Random attractors and the preservation of synchronization in the presence of noise

Transcription

1 Random attractors and the preservation of synchronization in the presence of noise Peter Kloeden Institut für Mathematik Johann Wolfgang Goethe Universität Frankfurt am Main

2 Deterministic case Consider the ordinary differential equations in R d dx dt = f(x), dy dt = g(y), f, g regular and satisfy the one-sided Lipschitz conditions on R d for some L > 0, x 1 x 2, f(x 1 ) f(x 2 ) L x 1 x 2 2, y 1 y 2, g(y 1 ) g(y 2 ) L y 1 y 2 2, globally asymptotically stable equilibria x and ȳ.

3 Now consider the dissipatively coupled system dx dt = f(x) + ν(y x), dy dt = g(y) + ν(x y) with ν > 0. globally asymptotically stable equilibrium ( x ν, ȳ ν ). ( x ν, ȳ ν ) ( z, z) as ν, where z is the unique globally asymptotically stable equilibrium of the averaged system dz dt = 1 2 This is known as synchronization (f(z) + g(z)).

4 Stochastic case What is the effect of environmental noise on synchronization? Coupled Ito stochastic differential equations with additive noise dx t = (f(x t ) + ν(y t X t )) dt + α dwt 1, dy t = (g(y t ) + ν(x t Y t )) dt + β dw 2 t where W 1 t, W 2 t are independent two-sided scalar Wiener processes and α, β R d are constant vectors. unique stochastic stationary solution ( X ν t, Ȳ ν t ), which is pathwise globally asymptotically stable.

5 Moreover on finite time intervals [T 1, T 2 ] of R ( X ν t, Ȳ ν t ) ( Z t, Z t ) as ν pathwise where Z t is the unique pathwise globally asymptotically stable stationary solution of the averaged SDE dz t = f(z t) + g(z t ) 2 dt + α 2 dw 1 t + β 2 dw 2 t. [Caraballo & Kloeden, Proc. Roy. Soc. London (2005)]

6 Recall: The solutions of Ito stochastic differential equations are pathwise continuous, but not differentiable. Ito SDEs are really stochastic integral equations with stochastic integrals defined in the mean-square or L 2 sense. How do we apply the Lipschitz properties to obtain pathwise estimates?

7 A technical detour : Consider the Ito SDE dx t = f(x t ) dt + α dw t where f satisfies the one-sided Lipschitz condition. i.e., the stochastic integral equation X t = X t0 + t t 0 f(x s ) ds + α t t 0 dw t The difference of any two solutions satisfies pathwise X 1 t X 2 t = X 1 t 0 X 2 t 0 + t [ f(x 1 s ) f(xs) ] 2 t 0 }{{} continuous paths ds

8 Fundamental theorem of calculus X 1 t X 2 t pathwise differentiable. d dt [ X 1 t X 2 t ] = f(x 1 t ) f(x 2 t ) pathwise Apply the one-sided Lipschitz condition d dt X 1 t X 2 t 2 = 2 X 1 t X 2 t, f(x 1 t ) f(x 2 t ) 2L X 1 t X 2 t 2 X 1 t X 2 t 2 X 1 t 0 X 2 t 0 2 e 2L(t t 0 ) 0 as t i.e. all solutions converge pathwise together but to what?

9 Special case : Ito SDE with linear drift f(x) = x dx t = X t dt + α dw t explicit solution X t = X t0 e (t t 0) + αe t t t 0 e s dw s The forward limit as t does not exist moving target! But the pullback limit as t 0 with t fixed does exist: t lim X t = Ō t := αe t e s dw s t 0 (pathwise)

10 x(, t 0, x 0 ) φ x 0 x(, t 0, x 0 ) φ t 0 t (a) Forward (b) pullback The Ornstein-Uhlenbeck stochastic stationary process Ō t is a solution of the linear SDE and all other solutions converge pathwise to it in the forward sense X t Ōt 0 as t (pathwise)

11 Random dynamical systems : Let (Ω, F, P) be a probability space and (X, d X ) a metric space. A random dynamical system (θ, φ) on Ω X consists of a metric dynamical system θ on Ω, which models the noise, a cocycle mapping φ : R + Ω X X, which represents the dynamics on the state space X and satisfies 1). φ(0,ω, x 0 ) = x 0 (initial condition) 2). φ(s + t,ω, x 0 ) = φ(s, θ t ω, φ(t,ω,x 0 )) (cocycle property) 3). (t,x 0 ) φ(t,ω, x 0 ) is continuous (continuity) 4). ω φ(t,ω,x 0 ) is F-measurable (measurability) for all s, t 0, x 0 X and ω Ω.

12 Random attractors A random attractor is a family of nonempty measurable compact subsets of X Â = {A(ω) : ω Ω} which is φ-invariant φ(t, ω, A(ω)) = A(θ t ω) for all t 0, pathwise pullback attracting in the sense that dist X (φ (t, θ t ω, D(θ t ω)), A(ω)) 0 for t + for all suitable families D = {D(ω) : ω Ω} of nonempty measurable bounded subsets of X.

13 Theorem (Crauel, Flandoli, Schmalfuß etc) Let (θ, φ) be an RDS on Ω X such that φ(t,ω, ) : X X is a compact operator for each fixed t > 0 and ω Ω. If there exists a pullback absorbing family B b = {B(ω) : ω Ω} of nonempty closed and bounded measurable subsets of X, i.e. there exists a T bd,ω 0 such that φ(t, θ t ω,d(θ t ω)) B(ω) for all t T bd,ω for all b D = {D(ω) : ω Ω} in a given attracting universe. Then the RDS Ω X has a random attractor A b with component subsets given by A(ω) = \ [ φ(t,θ t ω,b(θ t ω)) for each ω Ω. s>0 t s

14 General case again Substract the integral version of the linear SDE for Ōt from the integral version of the nonlinear SDE dx t = f(x t ) dt + α dw t to obtain X t Ōt = X t0 Ōt 0 + t [ ] f(xs ) + Ōs ds t 0 V t := X t Ō t is pathwise differentiable and satisfies the pathwise ODE d dt V t = f(v t + Ō t ) + Ō t (pathwise)

15 Apply the one-sided Lipschitz condition pathwise to d dt [ Xt Ō t ] = [ f(xt ) f(ō t ) ] + [ f(ō t ) + Ō t ] (pathwise) to obtain the pathwise estimate V t 2 V t0 2 e L(t t0) + 2 t L e Lt e Ls ( f(ō s ) 2 + Ō s 2) ds t 0 Take pathwise pullback convergence as t 0 to obtain X t Ōt R t := L e Lt t e Ls ( f(ōs) 2 + Ōs 2) ds for t T depending on suitable bounded sets of initial values.

16 i.e., there exists a family of compact pullback absorbing balls centered on Ōt with random radius R t. Dynamical systems limit set ideas there exists a compact setvalued stochastic process A t inside these absorbing balls which pathwise pullback attracts the solutions. BUT the solutions converge together pathwise in forwards sense, so the sets A t are in fact all singleton sets stochastic stationary solution X t.

17 General Principles All regular Ito SDE in R d can be transformed into pathwise ODE [Imkeller & Schmalfuß (2001), Imkeller & Lederer (2001,2002)] and generate random dynamical systems pathwise theory and numerics for Ito SDE Pullback convergence enables us to construct moving targets.

18 Stochastic stationary solutions are a simple singleton set version of more general random attractors theory of random dynamical systems e.g., Ludwig Arnold (Bremen) parallel theory of deterministic skew product flows e.g., almost periodic ODE : George Sell (Minneapolis) A theory of nonautonomous dynamical systems e.g., pullback attractors

19 Effects of discretization on synchronization Numerical Ornstein-Uhlenbeck process For the linear SDE with additive noise, dx t = X t dt + α dw t, the drift-implicit Euler-Maruyama scheme with constant step size h > 0 is X n+1 = X n hx n+1 + α W n, n = n 0, n 0 + 1,..., which simplifies algebraically to X n+1 = h X n + α 1 + h W n, Here the W n = W h(n+1) W hn are mutually independent and N(0, h) distributed

20 It follows that X n = 1 (1 + h) n n X n α 1 + h n 1 1 (1 + h) n 1 j W j j=n 0 and the pathwise pullback limit, i.e. with n fixed and n 0, gives the discrete time numerical Ornstein-Uhlenbeck process Ô (h) n := α 1 + h n 1 j= 1 (1 + h) n 1 j W j, n Z. (1) which is an entire solution of the numerical scheme and a discrete time stochastic stationary process. One can show that it converges pathwise to the continuous time Ornstein- Uhlenbeck process in the sense that Ô (h) 0 Ô 0 as h 0.

21 Discretization of an isolated stochastic system Consider the nonlinear SDE in R d with additive noise, dx t = f(x t ) dt + α dw t, where the drift coefficient f is continuously differentiable and satisfies the one-sided dissipative Lipschitz condition with constant L. The drift-implicit Euler-Maruyama scheme with constant step size h > 0 applied to this SDE is X n+1 = X n + hf(x n+1 ) + α W n, which is, in general, an implicit algebraic equation and must be solved numerically for X n+1 for each n.

22 The difference of any two solutions X n+1 X n+1 = (X n X n) + h ( f(x n+1 ) f(x n+1) ), does not contain a driving noise term. Then X n+1 X n+1 2 = X n+1 X n+1, X n X n +h X n+1 X n+1, f(x n+1 ) f(x n+1) X n+1 X n+1 X n X n hl X n+1 X n+1 2, X n X n Xn+1 X n Lh X n X n, 1 (1 + Lh) n X 0 X 0 0 as n. i.e. all numerical solutions converge pathwise to each other forward in time.

23 Change variables to U n := X n Ô(h) n, where Ô(h) n is the numerical Ornstein-Uhlenbeck process, to obtain the numerical scheme U n+1 = U n + hf ( ) U n+1 + Ô (h) n+1 + hô (h) n. Taking the inner product of both sides with U n+1 we obtain ( ) U n+1 2 = U n+1, U n + h U n+1, f U n+1 + Ô(h) n+1 + h U n+1, Ô(h) n U n+1 U n + h U n+1, f ( ) U n+1 + Ô(h) n+1 + h U n+1 Ô(h) n Rearranging, using the one-sided Lipschitz condition and simplifying gives (Ô(h) U n+1 U n Lh U n+1 + h f n+1) + h Ô(h) n..

24 where U n+1 B n (h) := f Lh U n + h 1 + Lh B(h) n, (Ô(h) n+1) + Ô(h) n, U n 1 (1 + Lh) n n U n h 1 + Lh n 1 j. 1 B(h) (1 + h) n 1 j j=n 0 Taking the pullback limit as n 0 with n fixed, it follows that U n is pathwise pullback absorbed into the ball B d [0, R n ] in R d centered on the origin with squared radius R 2 n := 1 + h 1 + Lh n 1 j= 1 B(h) (1 + h) n 1 j j. Note that R n is random, but finite.

25 From the theory of random dynamical systems we conclude that the discrete time random dynamical system generated by drift-implicit Euler- Maruyama scheme has a random attractor with component sets in the corresponding balls B d [0, R n ]. Since all of the trajectories converge together pathwise forward in time, the random attractor consists of a single stochastic stationary process which we shall denote by Û(h) n. Transforming back to the original variable, we have shown that the driftimplicit Euler-Maruyama scheme applied to the nonlinear SDE has a stochastic stationary solution X (h) n := Û (h) n + Ô (h) n, n Z, taking values in the random balls B d [Ô (h) n, R n ], which attracts all other solutions pathwise in both the forward and pullback senses.

26 Discretization of the coupled stochastic systems Consider the coupled stochastic system in R 2d (now α, β are nonzero scalars) dx t = (f(x t ) + ν(y t X t )) dt + α dw 1 t, dy t = (g(y t ) + ν(x t Y t )) dt + β dw 2 t, The corresponding drift-implicit Euler-Maruyama scheme with constant step size, X n+1 = X n + h (f(x n+1 ) + ν(y n+1 X n+1 )) + α W 1 n, Y n+1 = Y n + h (g(y n+1 ) + ν(x n+1 Y n+1 )) + β W 2 n, can be written as the 2d-dimensional vector system X n+1 = X n + h (F(X n+1 ) + νbx n+1 ) + A W n

27 with the 2d-dimensional vectors ( ) ( x W 1 X =, W y t = t Wt 2 ), F(X) = ( f(x) g(y) ), and the 2d 2d-matrices [ αid A = 0 ] 0, βi d B = where I d is the d d identity matrix. The function G := R 2d R 2d defined by G(X) := F(X) + νbx [ ] Id I d, I d I d satisfies a dissipative one-sided Lipschitz condition with constant L. i.e. the vector scheme has essentially the same structure as the scheme for the uncoupled nonlinear equation, but in a higher dimensional space.

28 The previous analysis can be repeated almost verbatim to give the existence of a unique stochastic stationary process ( ) X X (h,ν) n (h,ν) n = Ŷ n (h,ν), n Z, which attracts all other solutions pathwise in both the forward and pullback senses. X (h,ν) n Moreover, the take values in the random balls B 2d [Ô(h) n, ˆR n ] for appropriately defined ˆR n (which are independent of ν), where Ô(h) n is the discrete time Ornstein-Uhlenbeck stochastic stationary solution for the discrete time 2d-dimensional linear system X n+1 = h X n h A W n.

29 Theorem 1 ( X(h,ν) n Ŷ (h,ν) n ) ( Ẑ(h, ) n Ẑ (h, ) n ) pathwise uniformly on bounded integer intervals [N 1, N 2 ] as ν, where (Ẑ(h, ) n ) n Z is the discrete time stationary stochastic solution of the drift-implicit Euler-Maruyama scheme with constant step size Z n+1 = Z n h (f(z n+1) + g(z n+1 )) α W 1 n β W 2 n applied to the averaged SDE dz t = 1 2 (f(z t) + g(z t )) dt αdw 1 t βdw 2 t.

30 Synchronization of SDE with linear noise A Stratonovich stochastic differential equation with linear noise dx t = f(x t ) dt + α X t dw t can be transformed to the pathwise random ordinary differential equation dx dt = F(x, Ō t (ω)) := e Ō t (ω) f using the transformation ( eōt(ω) x ) + Ō t (ω) x x(t, ω) = e Ōt(ω) X t (ω). with the Ornstein-Uhlenbeck process Ōt := αe t t es dw s. NOTE: F satisfy the one-sided Lipschitz condition if f does.

31 Similar a pair of Stratonovich SDEs can be transformed to the RODEs dx t = f(x t ) dt + α X t dw 1 t, dy t = g(y t ) dt + β Y t dw 2 t, dx dt = F(x, Ō1 t (ω)) := e Ō1 t (ω) f ( eō1 t (ω) x dy ( ) dt = G(y, Ō2 t (ω)) := e Ō2 t (ω) g eō2 t (ω) y with the transformations ) + Ō1 t (ω) x, + Ō 2 t (ω) y, x(t, ω) = e Ō1 t (ω) X t (ω) Ō 1 t := αe t t e s dw 1 s y(t, ω) = e Ō2 t (ω) Y t (ω). Ō 2 t := βe t t e s dw 2 s

32 The coupled system of random ordinary differential eqautions (RODEs) dx dt = F(x, Ō1 t (ω)) + ν(y x), dy dt = G(y, Ō2 t (ω)) + ν(x y) has a pathwise asymptotically stable random attractor consisting of single stochastic stationary process ( x ν (t, ω), ȳ ν (t, ω)) with ( x ν (t, ω), ȳ ν (t, ω)) ( z(t, ω), z(t, ω)) as ν,

33 where z(t, ω) is the pathwise asymptotically stable stochastic stationary process of the averaged RODE dz dt = 1 2 [ F(z, Ō 1 t ) + G(z, Ō2 t ) ] i.e. dz dt = 1 2 [ e Ō1 t f (Ō1 t z ) + e Ō2 t f (Ō2 t z )] [Ō1 ] t + Ō2 t z or the equivalent Stratonovich SDE dz t = 1 2 [ e η t f(e η t Z t ) + e η t g(e η t Z t ) ] dt α dw 1 t β dw 2 t. where η t := 1 2 (Ō1 t Ō 2 t ).

34 Direct synchronization of the SDE The corresponding system of coupled SDE is dx t = dy t = ( f(x t ) + ν ( g(y t ) + ν (eō1 t Ō2 t Yt X t )) (e Ō1 t +Ō2 t Xt Y t )) has a unique stochastic stationary solution ( Xν t (ω), Ȳ ν t (ω) ) dt + α X t dw 1 t, dt + β Y t dw 2 t. which is pathwise globally asymptotically stable with ( X ν t (ω), Ȳ ν t (ω) ) ( z(t, ω)e Ō1 t (ω), z(t, ω)e Ō2 t (ω)) as ν,

35 Stochastic reaction-diffusion system on a thin two-layer domain Let D 1,ε and D 2,ε be thin bounded domains in R d+1, d 1, D 1,ε = Γ (0, ε), D 2,ε = Γ ( ε, 0), with 0 < ε 1 and Γ a bounded C 2 -domain in R d. Write x D ε := D 1,ε D 2,ε as x = (x, x d+1 ) where x Γ and x d+1 ( ε, 0) (0, ε).

36 Consider the system of Ito stochastic PDE t Ui ν i U i + au i + f i (U i ) + h i (x) = Ẇ(t, x ), t > 0, x D i,ε, i = 1, 2, where Ẇ(t, x ) white noise depending only x Γ. [Deterministic model: Chueshov & Rekalo (Matem. Sbornik, 2004)] [Stochastic model: Caraballo, Kloeden & Chueshov (SIAM J. Math.Anal., 2007)]

37 Neumann boundary conditions ( U i, n i ) = 0, x Di,ε \ Γ, i = 1, 2, on the external part of the boundary of the compound domain D ε, where n is the outer normal to D ε Matching condition on Γ U ( ν 1 ) 1 + k(x, ε)(u 1 U 2 ) x d+1 U (ν 2 ) 2 + k(x, ε)(u 2 U 1 ) x d+1 Γ Γ = 0, = 0.

38 Synchronization as ε 1 with the averaged system t U ν ΓU + au + f(u) + h(x ) = Ẇ(t, x ), x Γ, on the spatial domain Γ with the Neumann boundary conditions on Γ and with ν = ν 1 + ν 2 2, f(u) = f 1(U) + f 2 (U) 2, h(x ) = h 1(x, 0) + h 2 (x, 0). 2 Method : Transform Ito SPDE into a pathwise random PDE.

39 Assumptions f i C 1 (R) such that f i (v) c for all v R and vf i (v) a 0 v p+1 c, f i (v) a 1 v p 1 + c, a j and c positive constants and 1 p < 3; h i H 1 (D i,1 ), i = 1, 2; k(, ε) L (Γ), k(x, ε) > 0, x Γ, ε (0,1], lim ε 0 ε 1 k(x, ε) = +, x Γ; W(t), t R, two-sided L 2 (Γ)-valued n o Wiener with covariance operator K = K 0 such that for some β > max 1, d 4 h tr K ( N + 1) 2β 1i <, N Laplacian in L 2 (Γ) with Neumann boundary conditions on Γ. (Ω, F, P) the corresponding probability space

40 Theorem 2 Under the above Assumptions the following assertions hold. 1. The coupled SPDE generates an RDS (θ, φ ǫ ) in the space H ǫ = L 2 (D 1,ǫ ) L 2 (D 2,ǫ ) L 2 (D ǫ ) given by φ ǫ (t, ω)u 0 = U(t, ω), where U(t, ω) = (U 1 (t, ω); U 2 (t, ω)) is a strong solution to the problem and U 0 = (U 1 0 ; U2 0 ). 2. Similarly, the averaged SPDE generates an RDS (θ, φ 0 ) in the space L 2 (Γ). 3. Cocycles φ ǫ converge to φ 0 in the sense Z lim sup 1 φ ǫ (t, ω)v φ 0 (t, ω)v 2 dx = 0, ǫ 0 t [0,T] ǫ D ǫ ω, for any v(x) H ǫ independent of the variable x d+1, and any T > These RDS (θ, φ ǫ ) and (θ, φ 0 ) have random compact pullback attractors {Āǫ (ω)} and {Ā0 (ω)} in their corresponding state spaces. Moreover, if K is non-degenerate, then {Ā0 (ω)} is a singleton, i.e. Ā 0 (ω) = { v 0 (ω)}, where v 0 (ω) is an L 2 (Γ)-valued tempered random variable.

41 5. The attractors {Āǫ (ω)} are upper semi-continuous as ǫ 0 in the sense that for all ω Ω lim ǫ 0 j sup v Āǫ (ω) 6. In addition, if inf v 0 Ā0 (ω) 1 ǫ ν 1 = ν 2 := ν, Z ff v(x, x d+1 ) v 0 (x ) 2 dx = 0. D ǫ f 1 (U) = f 2 (U) := f(u), h 1 (x, x d+1 ) = h(x ) = h 2 (x, x d+1 ); f(u) is globally Lipschitz, i.e. there exists a constant L > 0 such that and also that f(u) f(v ) L U V, U, V R, k(x, ǫ) > k ǫ for x Γ, ǫ (0,1] and lim ǫ 0 ǫ 1 k ǫ = +, then, there exists ǫ 0 > 0 such that for all ǫ (0, ǫ 0 ] the random pullback attractor {Āǫ (ω)} for (θ, φ ǫ ) has the form Ā ǫ (ω) v(x, x d+1 ) v 0 (x ) : v 0 Ā0 (ω), where {Ā0 (ω)} is the random pullback attractor for the RDS (θ, φ 0 ).

42 References L. Arnold, Random Dynamical Systems, Springer-Verlag, A.N. Carvalho, H.M. Rodrigues & T. Dlotko, Upper semicontinuity of attractors and synchronization, J. Math. Anal. Applns., 220 (1998), H.M. Rodrigues, Abstract methods for synchronization and applications, Applicable Anal., 62 (1996), P.E. Kloeden, Synchronization of nonautonomous dynamical systems, Elect. DJ. Diff. Eqns., 2003, T. Caraballo & P.E. Kloeden, The persistence of synchronization under environmental noise, Proc. Roy. Soc. London A461 (2005), T. Caraballo, P.E. Kloeden and A. Neuenkirch, Synchronization of systems with multiplicative noise, Stochastics & Dynamics 8 (2008), P.E.Kloeden, A. Neuenkirch and R. Pavani, Synchronization of noisy dissipative systems under discretization, J. Difference Eqns. Applns. (to appear) T. Caraballo, I. Chueshov & P.E. Kloeden, Synchronization of a stochastic reactiondiffusion system on a thin two-layer domain, SIAM J. Math.Anal., 38 (2007), I.D. Chueshov and A.M. Rekalo, Global attractor of contact parabolic problem on thin two-layer domain, Sbornik Mathematics, 195 (2004),