Professional Documents
Culture Documents
2302 01001
2302 01001
ON THE SPHERE
arXiv:2302.01001v1 [math.PR] 2 Feb 2023
The authors have been partially supported by the Generalitat de Catalunya (grant 2017 SGR
359) and the Spanish Ministerio de Ciencia, Innovación y Universidades (project MTM2017-83499-
P).
1
2 VÍCTOR DE LA TORRE AND JORDI MARZO
logarithmic energy points s∗ = 3, see next section for definitions. The expected
worst-case error of some random configurations was also studied in [BrSaSlWo].
Definition 2. Let (XN ) be a sequence of random N −point configurations on Sd
following some distribution and let s > d/2. As in the deterministic case, we allow
the subindex N to follow a subsequence converging to +∞. We say that (XN ) is a
sequence of QMC designs for Hs (Sd ) on average (or s-QMC designs on average) if
there exists Cd,s > 0 such that
q
(2) E[wce(XN , Hs (Sd ))2 ] ≤ Cd,s N −s/d .
With obvious changes in the proof of Lemma 23 in [BrSaSlWo], the following
property shows that if a sequence of random point configurations is a sequence of
s-QMC designs on average, for some s > d/2, then it is also of s′ -QMC designs for
d/2 < s′ < s.
Proposition 1. Given s > d/2, if E[wce(XN , Hs (Sd ))2 ] ≤ 1, then there exists a
constant Cd,s′ ,s > 0 such that
′ s′ /s d
E[wce(XN , Hs (Sd ))2 ] ≤ Cd,s′ ,s E[wce(XN , Hs (Sd ))2 ] , < s′ < s.
2
This last result allows us to define the average QMC strength, as in the deter-
ministic case, as the maximal value of s > d/2 for which (2) holds.
It was shown in [BrSaSlWo, Theorem 7] that uniform i.i.d. points on the sphere
are not s-QMC designs on average for any s > d/2. On the other hand, [BrSaSlWo,
Theorem 21,Theorem 22] shows that points from jittered sampling (i.e. uniform
i.i.d point taken with respect to an area regular partition) have average strength
d/2 + 1. Observe that in this last case the average strength matches the numerically
conjectured in [BrSaSlWo] for the related equal area points in S2 .
1.1. Harmonic ensemble. In our first result we show that points from the har-
monic ensemble have average strength d+1 2 . The mode of this distribution corre-
sponds to the Fekete points, see next section, for which it was conjectured strength
3/2 in [BrSaSlWo]. The expected worst case error of this process was previously
studied in [Hi].
Theorem 1. Let (XN ) be a sequence where XN is an N −point set drawn from the
harmonic ensemble in Sd . Observe that N must be of the form d(L) for some natural
L. Then (XN ) forms a sequence of QMC designs on average for d2 < s < d+1 2 .
Moreover
d+1 d+1
(3) lim N d E[wce(XN ; H 2 (S2 ))2 ] = +∞,
N →+∞
d+1
therefore (XN ) is not a QMC design on average if s > 2 and the average QMC
strength is d+1
2 .
For the harmonic ensemble we can deduce, from results in [BrSaSlWo, BeMaOr],
see also [Be], almost sure optimality of the worst-case error up to a logarithmic
factor.
d d+1
Corollary 1. For every M > 0 and 2 <s< 2 , there exists Cd,s,M > 0 such
that
2s
!
s d (log N ) d+1 1
(4) P wce(XN ; H (S )) ≤ Cd,s,M s ≥1− ,
Nd NM
4 VÍCTOR DE LA TORRE AND JORDI MARZO
where XN is an N -point set drawn from the harmonic ensemble. Therefore, for
fixed d2 < s < d+12 there exists Cd,s > 0 such that, with probability 1 and for N
large enough,
2s
(log N ) d+1
wce(XN ; Hs (Sd )) ≤ Cd,s s .
Nd
1.2. Spherical ensemble. The spherical ensemble is also a determinantal point
process on the sphere S2 . Applying results from [AlZa] it was shown in [Hi] that
points from the spherical ensemble are s-QMC designs on average for 1 < s < 2. One
can easily see that 2 is indeed the average strength. Observe that the mode of this
distribution is the set of elliptic Fekete points, i.e. minimizers of the logarithmic
energy. In this case there is no coincidence with the conjectured strength from
[BrSaSlWo], which was 3.
Theorem 2. Let (XN ) be a sequence where XN is an N −point set drawn from
the spherical ensemble. Then (XN ) is a sequence of QMC designs on average for
1 < s < 2, and for s ∈ (2, 3) there exists a constant C > 0 such that
N 2 E wce(XN ; Hs (S2 ))2 ≥ C,
2. Background
2.1. Riesz energy and worst-case error. The Riesz or logarithmic energy of a
set of N different points x1 , . . . , xN on the unit sphere Sd ⊂ Rd+1 is
X
Es (x1 , . . . , xN ) = fs (|xi − xj |),
i6=j
−s
where fs (r) = r for s 6= 0 and f0 (r) = − log r are, respectively, the Riesz and
logarithmic potentials. From now on, to simplify the notation, we write Es for
Es (x1 , . . . , xN ) when the set of points is clear from the context. This quantity has
QMC STRENGTH FOR SOME RANDOM CONFIGURATIONS ON THE SPHERE 5
a continuous version for measures which for the normalized surface measure σd and
0 6= s < d is
d+1
Γ d−s
d−s−1 Γ
Z Z
d 2 2
Vs (S ) = fs (|x − y|) dσ(x)dσ(y) = 2 √ .
Sd Sd π Γ d − 2s
1
wce(XN , Hs (Sd ))2 = − Ed−2s − Vd−2s (Sd )N 2 ,
(5) 2
N
while for d/2 + M < s < d/2 + M + 1, with M a positive integer,
(6)
N
1 X
wce(XN , Hs (Sd ))2 = 2 QM (xj · xi ) + (−1)M+1 (Ed−2s − Vd−2s (Sd )N 2 ) ,
N j,i=1
where
M
X (s) (d)
QM (xj · xi ) := (−1)M+1−l − 1 αl hl Pl (xj · xi ),
l=1
(d) (d)
with Pl (x) the Gegenbauer polynomial normalized by Pl (1) = 1,
Γ(x + n)
(x)n = , n ≥ 1.
Γ(x)
1 X N
D∞ (XN ) = sup χD(x,r) (xi ) − σd (D(x, r)),
d
x∈S ,r>0 N i=1
Observe that it follows from (5) that for s = (d + 1)/2 there is a direct relation
between energy, discrepancy and worst-case error
1/2
N
d+1 1 X
wce(XN ; H 2 (Sd )) = V−1 (Sd ) − 2 |xi − xj |
N i,j=0
s √
d πΓ( d2 )
(8) = D2 (XN ),
Γ( d+1
2 )
where the last equality follows from Stolarsky’s invariance principle, see [BrDi,
BrSaSlWo].
Observe that the mode of the distribution, in some sense the value that appears
most often in a set of data values sampled from this DPP, is the maximum of the
joint density
p(x1 , . . . , xd(L) ) = (1/d(L)!) det(KL (xi , xj ))1≤i,j≤d(L) ,
for x1 , . . . , xd(L) ∈ Sd . The mode points are therefore the Fekete points that maxi-
mize the absolute value of the determinant
VL (x1 , . . . , xd(L) ) = det(φi (xj ))1≤i,j≤d(L) ,
for φ1 , . . . , φd(L) a basis of the space PL (Sd ). Fekete points are called also extremal
fundamental systems, [Re]. Sloan and Womersley conjectured that they have all
positive cubature weights [SlWo] and were shown to be asymptotically uniformly
equidistributed in [MaOr, BeBoNy].
with respect to the surface measure in S2 . This shows that the mode of this point
process on the sphere is given by the minimizers of the logarithmic energy
X 1
E0 (x1 , . . . , xN ) = log ,
|xi − xj |
i6=j
The extension of this result to the Riesz s-energy for s < 4 has been recently
obtained in [DeMa]. In order to study the expected QMC strength we will use that
for x1 , . . . , xN ∈ S2 points drawn from zeros of elliptic polynomials mapped to the
sphere by the stereographic projection and s < 4, s 6= 0, 2,
21−s 2
(11) E[Es ] = N + C(s)N 1+s/2 + oN →∞ (N 1+s/2 ),
2−s
where
1 s s s s
C(s) = s
1+ Γ 1− ζ 1− .
2 2 2 2 2
In the particular case s = −2,
ζ(3) 1
(12) E[E−2 ] = 2N 2 − 8 + oN →∞ .
N N
3. Proofs
3.1. The harmonic ensemble (Theorem 1). We start with some preliminaries
in order to prove Theorem 1.
Let d2 < s < d2 + 1. From formula (5), for N = d(L) ∼ Ld points drawn from the
harmonic ensemble on the sphere we get
1
Z Z
E[wce(XN ; Hs (Sd ))2 ] = 2 KL (x, y)2 |x − y|2s−d dσ(x)dσ(y)
N Sd Sd
d(L)2
Z Z
(1+λ,λ)
= (1+λ,λ) PL (hx, yi)2 |x − y|2s−d dσ(x)dσ(y)
PL 2 2
(1) N S S d d
Cd
Z
(1+λ,λ)
= (1+λ,λ) PL (hx, ni)2 |x − n|2s−d dσ(x)
PL (1)2 Sd
Z 1
Cd (1+λ,λ) d
= (1+λ,λ) PL (t)2 (1 − t)s−1 (1 + t) 2 −1 dt,
PL (1)2 −1
where n stands for the north pole of Sd .
From the asymptotic property of the gamma function
Γ(n + α)
lim = 1, α ∈ R,
n→∞ Γ(n)nα
we get that
L + d2
(1+λ,λ) 1
PL (1)= ∼ d
Ld/2 .
L Γ( 2 + 1)
Therefore we have that for some constant Cd,s > 0,
Cd,s 1 (1+λ,λ) 2
Z
d
(13) E[wce(XN ; Hs (Sd ))2 ] = d PL (t) (1 − t)s−1 (1 + t) 2 −1 dt.
L −1
The following lemma is an extension to −1 < a < d of a result proved in
[BeMaOr] for 0 < a < d.
Proposition 2. Given −1 < a < d,
Z 1 Z ∞
1 (1+λ,λ) a a J1+λ (t)2
lim a PL (t)2 (1 − t)λ− 2 (1 + t)λ dt = 2 2 +d dt
L→∞ L −1 0 t1+a
and the last integral converges.
QMC STRENGTH FOR SOME RANDOM CONFIGURATIONS ON THE SPHERE 9
Proof. The proof is essentially the same as in [BeMaOr, Proposition 6] but with a
few changes in the last estimates. We split the integral
Z 1
(1+λ,λ) a
L−a PL (t)2 (1 − t)λ− 2 (1 + t)λ dt
−1
"Z c c
#
− cos
Z L
Z cos L 1
(1+λ,λ) a
= + + L−a PL (t)2 (1 − t)λ− 2 (1 + t)λ dt
c c
−1 − cos L cos L
This shows that points drawn from the harmonic ensemble form an s-QMC design
on average for d2 < s < d+1
2 .
To get (3) we use again (13) for s = d+1
2 and the representation of the integral
in terms of generalized hypergeometric function [ErMaObTr, p. 288]
1 d d
s d 2 −d Γ 2 + L Γ 2 + L Γ 2 + L + 1
E[wce(XN ; H (S )) ] ∼ L 2
Γ(L + 1)2 Γ d + 21 + L
d+1 1 d 1 1
× 4 F3 −L, d + L, , ; + 1, d + + L, −L + ; 1 ,
2 2 2 2 2
where all the constants depend only on d.
It is easy to see (by induction) that the quotient
(−L)n (d + L)n
,
(d + 12 + L)n (−L + 21 )n
is increasing as a function of 0 ≤ n ≤ L and therefore
L d+1
1
d+1
s d 2
X
2 n 2 n 1
N d E[wce(XN ; H (S )) ] & d
.
n=0 2 +1 n
n!
Finally, this last series diverges when L → ∞ by Gauss test taking a = (d + 1)/2,
b = 1/2 and c = a + b, because
(a)n (b)n
(c)n n! (c + n)(n + 1) 1 Cn
(a)n+1 (b)n+1
= =1+ + 2,
(a + n)(b + n) n n
(c)n+1 (n+1)!
Proof. (Corollary 1) Recall that a measure of the uniformity of a finite set of points
is the L∞ spherical cap discrepancy D∞ (XN ) where XN = {x1 , . . . , xN } ⊂ Sd . It
was proved in [BeMaOr] that for every M > 0 there exist CM > 0 such that
log N 1
P D∞ (XN ) ≤ CM d+1 ≥ 1 − M ,
N 2d N
where XN is an N -point set drawn from the harmonic ensemble.
Now it follows from the result above, formula (8) and D2 (XN ) . D∞ (XN ) that
there exists (another) CM > 0 such that
d+1 log N 1
P wce(XN ; H 2 (Sd )) ≤ CM d+1 ≥ 1 − M .
N 2d N
To get (4) it is enough to apply the interpolation result from [BrSaSlWo, Lemma
23], from which we get for d/2 < s < (d + 1)/2 a constant Cd,s > 0 such that
d+1 d+1 d+1
wce(XN ; Hs (Sd )) 2s ≤ Cd,s wce(XN ; H 2 (Sd )) if wce(XN ; H 2 (Sd )) ≤ 1, see [Be,
Section 1.5].
QMC STRENGTH FOR SOME RANDOM CONFIGURATIONS ON THE SPHERE 11
Proof. (Theorem 2) The first part is due to [Hi]. We include it for the sake of
completeness. Let s ∈ (1, 2). Then the worst-case error is given by (5). Taking
expectations and using (9) with s′ = 2 − 2s,
Ns
N s E wce(XN , Hs (S2 ))2 = − 2 E [E2−2s ] − V2−2s (S2 )N 2
N
22s Γ(s) Γ(N )
= Ns
4 Γ(N + s)
22s Γ(s)
−−−−→ ,
N →∞ 4
Γ(N ) −s
since Γ(N +s) ∼ N by the asymptotic property of the gamma function. Then
s s 2 2
N E wce(XN , H (S )) is bounded and (XN ) is a sequence of s-QMC designs on
average for s ∈ (1, 2).
Now we show that s∗ = 2. Let s ∈ (2, 3). The expression for the worst-case error
is (6) with M = 1:
N
1 X
wce(XN ; Hs (S2 ))2 = 2 Q1 (xj · xi ) + E2−2s − V2−2s (S2 )N 2 ,
N j,i=1
with
(s)
Q1 (xj · xi ) = −6α1 xj · xi ,
(2)
where we have used that P1 (x) = x. Then
(14)
N
1 (s)
X
wce(XN ; Hs (S2 ))2 = 2 −6α1 xj · xi + E2−2s − V2−2s (S2 )N 2
N j,i=1
N 2
1 (s)
X |xj − xi |
= 2 −6α1 1− + E2−2s − V2−2s (S2 )N 2
N j,i=1
2
1 h
(s)
i
= 2 3α1 E−2 − 2N 2 + E2−2s − V2−2s (S2 )N 2
N
12 VÍCTOR DE LA TORRE AND JORDI MARZO
Γ(N ) 2 2−s
. Hence, N 2 E wce(XN ; Hs (S2 ))2 is bounded below by a
since Γ(N +s) N ∼ N
positive constant.
Proof. (Theorem 3) Let s ∈ (2, 3). We have already seen in (14) that the expression
for the worst-case error is
1 h (s) i
wce(XN ; Hs (S2 ))2 = 2 3α1 E−2 − 2N 2 + E2−2s − V2−2s (S2 )N 2 .
N
Taking expectations and using (12) for E[E−2 ] and (11) with s′ = 2 − 2s,
N s h (s) i
N s E wce(XN ; Hs (S2 ))2 = 2 3α1 E [E−2 ] − 2N 2 + E [E2−2s ] − V2−2s (S2 )N 2
N
(s) 1 1
= N s−2 3α1 −8ζ(3) + o
N N
+ C(2 − 2s)N 2−s + o(N 2−s )
(s)
−8ζ(3)N s−3 + o N s−3
= 3α1 + C(2 − 2s) + o(1)
−−−−→ C(2 − 2s).
N →∞
Then N E wce(Q[XL ]; H (S2 ))2 is bounded for s ∈ (2, 3). For 1 < s ≤ 2, the
s s
result holds automatically from Proposition 1.
Now we see that the strength is s∗ = 3. Let s ∈ (3, 4). By (6) with M = 2, the
square of the worst-case error is
N
1 X
wce(XN ; Hs (S2 ))2 = 2 Q2 (xj · xi ) − (E2−2s − V2−2s (S2 )N 2 ) ,
N j,i=1
(2)
where, using that P2 (x) = 12 (3x2 − 1),
(s)
Q2 (xj · xi ) = −5α2 [3(xj · xi )2 − 1]
(s) 2 3 4
= −5α2 2 − 3|xj − xi | + |xj − xi | .
4
Then
1 h (s)
3
wce(XN ; Hs (S2 ))2 = − 5α2 2N 2
− 3E−2 + E−4
N2 i 4
− (E2−2s − V2−2s (S2 )N 2 ) .
QMC STRENGTH FOR SOME RANDOM CONFIGURATIONS ON THE SPHERE 13
Taking expectations and using (12) for E[E−2 ] and (11) with s′ = 2 − 2s,
1 h (s)
3
E wce(XN ; Hs (S2 ))2 = 2 − 5α2 2N 2 − 3E[E−2 ] + E[E−4 ]
N i 4
− (E[E2−2s ] − V2−2s (S2 )N 2 )
(
1 (s) 2 2 1 1
= 2 − 5α2 2N − 3 2N − 8ζ(3) + o
N N N
3 32 2 1 1
+ N + 64ζ(3) + o
4 6 N N
)
2−s 2−s
− C(2 − 2s)N +o N
( )
1 (s) 1 1 2−s
= 2 − 5α2 72ζ(3) + o − C(2 − 2s)N .
N N N
Therefore,
(s)
N 3 E wce(XN ; Hs (S2 ))2 = −5α2 [72ζ(3) + o (1)] − C(2 − 2s)N 3−s
(s)
−−−−→ −360ζ(3)α2 > 0.
N →∞
Then there exists a constant B > 0 such that
N 3 E wce(XN ; Hs (S2 ))2 ≥ B.
References
[AlZa] K. Alishashi, K., Zamani, M. S.: The spherical ensemble and uniform distribution of points
on the sphere. Electron. J. Probab. 20, no. 23 (2015).
[AnGuZe] Anderson, G.W., Guionnet, A., Zeitouni, O.: An introduction to random matrices.
Cambridge Studies in Advanced Mathematics, 118. Cambridge University Press, Cambridge
(2010).
[ArBeSh] Armentano, D., Beltrán, C., Shub, M.: Minimizing the discrete logarithmic energy on
the sphere: The role of random polynomials. Trans. Amer. Math. Soc., 363(6), 2955-2965
(2011).
[BeMaOr] Beltrán, C., Marzo, J., Ortega-Cerdà, J.: Energy and discrepancy of rotationally in-
variant determinantal point processes in high dimensional spheres, J. of Complex., 37, 76-109
(2016).
[Be] Berman, R.: The spherical ensemble and quasi-Monte-Carlo designs, arXiv:1906.08533v2
[math.PR] (2021).
[BeBoNy] Berman, R., Boucksom, S., Nyström, D.W.: Fekete points and convergence towards
equilibrium measures on complex manifolds, Acta Math, 207, 1-27 (2011).
[BeBoLe92] Bogomolny, E., Bohigas, O., Leboeuf, P.: Distribution of roots of random polynomials.
Phys. Rev. Lett., 68, 2726-2729 (1992).
[BeBoLe96] Bogomolny, E., Bohigas, O., Leboeuf, P.: Quantum chaotic dynamics and random
polynomials. J. Statist. Phys., 85(5), 639-679 (1996).
[BoRaVi] Bondarenko, A., Radchenko, D., Viazovska, M.: Optimal asymptotic bounds for spher-
ical designs. Ann. Math. 178(2), 443-452 (2013).
[BoHaSa] Borodachov, S., Hardin, D., Saff, E.: Discrete Energy on Rectifiable Sets, Springer New
York (2019).
[BrChCoGiSeTr] Brandolini, L., Choirat, C., Colzani, L., Gigante, G., Seri, R., Travaglini, G.:
Quadrature rules and distribution of points on manifolds. Ann. Sc. Norm. Super. Pisa Cl. Sci.
(5), 889-923 (2014).
14 VÍCTOR DE LA TORRE AND JORDI MARZO
[BrDi] Brauchart, J. S., Dick, J.: A characterization of Sobolev spaces on the sphere and an
extension of Stolarsky’s invariance principle to arbitrary smoothness. Constr. Approx., 38(3),
397-445 (2013).
[BrHe] Brauchart, J. S., Hesse, K.: Numerical integration over spheres of arbitrary dimension,
Constr. Approx. 25, no. 1, 41-71 (2007).
[BrSaSlWo] Brauchart, J.S., Saff, E.B, Sloan, I.H., Womersley, R.S.: QMC designs: optimal order
quasi Monte Carlo integration schemes on the sphere. Math. Comp. 83, no. 290, 2821-2851
(2014).
[BrGrKuZi] Brauchart, J.S., Grabner, P.J., Kusner, W., Ziefle, J.: Hyperuniform point sets on
the sphere: probabilistic aspects. Monatsh Math 192, 763-781 (2020).
[DeMa] de la Torre, V., Marzo, J.: Expected energy of zeros of elliptic polynomials,
arXiv:2211.07599 (2022).
[DL] NIST Digital Library of Mathematical Functions. http://DL.nist.gov/, Release 1.1.6 of 2022-
06-30.
[ErMaObTr] Erdélyi, A., Magnus, W., Oberhettinger, F., Tricomi, F.G.: Tables of integral trans-
forms. Vol. II. McGraw-Hill Book Company, Inc., New York-Toronto-London (1954).
[He] Hesse, K.: A lower bound for the worst-case cubature error on spheres of arbitrary dimen-
sion, Numer. Math. 103, no. 3, 413-433 (2006).
[HeSl] Hesse, K., Sloan, I. H.: Optimal lower bounds for cubature error on the sphere S2 , J.
Complexity 21, no. 6, 790-803 (2005).
[Hi] Hirao, M.: QMC designs and determinantal point processes, Monte Carlo and quasi-Monte
Carlo methods, 331-343, Springer Proc. Math. Stat., 241, Springer, Cham (2018).
[HoKrPeVi] Ben Hough, J., Krishnapur, M., Peres, Y., Virág, V.: Zeros of Gaussian Analytic
Functions and Determinantal Point Processes. American Mathematical Society, Providence,
RI (2009).
[Kr] Krishnapur, M.: From random matrices to random analytic functions. Ann. Probab.,
37(1):314-346 (2009).
[Ma] Macchi, O.: The coincidence approach to stochastic point processes. Advances in Appl.
Probability, 7:83-122 (1975).
[MaOr] Marzo, J., Ortega-Cerdà, J.: Equidistribution of the Fekete points on the sphere, Constr.
Approx., 32, 3, 513-521 (2010).
[PePe] Pemantle, R., Peres, Y.: Concentration of Lipschitz Functionals of Determinantal and
Other Strong Rayleigh Measures. Combinatorics, Probability and Computing, 23(1), 140-160
(2014).
[Re] Reimer, M.: Multivariate Polynomial Approximation, Volume 144, Springer (2003).
[SlWo] Sloan, I.H., Womersley, R.S.: How good can polynomial interpolation on the sphere be?
Advances in Computational Mathematics 14, 195-226 (2001).
[SmSh] Smale, S., Shub, M.: Complexity of Bézout’s theorem III. Condition number and packing.
J. Complexity 9, no. 1, 4-14. Festschrift for Joseph F. Traub, Part I. (1993).
[Sz] Szegö, G.: Orthogonal polynomials, American Mathematical Society, Colloquium Publica-
tions, vol. 23 (1939).