• Tidak ada hasil yang ditemukan

Bounded condition numbers

Dalam dokumen Inference, Computation, and Games (Halaman 75-85)

Chapter IV: Proving exponential decay of Cholesky factors

4.5 Bounded condition numbers

In this section, we will bound the condition numbers of๐ต(๐‘˜) based on the following condition, which we will show to be satisfied for Examples1and2.

Condition 2. Let๐ป โˆˆ (0,1), ๐ถฮฆ โ‰ฅ 1be constants such that for1 โ‰ค ๐‘˜ < ๐‘™ โ‰ค ๐‘ž, ๐œ†min ฮ˜(๐‘˜)

โ‰ฅ 1 ๐ถฮฆ

๐ป2๐‘˜, (4.25)

๐œ†max ฮ˜(๐‘ž)

๐‘™ ,๐‘™ โˆ’ฮ˜(๐‘ž)

๐‘™ ,1:๐‘˜ฮ˜(๐‘ž),โˆ’1

1:๐‘˜ ,1:๐‘˜ฮ˜(๐‘ž)

1:๐‘˜ ,๐‘™

โ‰ค ๐ถฮฆ๐ป2๐‘˜. (4.26) Theorem 5. Condition2implies that, for all1โ‰ค ๐‘˜ โ‰ค ๐‘ž,

๐ถโˆ’1

ฮฆ ๐ปโˆ’2(๐‘˜โˆ’1)Idโ‰บ ๐ต(๐‘˜) โ‰บ ๐ถฮฆ๐ปโˆ’2๐‘˜Id, (4.27)

and, for๐œ… B ๐ปโˆ’2๐ถ2

ฮฆ,

cond ๐ต(๐‘˜)

โ‰ค ๐œ… . (4.28)

Proof. The lower bound in (4.27) follows from (4.26) and ๐ต(๐‘˜) = ฮ˜(๐‘ž)

๐‘˜ , ๐‘˜โˆ’ฮ˜(๐‘ž)

๐‘˜ ,1:(๐‘˜โˆ’1)ฮ˜(๐‘ž),โˆ’1

1:๐‘˜ ,1:๐‘˜ฮ˜(๐‘ž)

1:(๐‘˜โˆ’1), ๐‘˜

โˆ’1

. (4.29)

The upper bound in (4.27) follows from (4.25) and ๐ต(๐‘˜) = ฮ˜(๐‘˜)โˆ’1

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 constants๐ถฮฆ โ‰ฅ 1and๐ป โˆˆ (0,1)satisfy (1) ๐ถ1

ฮฆ

๐ป2๐‘˜ โ‰ค k๐œ™kโˆ—2

|๐›ผ|2, for๐›ผ โˆˆR๐ผ

(๐‘˜)

and๐œ™=ร

๐‘–โˆˆ๐ผ(๐‘˜) ๐›ผ๐‘–๐œ™๐‘–; and (2) min๐œ‘โˆˆspan(๐œ™๐‘–)

๐‘–โˆˆ๐ผ(๐‘˜โˆ’1)

k๐œ™โˆ’๐œ‘kโˆ—2

|๐›ผ|2 โ‰ค ๐ถฮฆ๐ป2(๐‘˜โˆ’1), for ๐›ผ โˆˆ R๐ฝ

(๐‘™), ๐‘˜ < ๐‘™ โ‰ค ๐‘ž, and ๐œ™ =ร

๐‘–โˆˆ๐ฝ(๐‘™) ๐›ผ๐‘–๐œ™๐‘–.

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:

๐›ผ>

ฮ˜๐‘™ ,๐‘™โˆ’ฮ˜(๐‘ž)

๐‘™ ,1:๐‘˜ฮ˜(๐‘ž),โˆ’1

1:๐‘˜ ,1:๐‘˜ฮ˜(๐‘ž)

1:๐‘˜ ,๐‘™

๐›ผ= inf

๐›ฝโˆˆR๐ผ(๐‘˜)

(๐›ผโˆ’๐›ฝ)>ฮ˜(๐‘ž)(๐›ผโˆ’ ๐›ฝ) (4.30)

= min

๐œ‘โˆˆspan{๐œ™๐‘–|๐‘–โˆˆ๐ผ(๐‘˜)}

k๐œ™โˆ’๐œ‘kโˆ—2 โ‰ค ๐ถฮฆ๐ป2๐‘˜|๐›ผ|2. (4.31) We will now show that Examples1and2satisfy the conditions of Theorem6. For simplicity, for หœฮฉ โŠ‚ ฮฉ and ๐œ™ โˆˆ ๐ปโˆ’๐‘ (ฮฉ), we still write ๐œ™ for the unique element

หœ

๐œ™ โˆˆ ๐ปโˆ’๐‘ (ฮฉ)หœ such that [๐œ™, ๐‘ขหœ ] = [๐œ™, ๐‘ข] for ๐‘ข โˆˆ ๐ป๐‘ 

0(ฮฉ)หœ . The following Fenchel conjugate identity [40, Ex. 3.27, p. 93] will be useful throughout this section.

k๐œ™k2

๐ปโˆ’๐‘ (ฮฉ) = sup

๐‘ฃโˆˆ๐ป๐‘ 

0(ฮฉ)

2[๐œ™, ๐‘ฃ] โˆ’ k๐‘ฃk2

๐‘ฃโˆˆ๐ป๐‘ 

0(ฮฉ). (4.32)

The first condition can be verified in a similar way as is done in [189].

Lemma 3. Letฮ˜be given as in Examples1and2. Then there exists a constant๐ถ depending only on๐›ฟ,๐‘ , and๐‘‘, such that

1 ๐ถฮฆ

โ„Ž2๐‘  ๐‘˜ โ‰ค k๐œ™kโˆ—2

|๐›ผ|2 , (4.33)

for๐ถฮฆ= kL k๐ถ,๐›ผ โˆˆR๐ผ

(๐‘˜)

, and๐œ™=ร

๐‘–๐›ผ๐‘–๐œ™๐‘–.

Proof. The proof can be found in Section.1.

In order to verify the second condition in Theorem 6, we will construct a ๐œ‘ such that๐œ™โˆ’๐œ‘integrates to zero against polynomials of order at most๐‘ โˆ’1 on domains of size โ„Ž๐‘˜. Then an application of the Brambleโ€“Hilbert lemma [65] will yield the desired factorโ„Ž๐‘˜ ๐‘ . To avoid scaling issues, we define, for 1 โ‰ค ๐‘˜ โ‰ค ๐‘žand๐‘– โˆˆ ๐ผ(๐‘˜),

๐œ™(

๐‘˜)

๐‘– B

๏ฃฑ๏ฃด

๏ฃด

๏ฃฒ

๏ฃด๏ฃด

๏ฃณ

๐œน๐‘ฅ๐‘–, in Example1, 1๐œ(๐‘˜)

๐‘–

/|๐œ(๐‘˜)

๐‘– |, in Example2,

(4.34)

noting that span{๐œ™(๐‘˜)

๐‘– | ๐‘– โˆˆ ๐ผ(๐‘˜)} = span{๐œ™๐‘– | ๐‘– โˆˆ ๐ผ(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 write๐‘ฃfor the zero extension of๐‘ฃ โˆˆ ๐ป๐‘ 

0(ฮฉ)to๐ป๐‘ (R๐‘‘)and๐œ™(

๐‘˜)

๐‘– for the extension of ๐œ™(

๐‘˜)

๐‘– โˆˆ ๐ปโˆ’๐‘ (ฮฉ) to an element of the dual space of๐ป๐‘ 

loc(R๐‘‘). We introduce new

54 measurement functions in the complement of ฮฉ as follows. For 1 โ‰ค ๐‘˜ โ‰ค ๐‘ž, we consider countably infinite index sets หœ๐ผ(๐‘˜) โŠƒ ๐ผ(๐‘˜). We choose points (๐‘ฅ๐‘–)๐‘–โˆˆ๐ผหœ(๐‘ž)\๐ผ(๐‘ž)

satisfying sup

๐‘ฅโˆˆR๐‘‘\ฮฉ

min

๐‘–โˆˆ๐ผหœ(๐‘˜)

dist(๐‘ฅ๐‘–, ๐‘ฅ) โ‰ค๐›ฟโˆ’1โ„Ž๐‘˜, min

๐‘–โ‰ ๐‘—โˆˆ๐ผหœ(๐‘˜)\๐ผ(๐‘˜)

dist(๐‘ฅ๐‘–, ๐‘ฅ๐‘—โˆช๐œ•ฮฉ) โ‰ฅ๐›ฟ โ„Ž๐‘˜. (4.35) We then define, for 1 โ‰ค ๐‘˜ โ‰ค ๐‘ž and ๐‘– โˆˆ ๐ผหœ(๐‘˜), ๐œ™(๐‘˜)

๐‘– B ๐›ฟ๐‘ฅ

๐‘– for Example 1, and ๐œ™(

๐‘˜)

๐‘– B

1๐ต ๐›ฟ โ„Ž๐‘˜(๐‘ฅ๐‘–)

|๐ต

๐›ฟ โ„Ž๐‘˜(๐‘ฅ๐‘–) | for Example2. LetP๐‘ โˆ’1denote the linear space of polynomials of degree at most๐‘ โˆ’1 (onR๐‘‘).

Lemma 4. Let ฮ˜ be as in Example 1 or Example 2. Given ๐œŒ โˆˆ (2,โˆž) and 1โ‰ค ๐‘˜ < ๐‘™ โ‰ค ๐‘ž, let๐‘ค โˆˆR๐ฝ

(๐‘™)ร—๐ผหœ(๐‘˜)

be such that

โˆซ

๐ต

๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–)

ยฉ

ยญ

ยซ

๐œ™๐‘–โˆ’ ร•

๐‘—โˆˆ๐ผหœ(๐‘˜)

๐‘ค๐‘– ๐‘—๐œ™(

๐‘˜) ๐‘—

ยช

ยฎ

ยฌ

(๐‘ฅ)๐‘(๐‘ฅ)d๐‘ฅ =0, for all๐‘ โˆˆ P๐‘ โˆ’1and๐‘– โˆˆ๐ฝ(๐‘™) (4.36) and๐‘ค๐‘– ๐‘— โ‰  0 โ‡’ supp

๐œ™(

๐‘˜) ๐‘—

โŠ‚ ๐ต

๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–). Then, for๐›ผ โˆˆ R๐ฝ

(๐‘™)

, ๐œ™ B ร

๐‘–โˆˆ๐ฝ(๐‘™) ๐›ผ๐‘–๐œ™๐‘– and ๐œ‘ Bร

๐‘–โˆˆ๐ฝ(๐‘™), ๐‘—โˆˆ๐ผ(๐‘˜) ๐›ผ๐‘–๐‘ค๐‘– ๐‘—๐œ™(๐‘˜)

๐‘— satisfy

k๐œ™โˆ’๐œ‘kโˆ—2 โ‰ค kLโˆ’1k๐ถ(๐‘‘ , ๐‘ )๐œŒ๐‘‘+2๐‘  ๐›ฟ๐‘‘

1+โ„Žโˆ’๐‘™ ๐‘‘๐œ”2

๐‘™ , ๐‘˜

โ„Ž2๐‘  ๐‘˜|๐›ผ|2, (4.37) with ๐œ”๐‘™ , ๐‘˜ B sup๐‘–โˆˆ๐ฝ(๐‘™)ร

๐‘—โˆˆ๐ผหœ(๐‘˜) |๐‘ค๐‘– ๐‘—| and k๐œ™kโˆ— B sup๐‘ขโˆˆ๐ป๐‘ 

0(ฮฉ)[๐œ™, ๐‘ข]/[L๐‘ข, ๐‘ข]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 ๐‘ข โˆˆ ๐ป๐‘ (ฮฉ), write D0๐‘ข B ๐‘ข and for 1 โ‰ค ๐‘˜ โ‰ค ๐‘ , write D๐‘˜๐‘ขfor the vector of partial derivatives of๐‘ขof order๐‘˜, i.e.

D๐‘˜๐‘ข B

๐œ•๐‘˜๐‘ข

๐œ•๐‘– 1ยทยทยท๐œ•๐‘–

๐‘˜

๐‘–1,...,๐‘–๐‘˜=1,...,๐‘‘

. The proof of Lemma4will use the following version of the Brambleโ€“Hilbert lemma:

Lemma 5 ([65]). Let ฮฉ โŠ‚ R๐‘‘ be convex and let ๐œ™ be a sublinear functional on ๐ป๐‘ (ฮฉ)for ๐‘ โˆˆNsuch that

(1) there exists a constant๐ถหœsuch that, for all๐‘ข โˆˆ ๐ป๐‘ (ฮฉ),

|๐œ™(๐‘ข) | โ‰ค๐ถหœ

๐‘ 

ร•

๐‘˜=0

diam(ฮฉ)๐‘˜kD๐‘˜๐‘ขk๐ฟ2(ฮฉ); (4.38)

(2) and๐œ™(๐‘)=0for all ๐‘ โˆˆ P๐‘ โˆ’1.

Then, for all๐‘ข โˆˆ๐ป๐‘ (ฮฉ),

|๐œ™(๐‘ข) | โ‰ค๐ถ ๐ถหœ (๐‘‘ , ๐‘ )diam(ฮฉ)๐‘ kD๐‘ ๐‘ขk๐ฟ2(ฮฉ). (4.39) The following lemma is obtained from Lemma5:

Lemma 6. For 1 โ‰ค ๐‘˜ < ๐‘™ โ‰ค ๐‘ž and ๐‘– โˆˆ ๐ฝ(๐‘™), let ๐œ™๐‘–, ๐‘ค๐‘– ๐‘— be as in Lemma 4 and Example2 and define ๐œ‘๐‘– B ร

๐‘—โˆˆ๐ผ(๐‘˜)๐‘ค๐‘– ๐‘—๐œ™(๐‘˜)

๐‘— . Then there exists a constant๐ถ(๐‘‘ , ๐‘ ) such that, for all๐‘ฃ โˆˆ๐ป๐‘ 

0(ฮฉ),

โˆซ

๐ต ๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–)

(๐œ™๐‘–โˆ’๐œ‘๐‘–) (๐‘ฅ)๐‘ฃ(๐‘ฅ)d๐‘ฅ

โ‰ค๐ถ(๐‘‘ , ๐‘ )๐œŒ๐‘ โˆ’๐‘‘/2โ„Ž(๐‘ โˆ’๐‘‘/2)๐‘˜ยฉ

ยญ

ยซ

โ„Ž๐‘™ ๐‘‘/2+ ร•

๐‘—โˆˆ๐ผหœ(๐‘˜)

|๐‘ค๐‘– ๐‘—|ยช

ยฎ

ยฌ kD๐‘ ๐‘ฃk

๐ฟ2

๐ต ๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–).

(4.40)

Proof. We apply Lemma5to the linear functional๐‘ข โ†ฆโ†’โˆซ

๐ต

๐œŒ โ„Ž๐‘˜

(๐œ™๐‘–โˆ’๐œ‘๐‘–)๐‘ข. Since the second requirement of Lemma 5is fulfilled by definition, it remains to bound หœ๐ถ. We only execute the proof for Example1; the proof for Example2is analogous. We first note that while the sum in the definition of๐œ‘๐‘–only ranges over ๐‘— โˆˆ ๐ผ(๐‘˜), we can increase it to run over all of ๐‘— โˆˆ ๐ผหœ(๐‘˜), since for ๐‘— โˆˆ ๐ผหœ(๐‘˜) \๐ผ(๐‘˜), the support of๐œ™(

๐‘˜) ๐‘— is disjoint from that of๐‘ฃ โˆˆ๐ป๐‘ 

0(ฮฉ). Let๐‘ข โˆˆ ๐ป๐‘ (ฮฉ). Writing๐ถ(๐‘‘ , ๐‘ )for the continuity constant of the embedding of๐ป๐‘ (๐ต1(0))into๐ถ๐‘(๐ต1(0)), the inequalities

max

๐ต

๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–)|๐‘ข( ยท ) |= max

๐‘ฅโˆˆ๐ต1(0)

๐‘ข

๐œŒ โ„Ž๐‘˜(๐‘ฅโˆ’๐‘ฅ๐‘–)

โ‰ค๐ถ(๐‘‘ , ๐‘ )

๐‘ 

ร•

๐‘š=0

(๐œŒ โ„Ž๐‘˜)๐‘š

[D๐‘š๐‘ข] ๐œŒ โ„Ž๐‘˜( ยท โˆ’๐‘ฅ๐‘–) ๐ฟ2(๐ต1(0))

and

[D๐‘š๐‘ข] ๐œŒ โ„Ž๐‘˜( ยท โˆ’๐‘ฅ๐‘–)

๐ฟ2(๐ต1(0)) = (๐œŒ โ„Ž๐‘˜)โˆ’๐‘‘/2kD๐‘š๐‘ขk๐ฟ2(๐ต

๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–))

imply that

|๐œ™๐‘–(๐‘ข) โˆ’๐œ‘๐‘–(๐‘ข) | โ‰คยฉ

ยญ

ยซ

โ„Ž๐‘™ ๐‘‘/2+ ร•

๐‘—โˆˆ๐ผหœ(๐‘˜)

|๐‘ค๐‘– ๐‘—|ยช

ยฎ

ยฌ max

๐‘ฅโˆˆ๐ต ๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–)

|๐‘ข(๐‘ฅ) | (4.41)

โ‰ค๐ถ(๐‘‘ , ๐‘ )๐œŒโˆ’๐‘‘/2โ„Žโˆ’๐‘˜ ๐‘‘/2ยฉ

ยญ

ยซ

โ„Ž๐‘™ ๐‘‘/2+ ร•

๐‘—โˆˆ๐ผหœ(๐‘˜)

|๐‘ค๐‘– ๐‘—|ยช

ยฎ

ยฌ

๐‘ 

ร•

๐‘š=0

(๐œŒ โ„Ž๐‘˜)๐‘škD๐‘š๐‘ขk๐ฟ2(๐ต

๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–)).

(4.42)

Therefore the first condition of Lemma5holds with

หœ

๐ถ =๐ถ(๐‘‘ , ๐‘ )๐œŒโˆ’๐‘‘/2โ„Žโˆ’๐‘˜ ๐‘‘/2ยฉ

ยญ

ยซ

โ„Ž๐‘™ ๐‘‘/2+ ร•

๐‘—โˆˆ๐ผหœ(๐‘˜)

|๐‘ค๐‘– ๐‘—|ยช

ยฎ

ยฌ

, (4.43)

and we conclude the proof by writing๐ถ(๐‘‘ , ๐‘ )for any constant depending only on๐‘‘

and๐‘ .

56 We can now conclude the proof of Lemma4.

Proof of Lemma4. Write ๐œ‘ B ร

๐‘–โˆˆ๐ฝ(๐‘™)๐›ผ๐‘–๐œ‘๐‘– and ๐œ‘๐‘– B ร

๐‘—โˆˆ๐ผ(๐‘˜) ๐‘ค๐‘– ๐‘—๐œ™(

๐‘˜)

๐‘— . Equa- tion (4.32) implies that

k๐œ™โˆ’๐œ‘k2

๐ปโˆ’๐‘ (ฮฉ) = sup

๐‘ฃโˆˆ๐ป๐‘ 

0(ฮฉ)

ร•

๐‘–โˆˆ๐ฝ(๐‘™)

2๐›ผ๐‘–

โˆซ

๐ต

๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–)

(๐œ™๐‘–โˆ’๐œ‘๐‘–) (๐‘ฅ)๐‘ฃ(๐‘ฅ)d๐‘ฅ

!

โˆ’ k๐‘ฃk2

๐ป๐‘ 

0(ฮฉ). (4.44) The packing inequality ร

๐‘–โˆˆ๐ฝ(๐‘™) kD๐‘ ๐‘ฃk2

๐ฟ2 ๐ต

๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–) โ‰ค ๐ถ(๐‘‘) โ„Ž๐‘˜โˆ’๐‘™๐œŒ/๐›ฟ๐‘‘ k๐‘ฃk2

๐ป๐‘ 

0(ฮฉ) to- gether with Lemma6yields

k๐œ™โˆ’๐œ‘k2๐ปโˆ’๐‘ (ฮฉ)โ‰ค sup

๐‘ฃโˆˆ๐ป0๐‘ (ฮฉ)

ร•

๐‘–โˆˆ๐ฝ(๐‘™)

"

2|๐›ผ๐‘–|๐ถ(๐‘‘ , ๐‘ )๐œŒ๐‘ โˆ’

๐‘‘ 2โ„Ž(๐‘ โˆ’

๐‘‘ 2)๐‘˜ยฉ

ยญ

ยซ โ„Ž

๐‘™ ๐‘‘

2 + ร•

๐‘—โˆˆ๐ผ(๐‘˜)

|๐‘ค๐‘– ๐‘—|ยช

ยฎ

ยฌ

kD๐‘ ๐‘ฃk๐ฟ2(๐ต ๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–))

(4.45)

โˆ’ (๐ถ(๐‘‘))โˆ’1 โ„Ž๐‘˜โˆ’๐‘™๐œŒ/๐›ฟ

โˆ’๐‘‘

kD๐‘ ๐‘ฃk2

๐ฟ2

๐ต ๐œŒ โ„Ž๐‘˜(๐‘ฅ๐‘–)

#

. (4.46)

Applying the inequality 2๐‘Ž๐‘ฅโˆ’๐‘๐‘ฅ2 โ‰ค ๐‘Ž2/๐‘to each summand yields

k๐œ™โˆ’๐œ‘k๐ป2โˆ’๐‘ (ฮฉ) โ‰ค๐ถ(๐‘‘) โ„Ž๐‘˜โˆ’๐‘™๐œŒ/๐›ฟ

๐‘‘ ร•

๐‘–โˆˆ๐ฝ(๐‘™)

ยฉ

ยญ

ยซ

๐›ผ๐‘—๐ถ(๐‘‘ , ๐‘ )๐œŒ๐‘ โˆ’

๐‘‘ 2โ„Ž(๐‘ โˆ’

๐‘‘ 2)๐‘˜ยฉ

ยญ

ยซ โ„Ž

๐‘™ ๐‘‘

2 + ร•

๐‘—โˆˆ๐ฝ(๐‘˜)

|๐‘ค๐‘– ๐‘—|ยช

ยฎ

ยฌ ยช

ยฎ

ยฌ

2

(4.47)

โ‰ค๐ถ(๐‘‘ , ๐‘ )๐œŒ2๐‘  ๐›ฟ๐‘‘

1+โ„Žโˆ’๐‘™ ๐‘‘๐œ”2

๐‘™ , ๐‘˜

โ„Ž2๐‘  ๐‘˜|๐›ผ|2. (4.48)

Since, for all ๐‘“ โˆˆ๐ปโˆ’๐‘ (ฮฉ),

k๐‘“k2โˆ— =[๐‘“ ,Lโˆ’1๐‘“] โ‰ค k๐‘“k๐ปโˆ’๐‘ (ฮฉ)kLโˆ’1๐‘“k๐ป๐‘ 

0(ฮฉ) โ‰ค kLโˆ’1k k๐‘“k2

๐ปโˆ’๐‘ (ฮฉ), (4.49) we havek๐œ™โˆ’๐œ‘kโˆ— โ‰ค p

kLโˆ’1k k๐œ™โˆ’๐œ‘k๐ปโˆ’๐‘ (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 ๐œŒand the norm of weights๐‘ค๐‘–, ๐‘—. Lemma 7. There exist constants ๐œŒ(๐‘‘ , ๐‘ ) and๐ถ(๐‘‘ , ๐‘ , ๐›ฟ) such that for all 1 โ‰ค ๐‘˜ <

๐‘™ โ‰ค ๐‘ž, there exist weights ๐‘ค โˆˆ R๐ฝ

(๐‘™)ร—๐ผหœ(๐‘˜)

satisfying(4.36)and (with๐œ”๐‘™ , ๐‘˜ defined as in Lemma4)

๐œ”2

๐‘™ , ๐‘˜ โ‰ค โ„Ž๐‘™ ๐‘‘๐ถ(๐‘‘ , ๐‘ , ๐›ฟ). (4.50)

Proof. For Example1, (4.36) is equivalent to โ„Ž๐‘™ ๐‘‘/2๐‘(๐‘ฅ๐‘–) = ร•

๐‘—โˆˆ๐ผหœ๐œŒ(๐‘˜)

๐‘ค๐‘– ๐‘—๐‘(๐‘ฅ๐‘—),โˆ€๐‘ โˆˆ P๐‘ โˆ’1, (4.51)

where หœ๐ผ(๐‘˜)

๐œŒ B {๐‘— โˆˆ ๐ผหœ(๐‘˜) | ๐‘ฅ๐‘— โˆˆ ๐ต(๐‘ฅ๐‘–, ๐œŒ โ„Ž๐‘˜)}.

Fix๐‘– โˆˆ ๐ฝ(๐‘™), let๐œ† >0, and write๐‘ฅ๐œ†

๐‘— B

๐‘ฅ๐‘—โˆ’๐‘ฅ๐‘–

๐œ† . Write0 B (0, . . . ,0) โˆˆR๐‘‘. Since the function ๐‘( ยท ) โ†ฆโ†’ ๐‘(ยทโˆ’๐‘ฅ๐‘–

๐œ† )is surjective onP๐‘ โˆ’1, (4.51) is satisfied if โ„Ž๐‘™ ๐‘‘/2๐‘(0) = ร•

๐‘—โˆˆ๐ผหœ๐œŒ(๐‘˜)

๐‘ค๐‘– ๐‘—๐‘(๐‘ฅ๐œ†

๐‘—),โˆ€๐‘ โˆˆ P๐‘ โˆ’1. (4.52) For a multiindex ๐‘› = (๐‘›1, . . . , ๐‘›๐‘‘) โˆˆ N๐‘‘ and a point ๐‘ง = (๐‘ง1, . . . , ๐‘ง๐‘‘) โˆˆ R๐‘‘, write ๐‘ง๐‘› B รŽ๐‘‘

๐‘š=1๐‘ง๐‘›๐‘š

๐‘š . Use the convention0๐‘› =0 if๐‘› โ‰  0and00 =1. To satisfy (4.52), it is sufficient to identify a subset ๐œŽ of หœ๐ผ(

๐‘˜)

๐œŒ and ๐‘ค๐‘–,ยท โˆˆ R๐ผหœ

(๐‘˜)

such that #๐œŽ = ๐‘ ๐‘‘, ๐‘ค๐‘–, ๐‘— =0 for ๐‘— โˆ‰๐œŽ, and

โ„Ž๐‘™ ๐‘‘/20๐‘› =ร•

๐‘—โˆˆ๐œŽ

๐‘ค๐‘– ๐‘—(๐‘ฅ๐œ†

๐‘—)๐‘›,โˆ€๐‘› โˆˆ {0, . . . , ๐‘ โˆ’1}๐‘‘. (4.53) LetV๐œ† โˆˆR{0,1,...,๐‘ โˆ’1}

๐‘‘ร—๐œŽ

be the ๐‘ ๐‘‘ร—๐‘ ๐‘‘matrix defined by V๐œ†๐‘›, ๐‘— B

๐‘ฅ๐œ†

๐‘—

๐‘›

. (4.54)

For a multiindex ๐‘› โˆˆ N๐‘‘ and a point ๐‘ฅ โˆˆ R๐‘‘, ๐‘ฅ๐‘› B รŽ๐‘‘

๐‘š=1๐‘ฅ๐‘›๐‘š. Let w โˆˆ R๐œŽ be defined byw๐‘— B ๐‘ค๐‘–, ๐‘— for ๐‘— โˆˆ๐œŽ. Equation (4.53) is then equivalent to

โ„Ž๐‘™ ๐‘‘/2e=V๐œ†w, (4.55)

wheree โˆˆ R{0,1,...,๐‘ โˆ’1}

๐‘‘

is defined by e๐‘› B 0๐‘› for ๐‘› โˆˆ {0,1, . . . , ๐‘ โˆ’1}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œŽ and๐œ† > 0 such that๐œŽmin(V๐œ†) (the minimal singular value ofV๐œ†) is bounded from below by a constant depending only on๐‘ and๐‘‘.

For ๐›ผ โ‰ฅ 0, let (๐œ–๐‘—)๐‘—โˆˆ{0,1,...,๐‘ โˆ’1}๐‘‘ be elements of R๐‘‘ satisfying |๐œ–๐‘—| โ‰ค ๐›ผ for all ๐‘— โˆˆ {0,1, . . . , ๐‘ โˆ’1}๐‘‘. Let1B (1, . . . ,1) โˆˆR๐‘‘and, for ๐‘— โˆˆ {0,1, . . . , ๐‘ โˆ’1}๐‘‘, let ๐‘ง๐‘— B 1+ ๐‘— +๐œ–๐‘—. Observe that for ๐›ผ = 0, the points ๐‘ง๐‘— are on a regular grid. Let Vยฏ๐›ผ โˆˆ R{0,1,...,๐‘ โˆ’1}

๐‘‘ร—{0,1,...,๐‘ โˆ’1}๐‘‘ be the ๐‘ ๐‘‘ร—๐‘ ๐‘‘ matrix defined by ยฏV๐›ผ๐‘›, ๐‘— B ๐‘ง๐‘—๐‘› . Let ๐‘‰ be the ๐‘ ร— ๐‘ Vandermonde matrix defined by๐‘‰๐‘–, ๐‘— = ๐‘–๐‘—. Writing๐œŽmin(๐‘‰) for the minimal singular value of๐‘‰, we have for๐›ผ=0, by [119, Theorem 4.2.12],

๐œŽmin Vยฏ0

=(๐œŽmin(๐‘‰))๐‘‘. (4.56)

Since univariate polynomial interpolation on ๐‘  points with polynomials of degree ๐‘  โˆ’1 is uniquely solvable, we have ๐œŽmin(๐‘‰) > 0 and ๐œŽmin(Vยฏ0) > ๐ถ(๐‘‘ , ๐‘ ) > 0.

58 Therefore, the continuity of the minimal singular value with respect to the entries of ยฏV๐›ผimplies that there exists๐›ผโˆ—, ๐œŽโˆ— > 0 depending only on๐‘ , ๐‘‘such that ๐›ผ โ‰ค ๐›ผโˆ— implies ๐œŽmin(Vยฏ๐›ผ) > ๐œŽโˆ—. Since (by construction) the (๐‘ฅ๐‘–)๐‘–โˆˆ๐ผหœ(๐‘˜) form a covering of R๐‘‘ of radius โ„Ž๐‘˜, the (๐‘ฅ๐œ†

๐‘–)๐‘–โˆˆ๐ผหœ(๐‘˜) form a covering of R๐‘‘ of radius โ„Ž๐‘˜/๐œ† and for each ๐‘› โˆˆ {0,1, . . . , ๐‘ โˆ’1}๐‘‘, there exists an๐‘ฅ๐œ†

๐‘—๐‘›that is at distance at mostโ„Ž๐‘˜/๐œ†from๐‘›. Let ๐œŽ B {๐‘—๐‘› | ๐‘›โˆˆ {0,1, . . . , ๐‘ โˆ’1}๐‘‘} โŠ‚ ๐ผหœ(๐‘˜) be the collection of corresponding labels.

It follows from |๐‘ฅ๐œ†

๐‘—๐‘›| โ‰ค โˆš

๐‘‘ ๐‘ +โ„Ž๐‘˜/๐œ†that |๐‘ฅ๐‘—

๐‘›โˆ’๐‘ฅ๐‘–| โ‰ค ๐œ†

โˆš

๐‘‘ ๐‘ + โ„Ž๐‘˜, and ๐œŽ โŠ‚ ๐ผหœ(๐‘˜)

๐œŒ for

๐œŒ > 1+๐œ†

โˆš

๐‘‘ ๐‘ /โ„Ž๐‘˜. Selecting๐œ† = โ„Ž๐‘˜/๐›ผโˆ— implies that๐œŽmin(V๐œ†) > ๐œŽโˆ— and๐œŽ โŠ‚ ๐ผหœ๐œŒ(๐‘˜) for ๐œŒ >1+โˆš

๐‘‘ ๐‘ /๐›ผโˆ—. Defining

๐‘ค๐‘– ๐‘— B

๏ฃฑ๏ฃด

๏ฃด

๏ฃฒ

๏ฃด๏ฃด

๏ฃณ

(V๐œ†)โˆ’1โ„Ž๐‘™ ๐‘‘/2e

๐‘›

, if ๐‘— = ๐‘—๐‘›โˆˆ ๐œŽ,

0, otherwise,

(4.57)

the weights๐‘ค๐‘– ๐‘— satisfy๐œ”๐‘˜ ๐‘™ โ‰ค๐ถ(๐‘ , ๐‘‘)โ„Ž๐‘™ ๐‘‘/2and (4.36) with a๐œŒ depending only on๐‘  and๐‘‘. This concludes the proof for Example1. The proof is similar for Example2 with minor changes (the bound on๐œ” also depends on๐›ฟ).

The following lemma concerns the satisfaction of the second condition of Theorem6:

Lemma 8. In the setting of Examples1and2, there exists some constant๐ถ(๐‘‘ , ๐‘ , ๐›ฟ) >

0such that, for2โ‰ค ๐‘˜ < ๐‘™ โ‰ค ๐‘ž,๐›ผ โˆˆR๐ฝ

(๐‘™) and๐œ™ =ร

๐‘–๐›ผ๐‘–๐œ™๐‘–, min

๐œ‘โˆˆspan(๐œ™๐‘–)

๐‘–โˆˆ๐ผ(๐‘˜โˆ’1)

k๐œ™โˆ’๐œ‘kโˆ—2

|๐›ผ|2 โ‰ค๐ถ(๐‘‘ , ๐‘ , ๐›ฟ) kLโˆ’1kโ„Ž2๐‘ (๐‘˜โˆ’1). (4.58) Proof. Apply Lemma4with the bounds on ๐œŒand๐œ” 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 constant๐ถ(๐‘‘ , ๐‘ , ๐›ฟ)such that Condition2is fulfilled with๐ถฮฆ B max( kL k,kLโˆ’1k)๐ถ(๐‘‘ , ๐‘ , ๐›ฟ) and๐ป B โ„Ž๐‘ . Propagation of exponential decay

We will now derive the exponential decay of the Cholesky factors๐ฟ by combining the algebraic identities of Lemma 1with the bounds on the condition numbers of the ๐ต(๐‘˜) (implied by Condition2) and the exponential decay of the ๐ด(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‘‘: ๐ผร—๐ผ โˆ’โ†’R+is called ahierarchical pseudometricif (1) ๐‘‘(๐‘–, ๐‘–) =0, for all๐‘– โˆˆ๐ผ;

(2) ๐‘‘(๐‘–, ๐‘—) =๐‘‘(๐‘— , ๐‘–), for all๐‘–, ๐‘— โˆˆ ๐ผ;

(3) for all1 โ‰ค ๐‘˜ โ‰ค ๐‘ž, ๐‘‘( ยท, ยท )restricted to๐ฝ(๐‘˜) ร—๐ฝ(๐‘˜) is a pseudometric;

(4) for all 1 โ‰ค ๐‘˜ โ‰ค ๐‘™ โ‰ค ๐‘š โ‰ค ๐‘ž and ๐‘– โˆˆ ๐ฝ(๐‘˜), ๐‘  โˆˆ ๐ฝ(๐‘™), ๐‘— โˆˆ ๐ฝ(๐‘š), we have ๐‘‘(๐‘–, ๐‘—) โ‰ค 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 )and๐›พ โˆˆR+, let

๐‘๐‘‘(๐›พ) B sup

1โ‰ค๐‘˜โ‰ค๐‘™โ‰ค๐‘ž

sup

๐‘—โˆˆ๐ฝ(๐‘™)

ร•

๐‘–โˆˆ๐ฝ(๐‘˜)

exp(โˆ’๐›พ ๐‘‘(๐‘–, ๐‘—)). (4.59)

The following theorem states the main result of this section:

Theorem 8(Exponential decay of the Cholesky factors). Assume thatฮ˜fulfills Con- ditions1and2 with the constants๐›พ , ๐ถ๐›พ, ๐ป , ๐ถฮฆ and the hierarchical pseudometric ๐‘‘( ยท, ยท ). Then

(chol(ฮ˜))๐‘– ๐‘—

โ‰ค 2๐ถฮฆ๐‘๐‘‘(๐›พหœ/8)2

(1โˆ’๐‘Ÿ)2 4๐‘๐‘‘(๐›พหœ/4)๐ถฮฆ๐ถ๐›พ(๐‘๐‘‘(๐›พหœ/2))2 (1โˆ’๐‘Ÿ)2

!๐‘ž exp

โˆ’๐›พหœ 8๐‘‘(๐‘–, ๐‘—)

, (4.60) where ๐ถ๐‘… B max

n 1, 2

๐ถ๐›พ๐ถฮฆ 1+๐œ…

o

, ๐‘Ÿ B 1โˆ’๐œ…

โˆ’1

1+๐œ…โˆ’1, ๐›พหœ B โˆ’log(

๐‘Ÿ)

1+log(๐‘๐‘‘(๐›พ/2))+log(๐ถ๐‘…)โˆ’log(๐‘Ÿ) ๐›พ 2, and ๐œ… =๐ปโˆ’2๐ถ2

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 ๐ผ be an index set that is partitioned as ๐ผ = ๐ฝ(1) โˆช ยท ยท ยท๐ฝ(๐‘ž) and let ๐‘‘: ๐ผร—๐ผ โ†’Rโ‰ฅ0satisfy

๐‘‘(๐‘–1, ๐‘–๐‘›+1) โ‰ค

๐‘›

ร•

๐‘˜=1

๐‘‘(๐‘–๐‘˜, ๐‘–๐‘˜+1) for all1 โ‰ค ๐‘›โ‰ค ๐‘žโˆ’1and๐‘–๐‘˜ โˆˆ ๐ฝ(๐‘˜).

60 Let ๐‘€(๐‘˜) โˆˆ R๐ฝ

(๐‘˜)ร—๐ฝ(๐‘˜+1)

be such that |๐‘€(๐‘˜)

๐‘–, ๐‘— | โ‰ค ๐ถexp(โˆ’๐›พ ๐‘‘(๐‘–, ๐‘—)) for1 โ‰ค ๐‘˜ โ‰ค ๐‘žโˆ’1 and let

๐‘๐‘‘(๐›พ/2) B sup

1โ‰ค๐‘˜โ‰ค๐‘žโˆ’1

sup

๐‘—โˆˆ๐ฝ(๐‘˜+1)

ร•

๐‘–โˆˆ๐ฝ(๐‘˜)

exp

โˆ’๐›พ

2๐‘‘(๐‘–, ๐‘—)

for๐›พ โˆˆR+. (4.61) Then, for1โ‰ค ๐‘› โ‰ค ๐‘žโˆ’1,

๐‘›

ร–

๐‘˜=1

๐‘€(๐‘˜)

!

๐‘–, ๐‘—

โ‰ค (๐‘๐‘‘(๐›พ/2)๐ถ)๐‘›exp

โˆ’๐›พ

2๐‘‘(๐‘–, ๐‘—) .

Proof. Set๐‘–1 B๐‘–,๐‘–๐‘›+1 B ๐‘—. Then

๐‘›

ร–

๐‘˜=1

๐‘€(๐‘˜)

!

๐‘–, ๐‘—

โ‰ค๐ถ๐‘›

ร•

๐‘–2,...,๐‘–๐‘›โˆˆ๐ฝ(2),...,๐ฝ(๐‘›)

exp โˆ’๐›พ

๐‘›

ร•

๐‘˜=1

๐‘‘(๐‘–๐‘˜, ๐‘–๐‘˜+1)

!

โ‰ค๐ถ๐‘›exp

โˆ’๐›พ

2๐‘‘(๐‘–1, ๐‘–๐‘›+1) ร•

๐‘–2,...๐‘–๐‘›โˆˆ๐ผ

exp โˆ’๐›พ 2

๐‘›

ร•

๐‘˜=1

๐‘‘(๐‘–๐‘˜, ๐‘–๐‘˜+1)

!

โ‰ค (๐‘๐‘‘(๐›พ/2)๐ถ)๐‘›exp

โˆ’๐›พ

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 ๐ดโˆˆR๐ผร—๐ผ be symmetric and positive definite such that for๐ถ , ๐›พ >0 and a metric๐‘‘( ยท,ยท )on ๐ผwe have |๐ด๐‘–, ๐‘—| โ‰ค๐ถexp(โˆ’๐›พ ๐‘‘(๐‘–, ๐‘—)). It holds true that

(๐ดโˆ’1)๐‘– , ๐‘—

โ‰ค 4

k๐ดk + k๐ดโˆ’1kโˆ’1

(1โˆ’๐‘Ÿ)2exp โˆ’ log(๐‘Ÿ1)

(1+log(๐‘๐‘‘(๐›พ/2)) +log(๐ถ๐‘…)) +log(1๐‘Ÿ) ๐›พ 2๐‘‘(๐‘–, ๐‘—)

! (4.62)

where ๐‘๐‘‘(๐›พ/2) B sup๐‘—โˆˆ๐ผ

ร

๐‘–โˆˆ๐ผexp โˆ’๐›พ2๐‘‘(๐‘–, ๐‘—)

, ๐ถ๐‘… B max n

1, 2๐ถ

k๐ดk+k๐ดโˆ’1kโˆ’1

o

= maxn

1,2

๐ถk๐ดโˆ’1k 1+๐œ…

o , ๐‘Ÿ B

1โˆ’ 1

k๐ดk k๐ดโˆ’1k

1+ 1

k๐ดk k๐ดโˆ’1k

= 1โˆ’๐œ…โˆ’1

1+๐œ…โˆ’1, and ๐œ… B k๐ดk k๐ดโˆ’1k is the condition number of ๐ด.

Proof. On a compact set not containing 0, the function๐‘ฅ โ†ฆโ†’ ๐‘ฅโˆ’1can be accurately approximated by low-order polynomials in๐‘ฅ. 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. Let๐ต โˆˆR๐ผร—๐ผ 'R๐‘ร—๐‘be symmetric and positive definite with condition number ๐œ… and such that

๐ต๐‘–, ๐‘—

โ‰ค ๐ถexp(โˆ’๐›พ ๐‘‘(๐‘–, ๐‘—)) for some constant ๐ถ > 0 and some metric ๐‘‘ on ๐ผ. Let ๐ฟ be the Cholesky factor (in an arbitrary order) of ๐ตโˆ’1 (๐ตโˆ’1 =๐ฟ ๐ฟ๐‘‡). Then

๐ฟ๐‘– , ๐‘—

โ‰ค 4p k๐ตk k๐ตk + k๐ตโˆ’1kโˆ’1

(1โˆ’๐‘Ÿ)2 exp

log(๐‘Ÿ)

1+log(๐‘๐‘‘(๐›พ/2)) +log(๐ถ๐‘…) โˆ’log(๐‘Ÿ) ๐›พ 2๐‘‘(๐‘–, ๐‘—)

(4.63) where ๐‘๐‘‘(๐›พ/2) B sup๐‘—โˆˆ๐ผ

ร

๐‘–โˆˆ๐ผexp โˆ’๐›พ

2๐‘‘(๐‘–, ๐‘—)

, ๐ถ๐‘… B max n

1, 2๐ถk๐ต

โˆ’1k 1+๐œ…

o

, and๐‘Ÿ B

1โˆ’๐œ…โˆ’1 1+๐œ…โˆ’1.

Proof. Lemma2 implies that the Schur complements of ๐ตโˆ’1 can be expressed as inverses of sub-matrices of๐ต. 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 ฮ˜is the following lemma showing the stability of exponential decay under inversion for block-lower-triangular matrices (this operation appears in the definition of ยฏ๐ฟ in (4.14)):

Lemma 12. Let๐ผbe an index set that is partitioned as๐ผ =๐ฝ(1)โˆช ยท ยท ยท๐ฝ(๐‘ž)and assume that the matrix๐ฟ โˆˆR๐ผร—๐ผ is block-lower triangular with respect to this partition, with identity matrices as diagonal blocks. If๐‘‘( ยท,ยท )is a hierarchical pseudometric such that|๐ฟ๐‘– ๐‘—| โ‰ค๐ถexp(โˆ’๐›พ ๐‘‘(๐‘–, ๐‘—))(for some๐ถ โ‰ฅ 1and๐›พ > 0), then it holds true that

(๐ฟโˆ’1)๐‘– ๐‘—

โ‰ค 2๐‘ž(๐‘๐‘‘(๐›พ/2)๐ถ)๐‘žexp

โˆ’๐›พ

2๐‘‘(๐‘–, ๐‘—)

(4.64) with๐‘๐‘‘(๐›พ) Bsup1โ‰ค๐‘˜โ‰ค๐‘™โ‰ค๐‘žsup๐‘—โˆˆ๐ฝ(๐‘™) ร

๐‘–โˆˆ๐ฝ(๐‘˜)exp(โˆ’๐›พ ๐‘‘(๐‘–, ๐‘—)).

Proof. The Neumann series of a๐‘žร—๐‘ž block-lower-triangular matrix with identity matrices on the (block) diagonal can be written as

๐ฟโˆ’1=

๐‘ž

ร•

๐‘˜=0

(Idโˆ’๐ฟ)๐‘˜ . (4.65)

62 Since the sum terminates in๐‘žsteps, 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.

Dalam dokumen Inference, Computation, and Games (Halaman 75-85)