5.5 Alternating Direction Implicit (ADI) As motivation consider the two variable Poisson equation − u xx − u yy = f . Or − u xx = f + u yy . Fix y, and this is a sequence of odes with respect to x. − u yy = f + u xx Fix x, and this is a sequence of odes with respect to y. Solve the following sequences of odes: 1 − uxx m + 2 = f + u yy m 1 − u yy m +1 = f + uxx m + 2 Or, applying the FDM and solve a sequence of tridiagonal systems: 1 1 1 −ui -1, j m + 2 +2ui, j m + 2 − ui +1, j m + 2 ∆x 2 −ui , j -1 m +1 m +1 + 2ui , j ∆y 2 + −ui , j -1 m +2ui, j m −ui , j +1m m +1 − ui , j +1 = fi, j − ∆y 2 −ui -1, j m+ 1 2 m+ + 2ui, j ∆x 2 1 2 = fi , j − ui +1, j m++ 1 2 . → → There is a sequence of tridiagonal solves that are independent of j. → ↑ ↑ ↑ There is a sequence of tridiagonal solves that are independent of i. 2 Matrix Form. A = H +V Aˆ H= O ˆA 2 -1 ˆA = -1 O O O O 2I -I V = - I O O O O ADI Splittings. A = (α I + H ) - (α I -V ) A = (α I + V ) - (α I - H ) where α ∈¡ ,or α is a diagonal matrix. ADI Algorithm. A = H + V and Ax = d for m = 0, maxm solve (α I + H ) x m+ 1 2 = d + (α I -V )x m solve (α I + V ) x m +1 = d + (α I - H ) x test for convergence. m+ 1 2 3 Proposition 8. Let A = H + V and α >0. If A,H,V are SPD, then for all x0, xm+1→ x, where Ax = d. Proof. 1 m+ u m +1 = (α I + V ) -1 d + (α I - H ) u 2 -1 = (α I + V ) d + (α I - H )(α I + H )-1{d + (α I -V )u m } ˆ m + g where = Hu Hˆ = (α I + V ) -1 (α I - H )(α I + H )-1 (α I -V ) and g = (α I + V )-1 [ d + (α I - H )(α I + H )-1 d ]. ˆ ˆ < 1 for some norm. Show ρ (H)<1 or H Note the following: ˆ α I+V) -1 = (α I-H)(α I+H) -1 (α I-V)(α I+V) -1 (α I+V)H( = FG where F ≡ (α I-H)(α I+H) -1 G ≡ (α I-V)( α I+V)-1. ˆ σ (FG) σ (H)= Find so that F < 1 and G < 1. Consequently, FG ≤ F G < 1. 4 We will use the following facts: 1. F 2 Fx = sup x x 2 2 = ρ (F) where 2 ≡ (x T x)1/2 and F is symmetric. In order to show this, represent x by an orthonormal basis of eigenvectors of F. 2. H, V symmetric implies that F and G are symmetric. 3. Eigenvalues of F and G have the form α −λ where α +λ Hx = λx or Vx = λx λ >0 because H and V are SPD. Therefore, F 2 = (α I-H)( α I+H)-1 α−λ <1 α+λ 2 < 1 and G 2 ; λ >0 and α ≥ 0 α − λ α−λ = α + λ α+λ − α − λ α + λ α ≥λ α < λ. < 1 because 5 Application to 3D Problem. −u xx − u yy − u zz = f Apply the FDM and form A = H + V + W where W is associated with the z direction. A = (α I + H ) - (α I - V -W ) = (α I + V ) - (α I - H - W ) = (α I + W ) - (α I - H -V ). From these three splittings one can form the three stage ADI algorithm, and prove an analogue to Proposition 8.