Chapter IV: Proving exponential decay of Cholesky factors
4.2 Setting and notation
4.2.1 The class of elliptic operators
For our rigorous, a priori, complexity-vs.-accuracy estimates, we assume that G is the Greenβs function of an elliptic operator L of order 2π (π , π β N), defined on a bounded Lipschitz domainΞ© β Rπ, and acting onπ»π
0(Ξ©), the Sobolev space of (zero boundary value) functions having derivatives of order π in πΏ2(Ξ©). More
Figure 4.1: A regularity criterion. We measure the regularity of the distributions of measurement points as the ration ofπΏmin, the smallest distancebetween neighboring points or points and the boundary, andπΏmax, the radius of the largest ballthat does not contain any points.
precisely, writing π»βπ (Ξ©) for the dual space of π»π
0(Ξ©) with respect to the πΏ2(Ξ©) scalar product, our rigorous estimates will be stated for an arbitrary linear bijection
L: π»π
0(Ξ©) βπ»βπ (Ξ©) (4.4)
that issymmetric (i.e.β«
Ξ©π’Lπ£dπ₯ = β«
Ξ©π£Lπ’dπ₯), positive (i.e.β«
Ξ©π’Lπ’dπ₯ β₯ 0), and localin the sense that
β«
Ξ©
π’Lπ£dπ₯ =0 for allπ’, π£ βπ»π
0(Ξ©)such that suppπ’β©suppπ£=β . (4.5) Let kL k B supπ’βπ»π
0 kLπ’kπ»βπ /kπ’kπ»π
0 and kLβ1k B supπβπ»βπ kLβ1πkπ»π
0/kπkπ»βπ
denote the operator norms ofLandLβ1. The complexity and accuracy estimates for our algorithm will depend on (and only on)π , π ,Ξ©,kL k, kLβ1k, and the parameter
πΏ B πΏmin πΏmax B
minπβ πβπΌdist π₯π,{π₯π} βͺπΞ©
maxπ₯βΞ©dist(π₯ ,{π₯π}πβπΌ βͺπΞ©), (4.6) the geometric meaning of which is illustrated in Figure4.1.
46 4.2.2 Discretization in the abstract
Before talking about computation, we need to discretize the infinite-dimensional spaces π»π
0(Ξ©) and π»βπ (Ξ©) by approximating them with finite vector spaces. We first introduce this procedure in the abstract.
ForBa separable Banach space with dual space Bβ (such as π»π
0(Ξ©) andπ»βπ (Ξ©)), we write [ Β·, Β· ] for the duality product between Bβ and B. Let L: B β Bβ be a linear bijection and let G B Lβ1. Assume L to be symmetric and positive (i.e.
[Lπ’, π£] = [Lπ£ , π’]and[Lπ’, π’] β₯ 0 forπ’, π£ β B). Letk Β· kbe the quadratic (energy) norm defined bykπ’k2B [Lπ’, π’]forπ’ β Band letk Β· kβbe its dual norm defined by
kπkβ B sup
0β π’βB
[π, π’]
kπ’k =[π,Gπ]forπ β Bβ. (4.7) Let {ππ}πβπΌ be linearly independent elements ofBβ (known asmeasurement func- tions) and letΞβRπΌΓπΌ be the symmetric positive-definite matrix defined by
Ξπ π B [ππ,Gππ] forπ, π β πΌ. (4.8) We assume that we are given π β N and a partition πΌ = Γ
1β€πβ€ππ½(π) of πΌ. We representπΌ ΓπΌ matrices asπΓπ block matrices according to this partition. Given an πΌ Γ πΌ matrix π, we write ππ ,π for the (π , π)th block of π, and ππ
1:π2,π1:π2 for the sub-matrix of π defined by blocks ranging from π1 to π2 andπ1 toπ2. Unless specified otherwise, we write πΏ for the lower-triangular Cholesky factor of Ξand define
Ξ(π) B Ξ1:π ,1:π, π΄(π) B Ξ(π),β1, π΅(π) B π΄(π)
π , π for 1 β€ π β€ π. (4.9)
We interpret the{π½(π)}1β€πβ€πas labelling a hierarchy of scales withπ½(1)representing the coarsest andπ½(π) the finest. We write πΌ(π) forΓ
1β€π0β€ππ½(π
0).
Throughout this section, we assume that the ordering of the set πΌ of indices is compatible with the partition πΌ = Γ
π=1ππ½(π), i.e. π < π, π β π½(π) and π β π½(π) together implyπ βΊ π. We will write πΏ or chol(Ξ) for the Cholesky factor ofΞin that ordering.
4.2.3 Discretization ofπ»π
0(Ξ©)andπ»βπ (Ξ©)
While similar results are true for a wide range of measurements{ππ} β B =π»βπ (Ξ©) we will restrict our attention to two archetypical examples given by pointwise evaluation and nested averages.
We will assume (without loss of generality after rescaling) that diam(Ξ©) β€ 1. As described in Figure3.8, successive points of the maximin ordering can be gathered into levels so that after appropriate rescaling of the measurements, the Cholesky factorization in the maximin ordering falls in the setting of Example1.
Example 1. Let π > π/2. For β, πΏ β (0,1) let {π₯π}πβπΌ(1) β {π₯π}πβπΌ(2) β Β· Β· Β· β {π₯π}πβπΌ(π) be a nested hierarchy of points inΞ©that are homogeneously distributed at each scale in the sense of the following three inequalities:
(1) supπ₯βΞ©minπβπΌ(π) |π₯βπ₯π| β€ βπ, (2) minπβπΌ(π)infπ₯βπΞ©|π₯βπ₯π| β₯ πΏ βπ, and (3) minπ, πβπΌ(π):πβ π |π₯πβπ₯π| β₯ πΏ βπ.
Letπ½(1) B πΌ(1) andπ½(π) B πΌ(π) \ πΌ(πβ1) for π β {2, . . . , π}. LetπΉ denote the unit Dirac delta function and choose
ππ B β
π π
2 πΉ(π₯βπ₯π) forπ β π½(π) andπ β {1, . . . , π}. (4.10) The discretization chosen in Example 1 is not applicable for π < π/2 since in this case, functions in π»π
0(Ξ©) are not defined point-wise and thus πΉ β Bβ. A possible alternative is to replace ππ B β
π π
2 πΉ(π₯ βπ₯π) with ππ B ββ
π π
2 1π΅β(0)(π₯ β π₯π). However, while the exponential decay result in Theorem 4 seems to be true empirically for this choice of measurements, we are unable to prove it forπ < π/2.
Furthermore, the numerical homogenization result of Theorem 7 is false for this choice of measurements and π < π/2. However, our results can still be recovered by choosing measurements obtained as a hierarchy of local averages.
Given subsets ΛπΌ ,π½Λ β πΌ, we extend a matrix π β RπΌΛΓπ½Λ to an element of RπΌΓπ½ by padding it with zeros.
Example 2. (See Figure4.2.) Forβ, πΏβ (0,1), let(π(π)
π )πβπΌ(π) be uniformly Lipschitz convex sets forming a regular nested partition of Ξ© in the following sense. For π β {1, . . . , π}, Ξ© = Γ
πβπΌ(π)π(
π)
π is a disjoint union except for the boundaries.
πΌ(π) is a nested set of indices, i.e. πΌ(π) β πΌ(π+1) for π β {1, . . . , π β 1}. For π β {2, . . . , π} andπ β πΌ(πβ1), there exists a subset ππ β πΌ(π) such thatπ β ππ and π(
πβ1)
π =Γ
πβππ
π(
π)
π . Assume that eachπ(
π)
π contains a ball π΅
πΏ βπ(π₯(
π)
π )of centerπ₯(
π) π
and radiusπΏ βπ, and is contained in the ball π΅
βπ(π₯(π)
π ). Forπ β {2, . . . , π}andπ β
48
Figure 4.2: Hierarchical averaging. We illustrate the construction described in Example2in the caseπ =2. On the left we see the nested partition of the domain, and on the right we see (the signs of) a possible choice forπ1,π5, andπ6.
πΌ(πβ1), let the submatricesπ΄(π),π βR(ππ\{π})Γππ satisfyΓ
πβπππ΄(π),π
π, π π΄(π),π
π, π |π(π)
π | = πΏπ π andΓ
πβπππ΄(π),π
π , π |π(
π)
π | = 0for eachπ β ππ \ {π}, where |π(
π)
π |denotes the volume of π(
π)
π . Let π½(1) B πΌ(1) andπ½(π) B πΌ(π) \πΌ(πβ1) for π β {2, . . . , π}. Letπ(1) be the π½(1)Γ πΌ(1) matrix defined byπ(1)
π π B πΏπ π. Letπ(π) be theπ½(π) ΓπΌ(π) matrix defined byπ(π) BΓ
πβπΌ(πβ1) π΄(π),π for π >2, where we set ππ B ββπ π/2
Γ
πβπΌ(π)
π(π)
π, π 1π(π)
π
for eachπ β π½(π) (4.11) and define [ππ, π’] B β«
Ξ©πππ’dπ₯. In order to keep track of the distance between the differentππof Example2, we choose an arbitrary set of points{π₯π}πβπΌ β Ξ©with the property thatπ₯π βsupp(ππ) for eachπ β πΌ.
In the above, we have discretized the Greenβs functions of the elliptic operators resulting in the Greenβs matrixΞas the fundamental discrete object. The inverseπ΄of Ξcan be interpreted as the stiffness matrix obtained from the Galerkin discretization ofL using the basis given by
ππ B Γ
π
π΄π πG ππ
β B. (4.12)
These types of basis functions are referred to as gamblets in the prior works of [188β190] that form the basis for the proofs in this chapter. While exponentially
decaying, these basis functions are nonlocal and unknown apriori, hence they cannot be used to discretize a partial differential operator with unknown Greenβs function.
ForΞthe inverse of a Galerkin discretization of L in a local basis, analog results can be obtained by repeating the proofs of Theorems4and7in the discrete setting.
In the setting of Examples 1 and 2, denoting as πΏ the lower triangular Cholesky factor ofΞorΞβ1, we will show that
|πΏπ π| β€ poly(π)exp(βπΎ π(π, π)), (4.13) for a constantπΎ > 0 and a suitable distance measureπ( Β·, Β· ): πΌΓ πΌ βR.