-1- Unilateral small deviations of processes related to the fractional Brownian motion (FBM) Molchan G E-mail: molchan@mitp.ru St. Petersburg, September 12-18, 2005 1 Problem: find lim log pT / log T , T , T where pT P (t ) 1, t T , (t ), (0) 0 is a random process/field and T T is a convex domain, 0 . The interesting cases: self-similar (H-SS) processes with parameter of similarity H, HSS processes with stationary increments (H-SSSi); T (0, T ) and T (T , T ) . 2 Motivation Example 1. Sediment deposition model in geology [3]. The thickness of the sedimentary layer as a function of time is modeled as H-SSSi process (t ) . Due to erosion the sedimentary record is described as follows: h(t ) inf (s); s t. The question: what is Hausdorff dimension of episodes of deposition represented in the sedimentary record, i.e. dimsupport of dh(t ) D ? The answer: if ( s) FBM ( s) bs , b > 0 then D 1 ( FBM , T (0, T )) . Example 1(a). If M (t ) max( ( s) : 0 s t ) is the record function of FBM then dimsupport of dM (t ) 1 ( FBM , T (0, T )) . Example 2. 1-D Inviscid Burgers Equation ut uu x u xx , 0 . u (0, x) ( x) Solution: u (t , x) ( x a(t , x)) / t . Initial positions of particles which have not collisions up to moment t 1 form set S a : a(1, x), x R1 . -2- Question: dim S-? t Hopf ‘s algorithm: if c(a ) convex hull ( s ) ds t 2 / 2 0 then a (1, x) is inverse function to c' (a ) and S supp dc' (a). Conjecture (see [4]): if ( x) FBM then t 1 dim S ( s ) ds, T (T , T ) 0 Results M T sup (t ), t T , 3 Statistics: ZT sup t : (t ) 0, t T , GT arg sup (s), s T , S T 1 ( s ) 0 ds . T A typical trajectory of H-SS process (t ) with small value of M T perhaps have a small values of ZT , GT and ST . Therefore the following events AT , M T 1, ZT 1, |T 0 , GT 1, ST 1, GT T , are considered as characteristic in our problem. The following ( AT ) lim log PrAT / log T T is the index of AT . An example of strong correlation of characteristic events: if is Levy stable process, P( (1) 0) , 0 1 , then P(GT x | M T 1) F ( x), T where F (x) is non-trivial distribution, [2]. 4 Theorem 1. If (t ), (0) 0 is H-SS, stochastically continuous Gaussian process T H (T ) : T (t ) 1, t 1; T 2 o(ln T ) . ( H is the reproducing kernel Hilbert space associated with (s), s T ) then the indexes (M T ), (ZT ), (GT ) exist simultaneously and are equal. -3- t Examples: 1. Theorem 1 hold for FBM, IFBM FBM ( s) ds with intervals 0 T (0, T ) and ( T , T ) . In these cases one has ( A(M T )) ( A(Z T )) ( A(GT )) ( A(ST )) and T (s) 1, s 1, T (s) 0, s 1/ 2 , [1, 4]. 2. If T 0 is a compact convex domain in R d then Theorem 1 holds for the H-fractional Levy field but not for the H-fractional Brownian sheet. 3. If (t ) t e ts dw( s) , w is Brownian motion and T (0, T ) then 0 T (t ) t (1 T ) /(t T ), T 1/ ln T and T 2 O ( ln T ) . See (Dembo et al., J. A. M. S., 15, 857-892,2002) for applications of (t ) to random polynomials. Proof of Theorem 1 (sketch) pT P(M T 1, T ) P( (s) T 1 T , T ) E 1 1- ( ) where ( ) d ( T ) / d ( ) and is Gaussian measure associated with FBM on T .Applying Hölder’s inequality one has pT P( 1 T , t 1, T )1 exp ( T where 2 T 2 2 / 2 ) P( A( Z T ))1 T / 2 / ln T . P( A(ZT )) P( GT 1) , P( GT 1) P( GT 1, M 1 cT ) P(M 1 cT ) P(M T cT ) P(M 1 cT ) . Due to H-SS property of one has P(M T cT ) P(M T ' 1) where T ' TcT1 / H . If cT 2a ln T then P(M 1 cT ) o(T ac ) and T ' T 1o(1) . 5 H-SSSi Processes, T (T , T ) . Theorem 2, [1,4]. Let (t ), (0) 0 be a cadlag (right-continuous with limits to the left) H-SSSi-process and G arg sup (s), s (0,1). Then G has a continuous probability density (t ) in (0,1), ( x) x , (1 x) ( x) , 0 x 1 -4- i. e. 0 or 0 , if 0 then P GT 1, T (Ta, T (1 a) (a) /( 2T ) (1 o(1)), T , where 0 a 1 . Consequence: (H-FBM, T (T , T )) 1, 0 H 1 . Hint to Theorem 2 (). () (t 0 ), T (), T law ~ t 0 (Si property) law ~ GT ( ) GT ( (t 0 )) GT ( ) t 0 If t 0 aT and T (aT , T (1 a)) then T t 0 (0, T ) . Therefore ~ ~ P(| GT ( ) | 1) P(| GT ( ) aT | 1) P(| G ( , (0,1)) a | 1 / T ) (a) / 2T The second equality is the result of SS-property of . Theorem 3, [1]. If (t ), t R d is H-fractional Levy field i. e. is Gaussian and E | (t ) (s) | 2 | t s | 2 H then ( , T t T ) d . 6 H-SSSi processes, T (0, T ) . Lemma. Let (t ), (0) 0 be a cadlag SSSi process and for some 0 E max (t ) 1 ( 0,1) , E exp ( max (t )) , ( 0,1) then T I T E e ( s ) ds 0 1 HT (1 H ) E max ( s) (1 o(1)), T . (0,1) (see [8] for the case of Brownian Motion). Theorem 4 [1]. If (t ) H - FBM, T (0, T ) then P(ZT 1) LT I T P(M T 1) LT where LT and LT are slow variables. Therefore ( A(ZT )) 1 H ( A(M T )) and (H-FBM, T (0, T )) 1 H . Consequences. If (t ) H-FBM then dim (supp dM (t )) H . -5- duality of h-FBM with h = H and 1 – H: pTc P( (t ) 0, 1 t T | (s) 0, T s 1) T H o(1) pT P( (t ) 0, 1 t T ) T (1 H )o (1) and persistence: pTc pT if H 1 / 2 , pTc pT if H 1 / 2 . Conjecture. If (t1 ...t d ) is H-fractional Levy field and T (0 t i T , i 1,..., k ; | t j | T , k j d ) then ( , T ) d kH . This conjecture includes Th. 3, Th. 4 and it holds for the case H = 1 i. e. for (t ) t i i with i. i. d. Gaussian variables { i } . t 7 The integrated fractional Brownian motion, (t ) FBM ( s) ds . 0 The index ( , T ) exists for the cases T (0, T ), 0 H 1, T (T , T ), 1/ 2 H 1 . The proof is based on Li & Shao idea: if pT P( A(ZT )) then one has pTS pT p S , T , S 0 , due to positivity of E (t ) (s), t , s T , Slepian Lemma and SS-property of . ( , T (T , T )) 1 H , ([4]). Sketch of proof. Let c (t ) be the convex hull of (t ) t 2 / 2 then dim (supp dc (t )) H . Indeed, if t1 , t 2 S {supp dc} and | t i | k then c (t1 ) c (t 2 ) (t1 ) (t 2 ) t1 t 2 t1 t 2 H , 0 , a. s. Therefore dim( S ) H due to the Frostman’s Lemma. On the other hand if is index of A(M T ) for and T (T , T ) then dim ( S ) 1 . If H 1 / 2 then (see [6, 7]) 1 / 2, T (T , T ) . 1 / 4, T (0, T ) ( IFBM ) Numerical results [4, 5]. -6- T (T , T ) 1 H , . H (1 H ), T (0, T ) ( IFBM ) Conjecture: If IFBM , T (0, T ) then P(M T 1) cT H (1 H ) (log T ) ( H ) where ( H ) (1 H ) . 8 Non-power asymptotic. Theorem 5. If (t ) ( k cos kt k sin kt) f k exp( k / b) k 1 where 0 c1 f k c2 , b 0 and { k }, { k } are i. i. d. standard Gaussian variables then lim | log | 2 log P( (t ) t ) b . 0 This example came from analysis of Burgers equation with random force. 9 References. 1. Molchan G.M., Maximum of fractional Brownian motion: probabilities of small values. Commun. Math. Phys., 1999, 205:1, 97-111. 2. Molchan G.M., On a maximum of stable Levy processes. Theory of Prob. and its Applications, 2001, v.45, 343-349. 3. Molchan G.M., D.L. Turcotte, A stochastic model of sedimentation: probabilities and multifractality. Euro. Jln. Applied Math., v.13, part 4, 371-383 (2002). 4. Molchan G.M., Khokhlov A.V., Small values of the maximum for the integral of fractional Brownian motion, J. Stat. Phys. 114:3-4 (2004), 923-946. 5. Molchan G.M., Khokhlov A.V., Unilateral small deviations for the integral of fractional Brownian motion. (http://arXiv.org/math.PR/0310413, 2003). 6. Sinai, Ya.G., Distribution of some functionals of the integral of Brownian motion. Theor. Math. Phys., 90:3, 323 (1992) (in Russian). 7. Isozaki, Y., Watanabe, S., An asymptotic formula for the Kolmogorov diffusion and a refinement of Sinai’s estimates for the integral of Brownian motion. Proc. Japan Acad., 70:A-9 (1994). -7- 8. Kawazu, K. & Tanaka, H., On the maximum of a diffusion in a drifted Brownian environment. Seminaire de Probabilites, 27, Lecture Notes in Math. 1557, Berlin: Springer, 1993, 78-85.