivation of the full system
Finally,we need to close the problem(2.2)and(2.5)by an algebraic condition which is necessary condition for the Nash-point of functionals J’s.The classical Nash-point problem reads:For given u T,find v∈L∞(0,T;L∞(Ω;R M N))such that
for all z∈ L∞(0,T;L∞(Ω;R M))and corresponding m’s,the solutions2 We may define this also for other L p(L q)spaces once the uniqueness of m(v)is guaranteed. to(2.2)with v replaced by
The classical version treats the case where f i and g do not depend on m.In that case,the problem(2.7)is purely analytical(meaning stochastic free)formulation of a stochastic differential game driven by the dynamics
In recent years,interest came up to study cases with m-dependence of the pay-off f and/or the dynamics g.From PDE’s point of view,this leads to new interesting version of the Bellman system.
Here
Although,it is not known whether the problem(2.1)admits a Nash-point,we derive in what follows certain necessary conditions that must be fulfilled by the hypothetical Nash-point,which finally allow us to connect the pre-Bellman system(2.5)with the mean field equation(2.2)via the feed back formula(2.6)or its “equivalent”.
Under natural assumptions on the data(see Subsection 2.2)and,say,v∈L∞(0,T;L∞(Ω;R N M)),it is easy to see that the Gateaux derivatives of the J i and of m(v)exist.For
Horn 涡轮旋风切削可用于加工单线、多线螺纹和各种外形轮廓。而且,经过精密打磨的 S271 双刃可转位刀片是为不同的螺纹牙形和加工材料量身定制的。将可转位刀片安装在大小合适、高度稳定的刀片座内,刀片座既可以安装在新型模块化旋风铣刀头上,又可以直接安装在机床动力刀座上。
with arbitrary smoothΩ-periodic functionwe obtain that
with i=1,···,N satisfies
and is completed by the initial condition
Here,we use the subscript to abbreviate the notion of partial derivative,i.e.,g v(v,m):=∂v g(v,m)and g m(v,m):= ∂m g(v,m).Furthermore,assuming that v is the Nash-equilibrium,we have for all i=1,···,N that
for arbitrary smoothΩ-periodic function z.Notice that here m:=m(v),i.e.,m solves(2.2)with v.To evaluate the terms not involving explicitly z,we use the equations(2.5)and(2.8).First,multiplying(2.8)by u i,integrating over Q and using integration by parts(note that M i(0)=0),we deduce
Next,multiplying the i-th equation in(2.5)by M i,we observe
Finally,subtracting(2.11)from(2.10),we obtain the following identity:
Thus,using this relation in the necessary condition(2.9),we see that
for all i=1,···,N and all smooth Ω-periodic z’s.This consequently leads to the necessary compatibility condition
Thus,now we have a closed system of equations.Namely,(2.2),(2.5)and(2.13)forms a well-defined problem for which we want to establish our analytical result.Indeed,the first one will deal just with(2.2),(2.5)and(2.13)and leads to the uniform a prioriestimate for(m,v,u).
However,to obtain also the existence result,we shall require that for a given(m,∇u),we can find a unique v solving(2.13).For such a solution we define the feed back formula(2.6)as
and replacing v in(2.5)by the feed back formula,we obtain
which is supposed to be satisfied in Q,equipped with theΩ-periodic boundary conditions and completed by the initial condition
which does not contain a control function v.Nevertheless,this system is equivalent to(2.2)and(2.5)provided that v is defined such that it satisfies(2.13).The second result of the paper will be therefore established for(2.15),provided that the feed back formulaωis well defined.
2.2 Structural assumptions on f and g
We keep the notation from the introduction here.Through the whole text,we assume that all partial derivatives f v,g v,f m and g m with respect to v and m,respectively,exist and together with f and g are Carathéodory mappings,i.e.,are measurable with respect to(t,x)for all(v,m)and for almost all(t,x)they are continuous with respect to(v,m).This assumption will not be mentioned explicitly in what follows but we rather assume it implicitly in all statements.In what follows we also assume that K and C i are positive constants and the same for constants r,s≥0 which will be used to denote certain powers.Furthermore,we will frequently use C to denote a generic constant that may change from line to line but will depend only on data.In case,it will depend on some important quantity,and it will be clearly denoted in the text.
深圳港水上“巴士”航线的经济性由上述3个指标构成的成本价值量来决定,通过与其他路径经济性的比较,分析深圳港水上“巴士”的优劣势.
First,we state the assumptions forWe assume that f and f v satisfy the following growth conditions:
Please,observe here that if g is given by(2.23),then(2.13)reduces to
and also one sided sum coerciveness,i.e.,we assume
秦、晋围郑,郑国危在旦夕,此时佚之狐的推荐让郑伯看到了希望,烛之武面见郑伯时,先辞后许,寥寥数语,让情节平添了一层波澜。烛之武劝说秦王是整个故事的高潮部分,说辞层层深入、步步紧逼,让读者有一种揪心的紧张,为烛之武捏了一把汗,直到他成功地说服秦王撤军才让读者心神舒缓,稍感轻松。然而在秦军撤退之后,子犯建议晋文公攻打秦军,局势又顿时紧张了起来,最后晋文公以一番“不仁”“不知”“不武”的理性分析和清醒判断平息了一场虚惊。整个故事情节波澜起伏,险象环生,这样一来就使得文章具有极强的艺术感染力。
Concerning the structural assumptions on f we just assume that
We also prescribe the behaviour of f at 0,i.e.,for all i=1,···,N,we assume that
and finally the coerciveness ofi.e.,we assume that
Next,we focus on the assumptions on g.The standard ones are related to the growth estimates,which will be supposed to be given by
The forthcoming structural assumptions on g and also on f v are in fact the key restrictions of the paper.First,to simplify the further analysis,we shall assume that3 The linearity of g with respect to v is in fact not a necessary assumption.A more delicate here is just behaviour of g with respect to m.
where b0 does not depend on(t,x).The matrices are given functions of(t,x), with i=1,···,d and k=1,···,M and the meaning of A j v j is
Hence,as it is assumed we have a d-dimensional vector function g=(g1,···,g d).The inhomogeneities b0:R→R d and b1:R→R are given.Concerning the assumptions on functions b1 and b0,we require that
and for matrices A j we assume that
施工中需要注意的事项如下:①在导墙轴线放样时,应确保其内墙面与壁板桩轴线平行,导墙内侧净距应为壁板桩墙厚+40 mm;②为保证原土不被扰动,挖土可采用机械和人工相结合;③内模立模板、外模以土代模,用振捣器均匀浇捣密实,待砼达到70%强度后方可拆模;④拆模后在导墙顶部用10槽钢支撑,内部用80 mm×80 mm方木及时在墙间加撑,支撑间距为2.0 m,梅花型布置;⑤施工期间严格控制墙顶周边荷载,防止导墙受力向槽内挤压;⑥为保证精度,施工中应随时检查导墙和壁板桩的中心线是否保持一致,竖向面是否保持垂直,如发生偏差及时纠正。
Furthermore,one of the essentially required properties of A j is that they have the same range,i.e.,we assume that for all i,j=1,···,N,almost all(t,x)and all z ∈ R N d there holds
For derivatives of b0,we need that
Finally,for the derivative of b1 with respect to m,we introduce a certain“smallness assumption”:There existsδ∈ [0,1)such that for all b∈ R+and all v∈ R N M there holds
如图1所示,该线路为全高架车站线路,全线利用高架桥轨道梁结构中的钢筋设置了杂散电流收集网,并在收集网测量端子1 m内设置参比电极,采用杂散电流监测装置负责实现对杂散电流的数据监测,变电所负极通过回流电缆与钢轨相连,全线未设置排流柜。
Although the above assumption seems to be complicated,it naturally appears in the first a priori estimate of the problem.In the next subsection,we shall show that in many cases,the assumption(2.28)still allows very general behaviour of all quantities with respect to m,namely,the case when one can dominate the function g in certain sense.In addition,since(2.28)may seem to be complicated,we can replace it by
since then(2.28)follows easily from the previous assumptions.
全国层面,房地产开发企业商品房销售面积整体不断提升,由1987年的2 697.24万平方米逐步增长至2017年的169 408万平方米。如图5所示。
老和尚一本正经地对小和尚说,不好随便下山的,山下有老虎,老虎就藏在林子里,老虎是要吃人的。风影不以为然,那师父你为什么还要去桃花林?了空法师急了,大声道,山下的女人是老虎,老虎就是女人。风影更迷茫了,女人是女人,老虎是老虎,女人怎么会变成老虎的呢?看来,师父八成是吃错了什么药,疯了。
for all v and all m≥0.In addition,we assume the following one sided estimate for f i:There exists α ∈ [0,1)such that for all i=1,···,N there holds
which must be valid for all i=1,···,N and all j=1,···,M.
The above assumptions are sufficient for establishing formal a priori estimates.However,for getting also the existence of a weak solution,we need to give a well meaning to the feed back formula,or in other words,we need to guarantee the unique solvability of(2.30)for given m and∇u.There can be introduced many assumptions that would lead to such a goal but we follow here the most standard one,which is the monotone operator approach.For a given m,we define the mapping T:R N M→R N M by
and assume that it is continuous with respect to(m,v)and measurable with respect to(t,x)and the strictly monotone,i.e.,for all
that is supposed to be satisfied in the space-timecylinder Q,fulfills spatially periodic boundary conditions(with respect to the unit cubeΩ)and is completed by the initial data
and from the standard monotone operator theory,we can conclude the existence of a unique v solving(2.30)and also consequently,we obtain that the feed back law(2.6)is well defined.
2.3 Prototypical examp le
Next,we state the second main theorem of the paper,which is the existence result.
where B i’s are arbitrary bounded measurable matrices.Next,for b1 we consider
and for b0 and matrices A i,we just require(2.25)–(2.27).With such a choice,all assumptions(2.16)–(2.27)and also(2.31)are satisfied and we shall just to show what is the meaning of(2.28).A very direct computation leads to the necessary condition
which is surely satisfied whenever
2.4 Statement of main results
The main result of the paper is twofold.First,we give the uniform a priori estimate result which holds for all sufficiently smooth solutions provided that parameters satisfy the assumptions stated above.
Theorem 2.1 Let g and f satisfy(2.16)–(2.28).Then any sufficiently regular solution(m,v,u)to(2.2),(2.5)and(2.13)satisfies the following estimate:
where
provided that
3天的对峙,案情没什么进展,第三天晚饭后,王敬凯又坐到李桂明面前,以理解的口吻说:“你的案子我们公安已经查了几次,虽然你恨张秋,但张秋丢小孩那天,你是不占有作案时间。可是,后来张秋在报纸上找小孩,你就没做什么?没有什么想法?比如说要弄俩钱花?”
and
Our prototype example is the following.For f,we assume a structure
Theorem 2.2 Let g and f satisfy(2.16)–(2.28)and(2.31).Assume that(2.38)–(2.39)is fulfilled.Then for arbitrary u T ∈ L∞ and nonnegative m0∈ Lσ(Ω)with σ fulfilling(2.37)there exists a weak solution satisfying the estimate(2.36)and fulfilling or almost all t∈(0,T),
and completed by the relationship between v and(m,∇u),
To illustrate the power of the result,we just consider our prototypical example.First in case that s0=0,we see that the only restriction is
In the opposite extreme case,i.e.,if
then
3 Algebraic Estimates for Lagrangians and Hamiltonians
In this section,we derive basic algebraic inequalities that are satisfied for Lagrangians and consequently also for Hamiltonians provided that f and g satisfy the assumption introduced in Subsection 2.2,namely the assumptions(2.16)–(2.28).The key observation is that under these assumptions,the Lagrangians satisfy the lower sum coerciveness and the proper upper estimates.It will be also evident from the estimates below why we require 2s≤r in main results of the paper.
Lemma 3.1 Let f and g satisfy(2.16)–(2.28).Then there exists a constant C>0 and ansuch that for almost all(t,x),all(m,v,∇u)fulfilling(2.13)and all i=1,···,N there hold:
(i)Sum coerciveness
(ii)Upper bound
(iii)Global bound
Proof We start with the proof of(3.1).Using the definition of L in(2.5)we get the identity
Hence,using(2.18)for the first part and(2.22)for the second part,we observe that
where for the second estimate we used the Young inequality.This finishes the proof of(3.1).
Next,we look for(3.2).Using the definition of Lagrangians,we directly obtain for arbitrary ε>0 the following identity:
Next,using(2.17)–(2.18),(2.22)and the Young inequality,we have
It remains to estimate the term with f i.We use the convexity of f i with respect to v i and the assumption(2.20).Then to evaluate we also use the constraint(2.13),which however in our case reduces to(2.30)due to the structural assumptions(2.23).Doing so,we get
Then we estimate terms involving Aℓ∇u i with the help of(2.24),(2.26)and(2.30)as follows
Finally,we focus on estimate of v i.It follows from(2.21),(2.24),(2.26)and(2.30)that
where for the last inequality,we used the Young inequality and the structural constraint(2.30).Hence absorbing the corresponding term to the left-hand side,using(2.16)and the Young inequality,we get
Next,using the Young inequality in(3.5),we find
Finally,we substitute(3.6)into(3.7)to estimate term with v i and with the help of the assumption(2.16),we obtain
Consequently,(3.8)combined with(3.4)directly implies
Thus,if we choosesuch that
which is always possible and the maximal value of suchε0 depends only on the generic constant C,we obtain the desired estimate(3.2).
The estimate(3.3)is then a simple combination of(3.1)–(3.2).The proof is finished.
4 Uniform a Priori Estimates—Proof of Theorem 2.1
This section is devoted to the estimates for solution(m,u,v)of(2.2),(2.5)and(2.13)that depend only on data of the problem provided that the assumptions(2.16)–(2.28)are satisfied.Here,we proceed rather formally,considering that the solution is sufficiently regular.The justification of such a procedure then will be provided in the proof of the existence result,i.e.,in the proof of Theorem 2.2.
4.1 Estimates for m
We start with estimates for m.First,if m0≥0 almost everywhere inΩ,then the standard minimum principle for parabolic equations implies that(sufficiently smooth)solution satisfies m≥0 almost everywhere in Q as well.Next,settingϕ:=1 in(2.40),we get with the help of nonnegativity of m that
Consequently,we have
Next,in order to obtain estimates on(u,v)we need to improve the information about m since the Lagrangians(or Hamiltonians)depend heavily on m.Our goal is to prove the starting point inequality
where the constant C depends only on data given by assumptions on f and g.
Proof of(4.2)We first setϕ :=(m+1)σ−1 in(2.40),where σ is given by(2.37),i.e.,
With such a choice ofϕ,we get the identity
where the second equality follows from the structural assumption on g(see(2.23)).Clearly,the second term on the right-hand side vanishes due to the integration by parts and spatially periodic boundary conditions.For the first integral we use the Young inequality to conclude
Thus,using the bounds for b1 and A j,namely(2.24)–(2.25),we obtain after integration over(0,T)
Next,we setϕ:=u i in(2.40)and z:=m in the i-th equation in(2.41),integrate the result with respect to time and use integration by parts to obtain the following two identities:
and
Subtracting(4.4)from(4.5)we arrive at identity
where for the last inequality we used the Hölder inequality and the uniform bound(4.1).Next,we evaluate the last integral.Using the structural assumption(2.23)and the growth assumptions(2.24)–(2.27),we see that
To estimate I2 we use(2.26)and the constraint(2.30)to get
For the term I1 we use the Young and the Hölder inequalities to obtain for arbitraryε>0,
Finally,substituting these estimates into(4.6),summing the result over i=1,···,N and using(2.28),we obtain
Consequently,sinceδ<1 we can absorb the last term by the left-hand side.Therefore,we can use the sum coerciveness,i.e.,the assumption(2.18),to deduce that
Finally,using the estimate(4.8)in(4.3),we see that(recalling(2.37))
Thus,settingε>0 sufficiently small,we obtain the desired inequality(4.2).
Our next goal is to improve the information coming from(4.2)such that the estimate on E depends only on u and not on m.It means that we want to show that
provided that
The estimate(4.9)will be shown by a certain interpolation of(4.1)–(4.2).
Proof of(4.9)First,we recall the following parabolic interpolation inequality:
Next,we apply(4.12)onto the function withσgiven by(2.37)to get
Finally,we also use the a priori estimate(4.1)and the interpolation inequality
which is valid for allto obtain
where for the last inequality we used(4.13).
Next,we use(4.14)to handle the right-hand side of(4.2).Assuming that
we get from(4.14),
provided that
which by using of(2.37)can be shown to be equivalent to(4.10).Notice that(4.17)directly implies the validity of(4.15).Hence,we can absorb the first term on the right-hand side of(4.2)to get
In case that
we can use(4.1)to conclude(4.9)directly.If(4.19)is not true,we again use(4.14)to get from(4.18)
which after using the Young inequality leads to(4.9),provided that
Note that(4.20)is a stronger assumption that(4.19)and it can be shown by using(2.37)that(4.20)is equivalent to(4.11).Hence the proof of(4.9)is complete.
4.2 Estimates for u
This subsection is devoted to the uniform estimates on u,which will still depend on m.To be more precise,we want to show that for arbitrary p>d+2 we have the estimate
Proof of(4.21)We start with the estimates for below for the quantity
It is not difficult to observe from(2.5),(3.1)and the fact that 2s≤r,that w satisfies almost everywhere in Q,
Next,let us consider w1 a solution to
completed by zero initial condition,i.e.,w1(T)=0.Then by a standard parabolic estimate,we obtain that
whenever p>d+2.Then subtracting(4.23)from(4.22),we obtain
Hence from the theory for subsolutions to parabolic equation(see[8]),we obtain
which together with(4.24)and the assumption that u(T)∈L∞ leads to the final estimate from below
which is valid for arbitrary p>d+2.
Next,we focus on estimates from above.Keeping the notation for w,we can derive from(2.5)that
Hence,using(3.2),we get
where the second inequality follows from the assumption 2s≤r.Thus,we can repeat step by step the procedure for w and using the fact that u(T)∈ L∞(Ω),we obtain
Finally,we derive the uniform bound(4.21).First,summing(4.28)over i=1,···,N,we have
Since,we can combine this estimate with(4.26)to get
Consequently,it follows from(4.28)that
Finally to obtain also estimate from below for u i,we use(4.29)–(4.30)and get
which together with(4.30)implies the desired estimate(4.21).Hence the proof is complete.
4.3 Uniform L∞bounds
This subsection is devoted to the uniform bound for u,which directly implies the part of uniform estimates stated in(2.36).Here,we combine(4.14),(4.9)and(4.21)to obtain the final bound.We go back to(4.21)and estimate the right-hand side.Although we need to choose p>d+2,we formally provide all computation for p=d+2 and in the final restriction on the size of s0(orσ)we just use the strict inequality sign.Hence,we need to estimate the term on the right-hand side of(4.21),i.e.,the integral
If 2s0(d+2)<1 then the integral on the right-hand side of(4.21)is bounded due to(4.1)and therefore we immediately get
Hence,in what follows,we assume the opposite case.Assuming that
we can use(4.14)with q:=2s0(d+2)and we deduce
and it follows from(4.21)that
Inserting this estimate into the right-hand side of(4.9),we also deduce
Hence,in case
we can absorb the right-hand side by left-hand side to obtain
which implies a part of(2.36).Notice that(4.37)is a stronger assumption than(4.20)and therefore all needed assumptions,i.e.,the assumptions(4.17)and(4.37),are already encoded in(2.38).Furthermore,using(4.2)and(4.9),we obtain also the bound for m and∇m stated in(2.36).Finally,from(4.8),we deduce the bound for term with m(m r+1)|v|2 in(2.36).
4.4 Uniform estimates for∇u
This subsection is devoted to the last remaining part of(2.36),i.e.,the part of the estimate with∇u.Keeping the notation from the previous sections,we start with estimates for∇w.Using(3.1)and the fact that 2s≤r we have
Next,we multiply(4.39)byintegrate over Q and use integration by parts to obtain
Thus,we see that we can absorb the term with∇w by the left-hand side and due to the L∞bound for u(see(4.38)),and L 2s0(d+2)bound for m(see(4.34)),we deduce from(4.40)that
Notice that the first in(4.41)together with(4.8)leads to the estimate(2.36)for term involving|v|2.
Next,we use the inequality(4.27),i.e.,
which we multiply byand integrate over Q.Repeating step by step the procedure(4.40)–(4.41)and using uniform bounds on u and m,we get
Finally,combining(4.41)and(4.43),we have for all i=1,···,N,
which finishes the proof of(2.36).
5 Existence of Solution—Proof of Theorem 2.2
This section is devoted to the proof of the existence of a solution to(2.2),(2.5)and(2.30).Notice that due to the assumption(2.31)–(2.32),we know that(2.32)is equivalent to
with a Carathéodory mapping ω.Therefore we can omit(2.30)and replace v by ω(m,∇u)in(2.2)and(2.5)and solve the problem only for unknowns(m,u).In fact,this is also the way how one can get the existence of a solution to an approximative problem.Nevertheless,for the sake of simplicity and to simplify the notation,we keep writing v in what follows.
Second,we do not provide the complete and rigorous proof here.We rather emphasize those steps that are different from the known procedure for Bellman systems.To be more specific,we provide here the proof of weak sequential stability,which is the key property of the system of equations we have in mind.It means that we shall consider a sequence of(m n,u n,v n)of smooth solutions to(2.2),(2.5)and(2.30)(which is however equivalent to(2.15),once the mappingωis well defined)corresponding sequence of initial data
with nonnegativeOur goal is to show that
strongly inwhere the triple(m,u,v)solves again(2.2),(2.5)and(2.30)with initial data(u T,m0).Indeed such a result suggests that the rigorous existence proof is doable.Indeed,approximating Hamiltonians H by a sequence of bounded functions and similarly g by a sequence of boundedone may consider that for a such approximative system it is classical to obtain the existence of solution(m n,u n,v n)and the only remaining part of the proof is then the weak sequential stability.For details,how one can approximate Hamiltonians properly,we refer to[2].
5.1 Uniform a priori estimates
In this part we just use the result of Theorem 2.1,which holds for sufficiently smooth solutions.Indeed,we may assume that
such that(m n,u n,v n)satisfies for all
for all
and almost everywhere in Q,
with g given as
and L(m n,v n,u n)given as
Next,we focus on the estimate for the time derivatives.First,using(3.3)and uniform bounds(2.36),we see that
Consequently,we can deduce from(5.5)and also from(5.3)that for some q>d,
Similarly,using(2.22)and(5.3),we see that for some q∈ (1,∞),
5.2 Limit n→∞
Having(5.3),(5.1)and(5.10)–(5.11)and using the Aubin-Lions lemma,we can find subsequences that we do not relabel,such that for some q>d,
where M(K)denotes the space of Radon measures on K.Having(5.12)–(5.20),we can use the theory for parabolic equations with L 1 or measure right-hand side(see[4–5])to conclude that
Consequently,since the operator T in(2.31)–(2.32)is strictly monotone,it follows from(5.17)and(5.21)that
which in particular also implies(note that L and g are Carathéodory)
Thus,the convergence results(5.12)–(5.24)allow us to let n → ∞ in(5.4)–(5.7)to obtain that(m,v,u)solves(2.40),(2.41)and
for allThus,to finish the proof of Theorem 2.2,we need to show that
Indeed,having(5.26),we can read from the equation that the time derivative of u is better,i.e.,that
for some q>d and integrating by parts with respect to the time variable t in(5.25),we find that(2.41)holds true.
5.3 Identification of L and the proof of(5.26)
This last part is devoted to the proof of(5.26).We follow the procedure developed in[2]with the necessary changes due to the presence of the mean field variable m.We also proceed here slightly formally and refer the interested reader to[2],where the very similar procedure is made rigorously.Defining
it follows from(5.5)that in the sense of distributions we have
Next,we multiply(5.27)by
and obtain
Consequently,multiplying the resulting identity by arbitrary nonnegativeand integrating the result over Q,we get by using integration by parts that
Next,we letThanks to(5.12)–(5.14),it is easy to deduce that
Next,using(3.1),we get that
Consequently,we see that the right-hand side of(5.29)is bounded from below by a strongly convergent function,so using the point-wise convergence result(5.23)and the Fatou lemma,we get
Hence,combining(5.30)–(5.31),we see that for all nonnegative there holds
Finally,takingϕ:=e2Cwψwith arbitrarywe deduce that4 In fact we must mollify the test function with respect to the time variable and then to pas to the limit.Since such a procedure was explained in details in[2],we do not provide the complete proof here.
which implies that in the sense of distributions
Similarly,we deduce the opposite type inequalities.It follows from(5.5)that for all i=1,···,N,
Denoting z n:=(u i)n−ε0w n and multiplying(5.35)by e2Cz n we get
Hence,using(3.2),we see that the right-hand side is bounded by an convergent sequence and therefore we can proceed similarly as before by using the Fatou lemma to obtain
Thus summing with respect to i=1,···,N and dividing by(1 − ε0)we see that
which combined with(5.34)gives
Consequently also we obtain from(5.37)that
Finally using(5.38)–(5.39),we get
Hence,(5.39)implies that
and(2.41)follows.This completes the proof of Theorem 2.2.
References
[1]Bensoussan,A.,Breit,D.and Frehse,J.,Parabolic Bellman-systems with mean field dependence,Appl.Math.Optim.,73(3),2016,419–432.
[2]Bensoussan,A.,Bulíček,M.and Frehse,J.,Existence and compactness for weak solutions to Bellman systems with critical growth,Discrete Contin.Dyn.Syst.,Ser.B,17(6),2012,1729–1750.
[3]Bensoussan,A.,Frehse,J.and Yam,P.,Mean field games and mean field type control theory,Springer Briefs in Mathematics,Springer-Verlag,New York,2013.
[4]Boccardo,L.,Dall’aglio,A.,Gallouët,T.and Orsina,L.,Nonlinear parabolic equations with measure data,J.Funct.Anal.,147(2),1997,237–258.
[5]Boccardo,L.and Gallouët,T.,Nonlinear elliptic and parabolic equations involving measure data,J.Funct.Anal.,87(1),1989,149–169.
[6]Gomes,D.A.,Pimentel,E.A.and Sánchez-Morgado,H.,Time-dependent mean-field games in the subquadratic case,Comm.Partial Differential Equations,40(1),2015,40–76.
[7]Gomes,D.A.,Pimentel,E.A.and Voskanyan,V.,Regularity Theory for Mean-Field Game Systems,Springer International Publishing,New York,2016.
[8]Ladyzhenskaya,O.A.and Ural′tseva,N.N.,Linear and quasilinear elliptic equations,Translated from the Russian by Scripta Technica,Leon Ehrenpreis(eds),Academic Press,New York,1968.
[9]Lasry,J.-M.and Lions,P.-L.,Jeuxà champ moyen,I,Le cas stationnaire,C.R.Math.Acad.Sci.Paris,343(9),2006,619–625.
[10]Porretta,A.,On the planning problem for a class of mean field games,C.R.Math.Acad.Sci.Paris,351(11–12),2013,457–462.
[11]Porretta,A.,Weak solutions to fokker–planck equations and mean field games,ARMA,216(1),2015,1–62.