Anda di halaman 1dari 27

# a

r
X
i
v
:
0
8
0
8
.
0
1
8
0
v
2

[
m
a
t
h
.
N
A
]

1
5

A
u
g

2
0
0
8
CUBATURE FORMULA AND INTERPOLATION ON THE
CUBIC DOMAIN
HUIYUAN LI, JIACHANG SUN, AND YUAN XU
Abstract. Several cubature formulas on the cubic domains are derived using
the discrete Fourier analysis associated with lattice tiling, as developed in .
The main results consist of a new derivation of the Gaussian type cubature for
the product Chebyshev weight functions and associated interpolation polyno-
mials on [1, 1]
2
, as well as new results on [1, 1]
3
. In particular, compact
formulas for the fundamental interpolation polynomials are derived, based on
n
3
/4 +O(n
2
) nodes of a cubature formula on [1, 1]
3
.
1. Introduction
For a given weight function W supported on a set R
d
, a cubature formula of
degree 2n 1 is a nite sum, L
n
f, that provides an approximation to the integral
and preserves polynomials of degree up to 2n 1; that is,
_

f(x)W(x)dx =
N

k=1

k
f(x
k
) =: L
n
f for all f
d
2n1
,
where
d
M
denotes the space of polynomials of total degree at most n in d variables.
The points x
k
R
d
are called nodes and the numbers
k
R0 are called weights
of the cubature.
Our primary interests are Gaussian type cubature, which has minimal or nearer
minimal number of nodes. For d = 1, it is well known that Gaussian quadrature of
degree 2n 1 needs merely N = n nodes and these nodes are precisely the zeros
of the orthogonal polynomial of degree n with respect to W. The situation for
d 1, however, is much more complicated and not well understood in general.
As in the case of d = 1, it is known that a cubature of degree 2n 1 needs at
least N dim
d
n1
number of nodes, but few formulas are known to attain this
lower bound (see, for example, [1, 10]). In fact, for the centrally symmetric weight
function (symmetric with respect to the origin), it is known that the number of
nodes, N, of a cubature of degree 2n1 in two dimension satises the lower bound
(1.1) N dim
2
n1
+
_
n
2
_
,
known as M ollers lower bound . It is also known that the nodes of a cubature
that attains the lower bound (1.1), if it exists, are necessarily the common zeros
Date: August 15, 2008.
1991 Mathematics Subject Classication. 41A05, 41A10.
Key words and phrases. lattice, cubature, interpolation, discrete Fourier series.
The rst and the second authors were supported by NSFC Grant 10431050 and 60573023. The
second author was supported by National Basic Research Program grant 2005CB321702. The
third author was supported by NSF Grant DMS-0604056.
1
2 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
of n + 1
n
2
orthogonal polynomials of degree n with respect to W. Similar
statements on the nodes hold for cubature formulas that have number of nodes
slightly above M ollers lower bound, which we shall call cubature of Gaussian type.
These denitions also hold in d-dimension, where the lower bound for the number
of nodes for the centrally symmetric weight function is given in .
There are, however, only a few examples of such formulas that are explicitly
constructed and fewer still can be useful for practical computation. The best known
example is = [1, 1]
d
with the weight function
(1.2) W
0
(x) :=
d

i=1
1
_
1 x
2
i
or W
1
(x) :=
d

i=1
_
1 x
2
i
and only when d = 2. In this case, several families of Gaussian type cubature are
explicitly known, they were constructed ([13, 17]) by studying the common zeros of
corresponding orthogonal polynomials, which are product Chebyshev polynomials
of the rst kind and the second kind, respectively. Furthermore, interpolation
polynomials bases on the nodes of these cubature formulas turn out to possess
several desirable features (, and also ). On the other hand, studying common
zeros of orthogonal polynomials of several variables is in general notoriously dicult.
In the case of (1.2), the product Chebyshev polynomials have the simplest structure
among all orthogonal polynomials, which permits us to study their common zeros
and construct cubature formulas in the case d = 2, but not yet for the case d = 3
or higher.
The purpose of the present paper is to provide a completely dierent method
for constructing cubature formulas with respect to W
0
and W
1
. It uses the discrete
Fourier analysis associated with lattice tiling, developed recently in . This
method has been used in  to establish cubature for trigonometric functions on
the regular hexagon and triangle in R
2
, a topic that has been studied in [15, 16],
and on the rhombic dodecahedron and tetrahedron of R
3
in . The cubature on
the hexagon can be transformed, by symmetry, to a cubature on the equilateral
triangle that generates the hexagon by reection, which can in turn be further
transformed, by a nontrivial change of variables, to Gaussian cubature formula
for algebraic polynomials on the domain bounded by Steiners hypercycloid. The
theory developed in  uses two lattices, one determines the domain of integral and
the points that dened the discrete inner product, the other determines the space of
exponentials or trigonometric functions that are integrated exactly by the cubature.
In [10, 9] the two lattices are taken to be the same. In this paper we shall choose
one as Z
d
itself, so that the integral domain is xed as the cube, while we choose
the other one dierently. In d = 2, we choose the second lattice so that its spectral
set is a rhombus, which allows us to establish cubature formulas for trigonometric
functions that are equivalent to Gaussian type cubature formulas for W
0
and W
1
.
In the case of d = 3, we choose the rhombic dodecahedron as a tiling set and obtain
a cubature of degree 2n1 that uses n
3
/4 +O(n
2
) nodes, worse than the expected
lower bound of n
3
/6 +O(n
2
) but far better than the product Gaussian cubature of
n
3
nodes. This cubature with n
3
/4+O(n
2
) nodes has appeared recently and tested
numerically in . We will further study the Lagrange interpolation based on its
nodes, for which the rst task is to identify the subspace that the interpolation
polynomials belongs. We will not only identify the interpolation space, but also
give the compact formulas for the fundamental interpolation polynomials.
3
One immediate question arising from this study is if there exist cubature formulas
of degree 2n 1 with n
3
/6 + O(n
2
) nodes on the cube. Although examples of
cubature formulas of degree 2n 1 with N = dim
d
n1
= n
d
/d! + O(n
d1
) nodes
are known to exist for special non-centrally symmetric regions (), we are not aware
of any examples for symmetric domains that use N = n
d
/d! +O(n
d1
) nodes. From
our approach of tiling and discrete Fourier analysis, it appears that the rhombic
dodecahedron gives the smallest number of nodes among all other fundamental
domains that tile R
3
by translation. Giving the fact that this approach yields the
cubature formulas with optimal order for the number of nodes, it is tempting to
make the conjecture that a cubature formula of degree 2n 1 on [1, 1]
3
needs at
least n
3
/4 +O(n
2
) nodes.
The paper is organized as follows. In the following section we recall the result
on discrete Fourier analysis and lattice tiling in . Cubature and interpolation
for d = 2 are developed in Section 3 and those for d = 3 are discussed in Section 4,
both the latter two sections are divided into several subsections.
2. Discrete Fourier Analysis with lattice Tiling
We recall basic results in  on the discrete Fourier analysis associated with a
lattice. A lattice of R
d
is a discrete subgroup that can be written as AZ
d
= Ak :
k Z
d
, where A is a d d invertible matrix, called the generator of the lattice. A
bounded set
A
R
d
is said to tile R
d
with the lattice AZ
d
if

kZ
d

A
(x +Ak) = 1 for almost all x R
d
,
where
E
denotes the characteristic function of the set E. The simplest lattice is
Z
d
itself, for which the set that tiles R
d
is
:= [
1
2
,
1
2
)
d
.
We reserve the notation as above throughout the rest of this paper. The set
is chosen as half open so that its translations by Z
d
tile R
d
without overlapping. It
is well known that the exponential functions
e
k
(x) := e
2ikx
, k Z
d
, x R
d
,
form an orthonormal basis for L
2
(). These functions are periodic with respect to
Z
d
; that is, they satisfy
f(x +k) = f(x) for all k Z
d
.
Let B be a d d matrix such that all entries of B are integers. Denote
(2.1)
B
=
_
k Z
d
: B
tr
k
_
and

B
=
_
k Z
d
: k
B
_
.
It is known that [
B
[ = [

B
[ = [ det B[, where [E[ denotes the cardinality of the
set E. We need the following theorem [10, Theorem 2.5].
Theorem 2.1. Let B be a d d matrix with integer entries. Dene the discrete
inner product
f, g)
B
:=
1
[ det(B)[

jB
f(B
tr
j)g(B
tr
j)
4 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
for f, g C(), the space of continuous functions on . Then
f, g)
B
= f, g) :=
_

f(x)g(x)dx, (2.2)
for all f, g in the nite dimensional subspace
T
B
:= span
_
e
2i kx
: k

B
_
.
The dimension of T
B
is [

B
[ = [ det B[.
This result is a special case of a general result in , in which is replaced by

A
for an invertible matrix A, and the set
B
is replaced by
N
with N = B
tr
A
and N is assumed to have integer entries. Since we are interested only at the cube
[
1
2
,
1
2
]
d
in this paper, we have chosen A as the identity matrix.
We can also use the discrete Fourier analysis to study interpolation based on the
points in
B
. We say two points x, y R
d
congruent with respect to the lattice
BZ
d
, if x y BZ
d
, and we write x y mod B. We then have the following
result:
Theorem 2.2. For a generic function f dened in C(), the unique interpolation
function 1
B
f in T
B
that satises
1
B
f(B
tr
j) = f(B
tr
j), j
B
is given by
1
B
f(x) =

B
f, e
k
)e
k
(x) =

kB
f(B
tr
k)
B
(x B
tr
k), (2.3)
where
(2.4)
B
(x) =
1
[ det(B)[

j
B

e
2ij
tr
x
.
The proof of this result is based on the second one of the following two relations
that are of independent interests:
(2.5)
1
[ det(B)[

jB
e
2ik
tr
B
tr
j
=
_
1, if k 0 mod B,
0, otherwise,
and
(2.6)
1
[ det(B)[

B
e
2ik
tr
B
tr
j
=
_
1, if j 0 mod B
tr
,
0, otherwise.
For proofs and further results we refer to [10, 9]. Throughout this paper we will
write, for k Z
d
, 2k = (2k
1
, . . . , 2k
d
) and 2k + 1 = (2k
1
+ 1, . . . , 2k
d
+ 1).
3. Cubature and Interpolation on the square
In this section we consider the case d = 2. In the rst subsection, the general
results in the previous section is specialized to a special case and cubature formulas
are derived for a class of trigonometric functions. These results are converted to
results for algebraic polynomials in the second subsection. Results on polynomial
interpolation are derived in the third subsection.
5
3.1. Discrete Fourier analysis and cubature formulas on the plane. We
choose the matrix B as
B = n
_
1 1
1 1
_
and B
1
=
1
2n
_
1 1
1 1
_
.
Since B is a rotation, by 45 degree, of a constant multiple of the diagonal matrix,
it is easy to see that the domain
B
is dened by

B
= x R
2
: n x
1
+ x
2
< n, n x
2
x
1
< n,
which is depicted in Figure 1 below.
n
n
n
n
Figure 1. Rhombus
B
From the expression of B
tr
B
=

B
=:
n
, where

n
= j Z
2
: n j
1
+j
2
< n, n j
2
j
1
< n.
The cardinality of
n
is [
n
[ = 2n
2
. We further denote the space T
B
by T
n
, which
is given by
T
n
:= span
_
e
2i kx
: k
n
_
.
Theorem 3.1. Dene the set
X
n
:=
_
2k :
n
2
k
1
, k
2
<
n
2
2k + 1 :
n+1
2
k
1
, k
2
<
n1
2
_
.
Then for all f, g T
n
,
f, g)
n
:=
1
2n
2

kXn
f(
k
2n
)g(
k
2n
) =
_
[
1
2
,
1
2
]
2
f(x)g(x)dx.
Proof. Changing variables from j to k = 2nB
tr
j, or k
1
= j
1
+j
2
and k
2
= j
2
j
1
,
then, as j
1
and j
2
need to be integers and j
1
=
k1k2
2
, j
2
=
k1+k2
2
, we see that
(3.1) j
n
k = 2nB
tr
j X
n
.
Hence, as det(B) = 2n
2
, we conclude that f, g)
n
= f, g)
B
and this theorem follows
as a special case of Theorem 2.1.
The set
n
lacks symmetry as the inequalities in its denition are half open and
half closed. We denote its symmetric counterpart by

n
, which is dened by

n
:= j Z
2
: n j
1
+j
2
n, n j
1
j
2
n.
6 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
We also denote the counterpart of T
n
by T

n
, which is dened by
T

n
:= span
_
e
2i kx
: k

n
_
.
Along the same line, we also dene the counterpart of X
n
as
X

n
:=
_
2k :
n
2
k
1
, k
2

n
2
2k + 1 :
n+1
2
k
1
, k
2

n1
2
_
.
It is easy to see that [X
n
[ = [
n
[ = 2n
2
, whereas [X

n
[ = 2n
2
+ 2n + 1. We further
partition the set X

n
into three parts,
X

n
= X

n
X
e
n
X
v
n
,
where X

n
= X

n
(n, n)
2
is the set of interior points of X

n
, X
e
n
consists of those
points in X

n
that are on the edges of [n, n]
2
but not on the 4 vertices or corners,
while X
v
n
consists of those points of X

n
at the vertices of [n, n]
2
.
Theorem 3.2. Dene the inner product
(3.2) f, g)

n
:=
1
2n
2

kX

n
c
(n)
k
f(
k
2n
)g(
k
2n
), where c
(n)
k
=
_

_
1, k X

n
1
2
, k X
e
n
1
4
, k X
v
n
.
Then for all f, g T
n
,
_
[
1
2
,
1
2
]
2
f(x)g(x)dx = f, g)
n
= f, g)

n
.
Proof. Evidently we only need to show that f, g)
n
= f, g)

n
. Since c
(n)
k
= 1 for
k X

n
, the partial sums over interior points of the two sums agree. The set
X
e
n
of boundary points can be divided into two parts, X
e
n
= X
e,1
n
X
e,2
n
, where
X
e,1
n
consists of points in X
n
that are on the edges of [n, n)
2
, but not equal to
(n, n), and X
e,2
n
is the complementary of X
e,1
n
in X
e
n
. Evidently, if x X
e,1
n
,
then either x + (2n, 0) or x + (0, 2n) belongs to X
e,2
n
. Hence, if f is a periodic
function, f(x +k) = f(x) for k Z
2
, then

kX
e
n
c
(n)
k
f(
k
2n
) =
1
2

kX
e
n
f(
k
2n
) =

kX
e,1
n
f(
k
2n
).
Furthermore, for (n, n) X
n
, X

n
contains all four vertices (n, n). Since
a periodic function takes the same value on all four points,

kX
v
n
c
(n)
k
f(
k
2n
) =
f(
1
2
,
1
2
). Consequently, we have proved that f, g)
n
= f, g)

n
if f, g are periodic
functions.
As a consequence of the above two theorems, we deduce the following two cuba-
ture formulas:
Theorem 3.3. For n 2, the cubature formulas
(3.3)
_
[
1
2
,
1
2
]
2
f(x)dx =
1
2n
2

kX

n
c
(n)
k
f(
k
2n
) and
_
[
1
2
,
1
2
]
2
f(x)dx =
1
2n
2

kXn
f(
k
2n
)
are exact for f T

2n1
.
7
Proof. It suces to proof that both cubature formulas in (3.3) are exact for every
e
j
with j

2n1
. For this purpose, we rst claim that for any j Z
2
, there exist

n
and l Z
2
such that j = + Bl. Indeed, the translations of
B
by BZ
2
tile R
2
, thus we have j = x +Bl for certain x
B
and l Z
2
. Since all entries of
the matrix B are integers, we further deduce that := x = j Bl Z
2

B
=
n
.
Next assume j

2n1
. Clearly the integral of e
j
over is
j,0
. On the other
hand, let us suppose j = +Bl with
n
and l Z
2
. Then it is easy to see that
e
j
(
k
2n
) = e

(
k
2n
) for each k X

n
. Consequently, we obtain from Theorem 3.2 that

kX

n
c
(n)
k
e
j
(
k
2n
) =

kX

n
c
(n)
k
e

(
k
2n
) =

kXn
e

(
k
2n
)
=

kXn
e
j
(
k
2n
) =
_

(x)dx =
,0
.
Since = 0 implies j = Bl Z
2
which gives j = l = 0, we further obtain that

,0
=
j,0
. This completes the proof of (3.3).
We note that the second cubature in (3.3) is a so-called Chebyshev cubature;
that is, all its weights are equal.
3.2. Cubature for algebraic polynomials. The set

n
is symmetric with re-
spect to the mappings (x
1
, x
2
) (x
1
, x
2
) and (x
1
, x
2
) (x
1
, x
2
). It follows
that both the spaces
T
even
n
: = spancos 2j
1
x
1
cos 2j
2
x
2
: 0 j
1
+j
2
n,
T
odd
n
: = spansin 2j
1
x
1
sin 2j
2
x
2
: 1 j
1
+j
2
n
are subspaces of T

n
. Recall that Chebyshev polynomials of the rst kind, T
n
(t),
and the second kind, U
n
(t), are dened, respectively, by
T
n
(t) = cos n and U
n
(t) =
sin(n + 1)
sin
, t = cos .
They are orthogonal with respect to w
0
(t) = 1/

1 t
2
and w
1
(t) =

1 t
2
over
[1, 1], respectively. Both are algebraic polynomials of degree n in t. Recall the
denition of W
0
and W
1
in (1.2). Under the changing of variables
(3.4) t
1
= cos 2x
1
, t
2
= cos 2x
2
, (x
1
, x
2
) [
1
2
,
1
2
]
2
,
the subspace T
even
n
becomes the space
2
n
of polynomials of degree n in the variables
(t
1
, t
2
),

2
n
= spanT
j
(t
1
)T
kj
(t
2
) : 0 j k n
and the orthogonality of e
k
over implies that T
k
j
(t) := T
j
(t
1
)T
kj
(t
2
) are orthog-
onal polynomials of two variables,
1

2
_
[1,1]
2
T
k
j
(t)T
k

j
(t)W
0
(t)dt =
_

_
1, k = k

= j = j

= 0,
1
2
, (k, j) = (k

, j

) and (k j)j = 0,
1
4
, k = k

> j = j

> 0,
0, (k, j) ,= (k

, j

).
We note also that the subspace T
odd
n
becomes the space
_
1 t
2
1
_
1 t
2
2
p(t) : p

2
n1
in the variables t = (t
1
, t
2
), and the orthogonality of e
k
also implies that
8 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
U
k
j
(t) := U
j
(t
1
)U
kj
(t
2
) are orthogonal polynomials of two variables,
1

2
_
[1,1]
2
U
k
j
(t)U
k

j
(t)W
1
(t)dt =
1
4

j,j

k,k
.
The symmetry allows us to translate the results in the previous subsection to
algebraic polynomials. Since cos 2j
1
x
1
cos 2j
2
x
2
are even in both variables, we
only need to consider their values over X

n
x : x
1
0, x
2
0. Hence, we dene
(3.5)
n
:= (2k
1
, 2k
2
) : 0 k
1
, k
2

n
2
(2k
1
+ 1, 2k
2
+ 1) : 0 k
1
, k
2

n1
2
,
and, under the change of variables (3.4),
(3.6)
n
:= (z
k1
, z
k2
) : (k
1
, k
2
)
n
, where z
k
= cos
k
n
.
Furthermore, we denote by

n
:=
n
(1, 1)
2
the subset of interior points of
n
,
by
e
n
the set of points in
n
that are on the boundary of [1, 1]
2
but not on the
four corners, and by
v
n
the set of points in
n
that are on the corners of [1, 1]
2
.
The sets

n
,
e
n
and
v
n
are dened accordingly. A simple counting shows that
(3.7) [
n
[ = (
n
2
+ 1)
2
+ (
n1
2
+ 1)
2
=
n(n + 1)
2
+
_
n
2
_
+ 1.
Theorem 3.4. The cubature formula
(3.8)
1

2
_
[1,1]
2
f(t)W
0
(t)dt =
1
2n
2

kn

(n)
k
f(z
k1
, z
k2
),
(n)
k
:=
_

_
4, k

n
,
2, k
e
n
,
1, k
v
n
,
is exact for
2
2n1
.
Proof. We note that X

n
is symmetric in the sense that k X

n
implies that
(k
1
, k
2
) X

n
and (k
1
, k
2
) X

n
. Let g(x) = f(cos 2x
1
, cos 2x
2
). Then g
is even in each of its variables and g(
k
2n
) = f(z
k1
, z
k2
). Notice that f
2
2n1
implies g T

2n1
. Applying the rst cubature formula (3.3) to g(x), we see that
(3.8) follows from the following identity,

kX

n
c
(n)
k
g(
k
2n
) =

kn

(n)
k
f(z
k1
, z
k2
).
To prove this identity, let k denote the set of distinct elements in (k
1
, k
2
);
then g(
k
2n
) takes the same value on all points in k. If k X

n
, k
1
,= 0 and k
2
,= 0,
then k contains 4 points;

jk
c
(n)
k
g(
j
2n
) = 4g(
k
2n
) if k X

n
,

jk
c
(n)
k
g(
j
2n
) =
2g(
k
2n
) if k X
e
n
, and

jk
c
(n)
k
g(
j
2n
) = g(
k
2n
) if k X
v
n
. If k
1
= 0 and k
2
,= 0 or
k
2
= 0 and k
1
,= 0, then k contains 2 points;

jk
c
(n)
k
g(
j
2n
) = 2g(
k
2n
) if k X

n
and

jk
c
(n)
k
g(
j
2n
) = g(
k
2n
) if k X
e
n
. Finally, if k = (0, 0) then k contains 1
point and g(0, 0) has coecient 1. Putting these together proves the identity.
By (3.7), the number of nodes of the cubature formula (3.8) is just one more than
the lower bound (1.1). We can also write (3.8) into a form that is more explicit.
9
Indeed, if n = 2m, then (3.8) can be written as
1

2
_
[1,1]
2
f(t)W
0
(t)dt (3.9)
=
2
n
2
m

i=0
m

j=0
f(z
2i
, z
2j
) +
2
n
2
m1

i=0
m1

j=0
f(z
2i+1
, z
2j+1
),
where

means that the rst and the last terms in the summation are halved. If
n = 2m+ 1, then (3.8) can be written as
1

2
_
[1,1]
2
f(t)W
0
(t)dt (3.10)
=
2
n
2
m

i=0
m

j=0
f(z
2i
, z
2j
) +
2
n
2
m

i=0
m

j=0
f(z
n2i
, z
n2j
),
where

means that the rst term in the sum is divided by 2. The formula (3.10)
appeared in , where it was constructed by considering the common zeros of
orthogonal polynomials of two variables.
From the cubature formula (3.3), we can also derive cubature formulas for the
Chebyshev weight W
1
of the second kind.
Theorem 3.5. The cubature formula
(3.11)
1

2
_
[1,1]
2
f(t)W
1
(t)dt =
2
n
2

n
sin
2 k1
n
sin
2 k2
n
f(z
k1
, z
k2
)
is exact for
2
2n5
.
Proof. We apply the rst cubature formula in (3.3) on the functions
sin(2(k
1
+ 1)x
1
) sin(2(k
2
+ 1)x
2
) sin2x
1
sin 2x
2
for 0 k
1
+ k
2
2n 5, where t
1
= cos 2x
1
and t
2
= cos 2x
2
as in (3.4).
Clearly these functions are even in both x
1
and x
2
and they are functions in T

2n1
.
Furthermore, they are zero when x
1
= 0 or x
2
= 0, or when (x
1
, x
2
) are on the
boundary of X

n
. Hence, the change of variables (3.4) shows that the rst cubature
in (3.3) becomes (3.11) for U
k1
(t
1
)U
k2
(t
2
).
A simple counting shows that [

n
[ =
n
2

2
+
n1
2

2
=
(n1)(n2)
2
+
n
2
. The
number of nodes of the cubature formula (3.11) is also one more than the lower
bound (1.1). In this case, this formula appeared already in .
3.3. Interpolation by polynomials. As shown in , there is a close relation
between interpolation and discrete Fourier transform. We start with a simple result
on interpolation by trigonometric functions in T
n
.
Proposition 3.6. For n 1 dene
(3.12) I
n
f(x) :=

kXn
f(
k
2n
)
n
(x
k
2n
),
n
(x) :=
1
2n
2

n
e

(x).
Then I
n
f(
k
2n
) = f(
k
2n
) for all k X
n
.
10 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
Proof. For j
n
dene k = 2nB
tr
j. From the relation (3.1), j
n
is equivalent
to k X
n
with k = 2nB
tr
j. As a result, we can write I
n
f(x) as
I
n
f(x) =

jn
f(B
tr
j)
n
(x B
tr
j)
and the interpolation means I
n
f(B
tr
j) = f(B
tr
j) for j
n
. For k, j
n
,

n
(B
tr
(j k)) =
1
2n
2

n
e

(B
tr
(j k)) =
k,j
by (2.6).
For our main result, we need a lemma on the symmetric set X

n
and

n
. Recall
that c
(n)
k
is dened for k X

n
. Since the relation (3.1) clearly extends to
(3.13) j

n
k = 2nB
tr
j X

n
,
we dene c
(n)
j
= c
(n)
k
whenever k and j are so related. Comparing to (3.12), we
then dene
(3.14) I

n
f(x) :=

kX

n
f(
k
2n
)

n
(x
k
2n
), where

n
(x) :=
1
2n
2

n
c
(n)

(x).
We also introduce the following notation: for k X
e
n
, we denote by k

the point on
the opposite edge of X

n
; that is, k

X
e
n
and k

= k (2n, 0) or k

= k (0, 2n).
Furthermore, we denote by j

## the index corresponding to k

under (3.13).
Lemma 3.7. The function I

n
f T

n
satises
I

n
f(
k
2n
) =
_

_
f(
k
2n
), k X

n
,
f(
k
2n
) +f(
k

2n
), k X
e
n
,
f(
k
2n
) +f(
(k1,k2)
2n
) +f(
(k1,k2)
2n
) +f(
k
2n
), k X
v
n
.
Proof. As in the proof of the previous theorem, we can write I

n
f as
I

n
f(x) =

n
f(B
tr
j)

n
(x B
tr
j)
by using (3.13). Let S
k
(x) =

n
(B
tr
j). For all k, j

n
,
S
k
(B
tr
j) =
1
2n
2

n
c
(n)

(B
tr
(j k)).
Since e

(B
tr
j) = e

(B
tr
j) for any mod B, we derive by using a similar
argument as in Theorem 3.2 that
S
k
(B
tr
j) =
1
2n
2

n
e

(B
tr
(j k)).
By (2.6), S
k
(B
tr
j) =
k,j
if k, j
n
. If j

n

n
then j

n
, so that
if k
n
then S
k
(B
tr
j) =
k,j
. The same holds for the case of j
n
and
k

n

n
. If both k, j

n

n
, then S
k
(B
tr
j) =
k

,j
. Using the relation
(3.13), we have shown that

n
(
jk
2n
) = 1 when k j mod 2nZ
2
and 0 otherwise,
from which the stated result follows.
11
It turns out that the function

n
satises a compact formula. Let us dene an
operator T by
(Tf)(x) =
1
4
[f(x
1
, x
2
) +f(x
1
, x
2
) +f(x
1
, x
2
) +f(x
1
, x
2
)] .
For e
k
(x) = e
2ikx
, it follows immediately that
(3.15) (Te
k
)(x) = cos(2k
1
x
1
) cos(2k
2
x
2
) forall k Z
2
.
Lemma 3.8. For n 0,
(3.16)

n
(x) = 2 [D
n
(x) + D
n1
(x)]
1
4
(cos 2nx
1
+ cos 2nx
2
),
where
(3.17)
D
n
(x) :=
1
4

n
e

(x) =
1
2
cos (2n + 1)x
1
cos x
1
cos (2n + 1)x
2
cos x
2
cos 2x
1
cos 2x
2
.
Proof. Using the values of c
(n)

## and the denition of D

n
, it is easy to see that

n
(x) = 2 [D
n
(x) +D
n1
(x)]

v
e
v
(x).
Since
v
n
contains four terms, (n, 0) and (0, n), the sum over
v
n
becomes the
second term in (3.16). On the other hand, using the symmetry of

n
and (3.15),
D
n
(x) =
1
4

n
(Te

)(x) =

0j1+j2n
cos 2j
1
x
1
cos 2j
2
x
2
,
where

means that the terms in the sum are halved whenever either j
1
= 0 or
j
2
= 0, from which the second equal sign in (3.17) follows from [18, (4.2.1) and
(4.2.7)].
Our main result in this section is interpolation over points in
k
2n
: k
n
with

n
dened in (3.5).
Theorem 3.9. For n 0 dene
L
n
f(x) =

kn
f(
k
2n
)
k
(x),
k
(x) :=
(n)
k
T
_

n
(
k
2n
)

(x)
with
(n)
k
given in (3.8). Then L
n
f T
n
is even in both variables and it satises
L
n
f(
j
2n
) = f(
j
2n
) for all j
n
.
Proof. As shown in the proof of Proposition 3.7, R
k
(x) :=

n
(x
k
2n
) satises
R
k
(
j
2n
) = 1 when k j mod 2nZ
2
and 0 otherwise. Hence, if j

n
then
(TR
k
)(
j
2n
) =
1
4
R
k
(
j
2n
) = [
(n)
k
]
1

k,j
. If j
e
n
then the number of terms
in the sum of (TR
k
)(
j
2n
) depends on whether j
1
j
2
is zero; if j
1
j
2
,= 0 then
(TR
k
)(
j
2n
) =
1
4
_
R
k
(
j
2n
) + R
k
(
j

2n
)
_
=
1
2

k,j
= [
(n)
k
]
1

k,j
, whereas if j
1
j
2
= 0
then (TR
k
)(
j
2n
) =
1
2
R
k
(
j
2n
) = [
(n)
k
]
1

k,j
. For j = (n, 0) or (0, n) in
v
n
, we
have (TR
k
)(
j
2n
) =
1
2
_
R
k
(
j
2n
) +R
k
(
j

2n
)
_
=
k,j
; for j = (n, n)
v
n
we have
(TR
k
)(
j
2n
) =
1
4
_
R
k
(
(n,n)
2n
) +R
k
(
(n,n)
2n
) +R
k
(
(n,n)
2n
) +R
k
(
(n,n)
2n
)
_
=
k,j
; -
nally for j = 0
v
n
, it is evident that (TR
k
)(0) =
k,0
. Putting these together, we
12 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
have veried that
k
(
j
2n
) =
k,j
for all j, k

n
, which veries the interpolation of
L
n
f.
As in the case of cubature, we can translate the above theorem to interpolation by
algebraic polynomials by applying the change of variables (3.4). Recall
n
dened
in (3.6).
Theorem 3.10. For n 0, let
L
n
f(t) =

z
k
n
f(z
k
)

k
(t),

k
(t) =
k
(x) with t
i
= cos 2x
i
, i = 1, 2.
Then L
n
f
2
n
and it satises L
n
f(z
k
) = f(z
k
) for all z
k

n
. Furthermore,
under the change of variables (3.4), the fundamental polynomial

k
(t) satises

k
(t) =
1
2
T
_
D
n
(
k
2n
) +D
n1
(
k
2n
)

(x)
1
4
_
(1)
k1
T
k1
(t
1
) + (1)
k2
T
k2
(t
2
)

.
Proof. That L
n
f interpolates at z
k

n
is an immediate consequence of the change
of variables, which also shows that L
n
f
2
n
. Moreover, cos 2n(x
1

k1
2n
) =
(1)
k1
cos 2nx
1
= (1)
k1
T
n
(x
1
), which veries the formula of

k
(t).
The polynomial L
n
f belongs, in fact, to a subspace

n

2
n
of dimension
[
n
[ = dim
2
n1
+
n
2
+1, and it is the unique interpolation polynomial in

n
. In
the case of n is odd, this interpolation polynomial was dened and studied in ,
where a slightly dierent scheme with one point less was studied in the case of even
n. Recently the interpolation polynomials in  have been tested and studied
numerically in [3, 4]; the results show that these polynomials can be evaluated
eciently and provide valuable tools for numerical computation.
4. Cubature and Interpolation on the cube
For d = 2, the choice of our spectral set
B
and lattice in the previous section
ensures that we end up with a space close to the polynomial subspace
2
n
; indeed,
monomials in
2
n
are indexed by 0 j
1
+ j
2
n, a quarter of

n
. For d = 3,
the same consideration indicates that we should choose the spectral set as the
octahedron x : n x
1
x
2
x
3
n. The octahedron, however, does not tile
R
3
by lattice translation (see, for example, [6, p. 452]). As an alternative, we choose
the spectral set as rhombic dodecahedron, which tiles R
3
by lattice translation with
face centered cubic (fcc) lattice. In , a discrete Fourier analysis on the rhombic
dodecahedron is developed and used to study cubature and interpolation on the
rhombic dodecahedron, which also leads to results on tetrahedron. In contrast, our
results will be established on the cube [
1
2
,
1
2
]
3
, but our set
B
is chosen to be a
rhombic dodecahedron.
4.1. Discrete Fourier analysis and cubature formula on the cube. We
choose our matrix B as the generator matrix of fcc lattice,
B = n
_
_
0 1 1
1 0 1
1 1 0
_
_
and B
1
=
1
2n
_
_
1 1 1
1 1 1
1 1 1
_
_
.
The spectral set of the fcc lattice is the rhombic dodecahedron (see Figure 2). Thus,

B
= x R
3
: n x

< n, 1 < 3.
13
(0, 1, 0)
(
1
2
,
1
2
,
1
2
)
(
1
2
,
1
2
,
1
2
)
(
1
2
,
1
2
,
1
2
)
(
1
2
,
1
2
,
1
2
)
(0, 0, 0)
(1, 0, 0)
(0, 0, 1)
(0, 0, 1)
(1, 0, 0)
(
1
2
,
1
2
,
1
2
)
(
1
2
,
1
2
,
1
2
)
(
1
2
,
1
2
,
1
2
)
(
1
2
,
1
2
,
1
2
)
(0, 1, 0)
Figure 2. Rhombic dodecahedron
The strict inequality in the denition of
B
reects our requirement that the
tiling of the spectral set has no overlapping. From the expression of B
tr
, it follows
that
B
=:
n
is given by

n
:= j Z
3
: n j
1
+j
2
+j
3
, j
1
j
2
+j
3
, j
1
+j
2
j
3
< n.
It is known that [
n
[ = det(B) = 2n
3
. Furthermore,

B
=:

n
is given by

n
= Z
3

B
= k Z
3
: n k

< n, 1 < 3.
We denote the space T
B
by T
n
, which is given by
T
n
:= span
_
e
2i kx
: k

n
_
.
Then dimT
n
= [

n
[ = det(B) = 2n
3
.
Theorem 4.1. Dene the set
X
n
:=
_
2k :
n
2
k
1
, k
2
, k
3
<
n
2
_

_
2k + 1 :
n+1
2
k
1
, k
2
, k
3
<
n1
2
_
.
Then for all f, g T
n
,
f, g)
n
:=
1
2n
3

kXn
f(
k
2n
)g(
k
2n
) =
_
[
1
2
,
1
2
]
3
f(x)g(x)dx.
Proof. Changing variables from j to k = 2nB
tr
j, or j = B
tr
k/(2n), then, as
j
1
, j
2
, j
3
are integers and j
1
=
k2+k3
2
, j
2
=
k1+k3
2
, j
3
=
k1+k2
2
, we see that
(4.1) j
n
2nB
tr
j X
n
and

jn
f(B
tr
j) =

kXn
f(
k
2n
),
from which we conclude that f, g)
n
= f, g)
B
. Consequently, this theorem is a
special case of Theorem 2.1.
Just like the case of d = 2, we denote the symmetric counterpart of X
n
by X

n
which is dened by
X

n
:=
_
2k :
n
2
k
1
, k
2
, k
3

n
2
2k + 1 :
n+1
2
k
1
, k
2
, k
3

n1
2
_
.
A simple counting shows that [X

n
[ = n
3
+(n+1)
3
. The set X

n
is further partitioned
into four parts,
X

n
= X

n
X
f
n
X
e
n
X
v
n
,
14 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
where X

n
= X

n
(n, n)
2
is the set of interior points, X
f
n
contains the points in
X

n
that are on the faces of [n, n]
3
but not on the edges or vertices, X
e
n
contains
the points in X

n
that are on the edges of [n, n]
3
but not on the corners or vertices,
while X
v
n
denotes the points of X

n
at the vertices of [n, n]
3
.
Theorem 4.2. Dene the inner product
(4.2) f, g)

n
:=
1
2n
3

kX

n
c
(n)
k
f(
k
2n
)g(
k
2n
), where c
(n)
k
=
_

_
1, k X

n
1
2
, k X
f
n
1
4
, k X
e
n
1
8
, k X
v
n
.
Then for all f, g T
n
,
_
[
1
2
,
1
2
]
3
f(x)g(x)dx = f, g)
n
= f, g)

n
.
Proof. The proof follows along the same line as the proof of Theorem 3.2. We only
need to show f, g)
n
= f, g)

n
if fg is periodic. The interior points of X
n
and X

n
are the same, so that c
(n)
k
= 1 for k X

n
. Let
1
= (1, 0, 0),
2
= (0, 1, 0), and

3
= (0, 0, 1). Each point k in X
f
n
has exactly one opposite point k

in X
f
n
under
translation by n
i
and only one of them is in X
n
, so that f(x
k
) =
1
2
[f(x
k
)+f(x

k
)]
if f is periodic, which is why we dene c
(n)
k
=
1
2
for k X
f
n
. Evidently, only three
edges of X

n
are in X

n
X
n
. Each point in X
e
n
corresponds to exactly four points
in X
e
n
under integer translations n
i
and only one among the four is in X
n
, so
we dene c
(n)
k
=
1
4
for k X
e
n
. Finally, all eight corner points can be derived
from translations n
i
points, used repeatedly, and exactly one, (n, n, n), is in
X

n
X
n
, so that we dene c
(n)
k
=
1
8
for k X
v
n
.
We also denote the symmetric counterpart of

n
by

n
,
(4.3)

n
:= j Z
3
: n j

n, 1 < 3
and denote the counterpart of T
n
by T

n
, which is dened accordingly by
T

n
:= span
_
e
2i kx
: k

n
_
.
Theorem 4.3. For n 2, the cubature formulas
(4.4)
_
[
1
2
,
1
2
]
3
f(x)dx =
1
2n
3

kX

n
c
(n)
k
f(
k
2n
) and
_
[
1
2
,
1
2
]
3
f(x)dx =
1
2n
3

kXn
f(
k
2n
)
are exact for f T

2n1
.
Proof. As in the proof of Theorem 3.3, for any j Z
3
, there exist

n
and
l Z
3
such that j = +Bl.
Assume now j

2n1
. Clearly the integral of e
j
over is
j,0
. On the other
hand, let us suppose j = +Bl with
n
and l Z
3
. Then it is easy to see that
e
j
(
k
2n
) = e

(
k
2n
) for each k X

n
. Consequently, we get from Theorem 4.2 that

kX

n
c
(n)
k
e
j
(
k
2n
) =

kX

n
c
(n)
k
e

(
k
2n
) =

kXn
e

(
k
2n
)
=

kXn
e
j
(
k
2n
) =
_

(x)dx =
,0
.
15
Since = 0 implies j = l = 0, we further obtain that
,0
=
j,0
. This states that
the cubature formulas (4.4) are exact for each e
j
with j

2n1
, which completes
the proof.
4.2. Cubature formula for algebraic polynomials. We can also translate the
cubature in Theorem 4.3 into one for algebraic polynomials. For this we use the
change of variables
(4.5) t
1
= cos 2x
1
, t
2
= cos 2x
2
, t
3
= cos 2x
3
, x [
1
2
,
1
2
]
3
.
Under (4.5), the functions cos 2k
1
x
1
cos 2k
2
x
2
cos 2k
3
x
3
become algebraic poly-
nomials T
k1
(t
1
)T
k2
(t
2
)T
k3
(t
3
), which are even in each of its variables. The subspace
of T

n
that consists of functions that are even in each of its variables corresponds
to the polynomial subspace

n
:= spanT
k1
(x
1
)T
k2
(x
2
)T
k3
(x
3
) : k
1
, k
2
, k
3
0, k

+k

n, 1 < n.
Notice that X

n
is symmetric in the sense that if x X

n
then x X

n
for all
1, 1
3
, where (x)
i
=
i
x
i
. In order to evaluate functions that are even
in each of its variables on X

n
we only need to consider X

n
x : x
1
, x
2
, x
3
0.
Hence, we dene,
(4.6)
n
:= 2k : 0 k
1
, k
2
, k
3

n
2
2k + 1 : 0 k
1
, k
2
, k
3

n1
2

and, under the change of variables (4.5), dene
(4.7)
n
:= (z
k1
, z
k2
, z
k3
) : k
n
, z
k
=
k
2n
.
Moreover, we denote by

n
,
f
n
,
e
n
and
v
n
the subsets of
n
that contains interior
points, points on the faces but not on the edges, points on the edges but not on the
vertices, and points on the vertices, of [1, 1]
3
, respectively, and we dene

n
,
f
n
,

e
n
and
v
n
accordingly. A simple counting shows that
(4.8) [
n
[ = (
n
2
+ 1)
3
+ (
n1
2
+ 1)
3
=
_
(n+1)
3
4
+
3(n+1)
4
, n is even,
(n+1)
3
4
, n is odd.
Theorem 4.4. Write z
k
= (z
k1
, z
k2
, z
k3
). The cubature formula
(4.9)
1

3
_
[1,1]
3
f(t)W
0
(t)dt =
1
2n
2

kn

(n)
k
f(z
k
),
(n)
k
:=
_

_
8, k

n
,
4, k
f
n
,
2, k
e
n
,
1, k
v
n
,
is exact for

2n1
. In particular, it is exact for
3
2n1
.
Proof. Let g(x) = f(cos 2x
1
, cos 2x
2
, cos 2x
3
). Then g is even in each of its
variables and g(
k
2n
) = f(z
k
). Applying the rst cubature formula in (3.3) to g(x),
we see that (3.8) follows from the following identity,

kX

n
c
(n)
k
g(
k
2n
) =

kn

(n)
k
f(z
k
).
This identity is proved in the same way that the corresponding identity in Theorem
3.4 is proved. Let k denote the set of distinct elements in k : 1, 1
3
; then
g(
k
2n
) takes the same value on all points in k. If k X

n
, k
i
,= 0 for i = 1, 2, 3, then
k contains 8 points; if exactly one k
i
is zero then k contains 4 points; if exactly
two k
i
are zero then k contains one point; and, nally, if k = (0, 0, 0) then k
16 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
contains one point. In the case of k
i
,= 0 for i = 1, 2, 3,

jk
c
(n)
k
g(
j
2n
) = 8g(
k
2n
)
if k X

n
,

jk
c
(n)
k
g(
j
2n
) = 4g(
k
2n
) if k X
f
n
, and

jk
c
(n)
k
g(
j
2n
) = 2g(
k
2n
) if
k X
e
n
. The other cases are treated similarly. Thus, (4.9) holds for

2n1
.
Finally, the denition of

n

3
n
= spanT
k1
(x
1
)T
k2
(x
2
)T
k3
(x
3
) : k
1
, k
2
, k
3
0, 0 k
1
+k
2
+k
3
n
as a subspace. In particular,

2n1
contains
3
2n1
as a subset.
We note that

2n1
contains
3
2n1
as a subspace, but it does not contain
3
2n
since T
n
(x
1
)T
n
(x
2
) is in
3
2n
but not in
3
2n1
. Hence, the cubature (4.9) is of
degree 2n 1. A trivial cubature formula of degree 2n 1 for W
0
can be derived
by taking the product of Gaussian quadrature of degree 2n 1 in one variable,
which has exactly n
3
nodes. In contrast, according to (4.8), the number of nodes of
our cubature (3.8) is in the order of n
3
/4 +O(n
2
), about a quarter of the product
formula. As far as we know, this is the best that is available at the present time.
On the other hand, the lower bound for the number of nodes states that a cubature
formula of degree 2n1 needs at least n
3
/6 +O(n
2
) nodes. It is, however, an open
question if there exist formulas with number of nodes attaining this theoretic lower
bound.
Recall the cubature (4.9) is derived by choosing the spectral set as a rhombic
dodecahedron. One natural question is how to choose a spectral set that tiles R
3
by translation so that the resulted cubature formula is of degree 2n1 and has the
smallest number of nodes possible. Among the regular lattice tiling, the rhombic
dodecahedron appears to lead to the smallest number of nodes.
Just as Theorem 3.5, we can also derive a cubature formula of degree 2n 5 for
W
1
from Theorem 4.3. We omit the proof as it follows exactly as in Theorem 3.5.
Theorem 4.5. The cubature formula
(4.10)
1

3
_
[1,1]
3
f(t)W
1
(t)dt =
4
n
3

n
sin
2 k1
n
sin
2 k2
n
sin
2 k3
n
f(z
k
)
is exact for

2n5
; in particular, it is exact for
3
2n5
.
4.3. A compact formula for a partial sum. In order to obtain the compact
formula for the interpolation function, we follow  and use homogeneous coordi-
nates and embed the rhombic dodecahedron into the plane t
1
+ t
2
+ t
3
+ t
4
= 0
of R
4
. Throughout the rest of this paper, we adopt the convention of using bold
letters, such as t, to denote the points in the space
R
4
H
:=
_
t = (t
1
, t
2
, t
3
, t
4
) R
4
: t
1
+t
2
+t
3
+t
4
= 0
_
.
In other words, the bold letters such as t and k will always mean homogeneous
coordinates. The transformation between x R
3
and t R
4
H
is dened by
(4.11)
_

_
x
1
= t
2
+t
3
x
2
= t
1
+t
3
x
3
= t
2
+t
1

_
t
1
=
1
2
(x
1
+x
2
+x
3
)
t
2
=
1
2
(x
1
x
2
+x
3
)
t
3
=
1
2
(x
1
+x
2
x
3
)
t
4
=
1
2
(x
1
x
2
x
3
).
In this homogenous coordinates, the spectral set
B
becomes
(4.12)
B
=
_
t R
4
H
: 1 < t
i
t
j
1, 1 i < j 4
_
.
17
We now use homogeneous coordinates to describe

n
dened in (4.3). Let
Z
4
H
:= Z
4
R
4
H
and
H := j Z
4
H
: j
1
j
2
j
3
j
4
mod 4.
In order to keep the elements as integers, we make the change of variables
j
1
= 2(k
1
+k
2
+k
3
), j
2
= 2(k
1
k
2
+k
3
),
j
3
= 2(k
1
+k
2
k
3
), j
4
= 2(k
1
k
2
k
3
)
(4.13)
for k = (k
1
, k
2
, k
3
)

n
. It then follows that

n
in homogeneous coordinates
becomes
G
n
:= j H : j
1
j
2
j
3
j
4
0 mod 2, 4n j

4n, 1 , 4.
We could have changed variables without the factor 2, setting j
1
= k
1
+ k
2
+ k
3
etc. We choose the current change of variables so that we can use some of the
computations in . In fact, the set
(4.14) H

n
:= j H : 4n j

4n, 1 , 4
is used in . The main result of this subsection is a compact formula for the partial
sum
(4.15) D
n
(x) :=

n
e
k
(x) =

jGn
e
j
(t) =: D

n
(t), e
j
(t) := e
i
2
jt
,
where x and t are related by (4.11) and the middle equality follows from the fact
that

n
= G
n
under this change of variables. In fact, by (4.11) and (4.13), we
have
k x = k
1
(t
2
+t
3
) +k
2
(t
1
+t
3
) +k
3
(t
1
+t
2
)
= (k
2
+k
3
)t
1
+ (k
1
+k
3
)t
2
+ (k
1
+k
2
)t
3
=
1
4
[(j
1
j
4
)t
1
+ (j
2
j
4
)t
2
+ (j
3
j
4
)t
3
] =
1
4
j t
where in the last step we have used the fact that t R
4
H
. The compact formula of
D
n
(t) is an essential part of the compact formula for the interpolation function.
Theorem 4.6. For n 1,
D

n
(t) =
n+1
(t)
n
(t)
_

odd
n
(t)
odd
n2
(t)
_
,
where

n
(t) =
4

i=1
sin nt
i
sin t
i
,
and for n 1,

odd
n
(t) =
4

i=1
sin(n + 2)t
i
sin 2t
i
4

j=1
sin nt
j
sin(n + 2)t
j
, if n = even,
and

odd
n
(t) =
4

i=1
sin(n + 1)t
i
sin 2t
i
4

j=1
sin(n + 3)t
j
sin(n + 1)t
j
, if n = odd.
18 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
Proof. By denition, G
n
is a subset of H

n
that contains elements with all indices
being even integers. For technical reasons, it turns out to be easier to work with
H

n
G
n
. In fact, the sum over H

n
has already been worked out in , which is

jH

j
(t) =
4

i=1
sin(n + 1)t
i
sin t
i

i=1
sin nt
i
sin t
i
=
n+1
(t)
n
(t).
Thus, we need to nd only the sum over odd indices, that is, the sum
D
odd
n
(t) :=

jH
odd
n
e
j
(t), H
odd
n
:= H

n
G
n
.
Just as in , the index set H
odd
n
can be partitioned into four congruent parts, each
within a parallelepiped, dened by
H
(k)
n
:=
_
j H
odd
n
: 0 j
l
j
k
4n, l N
4
_
for k N
4
. Furthermore, for each index set J, J N
4
, dene
H
J
n
:=
_
k H
odd
n
: k
i
= k
j
, i, j J; and 0 k
i
k
j
4n, j J, i N
4
J
_
.
Then we have
H
odd
n
=
_
jN4
H
(j)
n
and H
J
n
=

jJ
H
(j)
n
.
Using the inclusion-exclusion relation of subsets, we have
D
odd
n
(t) =

JN4
(1)
|J|+1

kH
J
n
e
i
2
kt
.
Fix j J, using the fact that t
j
=

i=j
t
i
, we have

kH
J
n
e
i
2
kt
=

kH
J
n
e
i
2
P
lN
4
\J
(k
l
kj)t
l
=

kH
J
n

lN4\J
e
i
2
(k
l
kj)t
l
.
Since k H
J
n
implies, in particular, k
i
k
j
mod 4, we obtain

kH
J
n
e
i
2
kt
=

lN4\J

0k
l
kj4n
kH
J
n
e
i
2
(k
l
kj)t
l
=

lN4\J

0k
l
n
|k|J odd
e
2i k
l
t
l
,
where [k[
J
:=

lN4\J
k
l
. The last equation needs a few words of explanation:
if 4k

l
= k
l
k
j
, then using the fact that k
i
= k
j
, i, j J for k H
J
n
and
k
1
+k
2
+k
3
+k
4
= 0, we see that
1
4

lN4\J
(k
l
k
j
) = k
j
, which is odd by the
denition of H
J
n
; on the other hand, assume that

lN4\J
k

l
is odd, then we dene
k
j
=

lN4\J
k

l
for all j J and dene k
l
= 4kl

+k
j
, so that all components of
k are odd and k H
J
n
.
The condition that [k[
J
is an odd integer means that the last term is not a simple
product of sums. Setting
D
O
n
(t) :=
n

j=0,j odd
e
2ijt
=
e
2it
(1 e
4i
n+1
2
t
)
1 e
4it
,
D
E
n
(t) :=
n

i=0,i even
e
2ijt
=
1 e
4i
n+2
2
t
1 e
4it
,
19
we see that, up to a permutation, only products D
O
n
D
O
n
D
O
n
and D
O
n
D
E
n
D
E
n
are
possible for triple products ([J[ = 3), only D
O
n
D
E
n
is possible for double products
([J[ = 2), only D
O
n
is possible ([J[ = 1), and there is a constant term. Thus, using
the fact that abc (a 1)(b 1)(c 1) = ab +ac +bc a b c +1, we conclude
that
D
odd
n
(t) =

(i1,i2,i3)N4
D
O
n
(t
i1
)D
O
n
(t
i2
)D
O
n
(t
i3
)
+D
O
n
(t
1
)
_
D
E
n
(t
2
)D
E
n
(t
3
)D
E
n
(t
4
) (D
E
n
(t
2
) 1)(D
E
n
(t
3
) 1)(D
E
n
(t
4
) 1)

+D
O
n
(t
2
)
_
D
E
n
(t
1
)D
E
n
(t
3
)D
E
n
(t
4
) (D
E
n
(t
1
) 1)(D
E
n
(t
3
) 1)(D
E
n
(t
4
) 1)

+D
O
n
(t
3
)
_
D
E
n
(t
1
)D
E
n
(t
2
)D
E
n
(t
4
) (D
E
n
(t
1
) 1)(D
E
n
(t
2
) 1)(D
E
n
(t
4
) 1)

+D
O
n
(t
4
)
_
D
E
n
(t
1
)D
E
n
(t
2
)D
E
n
(t
3
) (D
E
n
(t
1
) 1)(D
E
n
(t
2
) 1)(D
E
n
(t
3
) 1)

,
where the rst sum is over all distinct triple integers in N
4
.
Assume that n is an even integer. A quick computation shows that
D
O
n
(t
1
)D
E
n
(t
2
)D
E
n
(t
2
)D
E
n
(t
4
) =
4

j=2
sin (n + 2)t
i
sin 2t
i
sin nt
1
sin 2t
1
.
Furthermore, we see that
D
O
n
(t
2
)D
O
n
(t
3
)D
O
n
(t
4
) D
O
n
(t
1
)(D
E
n
(t
2
) 1)(D
E
n
(t
3
) 1)(D
E
n
(t
4
) 1)
=
4

j=2
sin nt
i
sin 2t
i
_
e
int1

e
2int1
eint
1
1 e
4it1
_
=
4

j=2
sin nt
i
sin 2t
i
sin (n 2)t
1
sin 2t
1
.
Adding the two terms together and then summing over the permutation of the
sum, we end up the formula for D
odd
n
(t) when n is even. The case of n odd can be
handled similarly.
Let us write down explicitly the function D
n
(x) dened in (4.15) in x-variables.
Using the elementary trigonometric identity and (4.11), we see that
4
4

i=1
sint
i
= (cos (x
2
x
1
) cos x
3
)(cos (x
2
+x
1
) cos x
3
)
= cos
2
x
1
+ cos
2
x
2
+ cos
2
x
3
2 cos x
1
cos x
2
cos x
3
1,
so that we end up with the compact formula
D
n
(x) =

n+1
(x)

n
(x)
_

odd
n
(x)

odd
n2
(x)
_
, (4.16)
where

n
(x) =
cos
2
nx
1
+ cos
2
nx
2
+ cos
2
nx
3
2 cos nx
1
cos nx
2
cos nx
3
1
cos
2
x
1
+ cos
2
x
2
+ cos
2
x
3
2 cos x
1
cos x
2
cos x
3
1
,

odd
n
(x) =

n+2
2
(2x)
4

j=1
sin nt
j
sin(n + 2)t
j
, if n = even,
and

odd
n
(t) =

n+1
2
(2x)
4

j=1
sin(n + 3)t
j
sin(n + 1)t
j
, if n = odd,
20 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
in which t
i
is given in terms of x
j
in (4.11). As a result of this explicit expression,
we see that D
n
(x) is an even function in each x
i
.
4.4. Boundary of the rhombic dodecahedron. In order to develop the inter-
polation on the set X

n
, we will need to understand the structure of the points on
the boundary of

n
= Z
3

B
. As
B
is a rhombic dodecahedron, we need to
understand the boundary of this 12-face polyhedron, which has been studied in
detail in . In this subsection, we state the necessary denitions and notations on
the boundary of
B
, so that the exposition is self-contained. We refer to further
details and proofs to .
Again we use homogeneous coordinates. For i, j N
4
:= 1, 2, 3, 4 and i ,= j,
the (closed) faces of
B
are
F
i,j
= t
H
: t
i
t
j
= 1.
There are a total 2
_
4
2
_
= 12 distinct F
i,j
, each represents one face of the rhombic
dodecahedron. For nonempty subsets I, J of N
4
, dene

I,J
:=

iI,jJ
F
i,j
=
_
t
H
: t
j
= t
i
1, for all i I, j J
_
.
It is shown in  that
I,J
= if and only if I J ,= , and
I1,J1

I2,J2
=
I,J
if
I
1
I
2
= I and J
1
J
2
= J. These sets describe the intersections of faces, which can
then be used to describe the edges, which are intersections of faces, and vertices,
which are intersections of edges. Let
/ := (I, J) : I, J N
4
; I J = ,
/
0
:= (I, J) / : i < j, for all (i, j) (I, J) .
We now dene, for each (I, J) /, the boundary element B
I,J
of the dodecahedron,
B
I,J
:= t
I,J
: t ,
I1,J1
for all (I
1
, J
1
) / with [I[ +[J[ < [I
1
[ +[J
1
[ ;
it is called a face if [I[ +[J[ = 2, an edge if [I[ +[J[ = 3, and a vertex if [I[ +[J[ = 4.
By denition, the elements for faces and edges are without boundary, which implies
that B
I,J
B
I

,J
= if I ,= I
1
and J ,= J
1
. In particular, it follows that B
{i},{j}
=
F

i,j
and, for example, B
{i},{j,k}
= (F
i,j
F
i,k
)

## for distinct integers i, j, k N

4
.
Let ( = S
4
denote the permutation group of four elements and let
ij
denote
the element in ( that interchanges i and j; then t
ij
= t (t
i
t
j
)e
i,j
. For a
nonempty set I N
4
, dene (
I
:=
ij
: i, j I, where we take
ij
=
ji
and
take
jj
as the identity element. It follows that (
I
forms a subgroup of ( of order
[I[. For (I, J) /, we then dene
[B
I,J
] :=
_
GIJ
B
I,J
. (4.17)
It turns out that [B
I,J
] consists of exactly those boundary elements that can be
obtained from B
I,J
by congruent modulus B, and [B
I,J
] [B
I1,J1
] = if (I, J) ,=
(I
1
, J
1
) for (I, J) /
0
and (I
1
, J
1
) /
0
. More importantly, we dene, for 0 <
i, j < i +j 4,
B
i,j
:=
_
(I,J)K
i,j
0
[B
I,J
] with /
i,j
0
:= (I, J) /
0
: [I[ = i, [J[ = j .
(4.18)
21
Then the boundary of
B
can be decomposed as

H
=
_
(I,J)K
B
I,J
=
_
0<i,j<i+j4
B
i,j
.
The main complication is the case of [I[+[J[ = 2, for which we have, for example,
[B
{1},{2,3}
] = B
{1},{2,3}
B
{2},{1,3}
B
{3},{1,2}
. (4.19)
The other cases can be written down similarly. Furthermore, we have
B
{1},{2,4}
= B
{1},{2,3}

34
, B
{1,2},{4}
= B
{1,2},{3}

34
,
B
{1},{3,4}
= B
{1},{2,3}

24
, B
{1,3},{4}
= B
{1,2},{3}

23

34
,
B
{2},{3,4}
= B
{1,2},{3}

12

24
, B
{2,3},{4}
= B
{1,2},{3}

13

34
,
(4.20)
with
B
{1},{2,3}
=
_
(t, t 1, t 1, 2 3t) :
1
2
< t <
3
4
_
,
B
{1,2},{3}
=
_
(1 t, 1 t, t, 3t 2) :
1
2
< t <
3
4
_
.
(4.21)
If [I[ +[J[ = 2 then B
I,J
= B
{i},{j}
is a face and
B
1,1
= [B
{1},{2}
] [B
{1},{3}
] [B
{1},{4}
] [B
{2},{3}
] [B
{2},{4}
] [B
{3},{4}
]
If [I[ +[J[ = 3 then B
I,J
is an edge and we have
B
1,2
= [B
{1},{2,3}
] [B
{1},{2,4}
] [B
{1},{3,4}
] [B
{2},{3,4}
],
B
2,1
= [B
{1,2},{3}
] [B
{1,2},{4}
] [B
{1,3},{4}
] [B
{2,3},{4}
].
(4.22)
If [I[ +[J[ = 4, then
B
1,3
=
_
(
1
4
,
1
4
,
1
4
,
3
4
)

, B
2,2
=
_
(
1
2
,
1
2
,
1
2
,
1
2
)

B
3,1
=
_
(
3
4
,
1
4
,
1
4
,
1
4
)

.
(4.23)
Recall that G
n
is

n
= Z
3

B
in homogeneous coordinates. We now consider
the decomposition of the boundary of G
n
according to the boundary elements of
the rhombic dodecahedron. First we denote by G

n
the points inside G
n
,
G

n
:= j G
n
: 4n < j

< 4n, 1 , 4 =
_
j G
n
:
j
4n

B
_
.
We further dene, for 0 < i, j < i +j 4,
G
i,j
n
:=
_
k G
n
:
k
4n
B
i,j
_
(4.24)
The set G
i,j
n
describes those points j in G
n
such that
j
4n
are in B
i,j
of
B
. It is
easy to see that G
i,j
n
G
k,l
n
= if i ,= k, j ,= l and
_
0<i,j<i+j4
G
i,j
n
= G
n
G

n
.
4.5. Interpolation by trigonometric polynomials. We rst apply the general
theory from Section 2 to our set up with
B
as a rhombic dodecahedron.
Theorem 4.7. For n 1 dene
(4.25) I
n
f(x) :=

kXn
f(
k
2n
)
n
(x
k
2n
),
n
(x) :=
1
2n
3

n
e

(x).
Then for each j X
n
, I
n
(
j
2n
) = f(
j
2n
).
22 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
Proof. By (4.1), I
n
f(
j
2n
) = f(
j
2n
) for j X
n
is equivalent to I
n
f(B
tr
l) = f(B
tr
l)
for l
n
. Moreover, I
n
f can be rewritten as
I
n
f(x) =

jn
f(B
tr
j)
n
(x B
tr
j).
Hence, this theorem is a special case of Theorem 2.2.
Next we consider interpolation on the symmetric set of points X

n
. For this we
need to modify the kernel function
n
. Recall that, under the change of variables
(4.13),

n
becomes G
n
in homogeneous coordinates. We dene

n
(x) :=
1
2n
3

n

(n)

(x) =
1
2n
3

jGn

(n)
j
e
j
(t),
where x and t are related by (4.11),
(n)
k
is dened by
(n)
k
under the change of
indices (4.13), and
(n)
j
= 1 if j G

n
,
(n)
j
=
1
(
i+j
i
)
if j G
i,j
n
; more explicitly

(n)
j
:=
_

_
1, j G

n
1
2
, j G
1,1
n
,
1
3
, j G
1,2
n
G
2,1
n
,
1
4
, j G
1,3
n
G
3,1
n
,
1
6
, j G
2,2
n
.
For each k on the boundary of X

n
, that is,
k
2n
on the boundary of [
1
2
,
1
2
]
3
, let
(4.26) o
k
:= j X

n
:
j
2n

k
2n
mod Z
3
,
which contains the points on the boundary of X

n
that are congruent to k under
integer translations.
Theorem 4.8. For n 1 dene
(4.27) I

n
f(x) :=

kX

n
f(
k
2n
)R
k
(x), R
k
(x) :=

n
(x
k
2n
).
Then for each j X

n
,
I

n
f(
j
2n
) =
_

_
f(
j
2n
), j X

n
,

kSj
f(
k
2n
), j X

n
X

n
.
(4.28)
In homogeneous coordinates, the function

n
(x) =

n
(t) is a real function and it
satises

n
(t) =
1
4n
3
_

_
1
2
_
D

n
(t) +D

n1
(t)
_

1
3
4

=1
sin2
n1
2
t

sin2t

j=1
j=
cos 2(nt
j
+
n
2
t

1
3

1<4
cos 2n(t

+t

)
1
2
_

4
j=1
cos 2nt
j
, if n even
0 if n odd
_
_
,
(4.29)
23
from which the formula for

n
(x) follows from (4.11) and (4.16).
Proof. By (4.1), we need to verify the interpolation at the points B
tr
l for l

n
.
By denition, we can write
R
k
(B
tr
l) =
1
2n
3

n

(n)

(B
tr
(l k)).
It is easy to see that
tr
B
tr
l =
1
4n
(j
1
l
1
+j
2
l
2
+j
3
l
3
) if is related to j by (4.13).
Hence, as in the proof of Theorem 3.15 in , we conclude that
R
k
(B
tr
l) =
1
2n
3

n
e

(B
tr
(l k)),
Now, for l, k

n
, there exist p
n
and q Z
3
such that l k p B
tr
q.
Consequently, it follows from (2.6) that
R
k
(B
tr
l) =
1
2n
3

n
e

(B
tr
p) =
p,0
.
By (4.1), we have veried that
(4.30) R
k
(
j
2n
) =
_
1,
j
2n

k
2n
mod Z
3
,
0, otherwise,
which proves the interpolation part of the theorem.
In order to prove the compact formula, we start with the following formula that
can be established exactly as in the proof of Theorem 3.15 in :

n
(t) =
1
4n
3
_
1
2
(D

n
(t) +D

n1
(t))
1
6

kG
1,2
n
G
2,1
n

k
(t) (4.31)

1
4

kG
1,3
n
G
3,1
n

k
(t)
1
3

kG
2,2
n

k
(t)
_
.
Let us dene G
I,J
n
:= k G
n
:
k
4n
B
I,J
for I, J N
4
and also dene
_
G
I,J
n

:=
k G
n
:
k
4n
[B
I,J
]. It follows from (4.18), and (4.24) that
G
i,j
n
=
_
I,JK
i,j
0
_
G
I,J
n

and
_
G
I,J
n

=
_
GIJ
G
I,J
n
.
In order to compute the sums in (4.31), we need to use the detail description of the
boundary elements of
B
in the previous subsection. The computation is parallel
to the proof of Theorem 3.15 in , in which the similar computation with G
n
replaced by H
n
is carried out. Thus, we shall be brief.
Using t
1
+t
2
+t
3
+t
4
= 0 and the explicit description of B
{1},{2,3}
, we get

k[G
{1},{2,3}
n
]

k
(t) =

kG
{1},{2,3}
n
e
i
2
kt
+

kG
{2},{1,3}
n
e
i
2
kt
+

kG
{3},{1,2}
n
e
i
2
kt
=
n1

j=1,jeven
e
2ijt4
_
e
2ni(t1+t4)
+e
2ni(t2+t4)
+e
2ni(t3+t4)
_
=
sin 2
n1
2
t
4
sin2t
4
e
2i
n+1
2
t4
_
e
2in(t1+t4)
+e
2in(t2+t4)
+e
2in(t3+t4)
_
,
24 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
Similarly, we also have

k[G
{1,2},{3}
n
]

k
(t) =

kG
{1},{2,3}
n
e
i
2
kt
+

kG
{2},{1,3}
n
e
i
2
kt
+

kG
{3},{1,2}
n
e
i
2
kt
=
sin 2
n1
2
t
4
sin 2t
4
e
2i
n+1
2
t4
_
e
2in(t1+t4)
+e
2in(t2+t4)
+e
2in(t3+t4)
_
.
From these and their permutations, we can compute the sum over G
1,2
n
and G
2,1
n
.
Putting them together, we obtain

kG
1,2
n
G
2,1
n

k
(t) = 2
4

=1
sin 2
n1
2
t

sin 2t

j=1
j=
cos 2(nt
j
+
n
2
t

).
Using (4.23), we see that, G
2,2
n
= (2n, 2n, 2n, 2n) : ( and, if n is even
then G
1,3
n
= (n, n, n, 3n) : ( and G
3,1
n
= (3n, n, n, n) : (,
whereas if n is odd, then G
1.3
n
= G
3,1
n
= . As a result, it follows that

kG
2,2
n

k
(t) =

1<4
e
2in(t+t)
=

1<4
cos 2n(t

+t

),
where we have used the fact that t
1
+t
2
+t
3
+t
4
= 0, and

kG
1,3
n
G
3,1
n

k
(t) =
4

j=1
_
e
2intj
+e
2intj
_
= 2
4

j=1
cos 2nt
j
,
if n is even, whereas it is equal to 0 if n is odd.
Putting all these into (4.31) completes the proof.
Theorem 4.9. Let |I

n
|

## denote the norm of the operator I

n
: C([
1
2
,
1
2
]
3
)
C([
1
2
,
1
2
]
3
). Then there is a constant c, independent of n, such that
|I

n
|

c(log n)
3
.
Proof. Following the standard procedure, we see that
|I

n
|

= max
x[
1
2
,
1
2
]
3

kX

n
(x
k
4n
)

.
Using the formula of

n
in (4.29), it is easy to see that it suces to prove that
max
x[
1
2
,
1
2
]
3

kX

n
(x
k
2n
)

c(log n)
3
, n 0.
Furthermore, using the explicit formula of D
odd
n
(t) and (3.19) in , we see that
our main task is to estimate the sums in the form of
I
{1,2,3}
:=
1
2n
3
max
tQ

kX

sin n(t
1

k2+k3
2n
) sinn(t
2

k1+k3
2n
) sinn(t
3

k1+k2
2n
)
sin (t
1

k2+k3
2n
) sin(t
2

k1+k3
2n
) sin(t
3

k1+k2
2n
)

## and three other similar estimates I

{1,2,4}
, I
{1,3,4}
and I
{2,3,4}
, respectively, as well
as similar sums in which the denominator becomes product of sin 2(t
i

ki
2n
) and
n in the numerator is replace by n + 1 or n + 2. Here Q is the image of [1, 1]
3
under the mapping (4.11); that is,
Q = t R
4
H
:
1
2
t
1
+t
2
, t
2
+t
3
, t
3
+t
1

1
2
.
25
Changing the summation indices and enlarging the set X

n
, we see that
I
{1,2,3}
4 max
t[1,1]
_
1
2n
2n

k=0

sin n(t
k
2n
)
sin (t
k
2n
)

_
3
c(log n)
3
,
where the last step follows from the standard estimate of one variable (cf. [20, Vol.
II, p. 19]).
4.6. Interpolation by algebraic polynomials. The main outcome of Theorem
4.7 in the previous section is that we can derive a genuine interpolation by trigono-
metric polynomials based on the set of points in
k
2n
: k
n
dened at (4.6).
The development below is similar to the case of d = 2. We dene
Tf(x) :=
1
8

{1,1}
3
f(
1
x
1
,
2
x
2
,
3
x
3
).
Theorem 4.10. For n 0 dene
L
n
f(x) =

kn
f(
k
2n
)
k
(x),
k
(x) :=
(n)
k
T
_

n
(
k
2n
)

(x)
with
(n)
k
given in (4.9). Then L
n
f T
n
is even in each of its variables and it
satises
L
n
f(
j
2n
) = f(
j
2n
) for all j
n
.
Proof. As shown in (4.30), R
k
(x) :=

n
(x
k
2n
) satises R
k
(
j
2n
) = 1 when k j
mod 2nZ
3
and 0 otherwise. Hence, if j

n
then (TR
k
)(
j
2n
) =
1
8
R
k
(
j
2n
) =
[
(n)
k
]
1

k,j
. If j

n
, then we need to consider several cases, depending on
how many components of j are zero, which determines how many distinct terms
are in the sum (TR
k
)(
j
2n
) and how many distinct k can be obtained from j by
congruent in Z
3
. For example, if j
f
n
and none of the components of j are zero,
then there are 2 elements in o
j
, j and the one in the opposite face, and the sum
TR
k
(
j
2n
) contains 8 terms, so that (TR
k
)(
j
2n
) =
1
4

j,k
= [
(n)
k
]
1

k,j
. The other
cases can be veried similarly, just as in the case of d = 2. We omit the details.
The above theorem yields immediately interpolation by algebraic polynomials
upon applying the change of variables (4.5). Recall
n
dened in (4.7) and the
polynomial subspace

n
= spans
k1
1
s
k2
2
s
k3
3
: k
1
, k
2
, k
3
0, k
i
+k
j
n, 1 i, j 3.
Theorem 4.11. For n 0, let
L
n
f(s) =

z
k
n
f(z
k
)

k
(s),

k
(s) =
k
(x) with s = cos 2x.
Then L
n
f

n
and it satises L
n
f(z
k
) = f(z
k
) for all z
k

n
.
This theorem follows immediately from the change of variables (4.5). The explicit
compact formula of
k
(x), thus

k
(s), can be derived from Theorem 4.8.
The theorem states that the interpolation space for the point set
n
is exactly

n
, which consists of monomials that have indices in the positive quadrant of the
rhombic dodecahedron, as depicted in Figure 3 below.
The set
n
consists of roughly n
3
/4(1 + O(n
1
) points. The interpolation poly-
nomial L
n
f

n
is about a total degree of 3n/2. The compact formula of the
26 HUIYUAN LI, JIACHANG SUN, AND YUAN XU
(0, n, 0)
(
n
2
,
n
2
,
n
2
)
(n, 0, 0)
(0, 0, 0)
(0, 0, n)
Figure 3. Index set of

n
fundamental interpolation polynomial provides a convenient way of evaluating the
interpolation polynomial. Furthermore, the Lebesgue constant of this interpolation
process remains at the order of (log n)
3
, as the consequence of Theorem 4.9 and
the change of variables.
Corollary 4.12. Let |L
n
|

## denote the operator norm of L

n
: C([1, 1]
3
)
C([1, 1]
3
). Then there is a constant c, independent of n, such that
|L
n
|

c(log n)
3
.
References
 H. Berens, H. Schmid and Yuan Xu, Multivariate Gaussian cubature formula, Arch. Math.
64 (1995), 26 - 32.
 B. Bojanov and G. Petrova, On minimal cubature formulae for product weight function, J.
Comput. Appl. Math. 85 (1997), 113 -121.
 L. Bos, M. Caliari, S. De Marchi, and M. Vianello, A numerical study of the Xu polynomial
interpolation formula in two variables, Computing 76 (2006), 311-324.
 L. Bos, M. Caliari, S. De Marchi, and M. Vianello, Bivariate interpolation at Xu points:
results, extensions and applications, Electron. Trans. Numer. Anal. 25 (2006), 1-16.
 L. Bos, S. De Marchi and M. Vianello, On the Lebesgue constant for the Xu interpolation
formula. J. Approx. Theory 141 (2006), 134-141.
 J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices and Groups, 3rd ed. Springer,
New York, 1999.
 De Marchi, M. Vianello and Yuan Xu, New cubature formulae and hyperinterpolation in
three variables, preprint, (arXiv:math/0805.3529).
 C. Dunkl and Yuan Xu, Orthogonal polynomials of several variables, Encyclopedia of Math-
ematics and its Applications, Vol. 81, Cambridge Univ. Press, Cambridge, 2001.
 H. Li and Yuan Xu, Discrete Fourier Analysis on a dodecahedron and a tetrahedron, Math.
Comp. to appear.
 H. Li, J. Sun and Yuan Xu, Discrete Fourier analysis, cubature and interpolation on a hexagon
and a triangle, SIAM J. Numer. Anal. 46 (2008), 1653-1681.
 H. M oller, Kubaturformeln mit minimaler Knotenzahl, Numer. Math. 25 (1976), p. 185-200.
 H. M oller, Lower bounds for the number of nodes in cubature formulae, in Numerical Inte-
gration, 221-230, ISNM vol. 45, 1979, Birkhauser, Basel.
 C. R. Morrow and T. N. L. Patterson, Construction of algebraic cubature rules using poly-
nomial ideal theory, SIAM J. Numer. Anal., 15 (1978), 953-976.
 A. Stroud, Approximate calculation of multiple integrals, Prentice-Hall, Englewood Clis,
NJ, 1971.
27
 J. Sun, Multivariate Fourier series over a class of non tensor-product partition domains, J.
Comput. Math. 21 (2003), 53-62.
 J. Sun and H. Li, Generalized Fourier transform on an arbitrary triangular domain, Adv.
Comp. Math., 22 (2005), 223-248.
 Yuan Xu, Common Zeros of Polynomials in Several Variables and Higher Dimensional Quad-
rature, Pitman Research Notes in Mathematics Series, Longman, Essex, 1994.
 Yuan Xu, Christoel Functions and Fourier Series for multivariate orthogonal polynomials,
J. Approx. Theory, 82 (1995), 205-239.
 Yuan Xu, Lagrange interpolation on Chebyshev points of two variables, J. Approx. Theory,
87 (1996), 220-238.
 A. Zygmund, Trigonometric series, Cambridge Univ. Press, Cambridge, 1959.
Institute of Software, Chinese Academy of Sciences, Beijing 100080,China