Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Cubic function wikipedia , lookup
System of polynomial equations wikipedia , lookup
Elementary algebra wikipedia , lookup
Quartic function wikipedia , lookup
Bra–ket notation wikipedia , lookup
Basis (linear algebra) wikipedia , lookup
History of algebra wikipedia , lookup
Eigenvalues and eigenvectors wikipedia , lookup
System of linear equations wikipedia , lookup
4 Stabilization by noise of the Navier–Stokes equation Consider the Navier-Stokes equation Xt − ν0∆X + (X · ∇)X = fe + ∇p, ∇ · X = 0, in (0, ∞) × O, (4.1) X = 0, on (0, ∞) × ∂O, X(0) = x0, in O. in (0, ∞) × O, where O is an open and bounded subset of Rd, d = 2, 3, with smooth boundary ∂O. Here fe ∈ (L2(O))d is given. Let Xe = ye be an equilibrium solution to (4.1), i.e., −ν0∆Xe + (Xe · ∇)Xe = fe + ∇pe, in O, (4.2) ∇ · Xe = 0 in O, X = 0 on (0, ∞) × ∂O. e As seen earlier, if we replace X by X − Xe equation (4.1) reduces to Xt − ν0∆X + (X·∇)Xe+(Xe·∇)X+(X·∇)X = ∇p in (0, ∞) × O, (4.3) ∇ · X = 0 in O, X = 0 on (0, ∞) × ∂O, X(0) = x in O, where x = x0 − Xe. We can rewrite system (4.3) as ( Ẋ(t) + AX(t) + B(X(t)) = 0, t ≥ 0, (4.4) X(0) = x, where A and B were defined earlier in (1.9). 1 To stabilize (4.4), we associate the controlled stochastic system N X dX(t) + (AX(t) + B(X(t))dt = Vj (t)ψj dβj (t), (4.5) j=1 X(0) = x, where {βj }N j=1 is an independent system of real Brownian motions in a filtered probability space (Ω, P, F, {Ft}t>0). The main result, Theorems 4.1 below, amounts to saying that, e associated with H, under appropriate in the complexified space H assumptions on A (and, implicitly, on Xe), for each γ > 0 there e exist N ∈ N, {ψj }N j=1 ⊂ H, and an N -dimensional adapted process {Vj = Vj (t, ω)}N j=1 , ω ∈ Ω, such that for all x in a sufficiently small γt neighbourhood of the origin, t → e 4 X(t, ω) is decaying to zero for t → ∞ in a set Ω∗x of positive probability which is precisely estimated. Moreover, it turns out that the stabilizable controller arising in the right hand side of (4.5) is a linear feedback controller of the form (4.6) Vj (t) = ηhX(t), ϕ∗j iHe , ψj = P (mφj ), j = 1, ..., N, where |η| > 0 and ϕ∗j are the eigenfunctions of the dual Stokes-Oseen operator A∗ corresponding to eigenvalues λj with Re λj ≤ γ, {φj }N j=1 ∗ is a system of functions related to ϕj and m = 1lO0 is the characteristic function of O0 where O0 is a given arbitrary open subset of O. We may view (4.5) as the deterministic system (4.4) perturbed by P the white noise controller N j=1 Vj (t)ψj β̇j with the support in O0 . Notations Throughout in the following βj , j = 1, ..., N are independent real Brownian motions in a filtered probability space (Ω, P, F, {Ft}t>0) and we shall refer to [7] and [8] for definition and basic results on 2 stochastic analysis of differential systems and spaces of stochastic processes adapted to filtration {Ft}t>0. The scalar product of H is dee the complexnoted h·, ·iH and the norm | · |H . We shall denote by H ified space H + iH with scalar product denoted by h·, ·iHe and norm e is the space of all adapted square-mean by | · |He . CW ([0, T ]; L2(Ω, H)) e H-valued continuous processes on [0, T ]. The main result To begin with, let us briefly recall a few elementary spectral properties of the Stokes-Oseen operator A already mentioned in Section 1. e The Denote again by A the extension of A to the complex space H. operator A has a compact resolvent (λI − A)−1 and −A generates a e Consequently, A has a countable C0-analytic semigroup e−At in H. number of eigenvalues {λj }∞ j=1 with corresponding eigenfunctions ϕj each with finite algebraic multiplicity mj . Of course, certain eigenfunctions ϕj might be generalized and so, in general, A is not diagonalizable, i.e., the algebraic multiplicity of λj might not coincide with its geometric multiplicity. Also, each eigenvalue λj will be repeated according to its algebraic multiplicity mj . We shall denote by N the number of eigenvalues λj with Re λj ≤ γ, j = 1, ..., N, where γ is a fixed positive number. Denote by PN the projector on the finite dimensional subspace Xu = lin span{ϕj }N j=1 . e and We have Xu = PN H (4.7) 1 PN = − 2πi Z (λI − A)−1dλ, Γ where Γ is a closed smooth curve in C which is the boundary of a domain containing in interior the eigenvalues {λj }N j=1 . 3 Let Au = PN A, As = (I − PN )A. Then Au, As leave invariant e and the spectra σ(Au), σ(As) the spaces Xu and Xs = (I − PN )H are given by ∞ σ(Au) = {λj }N j=1 , σ(As ) = {λj }j=N +1 . Since σ(As) ⊂ {λ ∈ C; Re λ > γ} and As generates an analytic e we have C0-semigroup on H, (4.8) e t ≥ 0. |e−Astx|He ≤ Ce−γt|x|He , ∀x ∈ H, Herein, we shall assume that the following hypothesis holds. (H1) All eigenvalues λj , j = 1, ..., N, are semi-simple. Denote by ϕ∗j the eigenfunction of A∗ corresponding to the eigenvalue λj . We have, therefore, (4.9) Aϕj = λj ϕj , A∗ϕ∗j = λj ϕ∗j , j ∈ N. Since the eigenvalues {λj }N j=1 are semi-simple, it turns out that the ∗ N system consisting of {ϕj }N j=1 , {ϕj }j=1 can be chosen to form a e i.e, bi-orthonormal sequence in H, (4.10) hϕj , ϕ∗k iHe = δjk , j, k = 1, ..., N, where δjk is the Kronecker symbol. We notice also that the functions ϕj and ϕ∗j have the unique continuation property, i.e., (4.11) ϕj 6≡ 0, ϕ∗j 6≡ 0 on O0 for all j = 1, ..., N. Consider the following stochastic perturbation of the system (4.4) considered in the complex space N X dX + (AX + B(X))dt = η hX, ϕ∗j iHe P (mφj )dβj , (4.12) j=1 X(0) = x, 4 where |η| > 0 and m = 1lO0 is the characteristic function of the open e subset O0 ⊂ O. Here {φj }N j=1 ⊂ H is a system of functions to be specified later on. In two dimensions the stochastic differential equation (4.12) has a e for all T > 0. global solution X ∈ CW ([0, T ]; L2(Ω, H)) The closed loop system (4.12) can be, equivalently, written as dX(t) − ν0∆X(t)dt+(X(t) · ∇)Xedt +(Xe · ∇)X(t)dt+(X(t) · ∇)X(t)dt N X hX(t), ϕ∗j iHe φj dβj (t) = ηm (4.13) j=1 +∇p(t)dt in (0, ∞) × O, P-a.s. ¯ ¯ ∇ · X(t) = 0 in O, X(t) ¯ = 0, ∀t ≥ 0, P-a.s. ∂O X(0) = x in O. Hence, in the space (L2(O))d, the feedback controller {uj = ηmhX, ϕ∗j iHe φj }N j=1 has the support in O0. We shall define now φj , j = 1, ..., N, as follows. (4.14) φj (ξ) = N X αlj ϕ∗l (ξ), ξ ∈ O, l=1 where αlj are chosen in such a way that N X αlj hϕ∗l , ϕ∗k i0 = δjk , j, k = 1, ..., N. l=1 (By Lemma 2.2, the system {ϕ∗` }N `=1 is linear independent on O0 and so, the Gram matrix {hϕ∗l , ϕ∗k i0}N l,k=1 is not singular, which makes this 5 possible.) With this choice, we have hφj , ϕ∗k i0 = δkj , k, j = 1, ..., N. R Here, we have used the notation hu, vi0 = O0 u(ξ)v̄(ξ)dξ. In the following, we shall denote by Aα , α ∈ (0, 1), the fractional power of order α of A, by D(Aα ) its domain and set |x|α = |Aα x| for 1 all x ∈ D(Aα ). Moreover, we shall denote by W the space D(A 4 ) if 1 d = 2 and D(A 4 +²) if d = 3 where ² > 0 is small. (4.15) Theorem 4.1 Let d = 2, 3, Xe ∈ C 2(O) and p (4.16) |η| ≥ max 6γ − 2Re λj . 1≤j≤N Then, there is C ∗ > 0 independent of ω such that, for each x ∈ W , |x|W ≤ (C ∗)2, there is Ω∗x ⊂ Ω with ³ ´− γ 2 1 − 2(ηN ) (4.17) P(Ω∗x) ≥ 1 − 2 C ∗|x|W2 − 1 , the solution X(t, x) to (4.12) satisfies ´ ³ γt (4.18) lim e 4 |X(t, x)|He = 0, t→∞ P-a.s. in Ω∗x. In particular, Theorem 4.1 implies that, if |x|W ≤ ρ0 < (C ∗)−2, then X = X(t, x) is exponentially decaying to 0 on a set Ω∗x of probability greater than ³ ´− γ 2 − 21 2(ηN ) ∗ 1 − 2 C |x|W − 1 . The constant C ∗ depends of Xe only. For the linearized Navier–Stokes equation, that is, if one takes B = 0, the exponential decay in (4.12) occurs with probability one. In fact, as seen from the proof of Theorem 4.1, the constant C ∗ comes 6 out from estimates on the nonlinear inertial term B and so, it is zero if this term is absent from the equation. As mentioned earlier, system (4.13) is written here in the complex e space H. If we set X1(t) = Re X(t), X2(t) = Im X(t), it can be rewritten as a real system in (X1, X2). In this case, the feedback controller is an implicit stabilizable feedback controller with support in O0 for the real Navier–Stokes equation (4.3). Of course, if λj , j = 1, ..., N , are real, then we may view X(t) as a real valued function and so, in (4.17), |X|He = |X|H . In particular, by Theorem 4.1 we have the following corollary. Corollary 4.2 Under the assumptions of Theorem 4.1 the feedback controller (4.19) ηm N X hX − Xe, ϕ∗j iHe φj j=1 stabilizes exponentially the stationary solution Xe, P-a.e in Ω∗x. Proof of Theorem 4.1 The idea of the proof is to transform equation (4.12) in a deterministic equation with random coefficients via substitution (4.20) y(t) = N Y e−βj (t)Γj X(t), t ≥ 0, j=1 e →H e is the linear operator where Γj : H (4.21) Γj x := ηhx, ϕ∗j iHe P (mφj ), 7 e j = 1, ..., N x ∈ H, e H) e is the C0-group generated by Γj , i.e., and esΓj ∈ L(H, (4.22) d sΓj e x − Γj esΓj x = 0, ds e ∀ s ∈ R, x ∈ H. We have by (4.21) and by (4.14) that (4.23) Γj Γk x = η 2hx, ϕ∗j iHe P (mφj )δjk , ∀ j, k = 1, ..., N and therefore the operators Γ1, ..., ΓN commute, because the Leray operator P is self-adjoint. Then, (4.12) reduces to N X 1 dy(t) 2 + Ay(t) + Γ j y(t) + F (t)y(t) dt 2 j=1 ´ ³ PN PN β (t)Γ − β (t)Γ j j j j (4.24) y(t) = 0, e j=1 B e j=1 ∀ t ≥ 0, P-a.s. y(0) = x, where F (t)y(t) = e − PN j=1 βj (t)Γj ³ PN ´ β (t)Γ A e j=1 j j y(t) − Ay(t). By a solution to (4.24) we mean a function 1 y ∈ C([0, ∞); D(A 4 )) ∩ L2(0, ∞; D(A)), which fulfills (4.24) P-a.s. in the mild sense (see Lemma 4.5 below). Conversely, if y is a solution to (4.24), then it is an adapted process and so (4.25) X(t) = N Y eβj (t)Γj y(t), j=1 8 t ≥ 0, 3 1 belongs to CW ([0, T ]; L2(Ω, P; D(A 4 )) ∩ L2(Ω, P, C([0, T ]; D(A 4 ))) and satisfies equation (4.12). Then we shall confine in the following to study existence and exponential convergence in probability to solutions y to equation to (4.24). We notice first that, as easily follows by (4.21) and (4.23), we have (4.26) esΓj y=η −1Γj y(eηs−1)+y=(eηs−1)hy, ϕ∗j iHe P (mφj )+y, ∀ s > 0, j = 1, ..., N, y ∈ H. respectively e−sΓj y = η −1Γj y(e−ηs − 1) + y = (e−ηs − 1)hy, ϕ∗j iHe P (mφj ) + y, ∀ s > 0, j = 1, ..., N, y ∈ H. This yields N ³ ´ X βj (t) (4.27) F (t)y = e − 1 hy, ϕ∗j iHe (AP (mφj ) − λj P (mφj )). j=1 We consider the operator (4.28) N 1 X 2 AΓy := Ay + Γ y, 2 j=1 j ∀ y ∈ D(A) e is and notice that the C0-semigroup e−AΓt generated by −AΓ on H analytic. The operator AΓ + F (t) generates an evolution operator e that is U (t, τ ) on H, d U (t, τ ) + (AΓ + F (t))U (t, τ ) = 0, 0 ≤ τ ≤ t dt U (τ, τ ) = I. 9 Lemma 4.3 Let p γ the number fixed as in Notation. We have for η ≥ max1≤j≤N 6γ − 2Re λj µ Z t ¶ −γ(t−τ ) kU (t, τ )kL(H, (1+η 2)|x| 1+ e−γ(τ +2s)ζ(s)ds , e H) e ≤Ce (4.29) τ ∀ t ≥ τ, P-a.s., P βj (t) . where C is independent of ω and ζ(t) = N j=1 e Proof. We shall use the spectral decomposition of the system dy + AΓy + F (t)y = 0, t ≥ τ dt (4.30) y(τ ) = x. in the direct sum Xu ⊕ Xs of γ-unstable and γ-stable spaces of the operator A. Namely we set yu = PN y, ys = (I − PN )y and, since by (4.27), PN F (t)y = 0, we may rewrite system (4.30) as N X dy 1 u +A y + P Γ2j yu = 0, t ≥ τ, u u N dt 2 (4.31) j=1 y (τ ) = P x. u and (4.32) N N X dy 1 s 2 dt + Asys + 2 (I−PN ) Γj yu+(I−PN )F (t)yu=0, j=1 t≥τ ys(τ ) = (I − PN )x. 10 We have y = yu + ys, yu = PN j=1 yj ϕj Auϕj = λj ϕj , and, by (4.9), j = 1, ..., N. Recalling that Γ2j y = ηΓj y = η 2hy, ϕ∗j iHe P (mφj ), we may rewrite (4.31) as 1 dyj + λj yj + η 2yj hP (mφj ), ϕ∗j iHe = 0, dt 2 y (τ ) = hx, ϕ∗i . j e j H t ≥ τ, j = 1, ..., N, Taking into account (4.15), it follows that 1 dyj + λj yj + η 2yj = 0, t ≥ τ, j = 1, ..., N, dt 2 y (τ ) = hx, ϕ∗i . j e j H This yields 1 2 yj (t) = e−(λj + 2 η )thx, ϕ∗j iHe , j = 1, ..., N, t ≥ 0. Hence, for η 2 ≥ 6γ − 2 Re λj , j = 1, ..., N, we have (4.33) |yu(t)|He ≤ Ce−3γ(t−τ )|x|He , ∀ t ≥ τ. Now, coming back to system (4.32), we shall rewrite it as N X dy 1 s + Asys + η 2 yj (I − PN )P (mφj ) dt 2 j=1 X N ³ ´ β (t) j (4.34) + e −1 yj (I−PN )(AP (mφj )−λj P (mφj ))=0, j=1 t ≥ τ, ys(τ ) = (I − PN )x. 11 Then by (4.33) and (4.8) we have that |ys(t)|He ≤ |e−As(t−τ )(I − PN )x|He Z N ´ 1 2 t X ³ βj (s) + η e − 1 |e−As(t−s)yj (s)(I − PN ) 2 τ j=1 ×|P (mφj ) + AP (mφj ) − λj P (mφj )|He ds Z tX N 2 η −γ(t−τ ) ≤ Ce |x| + C |x|H |e−3γse−γ(t−s)ζ(s)ds 2 0 j=1 Z t ≤ Ce−γ(t−τ )(1+η 2)|x|H e−γ(τ +2s)ζ(s)ds, ∀ t ≥ 0, P-a.s., τ for some constant C independent of x and ω ∈ Ω. This completes the proof. Now, we fix η. Lemma 4.4 We have Z ∞ eγ(t−τ )|U (t, τ )x|2W dt τ µ ¶2 Z ∞ (4.35) ≤ C|x|2W 1 + e−γ(τ +2t)ζ(t)dt , ∀ x ∈ W, τ where C is independent of ω ∈ Ω, 0 ≤ ² < 14 . Proof. We set z(t) := e γ 2 (t−τ ) U (t, τ )x, Then, by Lemma 4.3, we have µ Z ∞ Z |z(t)|2He dt ≤ C|x|2 1 + τ ∞ τ 12 0 < τ < t. ¶2 e−γ(τ +2t)ζ(t)dt , ∀ x ∈ H, while N dz 1 X 2 γ + ν0Az + A0z + Γj z + F (t)z = z, dt 2 j=1 2 t ≥ τ. 1 e we get (4.35). Multiplying the latter by z and A 2 +2²z (scalarly in H) We come back to (4.24) and set G(t, y) := e − PN j=1 βj (t)Γj B(e PN j=1 βj (t)Γj e t ≥ 0. ∀ y ∈ H, y), Recalling (4.21) and (4.26), we see that (4.36) B(eβj (t)Γj y) = B(y) + hy, ϕ∗j i2He (eηβj − 1)2B(P (mφj )) = (eηβj (t) − 1)hy, ϕ∗j iHe [B1(y, P (mφj )) + B2(y, P (mφj ))], where B(y) = P ((y · ∇)y) and (4.37) B1(y, z) = P ((y · ∇)z), B2(y, z) = P ((z · ∇)y), ∀ y, z ∈ D(A). Then by (4.26),(4.27) and (4.36) we have, for all j, k = 1, ..., N , e−βk (t)Γk B(eβj (t)Γj y) = e−βk (t)Γk [B(y) + hy, ϕ∗j i2He (eηβj − 1)2B(P (mφj )) +(eηβj (t) − 1)hy, ϕ∗j iHe [B1(y, P (mφj )) + B2(y, P (mφj ))]. We write (4.24) as dy(t) + AΓy(t) + G(t, y(t)) + F (t)y(t) = 0, dt 1 ∀ t ≥ 0, P-a.s. We set z(t) = e 2 γty(t) and rewrite it as γ 1 dz(t) + (AΓ − γ)z(t) + e− 2 tG(t, z(t)) + F (t)z(t) = 0 (4.38) dt 2 z(0) = x. 13 Equivalently, Z (4.39) t z(t) = S(t, 0)x − γ S(t, s)e− 2 sG(s, z(s))ds, ∀ t ≥ 0, 0 where 1 S(t, τ ) = U (t, τ )e− 2 γ(t−τ ). Lemma 4.5 There is Ωx ⊂ Ω, with ´− γ 2 ³ 1 − 8(ηN ) P(Ωx) ≥ 1 − C ∗|x|W2 − 1 , with C ∗ > 0 independent of ω and x such that for each x ∈ X with |x|W ≤ C ∗ equation (4.39) has a unique solution z ∈ C([0, ∞); W )) ∩ L2(0, ∞; Z). Here 1 W = D(A 4 ), 3 Z = D(A 4 ) 1 if d = 2 and 3 W = D(A 4 +²), Z = D(A 4 +²) if d = 3. Proof. We shall proceed as in the proof of Theorem 3.6. Namely, we rewrite (4.39) as z(t) = S(t, 0)x + N z(t) := Λz(t), t ≥ 0, where N : L2(0, ∞; Z) is the integral operator Z t − γ2 s N z(t) = − S(t, s)e G(s, z(s))ds. 0 We set U(ω) := {z ∈ L2(0, ∞; Z) : |z|L2(0,∞;Z) ≤ R(ω)}, 14 where R : Ω → R+ is a random variable such that 2C1∗|x|W p 1 + 1 − 4(C1∗)2|x|W (1 + µ)2 (4.40) 2C1∗|x|W p ≤ R(ω) ≤ , ∗ 2 2 1 − 1 − 4(C1 ) |x|W (1 + µ) Then, as easily follows for ω ∈ Ω. |x|W ≤ ρ1(ω) := [8(1 + µ(ω)2)(C1∗)2]−1, (4.41) we have ΛU(ω) ⊂ U(ω). Now, we shall apply the Banach fixed point theorem to Λ on the set U(ω). Let z1, z2 ∈ U (ω). We find that |N z1 − N z2|L2(0,∞;Z) µ Z ∞ ¶ Z ∞ ≤ C1∗ e−γt|G(t, z1) − G(t, z2)| 1 dt 1 + e−2γsζ(s)ds 4 0 Z 0 ∞ ≤ C1∗C2∗ (1 + δ(t))e−γt|z1(t) 0 µ Z ∞ ¶ −z2(t)|Z (|z1(t)|Z + |z2(t)|Z )dt 1 + e−2γsζ(s)ds µZ ≤ C1∗C2∗ µZ 0 ∞ × ¶ 21 ∞ |z1(t) − z2(t)|2Z dt e−γt(|z1(t)|2Z + |z2(t)|2Z )dt 0 ¶ 12 (1 + µ(ω))2 0 ∗ ∗ ≤ 2C1 C2 (1 + µ(ω))2R(ω)|z1 − z2|L2(0,∞;Z), where C1∗, C2∗ are independent of ω. Now, we take (4.42) |x|W ≤ ρ(ω) := min{ρ1(ω), ρ2(ω)} = ((C ∗)2(1 + µ)2)−1, 15 where C ∗ is a suitable chosen constant independent of ω. Then, for x satisfying (4.42), N is a contraction on U(ω) and maps U(ω) on itself. We set (4.43) Ωx = {ω ∈ Ω : |x|W ≤ ρ(ω)}. Hence, for each ω ∈ Ωx, the equation (4.39) has a unique solution z satisfying conditions in Lemma 4.5. On the other hand, by (4.42) and (4.43) we see that ³ ´− γ 2 1 − 8(ηN ) P(Ωx) ≥ 1 − C ∗|x|W2 − 1 , as claimed. Comments The main result, Theorem 4.1, is taken from [5]. A similar result for the linearized Oseen–Stokes system was earlier obtained in [1]. Noise stabilizable feedback controller were designed in [3] for periodic flows in a 2 − D channel and also in [4] for controllers with support in a discrete set of points. For the Oseen-Stokes systems is designed in [6] a stabilizable Stratonovich controller of the form u = 1lO0 M X Rk (X) · β̇k , k=1 which has the advantage to be stable to the perturbation of βk . 16 References [1] V. Barbu, The internal stabilization by noise of the linearized Navier-Stokes equations, ESAIM COCV, (2011). [2] V. Barbu, Stabilization of Navier–Stokes Flows, SpringerVerlag, London, 2010. [3] V. Barbu, Stabilization of a plane periodic channel flow by noise wall normal controllers, Systems & Control Letters, vol. 59 (2010), 608-618. [4] V. Barbu, Exponential stabilization of the linearlized Navier– Stokes equation by pointwise feedback controllers, Automatica, vol. 46 (2010), 2022-2027. [5] V. Barbu, Internal stabilization of the Oseen-Stokes equations by Stratonovich noise, Systems & Control Letters (to appear). [6] V. Barbu, G. Da Prato, Internal stabilization by noise of the Navier-Stokes equations, SIAM J. Control & Optim., vol. 49 (2011), 1-21. [7] G. Da Prato, An Introduction to Infinite Dimensional Analysis, Springer-Verlag, Berlin, 2006. [8] G. Da Prato, J. Zabczyk, Ergodicity for Infinite Dimensional Systems, London Mathematical Society Lecture Notes, n. 229, Cambridge University Press, 1996. 17