A <strong>FETI</strong>-<strong>DP</strong> FORMULATION FOR THE STOKES PROBLEM 5where S is given by⎛⎞K II K I∆ BIT(2.7) S = ⎜⎝KI∆ T K ∆∆ B∆T ⎟⎠ .B I B ∆ 0Substituting (u I , u ∆ , p) into (2.5) and <strong>the</strong>n solving <strong>for</strong> û Π⎛ ⎞K IΠ(2.8) S ΠΠ û Π = f Π − ⎜⎝K ∆Π⎟⎠B Πwe obtain <strong>the</strong> resulting algebraic system on λ,(2.9) F <strong>DP</strong> λ = d,where(2.10)⎛ ⎞0F <strong>DP</strong> = ⎜⎝J∆T ⎟⎠0⎛ ⎞0d = ⎜⎝J∆T ⎟⎠0TTT⎛⎛⎞ ⎛ ⎞ ⎞f I 0S −1 ⎜⎜⎝⎝f ∆⎟⎠ − ⎜⎝J∆T ⎟⎠ λ ⎟⎠ ,0 0⎛ ⎞ ⎛ ⎞T⎛ ⎞ ⎛ ⎞T⎛ ⎞0 0 K S −1 ⎜⎝J∆T ⎟⎠ + IΠ K IΠ0⎜⎝J∆T ⎟⎠ S −1 ⎜⎝K ∆Π⎟⎠ S−1 ⎜ΠΠ ⎝K ∆Π⎟⎠ S −1 ⎜⎝J T ⎟∆⎠ ,0 0B Π B Π 0⎛⎛⎞ ⎛ ⎞ ⎛ ⎛ ⎞T⎛ ⎞⎞⎞f I K S −1 ⎜⎜⎝⎝f ∆⎟⎠ − IΠK ⎜⎝K ∆Π⎟⎠ S−1 ⎜ΠΠ ⎝ f IΠ f IΠ − ⎜⎝K ∆Π⎟⎠ S −1 ⎜⎝f ∆⎟⎟⎟⎠⎠⎠ ,0 B Π B Π 0and⎛ ⎞K IΠS ΠΠ = K ΠΠ − ⎜⎝K ∆Π⎟⎠B ΠT⎛ ⎞K IΠS −1 ⎜⎝K ∆Π⎟⎠ .B ΠThe resulting system on λ is symmetric and positive semidefinite. In a more detail, when<strong>the</strong> velocity <strong>unknowns</strong> at subdomain vertices are selected as <strong>the</strong> <strong>primal</strong> <strong>unknowns</strong>, it has onenull space component which is given by(2.11)( )(1) µ ∣∣Γij0µ (2) =0( )ζij n (1)ijζ ij n (2) , ∀Γ ij .ijHere, µ (1)0 and µ (2)0 are Lagrange multipliers related to each x and y- components <strong>of</strong> velocity<strong>unknowns</strong>, n (k)ij are each component <strong>of</strong> n ij , <strong>the</strong> unit normal vector to Γ ij , and∫(2.12) ζ ij (x l ) = φ l (x(s), y(s)) ds,Γ ij
6 H.H. Kim, C.-O. Lee, and E.-H. Parkwhere φ l is <strong>the</strong> velocity basis element related to <strong>the</strong> node x l at Γ ij . For <strong>the</strong> details, we refer[5, Section 2.2].We now introduce a subspace <strong>of</strong> M, which is orthogonal to <strong>the</strong> null space <strong>of</strong> F <strong>DP</strong> ,⎧⎫⎨M c =⎩ µ ∈ M : ∑ ⎬µ ij · ζ ij n ij = 0⎭ ,ijwhere µ ij = µ| Γij . Then F <strong>DP</strong> is positive definite on M c . The system in (2.9) is <strong>the</strong>n solvedby <strong>the</strong> conjugate gradient method with a lumped preconditioner <strong>of</strong> <strong>the</strong> <strong>for</strong>m,(2.13) ̂M −1 = J ∆ K ∆∆ J T ∆.In our previous work [5], we proved <strong>the</strong> following condition number bound <strong>for</strong> <strong>the</strong> <strong>FETI</strong>-<strong>DP</strong> algorithm equipped with <strong>the</strong> lumped preconditioner and with <strong>the</strong> velocity <strong>unknowns</strong> atsubdomain vertices as <strong>primal</strong> <strong>unknowns</strong>,κ(̂M −1 F dp )) ≤ C(H/h)(1 + log(H/h)),which determines <strong>the</strong> convergence <strong>of</strong> <strong>the</strong> conjugate gradient iteration. The same bound hasbeen proved to be optimal <strong>for</strong> <strong>the</strong> <strong>FETI</strong>-<strong>DP</strong> algorithm <strong>of</strong> <strong>the</strong> elliptic problems with a lumpedpreconditioner, see [13].In <strong>the</strong> work by Li and Widlund [11], both <strong>the</strong> velocity <strong>unknowns</strong> at <strong>the</strong> subdomain verticesand <strong>the</strong> velocity averages on subdomain edges are used as <strong>the</strong> <strong>primal</strong> velocity <strong>unknowns</strong>.In addition, <strong>primal</strong> pressure <strong>unknowns</strong> are included in <strong>the</strong>ir <strong>FETI</strong>-<strong>DP</strong> <strong>for</strong>mulation. They introduceda quite expensive Dirichlet preconditioner and obtained a condition number boundC(1 + log(H/h)) 2 . Due to <strong>the</strong> introduction <strong>of</strong> <strong>the</strong> <strong>primal</strong> pressure <strong>unknowns</strong>, <strong>the</strong>ir approachneeds both <strong>of</strong> <strong>the</strong>m, i.e., velocity <strong>unknowns</strong> at subdomain vertices and averages <strong>of</strong> <strong>the</strong> velocityon edges, to provide <strong>the</strong> stability <strong>of</strong> <strong>the</strong> coarse problem matrix as well as to make<strong>the</strong>m satisfy zero flux condition across subdomain interfaces. In <strong>the</strong>ir experimental work, a<strong>FETI</strong>-<strong>DP</strong> algorithm with <strong>primal</strong> velocity <strong>unknowns</strong> at subdomain vertices is tested. Its convergencedepends on <strong>the</strong> number <strong>of</strong> subdomains and additional <strong>primal</strong> <strong>unknowns</strong> are requiredto achieve a scalable algorithm.<strong>On</strong> <strong>the</strong> o<strong>the</strong>r hand, <strong>for</strong> two dimensional elliptic problems, it is well known that ei<strong>the</strong>r<strong>primal</strong> <strong>unknowns</strong> at subdomain vertices or <strong>primal</strong> <strong>unknowns</strong> related to <strong>the</strong> averages on subdomainedges are enough to obtain a scalable condition number bound, which means that <strong>the</strong>condition number bound only depends on <strong>the</strong> local problem size, see [12].No pressure <strong>primal</strong> <strong>unknowns</strong> in <strong>the</strong> <strong>FETI</strong>-<strong>DP</strong> algorithm <strong>of</strong> our work [5] resulted in ascalable method <strong>for</strong> <strong>the</strong> Stokes problems with only <strong>the</strong> <strong>primal</strong> velocity <strong>unknowns</strong> at subdomainvertices. Its condition number bound is <strong>the</strong> same as that <strong>of</strong> elliptic problems with aninexact lumped preconditioner, see [13]. We note that in <strong>the</strong> work [13] a better condition
- Page 1 and 2: On the selection of primal unknowns
- Page 3 and 4: 2 H.H. Kim, C.-O. Lee, and E.-H. Pa
- Page 5: 4 H.H. Kim, C.-O. Lee, and E.-H. Pa
- Page 9 and 10: 8 H.H. Kim, C.-O. Lee, and E.-H. Pa
- Page 11 and 12: 10 H.H. Kim, C.-O. Lee, and E.-H. P
- Page 13 and 14: 12 H.H. Kim, C.-O. Lee, and E.-H. P
- Page 15 and 16: 14 H.H. Kim, C.-O. Lee, and E.-H. P