7.4 Unsymmetric Weak Recovery
In this section we show how suitable sampling inequalities can be used to derive upper bounds for a weak unsymmetric recovery process from [47]. We briefly recall this process.
Since we are going to apply a non-symmetric method, we have to distinguish carefully between the tRial and the teSt side. As outlined in Section 2.3.1 we shall denote the test kernel withS:Rd×Rd→Rand the discretization byXs={x1, . . . , xNs} ⊂Ωwith the associated fill distancehs. On the tRial side we denote the kernel withR:Rd×Rd→R and the discretization withYr = {y1, . . . , yNr} ⊂ Ωwith fill distancehr. According to [45, Chapters 3.4 and 3.5] we use weak data of the form given in (2.1.7), where we replace KbyS, namely
λj(u) = Z
Ω
S(x, xj)u(x)dx , forxj ∈Xs, (7.4.1) where the kernelS:Rd×Rd→Rneeds to fulfill the following properties,
1. R
ΩS(x, xj)dx= 1 for all xj ∈Xs,
2. supp(S(·, xj)) =Vj , diam(Vj) =δVj ≈hs, 3. kS(·, xj)kLp(Ω) ≤CδV−jd/q≈h−sd/q, 1p+ 1q = 1.
Under these conditions we may apply Theorem 7.2.1, i.e., there is a sampling inequality of the form
|u|Wp|α|(Ω) ≤Chτ+d/p−|α|s |u|Wpk+s(Ω)+ ˜Ch−|α|s max
1≤j≤N|λj(u)|
for all u ∈ W2τ(Ω)and all discrete test sets Xs. Now we consider the following prob-lem. An unknown function f ∈ W21(Ω) has to be recovered approximately from its data (λ1(f), . . . , λN(f))T. From the previous chapters we know that there is a good but unknown approximation. It is given by the best approximation from the trial space sf ∈ VR,Yr = span{R(· −yj) |yj ∈Yr} to the functionf ∈ W21(Ω). In Chapter 8 we obtain under certain conditions an error estimate of the form
kf −sfkL2(Ω)≤hrkfkW21(Ω) , (7.4.2) showing linear approximation order, which is optimal [48]. Unfortunately, this best appro-ximation is in general numerically unavailable. The idea from [47] is to solve the system
findsr∈VR,Yr :λj(f−sr) = 0for all1≤j≤Ns
to produce an approximation. Since this, however, is an unsymmetric kernel method, we first have to prove that this system is solvable. To do so, we proceed along the lines of [47], which we briefly outlined in Section 2.3.1. We assume an inverse inequality of the form
kskW21(Ω) ≤Cγ(Yr)kskL2(Ω) for alls∈VR,Yr .
Unfortunately, the value ofγ(Yr)is in general not known. There is a result in this direction in Chapter 6, namely ifRis a radial basis function with algebraic smoothnessτ > d/2and
ifYr is separated from the boundary thenγ(Yr) ≈ qr−τ. However, we can always make sure that the test-meshnorm is small enough to stabilize the reconstruction, i.e., we shall assume
hsγ(Yr)≤ 1
2C , C >1. (7.4.3)
Now we can prove the full-rank of the unsymmetric reconstruction of a function f ∈ W21(Ω)by an approximationsr ∈VR,Yr.
Lemma 7.4.1 Under the condition (7.4.3) the system
λj(f−sr) = 0, for all1≤j≤Ns. (7.4.4) is of full rank.
Proof:We have forsr=PNr
ℓ=1aℓΦ (· −yℓ)the equivalence λj(f−sr) = 0for all1≤j≤Ns ⇔
λj(f) =
Nr
X
ℓ=1
aℓλj(Φ (· −yℓ)) for all1≤j≤Ns. (7.4.5) If we introduce the notation
A := (Aℓ,j)1≤ℓ≤Nr 1≤j≤Ns :=
Z
Ω
R(x−yℓ)S(x−xj)dx
1≤ℓ≤Nr 1≤j≤Ns
∈RNr×Ns ,
F :=
λ1(f) ... λNs(f)
∈RNs ,
equation (7.4.5) leads to the unsymmetric linear system
Aa=F . (7.4.6)
Now we can compute
kurkL2(Ω) ≤ C
hskurkW21(Ω)+ max
1≤j≤Ns|λj(ur)|
≤ C
hsγ(Xr)kurkL2(Ω)+ max
1≤j≤Ns|λj(ur)|
≤ 1
2kurkL2(Ω)+C max
1≤j≤Ns|λj(ur)| . This yields
kurkL2(Ω)≤2C max
1≤j≤Ns|λj(ur)| .
7.4. UNSYMMETRIC WEAK RECOVERY 103 Therefore,max1≤j≤Ns|λj(ur)| = 0implieskurkL2(Ω) = 0. Sinceur ∈ VR,Yr we have ur ≡0. Thus rank ofA= #Yr ≤#Xs, which means that the system is of full rank. 2 Equation (7.4.2) implies that we can theoretically solve the system (7.4.4) to some pre-scribed accuracyδ(r, s). This is due to the fact sf is a good candidate for a solution, and by choosinghssmall enough we can achieve every prescribed valueδ(r, s). Practically we can apply some residual minimization techniques to find an approximate solution, such that the residual is smaller thanδ(r, s). We denote this approximate solution byur,s ∈ VR,Xr, i.e., we have
|λj(f−ur,s)| ≤δ(r, s)for all1≤j≤Ns. The exact error bound is given in the following theorem.
Theorem 7.4.2 We denote byur,s ∈ VR,Yr the approximate solution of the system (7.4.4) to the accuracyδ(r, s). Then we have an error bound of the form
kf −ur,skL2(Ω)≤C
2hs+ 1 C1hr
kfkW21(Ω)+δ(r, s). If we chooseδ(r, s) =
2hs+ C1
1hr
kfkW21(Ω), we get a final bound of the form kf−ur,skL2(Ω)≤C
2hs+ 1 C1hr
kfkW21(Ω) .
Proof:Let us denote byurthe unknown best trial approximation. Then we can compute 1
C kf−ur,skL2(Ω) ≤hskf−ur,skW21(Ω)+ max
1≤j≤Ns|λj(f −ur,s)|
≤ hskfkW21(Ω)+hskur,skW21(Ω)+ max
1≤j≤Ns|λj(f −ur,s)|
≤ hskfkW21(Ω)+hskurkW21(Ω) + +hskur,s−urkW21(Ω)+ max
1≤j≤Ns|λj(f−ur,s)|
≤ 2hskfkW21(Ω)+hsγ(Xr)kur,s−urkL2(Ω)+ max
1≤j≤Ns|λj(f−ur,s)|
≤ 2hskfkW21(Ω)+ 1
2C1kur,s−urkL2(Ω)+ max
1≤j≤Ns|λj(f−ur,s)|
≤ 2hskfkW21(Ω)+ 1 2C1
kur,s−fkL2(Ω)+kur−fkL2(Ω)
+
+ max
1≤j≤Ns|λj(f−ur,s)|
≤ 2hskfkW21(Ω)+ 1
2C1 kur,s−fkL2(Ω)+ 1
2ChrkfkW21(Ω)
+ max
1≤j≤Ns|λj(f−ur,s)| . This shows
kf−ur,skL2(Ω)≤C
2hs+ 1 2C1
hr
kfkW21(Ω)+ max
1≤j≤Ns|λj(f −ur,s)| ,
giving the expected error bound kf −ur,skL2(Ω)≤C
2hs+ 1 2C1
hr
kfkW21(Ω)+δ(r, s).
The second inequality follows from the special choice forδ(r, s). 2 This generalizes the results of [47] in two ways. Our error analysis does not assumef to be known on a slightly larger domainΩ. Furthermore, we obtain results for the˜ L2-norm, not for negative order Sobolev norms.
Chapter 8
Galerkin Methods
In this chapter we shall combine the well-known theory of variational problems with the idea of sampling inequalities. Galerkin methods are a powerful tool for solving elliptic partial differential equations. We shall extend the theory given in [62] and show a new application in the theory of weak recovery [47]. Unfortunately, the error estimates in [62]
are given in theW21(Ω)-norm, and they require the solutionuto be in a native space of a radial basis function, which automatically forcesuto be continuous. In contrast, we shall give error estimates in theL2-norm, hence we do not assumeuto be continuous.
8.1 Model Problem: Elliptic Partial Differential Equations
We shall concentrate on problems of the following form. LetΩbe a bounded domain with boundary∂Ω∈C1. We shall consider partial differential equations of the form
− Xd i,j=1
∂
∂xi
ai,j(x) ∂u
∂xj(x)
+c(x)u(x) = f(x) , x∈Ω Xd
i,j=1
ai,j|∂Ω(x)∂u(x)
∂xj νi(x) +h(x)u(x) = g(x) , x∈∂Ω, (8.1.1) where we assumeai,j, c ∈L∞(Ω),f ∈ L2(Ω),ai,j|∂Ω, h ∈L∞(∂Ω)andg ∈ L2(∂Ω) fori, j = 1, . . . , d. Further, we denote byν the outward pointing unit normal vector to the boundary∂Ω.
The weak formulation of (8.1.1) is given in terms of the bilinear form apde(u, v) :=
Z
Ω
Xd i,j=1
ai,j∂u(x)
∂xj
∂v(x)
∂xi
+cuv
dx+ Z
∂Ω
huvdS (8.1.2)
and the linear form
Fpde(v) :=
Z
Ω
f vdx+ Z
∂Ω
gvdS (8.1.3)
105
for a functionv ∈W21(Ω). The matrixApde(x) := (ai,j(x))is assumed to be uniformly elliptic onΩ, i.e., there is a constantγpdesuch that
γpdekαk22 ≤αTA(x)α
for allα∈Rdand allx∈Ω. This gives rise to the variational problem
find u∈W21(Ω) : apde(u, v) =Fpde(v) for allv∈W21(Ω) . (8.1.4) If furtherc, h ≥0, and if at least one of them is uniformly bounded from zero on a subset of nonzero measure of Ω or ∂Ω, we know that apde(·,·) : W21(Ω)×W21(Ω) → R is strictly coercive, and thatF :W21(Ω)→Ris continuous. We point out that the boundary conditions are part of the weak formulation. This unfortunately excludes Dirichlet boundary conditions at the moment.