Probabilistic Solution of the Navier-Stokes Equations on a 2D Torus April 23, 2014 1 NS Equations The Navier-Stokes (NS) PDEs model the time evolution of P components of the velocity, u : D → RP , of an incompressible fluid on a spatial domain X . The NS boundary value problem on the spatio-temporal domain D = X × T is given by: ( ) ∂ ∂t u − θ ∆u + u · ∇ u ∇·u ∫ (j) u dx u = f − ∇p, (x, t) ∈ D, = 0, (x, t) ∈ D, = 0, (x, t) ∈ D, j = 1, 2, = uB , (x, t) ∈ X × {0}, (1) where ( ∆u = ∂2 ∂x(1) 2 u1 + ∂2 ∂x(2) 2 u1 , ∂2 ∂x(1) 2 u2 + ∂2 ∂x(2) )⊤ 2 u2 = 2 ∑ m=1 ( 2 ∑ ∂2 m′ =1 ∂x(m′ ) ) 2 em u and ( ) ( 2 ( )⊤ ∑ 2 ∑ ∂ ∂ ∂ ∂ ∂ (m′ ) u · ∇ ũ = u1 (1) ũ1 + u2 (2) ũ1 , u1 (1) ũ2 + u2 (2) ũ2 u em ũ(m) = (m′ ) ∂x ∂x ∂x ∂x ∂x ′ m=1 ) m =1 This model is parameterized by the viscosity of the fluid, θ > 0; the pressure function p : D → R; and the external time-homogeneous forcing function f : X → R. We consider the NS equations over a 2-dimensional torus shaped domain, X = [0, 2π) × [0, 2π), expressed in spherical coordinates. We further assume periodic boundary conditions (second and third lines of equation 1), and viscosity θ = 1 ×10−3 in the turbulent regime. For simplicity, we consider the unforced NS equations (f = 0). Often, the quantity of interest is the vorticity, or local spinning motion of the incompressible fluid, which we define as, ϖ(x, t) = −∇ × u(x, t), where clockwise rotation corresponds to positive vorticity. Vorticity can also help to better visualize the solution of the NS system by summarizing the two components of velocity by a one-dimensional 1 function. 2 Spectral projection for NSE The phase space of the solution to (1) at a given time point is the Sobolev space, { H := v ∈ L2 (X ) : ∂ v ∈ L2 (X ), ∇ · v = 0, ∂x(i) ∫ v (j) dx = 0 } and let H be its closure. This is a subspace of L2 , endowed with the L2 inner product. Also define ( )2 P : L2 (X ) → H. We want to project (1) to the subspace H. Now we can rewrite (1) as du − θ Au + B (u, u) = −P + f, u(0) = v 0 ∂t (2) where P is the solution to the Poisson equation, ( ) ∇· u·∇ u=− 2 ∑ ∂ ∂x(m) m=1 ( 2 ∑ u (m′ ) m′ =1 ∂ ∂x(m′ ) ) u(m) = − 2 ∑ m=1 2 ∑ ∂2 ∂x(m) ∂2 (m) m=1 ∂x 2p 2p Next, we replace u(x, t) in (2) by u(x, t) = ûk (t) ϕk (x), were ûk (t) = u(x, t) ϕk (x), k = Z2 \{0}, and hereafter omit spatial and temporal dependence in the Fourier modes and coefficients from the notation. We then project (2) on the Fourier bases to obtain, (j) û˙ ℓ = −θ 2 ( 2 ) ∑ ∑ (m) 2 (j) (m) (m) (j) (j) wℓ ûℓ − i ûk wℓ−k ûℓ−k + fˆℓ + ∑ 2 m=1 1 ∑ (m) 2 m,m′ m=1 wℓ m=1 (m′ ) (m) ûℓ−k wℓ (m′ ) (m) ûk wk (3) with initial condition (j) (j) ûℓ (0) = ⟨v, ϕℓ ⟩ = v̂ℓ . 2.1 Derivation of expression (3) The Fourier bases, (j) (1) (2) ϕk := exp(iwk x(1) ) exp(iwk x(2) ) = exp(iwk · x), 2 with wavenumbers, 2 ∑ −1 wk = (2π) em k, k ∈ Z2 \{0}, m=1 form an orthonormal basis for H, where ej is the unit vector in the jth direction We will use the partial derivatives of these Fourier modes, ∂ (j) (m) (j) (m) ϕ = iwk exp(iwk · x) = iwk ϕk ∂x(m) k ( ) ∂ 2 (j) (m) 2 (m) (j) ϕ = − w exp(iwk · x) = −wk ϕk . k ∂x(m) k and the orthonormality property, { (j) (j) ⟨ϕk , ϕℓ ⟩ = 0 if k ̸= ℓ; 1 if k = ℓ. where ⟨·, ·⟩ represents the L2 norm. We also use the fact that { (j) (j) ⟨ϕk ϕk′ , ϕℓ ⟩ = 0 if k + k′ ̸= ℓ; 1 if k + k′ = ℓ. We need to project (2) to the ℓth Fourier mode. Let us do this for each term of the equation separately. Denote by the subscript j the jth component of each vector. Therefore, ) ( d (j) ∂ (j) (j) (j) (j) ϕk , ϕ ℓ ⟩ û ⟨ u , ϕℓ ⟩ = ⟨ ∂t dt k d (j) (j) (j) = ûk ⟨ϕk , ϕℓ ⟩ dt d (j) (j) = ûℓ := û˙ ℓ dt The projection of the forcing function is just its Fourier transform, (j) (j) ⟨f (j) , ϕℓ ⟩ = fˆℓ The projection of the jth component of the Laplacian A on the ℓth Fourier mode is, ⟨Au (j) (j) , ϕℓ ⟩ =⟨ 2 ∑ ∂2 m=1 2 ∂x(m) 3 (j) u(j) , ϕℓ ⟩ = ⟨− 2 ( ) ∑ (m) 2 (j) (j) (j) wk ûk ϕk , ϕℓ ⟩ m=1 2 ( ) ∑ (j) (j) (m) 2 (j) =− wk ûk ⟨ϕk , ϕℓ ⟩ =− m=1 2 ( ∑ ) (m) 2 wℓ (j) ûℓ m=1 The projection of the jth component of B on the ℓth Fourier mode is, (j) ⟨B(u, u(j) ), ϕℓ ⟩ = ⟨ 2 ∑ m=1 2 ∑ =⟨ =i =i m=1 2 ∑ m=1 2 ∑ u(m) ∂ (j) u(j) , ϕℓ ⟩ ∂x(m) (m) (m) (m) (j) (j) (j) ûk ϕk iwk′ ûk′ ϕk′ , ϕℓ ⟩ (m) (m) (j) (m) (j) (m) (m) (j) (j) ûk wk′ ûk′ ⟨ϕk ϕk′ , ϕℓ ⟩ ûk wℓ−k ûℓ−k m=1 where the second to last line uses the fact that k′ = ℓ − k. In order to compute the projection of the jth component of pressure P , we first note that the assumption of a constant body force and divergence free condition is satisfied by letting 2 ∑ ∂ ∂x(m) m=1 ( 2 ∑ u (m′ ) m′ =1 ∂ ∂x(m′ ) ) u(m) = − 2 ∑ ∂2 m=1 ∂x(m) 2p we then project both sides to the ℓth Fourier mode as follows: 2 ∑ ∂ ⟨ ∂x(m) m=1 ( 2 ∑ m′ =1 u (m′ ) ∂ ∂x(m′ ) ) (j) u(m) , ϕℓ ⟩ ( 2 ) ∑ (m′ ) (m′ ) (m′ ) (m) (m) ∂ (j) = ⟨i ûk′ ϕk′ wk ûk ϕk , ϕℓ ⟩ (m) ∂x m=1 m′ =1 ∑ (m′ ) (m′ ) (m) ∂ (m′ ) (m) (j) ϕk ′ ϕk , ϕ ℓ ⟩ = ⟨i ûk′ wk ûk (m) ∂x m,m′ ) ∑ (m′ ) (m′ ) (m) ( (m′ ) (m′ ) (m) (m) (m′ ) (m) (j) = ⟨− ûk′ wk ûk wk′ ϕk′ ϕk + wk ϕk′ ϕk , ϕℓ ⟩ 2 ∑ m,m′ 4 =− ∑ (m′ ) (m′ ) (m) ûk ( ) (m′ ) (m′ ) (m) (j) (m) (m′ ) (m) (j) wk′ ⟨ϕk′ ϕk , ϕℓ ⟩ + wk ⟨ϕk′ ϕk , ϕℓ ⟩ (m′ ) (m′ ) (m) ûk ( ) (m′ ) (m) wℓ−k + wk (m′ ) (m) ûk′ wk m,m′ =− ∑ ûℓ−k wk m,m′ =− ∑ ûℓ−k wℓ (m′ ) (m) ûk wk m,m′ where the second to last line uses the fact that k′ = ℓ − k. And, ⟨− 2 ∑ m=1 ∂2 ∂x (j) p, ϕℓ ⟩ (m) 2 =⟨ 2 ∑ (m) 2 wk m=1 2 ∑ = p̂ℓ (j) (j) p̂k ϕk , ϕℓ ⟩ ( ) (m) 2 wℓ m=1 We can then equate these expressions and solve for p̂ℓ : − ∑ (m′ ) (m) (m′ ) (m) ûℓ−k wℓ wk ûk m,m′ = p̂ℓ 2 ∑ (m) 2 wℓ m=1 ∑ (m′ ) (m) (m′ ) (m) 1 p̂ℓ = − ∑ ûℓ−k wℓ wk ûk ( )2 (m) 2 ′ m,m m=1 wℓ 3 Discretization in space The Navier-Stokes model (1) was reduced to a set of 128 × 128 stiff coupled ODEs with associated constraints through a pseudo-spectral projection in Fourier space, described above. The spatial discretization grid was equally spaced. The initial velocity field is generated from a bivariate normal distribution at each of the mesh points. 4 Exponential time differencing Since each of the 128 × 128 ODEs have both linear and nonlinear terms, the system is stiff and the numerical solutionwill require prohibitively small time-steps. To overcome this problem, we use exponential time-differencing to solve the linear part of each ODE exactly, leaving only the nonlinear part. Note that the projected equation (3) can be written as follows: 5 ( (j) û˙ ℓ −θ = 2 ( ∑ ) (m) 2 (j) wℓ ûℓ ) + (j) fˆℓ m=1 + −i 2 ∑ m=1 (m) (m) ∑ (m′ ) (m) (m′ ) (m) 1 ûℓ−k wℓ wk ûk ( )2 (m) ′ w m,m m=1 ℓ (j) ûk wℓ−k ûℓ−k + ∑ 2 (j) := Lûℓ + N (û(ℓ−k) , ûk ; t), Therefore we can apply the method of exponential time-differencing by using the following transformation at each step: w = exp{−Lt} u which gives transforms the system to the following completely nonlinear ODE, d w = exp{−Lt} N (w; t) dt 6