This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
') ]
>.
.
Substituting this in (2.4) , we arrive at the inequality [E
sup I~(u) - ~(v)IP ]
l ip
~
which proves the lemma, since by separability of [E
I~(u)
sup tl ,vE[O ,T]
-
CO"(T)
1J.,vETs
~(v)IP ]
l/P
= lim sup [ E 8-+00
~(t)
sup I~(u) - ~(v)IP
]l lP
u, vETs
.
n
Lemma 2.1 almost immediately results in the following fact: LEMMA 2 .2. Let ~(t) be an ordered process satisfying (2.1, 2.2) and such that ~ (O) = O. Then there exists a constant C depending on >. such that for all "( > 0
Esup[~(t) _ "(O"q(t)]P ~ C[2q(p + 2) - 4]q(P+2 )-2, t2:0
+
(2.5)
"(p l(q-l )
where [xl+ = max(O, x ). Proof We will use the following form of the Markov inequality
(2.6) which immediately results from the banal inequality
Without loss of generality, we may assume that 0"2(t) is continuous and such that limt-+oo 0"2 (t) = 00 . Then for any integer k ~ 0 we can find tk b) such that
YURIGOLUBEV
114
Using that f (x ) = x Pl {x > xo} is monotone in x > 0, we have E sup [~ (t) - I'aq(t)]: t ~O
00
s 2:E k=O
e (t) l{~ (t) ~ I'aq(t) }
sup t E [t k(-r ),t k+l (-Y)]
00
:::; 2: E
k=O
e (t )l{ ~(t) ~ I'aq(tk(,,)) }
sup t E [t k(-y) ,tk+ l(-y)]
(2.7)
00
:::; 2:E
k=O
:::; E
e (t )l{
sup tE [tk (-y) ,t k+l (-Y )]
sup tE [t k(-y) ,tk+l (-Y)]
«o ~ I'aq(t k(,,)) }
1~ (t)IP
sup °991 (-Y) 00
+ 2: E k =l
sup
0::; t::; tk+ 1(-y)
e(t) l{
sup
O::;t::;tk+l (-y)
~(t) ~ I'aq(tk(,, ))} '
By Lemma 2.1, the first term at t he right-hand side of the ab ove inequality is bounded as follows
whereas the second one , in view of (2.6), is controlled by 00
2: E
e (t) l{
su p O::;t::;tk+ l(-Y)
k= l
C
:::;
sup O::;t::;t k+ l (-y)
~(t) ~ I'aq(tk(,,)) }
d p+d 00 aP+d(tk+l(,,)) _ C(p + d)p+d 00 (k + l )p+d (p +) 2: baq (tk (,,) )]d I'p/(q- l) 2: . kqd/(q-l) k= l
:::;
k=l
C[2 (p + d)] p+d 00 1 I'p/(q- l) 2: kd/ (q -l )-p ' k=l
Setting d = (q -l )(p + 2) in the above inequ ality and using (2.7) toget her 0 with (2.8), we prove (2.5).
2.2. Some examples of ordered processes . The simplest example of an ordered process is ~ (t) = ~t , where ~ is a zero mean random variable with a finite exponential moment E cosh ( A~ ) < 00 for som e A > O. As we have already mentioned, the Wiener process W (t) is an ordered process. At the first glance, ~t and W (t ) are qu ite different, but from the viewpoint of Lemma 2.2 they ar e equivalent. Of cours e, t he distribution of maxt ~o [W (t ) - I't ] is well-known
P {max[W(t ) t~O
I' tl ~ x} = exp ( - 2I'x ).
115
ON ORACLE INEQUALITIES
The next two examples play an essential role in adaptive estimation. Let H, (.) be a family of ordered smoothers (see Definition 1.1). Consider the following Gaussian processes n
e+(t) = 2:)Hto(Ak) - Hto+t(Ak)]bke(k) , k=1
t
>: 0
n
e-(t) = LlHto(Ak) - Hto-t(Ak)]bke(k), k=1
0 ~ t ~ to,
where e(k) are i.i.d. N(o , 1) and 2::1 b; < 00. It is easy to see that e+(t) and e-(t) are ordered processes. Indeed, in view of (1.12) we have for
t2 :::: t1 n
Ee~(h) = L lHto(Ak) - HtO+t1(Ak)][Hto(Ak) - HtO+tl (Ak)]b~ k=l n
~ L [Hto(Ak) - HtO+tl (Ak) ][Hto(Ak) - HtO+t 2(Ak)]b~ k=1 and similarly, n
Ee~(t1) = LlHto- t1(Ak) - Hto(Ak)][Hto-tl (Ak) - Hto(Ak)]b~ k=l n
~ L[Hto- t1(Ak) -
u.; (Ak)][Hto- t2(Ak) - n.;(Ak) ]b~
k=l
Therefore with Lemma 2.2 we get
E sup [e+(a) - , i)H"o(Ak) "~" O
E sup [e-(a) - , f)H" 0 (Ak) "~"o
k=1
H"(AkWb~] p ~ C(~), +
k=1
"[
H"(AkWb~]P ~ C(~) , +
,
thus arriving at LEMMA 2.3 .
Let {H,,(-) , a :::: O} be a fam ily of ordered smoothers,
th en for any , > 0
E~~~[~lH"o(Ak) -, 't[H"o(Ak) k=l
H,,(Ak)]he(k)
H"(Ak)fb~] P ~ C(~) . +
,
(2.9)
YURIGOLUBEV
116
The next imp ort ant orde red pro cess is defined by n
T/(t ) = I: H 1 / t (>\k)(e (k) -1 ), k= 1
where e(k) are i.i.d. N (O, l) and {HtO , t 2: O} is a family of ordered smoothers. It is easy to check t hat
So, in order to apply Lemma 2.2, it remains to check (2.2). Denoting for brevity n
IIH"'2 - H"'l 11 2= I:lH"'2(>"k ) -
u.; (>"k)]2,
k= 1
we have
(2.10)
Since obviously
t hen using the Taylor expansion for 10g(1 - .) at the right-hand side of (2.10) , we get for>" ::; 1/2 E exp[>".6. e(02 ' 01)]
s exp (C>..2),
t hus pr oving (2.2). Therefore using Lemma 2.2, we obt ain t he following fact . LEMMA 2.4. Let {H", O , 0 2: O} be a f amily of ordered smoothe rs, then for all , > 0
117
ON ORACLE INEQUALITIES
2.3. Proof of Theorem 1.1. Denote for brevity ek = (e, ¢k ). We begin with a simple auxiliary lemma that is cornerstone for the proof. LEMMA 2.5. Let {H,A ·), 0:: 2': O} be a family of ordered smoothers. Then there exists a constant C such that for any data-driven smoothing parameter 6
(2.12)
00
::; CEo m:xAkHl(>\k)Eo 2:)1
-
He,(Ak)]2e~.
k=l
and
(2.13)
00
::; CEo m:XAkHl(Ak)Eo
2:[1 - He,(Ak)]2e~ . k=l
Proof Let
0::0
be a given smoothing parameter. We obviou sly have
00
Eo
2:[1 - He,(Ak)h!\~ek~(k)
(2.14)
k=l 00
= Eo 2:[Hc>o(Ak) - He,(Ak)]Aek~(k) . k=l
It follows immediately from (2.9) that
lEo ~[Hc>o(Ak) - He,(Ak)]Aek~(k)1 ::; l Eo f)Hc>o (Ak) - He, (Ak)]2 Ake~
+ C.
k=l
I
Therefore minimizing the right-hand side in I ' we obtain
To bound from above the right-hand side at the above display, we use once again that Hc>(-) are ordered smoothers. So, when 6::; 0::0 we obtain
YURI GOLUBEV
118
00
~ m:XAkH~o()\k)L [l-H& (AkWB~ k=l and simil arly for &
~ 0:0
00
00
L[Hao(Ak) - H& (Ak)]2AkB~ ~ m:XAkH&(Ak ) L[l - Hao (Ak) ]2B~. k=l
k= l
Therefore combining these inequalities with (2.14) and (2.15), we get
Using the element ary inequality 2ab ~ ab ove display as follows
f.W2
+ b2 / u ; we
can conti nue the
i- ~[1- H& (Ak)]ABk~ (k)1 ~[
::; fL L..t 1 -
k=l +fLEo
Hao (I\k \ )] 2112 CmaXk AkH~o (Ak ) k + ----::::-".-.:..--'-(l
I:[l -H& (Ak)]2B~ +
fL
CEo maXk AkHl (Ak) .
k= l Therefore minimizing the right-hand side in
fL 0:0 ,
we get
lEo ~[1 - H& (Ak)]ABk~(k)1 . f{ fL Z:: ~ [ 1 - H aD (\I\k )J2 Bk2 + ----'--"--CmaXk AkH~o(Ak )} ::; m ~
~l
I:[l -H& (AkWB~ + I:[l- H&(Ak )] 2B~ +
+EO{fL
k=l
::; 2EO{fL
~l
fL
CmaXk AkHl(Ak)} fL CmaXk AkH l (Ak)}. fL
119
ON ORACLE INEQUALITIES
To finish the proof of (2.12) it suffices to minimize the right-hand side in f-L . Inequality (2.13) follows from (2.12) since Hcx(A) = 2Hcx(A) - H;(A) are ordered smoothers and we can apply (2.12) with HcxO = HcxO. 0 In view of the definition of the empirical risk and (1.9) , we have n
Rpen [Y, &] =
,
'2
IIBo - B&II + a
2
2"'"
Pen(&) - o Z:: Ak
k=l n
= L[l - H&(Ak)]2B~
+ (]'2 Pen(&)
k=l n
+ L[H~(Ak) - 2H&(Ak)]Ake(k ) k=l n
+2(]' L[l - H&(Ak)]2.j:\;Bk~(k)
k=l and n
liB - 8&11 2 = L[B(k) - H&(Ak)y(kW k=l n
00
= L[l - H&(Ak)]2B~
k=l
+ (]'2 L AkH~(Ak)e(k) k=l
n
-2(]' L [l - H&(Ak)]Bk.j:\;H&(Ak)~(k) .
k=l Therefore for the excess risk we have
n
-2(]' L[l - H&(Ak)lBk.j:\;~(k)
k=l
(2.16)
n
-2f-L(]' L[l - H&(Ak)]2Bk.j:\;~(k) k=l
+ Cmaxk (]'2AkH~(Ak) + 2(1+f-L)(]'2 f-L - (1
+ f-L)(]'2 Pen(&)
t
AkH&(Ak)e (k)
k=l - f-L(]'2
~ AkH~(A k)e(k)}.
YURIGOLUBEV
120
The last two lines can be bounded by cr 2 !:l~en (p,). Indeed,
E{ Ccr 2
maXk
AkH~()'k) + 2(1 + J-L)cr 2 t
AkH&(Ak)e(k)
~l
J-L
-(1 + J-L)cr 2Pen(&) - J-Lcr 2 t AkH~(Ak)e(k) } k=l Ccr2 maxi, AkH~(Ak)
::; E sup {
'"
+2(1 + J-L)cr 2
J-L
n
(2.17)
L Ak H", (Ak)e(k) k=l
-(1
+ J-L)cr 2Pen (ex) -
2
J-Lcr 2 t AkH;(Ak)~2(k) } k=l
= cr !:l~en (J-L) .
Finally, with Lemma 2.5 we obtain Eo {-J-L t[l _
k=l
H&(Ak)]2B~
_ Ccr
2 maxk
AkH~(Ak)
J-L
-2cr ~[l-H&(>\k)]Bk~~(k) - 2J-Lcr ~[1-H&(Ak)]2Bk~~(k)} ::; O. This inequality together with (2.16) and (2.17) completes the proof of the theorem. 2.3.1. Proof of Theorem 1.2. In view of Theorem 1.1, it suffices to check that
(2.18) where
121
ON ORACLE INEQUALITIES
We begin the proof of (2.18) with the deterministic term. By (1.13) we get su p { CmaXk
)..kH~(>'k) -
!!:-
fl,
,,>0
~ sup {
1
CA
-
K
_1_
,,>0
fl,
[Ln AkH~(Ak) ]K k=1
~ su p { C)..~-K xK _ x:2:0
t AkH~(Ak)}
J.L
(2.19)
2 k=1
!!:-x} =
~
n AkH~(Ak) } L k=1
C 1/(1- K) A1fl,(K+1) /(K-1).
2
Denote for brevity
H (A) = 2(1 + J.L)H,,(Ak) - J.LH~(Ak) " 2 + J.L . Our next step is to show that
E
~~~{ (2 + J.L) ~ AkH,,(Ak)[e(k) - 1] _ .~ ~ AkH~(Ak)}
(2.20)
:S C 1/(1-K)A1 (1- ,,")-1/(1-K)J.L(K+l) /(K-l) . It is easy to see that in view of (1.14)
Next notice that if {H",( ·), a?: O} is a family of ordered smoothers, then {H,,(-), a?: O} is also a family of ordered smoothers. Therefore by Lemma 2.4, for any J.L > 0 we obtain E
s~p{ (2 + J.L) ~ AkH,, (Ak)[e(k) -
1] -
~ ~ AkH~(Ak)}
~ (2 + J.L)ES~P{~ AkH",(Ak)[e(k) -1 ] _ ~
K J.L A1 [0"2(a) (1 - ,,")(2 + J.L)2] 1/ (1+ )} 2(2 + J.L) 2Ai(2 + 2J.L)2
C(l
+ J.L )2/ (1-1< )2(4+K) / (1-K)A1(1- ,,")-1/ (1-K)J.L (K+1 )/ (K-1).
thus proving (2.20) . Thus (2.18) follows obviously from (2.19) and (2.20) .
122
YUR IGO LUBEV
REFERENCES [1J L . CAVALIER AND Y U. GO LUBEV, R isk hull m et hod and regularization by projections of ill-posed inverse problem s, A nn . of St at . (2006), 34: 1653-1677. [21 L . C AVALIER, G .K. GOLUBEV, D. PICARD , AND A .B . T SYBAKOV, Oracle in equalities fo r inverse problems, Ann . of St a t . (2002), 30: 843-874. [3J H .W . E NGL, M . H ANKE, AND A . N EUBAUER, R egulari zat ion of Inverse Pro blem s, Kl uwer Academic Publishers , 1996. [4J L A . IBRAGIMOV AND R.Z. KH ASMINSKII, Statisti cal Estimation. A symptoti c T heory , Springer-Verlag , NY, 198!. [5] A . K NEIP, Ordered lin ear smo others, Ann. Statist. (1994), 22 : 835-866. [6J B . M AIR AND F .H . R UYMGAART, Statistical estimation in Hilbert scale. SI AM J . Appl. M a th. (1996) , 56 : 1424-1444. [7J A .N . TIKHONOV AND V .A. ARSENIN, Solution of Ill-pos ed Problems , Winston & Sons, Washington, 1977 . [8] A . VAN DER VAART AND J . W ELLNER Weak convergenc e and em pirica l process es. Springer-Verlag, NY , 1996.
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTIONS ALTERNATIVE OLEC V. LEPSKI* AND CHRISTOPHE F . POUET*t Abstract . In this paper, we consider the problem of the minimax hypothesis testing in the multivariate white gaussian noise model. We want to test the hypothesis about the absence of the signal against the alternative belonging to the set of smooth composite functions separated away from zero in sup-norm . We propose the test procedure and show that it is optimal in view of the minimax criterion if the smoothness parameters of the composition obey some special assumption. In this case we also present the explicit formula for minimax rate of testing. If this assumption does not hold, we give the explicit upper and lower bounds for minimax rate of testing which differ each other only by some logarithmic factor. In particular, it implies that the proposed test procedure is " almost " minimax. In both cases the minimax rate of testing as well as its upper and lower bounds are completely determined by the smoothness parameters of the composition. Key words. nonparametric hypothesis testing, separation rate, minimax rate of testing, composite functions , structural models, metric entropy, gaussian random function, Implicit Function Theorem. AMS(MOS) subject classifications. 62ClO.
1. Introduction. In this paper, we study the problem of minimax hypothesis testing in the multidimensional gaussian white noise model
dXe(t)
= g(t)dt + cdW(t),
t
= (tI, . . . , td) E'D o
(1.1)
where 'Do is an open interval in lR d , d 2: 1, W is the standard Brownian sheet in lR d and 0 < e < 1 is the noise level. Our goal is to test the hypothesis on the absence of the signal g, i.e.
H:
g=O,
against the alternative written in the following form
G(1Pe) = {g
E
A:
9 E
G(1Pe) ,
G:
Ilglloo
~ sup Ig(t)12:1Pe} . tED
Here G is a compact subset of 1L 2 ('Do) endowed by the metric generated by 11 ·1100 , 'D c 'Do is an open interval in lR d and 1Pe -> 0, E -> 0, is the separation sequence. We consider the observation set 'Do which is larger than 'D in order to avoid the discussion of the boundary effects. Without loss of generality we will assume that 'Do = [-1, l]d and 'D = [- 1/2, 1/2]d. We define a decision rule to be any measurable function of the observation {Xe(t), t E 'Do} taking the values 0 and 1. *Laboratoire d'Analyse, Topologie et Probabilites, UMR CNRS 6632, Universit e d 'Aix-Marseille 1, 39, rue F . Joliot-Curie, 13453 Marseille cedex 13, France (lepski~cmi.univ-mrs. fr ). t (p ouet ccmt . univ-mrs . fr ). 123
124
OLEG V. LEPSKI AND CHRISTOPHE F. POUET
1.1. M in imax approach. To measure the performance of any decision rule we will use so-called minimax criterion. Let J!l'9 be the probability measure on the Borel a-algebra of C (D o) generated by the observation (1.1). For any decision rule ii we int roduce the risk function
1<.(ii ,1fe)=J!l'o{ii=1}+
sup J!l'g{ii=o} ,
(1.2)
gEG C1/J. )
which is the sum of the first and t he second errors probabilities. We say that the separation sequence 'Pe is the minimax rate of testing and ii is t he minimax decision rule if 1. limc~oo limsuPe->o 1<.(ii, Cepe) = 0; 2. limc~oliminfe~oinf}. 1<.(ii ,C'Pe) = 1, whe re inf is taken over all possib le decision rules. It is worthy of mentioning t hat t he minimax rate of testing as well as t he minimax decision rule usu ally dep end on the parameter set G t hat makes its choice very import ant. In t he next section we discuss t he choice of G for a model described by a mult ivar iate function .
1.2. Choice of parameter set. Structural assu m ptio ns. It is well known t hat the main difficulty in testing under alternatives being multivariate functions is the curse of dimensionality: the best attainable rate of t esting decreases very fast , as the dimension grows . To illustrate this effect , suppose, for example, that the underlying function 9 belongs to G = lHId( V, L ), u :» O,L > 0, where lHId(V, L) is an isotropic Holder class of fun ctions. We give the exact definition of this functional class later . Here we only mention that lHId(V, L) consists, in particular, of functions 9 with bounded partial derivatives of order less or equal LvJ and such that , for all x, y E Do,
where Pg(x ,y - x) is the Taylor polynomial of order LvJ obtained by expansion of 9 around the point x , and II . II is the Euclidean norm in JRd. P arameter u characterizes the isotropic (i.e., t he same in each direction) smoothness of function g . If we use the risk (1.2) , t hen ( [11], [16]) the r at e of testing is given by
It is clear that if v is fixed and d is large enough this asymptotics of the rate is too pessimistic to be used for real data: the valu e 'Pe ,d(V) is small only if the noise level e is unreasonably small. On the other hand, if the noise level e is realistically small the above asymptotics might be of no use alr ead y in dime nsion 2 or 3. This problem arises because the d-dimensional Holder class lHId(V, L ) is t oo massive . A way t o overcome the curse of dimensionality is to consider
HYPOTHESIS TESTI NG UNDER COMP OSIT E FUNCTI ONS
125
models wit h poorer functional classes G. Clearly, if the class of candidate functions g is smaller , t he rat e of testing is fast er. Note t hat t he "poverty" of a funct ional class can be describ ed in t erms of rest ricti ons on its met ric ent ropy and there are several ways t o do it. The way we will follow in the pres ent paper consists in imposing a structural assumption on the function g . The classical exa mples are pr ovided by the single ind ex additive and projection pursuit structures ([3], [4], [6], [7], [8], [9], [20] and [21] among others). (i) [Sin gle-index m odel.) Let e be a dire ction vecto r in]Rd, and assume that g(x) = f (eT x) for some unknown univari ate functi on f. (ii) [Additive m odel.). Assum e that g(x ) = L~=l f i (Xi ) , where fi are unknown univariat e functions. (iii) [P rojec ti on pu rsuit regressi on.) Let e 1 , .. . ,ed be dir ect ion vectors in ]Rd , and assume that g(x) = L~=lfi(eTx) , wh ere I, ar e as in (ii). (iv) [Multi-in dex model.) Let e1, . . . , em, m < d are dir ect ion vectors and assume that g(x ) = f eeT x , . . . , e;;'x ) for some unknown rndimension al function f. In genera l, und er structural assumpt ions the rate of testi ng improves, as compared t o the slow d-dim ensional rat e 'Pe.,d(l/). On t he other hand, the assumption that the underlying fun ct ion g belongs to a poor functional class can lead t o inad equ ate model. In general, it is qu it e restrict ive to ass ume t hat g has some parametric str ucture, i.e. the str uct ure described by a finit e dimension al par amet er and remained un chan geable in the whole doma in of observati on. Thus, we seek a rather general (nonparametric) str uct ural restriction which would ad mit powerful testing pro cedures wit hout sacrificing flexibility of t he mod elin g. We argue t hat t his program can be realized if the un derlying function g is a composition of two smooth funct ions . 1.3. Composite functions. We now define our nonparametric struc ture imp osed on the model. We will assume that g is a com posite fun ction, i.e., t hat get) = f (G (t )), where f : [0, 1] -; ]R and G : V o -; [0,1 ]. All our results remain valid if we replace [0, 1] in the definition of f and G by some other bounded interval in R We will further suppose that f and G are smooth functi ons such that f E lHhb, L 1 ) and G E lHId(,8 , L2) where "( , L 1 ,,8, L 2 are positive const ants . Here and in wh at follows JH[1 b, L 1 ) and lHId (,8, L 2 ) ar e t he Holder class on [0, 1] and t he isotropic Holder class on V o respectively (see Definiti on 1.1 below). The class of composite funct ions g with su ch f and G will be denoted by lHI(a, £), where a = b,,8) E lR~ and £ = (L 1, L 2 ) E ]R~. The performan ce of a t est ing pro cedure will be measured by the risk function (1.2) where we set G = lHI(a, £). We will see t hat t he val ue of a det ermines th e quality of testing associated to our mo del, i.e. , t he rate of testing.
126
OLEG V. LEPSKI AND CHRISTOPHE F . POUET
We now give the definitions of anisotropic Holder ball and discuss some trivial cases of testing under alternatives being composite functions. DEFINITION 1.1. Fix v > 0 and L > 0 and an interval V' in ]Rd. Let lv j be the largest integer which is strictly less than v and for f = (k l , .. . , kd) E N d set IfI = k l + .. .+ kd. The ISOTROPIC HOLDER CLASS lHId(V, L) ON Viis the set of all functions G : V' ~ ]R having on V' all partial derivatives of order lv J and such that J
-
"" I fJlkIG(y)~ kd L..J ~ kl
Ikl=Lvj
uX'" uX d 1
I
fJ1kIG(z) < Lily - zllv-Lv j I ~ kl !:l kd , V Y, z E V . UX I ' " uXd
where Xj and Yj are the jth components of x and y . REMARK 1.1. It is easily to see if G E lHId(V, L) then
rr d
G(Y) - ""
I
L..J
O~lk'~ Lv j
fJ 1kIG(x) fJxk1 ... fJx kd I
d
(Yj_Xj)k k .!
j=1
j
I -< Lily-xliV ,
V X,Y E V'.
J
where Xj and Yj are the jth components of x and y. REMARK 1.2 . It is also evident that if G E lHId(V, L) then fJG
~ E
lHId(V - 1, L) ,
UXI
Trivial cases. Zone of super-slow rate: 0 < "(, (3 < 1. Clearly, in this zone JHI(a, £) c lHI d ('y(3 , L 3 ) , where L 3 is a positive constant depending only on "(, j3 and £. Due to this inclusion, a standard testing procedure [11] converges with the rate
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTIONS
127
We finally remark that if 13 ~ 1 the composite function 9 is rather non-smooth. The minimax rate of testing on lHI(a, £) is the same as on the Holder class lHId«l/\ ,)13, .). This is a very slow rate 'Pe,d«l/\ ,)13). Therefore, only for 13 > 1 we can expect to find decision rules with powerful statistical properties and later on only the case 13 > 1 will be studied. Moreover, all results in the paper are proved under assumption d = 2. Below we discuss in detail this restriction.
2. Test procedure and main results. 2.1. Test procedure. Our test procedure uses two different constructions of the decision rules. Both of them are based on the kernel estimators of the underlying function 9 = f(G). Let K : JR --. JR be a function satisfying the following assumption. Assumption (K). 1. K(u) = 0, Vu ef. [-1/2,1 /2].
JK(u)du=1.
2.
3N E N* such that JujK(u)du = 0, Vj = 1,N. 4. KElHI 1(1, 1) on JR. Put K(v) = K(Vl)K(V2) , v E JR2 . We will say that K is a kernel. First construction. Put 3.
213
h*
=
e 2-Y13 + 13 + 1
if
13>2,-1;
if
f3~2,-1;
'213
{
[eJln (l /e))
where a =
13 - lf3J,
(13+1)2 ,
1, a,
if if
e-: 2; 13>2;
and define VA > 0
h[A] We set for any r
X= {
= (q, A, x),
K-13
= h*A7J+f.
where q : JR --. JR, A> 0, x E
1),
Later on we will see that ge,i(r)(IIKI12/vIXh),i = 1,2, with the special choice of the function q and parameter A can be viewed as the kernel estimator of the function 9 at the point x. Set _ { e2-yS+J+1 / Jln (l/e), c; -
if
13 > 2, - 1;
if
13 ~ 2,- 1,
13--y
[eJln (l/e)] :rwm ,
128
OLEG V. LEPSKI AND CHRISTOPHE F. POUET
A=
{Aj =
(~2jr,
1/2 < r;2 J
j = O,l},
~ 1,
where K, = 1/({3 - 1) VI,. Put 'I' = lHh ({3, S) ® A ® 1), where S is the constant from Lemma D.I. Introduce the random events
{~~~ Ig""i(T) (h[A]/Axf/2131 ~ 3eL({3,L2)} , i = 1,2,
A =
where £({3, L 2 ) = yC(8S) 21 [1 - II {31 and the constant C is defined in Lemma A.I. The first decision rule is defined as RA1 V RA2 . Second construction. Put 13+1
f-l
=
e~ Jln(~/e),
if
{3 > 2, - 1;
[eJln (lie))
if
{3
{
and define for any x E
- ( )-
g",
X
-
i3+T,
~
2, - 1,
1)
1
f-lII K ii2
J (t K
1 - Xl
f-l'
t2 -f-l X2) dX""tl" ( '\
Clearly, that g",(x)[IIKlldf-l] is the standard kernel estimator with the bandwidth f-l . Introduce the random event
The second decision rule is defined as RB . Our final decision rule is simply the maximum between the first and second one, i.e .6.; = RA1 V RA2 V RB .
2.2. Main results. For any separation sequence ¢'" us denote
-t
0, e
-t
0 let
and consider the separation sequence
=
e2")~~+l
if {3 > 2, - 1; '13
{ [eJln (l ie))
i3+T,
if {3 ~ 2, - I.
THEOREM 2.1. Let 11 E JR~ and £ E JR~ be fixed. There exists C* = C(ll, £) > 0 such that VC > C*
lim sup ",--0
[lP'o{A IUA2 U B } +
sup 9ElHI(C'P
lP'g{;hn,,42nB}]
=0 .
HYPOTHESIS TESTING UNDER COMPOS ITE FUNCT IONS
129
The explicit expression of the constant C* can be found in the proof of t he theorem. THEOREM 2 .2. For any given n E JR~ and i: E JR~ lim inf lim inf il}f [JPlo{Li e = C-+O
e-+O
6.
•
1} +
sup ( g EH C
( l/ e ) l ~
)
IP'g{tS e = O}]
=
1,
where inf is taken over all possible decision rules. Here T/ = 0 if (3 ::; 2'Y - 1 and T/ > 0 is an arbitrary number if (3 > 2'Y - 1. R EM A RK 2.1. (Upper bound) 1. The assertions proved in Th eorem 2.1 and Theorem 2.2 allow us to state that 6; is minimax decision rule and 'Pe(n) is the minimax rate of testing in the case (3 ::; 2, - 1. 2. Moreover, if (3 ::; 2'Y - 1 the rate of testing 'Pe( n) does not depend on , and is equal to the minim ax rate 'Pe,2((3) associated to the isotropic Holder class lHId((3, .) . This is rather surprising: the quality of the test under alternatives written in the form the 9 = f (G) is the same as for the altern atives described by function G independently of function f. Such a behavior cannot be explained in the terms of smoothness: f (G ) E lHI2 b ,.) and does n ot belong to lHI2((3, .) 3. Also , it uiorihs to m ention that the method of obtaining of lower bound in the case (3 ::; 2, - 1 can be simply generalized for an arbitrary dimension d > 2 . The corresponding domain would be (3 ::; db - 1) + 1. REM ARK 2.2 . (Lower bound) 1. In the case (3 > 2'Y -1 the separation sequences in Theorem 2.1 an d in Theorem 2.2 do not coincide. However, they differ from each other only by the fa ctor [In (1/ gj'7 , where T/ is an arbitrary positive number. It allows us to say that the decision rule 6; is "alm ost " minimax decision rule . We are definitely sure that the separation sequence 'Pe(n) from Theorem 2.1 is the minimax rate of testing and, therefore, the lower bound result should be improved. 2. Let us also note that the lower bound construction used in the case (3 > 2, - 1 is heavily based on the assumption d = 2. In t he context of t he las t remark it is interesting to compare our results with the resul ts recently obtained in [13] for t he estimation of composite function. In [13] t he authors study t he minimax rate of convergence of estim ators, i.e. t he asymptotics of minimax risk whic h is defin ed as
where inf is taken over all possible estimators . The lower bo und result proved in [13] states t hat for any dimension d 2: 2 minimax rat e of conver gence ca n not be fast er than
130
OLEG V. LEPSKI AND CHRISTOPHE F. POUET 213-r
"l/J,,(a) =
{
[eJln(1/e)]2 "Y1'+i3+d 21' [eJln (l ie)] 2 i3 +d,
1 ,
if (3
> db - 1) + 1;
if (3::; db
-
1) + 1.
Comparing this result with the results given by Theorem 2.1 we can conclude that the minimax rate of convergence of estimators and minimax rate of testing differ if (3 > 2,-1 and d = 2. This result is very unusual. As far as we know it is the first problem described in the literature where the minimax rate of convergence of estimators and the minimax rate of testing are different in the case of sup-norm losses. In particular, the upper bound result proved in [13] under additional assumption, E (0,2) , (3 E (1,2] shows that "l/Jr;;(a) is the minimax rate of convergence of estimators if j3 ::; db - 1) + 1. As we see both rates, CPr;; (a) and "l/Jr;; (a) coincide if (3 ::; 2, - 1 and d = 2. 2.3. Open problems. OTHER STATISTICAL MODELS AND PROBLEMS .
1. We considered the gaussian white noise model because it is the simplest and idealized object to study, its analysis requires a minimum of technicalities. Composition structures can be studied in the same spirit for more realistic models, such as nonparametric regression with random design , nonparametric density estimation and classification. Note that our theorems can be directly transposed to gaussian nonparametric regression model with fixed equidistant design using the equivalence of experiments argument
[2], [18]. 2. We restrict our study to the sup-norm loss and to the Holder smoothness classes . A natural extension would be to consider models where the risk junction is described by other norms and other smoothness classes, such as Sobolev and Besov classes, or, for instance, by the classes of monotone or convex functions. The case of functional classes with anisotropic smoothness is of interest as well. 3. We consider only the simplest composition f(G), where f : JR -+ JR and G : JRd -+ R A more general description could be , G k ) , where! : JRk -+ JR and G s : JRd. -+ JR, s = K" and ! (G 1, d1 + + dk = d. 4. A related more complex modeling can be based on Kolmogorov's theorem of representation of a continuous function of several variables by compositions and sums of functions of one variable addition [14], [19] . REFINEMENT OF ASSUMPTIONS.
1. In the present paper we are able to treat only the case d = 2. We believe that our upper bound result (Theorem 2.1), including t esting procedure, with minor changes can be directly applied in any dim ension in the case (3 ::; d(, - 1) + 1. It will require onl y
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTIONS
131
to prove smoothness properties of the implicit function of d - 1 variables analogously to result obtained in Lemma D.l. This conjecture is partially confirmed by the results from [13] for the case "( E (0,2),,6 E (1,2]. To construct the decision rule one can apply the estimation procedure proposed in [13] which is quite different from the constructions used in the present paper. In view of Remark 2.2 this approach would bring a minimax decision rule. 2. However, the lower bound result given by Theorem 2.2 if ,6 > 2"(- 1 uses rather sophisticated construction of random walk on JR recently proposed in [5]. As far as we know such result does not exist in the dimension larger than 1, that restricts the use of our lower bound construction in the dimension larger that 2. Moreover, as it has been already mentioned in Remark 2.2, even if d = 2 the lower bound is not exact. As a consequence, in the case ,6 > d("t - 1) + 1 the extension to an arbitrary dimension as well as the exact lower bound for d = 2 for the risk function (1.2) remain open problems. 3. Proofs. 3.1. Proof of Theorem 1. I. Let us find the upper bound for the first error probability, i.e. for lPo{AI U A 2 U B}.
For given and put
x E A let
us denote 'I.x
=
{T = (q,.\, x),
q E lHl(,6, S),
XED}
Note that
lPO{Al
U
A2 U B} :S lPo{B} + 2
L
lPo{A.x,t},
(3.1)
.xEA
since the distributions of the gaussian random functions cide under lPo. Put also
First we find the upper bounds for IE sUPrE'I".>. IE sUPxE'D ((x). I. La. Upper boundjorIEsuPrE'I"€(T).
90:,1
«T),.\
and
90:,2
coin-
E A, and for
132
OLEG V. LEPSKI AND CHRISTOPHE F. POUET
Let us fix A E A, T = (q,A,x) E'I A, T' = (q',A,X') E'I A and consider the intrinsic semi-metric generated by gaussian random function €(-) :
Putting 'it E JR
we have
and, therefore , h(T, T')
~ IIK~x - K~\xIl2 + IIK~\x - K~\x'112 A
(3.2)
yIJ; + yIf;.
IIK;,xIl
Taking into account that = 1, 'iT = (q, A, x) E 'I and the definition 2 of the kernel K we obtain by direct calculations
It =
2[1- (K; ,,,, K;,,,, )] = -2(K;,,,,K;,,,,-K;,,, )
= -I I;II~
I x::
2V e 2)K(v l) [X::(Vl+eA"'1h[A])[QeV2)-Q'eV2) J)-K(Vl)] dv 1dv2 .
Applying the assumption (Kl), (K2) and (K4) we get
h
~2 [::~::~] [~:]] Ilq-q'lloo ~2 [~:]] Ilq-q'lloo A Q(A)llq-q'lloo. (3.3)
Putting q~ , x'(-) = q'(-) -
q'(. +(X2 - X2)/A)
we have similarly to (3 .3)
Taking into account that q' E IHI(.8, S) and, therefore , 2 llq'x ,x ' II < sIx -A x21 (X)
-
(3.5)
133
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTIONS
applying assumptions (Kl), (K2) and (K4) we obtain from (3.4) and (3.5)
where
II· Ih
is lLl-norm on lR 2 . Thus, we have
2V2[h* A~/2rl [1 + ~ CEllx - x'/h.
12 :::;
diam(D) + [S VI] diam(D)] Ilx
- x'iiI
(3.6)
Finally, we obtain from (3.2), (3.3) and (3.6)
fJ>..(T,r ') :::; VQ(A)llq - ql/loo + /CEllx - x'iiI.
(3.7)
For any 8> 0 let us denote by E~oo)(lHh(,8,S)), Ey)(D) and E~ji).)('I,,),A E A, the 8-entropy OfJH[l (,8, S) w.r.t uniform norm, 8-entropy oi D w.r.t. 11·111norm, and 8-entropy of 'I" w.r.t. intrinsic semi-metric p" respectively. Clearly, that \/8 > 0
E~l) (D) :::; 2ln (
dia;(D))
(3.8)
Note also that (3.7) implies Yu > 0 and YA E A
E~ji",)('I,,) :::; ECc::} (JH[(,8,S))
+ E~l
4Q("')
(D).
(3.9)
4C<
Therefore, we deduce from (3.8), Lemma A.l and (3.9) that \/u > 0 and YA E A
E~ii>. )('I,,) :::;C[4SQ(A]*U-~ +4ln(l /u) +2ln(4C diam(D)) E
Noting that IEt 2(T) = 1, Yr E 'I, and applying Lemma B.l with obtain from (3.10) YA E A
V' J-k
IE 7"s:~K t(r) :::; £(,8 , L 2 ) [ h[A] where £(,8 , L 2 )
+
(J"
(3.10)
= 1 we
~ + £,
= VC(8S) -k [1 - 1/,8] and £ = J2ln (4 diam(D)) + 2.
134
OLEG V. LEPSKI AND CHRISTOPHE F. POUET
Simple calculations show that In Cc; :=: In (1/ c) and (3.11) where a > 0 depends on (3 only. Thus, we have for all and VA E A
E
> 0 small enough
(3.12) Ll.b. Upper bound for IE SUPxE'D ~(x). Putting Vx, x' E V
we obtain similarly to (3.6) Vx, x' E V
p(x ,x /) :::; f.l-1J diam(V) llx - x'iiIThis together with 1E~2(x) = 1, Vx E V, (3.8) and Lemma B.1 gives (3.13) Note that VA E A in view of (3.12) and (3.11) we have
, 1/2(3} lP'O{AA,I} = lP' { sup 1~(r)1 ? 3£((3 , £2) ( AX) / h[AJ rE'!A ) 1/2(3} , = 2lP' { sup ~(r) ? 3£((3 , £2) ( AX / h[AJ rE'!;,
,
,
( ) 1/2(3 }
:::; 2lP' { sup ~(r) - IE sup ~(r) ? £((3, £2) AX / h[AJ rE'!A rE'!;, :::; 2lP' { sup rE'!A
~(r) -
Applying Lemma B.2 with
(J
IE sup €(r) ?
a«,£2)C- a} .
rE'! A
= 1 we get VA E A
and , therefore
L lP'o{ AA,I} :::; 2log2 (1/() exp { _~£2(fJ, £2)C- 2a} . AEA
(3.14)
HYP OTHESIS T ESTI NG UNDER COMPOSIT E FUNCT IONS
13 5
Using t he same arguments we have from (3 .13)
Jl!'o{B} ::;
2J1!'{SUP~ (X) - lEsup~(x) 2: V ln( I //1) } xE D x ED
::; 2.Jii. (3.15)
Taking together the bounds found in (3.14) and (3.15) we get from (3.1)
Jl!'o { A l U A2 U B}
--+
0, as
e --+ O.
(3 .16)
II. Let us find t he upper bound for t he second err or probabilit y, i.e. for
Jl!'g {A l n A2n B}.
sup gE IHl (C cp c (a ))
For any g E lHI(a, ..c) , r E 'I' and x E 'D let us denote by ,
1
Bg(T) = ,\h[,\J
/
K
Bg(x ) = : 2/ K
(tI
-Xl - '\"'Q((t2- X2)/'\ ) h-X2) h['\]
, -,\-
f(G (t ))dt - f(G (x )) ;
Cl ~ Xl ,t2 ~ X2 ) f (G(t ))dt - f (G(x )).
It is clear that
[%&r] -
Bg(r ) =
s,
Bg(x ) =
z, [ge (X)~KI I2]
f (C(x)) , (3.17)
- f (C(x)) .
Let M ~ M ({3 , L 2 ) be t he constant from Lemm a D.l. Put j = 1, J
Sjo(x) =
{gE lHI(a, ..c) :
Sj j (x) = {
g E
IC 1 ,o(x)! V !CO, l (X)! ::;
A' 1]1/'"::; lHI( a, ..c) : [ ~
+ 1,
M~/"' } ;
IC1,o(x)1V IC ,l (X)! ::; O
[A.] ~ l /"'} ,
where, AJ+l ~ M (L 2) "', C l ,o(x) and CO,l(X) ar e [8C/8t l ] (x ) and [8C/8 t2] (x) respe ctively. First, for given x . E 'D let us find the upper bound for sup g Eii o(x )
IBg(x) l,
sup
!Bg(r )l , r = (q,Aj,X ), j = I ,J + 1.
gEiij (x )
vVe denote by 3 1 , 3 2 , . . . , the functions uniformly bounded by 1 on 'D and by Gl , G2 , ... , t he absolute constants and wit hout loss of genera lity we will assume that (3 .18)
136
OLEC V. LEPSKI AND CHRISTOPHE F. P OUET
II.l.a. Upper boundforsuP gEYJ o(x) IBg(x )!. In view of R em ark 1.1 we have \:It E V
where
Since f E lHI 1 (r, L l ) we have 'Vt E V (t he summation with respect t o em pty set of indexes is supposed to be 0)
bJ
L jCl )(G(x)+Px(t- x )) [5 1(t)L21It- xlli3f
f( G(t )) = f(G(x)+ Px(t - x )) +
1=1
= f (G(x ) + Px(t - x ))
bJ
= f (G(x ))
+L
f (l )(G(x )) [Px(t - x )]l + L 154 (t) [Px (t - x)f
1= 1
>
wher e Xl = 0 if , :5 1 and Xl = 1 if , Tak ing into account that 'Vt E (1/ 2)/L, X2 + (1/ 2)/L] and 'Vg E 5)o(x ),
l.
[X l -
2
IPx(t - x )I :5 C2C;/L + X2 C3L2/L,
X2
(1/ 2)/L, X l
=
{
+ (1/ 2)/L]
X
[X2 -
0, if 1 < f3 :5 2; 1, if f3 > 2,
and using the assumptions (Kl), (K2) and (K3) with N ~ f3 2 we obtain from (3.19) : sup
!B g(x) l :5 Cl (a, £ ) [(C;/L )"
+ X2 /L2'Y + Xl/L13 + /L'Y I3] ,
(3.20)
gE.l'io (x )
where Cl (a, £ ) is the constant depe nding only on " f3, L l and L 2. P ut ting 5 = 2, /\ f3 /\ ,f3 we note t hat (3 .20) implies Vu E 2t sup
gEJ'J o(x)
!B g (x)!:5 3C 1 (a, £ ) [(C;/L )"
+ /L s ]
:5 6C l (a, £ )
(3.21)
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTIONS
ILl.b. Upper bound for SUP9Ej'Jj(X) 1139 (T) I, j = 1, J
137
+ 1.
Let us denote by qxO the implicit function being the solution of the equation G(t) = G(x). In view of Lemma D.1 VG E 5'j j(x) the function
qx(.) is uniquely defined on the interval t, and I J + 1 = [x2 -
=
[x2 - Aj , X2
+ Aj] , j = 1, J
1,X2 + 1] .
In view of Remark 1.1, taking into account that G(qx(t2) ,t2)
G(x), Vt2 E Ij we have "It = (tl, t2) E V such that t2 E I j
(3.22) Therefore, repeating all lines in (3.19) we obtain "It E I j 0 I j bJ
f(G(t))
=
f(G(x)) +
L
f(l )(G(x)) ['-Px(tl - qx(t2) ,t2) ]1
1=1
+X 1 C4 S 7 (t)L l [L2 V 1rltl
- qx(t 2)1 13
(3.23)
+LlSS (t)['-Px(tl - Qx(t2), t2)r + S6(t)L lL;ltr - qx(t2)1 , 13, Set (3.24) and let Tx ,j
= (qj,x,Aj,x). Note that
Here we used , in particular , the definition of the kernel K . Let us also remark that the kernel K is vanished on lR \ [- 1/2, 1/2] and, therefore, the integration w.r.t t2 in (3.25) is done on the domain of definition of qxO (see Lemma D.1 and the definition of Aj). Changing the variables in (3.25) u = (tl - Qx(t2)) /h [Aj ], v = (t2 X2) /Aj , noting that the functions Zk ,x( ' ) in (3.22) are independent of u and using the assumptions (Kl), (K2) and (K3) with N ;::: (32, we obtain from (3.23) and (3.25) similarly to (3.20)
138
OLEC V. LEPSKI AND CHRISTOPHE F. PO UET
sup
IBg (Tj,x)1
gEnj(x)
~ 02(a,.£){ ([Ajf /l
(3.26)
~ 302(a,.£) {([Aj]l/l
and,t herefore , "It E 'D such that It 1
-
qx(t2)1 ~ (1/ 2)h[Aj] and t2 E [X2 -
Aj , X2+ Aj]
Now let us return to (3.26). First, we note that
Next,
if {3 > 2, - 1; if {3 ~ 2, - 1, where
!!Z
() - { u -
[1 _K,({3 - x )({3 - 1)]
,
~
{3 + 1
[1-
2{3({3
if {3 > 2, - 1;
+ 1)
K({3 -X)({3- , )] 2, {3
if (3
~
2, - 1.
and y = y(a) > 0 is the given number. 1. Case , ~ 1 , 1<{3 ~ 2. {3 - 1
z( a) = (3 [ 1 - 2{3({3 + 1)
]
=
2{32 + (3 + 1 2({3 + 1)
> 1.
(3.28)
139
HYPOTHESIS TESTING UNDER COMP OSIT E F UNCTIONS
2. Case "I :::; 1, {3 > 2.
z(a) = 2 [1 _ ({3 - a )({3 - 1)] > 2 _ /3 - 1 > 1. /3+ 1 2/3({3 +1 ) Note also that in t hese cases necessaril y /3 > 2"1 - 1. 3. Case 1 < "I < /3 :::; 2, /3 > 2"1 - 1. z(a)
f3 -1 ] =;;f3 [ 1- 2f3(f3+1) ~
1]=
2f3 [ f3 f3+1 1- 2f3(f3+ 1)
2f32 + f3 + (f3+ 1)2
1>
l.
4. Case 1 < "I < /3 ::; 2, /3 :::; 2"1 - 1.
/32 ~/3"1 ] ?
z(a) = / : 1 [1 -
2~/3-: 1)] > 1.
/ : 1 [1 -
5. Case "I> 1, (3 > 2, /3 > 2"1 - 1.
=
z(a)
[2 1\ ~]
(/3~~~~ D 1)] ? /32: 1 [1- 2~~\)]
[1-
/32 + 3/3
= ({3 + 1)2 > 1. 6. Case "I > 1, /3 > 2, /3 :::; 2"1-1.
z(a) =
2+3{3 [1 - ({3 -a)({3 - 'Y) ] > ~ [~_ P.-] > /3 > 1. /3+ 1 2"1/3 - /3+ 1 2 2"1 - ({3+1)2
~
T hus, we have from (3.26), (3.27) and (3.28) t hat sup
sup
!B (7j ,x)l ::; 6C (a, £ )epe(a).
(3.29)
2
g
j =I,J+ I gESij(x )
°
11.2. Let C > be t he constant the choice of which will be done lat er. Since 9 E lHI(Cepe(a)) there exists x = x(g) E'D such that (3.30) Let us introduce the following notations:
) = ) (g) = {j = 0, J + 1: q(v) = q3,x(v ) = [qx(X2
9 E lHI (Cepe( a))
+ VA3 ) - Xl]/
n Sjj (x )} ;
(A3(;
T = 73,x'
where qj,x is defined in (3.24) . Note also that ) is correctly defined becau se "Ix E 'D J+ I
lHI(a, £ ) =
USj j (x ),
j =O
Sj j (x ) nSjj/(x) = 0, j
=1=
1',
j ,j'
= 0, J + 1.
140
OLEG V. LEPSKI AND CHRISTOPHE F. POUET
In view of Lemma D.1 (1) and (3) QEJH[I(,8,S) , on [-1 ,1] .
(3.31)
Put
and note that due to (3.31)
nA"j ,i
J+l
Ai =
t,
S;; ...
13 S;; B.
j=1
if if
3=f: 0, 3= O.
(3.32)
Without loss of generality we will assume that (3.18) is true for x and, therefore it is sufficient to find the upper bounds for JP9 {AI} and JP 9 {B} . II.2.a. Upper bound for JPg{AI}. First let us note that V).. > 0
Next ,
9",1(1')
=
(3.33)
=
[f(G(x)) +.8g (1')] +s€(1').
Thus, we have from (3.29), (3.11) , (3.33) and the definition of l'
Al =
{
<;;; {
(h[>-] />-x) kl§",l(f) 1:S 3C£((3,£2)}
EII~llf:) [I f (G(f)) 1- IB
g(
<;;;
{It (f) [ ~ (>-X /h[>-]) $
<;;;
{ ltcr)1 ~ l~~ (AX /h[A]) k
<;;;
{ It(f)1 ~ E- a [(0/2 -
f) I] -E (h[>-J/>-x)
kIt(
f) I:s 3C£((3, £2) }
[(0/2 - 602(a,,C))/ IIKI12 - 3£((3, £2)] } [(0/2 - 602(a, 'c)) /IIKI12 - 3£((3, £2)] }
602(a, ,C)) /IIKI12 - 3£((3,£2)] } ,
141
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTION S
where 0 is chosen such that 0 > 1202(a, £) + 31IKI 12.c(.8 , L 2) and a is defined in (3.11). Taking into account that is the standard normal random variable and denoting by its distribution function we obtain vs such that 1= 0
ten
IP'9 {AI}
3
:::; (e- a [ ( 0 /2 - 6C2(a, £)) /IIK II 2 -
II.2.b. Upper bound for First, let us note that
3e.c (,B, L 2)]) .
(3.34)
IP'g{ B}. In (1/ p,) ~ Z (a) In (l/e),
p, = e Jln (1/e)
(3.35)
where /3+ 1
2~,8+,8+1
Z( a) = {
,8+1
if
,B> 2, - 1,
if
,B:::; 2, - 1.
Next,
ge(X)
= 11:1 12
[f(G (X)) + Bg(x)] +e~(x) .
Similarly to previous calc ulation we get from (3.21) and (3.35)
B = { ge(X) <
~ { 11:
[3Jln (1/p,) + 2] e }
112 [If (G(x))I-
IBg(x)l] -
el~(x)1 :::; [3Jln (1 /p, ) + 2] e }
{
~ :Y:~I%~; [If(G(X))1-I Bg (x)l] - el~(x)1 :::; [3Jln (1 /p,) + 2]e } ~ { 1~(x) l 2: J ln (1/e ) [(0 /2-601 (a, £)) / IIKI12-3 JZ (a)] } , where 0 is chosen such that 0 > 120 1(a, £) + 31IKI12JZ (a). Taking int o acco unt that ~(x) is t he standar d nor mal random variable we obtain Yg such t hat = 0
3
IP' g{ B} :::; ( J ln (l/e) [( 0 /2 - 601(a, £ )) /II KI12 - 3JZ (a)J) .
(3.36)
It remains to note that upper bounds in (3.34) and (3.36) ar e independent of 9 and therefore, we finally obtain from (3.32) lim sup
sup
e ~O
9E IlI( C 'P, (a l )
IP'g{A 1 n A2 n B} =
0,
142
OLEC V. LEPSKI AND CHRISTOPHE F. POUET
3.2. Proof of Theorem 2. There are two different constructions for the lower bound. The first one is used in the case 13 > 2"( - 1 and it is based on a rather sophisticated random walk. As to the second construction, used in the case 13 :S 2"( - 1, it is absolutely standard. Putting f(t) = t the problem is reduced to the testing under alternative described by the function G and, after that, one can use the result from [11] . Thus, only the case 13 > 2"( - 1 will be considered. In the proof of the theorem C 1 , C2 , . . • denote absolute constants. Each time they appear, we explain the link between these constants and the constants a and 'c. I General remarks As usual (see, for example, [12]), we will construct a lower bound for inf},. R(fS., C
where JPl 1r (A) =
Ie JPl
fe
1
2: 1 - "2!JPlo - JPl1r 12 ,
(A) dat (8) and !JPlo- JPl1r
I; =
!Epo
(~~
_
1)
2.
The proof is done in the following way. First, the parametric subset in the alternative is defined. Next , we do some calculation on the L 2-distance. Finally, the probability tt is chosen such that it leads to the expected rate of testing up to a logarithmic factor. 1.1 Parametric subset of alternatives. Let us choose a function F : lR. - + lR.+ such that 1. F(~»O.
2. F(1 /2+u)=F(1/2-u). F(u) = 0, u't [0, 1]. 4. FE1Hh("(,1) .
3.
Let us also choose a function r : lR. -+ lR. such that (i) . The function r belongs to the Holder space lHI 1 (13, L 3 ) with L 3 a constant compatible with property (iii). (ii) . The function t7 is non-decreasing. (iii) . t7 (0) = 0, r (1) = 1 and Vl < m = Lj3J : t7(I) (0) = t7(I) (1) = O. Introduce the parameters ,x, hand T such that
1h L L"Y "Y F (1/ 2) = C 3
C{Je(a)
(In (1/c)2+1J) 2"113+13+1 2 -y{3
h=,X(3,
Let~=(6" .. ,~T)E3T={0, l}T . Define k
s, (0 =
L ~i , i=l
So = 0.
T=[l /'x].
HYP OTH ESIS T ESTING UNDER COMPOSITE FUNCTIONS
Set D.. k = [- 1 + (k - 1»., -1
+ k>.] , k = 1, T
and define
14 3
v» E [-
1, 0]
~[Sk - l + ~k t ( y- (k-1 x ».)] lI6.k (Y)·
g~ (y) = ~
Finally, we put g~ (y ) = [g ~(y) + g ~ (-y ) ] /2 , y E [- 1, 1]. The par ametric subset of alte rnatives is defined as follows.
r~ (x) = L 1h"l F Clearly that
r~ =
( Xl -
>.:g~ (X2) ), ~ E 2T.
f (G), where
L 1 h"lF (Lh3 +2~) . f( u ) = £"I U
3
f
By const ruction G E lHI 2( ,B, L 2 ) ,
E
lHI 1('y , L 1 ) . and, moreover ,
1.2 Calculation of the B ayesian likelihood ratio. Let 1r be a probability measu re on 2T which will be chosen later. According t o Gir sanov formula , t he Bayesian likelihood rat io unde r lP'o is
t; = z, ( ex p
( ~J r~ (x)
dW (x ) -
J
2~2 r~ (x) 2 dX) )
.
Put
where the absolute constant C 1 depends only on L 1 , L 2 and L 3. Let €be an independent version of~ . Then,
lEo (l;)=lE>T x>TlEo exp(~J[t~ (X)+tf(X)] dW(x) - 2~2J[t~ (X) +t~(x)]
=lE", x >T exp{c-2/t~ ~ lE>T X>TexP { [172/c2] ;, \ ,ti)} , L.. u\Sk(
dX)
_\}
O - Sk (O $1
k= l
~ tA;E>TX>T ( 19:1<.L.. . =0
where A" = exp (6 2/0"2)
I lsk1-1W....Skl-1(f )19 · . . I ISk' _lW-SkS-1(f)19) '
< k.$1'
- 1.
144
OLEC V. LEPSKI AND CHRISTOPHE F. POUET
Note that up to this point, all formulas hold for any prior distribution on 2T. 1.3 Prior distribution and final calculation. In order to get the rate of testing, we choose the prior distribution 1f as the distribution of the random walk defined in Lemma C.l. Putting
1f
we have
L
lE"x" (
lllskl_ICO-skl_l(e)!$l'"
l$k , < ...
L
L
lE"x" (
l:5kl < .. .
L
llISkS_'(~)-Sk,-1(e)19)
L s,xllls k1_1(e)-xlI9 " ' ][ISkS- l (O- XS!9 )
Xl ,""X S
L
lE"x" (
1:5k 1 < ...
L s ,x][ lskl _l (e)- xd 9 ' "
X l, ' '' 'X s
][ISks_ 1- l (t) - XS- 119
- k s - 1 + l)l+ry k, - k s - l
3c1n (k s
3cln (k s -k 1 +1)1+'7 k s-k 1
3cln(kl)1+7 3cln(k 2 - k1+1)1-h7 l$k , < .. .
k 1-1
k2-kl
S (3cln(T)2+'7)' S (cRwln(T))s (2+'7 ). Thus, the following upper bound holds T
lEo (l;) ~
LA: (cRwInT)(2+
7J )s
s=o
<
- 1- (cRwInT)
2+
1 '7J(exp[O' 2/c: 2] -1)
.
(3.37)
The right-hand side of (3.37) is of order 1 + (CRW In T) (2+'7J ) 0'2 / c: 2 . Recall that
h "Y =
C
L~
3
( / [I
L 1P (1/2) c:
n
1/] 1+'7J/ 2) e
2~
2..ri3+13+1
.
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTIONS
145
Therefore, 1
l-(cRwlnT)
(2+ ) T)
(exp[0'2/c2] -1)
-
1. (3.38)
Thus,
liminfiI}fR(.6. ,Ccpe) e-+O
fl.
~ 1- -2
1
C
2JI3+/+1
1- C
-y
C3
2JI3+/+1 -y
C3
,
(3.39)
where the absolute constant C 3 depends only on L 1 , L 2 , L 3 , (3, I , F (1/2) and CRW. It remains to note that the left-hand side of (3.39) goes to 1 if C -> O. 0
APPENDIX
A. Metric entropy. Let ('1', p) be metric space and let T c '1' be a compact set. For any 5 > 0 we denote by Ef(T) 5-entropy of T W.r.t p. The result below can be found in [10]. LEMMA A.1. If P = 11 ·1/00 (uniform norm on [-1 ,1]) then V5 > 0
C is the constant independent on 0 and L. The upper bound given in the lemma is sharp. We do not present the lower bound because it is not used in the paper.
B. Gaussian random functions. In this section we present some results concerning the large deviation probability of the extrema of gaussian random function . These results can be found, for example, in [1] and [17]. LEMMA B.1. Let T)(t), t E T be a centered gaussian random function and p(s, t) = jEiTJ(t) - T)(s)1 2 its intrinsic semi-metrics. Then EsupT)(t)S;4V2 tE'[
r VE[;(T)du,
l;
where 0' = VSUPtE'[ Var(T)(t)) . Moreover, if Dudley integral is finite then T)(-) is a.s . bounded on T . LEMMA B .2. Let ry(t), t E T be a centered gaussian random function a.s. bounded on T. Then Vx > 0
lP'{SUPT)(t) > ESUP17(t)+x} S; exp{ -x 2/ 20'2} . tE'[
tE'[
146
OLEG V. LEPSKI AND CHRISTOPHE F. POUET
C. Random walk. T he following lemma deals with the existence of a random walk whic h is more unpredictable than the simple random walk . It is a dir ect consequence from [5] (T heorem 1.4). Let ( = {(j , j E N} denote a sequence of random variables taking values in {O, 1} and Let §n (() denote the partial sum =1 ( j . LEMMA C.l. For any 1] > 0, there exists a probability tt on [{O}, {l} ]OO and a constant c > 0 independent of n such that for any k > 0 and any x ~0 :
I:7
D . Implicit functions. In t his sect ion we establish some smoothness properties of implicit functi ons. In spite of the fact t hat the main argume nt here is Implicit Funct ion Theorem [15], we were unable to find req uired results in t he existing lit erature and we give below t heir direct proof. Let G E JHlz (,8, L ) on 'Do ~ IR 2 , Xo E 'Do be the fixed point , and let us consider the equation (D .1)
G(t) = G (xo) Without loss of generality we will suppose t hat
LEMMA D .l. There exist the positive constant M = M (,8, L ) and 5 = 5 (,8, L ) such that VG E 1HI2 (,8, L ) the implicit function qO satisfying (D.1) 1. exists and is uniquely defined on the set [YO - Ma t<; ,Yo + Mat<;] ;
2.
• takes its values in [xo - M (4a)t<; , Xo
+M
(4a) t<;] ;
• Iq' (y )l :::; tc; Vy E [Yo -Mat<;,Yo + Mat<;] , where the constant K 1 depend only on ,8; 3. belongs to lHI l (,8, 5 a""-I3 ) on [yO- M at<; ,Yo + Mat<; ] , where, remind, '" = [1/(,8 - 1)] V 1. Proof. We will use the following not at ions: K t , K 2 , . .. denote absolute constants depending only on ,8, 8 t O, 8 2 0 ,.. . de note bounded functions and F' denotes the derivative of the function F . Also wit ho ut loss of generality we put (x o, yo) = (0, 0) and ass ume that Gl ,o(O, 0) = a. Since Gl ,o E lHI 2 (,8 - 1, L ) the following inequality holds: Gl ,o(x , y) ~ a/2 ,
Vjx l, Iyl :::; (a/ 2Lm)'\ m = ,8 - a .
This follows from Gl ,o(x , y) ~ G t,o(O, 0) - Lm llx _ Yllt / t<; .
(D.2)
HYPOTHESIS TESTING UNDER COMPOSITE FUNCTIONS
147
Put 5 = (a/4{3Lmt. We have
G(5,0 ) = G (0 ,0 ) +G l ,o ( 0 ,0 ) 5 + S1 ( 5) 513f\ 2~ a5_L513 f\ 2 ~ al<+!( I/ LtA{3), where p({3 ) is constant depending on {3 only. In the same way as above, we have
As the function G E IHb ({3, L ), the following exp ansion at point (5, 0) holds
This implies
G(5,y) ~ p({3)al<+l(1 /Lt G( -5, y ) ::; -p({3)al<+l (l / Lt Thus, G(5,y) > 0, G(-5 ,y ) < 0,
'lilyl ::;
alyl- Lmlyl13 f\ 2; + alyl + Lmlyl13f\ 2 . (a/16{3mLt.
Let us fix y E [-(a/ 16{3m Lt, (a/16{3m Lt ] and consid er the function R(-) = G(·,Y). We have
R (5) > 0, R (- 5) < 0, R' (x ) ~a/ 2,
'lixE [- (a/4{3mL)1/13-1, (a/ 4{3mLt].
Then , there exists a unique point x such that R(x) = 0. The implici t function q(-) is defined by
q(y) = x(y). Clearly it exists on the set [ - (a/16{3m L)I< , (aj I6{3m L)I<] and t akes its value in [ - (aj 4{3mL)I< , (aj 4{3mL )1<]. The implicit functi on theorem [15] asserts that the fun ct ion q(.) is m tim es differentiable and we can calculate its derivatives using formula G(q(y ), y) == 0. In parti cular ,
q'(y) = _ GO,l (q(y), y) . G1 ,o(q(y) , y) and in view of (D.2) Iq' (y )1 ::; K 1, 'liy E [ - (aj I6{3m L) I<,( aj 16{3m L)I<] . where K 1 depends only on (3. Thus, th e assertions (1) and (2) of the lemma are proved . Let us now study the regulari ty of the implicit funct ion q(.). Put ting Gk r = Xl X k, r = 1, m , and different iating the identity G(q(y ), y ) = 2 we first prove that for n = 1, .. . , m : I
ez:
°
q(n)(y)G1,o(q(y) , y) + P n (q ( l ) (y), ... ,q(n- l )(y) , y)+ GO,n(q(y ), y) = 0, (D.3)
148
OLEG V. LEPSKI AND CHRISTOPHE F. POUET
where Pn (Zl , . . . , zn- l , y) is a polynomial such that each term is ofthe form n-l
n-l
i=l
i=l
Gk,r(q(y), y)
IT z~ ;, I: ili :::; n,
k+r:::; n.
The relation is true for n = 1 and n = 2. Assume that the relation is true up to n . Then the derivative of each term in the polynomial Pn is also a polynomial such that each term is of the form
Gk,r(q(y), y)
n
n
i=l
i=l
IT z~; , I: u. s n + 1,
k+r:::;n+1.
Thus, we have
where Pn+1 (Zl , ... ,Zn-l ,zn,y) is a polynomial such that each term is of the form
Gk,r(q(y), y)
n
n
i=l
i=l
IT z~; , 2:= il, :::; n,
k + r :::; n .
Therefore the relation holds for any n = 1, . . . , m. In the same way as above we prove that Iq(n)(y)1 :::; K (L /at- 1 . Indeed , this relation is true for n = 1 and n = 2. Let us assume now that the relation is true up to n. Then, each term in the polynomial Pn + 1 is bounded
Gk,r(q(y), y)
n
n
i=l
i=l
IT q(i)(y)l; :::; K 4L IT (L /a)(i-l )l; :::; K 4L (L /at .
It leads to Iq (n+l )(y)! :::; K s (L /at . Let us now study the regularity of points y and z and proceed as follows.
q(m).
We use equality for n = m at
m-1
IGk,r(q(y) ,y)
II q(i)(vd; -
m- 1
Gk,r(q(y),y)
m-1
::; L
II (Lla)(i- 1)1; (L iar jy -
II q(i)(Wi)l; I i=l
i=l
z lK 6 (L la) (r-1 )(lr- 1) ::; L (L la)m-1
Iy -
z];
i=1 i¥r
n
,.",...1
!Gk,r(q(y), y)
i= l
,.",...1
q(i)(z)I ;-Gk,r(q(Z) ,z)
II q(i)(z)I ;I::; K 7Lly- z i=l
IGo,m(q(y) , y) - GO,m(q(Z) ,
z)1 ::; KsLly -
z l,B-m.
l
(L la )""""1 i
HYPOTHES IS TESTING UNDER COMPOSITE FUNCTIONS
149
Therefore summing all the differences and dividing by a we get
where Kg is a constant depending only on {3. These results entail t hat the implicit function is in H; ({3, Kg (L/a)m). 0 The authors are grateful to Fabienne Castell for suggesting to use the resu lts from [5] .
REFERENCES [lJ R .J . ADLER, An Introduction to Continuity, Extrema and Related Topics for Gen[2J [3J
[4] [5J
[6] [7J [8J [9] [10J
[11]
[12J [13] [14J
[15]
[16J [17] [18] [19]
eral Gaussian Processes, IMS Lecture, Notes-Mongraph Series, 12 Institute of Mathematical Statistics, Hayward CA, 1990. L. BROWN AND M . Low, Asymptotic equivalence of nonparametric regression and white noise, Ann. Statist., 2 4 (1996), pp. 2384-2398. H . CHEN, Estimation of a projection-pursuit type regression model, Ann. Statist., 19 (199 1), pp. 142-157. G.K. GOLUBEV, Asymptotically minimax estimation of a regression function in an additive model, Problems Infor m . Transmission, 28 (1992), pp. 101-112. a. HAGGSTROM AND E. MOSSEL, Nearest-neighbor walks with low predictability profile and percolation in 2 + E dimensions, Ann. Probab., 26 (1998), pp. 1212-1231. P . HALL, On projection-pursuit regression, An n. Statist., 1 7 (1989), pp . 573-588. M. HRISTACHE, A. JUDITSKY , AND V . SPOKOINY, Direct estimation of the index coefficient in a single-index model, Ann. Statist. , 29 (2001), pp. 595-623. M . HRISTACHE, A. JUDITSKY, J. POLZEHL, AND V. SPOKOINY, Structure adaptive approach for dimension reduction, Ann. Statist., 29 (2001), pp . 1537- 1566 . P. HUBER, Projection pursuit. With discussion., Ann. Statist., 1 3 (1985), pp . 435-525 . LA . IBRAGIMOV AND R.Z. HASMINSKII, Statistical Estimation. Asymptotic Theory, Springer-Verlag, 1981 (Originally published in Russian in 1979) . Yu.1. INGSTER Asymptotically minimax testing of nonparametric hypotheses, Probability theory and mathematical statistics, Vol. I (Vilnius, 1985) , pp. 553-574, VNU Sci. P ress, Utrecht, 1987 . Yu .1. INGSTER AND LA . SUSLlNA, Nonparametric Goodness-of-Fit Testing Under Gaussian Models, Lecture Notes in Statistics, Springer-Verlag, New York, 2003. A. IOUDITSKI, a.v. LEPSKI, AND A .B. TSYBAKOV, Statistical estimation of com posite functions, Manuscript, 2006. A .N . KOLMOGOROV, On representation of continuous functions of several variables as superposition of continuous functions of one variable and addition, Dokl. Akad. Nauk SSSR, 11 4 (1957) , pp. 369-373. S . LANG, Analysis II, Addison-Wesley, Reading, 1969. O .V. LEPSKI AND A .B. TSYBAKOV, Asymptotically exact nonparametric hypothesis testing in sup-norm and at a fixed point, Probab. Theory and Related Fields, 117 (2000), pp . 17-48. M. LIFSHITS, Gaussian Random Functions, Kluwer Academic Publishers, Dordrecht , 1995 . M . REISS , Asymptotic equiva lence for nonparametric regression with multivariate and random design (2006), Submitted. D .A . SPRECHER, An improvement in the superposition theorem of Kolmogorov,. J . Math. Anal. Appl. , 38 (1972), p p. 208-213.
150
OLEG V. LEPSKI AND CHRISTOPHE F. PO UET
[20] C. J . STONE, Optimal global rates of convergence for nonparametric regress ion, Ann. Statist., 10 (1982) , pp. 1040--1053. [21] C .J . STONE, Additive regression and other nonparametric models, Ann. Statist., 13 (1985) , pp. 689-705.
PART
III:
STOCHASTIC
PARTIAL DIFFERENTIAL EQUATIONS
ON PARABOLIC PDES AND SPDES IN SOBOLEV SPACES W~ WITHOUT AND WITH WEIGHTS NICOLAI V. KRYLOV " Abstract. 'We present a "st reamlined" t heory of solvabili ty of parabolic P DEs and SPDEs in half sp aces in Sobolev spaces wit h weights . T he approach is based on inte rio r estimates for equatio ns in the whole space and is easier t han and quite di fferent from t he standard one . Key words. St och ast ic partial differential eq uat ions , Sobolev spaces wit h weights, interi or estimates . AMS(MOS) subject classifications. Primary 60H1 5, 35K 20.
1. Introduction. Let jRd be a d-dimensional Euclidean sp ace consisting of points x = (xl, ..., x d ) . We denote
8 x'
D , = -8"
Dij =
u.o;
Introduce D u as the gra dient of a function u wit h resp ect to x and D 2u as its Hessian matrix. In domains G C jRd we are considering t he following par abolic equation (1.1) where T E (0, 00] and L tu = ( 1 /2)a~j Dij u
+ b~DiU + CtU,
an d t he stochastic par t ial different ial equat ion
where r is a stopping t ime
Here and everywhere below we are using t he summation convention. The precise framework in which th ese equat ions ar e treated is given late r in the ar t icle. We present t hree sets of results : Set 1 . Int erior Wi-esti mates for solut ions of (1.1) (Sect ion 3) and (1.2) (Sect ion 2) when G = jRd . "127 Vi nce nt Hall, University of Minnesota, Min nea polis, MN, 55455 . The work was partially suppo rted by the NSF Grants DMS-0140405 and DMS-0653121. 151
152
NICOLAI V. KRYLOV
Set 2 . Boundar y estimates for solutions of (1.1) (Section 5) and (1.2) (Sect ion 4) in spaces like with weights when G = IRt, where
Wi
IRt
= {x = (x l , x' ) E IRd
:
x l> O,x' E IR d -
I
}.
Set 3 . Exist ence and uniqueness t heorems for (1.1) (Sect ion 7) and (1.2) (Section 8) in spaces like with weights when G = IRt. In case of equatio n (1.1) we take p E (1,00) and in case of (1.2) we restrict p to [2, 00).
Wi
This is, of course , a quite nar row subclass of t he class of pro blems treated in t he modern t heory of SP DEs and we do not try to overview t his very wide ly develop ed theory referring the reader t o t he collection of articles [4]' to [6]' [16], [20]' and many references t herei n. Our main resul ts collected in Set 3 are known from [7] and [9] (apart from Theorem 8.1) even for spaces of Bessel pot enti als HJ, which coincide with W ;r if 'Y is a nonnegative int eger . However , t he proofs here are qui t e differe nt and mu ch more eleme ntary albeit valid only for 'Y = 2 (see however Rem ark 2.1). Conce rn ing earlier res ults on stochastic parti al differ enti al equations in domains t he reader can consult [2]' [1], [3]' [5]' [17]' [18]' an d [19]. Notice that t he ap proach in [2]' [1], [3], and [5] is based on semigroup theory and req uires , sometimes imp licitl y, certain com patibility of the data on the bou ndary. We work in Sobolev spaces wit h weights allowi ng the derivatives of solutions to blow up near t he boundary, which enables us to avoid any com patibility conditions. In a sense the result of Section 3 also belongs to "well-known results" . However , we could not find it in the form we need and therefore give a complet e proof. To the best of our knowledge, all other results of t he paper are new. \Ve derive the results in Set 3 from t hose in Set 2, whic h in t urn are derived from t he resu lts in Set 1. The approach in [7]' [9], [17]' [18]' and [19] requires first developing t he full t heory of solvab ility in 5j;,e(T)-s paces (see t he definition of 5j;,e(r) in Sect ion 4) for equations in IRt wit h coefficients independent of x . In case of (1.2) this leads t o t he fact t hat from t he start one needs t he followin g restriction on t he par amet er e calibrat ing the rat e wit h which t he deri vati ves of solutions can blow up near GIRt:
d - 1 + P [1 -
1_
p( l - <5)
_] < o < d - 1 + p ,
(1.3)
+ <5
where J E (0, 1] is t he constant of "parabolicity" of (1.2) in t he direction of x l axis (see condition (8.6) below) . By the way, observe that (1.3) is obviously satisfied if
d - 2+p
~
e < d - 1 + p.
Also if J = 1, t hen (1.3) becomes a condition necessary and sufficient for solvability theory of equations like (1.1) in domains: d - 1 < < d - 1 + p.
e
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
153
e
Here from the very beginning we treat variable coefficients, allowing to be any number, and introduce condition (1.3) in the very end. Actually, apart from Section 8 and a few auxiliary results in Section 6 there is no restriction on e whatsoever. We will see that (1.3) is needed for only one purpose: to estimate the zeroth order norm of solutions of equations with variable coefficients. Therefore, if for a particular equation and particular T one can prove such an estimate for 0 beyond the range in (1.3), then automatically, the solvability theory becomes available for this equation on (0,T) (see Theorem 8.1). We will present some examples when it really happens in a subsequent article. There we will show an example in which for a (beyond the range in (1.3)) for the coefficients frozen at some points Xo the equation is not solvable in Sj;,O(T), but the original equation with variable coefficients does admit a unique solution in Sj;,O(T). One of motivations for writing this paper was to prepare necessary tools for extending to the case of variable coefficients the pointwise boundary Holder regularity result from [13] proved there for equations with constant coefficients. For that purpose we prove Theorem 4.1 and say more about it in Remark 4.3. One of advantages of our approach is seen from the proof of the uniqueness of solutions of SPDEs on random time intervals [0, T]. This issue was not properly addressed in quite a few articles. In [8J uniqueness for random time intervals was obtained by proving it first for deterministic times and then using a method of continuation of arbitrary solution on [0, TJ beyond T. We do not use this roundabout here. We have outlined above the contents of all sections in the article apart from Section 6 in which we collect some results needed to justify the way we prove the estimates of the zeroth norm of solutions.
e
2. Interior estimates for SPDEs in the whole space. Let (D,:F, P) be a complete probability space with a given filtration (:Ft , t ~ 0) of a-fields :Ft C :F complete with respect to :F,P. Let do ~ 1 be an integer, possibly taking the value 00 and let w~, defined for finite integers 1 ~ k ~ do, be independent one-dimensional processes each of which is a Wiener process with respect to the filtration {:Ft , t ~ O}. Let P be the predictable a-field on D x lR+, where jR+ = (0,00). Introduce £2 as the Hilbert space of sequences y = (yk, 1 ~ k ~ do , k < 00) of real numbers with norm defined by do
lyl~2
=
L
ly k l2.
k=l
Let d ~ 1 be an integer and assume that for each x E jRd and i, j = 1, ..., d we are given a~j (x), b~(x), Ct(x), which are real-valued processes defined for t > O. We also assume that for each x E jRd and i = 1, ... ,d we are given £2-valued processes a;(x) = (a;k(x) , k = 1,2, ...) and lIf(x) = (lI~(x), k = 1,2, ...) defined for t > O.
154
NICOLAI V. KRYLOV
A SSUMPT ION 2 .1. (i) The functions a,b , c, a , and v are P x B (JRd)m easurable; (ii) Th e functions a i k and v k are continuously differentiable in x; (iii) For som e constants 80,81 , J E (0, 1] for all values of the indices and arguments we have a i j = a j i and for all unit ); E JR d ,
(2.1) where
This assumption is supposed to be satisfied throughout the st ochast ic part of t he ar ticle and in this sect ion we are considering equat ion (1.2). Equation (1.2) is understood in the sense introduced in [11] or [10] and the solutions are looked for in the space 'H.~,o (r) with f E IL p (r ) and 9 = (gk , k = 1,2 , ...) E lHI~ (r). We remind bri efly the definition of these spaces . Everywhere in the stochastic part of the article unless spe cifically ind icat ed otherwise we t ake p ~ 2.
For a sto pping t ime r and integer n W; (JRd) and we set
W; =
(O , rn = {(w,t ):
= 0,1 ,2 we take Sobolev sp aces
°< t::; r (w),t < oo},
In an obvious way we int rod uce t he norms in t hese spa ces and extend these definiti ons for £2- valued funct ions su ch as g = (gk, k = 1, 2, ...). If for any (w, t ) E (0, r n we are given a generalized functi on u = Ut (w) = Ut on JRd such t hat U E n O
(Ut,¢) = ( (fs , ¢ )ds +
Jo
do
t
L ( (g;,¢) dw: k=l
Jo
(2.2)
holds for all finit e t ::; r with pr obability 1. We set
If (2.2) holds we write
(2.3)
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
155
and this explains the sense in which (1.2) is understood. If G is a domain in ]Rd and (2.2) holds for any 1> E C!)(G), then we say that (2.3) holds in G. To describe the first main result of this section we take an integer d 1 E [1, d] and for x E ]Rd introduce X,
= (1 X ,
... ,xd 1 )
,
Bk =
{x
E]Rd :
Ix'i < R}.
THEOREM 2.1. Take some s , R E (0, 00) and a stopping time rand assume the following: (i) There exists a constant K E (0,00) such that for x E Bk and y E ]Rd, such that Ix - yl ::; eR, and all values of the indices and other arguments
RIDO";(y)le2
+ Rlb~(y)1 + RIVt(y)le2 + R 2!ct(y)1 + R2IDVt(y)le2 ::; K , 10"; (x) - 0"~(y) le2 + la~j(x) - a~j(y)1 ::; 130,
where 130 = 13o(d,P,OO ,Ol) E (0,1] is a (small) constant from Lemma 2.1 below. (ii) We are given a function U E 'H.;,o(r) which is a solution of (1.2) with some f E lLp(r) and g = (gk, k = 1,2, ...) E JH[~(r). Finally, assume that Ut(x) = 0 if x rf. Bk. Then there exist constants N = N(d ,p,oo,od and N* = N*(d,p ,oO ,Ol,C,K).such that
::; N(llf lllLp(r)
IID 2uIILp(r) + R-11IDuIILp(r ) 1 2 + IID gII Lp(r ) + R- IlgIILp(r)) + N* R- lIuliLp(r) '
(2.4)
To prove this theorem we need a lemma. LEMMA 2.1. Let r be a stopping time and assume that we are given an U E 'H~ ,O(T) which satisfies (1.2) with some f E lLp(r) and g = (gk, k = 1,2, ...) E JH[~(r). Assume that for a domain G c]Rd (perhaps, G =]Rd) we haveut(x) = 0 if x rf. G. Also assume that there are constantsc ,K E (0,00) such that for all x E G and y E ]Rd, such that Ix - yl ::; c, and all values of other arguments we have
IDO";(y )le2 + Ib~(y)1 + IVt(y)le2 + ICt(y)1 + IDvt(y) le2 ::; K , 10"; (x) - 0";(y)!e 2 + la~j (x) - a~j (y)1 ::; 130 , where 130 = 130 (d, p, 0o, 81) E (0, 1] is a (small) constant an estimate from below for which can be obtained from the proof. Then there exist constants N such that
= N(d ,p,oO,Ol),
N*
= N*(d,p,oO,Ol, €,K)
156
NICOLAI V. KRYLOV
P roof. Take a nonnegative ( E CO' ( jRd) with support in t he ball of rad ius f: centered at the origin and such that
Introdu ce (Y(x ) = ( (x - y ) and observe that with N = N (d, p)
ID2Ut (x )IP = { I( Y(x )D2Ut(x )IP dy :s; N { ID 2((Y(x)u(x) )IP dy
l ad
+N
l ad
(2.6)
{ ID(Y(x)IPI Dut(x)IP dy + N { ID2(Y(x )IP lut(x)IP dy.
l ad
lad
For each y E jRd the function (Y Ut satisfies
d((YUt) = (L¥((Y ut)
+ in dt + (A¥k((y Ut) + g¥k) dw~,
where
L¥ = ( 1/2) a~j (y) Dij , i f = ( Yi t + ( YLtUt - L¥(( Yut),
A¥k = aik(y )D i ,
g¥k = (Yg~
+ ( YA~ut -
A¥k(( yut ).
By the results for equati ons with coefficients ind ependent of x (see [10] or [11]), for each y
IID2(( Yu )lI i
p
(T)
:s; N( lI i Ylli
p
(T )
+ II DgYll t(T» )'
(2.7)
where here and below N = N( d, p, 50, 51 )' Below we also denote by N* generic constants depending only on d,p,5 0 ,5 1 ,f: , and K . Observe that
(Y (x) LtUt(x) - L¥(( Yut )(x ) = (Y (x) (1 /2 ) [a~j (x) - a~j (y)]Dij Ut(x)
+ (Y(x)[ Ltut(x) -
( 1/2) a~j(x) Dij ut(x)]
+ [(Y(x)( 1 /2)a~j(y) Dij ut(x) - L¥((Yut )(x )]. Here the right-hand side is zero unless x E G and Ix- yl :s; f:
and it follows
from our assumptions that
where
Therefore,
lI i Y lli p(T) s N II(Y i ll i p(T) + N,Bg II (Y D2u tlIC (T) + N* I\77 Y (IDu t l + !Utl)lIi (T) ' p
(2.8)
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
157
Furthermore,
D((Y Atut - A¥((Yut}) = (YAtDUt - A¥((YDUt) + [D((YAtut) - (YAtDUtl + [A¥((Y DUt) - DA¥((Y ut)], and
A¥((YDUt)I~2 ::; N,B5'I(YIPID2utIP + N*lryYIPIDutIP , ID((YAtut) - (YAtDutI~2 ::; N*lryYIP(IDuti P + IUtIP), IA¥((Y DUt) - DA¥((YUt)I~2 ::; N*lryYIP(IDutIP + IUt IP).
I(Y AtDut -
Hence,
II D gYII[p(r) ::; N(II(Y Dgllt(r)
+ IlryYglI[p(r») + N,Bb'II(Y D 2Utll[p(r) +N*II77Y(I Dutl + lutI)ll[p(r)"
Coming back to (2.7) we conclude that for each y E ]Rd
IID 2(( Yu)ll[p(r)::; N(II(Yfllt(r) + II(YDgll[p(r) + IlryYgll[p (r») +N,B5'II(YD2Utll[p(r) + N*II77 Y(IDutl + IUtJ)ll[p(r)" We now integrate both part of (2.6) with respect to also use simple observations like
r II(YD2Utll[
JIRd
(r ) p
dy =
X,
t, and w . We
Err ( r (P(x-y)ID2Ut(xW dY)dxdt Jo JJRd JIRd
1Ld r
= E
ID 2Ut(xW dxdt
2ulI[p(r)'
= IID
Ld IlryY( IDutl + IUtl) II~p(T) dy ::; N* (1IDullt(r) + Ilull[ p(r»)· Then we find
2ulI[p(r) IID ::; N(llfll[p(r) + Ilg ll~~(r») +N1,Bb'II D 2ullt(r ) + N*l lul l~~(r)'
(2.9)
where N 1 = N 1(d,p , 60, 81 ) . We can now specify the value of ,Bo by taking it such that
N 1 ,B5' ::; 1/2. Then by collecting like terms in (2.9) we come to
158
NICOLAI V. KRYLOV
Here by interpolation inequalit ies, for any I E (0, 1],
Il u ll~~ C"r) ::; IIID2u IIC(T) + N (d,p)t-l llulli p(T)' so t hat by choosing I to satisfy Ni l::; 1/2 we get (2.5) from (2.10). The lemma is pr oved. 0 Proof of Theorem 2.1. If R = 1 the result follows directly from Lemma 2.1. The case of general R we reduc e t o t he par t icular one by using dilations. Introduce
,f-:F k J"t = R2t, T• = R- 2 T , W· tk = R -1 WR2t' (at, i; Ct, O"t , Vt)(x ) = (aR2t, Rb R2t , R 2cR2t, (jR2t, RVR2t )(Rx), Ut(x ) = UR2t(Rx ), f t(x ) = R 2fR 2t(Rx) , gf (x ) = Rg~2 t (Rx) . Also introduce the operators i; and A~ constructing them from the above introduced coefficients. It is easily seen that w~ are indepen dent Ft - Wiener processes, f is an Ft-stopping time, all the above pr ocesses with hats are pr edictable wit h respect to the filtr ation {Ft } , and U E H;,o(-T) , f E Lp(f) , 9 E Jfu~(f) , where the sp aces with hats are defined on t he basis of {Ft} . Ob serve that for t ::; f
£ tUt(X) = (1/2)a i j (x )DijUt(x ) + b~ (x)DiUt (X) + Ct(X)Ut(x) ij i R 2 [( 1/2)aR2t Dij uR2 t + bR2tD iUR2t + CR2tUR2t1 (Rx) = R2LR2tUR2t(Rx ),
=
1 t
[£ sUs(X) + f s(x )]ds =
1
~t
[Lsus( Rx ) + f s(R x )]ds.
Of course , we un derst and t his equality in the sense of distributions:
for any ¢ E CO" (lR d ) . One also knows that if ht is an Ft-predictable process satisfying a natural integrability condition with resp ect to t , then (a.s.). Therefore, (a.s.) for finit e t ::; f
it[A:u + s
it[A~2sUR2S + g~2s] (Rx) 1
g: ](x ) dill: = R
2
R t
=
[A: us+ g:](Rx) dw: .
dw:
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS ~
It follows that (a.s.) for finite t
it
[tsUs(X) + ls(x)] ds +
159
f
it [A~us +
g;](x) dw;
= UR2t(Rx) = Ut(X),
so that u satisfies equation (1.2) with new operators and free terms. It is also easy to see that our objects with hats satisfy the assumptions of the theorem with R = 1. Therefore, by the result for R = 1
Now it only remains to notice that changing variables shows that this inequality is precisely (2.4). The theorem is proved. Here is an interior estimate. THEOREM 2.2 . Take some e.R E (0 ,00) and a stopping time T and suppose that the assumptions (i) and (ii) of Theorem 2.1 are satisfied. Then, for any r E (0, R) , we have
IIIB~D2UlllLp (7")
+ R-IIIIB~DulllLp (7")
~ N(II IBjJ lllL p(7")
+ IIIBnD9IiJLp(7") (2.11)
+ (R -
r)-IIIIBn9 11ILp (7"»)
+ N*(R -
r)-21IIBnulllLp (7" ) ,
where N = N(oo, 01, d, p) and N* = N*(K, e, 00 , 01 , d,p) . Proof. We follow a usual procedure taken from the theory of PDEs. Let x( s) be an infinitely differentiable function on lR such that x( s) = 1 for s ~ 1 and X(s) = for s ~ 2. For m = 0,1,2 , ... introduce, (ro = r)
°
m
rm = r
+ (R -
r) ~ Ti,
~m(x)
= X(2 m+l(R -
r)-l(lx'j- r m) + 1).
i=1
As is easy to check, for
Q(m) = B~ , '"
it holds that (m
= 1 on Q(m) ,
(m
=
°
outside
Q(m + 1).
Also (observe that N2 m+l = N 12 m with N l = 2N)
ID(ml ~ N2 m(R -
r)-I,
ID2( m l ~ N2 2m (R _ r)-2.
Next, the function (mUt is in 'H.~,O(T) and satisfies
where
160
NICOLAI V. KRYLOV
Since (mUt(x) = 0 for x
rt Bk, by Theorem 2.1
Am := IID 2((m u) IIE
(r)
p
+ R-PIID((mu)IIE
1 r
+ N* R- 2pUO + N E
(B mt + c';
(r ) :::;
p
N(P' + R-PCO) (2.12)
+ G;"t + R-PC~t) dt ,
where
P' =
IIIBjJIIEp(r)'
GO =
IIIB R91 It(r)'
UO =
IIIB RuIIEp(r)'
B mt = Ilu(Lt - Ct)(m - a~j DiUtDj(mllt (IRd), C;'t = I ID(UtO"~Di(m)llip (IRd )'
G;"t = IID ((m9t)llt(IRd) ,
C~t = IlutO"~Di(mllip(IRd )" Observe that by the above mentioned properties of (m and the assumption on b, we have
It follows that
e.; :::; N*2 2mp(R -
r)-2 PU? + N*2
mp(R
- r)-PIIDutllip(Q(m+l»'
where
Furthermore, by interpolation inequalities for any "I >
II DUtIlip(Q(m+l»
a
:::; IID( (m+l Ut) Ilip (IRd)
:::; 'Y (R - r)P2-mpIID2((m+1Ut) llip(Q(m+2» + N'Y-12mp(R - r)- PU? , so that for "I E (0,1) (with "I, perhaps, different from the one above)
B mt :::; 'Y II D 2((m+1Ut)lli p(Q(m+2»
+ N*'Y-122mp(R - r)-2 PU?
Simil arly, for "I E (0, 1)
C;'t :::; 'YIID2((m+lUt)llip (Q(m+2»
+ N*'Y-122mp(R - r)-2 PU?
and almost obviously
:::; N2 mp(R - r)-PU? :::; RPN 'Y-122mp(R - r)-2 PU? mp(R G;"t :::; N( IID9t llip(BR ) + 2 - r)-PI19tllip(BR »)·
C~t
Hence (2.12) yields
Am:::; 'YAm+l
+ NP + N2 mp(R - r)-PGO + N*'Y-122mp(R - r)-2 PUO ,
161
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
where
Now we take "( = 8- P and get "(m Am::; "(m+1 A m+ 1
+ N"(mp + N"(m2mp(R _
r)-PC o
+N*"(m"(-122mp(R _ r)- 2PUO, 00
Ao +
L
(2.13)
00
"(mAm ::;
m=l
L
"(m Am
+ NP + N(R -
r)-PCo
m=l
+N*(R - r)-2 PUO. In order to cancel like terms in (2.13) we need the series to converge. Observe that
and therefore 00
L
"(m A m ::; N(l
+ (R -
r)-2P)llull~~(r)"
m=l
However, the right-hand side may be infinite. In order to circumvent this difficulty we take in the beginning of the proof T 1\ T in place of T , where T E (0,00) . Then by the definition of'H~(T) we have
lIull~~(r/\T) < 00 and we get from (2.13) that A o for modified T is less than the right-hand side of (2.11). Since its left-hand side, with TI\T in place of r , is obviously less than A o, we obtain (2.11) with T on the left replaced with T 1\ T. After this it only remains to let T ~ 00. The theorem is proved. 0 REMARK 2.1. One can prove interior estimates for higher order derivatives of solutions if the coefficients are more regular. This is a routine matter and is achieved by considering equations for derivatives of solutions, which one obtains by differentiating the equation. On this way one would obtain more regular solutions of equations in half spaces. However, for brevity we do not pursue this issue. 3. Interior estimates for solutions of parabolic PDEs in the whole space. Assume that for i , j = 1, ... , d we are given real-valued Bor el measurable functions a~j (x), b~(x), Ct(x) defined for (t, x) E JR+ X JR d , where
JR+ = (0, (0).
ASSUMPTION 3.1. For all values of the indices and arguments a ij j i a and, for a constant 50 E (0,1] and all unit>' E JR d ,
5- 1 > aij>.i>.j > 5 .
° _
t
_
°
=
162
NICOLAI V. KRYLOV
This assumption is supposed to be satisfied throughout the deterministic part of the paper. Here we take a constant T E (0,00] and consider equation (1.1) The solutions of (1.1) are looked for in the space H;,o(T) with f E lLp (T). We remind briefly the definition of these spaces. Everywhere in the deterministic part of the article p>1.
= 0, 1,2 we take Sobolev spaces W; = W;(JRd )
For an integer n we set
and
In an obvious way we introduce the norms in these spaces. Observe that we are using the same notation as in the stochastic case . Deterministic spaces are subspaces of stochastic ones and there is no real danger of confusion. If for any finite t E [0, T] we are given a generalized function U = Ut on ]Rd such that U E no<s
holds for all finite t
~
T. In that case we write
a
at Ut = ft, which explains the sense in which (1.1) is understood and set
We take d 1 , x' , and Bk from Section 2 and state the first main result of this section. THEOREM 3.1. Take some e, R E (0,00) and assume the following : (i) There exists a constant K E (0,00) such that for x E Bk and y E ]Rd, such that Ix - yl ::; eR, and all values of the indices and other arguments Rlb~(y)1
+ R2Ict(y)1 ~ K ,
la~j (x) - a~j (y)1 ~ 130 ,
where 130 = 13o(d, p, 50, 5d > a is a (small) constant from Lemma 3.1 below. (ii) We are given a function U E H; ,o(T) which is a solution of (1.1) with an f E lLp(T). Finally , assume that Ut(x) = a if x ¢ Bk. Then there exist constants N = N(d,p,5 0) and N* = N*(d,p,5 0,e: ,K) such that
IID2U lllLp (T)
+ R-11 IDuliJL p(T) ~
Nllf lllLp (T) + N* R-21 IulIlLp (T)'
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
163
Similarly to T heore m 2.1 t his t heorem is obtained by using the parabolic dilations from the following lemma in which one can take G = Bk. LEMMA 3 .1. As sum e that we are given a U E 1i~, o (T ) which satisfies (1.1) with an f E Lp(T) . Assume that for a domain G C IR d (perhaps, G = IR d ) we have Ut(x ) = 0 if x rf. G. Also assume that there are constants e, K E (0, 00) such that for all x E G and y E IRd , with Ix - yj :S e, and all values of other arguments we have
I b~ (y) 1
l a~j (x) - a~j (y)1:S (30,
+ ICt(y)1:S K ,
where (30 = (3o(d, p, 00) > 0 is a (small) constant an estimate from below for which can be obtained from the proof. Then there exist constants
= N(d,p,oo),
N
N*
= N*(d, p,oo,c, K)
such that (3 .1 )
Proof. We take a nonnegative ( E Off (lR d ) from the proof of Lemma 2.1 and use t he facts that (2.6) holds and for each y E IR d the function (YUt satisfies
where
By classical results for equations wit h coefficients independent of z , for each y (3.2) wher e here and below N = N(d , p, 00) ' Below we also denote by N* generic constants de pendi ng only on d, p, 00, e, and K . By combining (2.8) wit h (3.2) we see that for eac h y E IR d
IID2(( Yu )llf p(T) :S N(II(Y fl lfp(T) + N(3b'II(Y D 2Ut llf p(T) + N *II T)Y(I DUt l + IUtI) IIC (T)' We now integrate both par t of (2.6) wit h respect to t , x. We also use simple observations like T II ( YD2Ut li r (T) dy = (P(x - y )ID2Ut(x )IP dy) dxdt
r
i;
p
r r(r
Jo J'iR d J'R,d
= iT
ld
ID 2Ut(x)IPdxdt =
II D2ullf p(T)'
164
NICOLAI V. KRYLOV
Then we find
where N 1 = N1(d ,p, bo). We can now spe cify the value of 13o by taking it such that N l13g
:s; 1/2.
Then by collecting like terms in (3.3) we come to (3.4) Here by interpolation inequalities, for any v E (0,1 ]'
so that by choosing 'Y such that Ni, :s; 1/2 we get (3.1) from (3.4). The lemma is proved. Our next result is the following interior estimate. THEOREM 3.2. Take some € , R E (0,00) and suppose that the assumptions (i) and (ii) of Theorem 3.1 are satisfied. Then, for any r E (0, R) , we have
I I IB~ D2 u ll lLp(T)
+ R-IIIIB~Dul llLp (T)
:s; N IIIBjJ IIlLp(T) + N*(R - r)-2 1IIBkU lllLp (T),
(3.5)
where N = N(bo,d,p) and N* = N*(K,€,bo ,d,p) . Proof. We take T'm , (m , and Qm from the proof of Theorem 2.2 and notice that the function (m Ut is in 1{;,0 (T) and satisfies
where
By Theorem 3.1 2 Am := IID (( mu )llt
(T) + R-P IID((mu)ll r p(T)
:s; N F + N* R- 2
pU O
+N
iT e.; dt ,
(3.6)
ON SPDES IN SOBOLEV SPACES WITH WEI GHTS
16 5
where
F = II IBiJlltcT)' Be«
=
Uo = II IBRull[pCT)'
II U( Lt - Ct )(m - a~j DiUtDj (m ll ~p (Rd )"
Observe that by t he propert ies of ( m and its derivatives and t he assumption on b, we have
It follows that
where
Fur thermore, by interpolat ion inequalities for any , >
°
II Dutll ~pCQ(m+l)) 5 II D((m+l Ut)ll ~p(lRd) 5 ,(R - r)pTmp IID2 ( (m+l Ut ) l l ~ p (Q(m+2) ) + N ,-12mp(R - r )- PU?, so t hat for , E (0, 1) (wit h " perhap s, different from the one ab ove)
s.; 5
,II D2((m+l Ut)l l ~p(Q (m+2) )
+ N *, - 122mp(R -
r)-2 PU?'
Hence (3.6) yields
Am 5 , Am + 1 + N F
+ N *, - 122mp(R -
r )-2PUO.
Now we t ake , = 8- P and get
, mAm 5 , m+ 1A m+ 1 + N ,mF 00
+ N *,m, - 122mp(R _ r )-2PUO,
00
A o+ '2::,mA m 5 '2:: ,mAm+NF+N*(R-r )-2 PU O. m=l m=l
(3.7)
In order to ca nc el like terms in (3.7) we need the series to converge . Observe that
and therefore 00
'2:: ,mAm 5 N( l + (R - r)- 2P ) lI u l l ~~ cT)"
m=l
166
NICOLAI V. KRYLOV
However, the right-hand side may be infinite if T = 00. In order to circumvent this difficulty in that case we take in the beginning of the proof S E (0,00) in place ofT = 00. Then by the definition ofH~(oo) we have
Ilull~~(s) < 00 and we get from (3.7) that A o for modified T is less than the right-hand side of (3.5) . Since its left-hand side, with S in place of T, is obviously less than A o, we obtain (3.5) with T on the left replaced with S. After this it only remains to let S -- 00. The theorem is proved. REMARK 3 .1 . One can prove interior estimates for higher order derivatives of solutions if the coefficients are more regular. This is a routine matter and is achieved by considering equations for derivatives of solutions, which one obtains by differentiating the equation. On this way one would obtain more regular solutions of equations in half spaces. However, for brevity we do not treat this issue . 4. Local regularity near the boundary for SPDEs in half spaces. In the setting of Section 2 we will be considering (1.2) in !Ri . We need the stochastic Sobolev spaces JH[;,ii(r), n = 0,1,2, and 5);,ii(r) introduced in [15] in the following way. We take a number e E !R and introduce Lp,ii as the space of functions on !Ri having finite norm given by
Then H~,ii is the set of functions such that u, MDu E Lp,ii, where M is the operator of multiplying by xl, and H;,ii is the set of functions such that u, M Du, 1\1[2 D2u E Lp,ii' The norms in these spaces are introduced in a natural way (cf. Remark 4.2 below) and, as before, we extend these definitions for .e2 -valued functions such as g. Naturally, we write H~,ii =
i.:
These definitions are used for all p E (1, (0) . REMARK 4.1. In [12] the spaces H;,ii are introduce for all real, and it is proved that CQ'(!Ri) is dense in H;,ii' In our case that, = 0,1 ,2 this is a rather simple fact. To prove it , it suffices to show that the subset of H;:'ii ' n = 0,1 ,2, consisting of functions on !R~ with compact support is dense in H;:'ii ' To do that, we take some nonnegative ~ E CQ'(JR) and 7] E CQ'(!Rd-l) such that ~(O) = 1, e(t) = for ItI ~ 1, and 7](0) = 1 and for m = 1,2, ... define
°
(m(x) = ~(m-Ilnxl)7](e-2mx') ,
x E !Ri.
Notice that
Ix l Dl(m(x)\ Ixl Di(m(x)1
= Im-I((m-Ilnxl)7](e-2mx')I, = xle-2ml~(m-llnxl)(Di7])(e-2mx')I,
i ~ 2,
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
167
and if ~(m-qnxl) i 0, then m-qnx 1 ~ 1 and xl ~ em. It follows that Ix1D(m(x)1 tend to zero as m -+ 00 uniformly on The following formulas in which i ,j ?: 2
Ri.
l(x l)2 DU(m(x)! = Im- 2C(m-1lnx1) - m-1nm-1ln x 1)ITJ(e- 2mx'), I(x l)2 DU(m(x)1 = xle-2mm-Ilnm-lln x) (D i TJ )(e- 2mx')I, l(x l)2D ij(m(x)1 = (xl)2e-4ml~(m-llnxl)(DijTJ)(e-2mx') 1
Ri.
show that l(x 1)2D2(m(x)1 tend to zero as m -+ 00 uniformly on Finally, observe that (m are uniformly bounded and tend to 1 pointwise. Now the dominated convergence theorem and the formulas lu - u(ml = 11 - (mllul, IMDu - MD((mu)1 ~ 11- (ml lDul + luMD(ml, 1M2D 2u - M2D2((mu)1 ~ 11- (m11M2D 2ul + 2IMD(m l tMDul + luM 2D 2(ml easily show that , if n E {O, 1, 2} and u E H;,o, then (mu -+ U in H;,o as m -> 00 . REMARK 4.2. A few times in the future we will be dealing with functions u such that M-Iu E H;,o' In connection with this it is useful to have in mind that M D i(M-Iu) = DiU - JliM-1u , M 2Dij(M-Iu) = M Diju - JUDju - Jlj DiU + 2JUJlj M-Iu implying that
IIM- I u IIHl = IIMD(M-1u)IIL e + IIM - Iu IIL e ~ N(IIDu IILp ,e + IIM - I uIILp, e), IIDuIILp,e::; NIIM-Iu IIH~ ,e ' IIM- I u IIH2 = IIM 2D2(M-1u)IIL e + IIM-IuIlHl ~ N(IIMD2uIILp ,e + IIDuIILp ,e + IIM-IuIILp,e), p,e
p ,e
p,
p,
p,
p ,e
II M D 2u liLp , e + IIDu llLp, e + IIM-1u IIL p, e ~ N IIM- 1u IIH2p,a . It follows that the H;,o norm of M-1u is equivalent to
Next , for n = 0,1 ,2 and stopping times 7 we define JHI;,O(7) as the set of functions f = ft = ft(w) on (0,7] , t < 00 , with values in the set of distributions on which are P-measurable and have finite norm given by
Ri
I lfll~;,e (7") = E 17" 11ft II iI;,e dt.
168
NICOLAI V. KRYLOV
Define Lp,l:I (r) = lHI~, I:I (r) . For a function U = Ut = Ut(w) given on (0, for finite t wit h values in t he set of distribut ions on lRi we write U E 55; ,I:I ,o(r ) if and only if M -Iu E lHI;,I:I(r) and t here exist a real valued f E Lp,l:I (r) and an £2valued 9 = (gk, k = 1, 2, ...) E lHI~,I:I (r) such that for any ¢ E CO" (lRi ) wit h prob ability 1 we have
rn
f
(Ut, ¢) = t (M-Ifs, ¢ )ds+ t (g: , ¢ )dw: k=IJO Jo for all finite t E (0, r]. We set
I l ull jJ~.e (T) = I IM-Iu lI lH~, e (T) + IIfIlLp,e(T) + I l g l llHI~ ,e (T) ' Recall that Assumptions 2.1 is supposed to be satisfied throughout the article. For r > 0 denote QT={ X ElRd :O<x l < r }.
THEOREM 4.1. Take an R E (0, 00] and a stopping time -r and assum e the f ollowing. (i) For som e constants s E (0, 1] and K E (0, 00) we have ' 1 IxI b~. (x) 1 + Ix I DC7;(x )ll2 + Ix Vt(X)!l2
(4.1) + l(x l f DVt(x )ll2 + l(x l ) 2 Ct( x) j ::; K , l a~j (x) - a~j (y)1 + 1C7:(X) - C7;(y )ll2 ::; (30, whenever x ,y E lRi, x l ,yl ::; R , Ix - y l::; s( x l/\ yl ), i, j = 1, ... , d, t > 0, where{3o = (30(d, p, 50, 5r) E (0, 1] is the constant from Theorem 2.1; (ii) We have a fun ction U such that ¢u E 1{~ , o (r ) for any ¢ E CO" (QR ) and u satisfies (1.2) in lRi with some f E Lp,l:I(r) , g = (gk, k = 1,2 , ...) E lHI~,I:I(r) .
Then for any r E (0, R /4 ) II IQrM D 2u llLp ,e(T) + II I QrDu lhLp, e(T) ::; N IIIQR M f liJL p,e(T)
(4.2)
+ NI IIQRM Dg IIJL p,e(T) + N IIIQ RgII JLp ,e (T ) + N *I IIQ RM-l u II JLp.e(T)' where N = N(d,p,5 0,5 l) and N * = N*(d,p,5 0, 51 , s , K ). Proof. We are going to appl y Theorem 2.2 t o shifted B Rwhen d l = 1. For n = -1 ,0,1 , ..., set r n = 2- n / 3r . Observe that if n 2: 0, t hen the half widt h of Qrn_l \ Qrn+2 equals Pn := rn+2/2 and r n-l
+ Pn ::; 2r -1 < 4r < R ,
r n+2 - Pn = Pn·
It follows th at for x E Qrn_l \ Qr n+2 and y , such that Ix - yj ::; SPn, we have
R > rn-l 2: x l 2: r n+2 2: Pn, yl::; Xl + cPn < R , yl 2: x l - SPn 2: Pn, Pn ::; Xl /\ v', Ix _ yl ::; c (X l /\ yl ),
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
169
so that by our assumptions
Furthermore, if n :::: 0, ( E 00((0 , R)) and «(z) = 1 for rn + 2 ::; z ::; rn-l , then (u satisfies (1.2) in IR d with certain f and g which on Qrn-l \ Qr n+2 coincide with the original ones. Finally, if n :::: 0, then the distance between the boundaries of Qr n \ Qrn+l and Qrn_l \ Qr n+2 is (21 / 3 - l)rn + 2 . It follows by Theorem 2.2 that for n :::: 0
We multiply both parts by r~t~-d and use the facts that rn-l on Qrn_l \ Qr n+2 the ratio xl/r n + 2 satisfies 1 ::; xl / r n + 2
::;
= 2r n+ 2 and
2.
Then we obtain
Upon summing up these inequalities over n :::: 0 we conclude
IIIQ~MD2ull[p,O(T) + IIIQrDull[p ,o(T) ::; N(IIIQ~_l M f ll[p,o(T )
+ I IIQ~_l M Dgll[p,O(T) + Ilh Qr_ gllt ,o(T)) + N*llhQ~_l M-lullt,o(T)' 1
which is somewhat sharper than (4.2). The theorem is proved. 0 By letting r ~ 00 in (4.2) we get the following . COROLLARY 4.1. If the assumptions of Theorem 4.1 are satisfied with R = 00, then
+ IIDulllLp.O(T) ::; NIIM fl llLp,o(T) + NllglllHI;,o (T) + N*IIM-lUlllLp,o(T) '
11M D2u lllLp,O(T)
REMARK 4.3. This corollary for equations with constant coefficient is known from Lemma 3.6 of [14] for d = 1 and Lemma 3.8 of [15] for d :::: 1. For variable coefficients it is Lemma 4.1 of [8] if d = 1. However, there is a
170
NICOLAI V. KRYLOV
very imp ortant distinction betwee n Corollary 4.1 and t he above mentioned references, where from t he st art it is assumed t hat u E .fj ~, II ,O (T ). As we have mentioned in t he Introduction, we intend to use Theorem 4.1 to extend to t he case of variable coefficients t he pointwise boundar y Holder regularity result from [13] proved there for equations with constant coefficients. In a subseq uent article we will show that solutions in .fj ~ , II ,O (T ) spaces wit h rather larg e e admi t estimates of M-I u in lLp ,I'(T) with much smaller J.L if T is chosen in a special way. This, Theorem 4.1, and embedding t heorems will lead to proving t he pointwise Holder continu it y of solutions near t he boundary. REMARK 4.4. We discuss assumption (4.1) in case that R = 00. It impli es that l a~j (x) - a~j (y)1 ::; (3 for a small (3 > 0 and all x, y E JR~ satis fying Ix - yl ::; x l 1\ yl . If Xl = yl = r, then we need l a~j (r, x' ) a~j (r, y' )1 ::; (3 for lx' -y'l ::; r and when r becomes larger we need a~j (r, x' ) to be close to const ants on larg er and larger balls in JRd-l . Basically, we need a~j (r, x ' ) to be independent of x' for large r. On the other hand , the behavior of a~j (x) near aJR~ can be quite irregular. For instance, take d = 1 and introduce the functi on
a(x ) = 2 + cosln x . If x ~ y > 0 and [z - yl ::; e(X 1\ y ), t hen x ::; (1 +e )y, In x ::; In y + lnfl and, for a ~ E (In y, In x), by the mean value theorem we have
+ e)
la(x) - a(y) 1= Isin ~ I (ln x -In y ) ::; In( 1 + e), which can be made arbitra rily small by choosing an appropriat e e. REM ARK 4. 5. It is worth emphasizing that in this section t here is no restrict ions on e, say like (1.3). 5. Local regularity near the boundary for parabolic PDEs in half spaces. In the setting of Secti on 4 we will be considering (1.1) in JR~ . We need the weighted Sobolev spaces lHI;,II (T ), n = 0,1 ,2 , and .fj ~ ,II (T ) introduced for p > 1 in t he following way. We take a number e E JR and t ake L p ,1I and H;,II from Section 4. Next, for n = 0,1,2 and T E (0, 00] we define lHI;,II (T ) as the set of functi ons i = it on (0, T ) wit h values in the set of distributions on JR~ which are measurable and have finite norm given by
I l fl l~;. 8(T) =
I TIIftlliI;.8
dt .
Define lLp ,II (T ) = lHI~, II (T) . For a function u = Ut given on [0, T ] n [0, 00) wit h values in t he set of distrib utions on JR~ we write u E .fj ~, II ,o (T ) if an d only if M -1u E lHI~,II (T)
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
and there exist a real valued we have
171
f E JLp,e(T) such that for any ¢ E Co(JRt)
for all finite t E (0, T]. In that case we write
and set
Recall that Assumptions 3.1 is supposed to be satisfied throughout the deterministic part of the article. For r > denote
°
Qr={XEJRd :O<X l
(i) For some constants e E (0, 1J and K E (0, (0) we have
whenever x, y E JRt, xl , yl ::; R, Ix - yl :$ c(x l 1\ yl), i, j = 1, .. ., d, t > 0, where f30 = f3o(d,p, (0) E (0,1) is the constant from Theorem 3.1; (ii) We have a function u such that ¢u E H~ ,o(T) for any ¢ E Co(QR) and u satisfies (1.1) in JRt with some f E JLp,e(T) . Then for any r E (0, R /4)
IIIQrM D2ulllLp ,o(T) ::; N IIIQR M flllLp ,o(T )
+ II I Qr D ul iJL p.o(T)
+ N*IIIQRM-lul llLp ,o(T),
(5.1)
where N = N(d,p,oo) and N* = N*(d,p,oo,c ,K) . Proof. As in the proof of Theorem 4.1 for n = -1,0,1 , ... we set r n = 2- n / 3 r and observe that if n 2: 0, then for x E Qrn_l \ Qr n +2 and y , such that Ix - y l ::; cpn, where Pn is the half width of Qrn _l \ Qr n +2' we have
Furthermore, if n 2: 0, ( E Co((O, R)) and ((z) = 1 for rn + 2 ::; z :$ rn-l, th en (u satisfies (1.1) in JRd with certain t, which on Qr n_l \ Qr n+ 2 coincides with the original one. Finally, if n 2: 0, then the distance between the boundaries of Qr n \ Qrn+l and Qrn-l \ Qr n+2 is (2 1/ 3 - 1)rn + 2 '
172
NI COL A I V. KRYLOV
It follows by Theorem 3. 1 that for n 2: 0
We multiply both p art s by r~t~-d and use the facts t hat rn-l = 2r n+ 2 and on Q rn-l \ Qr n+2 the ratio xl /rn + 2 satisfies
1 :::;
xl /rn + 2 :::; 2.
Then we obtain
Upon summing up these inequalities over n 2: 0 we concl ude
IIIQrMD2uIIEp.9(T) + III QrDu ll[ p.9(T) :::; N IIIQr_l Mf ll[p,9(T) + N' lIh~r_l M- l ull [ p,9(T)' which is somewhat sharper than (5.1) . The theor em is pr oved. 0 By letting r -+ 00 in (5.1) we get the following. COROLLARY 5 . 1. If the assumptions of Th eorem 5 . 1 are satisfied with R = 00, then
6 . Auxiliary results. In this section
p>l ,
,:= B - d - p+ 1.
He re we present some auxiliary facts needed for proving existence theorems in Sections 7 and 8. Several times in t he future we use t he following Scheffe 's lemm a , which we prove for com pleteness of presentation. LEMMA 6. 1. Let (E, E , J.1.) be a measure space, r E [1,00) , un ,u E L; (J.1.) 1 and Un -+ U in meas ure . Fin ally, let
Th en
(6.1)
173
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
Proof We have
Upon integrating through this equation and observing that (Iul r - Iunn+ :::; lulr we conclude by the dominated convergence theorem that
L
Ilul r -Iunn f.L(dx)
Next, if IU n r
(1 /2)lunl
-
ul ~ 31u/, then lunl + lui
~
~ O.
(6.2)
31ul, lui:::; (1/2)lunl, lulr
:::;
,
lunl r -Iul r
~ (1/2)lu n l r
r
,
rlu
IUn - ul ~ 2
ul :::; lun l + lui:::; 2lunl, r n l -Iun , n l :::; 4 IU n
r
-
r(lu
which along with (6.2) imply that
L
r
~ O.
Ie
r
~0
IUn - ul Ilu n-ul2':3Iul J-L(dx)
Furthermore,
IUn - ul Ilu n-ul<3I ul J-L(dx)
by the dominated convergence theorem. By combining the last two relations we come to (6.1) . The lemma is proved. 0 COROLLARY 6.1. Let (E,L"J-L) be a measure space, r,s E (1,00), r- 1 + s-l = 1, Un,U E L; (J-L) , Vn, V E Ls(f.L) , Un ~ u and Vn ~ v in measure. Finally, let
Then
Indeed, it suffices to use Holder 's inequality and the formul a
UnV n - UV = (un - u) v + (v n - v)u + (un - u)(vn - v). Next, we prove a version of Corollary 6.2 of [12]. LEMMA 6.2. If we are given a d x d matrix a, then , for any u we have
E
M H P. 2e,
174
NICOLAI V. KRYLOV
and if ii is nonnegative and symmetric and'Y < 0, then
where (if p < 2) by 0- 10 we mean O. Proof. If U has compact support, then, owing to the assumption that 'Y < 0, the result follows from Corollary 6.2 of [12]' which is proved by using integration by parts and Hardy's inequality. Observe that the case 1 < p < 2 is quite nontrivial. In the general case we take (m from Remark 4.1 apply (6.3) to U rn = (rn U in place of U and pass to the limit as m -> 00. To justify this procedure we claim that
(6.5) are continuous functions on M Lp,fJ and M H;,fJ' respectively. The first one is continuous because it is just the p-th power of the Lp ,o-norm of M- 1u. To prove the continuity of the second one, take Vn such that M-1 vn -> M-1 u in H;,o. Then
in Lp,fJ by Remark 4.2 . By Lemma 6.1 and Corollary 6.1 we also have p IM - 1v n l - 2M- 1vn
->
IM- 1ulp- 2M- 1u
in
L p/ Cp- 1) ,fJ,
{ (xl )1'+1IVnlp-2VnDijVn dx
),fiI·t
= { (Xl )fJ-d( IM- 1v n!p-2 M-1Vn)(MDijVn) dx
JRt
->
{
JRt
(xl )fJ-d(!M- 1u I P - 2M- 1vu)(M Diju) dx
and this proves the continuity of the second function in (6.5). If p ;:::: 2 then a similar argument shows that the last expression in (6.3) is continuous on M H;,o and this proves the lemma if p ;:::: 2. For arbitrary p > 1 we need to proceed differently. Since (6.3) holds with Urn in place of u, we conclude that
ON SPDE S IN SOBOLEV SPACES WITH WEIGHTS
17 5
where
nim .= . IM- I UmIP - 2 (D iUm)D jU m =
(!:.IM -IuIP-2(DiU) Dju + IM- Iu !p-2(M- Iu)( D
iu)(nM o«; + !M-IuIP-2(M-Iu)(Dju)(nM »«;
(6.7)
+IM- Iu IP(M D i( n)M o.c;
Here the fact ors of t he functi ons
are integr abl e against (Xl )iI- d dx, which follows from Holder 's inequ ality, and the functions themselves are un iformly bounded and tend t o zero on Therefore, (6.6) implies that
lRt.
By taking here (iij = rS ij and using Fat ou 's lemma we see t hat
Afte r t hat by (6.7) and t he domi nated converge nce t heore m we obtain
which along with (6.6) lead to (6.3). The above argument also shows that one can pass to the limit in (6.4) written for Urn in place of u . This proves the lemma. 0 C ORO LL A RY 6. 2. Den ote
Then
Q ::; N (d,p) P
+ (p _1 )- lp- I(r2 + 'Y)II M - I u lI ~
P ::; II M- I ul l ~~.~ 11MD u IIL p , B ' 2
p .B
,
(6.9) (6.10)
176
NICOLAI V. KRYLOV
Indeed, estimate (6.10) follows from Holder's inequality. By (6.3) for (iij = oij we have
(p - l)Q = "(b + l)p-IIIM-Iullt
::; "(b + l)p- I IIM- Iu llt
r
l p,9 - JJRd (X p +l lu JP- 2u6.u dx p
,9
+ N(d)P
and (6.9) follows. Next, we need an auxiliary function . Set
Bk =
E ~d-l :
{x'
Ix'i < R}.
LEMMA 6.3. For any 0> 0 there exists an R = R(o,d,p) > 1 and a nonnegative function fJ E CO"(~d) vanishing outside (1, R) x Bk such that In R ::; N(d,p)O-I/2 and
r
~fJP(x) dx =
JJRd X +
(6.11)
1,
Proof. Take a nonnegative ~ E C6"'(~d) with unit norm in L p , vanishing outside (-1 ,0) x B l , and then for some r > 0 set 'T)(x) = r-l /Pe-(d-l)r/P~(r-Ilnxl, e- r x') Then
r
~fJP dx = r- l
JJRd+ xl
=r- l
r
~e(r-Ilnxl, x') dx
JJRd+ Xl
r e(r- Ix l,x')dx=l,
JJRd
r 'T)p- 2IDlfJI 2xl dx = r- JJRdr (e-2IDI~12)(r-Ilnxl, x')~xl dx 3
JJRd+
+
r e-2IDI~ 12 dx,
= r- 2
Ja
d
where the last integral is finite even if p < 2 since ID~12 ::; N(d)~ sup ID2~1 because ~ is nonnegative and smooth. Also , since ~(x) = 0 for xl > 0 and by assumption r > 0, for i 2: 2 we have
r
JJRi
r (e-2IDi~ 12)(r-Ilnxl ,x')x dx JJRi = r- Ie- 2r r (e-2IDi~12)(r-Ixl , x')e 2X1 dx JJRd 2r ::; rr (e-2IDi~12)(r-lxl, x') dx JJRd
'T)p- 2IDifJI 2xl dx = r- Ie- 2r
1e-
= e- 2r
r (e-2IDi~12)(xl, x') dx.
J,ll{d
1
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
177
By observing that e- 2r :S Nr- 2 , we have both relations in (6.11) satisfied for an r = N(d,p)rS- 1 / 2 . Now we have to take care of the support of ry. The one constructed above has support in
(e- r , 1) x
B~r.
It turns out that conditions (6.11) are dilation invariant. Therefore, the function ry( e- r x) satisfies them as well. This function has support in
(1, er ) x
B~2r
and we see that we can take R = e2r . The lemma is proved. 0 Now we extend (6.4) for variable a by using a localization procedure. LEMMA 6.4 . Assume that, < and let€,(3,K E (0,00) be some constants and let a(x) be a measurable junction given on JR~ with values in the set oj symmetric nonnegative matrices and such that la ij I :S K and
°
(6.12) whenever x , y E JR~ and Ix - yl :S €(x 1 /\ yl) . Then jor any u E M H~,o and rS > 0,11; E (0,1] we have
1:=
r (xlrr+l!uIP-2aij(Diu)Djudx
lJR~
2:: (1 - 1I;)/2 p-2
r (x1p-1allluI
lJR~
P
dx - N(€-lR + 1)(3Q (6.13)
-tite:' R(3 + (3 + lI;-lrS) IIM-1ullt,9' where N = N(d,p,K ,B) , InR = N(d ,p)rS- 1/ 2 and Q is introduced in Corollary 6.2. Proof. Take the function ry from Lemma 6.3 and for any y E JR~ set
ryY(x) = ry(y1x 1, yl(y' _ X'))(yl)(d-2 )/ p, (Y (x) = ryp(y1x1,yl(y' - x'))(yl)d-2 = [ryY(x )]p. Observe that , for any x E JRi and
0:
E
JR
(6.14)
For
0:
°
= this yields 1=
r I(y) dy ,
lJRd+
178
NICOLAI V. KRYLOV
where I(y)
= h(y) + 12(y)
h(y) = aij(y) h(y)
=
r (x }JRt
and with
y = ((yl)-l,y')
r (Xl)'+1(YluIP-2(DiU)Djudx ,
IiRt
Y+I(YluIP-2[a ij(x) - aij(y)](Diu)Djudx.
l)"
Here and below on many occasions we drop the argument x but not y. With R from Lemma 6.3 , if (Y(x) =1= 0, then 1 < ylx l < Rand ylly' _ xii < R implying that
yl < xl < Ryl, Iy' - xii < Ryl = R(x l /\ yl), 0< xl _ yl < Ryl, Ixl _ yll < R(x l/\ yl), Ix _ yl < 2R(x l /\ yl). In case 2R ~ e introduce n := [4Rc l] + 1 and find n points Xi, i = 1, ... , n, evenly spread on the straight segment connecting x and y and such that Xl = x, X n = y. Then n - 1 = [4Rc l] ~ 4Rc l - 1 ~ 2Rc l and
It follows that laij (x) - aij (Y)I ~ (n - 1){3 and
laij (x) - aij (Y)I ~ 4(C I R + 1){3, if 2R ~ c. Estimate (6.15) is also true if 2R 2R(x l /\ yl) ~ c(x l /\ yl) . Hence,
~
(6.15)
e, because then Ix - y! <
and by the construction of ( and the definition of Q
Thus , (6.16) To estimate
h (y)
we introduce
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
179
so that
where
Jj'(x) = (j,ij(Y)l vYIP-2(DiVY)DjvY,
J¥(x) = (j,ij(Y)lvYIP-2u2(Di1]Y)Dj1]Y .
It follows that
r h(y)dY"2(l-",) JJRir JJRtr (x1rr+1Jj'(x)dxdy _",-1 r r (Xl)1+1J~(x)dxdy. JIRt JIRt
JJRi
(6 .17)
Observe that
r (x 1)1+1J¥ dx s N r (x 1)1+l(7)Y)P-2ID1]YI 2Iu\P dx , JIRd'rJ.~t . r r (x 1rr+1J¥ dxdy s N r (xl)'Y+lluIP( r (1]Y)P- 2ID1]YI 2 dy) dx JIRt JIRt JIRt JIRt and by the construction of 1]
r (1]Y)P- 2ID1]YI 2dy JJRtr [1]P-2ID1]12](ylxl, yl(y' _ x'))(yl)d dy =
JIRi
=
r [1]p-2 ID1]12](ylxl ,y')yldy
JIRt
= (X l)-2
r 1]P- ID1]1 JIRt 2
2 l y
dy S O(xl )- 2,
r r (xlrr+lJ¥ dxdy S No JIRdr (Xl)1-lluI Pdx = NoI IM-1ulli.
JIRd JIRd +
+
+
. p,e
We now conclude from (6.17) that
r
JIRd
h(y) dY"2 (1- "')
+
r r
JIRd JIRd +
(x l)1+lJi(x) dxdy - ",- lNoI IM-lu lli.
. p,e
+
Next, apply Lemma 6.2 for each y to get
r (xl )1+1 Jj' dx "2 ,..2p-2(j,1l(y) r (Xl )1-llvY Pdx
JIRt = ,..2p - 2
r (xlrr-l(j,ll (y)(Ylu IP dx
JIRt
+ ,..2 p-2
JIRi
= ,..2p - 2
I
r (x l rr-l(j,ll(Y lu I dx P
JJRt
x l rr- l [(j,ll (y) _ all(x)](Y luIPdx. r ( JIRt
180
NICOLAI V. KRYLOV
We use again (6.15) on the support of (Y and then integrate with respect to y . Then we conclude
r r (x lp +1Jf dxdy 2: ·l p- 2 la~r (xl p - l a ll (Y lu IPdx
la~ j.iR~
- N (e- l R + l )f3 I1 M- Iu lli,p .9 ,
r h (y ) dy 2: (1 - K)'ip-2 At~r (x l p - l a ll lu IPdx
la~
- N[ (1 - K)(e- 1R + 1)13 + K- lol IIM-1u lli,p,9 ,
which in comb ination with (6.16) leads to (6.13). The lemma is proved. 0 We now generalize (6.3) for variable a. LEMMA 6.5 . Let e, 13 E (0,00) be some constants. Assume that we are given bounded m easurable functions a,ij(x) defi ned on lRi for i , j = 1, ..., d and such that (6.12) holds for all i , j = 1, ..., d and x, y E lRi such that Ix - yl ::; e(x l A yl) . Then for any u EM H; ,IJ we have
1:=
r (Xl P+ 1Iu jP- 2u ai j D ij u dx lad +
::; N f3(IIM D2u Il Lp, 9 1 I M- I u ll i,~.~ + e-2x IIM-1u llt.J
r (x lrr+1lu IP- 2ai j( Di u) Dj u dx lad+ + p-l , (-Y + 1) r (x l)"Y- Ia lll u IP dx ,
(6.18)
- (p -1 )
JR~
where X = (-y2 + 1)(1 + e2) and N = N( d, p) . P roof. As in the pr oof of Lemma 6.2 we may assume that t he support of u is a compact subset of lRi . W ithout losing generality we may also assume that a ij = a ji for all i,j. Next, take a nonnegat ive ( E C8"(lRi) with un it int egr al and such that ( (x) = 0 if xl rf. (1, 1 +e/2) or Ix'i 2: e/2. On ca n construct such functions in a unified way, first choosing such a function, say (1 corresponding to e = 1 and t hen for any e > 0 setting
Below we are also usin g the function 1](x)
=
(x lp +2(( x) .
One shows easily that for the indicator function 7f; of the set (1, 3/ 2) x we have
B~ / 2
ID 21](x)1 ::; Ne - 2- d (-y 2 + 1)(x I P [(x l )2 +e 2]7f;(1 + e- l(x l - 1), e - 1x ' ) ::; N (x 1 p e -
2
-
d
x 7f;(1 + e-l (x l - 1), e- Ix' ),
ON SPDES IN SOBOLE V SPACES WITH WEIGHTS
181
where N = N (d). In particular, (6.19) This estimate will be used later and now for y E ]Rd define
Observe that similarly to (6.14) we have
It follows that
1 = ( I (y) dy,
lad+
where I (y) = h ey) + 12 (y),
h ey) = { ( YluIP-2uaij (y)DijUdx ,
l Ri
12(y) =
y = « yl) - l,y'),
{ (YluIP- 2u[a ij (x ) - aij (Y)]Dijudx.
lai
By t he choice of ( we have t hat if (Y(x ) 1= 0, then 1 < yl x l < 1 + e/ 2 and yl lY' - x' i < c/ 2 implying that
yl < Xl < (1 + c/2)yl, Iy' - x' i < i/c/2 = (c/ 2)(x l /\ yl ), 0 < xl - yl < y le /2, Ix l - yl l < (e/2)(x l /\ ii) , [z - 111< c(x l /\ 11 1 ) , laij (x ) - aij(y)1:=:; (3. Hence,
and by (6.20) and Corollary 6.2
Ild1
2(y) dyl :=:;
+
We conclude
N(3I1MD2uIILp .9 1IM-I ulli~.~ ·
182
NICOLAI V. KRYLOV
To estimate h (y) we integrate by parts which is a very simple matter if p ~ 2 and is also justified for p E (1,2) in Lemma 2.17 of [12] . Then we find that h(y) = J1(y) + J2(y), where
J1(y)=- (p-1) J2(y) = -
r (YlujP-2aij(y)(D iu)Djudx,
JlR +d
r (Dj(Y)luIP-2uaij(y)DiUdx
JlR +d
=
_p-l
=
v:'
r (Dj(Y)a ij (y)DiluIP dx
JlR +d
r JlR
lulPai j (y)Dij(Y(x) dx.
d
+
As above
r
IJ1(y) + (p-1)
I
(YluIP-2aij(Diu)Djudxl:::; N (3
JlRi
r
(Ylu I P- 2IDuI 2dx,
JlRi
kd J1(y) dy + (p - kd 1)
+
(xlyr+1 luIP-2aij (DiU)DjUdxl :::; N(3Q ,
+
wher e Q is taken from Corollary 6.2. It follows that
I
kd J1(y) dy + (p -1) kd (Xlyr+lluIP-2aij(Diu)DjUdX! +
+
:::; N(3I IMD2uIILp ,eIIM-lulli~,~ +N(3(r2
+ li DII M - 1u ll i p, e'
Therefore, (6.21) yields I :::;N(3(IIMD2u IILp ,eI IM-lulli:,~
-(p-1)
+ (·l + 1)IIM-1ullip,e)
r (xlyr+l luIP-2aij(Diu)Djudx+ JlRir J2(y)dy. JlRi
(6.22)
Similarly, to the above estimates
jh(y) -
p-l
kdlu(x )IPaij Dij(Y dxl :::; N(3 kd lu1PjD 2( YI dx. +
+
To estimate ID 2(Y I introduce the function
e(z,y)
= (zl )'Y+2((zl,yly' _ z' )(yl )d-"Y - 3 .
Obviously, (Y(x) = e(y1x, y) . Therefore,
ID 2( Y(x )1 = (yl)2 1(D;e)(y 1x , y)\, I(D;e)(z, y)1 = I(D 21])(zl, yly' _ z') I(yl)d-"Y- 3 , ID 2(Y(x) 1= (yl )d-l-"YI(D 21])(X1yl , yl(y' - x'))I .
(6.23)
183
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
In particular , owing to (6.19)
where we have used Fubini's theorem which is possib le since the rig ht-hand side of (6.24) is finit e. Finally, owing to (6.20)
r
JRi
D ij(Y(x) dy
= D ij
r
JRi
(Y( x) dy
= D ij (X1)'Y+1 = -y(t + 1)(Xl p -l Jl iJl j ,
so that
By combining this with (6.22) we come to (6.19) and t he lemma is pr oved . 0 7. E x istence and u n iq ueness for parabolic PDEs in h alf s p a ces. We t ake pE (1, oo ),
TE (O , oo ),
d - 1
and prove the following version of Theorem 2.14 of [9]. THEOREM 7.1. Th ere exist constants {3 E (0, 1] and N <
(7.1 )
00,
depend-
ing only on 150 , d, p, and B, su ch that if
lRi,
whenever x, y E Ix-yl ::; X1 l\y l , i, j = 1, oo ., d, t E (O, T ), then for any f E M -1JLp,o(T) there exi sts a unique u E J) ~, o ,o (T ) satisfying (1.1) in Furthermore, fo r any u E J)~,o,o (T) we have
lRi ·
REM ARK 7.1. Part of condition (7.2) could have been repl aced with the requirement that for an e E (0, 1]
l a~j (x ) - a~j (y) j ::; e{3
184
NICOLAI V. KRYLOV
for all x , y satisfying Ix - yl ::; .s(x 1 /\ yl) and all i, i, t . However, this would not make the theorem more general as follows from the argument we used to obtain (6.15) . REMARK 7.2 . Theorem 7.1 is used in [9] to prove the solvability of parabolic equations in Sobolev spaces with weights in bounded Cl-domains. The proof for spaces like 5J;,o(T) with n = 2 is quite standard and is based on partitions of unity and flattening the boundary. Then equations are solved in small neighborhoods of the boundary, where conditions like (7.2) allow one to treat equations with quite irregular a i j and blowing up band c. The reader is sent to [9] for details. . To prove the Theorem 7.1 we need a lemma, in which G is a domain in ]Rd and by Lp ,loc(G) we mean the set of functions u such that u¢ E Lp(G) for any ¢ E CO'(G). LEMMA 7.1. Let Ut and ft be measurable junctions given jar t ::; T with values in Lp, loc(G). Assume that jar a continuous junction 7j; = 7j; (x ) given on G we have 7j; > 0]
and for t ::; T and each ¢ E
CO' (G)
we have
(7.4) Then
Proof. The inequality in (7.5) follows by Holder's inequality:
Next, we reduce the general situation to the one in which Ut has compact support in G . Let ( E CO'(G) and ( :::: O. Then
If (7.5) is true for UtC then iT
fc
(P7j;p-2IUtlp-2Udt dxdt = p-l
fc
7j;p-21(UTIP dx.
(7 .6)
185
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
The above estimates show that if we take ( = (n i 1, then (7.5) would follow from (7.6). Therefore, in the rest of the proof we assume that there is a compact set reG such that Ut(x) = ft(x) = 0 if x .;. r. In that case the values of'ljJ outside I' become irrelevant and we may assume that G = ]Rd, 'ljJ is bounded away from zero and infinity, Ut, ft are measurable and Lp-valued, and
We now take a nonnegative ( E eff (]Rd) which integrates to one and for an x E ]Rd substitute .s-d((C1(x - .)) in place of ¢ in (7.4) . Then by using the notation V(E) (x) =
s.r v(x-.sy)((y)dy
we find that for any x E ]Rd and t ::; T
U~E)(X) =
it f~€)(x)ds.
Then we have
p-llu~\xW =
iT
IU~E)(x)IP-2u~E)(x)f~E)(X)ds ,
Now we integrate through this relation over x E ]Rd after multiplying it by 'ljJp-2(x). It turns out that we can interchange all integrals since , say u~€)(x) is infinitely differentiable in x, measurable with respect to t, by Minkowskii's inequality (7.7) and the estimates as in the beginning of the proof are valid. conclude that iT
Ie 'ljJP-2Iu~E)IP-2u~E) f~€)
dxds
= p-l
Ie 'ljJP-2Iu~)IP
dx
Then we
(7.8)
We also observe that (for any t :::; T) we have U~E) ....... Ut in L p. Similar relation holds for ft. In particular, U~E) ....... Ut in measure
p,( dtdx) = 'ljJp-2 dxdt on [0, T] x ]Rd. Furthermore, (7.7) and the dominated convergence theorem imply that
186
NICO LAI V. KRYLOV
Simil ar relati on hold s for
f . By Lemma 6.1 and Corollary 6.1
lu(e) IP
lu(e)I P -
1
-t
lul
lulP in p
L 1(f-L),
1
in L pl (p- l )(f-L) , p 2u lu(e)IP-2u (e) - t lul in L pl (p_1) (f-L ), lu(e)IP- 2u (e)f ee) - t lulp- 2uf in L1(f-L ). -t
-
T hus , passing to t he limi t as e - t a in (7.8) lead s to (7.6). T he lemma is proved. 0 Proof of Theorem 7.1 . Rather elementary arg ume nts (cf. t he proof of Lemma 5.7 of [12]) show that if f E Cgo(JR+ x JRi), then the uni que class ical bounded solution of au/at = 6.u + fin JR+ x JRi with zero initi al and boundar y condit ion belongs to f:J ~ , e , o (T) . This gives us a starting point in t he method of cont inuity and , du e to the fact that f:J ~, e , o (T ) is a Ban ach space, shows t ha t t o prove the t heorem we need only prove t he apr iori estimate I luI ISj~ ,o(T) ::; N IIM (L u - au/ at )lllLp,o(T)
for any u E f:J ~ , e , o (T ) , where and below we denote by N generic constants dep ending only on 00, d, p , e, K . We will keep track of (3 up until we will see how to choo se it. By the definition of the norm in f:J ~ , e , o (T) we have
and
II M au/ at lllLp,o(T) ::; IIM (L u - au/at)liJLp ,o(T)
+ 11M Lul llLp,o(T),
wh ere by our ass umptions on b and c
with
F inally, by recalling t hat by Rem ark 4.2 the left-hand side of (7.3) is equivalent t o we see that we need only estimate jI l p in t erms of t he right-hand side of (7.3) . By Corollary 5.1 we have
iv»,
11M D 2u lllLp,o(T) + IIDulllLp,o(T) ::; N IIM fll lLp,o(T) + N II M -1ul llLp ,o(T), where f
= (a/at - L )u.
(7.9)
ON SPDES IN SOBOLEV SPACES WITH W EI GHTS
187
Next, we use Lemma 7.1 with G = lR~ , 1jJ = (Xl )-1, and M ((i - dJ/PUt and M (B-dJ/p ft in place of Ut and ft , resp ectively. We also obs erve that
Also
111jJ-l M (B -dJ/P(Ltut + ft)IILp (GJ = :::; NI IMD 2UtIILp,o + IIMlbtlDutllLp,o+
IIM (Lt ut + ft)IIL p,o IIM 2Ct (.!VI- IUt)liLp,o + II M f tII Lp,o'
so that by our assumptions and Remark 4.2
It follows that Lemma 7.1 is applicable and by noting that
1jJp-2I M( B- dJ/PUt lp- 2(M( B- dJ/PUt )(M (B- dJ/P(LtUt + ft)) = 1jJp- 2MB- dl ut lp-2 (L tut + ft) = M 1'+1Iutl p- 2(L t ut + ft), where I
=
e-
d - p + 1, we obtain
Observe that by (7.2)
2Iutlp- 2Ut(Ltut + it) :::; I Ut I P - 2 ut a~j DijUt + N j3M-Ilut IP-l IDuti +Nj3M-2Iut IP + Nlutlp-1 Ift l, and by Young's inequ ality
M-I IUtlp-lIDUtl = (M-2 (p-I J/Plut IP-1 )(M(P- 2J/ PI Dutl) :::; M-2 1ut1P + Mp- 2I DutIP, IUtlp-Ilft l :::; j3M - 2Iut IP + j3 1-p M 2(p -I JIf t iP. Furthermore,
188
NICOLAI V. KRYLOV
Therefore , coming back to (7.10) we get
By combining this with Lemma 6.5 we obtain
N (3J + N (3I-PIIMfl lr p,o(T)
+ p-I 'Y('r + 1) iT
ld
(xl)'-la:l lutIP dxdt
+
r
- (p-1) {
JJR~
Jo
(xl)'+l lutlP-2a~j(DiUt)DjUtdxdt;::: 0.
Lemma 6.4 with 8 E (0,1 ] and Corollary 6.2 allow us to est imat e the last t erm and we conclude that for any K E (0,1]
N(K - 18 + e NoT +v
l
2 /
(3 )J
+ N (3 -PIIM fl lt
,o(T)
r r (xl )'- la: l lutI dxdt ;::: 0, I; JJR~
(7.12)
P
where
v = p-I 'Y ('r + 1) - (p - 1)(1 - K)"(2p-2 = Kp-2(p - 1)"(2 + p-2(B - d - p + l)(B - (d - 1)) .
°
If we consider u as a function of K , then by virtue of (7.1) we have v < if K = 0. It follows that there exists a K = K(B,p, d) > such that v = - l/N , where N = N(B,p, d) E (0, (0) . Then (7.12) and the assumption that all ;::: 00 imply that
°
By using (7.9) we conclude
2
1 :S:; N I(8 + e N,o- ' / (3)1 + N(8
+ eN,O-, /2 (3 + (3 - P)II M f llt ,o(T)' (7.13)
Now we can specify the choice of 8 = 8(B, d,p, 80 , K), (3 (0, 1). We take them so that
s.s : 1/ 4,
= (3(B, d,p, 80 , K)
E
Nle N,o-, /2 (3 :s:; 1/ 4
and then (7.13) yields
which along with (7.9) lead to (7.3) and the theorem is proved.
D
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
189
8. Existence and uniqueness for SPDEs in half spaces. In t his section 7 is a fixed stopping t ime, p ~ 2. We start wit h a "conditional" res ult we were talking about in the Introdu ct ion. Here we only have a "nat ur al" ass umption on e. Of course, we are working in the setting of Sect ion 2. T HEOREM 8 .1. Let
d- 1
< e < d - 1 +p.
Assume that for a constan t K E (0,00) we have
Let 13 be the sma llest of the constants called Theorem 7.1 and assume th at
130 in Theorem 4.1 and 13 in
whenever x, y E lR.~, Ix - yl S; xl 1\ y 1, i,j = 1, ... , d, t E (0,7) . Fix an E lLp •Ii (7). A ssum e that there is a constant No < 00 such that f or any
f
(8. 1)
we have the apriori estimate
provided th at
in lR.~ (estimate (8.2) is not supposed to hold if there is no solution u E 5)~.1i ,0(7) of (8.3)). Then for any 9 E lHI~,1i(7) there exis ts a unique u E 5)~,1i,0( 7) satisfying (8.3) with A = 1. Furtherm ore, f or this solution
where N depends only on d,p, e, 60,61, K , and No. Proof. We know that 5)~, 1i , 0 (7 ) is a Banach space. From Theorem 7.1 we also know t hat for A = eq uation (8.3) is uniquely solvab le in 5)2p , e,0(7) . T here fore , by t he method of continuity to prove t he un iqu e solvab ility of (8.3) for A = 1 we only need to show that t here is a constant N such t hat for any objects in (8. 1) we have
°
(8.5)
190
NICOLAI V. KRYLOV
provided that (8.3) holds in ~t. By the definition of the norm in S)~,o ,o(T) we hav e
IlullSjz (r) = IIM- 1uIIIHIZ p, e
p ,o
(r)
+ IIM(Lu + f )IIILp,o(r ) + .xIIAu + gllIHI'
p,()
(r) ,
where as in the proof of Theor em 7.1
Also by our assumptions on a and v
IIAu + gilIHI1 oCr) p,
:::;
IIgII HI1oCr ) + N( IIM- 1u II ILp,0(r) + IIDu II ILp ,o(r) + 11M D2u IIILp,0(r)) ' p,
By combining this with Remark 4.2 we see that to prove (8.5) it suffices to prove (8.4) or else that IIM- 1u II ILp,0(r ) + IIDuI IILp ,o(r) :::; NI IMf II ILp,o(r)
+ II M D 2uIIILp,0(r )
+ N llgIIIHI; ,eCr)'
However, the lLp,o(T)-norms of MD 2u and Du ar e estimated through the lLp,o(T)-norm of M- 1u in Corollary 4.1 and the lat t er admits an est imate by assumption (8.2). The theorem is proved. D Here is a vers ion of one of the results of [7] in which we assume that for a const ant J E (0, 1] and all t and unit .x E ~d we have
ij (a t - C/t j ).xi.xj
>J_l_( ~a1j.xj)2 11 t ,
-
at
~
(8.6)
j
REMARK 8 .1. In light of (2.1) one can always take J = 01 in (8.6) becau se for the positive definite matrix (a~j) and TJ = (1,0, ..., 0) it holds that
(L aij.x j)
2
= (L
J
a~jTJi.xj) 2 :::; (a~jTJiTJj)a~j »» = ai1a~j »» .
J
On the ot her ha nd sometimes J = 1 and 01 may be less t ha n 1. This happens , for instance, if aij == 0 for all j and ai j == 0 for j =I- 1. THEOREM 8.2. Let (1.3) be satisfied and assume that for a constant K E (0,00) we have l
'
1 2
[z Da;( x) IRz + I(x ) DVt(x) IRz :::; K.
W e assert that th ere exist constants (3 > 0 and N < 00 depending only on oo,ol,J,d,p,(), and K , such that if
la~j (x) - a~j (y)1+ la;(x ) - a;(y)I Rz +IX1b~(x) 1 + IX1Vt(x)IRz + l(x 1)2 Ct(x)1 :::; (3,
(8.7)
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
191
wh en eve r x , y E JR~, Ix - y l :0:; Xl A y l , i , j = 1, ... ,d, t E (0,7) , th en for any f E M - llLp ,0(7), 9 = (gk, k = 1,2, ...) E JH[~,0(7) th ere exists a unique u E SJ ~, 0 ,O(7) satisfying (1.2) in JR~. Furthermore,
R EMARK 8 .2 . For d = do = 1 an exam ple of (J suitable for Theorem 8.2 can be obtained in the following way. Take a smooth fun ction f ~ 1 on (0, 00) such that f(x) = -ln x for small x and f(x) = 1 for x large . Then Ix!, (x )1 is bounded , say by a constant No . Next, for b E (0 ,1) introduce (J(x) = cos f O(x ). Then
1(J/(x)1 = bfO-l( x)lf'(x)llsinfO( x)I :O:; N ob x-l, and if a < y < x and Ix - yl :0:; x A y , then betw een x and y , so that y :0:; t;, we have
Ix -
yl :0:; y and , for a point
t;
which can be made arbitrarily small if b is chos en appropriately. R EMARK 8 .3. One could have introduced the parameter e as in Remark 7.1 and as there this would not make the theor em any more gener al. R EMARK 8.4. Theorem 8.2 is used in [7] to prove the solvabilit y of SPDEs in Sobolev spaces with weights in bounded Cl-domains. The proof for sp aces like 5);,0(7) with n = 2 is quite standard and is based on partitions of unity and flattening the boundary. Then equat ions are solved in sm all neighborhoods of t he boundary, where conditions like (8.7) allow on e to treat equat ions with quite irregular a ij , (Jik and blowing up b, c, and v (cf. Remark 4.4) . The reader is sent to [7] for details. To prove Theorem 8.2 we need the followin g counterpart of Lemma 7.1 in which we use the same notation. LEMMA 8.1. L et Ut = Ut (w), it, and gt = (g[, g; , .. .) be predictable proces ses given for t < 7 with values in Lp ,loc(G) . Assume that for a con ti nuous function 'IjJ = 'IjJ (x ) given on G we have 'IjJ > 0,
and
du;
=
ft dt
+ g~ dw~
in G for t < 7 in the sense that fo r each ¢ E Co(G) (a. s.) for all t E [0, 7) at once w e have
192
NICO LAI V. KRYL OV
Then
E
iTfc
,¢P- 2[2IUtlp- 2Ud t + (p -
1) IUt IP-2 Igtl~2 ] dxdt ~ O.
(8.8)
Proof. We mimic t he proof of Lemma 7.1. Fi rst , we obse rve t hat t he left-h an d side of (8.8) makes sense, since by Holder 's inequality (p ~ 2)
iTfc l'¢Ut IP-2Igtl ~2 (E iT I '¢utll~p(G) 1-2/p(E iT Il gtll~p(G) E
:'S
dt
dt)
dt )2/p
and t he t erm containing f t is taken care of as in the pr oof of Lemma 7.1. Next, as in that proof we redu ce t he general sit ua t ion to t he one in which Ut has compact support in G. More pr ecisely, in the rest of the proof we ass ume that there is a compact set I' c G such t ha t Ut(x) = ft (x ) = g~ ( x ) = 0 if x rf. f . In that case t he valu es of '¢ outside I' become irr elevant and we may ass ume that G = ]Rd, '¢ is bounded away from zero and infinity, Ut , f t ,gt are pr edi ct abl e and Lp-valu ed , and
E
I' (1Iut ll~p+ Ilftll~ p+ Il gtll~ p) dt < 00. Jo
In add it ion one can repl ace here p wit h 2 if one re places 7 wit h 7 /\ T , where T is a constant from (0,00). T hen by using t he not at ion from t he pr oof of Lemma 7.1 we find t hat for any x E ]Rd with probability one for all t < 7
u~"')(x) =
it
f ;"' )(x) ds + i t g~"')k (x) dW:.
By It o's formula for any constant T E (0,00) (here we use again that p ~ 2)
(2/ p) IUS~T(X)JP =
i
T
f\ T [2Iu~"' )(x) IP-2u~"' ) (x) f;"') (x) ds
+(p -1)lu~"')(x) IP-2Ig~"')1 ~21 ds + m tf\T(x ), where mt(x) is a local martingale start ing at zero . Ob serve t hat t he lefthand side is nonnegative and
because of t he sa me reasons as in t he beginning of t he pr oof and t he estimat e
ON SPDES IN SOBOLEV SPACES WITH WEIGHTS
193
that follows from Holder 's inequ ality. Then we find that
Now we integrate through this relation over x E jRd afte r multiplying it by 1jJp- 2(x ). It turns ou t that we can interchang e all integr als since, say u~e)(x) is infinitely differentiabl e in x, measurable with resp ect t o other variables, by Minkowskii's inequ ality (8.9)
and the est imate s as in the beginning of the proof are valid. Aft er int egrating with resp ect to x and set ting T -* 00 we obtain
Now notice that (for any wand t su ch that t < 7) we have u~e) in L p . Similar relations hold for It and g t . In particular, u~e) --> measure
p(dwdtdx)
-* Ut
Ut
in
= 1jJp-2 P(dw)dtdx
on (0,7n x jRd . Furthermore, (8.9) and the dominated convergen ce t heore m imply t hat
Simil ar relations hold for
I
Iu(e) IP- 2
lu(c)
and g. By Lemma 6.1 and Corollary 6.1 -->
IP- 2 Ig(e)
lulp - 2
1£2
L p/ (p-2 )(p) ,
in
2 -* luI P -
Igl£2 in t., (p ).
The t erm with I (c) in (8.10) is t aken care of in the sam e way as in the pr oof of Lemma 7.1 . Thus, passin g to the limit as c -* 0 in (8.10) leads to (8.8). The lemma is proved . 0 Proof of Theorem 8.2 . By Theorem 8.1 we onl y need to find (3 = (3 (00, 01,5, d,p, B, K) E (0 , 1] such that if the condit ions of the theorem are sat isfied with this (3, then (8.2) holds with No dep ending only on 00 ,01,5 , d,p, B, and K , whenever (8.1) and (8.3) ar e sat isfied. Ther efore, we concent rate on estimating 'f l / p .-
IIM- l u II IL
p
,9(T) '
194
NICOLA I V. KRYLO V
lRt,
In Lemma 8.1 we take G = 'ljJ (x ) = (xl) -l , and rep lace ut, lt , gt wit h M (£J-d)/PUt , M (lJ-d)/P(Ltut + I t ), M (£J-d)/P(Atut + gt), respectively. Vie also observe t hat as in t he proof of Theorem 7.1
11'ljJ - I M (O -d)/P (Ltut + I t)IILp(G) ::; NI IM - IUt IIH;,o + 11M I tlILp,o, where and below we denote by N generic constants depending only on 8o,8 1 ,d,p,B, K . We keep t rack of j3 up until we will see how to choose it. Finally,
IIM(O -d)/P(A t Ut
+ gt)IIL p(G)
IIAtut + gtllLp,o ::; N ll Dutll Lp,o + IIMlIt(M-Iut)IILp,o+ Ilgt llLp,o ::; N IIM- Iut IIH2 p ,o + IlgtllL o : =
p.
It follows that Lemma 8.1 is applicable and noting that
'ljJP- 2I M( O- d)/put lp-2 (M( O- d)/pud (M(O - d)/P(LtUt + It ))
= 'ljJp- 2M O-dl utIP-2(L tUt + It) = M ,+1lutlP-2(L tut + I t), 'ljJ p- 2I M (O- d)/putlp- 2IM (O- d)/P(Atut + gt )I;2
= 'ljJ p- 2M O-dl utlp- 2IAtut + gt l;2 = M ,+1lutlP- 2IAtut + gt l;2' where ,
=
B- d - p
+ 1, we obtain
E r r (x l )'+1 [2IUt /P-2Ut( LtUt + I t) Jo lrr{t p+ (p - 1)/utl 2I A tut + gtl;2] dxdt 2: O.
(8.11 )
Next , set
Jl / p := 11M D 2u II Lp,o(T) + II DuIILp,O(T ) + IIM- IuIIlLp,o(T) and note for t he future t hat by Corollar y 4.1 we have
JI / p ::; NIIM I IILp,O(T) + N ll g lllHI~ , o (T)
+ N tv».
(8.12)
We deal with the t erm containing LUt + I t in (8.11) as in t he proof of Theorem 7.1 and conclude that
N j3J + Nj3l- PIIM/II[
(T) p,o
+ E r r (xl ),+l [IUt IP- 2Uta~j o ;»,
Jo JJRd +
+ (p -1 )lutlp- 2IAtUt + gtl;J dxdt 2: O. By combining t his with Lemma 6.5 we obtain
-(p- 1)E
l' r (xl)'+ l lut IP-2[a~j(DiUt)Djut - IAtut+ gt l;2 ] dxdt Jo JJRt +p-l , (/+ l) E r
r (xl ),- lai 1IutJPdxdt
Jo JIRt
+ N j3J
+ N j3I-PIIM/ llt
,o(T ) 2: O.
(8.13)
195
ON SPD ES IN SOBOLEV SPACES WITH W E IGHTS
Her e
2 ij 2 2 2 IAtUt + gtl£2 = a t (DiUt)DjUt + IVtl£21Utl + Igtl£2
+ 2(O"ti ,Vt)£2 UtDiUt
+2(0":' gt)£2DiUt + 2(vt , gt)£2Ut· By our ass umpt ions ij
ij
-
11 -1
Ij
2
(at - at )(DiUt)DjUt ;::: J(at) (at Djud , IVt l;2 !UtIP::; ,BM - 2IUt IP, I U t l p- 2 ( 0"~, Vt) £2UtDiUt
::; N ,B(M- 2IUtIP
< N ,BM-IIUt !p-l IDUtl + M p- 2I DUtI P) ,
IUt IP-2 (O"~ , gt)e2DiUt < NIUt Ip- 2IDUt I!gt 1£2 ::; N ,Bl- pM P- 21gt 1 ~2 + N ,BM (2- p)/(p - l) IUt IP(p- 2) / (p- l ) IDUt IP/(p- l) ::; N ,Bl- pMP- 2 Igtl ~2 + N ,B(M-2IUtIP + Mp- 2I Dutl P) , pIUtlp- 2(Vt ,gt) £2Ut < N ,BM- 1I Utl 1Igtl £2 ::; N ,BM- 2IUtI P + N,BMP-2 Igtl ~2' By usin g the computations in (7.11) , adding that
E r r (x 1p +l MP-2 Igtl~2 dxdt =
Jo JlRt
Ilgll~
OCT)' P,
and con centrating on ,B ::; 1, we infer from (8.13) that
N ,BJ + N,B-P( IIMfll ~p , O (T)
+ IIg ll ~p , o (T))
b + l)E r r (xl)'Y-lail lutlP dxdt
+ p-l i
Jo JIRt
r (xlp+llutIP-2(ail)-I(aij D j Ut)2 dxdt > 0.
- (p -l)JE r
t; JlRt
Lemma 6.4 with J E (0,1 ] and Corollary 6.2 allow us to est imate the last t erm and we conclude that for any K, E (0,1]
N(K,-
IJ
+ eN 8-
1 /
2 ,B)J
+ N,B-P( IIMfl l~p , O (T) + Ilgllt,o(T) )
+vE
t'
r (xlp-laillut lpdxdt;:::O ,
(8.14)
Jo JIRt
wh ere
v = p-l i + p-2 [p(1 - J)
b + 1) - (p - l)J(l - K,)J2 p-2 = K,p-2(p - 1)J2J
+ J](O -
d - p + 1)
(0 - (d - 1) _ p[l _ p(l - 1_J) + J-1).
196
NICOLAI V. KRYLOV
If we consider v as a function of n; then by virtue of (1.3) we have v < 0 if", = O. It follows that there exists a '" = ",(B,J,p,d) > 0 such that u = -l iN , where N = N(B, J,p,d) E (0,00) . Then (8.14) and the assumption that a 11 ~ 50 imply that
By using (8.12) we conclude
I:::; N 1(5 + eNID-I /2 (3 )I
+ N(5 + eNID -I /2 f3 + f3- P )(IIM f llt .e(T) + Ilgll~~.O (T))·
(8.15)
Now we can specify the choice of 5, f3 E (0,1) depending only on
B,d,p,50,51 ,J, and K. We take them so that
and then (8.15) yields
which is (8.2) and the theorem is proved.
o
REFERENCES [1] Z. BRZEZNIAK, Stochastic partial differential equations in M-type 2 Banach spac es , Potential Anal. (1995) , 4(1) : 1-45. [2] Z. BRZEZNIAK, On stochastic convolution in Banach spaces and applications , Stochastics Stochastics Rep . (1997), 61(3-4) : 245-295. [3] G. DA PRATO AND A. LUNARDI, Maximal regularity for stochastic convolutions in LP spaces , Atti Accad. Naz . Lin cei Cl. Sci . Fis. Mat . Natur. Rend . Lincei (9) Mat. Appl. (1998) , 9(1) : 25-29. [4] G. DA PRATO AND L. TUBARO eds , "St ochas t ic partial differential equations and applications-VII" . Papers from the 7th Meeting held in Levico Terme, J anuary 2004 , Lecture Notes in Pure and Applied Mathematics, Vol. 245 (2006), Chapman & Hall/CRC, Boca Raton , FL. [5] F. FLANDOLI, Dirichlet boundary value problem for stochastic parabolic equations: Compatibility relations and regularity of solutions , Stochastics and Stochastics Reports (1990), 29(3) : 331-357. [6] M . HAIRER AND J .C . MATTINGLV, Ergodicity of the 2D Navier-Stok es equations with degenerate sto chastic forcing, Ann. of Math. (2), (2006) , 164(3) : 993-1032. [7J KVEONG-HUN KIM, On stochastic partial differential equations with variable coeffi cients in C1 domains , Stochastic Process. Appl. (2004), 112(2) : 261-283. [8] KVEONG-HuN KIM AND N.V. KRVLOV, On SPDEs with va riable coefficients in one space dimension, Potential Analysis (2004) , 21(3) : 209-239. [9] KVEONG-HuN KIM AND N .V. KRVLOV, On the Sobolev space theory of parabolic and elliptic equations in C 1 domains, SIAM J . Math. Anal. (2004), 36(2) : 618-642.
ON SPDES IN SOBOLEV SPACES W ITH WEIGHTS
197
[10J N .V. KRYLOV, On Lp-theory of stochastic partial differential equations in the whole space , SIAM J. Math. Anal. (1996), 27(2): 313-340. [l1 J N .V . KRYLOV , An analytic approach to SPDEs, pp. 185-242 in Stochastic P artial Differential Equations: Six Perspectives, Mathematical Surveys and Monographs , Vol. 64, AMS, Provide nce , R I, 1999. [12J N .V . KRYLOV , We ighted Sobolev spaces and Lap lace 's equation and the heat equ ations in a ha lf space , Comm. in PDE (1999) , 2 4 (9- 10): 1611-1653. [13J N.V. KRYLOV , Ma ximum principle fo r SPDEs and its app lications , in "St ochastic Differential Eq uations : T heory and Applications, A Volume in Honor of P rofessor Boris L. Rozovskii", P .H . Baxendale, S.V . Lototsky, eds ., Interdi scipli nary Mathematical Scie nces, Vol. 2 , World Scientific, 2007 . http:/ /arxiv .org/math .PR/0604125 . [14] N .V . KRYLOV AND S .V. LOTOTSKY, A Sobo lev space th eory of SPDEs with constant coefficients on a half lin e , SIAM J . Math. Anal. (1999) , 3 0 (2) : 298-325. [15J N.V . KRYLOV AND S .V . LOTOTSKY, A Sobo lev spa ce theory of SPDEs with con stan t coeffi cients in a half space, SIAM J . Mat h. Anal. (1999) , 31 (1) : 19-33. [16J S .B . KUKSIN , R emarks on the balan ce re lations for the two -dimensional NavierStokes equation with random forcing , J . Stat. Phys. (2006), 1 2 2(1) : 101-114. [17] S .V . LOTOTSKY, Di ric hlet problem for stochastic paraboli c equations in smooth domains , Stochastics and Stochastics Reports (1999) , 68 (1- 2): 145- 175. [18J S.V. LOTOTSKY, Sobo lev spaces with weights in domains and boundary value problems for degenerate elliptic equations , Methods a nd Applications of Analysis (2000), 1 (1) : 195- 204. [19] S .V . LOTOTSKY, Linear sto chastic para bolic equations, degenerating on the boundary of a domain, E lectron. J. Probab. (200 1) , 6 (24) : 14. [20J S . LOTOTSKY AND B. ROZOVSKII, Wiener chaos solutions of linear stochastic evolution equ ation s , Ann. Probab. (2006) , 34 (2) : 638-662.
STOCHASTIC PARABOLIC EQUATIONS OF FULL SECOND ORDER* SE RGEY
v.
LOTOTSKyt AND BORIS L. RO ZOVSKII+
A bst r act. A procedure is described for defining a generalized so lution for stochastic differential equations using t he Cameron-Martin version of t he W ien er C haos expansion. Existence and uniqueness of this Wiener Chaos so lution is established for parabolic stochastic PDEs such that both the drift and the d iffusion operators are of the second order.
1. Introduction. Conside r a stochastic evolution equation
du(t)
= (Au(t) + j (t ))dt + (Mu(t) + g(t))dW(t),
(1.1)
where A and M are differential operators , and W is a Wie ner process on a probability space (n, F , JP». Tr aditi on ally, t his equation was stud ied under t he followin g ass umptions : • AI. T he operator A is elliptic, t he order of the ope rator M is less t han t he order of A , and A - ~ M M* is ellipt ic (possibly deg enerate) ope rator , In fact, it is well known t hat unl ess ass um ption Al hold s, E quati on (1.1) has no solut ions in L 2 (n ; X ) for any reason abl e choice of t he state space X . It was shown recentl y (see [4, 5, 6] and t he referen ces t here in) t hat if only t he operator A is elliptic and t he order of M is sm aller than t he order of A , t he n t he re exists a unique generalized solution of E qu ati on (1.1). This solution is oft en referred to as Wie ner Chaos solution . It is given by t he Wi ener chaos expansion u (t) = 2: 01<00 U o (t) ~o , where { ~o }l o l < oo is the Cameron-Martin ortho normal basis in the space L 2 (n ; F W ; X ) of square integr able random elements in X measurable wit h resp ect to t o t he sigmaalgebra F W generated by t he W iener process. The Cameron-Martin basis {~o } is indexed by mul t iind ices a = (aI, a2, ...) . It was show n t hat for certain positive weights Q = {q (a)}lol
I lu l l~,x:=
L
q2 (a)
IluoIIL «o,T);X) < 00 ,
101 <00 *Sergey V . Lototsky acknow ledges support from NSF CAREER award DMS-0237724 . Bo ris 1. R ozovskii acknow ledges support from NSF Grant DMS 0604863, ARO Grant W9 1lNF-07-1-0044, a nd ONR G rant N00014 -07-1-0044 . tDepartment of Mathematics, USC , Los Angel es , CA 90089 (l ot ot s ky0 math. us c . edu) , http : / /www- rcf. usc .ed u /~ lototsky. +Division of A pp lied Mathematics , Brown University, P rovid ence, RI 029 12 (r oz ovs ky0dam. brown . edu ). 199
200
SERGEY V. LOTOTSKY AND BORIS L. ROZOVSKII
where X is the appropria t e Hilbert space characterizing t he "regularity" of the solut ion . Note that without assumpt ion Al
E
L
IluIIL«O,T);X) =
Ilu a
(t)IIL«o ,T);X) = 00.
lal <ex:>
In this pap er , we consider the Cauchy problem for the following stoch astic partial differential equation:
du = (aijDiD ju + biD iu + cu + f)dt + (PijDiD ju + (TiD iU + VU + g)dW,
t E [0 , TJ, x E lR.
(1.2)
In contrast to the pr evious work , t his is a par abolic SPDE of the full second order , in that the drift and diffusion op erators have the same order 2. We const ruct a scale of weighted Wi ener chaos spaces (re late d but not identical to Kondratiev 's spaces) and prove the existence and uniquene ss of t he solution in the spaces from this scale.
2. Constructing a solution: an example. Let IF = (D, F, {Fdo::;t::;T, lP') be a stochastic basis with the usual assumptions and W =
W(t), 0:::; t :::; T , a standard Wiener process on IF. For a Hilbert sp ace X , denote by L 2( W ; X) the collect ion of X-valued random eleme nt s that are square integrable (JEll , IIi < 00) and are measurable with resp ect to the sigm a-algebra generated by W(t) , t E [0, T ]. Consider the Ito equation u(t, x) =e-
x2 2 / +
it
Uxx(s, x)ds +
it
Uxx(s , x)dW(s) ,
(2.1)
t E [O ,T] , x E lR.
If there is a solution , its Fourier t ransform in sp ace , u(t , y) (1/ J27f) fIR e-iXYu(t , x )dx satisfies
y2 u(t , y) =e- / 2 - y2it u (s, y)ds - y2it u( s, y)dW(t) ,
t
E
[0, TJ,
(2 .2)
y E lR.
For each fixed y, (2.2) defines a geom etric Brownian motion:
u(t , y) = e- (l+t)y2_(y4/ 2)t- y2W(t).
(2.3)
Let H'Y(lR ) be the Sobolev space
(2.4) Since
(2.5)
SECOND-ORDER SPDES
201
the solution of (2.1 ) cannot be an element of L 2(W ; L 2 ((0, T) ; H i(JR))) for any I E JR, even though the initial condition is non-random and is an element of H i (JR) for every I E R Let us try another approach. Once again, assuming that the solution exists, we apply the Ito formula to the product u(t , X)£h(t), where
and h = h(t) is a smooth deterministi c function . Since
(2.7) we conclude that the fun ction
(2.8) if defined , must satisfy the heat equation
(2.9) If SUPt Ih(t)1 < 1, then this equ ation has a unique solution in every H i(JR) and
Uh(t, x )
= lEexp( -(X(t , x))2 / 2),
(2.10)
wh ere
1 t
X(t, x ) = x +
J2(1
+ h(s)) dW(s).
(2.11)
In other words , while existe nce of a solution of Equation (2.1) is st ill unclear, we now have a family of functions Uh(t, x ) defined by (2.10) . All we need now is a systematic procedure of relating the family of det erministic fun ctions ui, = Uh(t, x) to a random pro cess U = u(t , x) ; then this process is natural to call a solution of (2.1) . Here is a possible way of constructing a stochastic process from Uh . Let m = {mk ' k ~ 1} be the Fourier cosin e basis in L 2( (0, T) ): (2.12) Then
h(t)
= L hkmk(t). k 2::1
(2.13)
202
SERGEY V. LOTOTSKY AND BORIS 1. ROZOVSKII
For every fixed t E [0, T ] and I E JR, we can now interpret the fun ction (t , .) as a mapping from the set of sequenc es h = (h 1 , h 2 , . . .) to the space H'Y(JR d ) , and , as equalities (2.10) and (2.11) sug gest, this mapping is analyt ic in the region {h : L:k>l h~ < s ] for sufficiently small c. We will now compute the derivatives ofthis mapping. Let :J be the collect ion of multi-indices a = {ak ' k :::: I} . E ach a E :J has non-negative integer element s ak and
Uh
lal = L
ak <
(2.14)
00 .
k
We also use the notation
(2.15) and consider special multi-indices, lal = 1, ai = 1. For each a E :J defin e
a=
(0) with
la[ =
0 and
a = e. , with
(2.16) Then
(2.17) where
(2.18) On the other hand, by direct computation,
(2.19) where
(2.20) and
(2.21)
203
SECOND-ORDER SPDES
is n-th Hermite polynomial. It is a standard fact [1] that the collection {~o: , a E J} is an orthonormal basis in L 2(W;lR). The functions uo:(t, x), a E J, uniquely determine Uh(t, x) according to (2.17) . On the other hand , if
I: Iluo:(t)llk'Y(lR) <
00 ,
(2.22)
I: uo:(t,x)~o:
(2.23)
o:E:J
then the H 'Y(lR)-valued random process
u(t,x) =
o:E:J satisfies IE(u(t ,X)£h(t)) then also
Uh(t,X) ; if, in addition, u
=
u(t,x)
FF-adapted,
IS
I: uo:(t,x)~o:(t) .
=
(2.24)
o:E:J If condition (2.22) fails, then (2.23) is a formal series, which we define to be the stochastic process corresponding to the family Uh . As (2.5) suggests, if Uh is the solution of (2.9), then (2.22) fails for every v, Let us now see how fast the series diverges. Equality (2.9) implies
U(O)(t ,x)=et
x 2/ 2
r 82u(o)(s,x) 8x2 ds ,lal=O;
+io
it 82u~~~s ,x) -! +
UEi(t)=l
82u~~:,x) ds +
Uo: t -
82uo:(s , x) d 8 2 S
()-it o
~
(2.25)
82UO:_ Ek(S ,X) ()d 8 2 mk s s , lal > 1.
LJ yak
k=I
X
mi(s)ds , lal = 1;
0
X
Equations of the type (2.25) have been studied [4, Section 6 and References]. In particular, it is known that
I: Iluo:(t)llk'Y(lR) = lo:l=n
t~ II D;ntuo IIk 'Y(lR),
(2.26)
n.
where D x = 8/8x, t is the heat semigroup, and uo(x) = e- x 2 / 2. To simplify further computation, let us assume that "Y = O. Then, switching to the Fourier transform, II
D 2n x
t
U 11 2 d) 0 L2(lR
=
i rf*. !y I ne4
y2
(t+I )dy
= r (2n + ~) (1 + t)2n
.
(2.27)
Using Stirling's formula for the Gamma function I', (2.28)
204
SERGEY V. LOTOTSKY AND BORIS L. ROZOVSKII
wher e the numbers C(n) ar e uniformly bounded from above and below. Similar result holds in every H'Y(IR) . Thus, (2.22) does not hold , but instead, by (2.28), we have
c:
'" ~ 1 l lua(t)IIW' 2 (IR)
<
00.
(2.29)
«e.r
We denote by (£)o,o(W ; H'Y(IR)) the collection of formal seri es (2.24) satisfying (2.29); the reason for using (£)0,0 in the notation will become clear later . Note that we had equalit ies in all computations for Equation (2.1) that lead to (2.29) , which suggests that (£)o,o(W ; H'Y(IR)) is the natural solution space for Equation (2.1) . For a more gener al stochastic parabolic equat ion of full second order in IR d , the natural solution space turns out to be (£)p ,q(W;L 2((0 ,T); H'Y(lR d ) ) ) for suitable p, q :::; 0. In the next section we address the following questions: 1. How to define the spaces (£)p ,q(W; X) for p, q E IR without relying on an orthonormal basis in L 2((0, T)) ? 2. How to construct a solution of a general stochastic parabolic equations of full second order? 3. General constructions and the main result. As before, let IF = (0" F, {Ft}o:'O t:'OT, ]]D) be a stochastic basis with the usual assumptions and W = W(t), t :::; T , a standard Wiener process on IF. Denote by H' = HS((O , T)), s ~ 0, the Sobolev spaces on (0, T) with norm
°:: ;
(3.1)
wh ere A is the op erator
(3.2) with Neumann boundary conditions. This norm ext ends to functions of several variables via the tensor product of the spaces HS. DEFINITION 3 .1. Given real numbers p, q and a Hilbert space X , (£)p ,q(W; X) is the closure of the set of X -valued random elements
with respect to the norm (3.4)
where each TJk, k ~ 1, is a smooth symmetric function from [0, T]k to X .
205
SECOND-ORDER SPD ES REMARK
3 .1. (a) It is known [2, 7] that, for TJ of the type (3.3) , N 2
",1
2
JEllTJllx = IITJollx + 6
k! 1IIITJklio
2
Ilx·
(3.5)
k =l
(b ) The definition of each individual (£) p,q(W j X) inevitably involves arbitrary choices , such as the norm in Hq((O ,T)) . Further analysis shows that different choices result in shifts of the indices p , q , and the sp ace Up,q(£ )p,q(W ; X) does not dep end on any arbitrary choices. In the white noise setting, where n is the sp ace S' (JR.d) of the Schwartz distributions and IF' is the normalized Gaussian measure on S , the inductive limit Up,q (£) p,q (W ; JR.) is the Kondratiev sp ace (S)-l [2]. R EMARK 3.2 . If X is the Sobol ev sp ace H'Y(JR.d), then we denote t he norm II . IIp,q;x by II ·lIp,q;'Y: (3.6)
PROPOSITIO N 3.1.
Let TJ = f~O!.! where
f
E X and ~O!. is defined by
(2.20). Th en 2 _ 210!.Ip 2qO!. 1ITJ IIp,q; X - ~N Ilf
2
ll x,
(3.7)
where
N 2qO!.
=
II k
2q O!. k .
(3.8)
k21
Proof. Let
c;u = ~ ya!
where
E O!.
10:1 = n.
It is known [3] that
T r t: ... ('2 E U(Sl, "" Jo Jo Jo
Sn)dW(SI) .. . dW( Sn_l)dW(Sn) , (3.9)
is the symmetric function E O!. ( Sl , . .. , Sn)
=
L
mi, (Sa (l )) ... m i n ( Sa(n )) '
(3.10)
aEPn
In (3.10) , the summation is over all permutations of {I , . . . , n}, the functions mi, are defined in (2.12), and the positive integer numbers i 1 :::; i 2 :::; . . . in are such that , for every sequence (bk , k 2: 1) of positive numbers,
II b~k = b s.; .... . i, .
k 21
bi n '
(3.11)
206
SERGEY V. LOTOTSKY AND BORIS L. ROZOVSKII
For example, if a = (1 ,0 ,2,0,0 ,4,0,0, . . .), then 1001 = 7 and i l i 2 = i 3 = 3, i 4 = ... = i 7 = 6. Thus, in the notations of (3.4) , we have
TJk =
~Ee< f,
v a! { 0,
if k
= n,
l,
(3.12)
otherwise.
Not e that (3.13)
By definition (3.2) of the operator A we have
o
The result now follows. COROLLARY
3.1. A formal series (3.15)
with TJe< E X , is an eleme nt of (£ )p,q(W; X) if and only if (3.16)
Proof. This follows from (3 .14) and the equa lity
(3.17)
o
Deno t e by (£) p,q (W) the Hilbert space dual of (£) p,q(W ; IR) relative t o the inner product in L 2( W ; IR) , and by ((' , .)) the corresponding du ality. In the white noise setting, n p, q(£ )p,q (W ) is the space (Sh of the Kondratiev t est functions [2]. If TJ E (£)p,q(W ; X) and ( E (£)p,q(W) , then ((TJ, ( )) is defined and belongs to X. For h E L 2((0, T)) , define
e; =
Eh(T) = exp
(iT
h(s)dW(s) -
~
iT
Ih (SW dS) .
(3.18)
PROPOSITIO N 3.2 . Th e random variabl e Eh is an eleme n t of (£ )p,q (W)
if an only if (3.19)
SECOND-ORDER SPDES
207
Proof. Since
(3.20) it follows that
Eh
= 1 + 2::= CXl
k=l
iT1 ...1 8k
0
0
82
h(Sk) ' " h(sddW(sd '" dW(Sk_ddW (Sk)'
0
(3.21) By (3.4) and (3.5), Eh E (£)p,q(W) if and only if (3.22) that is, Ilh ll~q < 2P .
D
DEFINITIO N 3 .2. W e say that the fun ction h is suffic iently small if (3.19) holds for sufficiently large (positiv e) -p, -q . PROPOSITIO N 3.3. If u E U p,q(£)p,q(W; X) and h is sufficiently
small, th en Uh
= ((u, Eh ))
(3.23)
is an X -valu ed analytic function of h . P roof. For every u E Up,q(£) p,q(W; X) , there exist p , q such that u E (£) p,q(W ;X) ; by Proposition 3.2, Uh will indeed be defined for sufficiently small h . Similar to (2.17) we have (3.24) and this power series in hCY. converg es in som e (infinite-dimensional) neighborhood of zero. D From now on, D; = a/aXi, and the summation convention is in force: c.d, = 2:i Ci di, et c. Consider the linear equat ion in lR d
with initial condit ion u(O, x ) = v(x) , under the following assumptions: BO. All coefficient s are non-random. Bl. The fun ctions aij = aij(t,x) , Pij = Pij(t , X) are me asurable and bounded in (t, x) by a positive number Co , and (i) laij(t, x)- aij(t ,y)I+lpij(t,X)-Pij(t,y) 1::;Col x-yl , x , y E lR d , 0 < t ::; T;
208
SERGEY V. LOTOTSKY AND BORIS 1. ROZOVSKII
(ii) the matrix (ai j) is uniformly positive definite, that is, there exists a 8 > so that, for all vectors Y E Jltd and all (t, x), aijYiYj ~ 81y1 2 . B2. The functions b, = bi(t , x), c = c(t, x), a, = O'i(t, x) , and v = v(t ,x) are measurable and bounded in (t,x) by the number Co . B3.
°
p ,q
(3.26)
p ,q
For simplicity, we introduce the following notations for the differ ential operators in (3.25) :
A = aijDiDj + biDi + C, B = pijDiDj + a.D, + t/.
(3.27)
DEFINITION 3 .3. A solution u of (3.25) is an element of Up,q(£)p,q(W;L 2((0,T); H1(Jltd))) such that, for all sufficiently small hand all t E [0, T], the equality
Uh(t, x)
= Vh(X) +
it
(A
+ h(s)B)uh(s , x)ds
(3.28)
holds in H-l(Jltd) . The following theorem is the main result of this paper. THEOREM 3.1. Assume that, for some p > and q > 1, Uo E (£)p,q(W; L 2 (Jltd)) and I , 9 are elements of the space (£)p,q(W; L 2((0, T) ;H-l (Jltd))). Then there exist r , < such that Equation (3.25) has a unique solution U E (£)r,e(W; L 2((0, T) ; H1(Jltd))) and
°
e
l
T11u (t )II; ,t ;l dt < C·
°
I (llf (t)II;,q;_l + Ilg(t)II;,q;-l) T
(IIvll;,q;o +
°
dt).
(3.29)
e,
The number C > depends only on 8, Co ,p, q, r , and T. Proof. The proof consists of two steps: first, we prove the result for deterministic functions v, i, 9 and then use linearity to extend the result to the general case. Step 1. Assume that the functions v E L 2(Jltd), t, 9 E L 2((0, T); H-l(Jltd)) are deterministic. Then Vh = v, fh = I, gh = g , and classical theory of parabolic equations shows that, for sufficiently small h, Equation (3.28) has a unique solution ui, and the dependence of Uh on h is analytic. As in the previous section, we write U(t,x)
=
L ua(t,x)~a aE:J
(3.30)
SECOND-ORDER SPDES
209
where the coefficients u'" satisfy
U(O)(t, x) = v( x)
+ it (Au(o)(s, x) + f (s, x))ds,
UEk(t ,X) = i t AUEk(S,x)ds + i t (BU(O)(S, X) + g(s,x))mds)ds , (3.31) u",(s, x ) = t Au", (s, x)ds +
Jo
L y(ik Jot Bu"'_Ek (s , x)mk(s)ds , lal > 1. k
Denote by cP = cPs,t, t 2': s 2': 0 the semigroup gen erated by t he op erator AIt follows by ind ucti on on lal t hat
T herefore, using t he usual par abolic estimates,
(3.33)
and t hen (3.29) follows from (3.16). Step 2. As in Step 1, existence and uni quen ess of solution follows from un ique solvability of t he par abolic equation (3.28), and it remains to establish (3.29). Denote by u(t, x; V, F, G, ')') , ')' E J , t he solution of (3.25) with v = V(p f = g = If v = L "'EJ v",f;,,,,, etc., t he n
rc;
ce;
U(t,x) =
L
u(t ,x;v-y,f-y,g-y,')') ·
(3 .34)
-yEJ
It follows from (3.31) t hat u",(t, x; V, F , G, ')')
u",+-y(t, x ; V, F, G, ')') J(a + ')')!
u'"
= 0 if lal < ill and
(t,x;~ , -ftr,-7rr' (0)) va!
(3.35)
210
SERGEY V . LOTOTSKY AND BORIS L . ROZOVSKII
Using the results of Step 1,
1 T
<
~ (1IV'1'IILcJRd) +
Ilu(t,·; v'1',f.y,9'1') 11;,£;1 dt
I
T
(1If'1'(t)II~-lCJRd) + 119'1'(t)II~-lCJRd»)dt).
Now (3.29) follows from (3.34) by the triangle inequality.
(3.36)
o
REFERENCES [1] R.H. CAMERON AND W .T . MARTIN. The orthogonal development of nonlinear functionals in a series of Fourier-Hermite fun ctions. Ann. Math ., 48(2): 385-392, 1947 . [2] H. HOLDEN, B . 0KSENDAL, J. UB0E, AND T . ZHANG . Stochastic Partial Differential Equations. Birkhauser, Boston, 1996. [3] K. ITo. Multiple Wiener integral. J. Math. So c. Japan, 3 : 157-169, 1951. [4J S .V . LOTOTSKY AND B.L. ROZOVSKII. Stochastic differential eq uat ions: a Wiener chaos a p proach . In Yu . Kabanov , R. Liptser, and J . Stoyanov, editors, From stochastic calculus to mathematical finance: the Shiryaev f estschrift, pp. 433507. Springer, 2006 . [5] S .V . LOTOTSKY AND B .L . ROZOVSKII. Wiener chaos solutions of linear stochastic evolut ion equat ions . Ann. Probab ., 34(2) : 638-662, 2006 . [6J R . MIKULEVICIUS AND B .L . ROZOVSKII . Line ar parabolic stochastic PDE's and Wiener Chaos. SIAM J . Math . Anal., 292: 452-480, 1998. [7] D . NUALART. Malliavin Cal culus and Related Topics, 2nd Edition. Springer, New York, 200 6.