Computational Methods in Applied Mathematics

Vol. 12 (2012), No. 4, pp. 448–468

⃝
c 2012 Institute of Mathematics, NAS of Belarus
Doi: 10.2478/cmam-2012-0026

Polynomial of Best Uniform Approximation

to 1/x and Smoothing in Two-level Methods
Johannes Kraus · Panayot Vassilevski · Ludmil Zikatanov

In memory of Sergei Nepomnyaschikh - a pioneer in domain decomposition methods

Abstract — We derive defect correction scheme for constructing the sequence of poly-
nomials of best approximation in the uniform norm to 1/x on a finite interval with
positive endpoints. As an application, we consider two-level methods for scalar elliptic
partial differential equation (PDE), where the relaxation on the fine grid uses the afore-
mentioned polynomial of best approximation. Based on a new smoothing property of
this polynomial smoother that we prove, combined with a proper choice of the coarse
space, we obtain as a corollary, that the convergence rate of the resulting two-level
method is uniform with respect to the mesh parameters, coarsening ratio and PDE
coefficient variation.
2010 Mathematical subject classification: 65F10; 65N20; 65N30.
Keywords: polynomial smoothing; best uniform approximation; two-level methods; el-
liptic problems with large coefficient variation.

1. Introduction
The polynomial of best approximation in uniform norm to 1/x on a finite interval can be
found in different forms in many classical texts on approximation theory, for example, see [10,
1
p. 33, Eq. (4.25)], [11, Exercise 1.20]. In fact, the approximating polynomial for ,
t−a
a > 1, has already been discovered by Chebyshev in 1887, see [4]. Here, we derive a defect
correction algorithm for constructing the sequence of polynomials of best approximation
and show several results important for the applications, namely, sufficient conditions for
positivity and the monotonicity of this sequence.
As an application we study two-level methods with smoothers based on the polynomial
of best approximation to 1/x on a finite interval [λmin , λmax ], 0 < λmin < λmax , in the
∥ · ∥∞ (uniform) norm. We base our iterations on a three term recurrence which is given
Johannes Kraus
Johann Radon Institute for Computational and Applied Mathematics, Austrian Academy of Sciences Al-
tenberger Str. 69, 4040 Linz, Austria
E-mail: johannes.kraus@oeaw.ac.at.
Panayot Vassilevski
Center for Applied Scientific Computing, Lawrence Livermore National Laboratory, P.O. Box 808, L-560,
Livermore, CA 94550, USA.
E-mail: panayot@llnl.gov.
Ludmil Zikatanov
Department of Mathematics, The Pennsylvania State University, University Park, PA 16802, USA.
E-mail: ludmil@psu.edu.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 449

in [9]. For such smoothers we show a smoothing property of the polynomial and as a
corollary, based on an abstract two-level estimate we derive two-level (TL) or two-grid (TG)
convergence estimates in the case of discretized elliptic PDE with heterogeneous coefficients.
The estimate explicitly depends on the degree of the polynomial (or on the range of the
spectrum which needs to be resolved by the smoother). We prove that if coarse spaces with
stability and approximation properties that are robust with respect coefficient variation are
used, then the two-level methods with polynomial smoothers based on the polynomial of
best approximation to 1/x are robust with respect to the variation in the coefficients of
the PDE. Several examples of coarse spaces that provide the required contrast independent
approximation property are available in the literature, cf., e.g., [6], [12], and earlier [1] as
modified recently in [3]).
The paper is organized as follows. In Section 2 we introduce the three-term recurrence
relation for the polynomial of best approximation to 1/x and give description of algorithms
for computing such approximations with matrix arguments. We then discuss the properties
of the sequence of polynomials of best approximation to 1/x in Section 3. In Section 4
we discuss and prove the smoothing property of the polynomial smoother. This property
explicitly involves the polynomial degree and we use it in an abstract two-level convergence
result. As a corollary, we derive an estimate for the convergence rate in case of finite element
discretization of scalar elliptic PDE with coarse spaces that provide contrast independent
approximation resulting in a contrast independent two-grid convergence. This convergence
behavior is illustrated also with numerical tests in Section 5.

2. Best polynomial approximation to 1/x in uniform norm

We begin with notation and some simple and well known definitions related to Chebyshev
polynomials. We consider a finite interval, [λmin , λmax ], with 0 < λmin < λmax < ∞. We
denote
λmax 1 λmax + λmin κ+1
κ= , σ= , a= = . (2.1)
λmin λmax − λmin λmax − λmin κ−1
Note that a > 1 and σ > 0. The change of variables
( )
2 λmax + λmin
t= x− = 2σx − a,
λmax − λmin 2

1
maps the interval [λmin , λmax ] to [−1, 1]. The inverse map is x = (t + a), and hence,
2σ
1 2σ
= . We thus aim to find the polynomial of degree less than or equal to m of best
x t+a
1
approximation in the norm ∥ · ∥∞,[−1,1] of f (t) = , a > 1. We note that if Qm (t) is the
t+a
polynomial of best approximation to 1/(t + a) on [−1, 1], and the error of approximation is

1
E[−1,1] = min −Q ,
Q∈Pm t+a L∞ [−1,1]

then
1
qm (x) := 2σQm (2σx − a), and E = min −q = 2σE[−1,1] (2.2)
q∈Pm x L∞ [λmax ,λmin ]
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
450 Johannes Kraus et al.

are the polynomial of best approximation in L∞ -norm on [λmin , λmax ] and the error of ap-
proximation, respectively.
We denote the (first kind) Chebyshev polynomial of degree j by Tj . For Tj (ξ) ∈ Pj we
have
1[ √ √ ] 1[ √ √ ]
Tj (ξ) = (ξ + ξ 2 − 1)j + (ξ + ξ 2 − 1)−j = (ξ + ξ 2 − 1)j + (ξ − ξ 2 − 1)j .
2 2
We recall that
Tj (t) = cos j arccos(t), t ∈ [−1, 1]
and denote √
√ κ−1
δ := a − a2 −1= √ , η = −δ. (2.3)
κ+1
√
Evidently, 0 6 δ < 1, δ −1 = a + a2 − 1, η < 0 and δ = |η|.
With this notation in hand, we have the following identities,
1 1 2
a = − (η + η −1 ), = , (2.4)
2 t+a 2t − η − η −1

and directly from the expression for Tj (ξ) given above, we also have

Tj (a) = 12 (−1)k (η k + η −k ), Tj (−a) = 12 (η k + η −k ). (2.5)

2.1. Approximation error and three-term recurrence

1
We now give a representation of the best polynomial approximation to in the L∞ -norm
t+a
on the interval [−1, 1].
1
Theorem 2.1. For the polynomials of best uniform approximation to on [−1, 1],
t+a
the following three-term recurrence relation holds:

η −1 Qm+2 (t) − 2tQm+1 (t) + ηQm (t) = −2, m = 0, 1, . . . (2.6)

with
a 1 t
Q0 (t) = , Q1 (t) = √ − 2 .
a2 −1 a2 −1 a −1
The error of approximation is:

1 δm
E[−1,1] = min −Q = .
Q∈Pm t+a L∞ [−1,1] a2 − 1

Proof. The proof of the three term recurrence relation follows from [9, Theorem 1], where
1
such recurrence relation for the best approximation to is shown. How to derive the
t−a
1 1
approximation for from the one for is straightforward by showing that Qm (t), (the
t+a t−a
1 em (−t),
polynomial of best uniform approximation to on [−1, 1]) satisfies Qm (t) = −Q
t+a
where Qem (t) ∈ Pm , is the polynomial of best uniform approximation to 1 , a > 1 on
t−a
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 451

[−1, 1]. The error of approximation is also straightforward to estimate by showing that
Qm (t) also can be written as

1 2η m Rm+1 (t)
Qm (t) = − , (2.7)
t + a (η − η ) t + a
−1 2

where
Rm+1 (t) = η −1 Tm+1 (t) − 2Tm (t) + ηTm−1 (t). (2.8)

This form of Qm (t) is derived by showing that the right hand side of (2.7) satisfies the three
term recurrence given in (2.6).

Next lemma gives an estimate on |Rm+1 (t)| by a linear polynomial, which is used later
to derive a sufficient condition for the positivity of Qm (t).
Lemma 2.1. The following estimate holds for the polynomial Rm+1 (t) defined in the
proof of Theorem 2.1 (equation (2.8)):

−2(t + a) 6 Rm+1 (t) 6 2(t + a), t ∈ [−1, 1]. (2.9)

Proof. Recall that by the definition of η and δ (see (2.3)), we have that η < 0, and |η| = δ.
Let t = cos α, for α ∈ [0, π]. Then we find that

Rm+1 (t) + 2t − η − η −1 = η −1 (Tm+1 (t) − 1) − 2(Tm (t) − t) + η(Tm−1 (t) + 1)

m+1 m+1 m−1
= −2η −1 sin2 α + 4 sin α sin α
2 2 2
m−1
− 2η sin2 α
2
( )2
−1 m+1 m−1
= −2η sin α − η sin α
2 2
( )2
−1 m+1 m−1
= 2δ sin α + δ sin α > 0. (2.10)
2 2

In an analogous fashion we obtain

Rm+1 (t) − 2t + η + η −1 = η −1 (Tm+1 (t) + 1) − 2(Tm (t) + t) + η(Tm−1 (t) + 1)

m+1 m+1 m−1
= 2η −1 cos2 α − 4 cos α cos α
2 2 2
m−1
+ 2η cos2 α
2
( )2
−1 m+1 m−1
= 2η cos α − η cos α
2 2
( )2
−1 m+1 m−1
= −2δ cos α + δ cos α 6 0. (2.11)
2 2

Combining (2.10) and (2.11) and using 2t − η − η −1 = 2(t + a) yields the desired result.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
452 Johannes Kraus et al.

2.2. Stationary iteration with the polynomial of best approximation to 1/x

After rescaling, the result in Theorem 2.1 gives us the polynomial approximation on the
interval [λmax , λmin ]. Indeed, the recurrence relation for qm+1 (x) = 2σQm+1 (2σx − a) is:

Qm+1 (2σx − a) = η[−2 + 2(2σx − a)Qm (2σx − a) − ηQm−1 (2σx − a)].

Multiplying by 2σ then gives

qm+1 (x) = η[−4σ + 2σ(2σx − a)Qm (2σx − a) − 2σηQm−1 (2σx − a)]. (2.12)
This formula can be used to perform stationary iterations towards solving Au = f for
a given symmetric and positive definite matrix A and a given symmetric positive definite
preconditioner D to A. We first write η, σ and a in terms of µ0 = 1/λmax and µ1 = 1/λmin
and δ (defined in (2.3)). The reason for choosing these parameters is that the constants in
the algorithm, with the exception of δ, are symmetric with respect to µ0 and µ1 . From the
three term recurrence relation for qm (x), (2.12), it is straightforward to get the following
identity
4µ0 µ1
qm+1 (x) − qm (x) = δ 2 (qm (x) − qm−1 (x)) + √ √ [1 − xqm (x)] . (2.13)
( µ0 + µ1 )2
Note that this identity allows us to write qm+1 as a correction to qm and therefore computing
qm+1 looks like iteration in a defect-correction method: First computing the residual [1 −
qm (x) x], and then trying to correct it by adding an additional term. One can also easily see
that for any initial q0 and q1 , if the sequence qm (x) converges, then it converges to x−1 . In
other words, choosing q0 and q1 different from what they are above, will not generate the
sequence of best approximations to x−1 , but still this sequence will converge to x−1 .
We now move on to consider a standard stationary iterative method of the form: Given
an approximation v to the solution u of the linear system in hand, the next approximation
w is defined as
w = v + R (f − Av) .
A sequence of such approximations, approaching u (when the method is convergent) is
obtained by applying this iteration with w = uj+1 , v = uj , j = 0, . . ., and, with u0 , a given
initial guess.
In our focus are iterative methods with R which is polynomial in A and we define

R = qm (D−1 A)D−1 ,
] [
λ
where qm is the polynomial of best approximation to 1/x on the interval , λ with λ
κ
an upper bound for the largest eigenvalue of D−1 A and κ > 1, a parameter controlling the
length of the interval. In general, iterations with polynomial R are known as semi-iterative
methods (see [7] or earlier, [14]).
Algorithm 2.2 (Polynomial Preconditioning with R = qm (D−1 A)D−1 ).
Given r, the following steps provide Rr = qm (D−1 A)D−1 r:
0. Initially, compute r = D−1 r and set
1 1 √ √
v0 = (µ0 + µ1 )r and v1 = ( µ0 + µ1 )2 r − µ0 µ1 D−1 Ar.
2 2
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 453

1. For j = 1, 2, . . . , m − 1, compute the current and preconditioned residuals,

rj = r − Avj , r j = D−1 rj .

Next, vj+1 is computed based on the recurrence formula (2.13) with matrix argument
4µ0 µ1
vj+1 = vj + δ 2 (vj − vj−1 ) + √ √ rj .
( µ0 + µ1 )2

2. Finally, set Rr = vm .

3. Properties of the sequence of polynomials

λ
To simplify the presentation, we now set λ = λmax and in this notation we have λmin =
κ
1
(recall the definition of κ given in §2). We thus consider the best approximation qm (x) to
[ ] x
λ
on the interval , λ . We prove several results on the positivity of the polynomial qm (x),
κ
and the monotonicity of the sequence {qm } for sufficiently large m.
We first note the following identity

x qm (x) = 2σx Qm (2σx − a) = (t + a) Qm (t)

2η m 2(−1)m δ m (3.1)
= 1− R m+1 (t) = 1 − Rm+1 (t), t ∈ [−1, 1]
(η − η −1 )2 (δ − δ −1 )2
This gives
2(−1)m δ m
1 − xqm (x) = Rm+1 (t). (3.2)
(δ − δ −1 )2
[ ]
λ
Next Lemma shows that (1 − qm (x)x) > 0 for all x ∈ 0, .
κ
Lemma 3.1. Let qm (x) be the polynomial of degree less than or[ equal
] to m, which fur-
1 λ
nishes the best approximation to in the L∞ -norm on the interval , λ , κ > 1. Then the
x κ
following inequality holds:
( ]
λ
0 < 1 − xqm (x), ∀x ∈ 0, (3.3)
κ
Proof. Consider the polynomial

p(x) = 1 − xqm (x).

Note that p(x) is of degree at most (m + 1). Since we have

( )
1
p(x) = x − qm (x) ,
x
and
( x > 0 in)the intervals of interest, we may conclude that the sign changes in the function
1
− qm (x) are the same as the sign changes in p(x) for any x > 0. However, qm (x) is the
x
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
454 Johannes Kraus et al.

1
polynomial of best uniform approximation to , and hence there are at least (m + 2) points
[ ]x
λ
of Chebyshev alternance in the interval , λ . Thus, there exist points {xj }m+2
j=1 such that
κ
λ
6 x1 < x2 < . . . < xm+1 < xm+2 6 λ,
κ
and also such that
( ) ( )
1 1
− qm (xj ) = − − qm (xj+1 ) , j = 1, . . . , (m + 1).
xj xj+1
( )
1
We define now ε := − qm (x1 ) , and use the alternation property to get that
x1

p(xj )p(xj+1 ) = −xj xj+1 ε2 < 0, j = 1, . . . , (m + 1).

Hence, we may conclude that all the roots of p(x) are disjoint, and that each of them lies
in the open interval (xj , xj+1 ), j = 1, . . . ,((m +)1). We may also conclude that there are no
λ
roots of p(x) outside of the open interval , λ and there are no roots of its first derivative
κ
outside this interval. This is so by the Rolle’s theorem: the first derivative p′ (x) clearly
has m distinct roots, each lying between the roots [ ]of p(x). Hence, p(x) is either strictly
λ
increasing or strictly decreasing on the interval 0, and also it cannot have a zero in this
κ
interval. Recall that 0 < δ = −η < 1 and that Tj (−1) = (−1)j . Using the definition of
Rm+1 (t) from the proof of Theorem 2.1 (Equation (2.8)), and the relation (3.2) it follows
that
( )
λ 2(−1)m δ m
p = Rm+1 (−1)
κ (δ − δ −1 )2
2(−1)m δ m
= [(−δ −1 )(−1)m+1 − 2(−1)m + (−δ)(−1)m−1 ]
(δ − δ )
−1 2

2δ m −1 2δ m
= (δ + δ − 2) = < 1 = p(0).
(δ − δ −1 )2 (δ + δ −1 + 2)
Here we have used that

(δ − δ −1 )2 = [(δ 2 + δ − 2 )2 (δ 2 + δ − 2 )2 = (δ + δ −1 − 2)(δ + δ −1 + 2).

1 1 1 1
(3.4)
( ]
λ λ
We thus conclude that p(0) > p( ) and therefore p(x) must be decreasing on 0, , and
κ κ
this leads to ( )
2δ m λ
0< −1
=p 6 p(x) 6 1, (3.5)
(δ + δ + 2) κ
which concludes the proof.
[ ]
λ
Next lemma shows that for x ∈ 0, the sequence of polynomials of best approximation
κ
of increasing degree is monotone.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 455

[ ]
λ
Lemma 3.2 (Monotonicity). For all x ∈ 0, we have that qm (x) < qm+1 (x), where
κ
1
qj (x), j = m, (m + 1) is the best polynomial approximation of degree at most j to in the
[ ] x
λ
L∞ -norm on the interval , λ , κ > 1.
κ
[ ]
λ
Proof. The proof amounts to showing that [qm+1 (x) − qm (x)] > 0 for all x ∈ 0, . For
κ
λ
such values of x we have x 6 = µ−1 1 , and, hence
κ
1 √ 1
q1 (x) − q0 (x) = (µ0 + µ1 + 2 µ0 µ1 ) − µ0 µ1 x − (µ0 + µ1 )
2 2 √
√ √ √ −1 √ κ−1
= µ0 µ1 (1 − x µ0 µ1 ) > µ0 µ1 (1 − µ1 µ0 µ1 ) = > 0.
λ
Further, from (2.13) and Lemma 3.1 we have
4µ0 µ1
qm+1 (x) − qm (x) = √ √ [1 − qm (x) x] + δ 2 [qm (x) − qm−1 (x)]
( µ0 + µ1 )2
4κ
= √ [1 − qm (x) x] + δ 2 [qm (x) − qm−1 (x)]
λ(1 + κ)2
8κδ m
> √ 2 + δ 2 [qm (x) − qm−1 (x)].
λ(1 + κ) (δ + δ −1 + 2)
4κ
Noticing that (δ + δ −1 + 2) = then leads to:
κ−1
2 m+1
qm+1 (x) − qm (x) > δ + δ 2 [qm (x) − qm−1 (x)]. (3.6)
λ
Clearly,
qm+1 − qm (x) > 0, if qm (x) − qm−1 (x) > 0,
and a standard induction argument concludes the proof of the lemma.
Next lemma is a straightforward corollary of Lemma 3.1.
1
Lemma 3.3. Let qm (x) be the best polynomial approximation of degree at most m to
[ ] x
λ
in L∞ -norm on the interval , λ , κ > 1. Suppose that qm (x) is positive on the interval
[ ] κ
λ
, λ . Then qm (x) is positive on the whole interval x ∈ (0, λ].
κ
Proof. We ] already shown in the previous lemma that qm (x) > q0 (x)[> 0,]for all m > 1
[ have
λ λ
and x ∈ 0, . Since, by assumption qm (x) is positive on the interval , λ the proof is
κ κ
complete.
In the two-level method convergence estimates in the next section, we will use the fol-
lowing result (which also includes a sufficient condition for the positivity of qm (x)).
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
456 Johannes Kraus et al.

Lemma 3.4. Assume that κ and m are such that

(√ )m
κ−1 ω
√ 6 , for some ω ∈ (0, 2). (3.7)
κ+1 κ−1

Then the following inequality holds for all x ∈ (0, λ]:

{ }
1 κ+1 2−ω 1( ω)
min , 6 qm (x) 6 1+ . (3.8)
2 λ x x 2
] [
λ
Proof. Lower bound: We prove first the lower bound when x ∈ , λ . Let Rm+1 (t) be the
κ
polynomial that has been defined
[ in]Theorem 2.1. We use the relation (3.1) and Lemma 2.1.
λ
Note that −1 6 t 6 1 for x ∈ , λ , and we estimate below xqm (x) as follows
κ

2(−1)m δ m 2δ m
xqm (x) = 1 − R m+1 (t) > 1 − |Rm+1 (t)|
(δ − δ −1 )2 (δ − δ −1 )2
2δ m ( )
> 1− 2t + δ + δ −1
(δ − δ )
−1 2

2δ m ( )
> 1− 2 + δ + δ −1
(δ − δ )
−1 2

2δ m κ−1 2−ω
= 1− = 1 − δm > .
δ+δ −2
−1 2 2

In the last two steps we have used the identity (3.4)[ and ]the definition of δ, given in (2.3).
2−ω λ
We thus have shown that qm (x) > for all x ∈ , λ . Next, we apply Lemma 3.2 and
2x κ
we have that [ ]
κ+1 λ
qm (x) > q0 (x) = , for x ∈ 0, ,
2λ κ
which concludes the proof of the lower bound. ] [
λ
Upper bound: To prove the upper bound, we need to consider only the case x ∈ ,λ ,
[ ] κ
λ
because from Lemma 3.1 we already know that xqm (x) < 1 for x ∈ 0, . Now, for
[ ] κ
λ
x ∈ , λ , we apply an argument analogous to the one for the lower bound using the
κ
relation (3.1) and Lemma 2.1 (just changing “−” to “+”):

2(−1)m δ m κ−1 ω
xqm (x) = 1 − Rm+1 (t) 6 1 + δ m 61+ .
(δ − δ )
−1 2 2 2

Remark 3.1. Note that this lemma implies that the polynomial of best approximation is
positive on [0, λ] as long as (3.7) is satisfied with ω ∈ (0, 2).
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 457

To conclude this section, we discuss conditions relating κ and the degree of the polynomial
m so that (3.7) holds. In what follows, without loss of generality we assume that ln((κ −
1)/ω) > 1. In applications (particularly for analysis of convergence of two-level methods)
we are interested in large values of κ (resp. m). Since ω ∈ (0, 2), such condition is clearly
satisfied for κ > 2e + 1.
For fixed and sufficiently large κ, (as we assumed above), let m satisfy
√ √
κ+1 κ+1
ln[(κ − 1)/ω] 6 m 6 1 + ln[(κ − 1)/ω]. (3.9)
2 2
We will now show that the lower bound in (3.9) implies (3.7) (and therefore also the conclu-
sion of Lemma 3.4). Since 0 < δ < 1 we have

( )m [( )√κ+1 ] 12 ln[(κ−1)/ω]
2 2
δm = 1− √ 6 1− √
κ+1 κ+1

On the other hand, the function (1 − 2/ξ)ξ is increasing for all ξ > 2, and hence
[ ( )ξ ] 12 ln[(κ−1)/ω] ( )
2 κ−1 ω
m
δ < lim 1− = exp − ln = .
ξ→∞ ξ ω κ−1

Thus, if κ is given, the polynomial degree m for which (3.7) holds is bounded below by the
right hand side of (3.9).
In addition, it is easy to show that if (3.9) holds, then we also have
( )2 ( )2
1 ln m 1 ln(κ/2)
6 cω , with cω = sup . (3.10)
κ+1 m 2 κ>1;ω∈(0,2) 1 + ln(κ/ω)

Note that cω is finite. The inequality (3.10) is seen as follows. Since the logarithm is an
increasing function on its domain we get
√ 1
ln m > ln[ κ + 1)/2] + ln ln[(κ − 1)/ω] > ln(κ/2).
2
√
Also, from (3.9), since κ + 1 > 2 we have:
( √ )2 (√ )2
κ+1 κ+1
m 2
6 1+ ln[(κ − 1)/ω] 6 (1 + ln[(κ − 1)/ω]
2 2
1
6 (κ + 1)(1 + ln κ/ω)2 .
2
Combining the last two estimates then gives (3.10).

4. An application to two-level methods

We consider the linear system of equations

Au = f , (4.1)
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
458 Johannes Kraus et al.

where A ∈ IRN ×N is a symmetric and positive definite matrix, and f ∈ IRN is a given right
hand side vector. To describe a general two-level multiplicative method, we denote V = IRN ,
and also introduce a coarse space VH , VH ⊂ V , NH = dim VH , NH < N . In the following we
will always assume that VH = range(P ), where P : IRNH 7→ V and its matrix representation
in the canonical basis of IRNH is given by the coefficients in the expansion of the basis in
VH via the basis in V . Clearly, P is a full rank operator and its matrix representation is
oftentimes called prolongation or interpolation matrix. The restriction of A on the coarse
space is denoted by AH = P T AP .

4.1. Convergence rate estimates

In this subsection we prove convergence estimates for the classical multiplicative two-level
iteration, with polynomial smoother which is used to define a preconditioner B ≈ A−1 . In a
recent work [2] the properties of special polynomial smoothers have been exploited in order
to conduct an improved convergence analysis of smoothed aggregation algebraic multigrid
methods. Here, only for completeness, we include a two-level convergence result presented
in [3]. The only difference is that we use a polynomial smoother with polynomial defined
via Algorithm 2.2. As in [3] we show explicit dependence of the estimates on the degree of
the polynomial.
The results up to and including Theorem 4.3 hold for general SPD A, V and VH , provided
that the smoother is constructed using the polynomials of best approximation to 1/x on a
suitably chosen interval.
In this subsection, by ρ(X) we denote the spectral radius of a matrix X. If, in addition,
X is symmetric and positive definite matrix, we denote the X-norm by ∥v∥2X = v T Xv.
We define the two-grid (or TG) preconditioner using a classical two-level algorithm which
reads as follows.

Algorithm 4.1. Given w ∈ V which approximates the solution of (4.1) we define the next
approximation v ∈ V to u via the following two steps:

1. Coarse grid correction: y:=w + P A−1

H P (f − Aw)
T

2. Smoothing: v:=y + R(f − Ay).

We assume that R is symmetric and positive definite and A-norm convergent, namely

∥I − RA∥2A < 1. (4.2)

The error propagation operator for the two-level iteration above is

ETL = (I − RA)(I − πA ), πA = P A−1 T

H P A.

We then define the two-level preconditioner as:

∗
B = (I − ETL ETL )A−1 .
∗
Here ETL denotes the adjoint with respect to the inner product defined by A. Introducing
R̄ such that ( )
I − R̄A = (I − RA)2 and hence R̄ = 2R − RAR. (4.3)
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 459

it is straightforward then to compute that (see, e.g., [15]):

B = R̄ + (I − RA) P A−1
H P (I − AR) .
T
(4.4)

Recall a necessary and sufficient condition for R to be a convergent smoother in A-norm,

i.e., (4.2) to hold is that R̄ is SPD.
Our goal will be to prove a convergence rate estimate for the two-level method with
polynomial smoother. First, let us denote with D the diagonal of A and set

R = qm (D−1 A)D−1 ,

where qm (x) is the polynomial of best approximation to 1/x on a fixed interval [λ/κ, λ].
Both λ and κ are to be specified later.
One may also write R in the form
b −1/2 ,
R = D−1/2 qm (A)D b = D−1/2 AD−1/2 .
A (4.5)

Using the notation from Section 3, we set λ = ∥A∥ b ℓ∞ . In what follows, we hold λ fixed and
we vary κ and the degree of the polynomial m. However, κ and m do not vary independently
and we assume that κ and m satisfy the condition (3.7). With such choice of λ, κ and m,
one can easily show that R is a contraction (a convergent smoother) in A-norm and we do so
by showing that R̄ is SPD, which, as we mentioned earlier, is both necessary and sufficient
condition for (4.2) to hold. Clearly, R̄ can be written (see (4.3)) as
b − qm
R̄ = D−1/2 [2qm (A) 2 b b
(A)A]D−1/2 .

From the upper bound in Lemma 3.4 we immediately get that for all x ∈ (0, λ] we have
2+ω
xqm (x) 6 . Therefore, for all w ∈ V we get
2
b − [qm (A)]
wT (2qm (A) b 2 A)w
b > (2 − ∥xq(x)∥∞,(0,λ) )wT qm (A)w
b
2−ω T b
> w qm (A)w.
2
Applying the inequality above with w = D−1/2 y then shows that for all y ∈ V
2−ω T 2−ω
y T R̄y > y Ry > min qm (x) y T D−1 y. (4.6)
2 2 x∈(0,λ]

From the lower bound in Lemma 3.4, we conclude that R̄ is SPD.

We further note that each of the off-diagonal entries of (D−1/2 AD−1/2 ) is less than 1 and
the diagonal entry is equal to 1. Therefore, we have that

1 6 ∥D−1/2 AD−1/2 ∥ = ρ(D−1/2 AD−1/2 ) 6 ∥D−1/2 AD−1/2 ∥ℓ∞ = λ 6 nz , (4.7)

where nz is the maximal number of non-zeros in a row of A.

The convergence rate estimates are derived from the following theorem (two-level version
of the XZ-identity, cf. [5, 15]).
Theorem 4.2. Assume that R̄ is SPD. Then the following identity holds:

v T B −1 v = inf [∥vH ∥2A + ∥v − vH ∥2R̄−1 ]. (4.8)

vH ∈VH

Brought to you by | University of Iowa Libraries

Authenticated
Download Date | 5/30/15 11:54 AM
460 Johannes Kraus et al.

Based on Theorem 4.2, we now state and prove a convergence result involving the poly-
nomial smoother.
Theorem 4.3. Let A be a symmetric positive definite matrix and D be its diagonal. Let
λ = ∥D−1/2 AD−1/2 ∥ℓ∞ , and also κ > 1 and m satisfy (3.7). If R = qm (D−1 A)D−1 , with
qm (x) the polynomial of best approximation to 1/x on the interval [λ/κ, λ], then the following
estimate holds for all v ∈ V :
[ ]
T −1 4 λ 1
v B v6 inf ∥vH ∥A +
2
∥v − vH ∥D +
2
∥v − vH ∥A .
2
(4.9)
(2 − ω) vH ∈VH (κ + 1) 2−ω

Proof. First, we see that from (4.6) we have that

2−ω T 2
y T R̄y > y Ry and hence y T R̄−1 y 6 y T R−1 y. (4.10)
2 2−ω
Under the assumptions we made in the statement of the theorem we can apply Lemma 3.4,
and get that for all x ∈ (0, λ],
{ } ( )
1 λ x 2λ 2x
6 2 max , 6 + . (4.11)
qm (x) κ+1 2−ω κ+1 2−ω

Since Ab and qm (A)

b commute, and have the same set of orthonormal eigenvectors, we have
that for all w ∈ V we have

b −1 w 6 2λ ∥w∥2 + 2 ∥w∥2b .
wT [qm (A)]
κ+1 ℓ2
2−ω A

Taking y = D1/2 w in the inequality above and using the estimate given in (4.10)
[ ]
T −1 2 T −1 4 λ 1
y R̄ y 6 y R y6 ∥y∥D +
2
∥y∥A .
2
(4.12)
2−ω (2 − ω) (κ + 1) 2−ω

The proof is concluded by taking y = (v − vH ) and applying Theorem 4.2.

1
Without loss of generality, we set now ω = 1 and use that in equation (3.10) cω = c1 6 .
2
The estimate in the Theorem 4.3 takes the form.
Corollary 4.4. Under the assumptions of Theorem 4.3, with ω = 1 we have
[ ]
T −1 λ
v B v 6 4 inf ∥vH ∥A +
2
∥v − vH ∥D + ∥v − vH ∥A .
2 2
vH ∈VH (κ + 1)

In addition, if κ and m satisfy (3.9) we have

[ ]
−1 λ ln2 m
v B v 6 2 inf
T
2∥vH ∥A +
2
∥v − vH ∥D + 2∥v − vH ∥A .
2 2
vH ∈VH m2

To stress the fact that estimate (4.12) is purely algebraic, we formulate it separately, as
this is our main new result.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 461

Theorem 4.5. Let A be an s.p.d. matrix and D a given s.p.d. preconditioner for A such
that ∥D− 2 AD− 2 ∥ 6 λ. Consider the polynomial preconditioner
1 1

R = qm (D−1 A)D−1 ,
[ ]
λ
where qm is the polynomial of best approximation of 1/x over the interval , λ . The
κ
parameter κ is chosen depending on m such that (3.7) holds for a given ω ∈ (0, 2). Then
the following smoothing property holds for R and its symmetrized version R (see (4.3)):

2 − ω T −1 2λ 2
v R v 6 v T R−1 v 6 v T Dv + v T Av.
2 κ+1 2−ω

In addition, if κ and m satisfy (3.9), we have

2 − ω T −1 ln2 m T 2
v R v 6 v T R−1 v 6 2λ v Dv + v T Av.
2 m 2 2−ω

Remark 4.1. Theorem 4.5 provides an estimate which is different from the classical def-
inition of smoothing properties of semi-iteration methods found in [7, Section 6.2.5] and
[8, Section 10.8.1], see also [13, Section 4.1, and Section 4.3.2]. The above estimate com-
bined with a weak approximation property of the coarse space implies two-grid convergence,
whereas the classical smoothing property of Hackbusch combined with a strong approxi-
mation property of the Galerkin coarse-grid projection implies two-grid convergence that
improves with the number of smoothing steps. We note also that the parameter κ can be
used to control which part of the spectrum of A is resolved by the smoothing iteration. This
is so, because the properties of the sequence of polynomials given in Section 3 show that
the smoothing iteration
[ ] is convergent as long as the approximating polynomial is positive
λ
on the interval , λ . In another words, we can “approximate” well on a relatively small
κ
interval and this correspond to quick damping of the eigenmodes of A corresponding to large
eigenvalues (the ones close to λ), that is we have a good smoothing property. Of course,
other polynomial sequences can be used to provide such smoothing iterations. We refer to
Hackbusch [7, Section 6.2.5] and also to a recent work by Brezina and Vassilevski [3] for
further discussions. In the next section we show how κ, or, equivalently, the polynomial
degree, can be chosen in conjunction with the approximation properties of coarse spaces to
yield a uniformly convergent two-grid method for finite element discretization of a PDE.

4.2. Two-level method for discretized PDE

In this section we apply the abstract two–level result to the case of a two-level iterative
method with large coarsening ratio for the solution of a system of linear algebraic equations
arising from a discretization of scalar elliptic equation with heterogeneous coefficients sim-
ilarly to the presentation in [3], now for the case of a different polynomial smoother from
Theorem 4.5. We consider the following variational problem: Find u ∈ HD 1
(Ω), for a given
polygonal (polyhedral) domain Ω ⊂ IRd (d = 2 or 3) and a source term f ∈ L2 (Ω), such that
∫ ∫
a(u, v) ≡ α(x) ∇u · ∇v = f (x)v(x) = (f, v) , for all v ∈ HD 1
(Ω). (4.13)
Ω Ω
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
462 Johannes Kraus et al.

Here, Ω ⊂ IRd d = 2, 3 is a given domain whose boundary Γ = ∂Ω is partitioned as Γ =

ΓD ∪ ΓN . We assume that ΓD ̸= ∅ is closed as a subset of Γ and also has a nonzero (d − 1)
dimensional measure. We refer to ΓD as the Dirichlet part of the boundary and ΓN as the
1
Neumann part of the boundary. In the variational problem (4.13), HD (Ω) denotes the space
1
of functions in H (Ω) whose traces vanish on ΓD .
We are interested in the case when the diffusion coefficient α = α(x) is a piecewise
constant function, that may have large variations within Ω. We thus assume that Ω̄ =
∪ml=1 Ω̄l , with polygonal (polyhedral) subdomains Ωl , and that α(x) = αl , for all x ∈ Ωl and
0

l = 1, . . . , m0 . We introduce the following energy norm

∫ ∑
m0 ∫
∥v∥a =
2 2
α(x)|∇v| = αl |∇v|2 . (4.14)
Ω l=1 Ωl

We also need the weighted L2 norm

∫ ∑
m0 ∫
∥v∥20,α = 2
α(x)v = αl v2. (4.15)
Ω l=1 Ωl

We consider a standard discretization of the variational problem (4.13) with piecewise linear
continuous finite elements. To define the finite element spaces and the approximate solution,
we assume that we have a locally quasi–uniform, simplicial triangulation Th of Ω. We assume
that this triangulation also resolves Ωl , namely, for l = 1, . . . , m0 we have:

Ω̄ = ∪τ ∈Th τ, Ω̄l = ∪τ ∈TΩl τ, (4.16)

where TΩl ⊂ Th , for l = 1, . . . , m0 . The standard space of piecewise linear (w.r.t Th ) and
continuous functions vanishing on the boundary of Ω is denoted by Vh .
The discrete problem then reads: Find u ∈ Vh such that

a(u, v) = (f, v), for all v ∈ Vh . (4.17)

The notation and constructions in the previous section are suitable for the finite element
setting as well. Indeed, a coarse space corresponding to VH (denoted here with VH ) as
VH = range(P ), with the same P as before, but this time representing the coefficients in
the expansion of the basis in VH , {φH
j }j=1 via the canonical Lagrange basis {φj }j=1 in Vh .
NH N

Evaluating the bilinear form on the basis for Vh and the basis for VH defines the stiffness
matrix A and the matrix AH :

Akj = a(φj , φj ), (AH )kj = (P T AP )jk = a(φH H

j , φj ).

According to the considerations in the previous section, we use bold face to represent vectors
of degrees of freedom and normal font for functions. Thus a function v ∈ Vh is represented
by the vector v ∈ V .
We make the following assumption for the stability and approximation properties of the
coarse function space VH .
Approximation and stability assumption: For any v ∈ Vh there exists vH ∈ VH such that

H −2 ∥v − vH ∥20,α + ∥v − vH ∥2a 6 cas ∥v∥2a , (4.18)

where H is the diameter of the support of a typical basis function in VH , and the constant
cas is independent of the variations of the coefficient α(x).
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 463

Construction of coarse spaces satisfying this assumption is possible as already mentioned,

and we refer to [6], [12], and earlier [1] as modified recently in [3] for such constructions.
We next introduce a well-known inequality relating the weighted L2 norm on the function
space Vh and the norm provided by the diagonal of the stiffness matrix on the space of degrees
of freedom (nodal values of the piece-wise linear functions). Let {λj,T }d+1
j=1 be the barycentric
coordinates in an element T ∈ Th and αT be the value of the coefficient on T (recall that
α(x) is piece-wise constant). Let v ∈ Vh with corresponding vector of degrees of freedom
v ∈ V . We have the following simple inequality
∑ ∑
d+1 ∑ ∑
d+1
∥v∥2D = αT 2
vj,T |∇λj,T |2 6 cT h−2
T αT
2
vj,T |λj,T |2
T ∈Th j=1 T ∈Th j=1
∑
6 h−2
T αT cT cM,T ∥v∥L2 (T ) .
2

T ∈Th

In the inequalities above, we have used standard inverse inequality, and also that the local
mass matrix for an element T is equivalent to its diagonal with a bound cM,T independent
of the coefficient variation. Finally,
∥v∥2D 6 cM h−2 ∥v∥20,α , with cM = max cT cM,T . (4.19)
T ∈Th

It is also clear that ∥v∥A = ∥v∥a by the definition of the stiffness matrix.
We now formulate the spectral equivalence result for the two-level method when applied
to the discretized PDE (4.13).
Theorem 4.6. Let κ > 1, and m be such that (3.7) holds with ω = 1. Assume that VH
is such that the approximation and stability assumption holds. Then we have the following
spectral equivalence (for nz see (4.7)):
[ ( )2 ]
4c n H
v T Av 6 v T B −1 v 6 KT G v T Av, KT G = 8 + cas
M z
+3 . (4.20)
(κ + 1) h

Moreover, if κ (or equivalently the degree of the polynomial m) is sufficiently large the spectral
equivalence is uniform with respect to mesh parameters and coefficient variation.
Proof. The proof is the same as the one given in [3] however with a different smoothing
property provided by Theorem 4.5.
The lower bound is immediate, since ETL is a contraction in the A-norm. The upper
bound follows directly from Corollary 4.4 together used in conjunction with the simple
inequalities relating the function space Vh and V (see (4.19)). Given v ∈ Vh , let v ∈ V be
the corresponding vector of degrees of freedom. We have
[ ]
T −1 λ
v B v 6 4 inf ∥vH ∥A +
2
∥v − vH ∥D + ∥v − vH ∥A
2 2
vH ∈VH (κ + 1)
[ ]
cM nz h−2
6 4 inf ∥vH ∥A +
2
∥v − vH ∥0,α + ∥v − vH ∥a
2 2
vH ∈VH (κ + 1)
[ ]
cM nz h−2
6 4 inf 2
2∥v∥a + ∥v − vH ∥0,α + 3∥v − vH ∥a
2 2
vH ∈VH (κ + 1)
[ ( )2 ]
cas cM nz H
6 4 2+ + 3cas ∥v∥2a = KT G v T Av.
(κ + 1) h
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
464 Johannes Kraus et al.

√ H H
Clearly, for ( κ+1) > , and m satisfying (3.9), for example, m > ln(H/h), the spectral
h h
equivalence is uniform with respect to mesh size and coefficient variation.

5. Choice of coarse spaces and numerical tests

In this section, we present a number of tests that illustrate the robustness of the two–level
methods with the polynomial smoother analyzed in the present paper all in accordance
with Theorem 4.6. We consider the second order elliptic equation (4.13) with a mixture of
Neumann and Dirichlet boundary conditions. The Dirichlet boundary conditions are imposed
on the “east” and “west” vertical boundaries, i.e., ΓD = ΓE ∪ ΓW of Ω. As we pointed out,
the coefficient α(x) is piecewise constant and we assume that the fine triangulation of Ω is
aligned with (resolves) all the coefficient discontinuities. In Fig. 5.1 we show an example of
a fine grid Th , aligned with discontinuities.

Figure 5.1. Checkerboard coefficient distribution on a mesh with 25600 elements and 13041 vertices.

5.1. Coarse spaces

We use element agglomeration to define “coarse elements” as illustrated in Fig. 5.2. and a
variant of the spectral AMGe method (see, e.g., [15]) in the form presented in [3]. Briefly
the main steps in such coarse space construction are:

• Partitioning of the degrees of freedom as a union of non-overlapping sets, {A} called

aggregates. This is achieved by first partitioning the set of elements into agglomerated
elements {τ } (union of fine-grid elements). We use graph partitioner (metis) applied
to the graph having vertices the fine-grid elements with edges between two elements if
they share a common interface. Then, we form aggregates A, where each aggregate (a
set of fine degrees of freedom) corresponds a unique agglomerated element τ = τA by
distributing the shared fine degrees of freedom (fine-grid element vertices belonging to
two or more agglomerated elements) to a unique aggregate.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 465

(a) (b)

Figure 5.2. (a) An example of 10 aggregates; (b) the corresponding element agglomerates (unions of fine
grid elements) constructed on a mesh with 6400 elements and 3321 vertices. The distribution of discontinuity
of the PDE coefficient is not resolved by the agglomerates.

• Constructing a tentative interpolation matrix P , defined for an agglomerate τ . Con-

sider the local generalized eigenproblem,

Aτ φj = θj Dτ φj ,

where Aτ is the local stiffness matrix corresponding to the agglomerated element τ ,

Dτ is its diagonal and θ1 6 θ2 6 . . . 6 θnA with nA = |A| (cardinality of A). Given a
spectral tolerance θ, we select the eigenvectors {φj }nj=1
θ
, where nθ is the largest integer
for which the inequality θnθ < θ holds. Extended by zero outside each A, the vectors
{φj }nj=1
θ
form nθ columns of the global tentative interpolation operator P .

• Constructing the coarse space as the range of the interpolation matrix P , which is
defined as ( )
P = sm λ−1 D−1 A P .

Here, as in the previous section, D is the diagonal of A, λ > ∥D− 2 AD− 2 ∥ (e.g.,
1 1

λ = ∥D−1/2 AD−1/2 ∥∞ ), and sm (t) is the smoothed aggregation (SA) polynomial (cf.,
e.g., [2]) √
(−1)m T2m+1 ( t)
sm (t) = √ .
(2m + 1) t

5.2. Numerical tests

We recall some of the notations and definitions which are used in the tables and figures in
this section.

• N is the number of fine grid degrees of freedom;

• NH is the number of coarse degrees of freedom;

• nnz(X) is the number of the nonzero elements in a matrix X;

Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
466 Johannes Kraus et al.

• ϱeT G is the asymptotic convergence factor of the two grid method;

• oc(B) is the operator complexity measure of the two-grid preconditioner B, defined as

nnz(A) + nnz(AH )
oc(B) = .
nnz(A)

The first set of experiments are on a mesh with 102, 400 elements and N = 51, 681
vertices using 300 agglomerated elements (AEs). We stop the iterations when the relative
preconditioned residual norm is reduced by a factor of ε = 10−8 . The piecewise constant
coefficient α(x) is distributed in a checkerboard fashion with the values 1 (blue) and 106
(red) as illustrated in Fig. 5.1.
1
The experiments are performed for m = 2, 4, 6, 8, a ≡ = 0.158, 0.085, 0.055, 0.04, and
κ
a = 0.2, 0.1, 0.08, 0.06 respectively. They are chosen such that the inequality (3.7) (with
ω = 1) holds:
(√ )m ( √ )m
κ−1 1− a a 1 1
√ = √ < = , a= .
κ+1 1+ a 1−a κ−1 κ

The same degree m is used for the polynomial smoother in the two-level algorithm
( −1 )and the
smoother of the tentative interpolation matrix (it is smoothed out by sm D A ). The
number of non-zero entries of A, is nnz(A) = 359, 841. We also show how the spectral toler-
ance θ and the polynomial degree m influences the convergence versus operator complexity.
The results are presented in Tables 5.1–5.4. It is evident from the results that the method
can become fairly fast (in terms of convergence factors) at the expense of large operator
complexity.

ρeT G
θ NH nnz(AH ) oc(B) a = 0.158 a = 0.2

0.010 774 15,930 1.04 0.995 0.995

0.077 3,629 342,515 1.95 0.879 0.889
0.149 6,557 1,115,207 4.10 0.393 0.492

Table 5.1. Two-grid convergence, m = 2.

ρeT G
θ NH nnz(AH ) oc(B) a = 0.085 a = 0.1

0.010 774 22,092 1.06 0.985 0.986

0.077 3,629 472,907 2.31 0.531 0.538
0.149 6,557 1,538,845 5.28 0.188 0.084

Table 5.2. Two-grid convergence when m = 4.

Brought to you by | University of Iowa Libraries

Authenticated
Download Date | 5/30/15 11:54 AM
Polynomial of best uniform approximation to 1/x and two-level methods 467

ρeT G
θ NH nnz(AH ) oc(B) a = 0.055 a = 0.08

0.010 774 29,448 1.08 0.965 0.969

0.077 3,629 636,671 2.78 0.205 0.179
0.149 6,557 2,074,291 6.76 0.202 0.026

Table 5.3. Two-grid convergence when m = 6.

ρeT G
θ NH nnz(AH ) oc(B) a = 0.04 a = 0.06

0.010 774 37,618 1.10 0.926 0.933

0.077 3,629 808,357 3.25 0.197 0.111
0.149 6,557 2,632,755 8.32 0.193 0.028

Table 5.4. Two-grid convergence when m = 8.

In the last experiment shown in Table 5.5, we illustrate the behavior of the method
with respect to varying the contrast 10c again distributed in a checkerboard fashion. As it
is clearly seen, the two-grid method exhibits very good uniform two-grid convergence with
operator complexity less than two.

c -12 -9 -6 -3 0 3 6 9 12
NH 2336 2336 2336 2339 2322 2322 2322 2322 2322
oc(B) 1.94 1.94 1.94 1.94 1.93 1.93 1.93 1.93 1.93
nit 17 17 17 17 17 16 16 16 16
ρeT G 0.219 0.219 0.219 0.219 0.219 0.200 0.198 0.197 0.198

Table 5.5. Contrast independent two-grid convergence; coefficient jumps are 10c . The method corresponds
to spectral threshold θ = 0.045, m = 8, and a = 0.04.

Acknowledgments
The work of the first author has been supported by the Austrian Science Fund, Grant
P22989-N18. The work of the second author is performed under the auspices of the U.S.
Department of Energy by Lawrence Livermore National Laboratory under Contract DE-
AC52-07NA27344. The work of the third author is supported in part by the National Science
Foundation DMS-0810982, U.S. Department of Energy grant DE-SC0006903 and Lawrence
Livermore National Laboratory subcontract B595949.
The authors thank Delyan Kalchev from LLNL for his help with the numerical experi-
ments. We also thank Roman Andreev from ETH, Zurich, for pointing reference [9].
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 5/30/15 11:54 AM
468 Johannes Kraus et al.

References
[1] M. Brezina, C. Heberton, J. Mandel, and P. Vaněk, An iterative method with convergence rate chosen a
priori , Tech. Rep. 140, University of Colorado Denver, CCM, University of Colorado Denver, 1999, an
earlier version presented at the 1998 Copper Mountain Conference on Iterative Methods, April 1998.
[2] M. Brezina, P. Vaněk, and P. S. Vassilevski, An improved convergence analysis of smoothed aggregation
algebraic multigrid , Numerical Linear Algebra with Applications, 19 (2012), no. 3, pp. 441–469.
[3] M. Brezina and P. S. Vassilevski, Smoothed aggregation spectral element agglomeration AMG: SA-
ρAMGe, in: Large-scale scientific computing (I. Lirkov, S. Margenov, and J. Wasniewski, eds.), vol.
7116 of Lecture Notes in Comput. Sci., Springer, Berlin, 2012, pp. 3–15, also available as LLNL technical
report LLNL-PROC-490083.
[4] P. L. Chebyshev, Sur les polynômes réprésentant le mieux les valeurs des fonctions fractionnaires
élémentaires pour les valeurs de la variable contenues entre deu limites données, in: Oeuvres, vol. II,
St Petersburg, 1907, pp. 669–678.
[5] R. D. Falgout, P. S. Vassilevski, and L. T. Zikatanov, On two-grid convergence estimates, Num. Lin.
Alg. Appl., 12 (2005), pp. 471–494, also available as LLNL technical report UCRL-JC-150807.
[6] J. Galvis and Y. Efendiev, Domain decomposition preconditioners for multiscale flows in high-contrast
media, Multiscale Modeling & Simulation, 8 (2010), no. 4, pp. 1461–1483.
[7] W. Hackbusch, Multigrid methods and applications, vol. 4 of Springer Series in Computational Mathe-
matics, Springer-Verlag, Berlin, 1985.
[8] W. Hackbusch, Iterative solution of large sparse systems of equations, vol. 95 of Applied Mathematical
Sciences, Springer-Verlag, New York, 1994, translated and revised from the 1991 German original.
[9] V. N. Kutrunov, A polynomial of best uniform approximation in an iterative method for solving systems
of linear algebraic equations, Sibirsk. Mat. Zh., 33 (1992), no. 1, pp. 62–68, 219.
[10] G. Meinardus, Approximation of functions: Theory and numerical methods, Expanded translation of
the German edition. Translated by Larry L. Schumaker. Springer Tracts in Natural Philosophy, Vol.
13, Springer-Verlag New York, Inc., New York, 1967.
[11] T. J. Rivlin, An introduction to the approximation of functions, Dover Publications Inc., New York,
1981, corrected reprint of the 1969 original, Dover Books on Advanced Mathematics.
[12] R. Scheichl, P. S. Vassilevski, and L. T. Zikatanov, Weak approximation properties of elliptic projections
with functional constraints, Multiscale Modeling & Simulation, 9 (2011), no. 4, pp. 1677–1699.
[13] V. V. Shaidurov, Multigrid methods for finite elements, vol. 318 of Mathematics and Its Applications,
Kluwer Academic Publishers, Dordrecht, 1995.
[14] R. S. Varga, Matrix Iterative Analysis, Prentice-Hall Series in Automatic Computation, Prentice-Hall,
Inc., Englewood Cliffs, New Jersey, 1962.
[15] P. S. Vassilevski, Multilevel block factorization preconditioners. Matrix-based analysis and algorithms
for solving finite element equations, Springer, New York, 2008.

Brought to you by | University of Iowa Libraries

Authenticated
Download Date | 5/30/15 11:54 AM