Chapter IV: Proving exponential decay of Cholesky factors
4.5 Bounded condition numbers
In this section, we will bound the condition numbers of๐ต(๐) based on the following condition, which we will show to be satisfied for Examples1and2.
Condition 2. Let๐ป โ (0,1), ๐ถฮฆ โฅ 1be constants such that for1 โค ๐ < ๐ โค ๐, ๐min ฮ(๐)
โฅ 1 ๐ถฮฆ
๐ป2๐, (4.25)
๐max ฮ(๐)
๐ ,๐ โฮ(๐)
๐ ,1:๐ฮ(๐),โ1
1:๐ ,1:๐ฮ(๐)
1:๐ ,๐
โค ๐ถฮฆ๐ป2๐. (4.26) Theorem 5. Condition2implies that, for all1โค ๐ โค ๐,
๐ถโ1
ฮฆ ๐ปโ2(๐โ1)Idโบ ๐ต(๐) โบ ๐ถฮฆ๐ปโ2๐Id, (4.27)
and, for๐ B ๐ปโ2๐ถ2
ฮฆ,
cond ๐ต(๐)
โค ๐ . (4.28)
Proof. The lower bound in (4.27) follows from (4.26) and ๐ต(๐) = ฮ(๐)
๐ , ๐โฮ(๐)
๐ ,1:(๐โ1)ฮ(๐),โ1
1:๐ ,1:๐ฮ(๐)
1:(๐โ1), ๐
โ1
. (4.29)
The upper bound in (4.27) follows from (4.25) and ๐ต(๐) = ฮ(๐)โ1
๐ , ๐.
The following theorem shows that (4.26) is a Poincarรฉ inequality closely related to the accuracy of numerical homogenization basis functions [120, 163, 191] and (4.25) is an inverse Sobolev inequality related to the regularity of the discretization ofL:
Theorem 6. Condition2holds true if the constants๐ถฮฆ โฅ 1and๐ป โ (0,1)satisfy (1) ๐ถ1
ฮฆ
๐ป2๐ โค k๐kโ2
|๐ผ|2, for๐ผ โR๐ผ
(๐)
and๐=ร
๐โ๐ผ(๐) ๐ผ๐๐๐; and (2) min๐โspan(๐๐)
๐โ๐ผ(๐โ1)
k๐โ๐kโ2
|๐ผ|2 โค ๐ถฮฆ๐ป2(๐โ1), for ๐ผ โ R๐ฝ
(๐), ๐ < ๐ โค ๐, and ๐ =ร
๐โ๐ฝ(๐) ๐ผ๐๐๐.
Proof. Inequality (4.25) is a direct consequence of the first assumption of the the- orem, whereas (4.26) follows from the variational property [258, Theorem 5.1] of
the Schur complement:
๐ผ>
ฮ๐ ,๐โฮ(๐)
๐ ,1:๐ฮ(๐),โ1
1:๐ ,1:๐ฮ(๐)
1:๐ ,๐
๐ผ= inf
๐ฝโR๐ผ(๐)
(๐ผโ๐ฝ)>ฮ(๐)(๐ผโ ๐ฝ) (4.30)
= min
๐โspan{๐๐|๐โ๐ผ(๐)}
k๐โ๐kโ2 โค ๐ถฮฆ๐ป2๐|๐ผ|2. (4.31) We will now show that Examples1and2satisfy the conditions of Theorem6. For simplicity, for หฮฉ โ ฮฉ and ๐ โ ๐ปโ๐ (ฮฉ), we still write ๐ for the unique element
ห
๐ โ ๐ปโ๐ (ฮฉ)ห such that [๐, ๐ขห ] = [๐, ๐ข] for ๐ข โ ๐ป๐
0(ฮฉ)ห . The following Fenchel conjugate identity [40, Ex. 3.27, p. 93] will be useful throughout this section.
k๐k2
๐ปโ๐ (ฮฉ) = sup
๐ฃโ๐ป๐
0(ฮฉ)
2[๐, ๐ฃ] โ k๐ฃk2
๐ฃโ๐ป๐
0(ฮฉ). (4.32)
The first condition can be verified in a similar way as is done in [189].
Lemma 3. Letฮbe given as in Examples1and2. Then there exists a constant๐ถ depending only on๐ฟ,๐ , and๐, such that
1 ๐ถฮฆ
โ2๐ ๐ โค k๐kโ2
|๐ผ|2 , (4.33)
for๐ถฮฆ= kL k๐ถ,๐ผ โR๐ผ
(๐)
, and๐=ร
๐๐ผ๐๐๐.
Proof. The proof can be found in Section.1.
In order to verify the second condition in Theorem 6, we will construct a ๐ such that๐โ๐integrates to zero against polynomials of order at most๐ โ1 on domains of size โ๐. Then an application of the BrambleโHilbert lemma [65] will yield the desired factorโ๐ ๐ . To avoid scaling issues, we define, for 1 โค ๐ โค ๐and๐ โ ๐ผ(๐),
๐(
๐)
๐ B
๏ฃฑ๏ฃด
๏ฃด
๏ฃฒ
๏ฃด๏ฃด
๏ฃณ
๐น๐ฅ๐, in Example1, 1๐(๐)
๐
/|๐(๐)
๐ |, in Example2,
(4.34)
noting that span{๐(๐)
๐ | ๐ โ ๐ผ(๐)} = span{๐๐ | ๐ โ ๐ผ(๐)}. To obtain estimates independent of the regularity ofฮฉ, for the simplicity of the proof and without loss of generality, we will partially work in the extended space R๐ (rather than onฮฉ).
We write๐ฃfor the zero extension of๐ฃ โ ๐ป๐
0(ฮฉ)to๐ป๐ (R๐)and๐(
๐)
๐ for the extension of ๐(
๐)
๐ โ ๐ปโ๐ (ฮฉ) to an element of the dual space of๐ป๐
loc(R๐). We introduce new
54 measurement functions in the complement of ฮฉ as follows. For 1 โค ๐ โค ๐, we consider countably infinite index sets ห๐ผ(๐) โ ๐ผ(๐). We choose points (๐ฅ๐)๐โ๐ผห(๐)\๐ผ(๐)
satisfying sup
๐ฅโR๐\ฮฉ
min
๐โ๐ผห(๐)
dist(๐ฅ๐, ๐ฅ) โค๐ฟโ1โ๐, min
๐โ ๐โ๐ผห(๐)\๐ผ(๐)
dist(๐ฅ๐, ๐ฅ๐โช๐ฮฉ) โฅ๐ฟ โ๐. (4.35) We then define, for 1 โค ๐ โค ๐ and ๐ โ ๐ผห(๐), ๐(๐)
๐ B ๐ฟ๐ฅ
๐ for Example 1, and ๐(
๐)
๐ B
1๐ต ๐ฟ โ๐(๐ฅ๐)
|๐ต
๐ฟ โ๐(๐ฅ๐) | for Example2. LetP๐ โ1denote the linear space of polynomials of degree at most๐ โ1 (onR๐).
Lemma 4. Let ฮ be as in Example 1 or Example 2. Given ๐ โ (2,โ) and 1โค ๐ < ๐ โค ๐, let๐ค โR๐ฝ
(๐)ร๐ผห(๐)
be such that
โซ
๐ต
๐ โ๐(๐ฅ๐)
ยฉ
ยญ
ยซ
๐๐โ ร
๐โ๐ผห(๐)
๐ค๐ ๐๐(
๐) ๐
ยช
ยฎ
ยฌ
(๐ฅ)๐(๐ฅ)d๐ฅ =0, for all๐ โ P๐ โ1and๐ โ๐ฝ(๐) (4.36) and๐ค๐ ๐ โ 0 โ supp
๐(
๐) ๐
โ ๐ต
๐ โ๐(๐ฅ๐). Then, for๐ผ โ R๐ฝ
(๐)
, ๐ B ร
๐โ๐ฝ(๐) ๐ผ๐๐๐ and ๐ Bร
๐โ๐ฝ(๐), ๐โ๐ผ(๐) ๐ผ๐๐ค๐ ๐๐(๐)
๐ satisfy
k๐โ๐kโ2 โค kLโ1k๐ถ(๐ , ๐ )๐๐+2๐ ๐ฟ๐
1+โโ๐ ๐๐2
๐ , ๐
โ2๐ ๐|๐ผ|2, (4.37) with ๐๐ , ๐ B sup๐โ๐ฝ(๐)ร
๐โ๐ผห(๐) |๐ค๐ ๐| and k๐kโ B sup๐ขโ๐ป๐
0(ฮฉ)[๐, ๐ข]/[L๐ข, ๐ข]12 as in (4.7).
We proceed by proving Lemma 4 in the setting of Example 1. The proof in the setting of Example 2can be found in Section .1. For ๐ข โ ๐ป๐ (ฮฉ), write D0๐ข B ๐ข and for 1 โค ๐ โค ๐ , write D๐๐ขfor the vector of partial derivatives of๐ขof order๐, i.e.
D๐๐ข B
๐๐๐ข
๐๐ 1ยทยทยท๐๐
๐
๐1,...,๐๐=1,...,๐
. The proof of Lemma4will use the following version of the BrambleโHilbert lemma:
Lemma 5 ([65]). Let ฮฉ โ R๐ be convex and let ๐ be a sublinear functional on ๐ป๐ (ฮฉ)for ๐ โNsuch that
(1) there exists a constant๐ถหsuch that, for all๐ข โ ๐ป๐ (ฮฉ),
|๐(๐ข) | โค๐ถห
๐
ร
๐=0
diam(ฮฉ)๐kD๐๐ขk๐ฟ2(ฮฉ); (4.38)
(2) and๐(๐)=0for all ๐ โ P๐ โ1.
Then, for all๐ข โ๐ป๐ (ฮฉ),
|๐(๐ข) | โค๐ถ ๐ถห (๐ , ๐ )diam(ฮฉ)๐ kD๐ ๐ขk๐ฟ2(ฮฉ). (4.39) The following lemma is obtained from Lemma5:
Lemma 6. For 1 โค ๐ < ๐ โค ๐ and ๐ โ ๐ฝ(๐), let ๐๐, ๐ค๐ ๐ be as in Lemma 4 and Example2 and define ๐๐ B ร
๐โ๐ผ(๐)๐ค๐ ๐๐(๐)
๐ . Then there exists a constant๐ถ(๐ , ๐ ) such that, for all๐ฃ โ๐ป๐
0(ฮฉ),
โซ
๐ต ๐ โ๐(๐ฅ๐)
(๐๐โ๐๐) (๐ฅ)๐ฃ(๐ฅ)d๐ฅ
โค๐ถ(๐ , ๐ )๐๐ โ๐/2โ(๐ โ๐/2)๐ยฉ
ยญ
ยซ
โ๐ ๐/2+ ร
๐โ๐ผห(๐)
|๐ค๐ ๐|ยช
ยฎ
ยฌ kD๐ ๐ฃk
๐ฟ2
๐ต ๐ โ๐(๐ฅ๐).
(4.40)
Proof. We apply Lemma5to the linear functional๐ข โฆโโซ
๐ต
๐ โ๐
(๐๐โ๐๐)๐ข. Since the second requirement of Lemma 5is fulfilled by definition, it remains to bound ห๐ถ. We only execute the proof for Example1; the proof for Example2is analogous. We first note that while the sum in the definition of๐๐only ranges over ๐ โ ๐ผ(๐), we can increase it to run over all of ๐ โ ๐ผห(๐), since for ๐ โ ๐ผห(๐) \๐ผ(๐), the support of๐(
๐) ๐ is disjoint from that of๐ฃ โ๐ป๐
0(ฮฉ). Let๐ข โ ๐ป๐ (ฮฉ). Writing๐ถ(๐ , ๐ )for the continuity constant of the embedding of๐ป๐ (๐ต1(0))into๐ถ๐(๐ต1(0)), the inequalities
max
๐ต
๐ โ๐(๐ฅ๐)|๐ข( ยท ) |= max
๐ฅโ๐ต1(0)
๐ข
๐ โ๐(๐ฅโ๐ฅ๐)
โค๐ถ(๐ , ๐ )
๐
ร
๐=0
(๐ โ๐)๐
[D๐๐ข] ๐ โ๐( ยท โ๐ฅ๐) ๐ฟ2(๐ต1(0))
and
[D๐๐ข] ๐ โ๐( ยท โ๐ฅ๐)
๐ฟ2(๐ต1(0)) = (๐ โ๐)โ๐/2kD๐๐ขk๐ฟ2(๐ต
๐ โ๐(๐ฅ๐))
imply that
|๐๐(๐ข) โ๐๐(๐ข) | โคยฉ
ยญ
ยซ
โ๐ ๐/2+ ร
๐โ๐ผห(๐)
|๐ค๐ ๐|ยช
ยฎ
ยฌ max
๐ฅโ๐ต ๐ โ๐(๐ฅ๐)
|๐ข(๐ฅ) | (4.41)
โค๐ถ(๐ , ๐ )๐โ๐/2โโ๐ ๐/2ยฉ
ยญ
ยซ
โ๐ ๐/2+ ร
๐โ๐ผห(๐)
|๐ค๐ ๐|ยช
ยฎ
ยฌ
๐
ร
๐=0
(๐ โ๐)๐kD๐๐ขk๐ฟ2(๐ต
๐ โ๐(๐ฅ๐)).
(4.42)
Therefore the first condition of Lemma5holds with
ห
๐ถ =๐ถ(๐ , ๐ )๐โ๐/2โโ๐ ๐/2ยฉ
ยญ
ยซ
โ๐ ๐/2+ ร
๐โ๐ผห(๐)
|๐ค๐ ๐|ยช
ยฎ
ยฌ
, (4.43)
and we conclude the proof by writing๐ถ(๐ , ๐ )for any constant depending only on๐
and๐ .
56 We can now conclude the proof of Lemma4.
Proof of Lemma4. Write ๐ B ร
๐โ๐ฝ(๐)๐ผ๐๐๐ and ๐๐ B ร
๐โ๐ผ(๐) ๐ค๐ ๐๐(
๐)
๐ . Equa- tion (4.32) implies that
k๐โ๐k2
๐ปโ๐ (ฮฉ) = sup
๐ฃโ๐ป๐
0(ฮฉ)
ร
๐โ๐ฝ(๐)
2๐ผ๐
โซ
๐ต
๐ โ๐(๐ฅ๐)
(๐๐โ๐๐) (๐ฅ)๐ฃ(๐ฅ)d๐ฅ
!
โ k๐ฃk2
๐ป๐
0(ฮฉ). (4.44) The packing inequality ร
๐โ๐ฝ(๐) kD๐ ๐ฃk2
๐ฟ2 ๐ต
๐ โ๐(๐ฅ๐) โค ๐ถ(๐) โ๐โ๐๐/๐ฟ๐ k๐ฃk2
๐ป๐
0(ฮฉ) to- gether with Lemma6yields
k๐โ๐k2๐ปโ๐ (ฮฉ)โค sup
๐ฃโ๐ป0๐ (ฮฉ)
ร
๐โ๐ฝ(๐)
"
2|๐ผ๐|๐ถ(๐ , ๐ )๐๐ โ
๐ 2โ(๐ โ
๐ 2)๐ยฉ
ยญ
ยซ โ
๐ ๐
2 + ร
๐โ๐ผ(๐)
|๐ค๐ ๐|ยช
ยฎ
ยฌ
kD๐ ๐ฃk๐ฟ2(๐ต ๐ โ๐(๐ฅ๐))
(4.45)
โ (๐ถ(๐))โ1 โ๐โ๐๐/๐ฟ
โ๐
kD๐ ๐ฃk2
๐ฟ2
๐ต ๐ โ๐(๐ฅ๐)
#
. (4.46)
Applying the inequality 2๐๐ฅโ๐๐ฅ2 โค ๐2/๐to each summand yields
k๐โ๐k๐ป2โ๐ (ฮฉ) โค๐ถ(๐) โ๐โ๐๐/๐ฟ
๐ ร
๐โ๐ฝ(๐)
ยฉ
ยญ
ยซ
๐ผ๐๐ถ(๐ , ๐ )๐๐ โ
๐ 2โ(๐ โ
๐ 2)๐ยฉ
ยญ
ยซ โ
๐ ๐
2 + ร
๐โ๐ฝ(๐)
|๐ค๐ ๐|ยช
ยฎ
ยฌ ยช
ยฎ
ยฌ
2
(4.47)
โค๐ถ(๐ , ๐ )๐2๐ ๐ฟ๐
1+โโ๐ ๐๐2
๐ , ๐
โ2๐ ๐|๐ผ|2. (4.48)
Since, for all ๐ โ๐ปโ๐ (ฮฉ),
k๐k2โ =[๐ ,Lโ1๐] โค k๐k๐ปโ๐ (ฮฉ)kLโ1๐k๐ป๐
0(ฮฉ) โค kLโ1k k๐k2
๐ปโ๐ (ฮฉ), (4.49) we havek๐โ๐kโ โค p
kLโ1k k๐โ๐k๐ปโ๐ (ฮฉ), and this completes the proof.
The following geometric lemma shows that the assumption (4.36) of Lemma4can be satisfied with a uniform bound on the value of ๐and the norm of weights๐ค๐, ๐. Lemma 7. There exist constants ๐(๐ , ๐ ) and๐ถ(๐ , ๐ , ๐ฟ) such that for all 1 โค ๐ <
๐ โค ๐, there exist weights ๐ค โ R๐ฝ
(๐)ร๐ผห(๐)
satisfying(4.36)and (with๐๐ , ๐ defined as in Lemma4)
๐2
๐ , ๐ โค โ๐ ๐๐ถ(๐ , ๐ , ๐ฟ). (4.50)
Proof. For Example1, (4.36) is equivalent to โ๐ ๐/2๐(๐ฅ๐) = ร
๐โ๐ผห๐(๐)
๐ค๐ ๐๐(๐ฅ๐),โ๐ โ P๐ โ1, (4.51)
where ห๐ผ(๐)
๐ B {๐ โ ๐ผห(๐) | ๐ฅ๐ โ ๐ต(๐ฅ๐, ๐ โ๐)}.
Fix๐ โ ๐ฝ(๐), let๐ >0, and write๐ฅ๐
๐ B
๐ฅ๐โ๐ฅ๐
๐ . Write0 B (0, . . . ,0) โR๐. Since the function ๐( ยท ) โฆโ ๐(ยทโ๐ฅ๐
๐ )is surjective onP๐ โ1, (4.51) is satisfied if โ๐ ๐/2๐(0) = ร
๐โ๐ผห๐(๐)
๐ค๐ ๐๐(๐ฅ๐
๐),โ๐ โ P๐ โ1. (4.52) For a multiindex ๐ = (๐1, . . . , ๐๐) โ N๐ and a point ๐ง = (๐ง1, . . . , ๐ง๐) โ R๐, write ๐ง๐ B ร๐
๐=1๐ง๐๐
๐ . Use the convention0๐ =0 if๐ โ 0and00 =1. To satisfy (4.52), it is sufficient to identify a subset ๐ of ห๐ผ(
๐)
๐ and ๐ค๐,ยท โ R๐ผห
(๐)
such that #๐ = ๐ ๐, ๐ค๐, ๐ =0 for ๐ โ๐, and
โ๐ ๐/20๐ =ร
๐โ๐
๐ค๐ ๐(๐ฅ๐
๐)๐,โ๐ โ {0, . . . , ๐ โ1}๐. (4.53) LetV๐ โR{0,1,...,๐ โ1}
๐ร๐
be the ๐ ๐ร๐ ๐matrix defined by V๐๐, ๐ B
๐ฅ๐
๐
๐
. (4.54)
For a multiindex ๐ โ N๐ and a point ๐ฅ โ R๐, ๐ฅ๐ B ร๐
๐=1๐ฅ๐๐. Let w โ R๐ be defined byw๐ B ๐ค๐, ๐ for ๐ โ๐. Equation (4.53) is then equivalent to
โ๐ ๐/2e=V๐w, (4.55)
wheree โ R{0,1,...,๐ โ1}
๐
is defined by e๐ B 0๐ for ๐ โ {0,1, . . . , ๐ โ1}๐. We will now identifyw by inverting (4.55). To achieve this while keeping the norm of w under control, we will seek to identify the subset ๐ and๐ > 0 such that๐min(V๐) (the minimal singular value ofV๐) is bounded from below by a constant depending only on๐ and๐.
For ๐ผ โฅ 0, let (๐๐)๐โ{0,1,...,๐ โ1}๐ be elements of R๐ satisfying |๐๐| โค ๐ผ for all ๐ โ {0,1, . . . , ๐ โ1}๐. Let1B (1, . . . ,1) โR๐and, for ๐ โ {0,1, . . . , ๐ โ1}๐, let ๐ง๐ B 1+ ๐ +๐๐. Observe that for ๐ผ = 0, the points ๐ง๐ are on a regular grid. Let Vยฏ๐ผ โ R{0,1,...,๐ โ1}
๐ร{0,1,...,๐ โ1}๐ be the ๐ ๐ร๐ ๐ matrix defined by ยฏV๐ผ๐, ๐ B ๐ง๐๐ . Let ๐ be the ๐ ร ๐ Vandermonde matrix defined by๐๐, ๐ = ๐๐. Writing๐min(๐) for the minimal singular value of๐, we have for๐ผ=0, by [119, Theorem 4.2.12],
๐min Vยฏ0
=(๐min(๐))๐. (4.56)
Since univariate polynomial interpolation on ๐ points with polynomials of degree ๐ โ1 is uniquely solvable, we have ๐min(๐) > 0 and ๐min(Vยฏ0) > ๐ถ(๐ , ๐ ) > 0.
58 Therefore, the continuity of the minimal singular value with respect to the entries of ยฏV๐ผimplies that there exists๐ผโ, ๐โ > 0 depending only on๐ , ๐such that ๐ผ โค ๐ผโ implies ๐min(Vยฏ๐ผ) > ๐โ. Since (by construction) the (๐ฅ๐)๐โ๐ผห(๐) form a covering of R๐ of radius โ๐, the (๐ฅ๐
๐)๐โ๐ผห(๐) form a covering of R๐ of radius โ๐/๐ and for each ๐ โ {0,1, . . . , ๐ โ1}๐, there exists an๐ฅ๐
๐๐that is at distance at mostโ๐/๐from๐. Let ๐ B {๐๐ | ๐โ {0,1, . . . , ๐ โ1}๐} โ ๐ผห(๐) be the collection of corresponding labels.
It follows from |๐ฅ๐
๐๐| โค โ
๐ ๐ +โ๐/๐that |๐ฅ๐
๐โ๐ฅ๐| โค ๐
โ
๐ ๐ + โ๐, and ๐ โ ๐ผห(๐)
๐ for
๐ > 1+๐
โ
๐ ๐ /โ๐. Selecting๐ = โ๐/๐ผโ implies that๐min(V๐) > ๐โ and๐ โ ๐ผห๐(๐) for ๐ >1+โ
๐ ๐ /๐ผโ. Defining
๐ค๐ ๐ B
๏ฃฑ๏ฃด
๏ฃด
๏ฃฒ
๏ฃด๏ฃด
๏ฃณ
(V๐)โ1โ๐ ๐/2e
๐
, if ๐ = ๐๐โ ๐,
0, otherwise,
(4.57)
the weights๐ค๐ ๐ satisfy๐๐ ๐ โค๐ถ(๐ , ๐)โ๐ ๐/2and (4.36) with a๐ depending only on๐ and๐. This concludes the proof for Example1. The proof is similar for Example2 with minor changes (the bound on๐ also depends on๐ฟ).
The following lemma concerns the satisfaction of the second condition of Theorem6:
Lemma 8. In the setting of Examples1and2, there exists some constant๐ถ(๐ , ๐ , ๐ฟ) >
0such that, for2โค ๐ < ๐ โค ๐,๐ผ โR๐ฝ
(๐) and๐ =ร
๐๐ผ๐๐๐, min
๐โspan(๐๐)
๐โ๐ผ(๐โ1)
k๐โ๐kโ2
|๐ผ|2 โค๐ถ(๐ , ๐ , ๐ฟ) kLโ1kโ2๐ (๐โ1). (4.58) Proof. Apply Lemma4with the bounds on ๐and๐ obtained in Lemma7.
The following theorem is a direct consequence of Theorems 6, Lemma 3 and Lemma8.
Theorem 7. In the setting of Examples1and2, there exists a constant๐ถ(๐ , ๐ , ๐ฟ)such that Condition2is fulfilled with๐ถฮฆ B max( kL k,kLโ1k)๐ถ(๐ , ๐ , ๐ฟ) and๐ป B โ๐ . Propagation of exponential decay
We will now derive the exponential decay of the Cholesky factors๐ฟ by combining the algebraic identities of Lemma 1with the bounds on the condition numbers of the ๐ต(๐) (implied by Condition2) and the exponential decay of the ๐ด(๐) (specified in Condition1). The core of our proof is based on a combination/extension of the
results of [31,32,34,68,128,144] on decay algebras. The pseudodistance๐( ยท,ยท ) appearing in (4.13) is not a pseudometric because it does not satisfy the triangle inequality. However, to prove (4.13), we will only need the following weaker version of the triangle inequality:
Definition 8. A function๐: ๐ผร๐ผ โโR+is called ahierarchical pseudometricif (1) ๐(๐, ๐) =0, for all๐ โ๐ผ;
(2) ๐(๐, ๐) =๐(๐ , ๐), for all๐, ๐ โ ๐ผ;
(3) for all1 โค ๐ โค ๐, ๐( ยท, ยท )restricted to๐ฝ(๐) ร๐ฝ(๐) is a pseudometric;
(4) for all 1 โค ๐ โค ๐ โค ๐ โค ๐ and ๐ โ ๐ฝ(๐), ๐ โ ๐ฝ(๐), ๐ โ ๐ฝ(๐), we have ๐(๐, ๐) โค ๐(๐, ๐ ) +๐(๐ , ๐).
Note that the ๐( ยท,ยท ) specified in (4.20) for Examples 1 and 2 is a hierarchical pseudometric. For a hierarchical pseudometric๐( ยท,ยท )and๐พ โR+, let
๐๐(๐พ) B sup
1โค๐โค๐โค๐
sup
๐โ๐ฝ(๐)
ร
๐โ๐ฝ(๐)
exp(โ๐พ ๐(๐, ๐)). (4.59)
The following theorem states the main result of this section:
Theorem 8(Exponential decay of the Cholesky factors). Assume thatฮfulfills Con- ditions1and2 with the constants๐พ , ๐ถ๐พ, ๐ป , ๐ถฮฆ and the hierarchical pseudometric ๐( ยท, ยท ). Then
(chol(ฮ))๐ ๐
โค 2๐ถฮฆ๐๐(๐พห/8)2
(1โ๐)2 4๐๐(๐พห/4)๐ถฮฆ๐ถ๐พ(๐๐(๐พห/2))2 (1โ๐)2
!๐ exp
โ๐พห 8๐(๐, ๐)
, (4.60) where ๐ถ๐ B max
n 1, 2
๐ถ๐พ๐ถฮฆ 1+๐
o
, ๐ B 1โ๐
โ1
1+๐ โ1, ๐พห B โlog(
๐)
1+log(๐๐(๐พ/2))+log(๐ถ๐ )โlog(๐) ๐พ 2, and ๐ =๐ปโ2๐ถ2
ฮฆ is defined as in Theorem5.
The remaining part of this section will present the proof of Theorem8. We will use the following lemma on the stability of exponential decay under matrix multiplica- tion, the proof of which is a minor modification of that of [128].
Lemma 9. Let ๐ผ be an index set that is partitioned as ๐ผ = ๐ฝ(1) โช ยท ยท ยท๐ฝ(๐) and let ๐: ๐ผร๐ผ โRโฅ0satisfy
๐(๐1, ๐๐+1) โค
๐
ร
๐=1
๐(๐๐, ๐๐+1) for all1 โค ๐โค ๐โ1and๐๐ โ ๐ฝ(๐).
60 Let ๐(๐) โ R๐ฝ
(๐)ร๐ฝ(๐+1)
be such that |๐(๐)
๐, ๐ | โค ๐ถexp(โ๐พ ๐(๐, ๐)) for1 โค ๐ โค ๐โ1 and let
๐๐(๐พ/2) B sup
1โค๐โค๐โ1
sup
๐โ๐ฝ(๐+1)
ร
๐โ๐ฝ(๐)
exp
โ๐พ
2๐(๐, ๐)
for๐พ โR+. (4.61) Then, for1โค ๐ โค ๐โ1,
๐
ร
๐=1
๐(๐)
!
๐, ๐
โค (๐๐(๐พ/2)๐ถ)๐exp
โ๐พ
2๐(๐, ๐) .
Proof. Set๐1 B๐,๐๐+1 B ๐. Then
๐
ร
๐=1
๐(๐)
!
๐, ๐
โค๐ถ๐
ร
๐2,...,๐๐โ๐ฝ(2),...,๐ฝ(๐)
exp โ๐พ
๐
ร
๐=1
๐(๐๐, ๐๐+1)
!
โค๐ถ๐exp
โ๐พ
2๐(๐1, ๐๐+1) ร
๐2,...๐๐โ๐ผ
exp โ๐พ 2
๐
ร
๐=1
๐(๐๐, ๐๐+1)
!
โค (๐๐(๐พ/2)๐ถ)๐exp
โ๐พ
2๐(๐, ๐) .
The proof of the following lemma (on the stability of exponential decay under matrix inversion for well-conditioned matrices) is nearly identical to that of [128] (we only keep track of constants; see also [68] for a related result on the inverse of sparse matrices).
Lemma 10. Let ๐ดโR๐ผร๐ผ be symmetric and positive definite such that for๐ถ , ๐พ >0 and a metric๐( ยท,ยท )on ๐ผwe have |๐ด๐, ๐| โค๐ถexp(โ๐พ ๐(๐, ๐)). It holds true that
(๐ดโ1)๐ , ๐
โค 4
k๐ดk + k๐ดโ1kโ1
(1โ๐)2exp โ log(๐1)
(1+log(๐๐(๐พ/2)) +log(๐ถ๐ )) +log(1๐) ๐พ 2๐(๐, ๐)
! (4.62)
where ๐๐(๐พ/2) B sup๐โ๐ผ
ร
๐โ๐ผexp โ๐พ2๐(๐, ๐)
, ๐ถ๐ B max n
1, 2๐ถ
k๐ดk+k๐ดโ1kโ1
o
= maxn
1,2
๐ถk๐ดโ1k 1+๐
o , ๐ B
1โ 1
k๐ดk k๐ดโ1k
1+ 1
k๐ดk k๐ดโ1k
= 1โ๐ โ1
1+๐ โ1, and ๐ B k๐ดk k๐ดโ1k is the condition number of ๐ด.
Proof. On a compact set not containing 0, the function๐ฅ โฆโ ๐ฅโ1can be accurately approximated by low-order polynomials in๐ฅ. Then, the spread of the exponential decay can be controlled by Lemma9. See Section.1for details.
By representing Schur complements as matrix inverses, Lemma 10 can also be used to show that the Cholesky factors of well-conditioned exponentially-decaying matrices are exponentially decaying. The following lemma appears in a similar form in [34] for banded matrices and in [144] without explicit constants.
Lemma 11. Let๐ต โR๐ผร๐ผ 'R๐ร๐be symmetric and positive definite with condition number ๐ and such that
๐ต๐, ๐
โค ๐ถexp(โ๐พ ๐(๐, ๐)) for some constant ๐ถ > 0 and some metric ๐ on ๐ผ. Let ๐ฟ be the Cholesky factor (in an arbitrary order) of ๐ตโ1 (๐ตโ1 =๐ฟ ๐ฟ๐). Then
๐ฟ๐ , ๐
โค 4p k๐ตk k๐ตk + k๐ตโ1kโ1
(1โ๐)2 exp
log(๐)
1+log(๐๐(๐พ/2)) +log(๐ถ๐ ) โlog(๐) ๐พ 2๐(๐, ๐)
(4.63) where ๐๐(๐พ/2) B sup๐โ๐ผ
ร
๐โ๐ผexp โ๐พ
2๐(๐, ๐)
, ๐ถ๐ B max n
1, 2๐ถk๐ต
โ1k 1+๐
o
, and๐ B
1โ๐ โ1 1+๐ โ1.
Proof. Lemma2 implies that the Schur complements of ๐ตโ1 can be expressed as inverses of sub-matrices of๐ต. The result then follows from Lemma10(see Proof.1
for details).
The last ingredient needed to prove the exponential decay of the Cholesky factors of ฮis the following lemma showing the stability of exponential decay under inversion for block-lower-triangular matrices (this operation appears in the definition of ยฏ๐ฟ in (4.14)):
Lemma 12. Let๐ผbe an index set that is partitioned as๐ผ =๐ฝ(1)โช ยท ยท ยท๐ฝ(๐)and assume that the matrix๐ฟ โR๐ผร๐ผ is block-lower triangular with respect to this partition, with identity matrices as diagonal blocks. If๐( ยท,ยท )is a hierarchical pseudometric such that|๐ฟ๐ ๐| โค๐ถexp(โ๐พ ๐(๐, ๐))(for some๐ถ โฅ 1and๐พ > 0), then it holds true that
(๐ฟโ1)๐ ๐
โค 2๐(๐๐(๐พ/2)๐ถ)๐exp
โ๐พ
2๐(๐, ๐)
(4.64) with๐๐(๐พ) Bsup1โค๐โค๐โค๐sup๐โ๐ฝ(๐) ร
๐โ๐ฝ(๐)exp(โ๐พ ๐(๐, ๐)).
Proof. The Neumann series of a๐ร๐ block-lower-triangular matrix with identity matrices on the (block) diagonal can be written as
๐ฟโ1=
๐
ร
๐=0
(Idโ๐ฟ)๐ . (4.65)
62 Since the sum terminates in๐steps, the thickening of the exponential decay can be
bounded using Lemma9. See Proof.1for details.
By applying the above results to the decomposition obtained in Lemma 1, we conclude the proof of Theorem8. See Proof.1for details.