21.06.2014 Views

H-Matrix approximation for the operator exponential with applications

H-Matrix approximation for the operator exponential with applications

H-Matrix approximation for the operator exponential with applications

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

96 I. P. Gavrilyuk et al.<br />

of <strong>the</strong> order k ≪ N = dim V h <strong>for</strong> σ × τ ∈ P far , see [13]. The reduction<br />

<strong>with</strong> respect to <strong>the</strong> operation count is achieved by replacing <strong>the</strong> full matrix<br />

blocks A τ×σ (τ × σ ∈ P far ) by <strong>the</strong>ir low-rank <strong>approximation</strong><br />

k∑<br />

H := a ν · c T ν , a ν ∈ R nτ , c ν ∈ R nσ ,<br />

A τ×σ<br />

ν=1<br />

{ ∫ { ∫<br />

where a ν =<br />

X(τ) a ν(x)ϕ i (x)dx}<br />

, c ν =<br />

i∈τ<br />

X(σ) c ν(y)ϕ j (y)dy}<br />

.<br />

j∈σ<br />

There<strong>for</strong>e, we obtain <strong>the</strong> following storage and matrix-vector multiplication<br />

cost <strong>for</strong> <strong>the</strong> matrix blocks<br />

( ) ( )<br />

N st A<br />

τ×σ<br />

H = k(nτ + n σ ), N MV A<br />

τ×σ<br />

H =2k(nτ + n σ ),<br />

where n τ =#τ, n σ =#σ. On <strong>the</strong> o<strong>the</strong>r hand, <strong>the</strong> <strong>approximation</strong> of <strong>the</strong><br />

order O(N −α ), α>0, is achieved <strong>with</strong> k = O(log d−1 N).<br />

3.3 The error analysis<br />

For <strong>the</strong> error analysis, we consider <strong>the</strong> uni<strong>for</strong>m hierarchical cluster tree T (I)<br />

(see [12,13] <strong>for</strong> more details) <strong>with</strong><strong>the</strong> depthL suchthat N =2 dL . Define<br />

P (l)<br />

2 := P 2 ∩ T2 l, where T 2 l is <strong>the</strong> set of clusters τ × σ ∈ T 2 suchthat blocks<br />

τ,σ belong to level l, <strong>with</strong> l =0, 1,...,L. We consider <strong>the</strong> expansions<br />

<strong>with</strong><strong>the</strong> local rank k l depending only on <strong>the</strong> level number l and defined by<br />

k l := min{2 d(L−l) ,m d−1<br />

l<br />

}, where m = m l is given by<br />

(3.4) m l = aL 1−q (L − l) q + b, 0 ≤ q ≤ 1, a,b > 0.<br />

Note that <strong>for</strong> q =0, we arrive at <strong>the</strong> constant order m = O(L), which<br />

leads to <strong>the</strong> <strong>exponential</strong> convergence of <strong>the</strong> H-matrix <strong>approximation</strong>, see<br />

[16].<br />

Introduce<br />

{<br />

N 0 = max max ∑<br />

max<br />

0≤l≤p<br />

τ∈T (l)<br />

τ:τ×σ∈P (l)<br />

2<br />

∑<br />

}<br />

1, max<br />

1 .<br />

σ∈T (l) σ:τ×σ∈P (l)<br />

2<br />

For <strong>the</strong> ease of exposition, we consider <strong>the</strong> only two special cases q =0<br />

and q =1. Denote by A h : V h → V<br />

h ′ <strong>the</strong> restriction of A onto <strong>the</strong> Galerkin<br />

subspace V h ⊂ L 2 (Ω) defined by 〈A h u, v〉 = 〈Au, v〉 <strong>for</strong> all u, v ∈ V h .The<br />

<strong>operator</strong> A H has <strong>the</strong> similar sense. The following statement is <strong>the</strong> particular<br />

case of [15, Lemma 2.4].<br />

Lemma 3.3 Let η =2 −α ,α>0, and<br />

|s(x, y) − s τσ (x, y)| η m l<br />

l 3−d dist(τ,σ) 2−d

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!