# Full text of "Suppression of unbounded gradients in a SDE associated with the Burgers equation"

## See other formats

SUPPRESSION OF UNBOUNDED GRADIENTS IN A SDE ASSOCIATED WITH THE BURGERS EQUATION "d^" O . SERGIO ALBEVERIO \ OLGA ROZANOVA O ' Abstract. We consider the Langevin equation describing a non- viscous Burg- f-\ ' ers fluid stochastically perturbed by uniform noise. We introduce a determin- ^^ ^ istic function that corresponds to the mean of the velocity when we keep fixed ^V( , the value of the position. We study interrelations between this function and the f^^ ' solution of the non-perturbed Burgers equation. Especially we are interested ^», -1 ' in the property of the solution of the latter equation to develop unbounded gra- dients within a finite time. We study the question how the initial distribution of particles for the Langevin equation influences this blowup phenomenon. We P^ ' show that for a wide class of initial data and initial distributions of particles ^^ ' the unbounded gradients are eliminated. The case of a linear initial velocity ^J is particular. We show that if the initial distribution of particles is uniform, r^ , then the mean of the velocity for a given position coincides with the solution of the Burgers equation and, in particular, it does not depend on the constant variance of the stochastic perturbation. Further, for a one space variable we get the following result: if the decay rate of the even power-behaved initial particles distribution at infinity is greater or equal |a;|~^, then the blowup is suppressed, otherwise, the blowup takes place at the same moment of time as C^ ' in the case of the non-perturbed Burgers equation. > m vn in 1. Introduction ^—v ■ It is well known that the non-viscous Burgers equation, the simplest equation QQ I that models the nonlinear phenomena in a force free mass transfer, a o Ut + {u,W)u^O, (1.1) where u{x,t) = {ui, ...,m„) is a vector-function R"+^ — s- M", before the formation of shocks is equivalent to the system of ODE C^ ■ x{t)=u{t,x{t)), u{t,x{t))^0. (1.2) The latter system defines a family of characteristic lines x = x(t) that can be interpreted as the Lagrangian coordinates of the particles. Given initial data u{x,0) =uo{x), (1.3) one can readily get an implicit solution of (1.1), (1.3), namely, u{t, x) — uo{x — tu(t, x)). Date: April 27, 2009. 1991 Mathematics Subject Classification. 35R60. Key words and phrases. Burgers equation, gradient catastrophe. Supported by Award DFG 436 RUS 113/823/0-1 and the special program of the Ministry of Education of the Russian Federation "The development of scientific potential of the Higher School", project 2.1.1/1399. 1 2 ALBEVERIO.ROZANOVA For special classes of initial data we can obtain an explicit solution. The simplest case is uo{x) = ax, a £ R, (1.4) where .(M).^^. (1.5) Thus, if a < 0, the solution develops a singularity at the origin as i ^ T, < T < oo, where T= . (1.6) a In the present paper we consider a 2 x n dimensional Ito stochastic differential system of equations, associated with (1.2), namely dXkit)^Ukit)dt, dUk{t) ^ a d{Wk)t, fc = l,..,n, X(0) = X, C/(0) = u, t > 0, where {X{t), U{t)) runs the phase space R" x R", cr > is constant, {W)k,t^ k = 1, ..., n, is the n - dimensional Brownian motion. Our main question is: can a stochastic perturbation suppress the appearance of unbounded gradients? The stochastically perturbed Burgers equation and the relative Langevin equa- tion were treated in many works (e.g. [T],[2]). The behavior of the space gradient of the velocity was studied earlier in other contexts in [3] , [4] , but this problem is quite different from the problem considered in the present paper. The analogous problem concerning the behavior of gradients of solutions to the Burgers equation under other types of stochastic perturbations was studied in [5] . Let us consider the mean of the velocity U{t) at time t when we keep the value of X{t) at time t fixed but allow U{t) to take any value it wants, namely J uP{t,x,u)du '^'^^^- ")pit,x,u)du ^ t>0,..R", (1.7) where P{t,x,u) is the probability density in position and velocity space, so that J P{t,x,u) dxdu ~ 1. This function obeys the following Fokker-Planck equation: dP{t,x,u) dt subject to the initial data ^ d 1 2 52 P(i,x,w), (1.8) P(0,a;,u) = Po{x,u). If we choose n Po{x, u) = 6{u - uo{x)) fix) = n '^("'^ - iM^))k) f{^), (1-9) fe=i with an arbitrary sufficiently regular /(x), then u{0,x) = uo{x). The function f{x) has the meaning of a probability density of the particle positions in the space at the initial moment of time and therefore f{x) has to be chosen such SUPPRESSION OF UNBOUNDED GRADIENTS that / f{x)dx — 1. If the latter integral diverges for a certain choice of /(x), we consider the domain il^ := [— L,_L]", L > and the re- normalized density (i V fhix) :— x{^l) f{x) I f{x)dx\ , where x(^l) is the characteristic function of r^i, we denote the respective probability density in velocity and position by PL{t,x,u) and modify the definition of u{t,x) as follows: / u PL{t, X, u) du u{t,x)^ lim "" ^ ,^ ——, t>0,xenL, (1.10) R" provided the limit exists. We apply heuristically the Fourier transform in the variables u and x to (1.8), (1.9) to obtain for P = P{t,\,0 P(0,A,C) - / /(s)e-*(«'"''("» e^^f^'") ds. (1.12) Thus, (1.11) and (1.12) give P(t,A,0 = e-^(l«+^*l''l«l') / /(s)e-*««+^*)^«"(^))e-^(^'^)ds, (1.13) Pit,x,u) -^^ J P(i,A,C)e*(«'")e*(''-)dAdC R2" V3 \ In,-, J.^ (3t^ («,«o(s)) + t^ |'[to(s)-tt|^+3|rE-s|2+3t(M+Mo(s),S-£c)) 7rcr2i2 ffc)g^^^^W'^ l«.«olsj;+t \uo(s)-u\ +6\x-s\ +ityu+ua'yS),s-x)j , (1.14) Now we substitute (1.14) in (1.8) or (1.10), integrate with respect to u and get the formula f , , s , ,, , s 3|tio(s)t+(s-x)|'' / {-uo{s)t - 3(s - x)) f{s) e i^^T^ ds u(t, x) ^ — — = , i > 0, a; e K", J J(sje 2CT^t3 ds R" (1.15) provided all integrals exist. Thus, we can compare u(t, x) with the solution u{t, x) of the non- viscous Burgers equation (1.1). 2. Exact results It is natural to begin with the case where the solution to the Burgers equation (1.1) can be obtained expHcitly. Let us choose uo{x) = ax, a < 0. (2.1) 4 ALBEVERIO, ROZANOVA One can see from (1.5), (1.6) that the gradient of the solution becomes unbounded as t ^ T. If the initial distribution of particles is either uniform or Gaussian, it is possible to get explicit formulas for u. Namely, for the uniform distribution f{x) = const both integrals in the numerator and the denominator in (1.15) can be taken and we get which coincides with (1.5). Therefore, the gradient becomes unbounded at T = — — . On the contrary, in the case of a Gaussian distribution, f{x) = I -j= r > 0, we get another explicit formula: One can see that the denominator does not vanish for all positive t, and at the critical time T we have u(t,x) = —^ax, that is the gradient becomes positive and tends to zero as t — > +oo. 3. ID CASE, SPECIFIC CLASSES OF INITIAL DISTRIBUTIONS OF PARTICLES AND INITIAL DATA Our main question is how the decay rate of the function f{x) at infinity relates to the property of u to reproduce the behavior of the solution of the non-perturbed Burgers equation at the critical time. For the sake of simplicity we dwell on the case of an one dimensional space, however the results can be extended to the higher dimensional space. Let us consider the class of initial distributions of particles f{x) which are intermediate between Gaussian and uniform. Our aim is to find a threshold rate of decay at infinity that still allows to preserve the singularity at the origin. We restrict ourselves to the class of smooth distributions f{x) and initial data Uo{x) satisfying the condition C(wo(C))'/(C)cxp(-7e^) dC <oo foraU m, ^ € N U {0}, 7 > 0. (3.1) As a representative of such a class of distributions we can consider f{x) ^ const • (1 + Ixl^f, fc e R. (3.2) Theorem 3.1. Let the initial be uo{x) be smooth, and for a certain fixed (3 < and all X (z R- (except for maybe a bounded set) 1^0(2;) — Px\ > 7 > 0. Moreover, assume that the distribution function f{x) is smooth, nonnegative, and the property (3.1) is satisfied. Then the mean u{t,x) has at the origin x ^ at the moment to = —-Q, f3 <0, a bounded derivative u'^{to,0). We remark that the initial data with a linear initial profile except for /3 = Uq{x) fall into the class of initial data that we have described above. Proof. First of all we perform a change of the time variable. Let e = t+4,/3<0. We expand u(t{€),x) given by (1.15) into Taylor series at the point t ^ to ^ —i {e — SUPPRESSION OF UNBOUNDED GRADIENTS 5 0), X = 0, taking into account that condition (3.1) guarantees the convergence of the integrals in the coefficients of the expansion. This expansion has the form J f{s)e^y^ y ds R / 3,33 f uq(s) \2 //(a2-4/32suo(s)+3s2/33+/3(uo(s))2) f{s)e^y^ V ds 3/3 E J{Ps-ua{s)) f{s)e^y ^ ) ds J {3ps - uo{s)) f{s)e^y ^ > ds 3/33 /..^'^^ N2 \ 2 / /(s)e^('^"') ds I (3.3) asa:;^0,e^O— (where ~ stands for the quotient of the left and right sides converging to 1). The theorem follows immediately from the asymptotics (3.3). D Let us notice that for even j{x) and odd uo(a^) the expansion (3.3) is less cum- bersome, namely, 00 3/33 /,yj(£) \2 / (f72 _ 4/32suo(s) + 3s2/33 + /3(mo(s))2) /(s) e-^\^^-n ds -{,{ \ \ 3/3 «(t(e),a;) ~ -^ ^ 3^3 /„„<., ^. ^. (3.3*) as X ^ 0, e ^- — . It can be readily calculated that if /3 ^ —00 (tg -^ 0), then (3.3*) yields ■u(t, a;) ^ ax, a; ^ 0, e ^ 0—, where a — Ux{0) (taking account of "°^ -^ ^ a, Uo(^) ^ a, ^ ^ 0). 3.1. Power-behaved distribution. Let us consider the specific class of even dis- tributions (3.2) and linear initial data. The case a linear initial function uq{x) = ax, a ^ /3, is particular. Indeed, we have from (3.3*) for .x ^ and for t ^ fo = — ;|, /3 < 0, the following asymptotic behavior: u{t,x) ^ A(/3)x, with 7 (^2 + (3s\a ~p){a~ 3/3)) f{s) e^ (''-")' ds ^^ 2a2 J.fis)e^^^--^-'ds We can see that if /3 < a (before the critical time T = — — , when the solution of the non-perturbed Burgers equation blows up) or /3 > a (after the time T) both integrals in (3.3*) converge and therefore the derivative {(^(t, 0) remains bounded. 6 ALBEVERIO.ROZANOVA Let US consider now the critical moment of time t ~ T, where p ~ a. In this case "°^^^ = (3 identically and we do not have a multiplier that guarantees the convergence of integrals of the form CfiO di for all TO e N, which is necessary for the validity of the asymptotics (3.3*). However, fortunately, due to the relative simplicity of f{x) we can compute u{t^x) in the vicinity of the origin directly, using the formula (1.15), which in this case takes the form / {-ast - 3(s - x)) (1 + s2)fe e-3|ast+(.-:.)p ^g Hit, x) — j^- TTT, Ti 777 TTi—, 7 i > 0, X G M. ^' ^ 2i / (l + s2)fcg-3|ast+(s-a;)p^g R+ (3.4) Computations show that for fc 7^ ^, to e Z, the asymptotic behavior of (3.4) as X — > 0, e ^ 0—, where e = t + — , can be expressed through the Gamma function and the generalized Laguerre functions L{vi, vi^ v-^), see [SI. It has the form u(t,x) ^ J: / ' ' X X, (3.5) F2(e,fc, a, tj) i^i(e, k, a, a) = Ai{k) e'^'^'^ + o(,-2fe-2) ^ ^^(^) ^0 ^ ^(^O)^ i^2(e, k, a, a) = A^{k) f-^fe-i + o(£-2fe-i) + yi4(A:) gO + 0(^0), where the coefhcients Ai{k), i — 1, ..,4, are as follows: 29/0+2 2fc ,'4^2 _ -,^ ;^ ^iW= Vm il+r /^ r(fc + i)^(fc,-fc + ^,o), JK |Q,|ofc+i COSTTK 2 -42(fc) = — V -T tan(7rfc) L(-. fc + -, 0), ^ '^ 2CT(fc + l) ^ ^4' 2' ^' 22fe+i 2fe/2i-_ n 1 -^-(^)- 3^|.p.Mfe + l)(^ + 2)L.fc ^(^ + ^)^(^'-^+2'»)' _ N/6.^H^(2fc + 3)r(fc + 3) '^'^''^" a(fc + i)(fc + 2)r(fc + |) *^"^"^^- Thus, if fc < —1, then the leading term of the numerator and denominator in (3.5) as e ^ 0— is A2 e" and (3.5) can be written as ^(''") ^ t(K°to(e"! "^(^^('^+"(^"^^"'"^°' ('■') where Bi(fc) = ^. This signifies that the derivative ii'^ (t, 0) tends to a finite limit as e ^ — . If— i>/s>— 1, then the leading term of the denominator is A4(fc)e°. Otherwise, if A: > — i, then this leading term is A3(fc)e~2'=-i. Thus we have for — i > k > —1 Ai(fc)e-2fe-2 + o(e-2fc-2) "^'^^^ A4(fc)e0 + o(e0) SUPPRESSION OF UNBOUNDED GRADIENTS and MUi,0)^i?2(fc)-^, B2(fc) = 4^, x^O,e^O-. (3.7) g2fc+2' --V-/ ^^ 4 At last for fc > — ^ we have "(^>^)- . /,.^,-2fc-l^^r.-2fc^l^ ^' a;^0, e^O- and ^Ui,0)^B3(fc)-e-\ S3(fc)-44lT^2fc + l. (3.8) If fc G Z, then the numerator and the denominator in the leading term in the expansion of (3.5) as a; ^ are expressed cither through rational functions (k > 0) or through a Gaussian distribution function (fc < 0). For k = ^^j^, I G Z, the coefficient of the leading term is expressed through a fraction of series consisting of the digamma functions. Anyway, the asymptotics (3.6) takes place also for k ^ ^, I £ Z, k ^ — i, and it can be found also as a limit k ^> k. For k < —1 the function u{t,x) behaves as in (3.6), where the coefficient Bi{k) can be calculated either independently or as lim xTkT- ^i'^ce for k = —1 the degrees in e-2fe-2 ^j^j e° coincide, then ^0 , , , (Ai(k) + A2(K))e" + o(e°) ,^ , n^^ Ht,x)^ lim ^ '^ ' ,oV , o^ ' x-{Bi + o{e°))x,x 0, e has the asymptotics (3.8) with the same value B^(k). An exceptional case is fc = — 2 , where where B4 = lim (Bi(k) + ^ij^] == ^ - ^^^^ ■ For fc > the function u{t,x) lal casi Fi(e,-l/2,a,(T)=.iie-i+o(e-i), Ai = lim Ai = ^^l^"l\ fe^-1/2 cr ^2(e, -1/2, a, (j) = As In(-e) + o (In(-e)) , ^Is = -^, e ^ - . Thus, for fc = — i we have ..(M)--^ + o(^),.-0-. (3.9) The following theorem summarizes our results: Theorem 3.2. Assume that in the case n — 1 the initial distribution Junction is fix) = const • (1 + Ixp)'', fc e M, and the initial velocity has the form uo{x) = ax, a < 0. Then the space derivative of the mean u{t, x) at the origin x ^ is bounded for all t > except for the critical time T = — — . At the critical time the behavior of the derivative depends on fc. Namely, for fc > — 1 the mean u{t, x) keeps the property of solutions to the non-perturbed Burgers equation to blow up at the critical time T at x = 0. The rates of the blowup for — ^ > fc> —1, fc> — i and fc == — i are indicated in (3.7), (3.8) and (3.9), respectively. Otherwise, if k < — 1, the derivative u'^(t,0) at the critical time remains bounded, i.e the singularity disappears. 8 albeverio.rozanova 4. Pressureless gas dynamics model and a limit case at vanishing NOISE Let us consider the pair (pt, u(-, t)),t > 0, where pt is the probability distribution of the random variable Xt, governed by the SDE Xt = Xq+L E[uo{Xq)\Xs] ds, t > 0, with a given random variable Xq with values in M and function uo{x), and u{t,x) = E[Mo(-'^o)|-'^i = x]. According to 7 , {pt,u{-,t),t > 0) is a weak solution to the pressureless gas dynamics model dtp + d^ipu) ^ 0, dtipu) + d.Apu^) = 0. (5.1) Therefore it is natural to expect that the limit as cr — > of the mean u (denoted by u(t, x)) takes part in the solution to (5.1). For smooth uo{x) and f{x) this can be readily shown. First of all, let us introduce the function p(t,x) = J P{t,x,u) du and notice that it satisfies the continuity equation dtp -\- dx{pu) = 0. Further, we check that the function v{t,x) satisfies the Burgers equation (1.1). Indeed, j^ exp f 3iM'^)^t+{^-^)f \ ^ 5(5 _ s{t, x)), as cr ^ in V , where s{t,x) is a solution to equation uo(s) + ^^ — 0, given in implicit form. This function exists and it is differentiable provided t 7^ — r-r^ (at this moment of time the solution to the Burgers equation blows up). Thus, u{t, x) — > uo(s(i, x)) as (7 — > 0. Now it is sufficient to substitute uo{s{t, x)) into (1.1) and compute the derivatives of s(i, x) by means of the implicit function theorem. Thus, for smooth initial data (f{x),UQ(x)) the pair {p,u) is a solution to the system dtp + dx{pu) = 0, dtipii) + dx{pu^) = A, A = - Px{t,x,u) {u- uf du, where A ^- as a ^ 0. The integral relaxation term A can be used instead of the traditional viscosity [8j. It is interesting to consider the Fokker-Plank equation (1.8) as a kinetic equation dP{t,x,u) ^-^ / dP{t,x,u) dukP{t,x,u) E/ ur-yi,x,u) uuk-r\i,x,u) \ _ \ finf,. fill I. I dt ^-^ V dxk du fc=i ^ (e.g. [9]), where the acceleration ii of the particles is due to external forces and the interaction forces with other particles. It can be readily calculated that in our case it = |(u — u). The authors would like to thank an anonymous referee for helpful suggestions. References [1] H. Risken, The Fokker-Planck Equation Methods of Solution and Applications, Second Edition, Springer- Verlag, 1989. [2] Weinan E, K. M. Klianin, A. E. Mazel, Ya. G. Sinai, Invariant measures for Burgers equation with stochastic forcing, Ann. of Math. (2) 151, no. 3, 877-960 (2000). [3] J.P.Bouchaud, M.Mezard, Velocity fluctuations in forced Burgers turbulence, Phys.Rev. E54, 5116 (1996). [4] V.Gurarie, Burgers equations retiisited, |arXiv:nlin/0307033y l [nln.CD] (2003). [5] S.Albeverio, O.Rozanova, The non-viscous Burgers equation associated with random posi- tions in coordinate space: a threshold for blow up behaviour, to appear in M'^AS (Mathe- matical methods and models in applied science) (2009), arXiv;0708.2320v2 [math.AP], SUPPRESSION OF UNBOUNDED GRADIENTS 9 [6] I.S. Gradshteyn, I.M. Ryzhik, Tabic of integrals, scries, and products. 6th ed. San Diego, CA: Academic Press, 2000. [7] A.Dcrmoune, Probabilistic interpretation for system of conservation law arising in advection particle dynamics C. R.Acad. Sci. Paris, 326, Serie I, 595-599 (1998) [8] D.Tan, T.Zhang, Y.Zheng, Delta-shock waves as limits of vanishing viscocity for hyperbolic system of conservation laws, J.Diff.Equat., 112 (1994), 1-32. [9] H.Struchtrup, Macroscopic Transport Equations for Rarefed Gas Flows: Approximation Methods in Kinetic Theory, Springer- Verlag Berlin Heidelberg, 2005. Q-) Universitat Bonn, Institut fur Angewandte Mathematik, Abteilung fur Stochastik, Wegelerstrasse 6, D-53115, Bonn; HCM, SFB611 and IZKS, Bonn; BiBoS, Bielefeld- Bonn, Germany (^) Mathematics and Mechanics Faculty, Moscow State University, Moscow 119992, Russia E-mail address, ^; albeverloOuni-bonn.de E-mail address, ^; rozanovaOmech. math. msu.su