THE TRANSFINITE DIAMETER OF THE REAL BALL AND SIMPLEX
T. BLOOM*, L. BOS, AND N. LEVENBERG
Abstract. We calculate the transfinite diameter for the real unit ball, B
d:= {x ∈ R
d: |x| ≤ 1} and the real unit simplex, T
d:= {x ∈ R
d+: P
dj=1
x
j≤ 1}.
1. Introduction
Suppose that K ⊂ C
dis compact. The transfinite diameter (and the associated notion of capacity) of K is a measure of the size of K, important in classical Complex Potential Theory (for d = 1) and Pluripotential Thoery (for d > 1). It is defined as follows. For n ∈ Z
+consider the monomials of degree at most n, z
γ, |γ| ≤ n which we order as {e
1(z), e
2(z), · · · , e
mn(z)}
so that the ordering respects the degree, i.e., |γ
j| < |γ
i| implies that j <
i, where e
i(z) = z
γi. Here m
n:=
n+dnis the dimension of the space of polynomials of degree at most n in d complex variables. Then for m
npoints z
j∈ K, 1 ≤ j ≤ m
n, the Vandermonde determinant of degree n is defined to be
vdm(z
1, z
2, · · · , z
mn) := det([e
i(z
j)])
1≤i,j≤mn).
The transfinite diameter of K is δ(K) := lim
n→∞
z1,··· ,z
max
mn∈K|vdm(z
1, z
2, · · · , z
mn)|
1/`nwhere `
n:=
d+1dnm
nis the degree of homogeneity of vdm(z
1, z
2, · · · , z
mn) considered as a polynomial on K
mn.
That the limit exists is a result of Zaharjuta [Z] where the following remarkable formula is proved:
(1.1) δ(K) = exp
1
vol(S
d) Z
Sd
log(τ (θ))dV
where
(1.2) S
d:= {θ ∈ R
d: θ
j≥ 0, 1 ≤ j ≤ d, and
d
X
j=1
θ
j= 1}
2010 Mathematics Subject Classification. Primary 32U15.
Key words and phrases. Transfinite Diameter, Real Ball, Real Simplex.
*Supported in part by an NSERC of Canada grant.
1
is a symmetric simplex (of dimension d − 1) in R
dand, for θ ∈ S
d, τ (θ) = τ (θ, K) := lim sup
i→∞, αi/|αi|→θ
inf{ke
i(z) + X
j<i
c
je
j(z)k
1/|αK i|}
is the so-called directional Chebyshev constant. We note that, for θ ∈ int(S
d), Zaharjuta showed that the lim sup in the definition of τ (θ) may be replaced by an ordinary limit.
In certain cases explicit formulas for the transfinite diameter have been calculated. J¸edrzejowski [J] has given the following formula for the unit complex euclidean ball, B
d:= {z ∈ C
d: |z| ≤ 1} :
(1.3) δ(B
d) = exp − 1
2
d
X
j=2
1 j
! .
From the product formula of Schiffer and Siciak [SS] we also have, for the unit cube Q
d:= [−1, 1]
d,
(1.4) δ(Q
d) = 1
2 .
Rumely [R] has given a beautiful integral formula for the transfinite diameter using notions of Pluripotential theory (see also [DR]). However, this appears to be difficult to explicitly evaluate, at least in the case of the real ball in which we are interested. In this paper we will use Zaharjuta’s formula (2.9) directly, by finding a formula for τ (θ) and then computing its integral over the simplex, to give a compact formula (Theorem 2.5 below) for the transfinite diameter of the real euclidean unit ball B
d:= {x ∈ R
d:
|x| ≤ 1} and the real unit simplex, T
d:= {x ∈ R
d+: P
dj=1
x
j≤ 1}.
2. The Ball
We begin with some standard facts about univariate orthogonal polyno- mials.
2.1. Univariate Gegenbauer Polynomials. Let C
nλ(x), n ∈ Z
+, λ >
−1/2, denote the Gegenbauer polynomials of degree n and parameter λ, i.e., the classical univariate polynomials orthogonal on [−1, 1] with respect to the inner product
(2.1) hf, gi
λ:=
Z
1−1
f (x)g(x)w
λ(x)dx where
w
λ(x) := (1 − x
2)
λ−1/2. It is known that (see e.g. [AS])
C
nλ(x) = k
nx
n+ lower degree terms
where
(2.2) k
n:= 2
nΓ(λ + n)
n!Γ(λ) and that
Z
1−1
(C
nλ(x))
2w
λ(x)dx = π2
1−2λΓ(n + 2λ) n!(n + λ)(Γ(λ))
2. If we let b C
nλ(x) :=
k1n
C
nλ(x) denote the associated monic orthogonal poly- nomials, it follows easily that
C b
nλ(x) = x
n+ lower degree terms and that
h(n, λ) :=
Z
1−1
( b C
nλ(x))
2w
λ(x)dx
= π2
1−2(n+λ)Γ(n + 1)Γ(n + 2λ) (Γ(n + λ))
2(n + λ) . (2.3)
We will make use of the following the nth root asymptotics of h(n, λ).
Lemma 2.1. Suppose that a
n, b
n≥ 0 are such that lim
n→∞a
n/n = a and lim
n→∞b
n/n = b with a + b > 0. Then
n→∞
lim h(a
n, b
n)
1/n= a
a(a + 2b)
a+2b2
2(a+b)(a + b)
2(a+b). Proof. From (2.3) we have
h(a
n, b
n) = π2
1−2(an+bn)Γ(a
n+ 1)Γ(a
n+ 2b
n) (Γ(a
n+ b
n))
2(a
n+ b
n) .
Clearly, the terms π
1/nand (a
n+ b
n)
1/nboth tend to 1 and hence can be ignored. Further,
2
1−2(an+bn)1/n= 2
1/n−2(an/n+bn/n)→ 2
−2(a+b)and hence we are left with showing that
n→∞
lim
Γ(a
n+ 1)Γ(a
n+ 2b
n) (Γ(a
n+ b
n))
2 1/n= a
a(a + 2b)
a+2b(a + b)
2(a+b).
But this is an easy consequence of Stirling’s formula, and we are done. 2.2. A Family of Orthogonal Polynomials on B
d. For µ > −1/2 and x = (x
1, x
2, · · · , x
d) ∈ B
dlet
W
µ(x) = (1 − |x|
2)
µ−1/2be a generalized Gegenbauer weight. Now, suppose that α ∈ Z
d+is a multi- index. We define (cf. Prop. 2.3.2 of [DX]) the polynomials
P
α(x) = b C
αλ11(x
1) × (1 − x
21)
α2/2C b
αλ22( x
2p1 − x
21)
!
× (1 − x
21− x
22)
α3/2C b
αλ33
( x
3p1 − x
21− x
22)
!
× · · · ×
×
(1 − x
21− · · · − x
2d−1)
αd/2C b
αλdd
( x
dq
1 − x
21− · · · − x
2d−1)
where
λ
1= µ + α
2+ α
3+ · · · + α
d+ (d − 1)/2, λ
2= µ + α
3+ α
4+ · · · + α
d+ (d − 2)/2, λ
3= µ + α
4+ α
5+ · · · + α
d+ (d − 3)/2,
·
· λ
d= µ, that is
λ
j:= µ + (d − j)/2 +
d
X
k=j+1
α
k.
Note that since the weight w
λ(x) is symmetric, the polynomials b C
nλ(x) are even for n even and odd for n odd. Hence the P
α, as defined above, are indeed d-variate polynomials.
Lemma 2.2. The polynomials P
α(x) are orthogonal with respect to the inner product
(2.4) hf, gi
µ:=
Z
Bd
f (x)g(x)W
µ(x)dx.
Proof. This is given in Prop. 2.3.2 of [DX].
Lemma 2.3. The polynomials P
α(x) are monic, i.e., of the form P
α(x) = x
α+ lower order terms
where the ordering of the monomials is graded-lexicographic with x
1≺ x
2≺
· · · ≺ x
d.
Moreover,
Z
Bd
(P
α(x))
2W
µ(x)dx =
d
Y
j=1
h(α
j, λ
j).
Proof. The mononoticity is obvious from the construction. The norm is easily calculated by writing
Z
Bd
f (x)W
µ(x)dx = Z
Bd−1
Z
√
1−|x0|2
−
√
1−|x0|2
f (x
0, x
d)W
µ(x
0, x
d)dx
ddx
0, where x
0:= (x
1, · · · , x
d−1), substituting y
d:= x
d/p1 − |x
0|
2, and proceed-
ing by induction.
2.2.1. The Measures W
µdx are Bernstein-Markov. A measure ν on a com- pact set K ⊂ C
dis said to be a Bernstein-Markov measure when there exists constants C(n) with the property that
n→∞
lim C(n)
1/n= 1 such that for all polynomials p of degree n,
kpk
K≤ C(n)kpk
L2(K;ν).
For many purposes, including the calculation of transfinite diameters, this means that the uniform norm can be substituted by the L
2norm with respect to the measure ν.
It turns out that, for the measures W
µ(x)dx, µ ≥ 0, there is a stronger statement. Specifically, from Lemma 1 of [B2], it follows that, for α = 0, (2.5) kpk
Bd≤
s 2 ω
dn + d d
+ n + d − 1 d
kpk
L2(Bd;W0dx)for all polynomials p of degree n. Here ω
dis the surface area of the d dimen- sional unit sphere in R
d+1. In other words, for this measure C(n) = O(n
d/2) is of polynomial growth. Similarly, the considerations of [B2] show that C(n) is also of polynomial growth for all µ ≥ 0.
2.3. The Directional Chebyshev Constants. Since the measures W
µ(x)dx, µ ≥ 0 are Bernstein-Markov measures, we may use any of the associated 2-norms,
kf k
µ:= hf, f i
1/2µ,
to compute the directional Chebyshev constants (cf. [B1, p. 320]). Specifi-
cally consider θ ∈ S
d, the unit simplex in R
d(see (1.2)). Then for n = |α| =
α
1+ · · · α
dτ (θ) = lim
α/n→θ
kP
αk
1/nµ= lim
α/n→θ
(
dY
j=1
h(α
j, λ
j) )
2n1by Lemma 2.3.
Suppose now that θ ∈ S
dwith θ > 0 (i.e., each component is positive).
Then under the hypothesis that α/n → θ we have
n→∞
lim α
j/n = θ
j, 1 ≤ j ≤ d, (2.6)
n→∞
lim λ
j/n =
d
X
k=j+1
θ
k, 1 ≤ j ≤ d.
(2.7)
Hence, by Lemma 2.1
n→∞
lim h(α
j, λ
j)
1/n= θ
θjj(θ
j+ 2 P
dk=j+1
θ
k)
θj+2Pdk=j+1θk2
2(Pdk=jθk)( P
dk=j
θ
k)
2(Pdk=jθk)=: H
j(θ).
(2.8)
We thus may conclude
Lemma 2.4. For θ ∈ S
dwith θ > 0, τ (θ) =
(
dY
j=1
H
j(θ) )
1/2.
2.4. Zaharjuta’s Formula for The Transfinite Diameter. We will use Zaharjuta’s formula for the transfinite diameter, given in the introduction:
(2.9) δ(B
d) = exp
1
vol(S
d) Z
Sd
log(τ (θ))dV
. Theorem 2.5. The transfinite diameter of the unit ball B
dis:
(a) for d even, δ(B
d) = 1
2 exp − 1 4
2d + 1 d
d
X
j=1
1 j + 1
2 + 1
2 log(2) + 1 4d
d
X
j=1
(−1)
jj
! , (b) for d odd,
δ(B
d) = 1
2 exp − 1 4
2d + 1 d
d
X
j=1
1 j + 1
2 + d − 1
2d log(2) − 1 4d
d
X
j=1
(−1)
jj
! . Remark. For d = 1 the formula reduces to the classical δ([−1, 1]) = 1/2 and for d = 2 we obtain δ(B
2) = 1/ √
2e, in agreement with the result of
[B1].
Proof. We use the Zaharjuta formula together with the formula for τ (θ) given in Lemma 2.1 to obtain
δ(B
d) = exp (d − 1)!
2 (C
d+ F
d− D
d− E
d)
where
C
d=
d
X
j=1
Z
Sd
log(θ
θjj)dV = − 1 (d − 1)!
d
X
j=2
1 j ,
D
d=
d
X
j=1
Z
Sd
log
2
2Pdk=jθkdV = d + 1
(d − 1)! log(2),
E
d=
d
X
j=1
Z
Sd
log
d
X
k=j
θ
k!
2(
Pdk=jθk)
dV = − 1 2(d − 2)! , and
F
d=
d
X
j=1
Z
Sd
log
θ
j+ 2
d
X
k=j+1
θ
k!(
θj+2Pdk=j+1θk)
dV
= 1 2
1 (d − 2)!
2 log(2) − 1 + Z
21
(x − 1)
d−2x log(x) dx
− 1 2
1 d!
d
X
j=2
1 j . The values for the integrals are given in a sequence of lemmas below. Note that vol(S
d) = 1/(d − 1)!; putting them together and simplifying gives the
result.
We can express integrals over the simplex S
das univariate integrals by means of B-splines.
Lemma 2.6. Suppose that a
1≤ a
2≤ · · · ≤ a
dwith d ≥ 2 and that f ∈ L
1[a
1, a
d]. Then
Z
Sd
f (
d
X
j=1
θ
ja
j)dV = 1
(a
d− a
1)(d − 2)!
Z
∞−∞
f (x)B(x | a
1, a
2, · · · a
d)dx where
B(x | a
1, a
2, · · · a
d) := (a
d− a
1)(· − x)
d−2+[a
1, a
2, · · · , a
d] is the B-spline of degree d − 2 and knots a
1, a
2, · · · , a
d. Here
y
+:= y y ≥ 0 0 y < 0
and f [a
1, a
2, · · · , a
d] denotes the divided difference of the function f at the
points a
i.
Proof. This is a standard formula of spline theory, based on the fact that B-splines are the Peano kernel for divided differences; see, for example, [deB,
p. 88].
We will need the following fact.
Lemma 2.7. For 1 < j ≤ d and d ≥ 2, we have B(x | 0, · · · , 0
| {z }
j−1
, 1, · · · , 1
| {z }
d−j+1
) =
d−2j−2
x
d−j(1 − x)
j−2, x ∈ [0, 1]
0, otherwise .
Proof. This is again a standard fact that follows easily from the recurrence formula for B-splines (see again [deB, p. 89]). Lemma 2.8. For 1 ≤ j ≤ d, we have
Z
Sd
θ
jlog(θ
j)dV = 1 (d − 2)!
Z
1 0(1 − x)
d−2x log(x)dx
= − 1 d!
d
X
j=2
1 j . Hence
C
d=
d
X
j=1
Z
Sd
θ
jlog(θ
j)dV = − 1 (d − 1)!
d
X
j=2
1 j .
Proof. By symmetry me need only consider j = d. Then, using the integral formula of Lemma 2.6, we have
Z
Sd
θ
dlog(θ
d)dV = 1
(1 − 0)(d − 2)!
Z
∞−∞
x log(x)B(x | 0, 0, · · · , 0
| {z }
d−1
, 1)dx
= 1
(d − 2)!
Z
1 0x log(x)(1 − x)
d−2dx since, by Lemma 2.7 with j = d − 1,
B(x | 0, · · · , 0
| {z }
d−1
, 1) = (1 − x)
d−2, x ∈ [0, 1]
0, otherwise . By Lemma 2.9 with m = d − 2 we obtain
1 (d − 2)!
(
− 1
(d − 1)d
d
X
j=2
1 j
)
and the result follows.
Lemma 2.9. For m ∈ Z
+we have Z
10
(1 − x)
mx log(x)dx = − 1
(m + 1)(m + 2)
m+2
X
j=2
1
j .
Proof. This is a special case of Formula 1. of §4.253 of [GR], however, for completeness, we provide an elementary proof. Let A
mdenote the integral in question. Then integrating
A
m+1= Z
10
(1 − x)
m+1x log(x)dx.
by parts with u = (1 − x)
m+1and v
0= x log(x) we easily obtain the recur- rence
A
m+1= m + 1
m + 3 A
m− 1
(m + 2)(m + 3)
2from which it is easy to verify the stated formula by induction. Lemma 2.10. For d ≥ 2 we have
D
d=
d
X
j=1
Z
Sd
log
2
2Pdk=jθkdV = d + 1
(d − 1)! log(2).
Proof. The j = 1 term is slightly different. In fact, for j = 1, P
dk=j
θ
k= 1 and the integrand is just 2 log(2) and the integral
(2.10)
Z
Sd
log
2
2Pdk=1θkdV = 2 log(2)vol(S
d) = 2 log(2) (d − 1)! . For the other terms we compute
d
X
j=2
Z
Sd
log
2
2Pdk=jθkdV
= 2 log(2)
d
X
j=2
Z
Sd
d
X
k=j
θ
k! dV
= 2 log(2)
d
X
j=2
1
(1 − 0)(d − 2)!
Z
1 0xB(x | 0, · · · , 0
| {z }
j−1
, 1, · · · , 1
| {z }
d−j+1
)dx
= 2 log(2) (d − 2)!
d
X
j=2
Z
1 0x d − 2 j − 2
x
d−j(1 − x)
j−2dx (by Lemma 2.7)
= 2 log(2) (d − 2)!
Z
1 0x
d
X
j=2
d − 2 j − 2
x
d−j(1 − x)
j−2dx
= 2 log(2) (d − 2)!
Z
1 0x
d−2
X
j=0
d − 2 j
x
(d−2)−j(1 − x)
jdx
= 2 log(2) (d − 2)!
Z
1 0x dx = log(2)
(d − 2)! .
Combining this with (2.10) gives
d
X
j=1
Z
Sd
log
2
2Pdk=jθkdV = 2 log(2)
(d − 1)! + log(2) (d − 2)!
and the result follows.
As pointed out by an anonymous reviewer, this integral can also be computed by completely elementary means. Specifically, first observe that by symmetry,
Z
Sd
θ
jdV = Z
Sd
θ
kdV = 1
d vol(S
d) = 1 d! . Then
D
d=
d
X
j=1
Z
Sd
log
2
2Pdk=jθkdV = 2 log(2)
d
X
j=1 d
X
k=j
1 d!
= 2
d! log(2)
d
X
j=1
(d − j + 1) = 2
d! log(2) d(d + 1) 2
= d + 1
(d − 1)! log(2).
Lemma 2.11. For d ≥ 2 we have
E
d=
d
X
j=1
Z
Sd
2
d
X
k=j
θ
k! log
d
X
k=j
θ
k!
dV = − 1 2(d − 2)! . Proof. The j = 1 term is 0 since P
dk=1
θ
k= 1 and so it can be ignored.
Hence we compute
d
X
j=2
Z
Sd
2
d
X
k=j
θ
k! log
d
X
k=j
θ
k! dV
= 2
d
X
j=2
Z
Sd
d
X
k=j
θ
k! log
d
X
k=j
θ
k! dV
= 2
d
X
j=2
1 (1 − 0)(d − 2)!
Z
1 0x log(x)B(x | 0, · · · , 0
| {z }
j−1
, 1, · · · , 1
| {z }
d−j+1
)dx
= 2
(d − 2)!
Z
1 0x log(x)
d
X
j=2
B(x | 0, · · · , 0
| {z }
j−1
, 1, · · · , 1
| {z }
d−j+1
)dx
= 2
(d − 2)!
Z
1 0x log(x) dx (as in Lemma 2.10)
= 2
(d − 2)!
− 1 4
and the result follows. Lemma 2.12. For d ≥ 2 we have
F
d=
d
X
j=1
Z
Sd
θ
j+ 2
d
X
k=j+1
θ
k!
log θ
j+ 2
d
X
k=j+1
θ
k! dV
= 1 2
1 (d − 2)!
2 log(2) − 1 + Z
21
(x − 1)
d−2x log(x) dx
− 1 2
1 d!
d
X
j=2
1 j .
The univariate integral above is evaluated in Lemma 2.13.
Proof. Let I
j:=
Z
Sd
θ
j+ 2
d
X
k=j+1
θ
k!
log θ
j+ 2
d
X
k=j+1
θ
k! dV.
We must compute P
dj=1
I
j. First note that, by Lemma 2.6, I
1=
Z
Sd
θ
1+ 2
d
X
k=2
θ
k!
log θ
2+ 2
d
X
k=2
θ
k! dV
= 1
2 − 1 1 (d − 2)!
Z
∞−∞
x log(x)B(x | 1, 2, · · · , 2
| {z }
d−1
)dx
= 1
(d − 2)!
Z
∞−∞
x log(x)B(x | 1, 2, · · · , 2
| {z }
d−1
)dx, (2.11)
while
I
d= Z
Sd
(θ
d) log(θ
d)dV
= 1
1 − 0 1 (d − 2)!
Z
∞−∞
x log(x)B(x | 0, · · · , 0
| {z }
d−1
, 1)dx
= 1
(d − 2)!
Z
∞−∞
x log(x)B(x | 0, · · · , 0
| {z }
d−1
, 1)dx
Further, for 2 ≤ j ≤ (d − 1), I
j=
Z
Sd
θ
j+ 2
d
X
k=j+1
θ
k!
log θ
j+ 2
d
X
k=j+1
θ
k! dV
= 1
2 − 0 1 (d − 2)!
Z
∞−∞
x log(x)B(x | 0, · · · , 0
| {z }
j−1
, 1, 2, · · · , 2
| {z }
d−j
)dx
= 1 2
1 (d − 2)!
Z
∞−∞
x log(x)B(x | 0, · · · , 0
| {z }
j−1
, 1, 2, · · · , 2
| {z }
d−j
)dx.
Hence 1 2 I
1+
d−1
X
j=2
I
j+ 1 2 I
d= 1 2
1 (d − 2)!
Z
∞−∞
x log(x)
d
X
j=1
B(x | 0, · · · , 0
| {z }
j−1
, 1, 2, · · · , 2
| {z }
d−j
)
dx.
However, just as Lemma 2.10, it can be shown (cf. [deB]) that
d
X
j=1
B(x | 0, · · · , 0
| {z }
j−1
, 1, 2, · · · , 2
| {z }
d−j
) ≡ 1
on its support, i.e., on (0, 2) in this case. Thus 1
2 I
1+
d−1
X
j=2
I
j+ 1
2 I
d= 1 2
1 (d − 2)!
Z
2 0x log(x)dx
= 1 2
1
(d − 2)! (2 log(2) − 1).
It follows that
d
X
j=1
I
j= 1 2
1
(d − 2)! (2 log(2) − 1) + 1
2 (I
1+ I
d).
But, by (2.11),
I
1= 1 (d − 2)!
Z
∞−∞
x log(x)B(x | 1, 2, · · · , 2
| {z }
d−1
)dx
= 1
(d − 2)!
Z
2 1(x − 1)
d−2x log(x)dx and from Lemma 2.8,
I
d= − 1 d!
d
X
j=2
1 j .
Putting these together yields the result.
Lemma 2.13. For m ∈ Z
+we have Z
21
(x − 1)
mx log(x)dx = a
mlog(2) + b
mwhere
a
m:=
2
m + 1 m even 2
m + 2 m odd
and
b
m:= − (−1)
m(m + 1)(m + 2)
(
(−1)
m−
m+2
X
k=1
(−1)
k/k )
.
Proof. Let A
mdenote the integral in question. Just as before we may integrate by parts (with u = (x − 1)
mand v
0= x log(x)) to show that
A
m+1= − m + 1
m + 3 A
m+ 4
m + 3 log(2) − 2m + 5 (m + 2)(m + 3)
2and from this one can easily verify the given formula by induction. 3. The Simplex
By the mapping result of Bloom and Calvi [BC] we immediately obtain Theorem 3.1. The transfinite diameter of the simplex T
dis given by
δ(T
d) = (δ(B
d))
2.
Proof. Just note that T
dis the image of B
dunder the mapping (x
1, x
2, · · · , x
d) 7→ (x
21, x
22, · · · , x
2d).
4. Concluding Remarks
Using our formulas for δ(B
d) and the formula (1.3) for δ(B
d) we easily find that
d→∞
lim δ(B
d) δ(B
d) = √
2
showing that the real and complex balls have finitely comparable transfinite diameters.
Further, for the inscribed cube 1
√ d Q
d⊂ B
dwe have
lim
d→∞
δ(B
d) δ(
√1d
Q
d) = p
2 exp(1 − γ)
where γ is Euler’s constant, while for the superscribed cube,
d→∞