MULTIDIMENSIONAL PSEUDO LEJA SEQUENCES

DIMITRI JORDAN KENNE

Abstract. The one-dimensional pseudo Leja sequences introduced in [5], as an alternative to

Leja sequences, provides us with good interpolation nodes for the approximation of holomor-
phic functions. We propose a definition of multidimensional pseudo Leja sequences associated
to a compact set K of the complex space Cp which generalises both the one-dimensional ver-
arXiv:2303.11871v2 [math.CV] 5 Dec 2023

sion and the multidimensional Leja sequences. We show that these sequences can be used
to calculate the transfinite diameter of K. We also present a relation to the pluricomplex
Green function associated to K. Subsequently, we show that the intertwinning of pseudo Leja
sequences is still a pseudo Leja sequence. We give a method to compute pseudo Leja sequences
with the help of discrete meshes.

1. Introduction
We are interested in the computation of goods points for approximation by multivariate poly-
nomial interpolation. Here, by good points, we mean arrays of interpolation nodes in a regular (
see below) compact set K for which the corresponding sequence of Lagrange interpolation poly-
nomials of any holomorphic function f on a neighbourhood of K, converges uniformly to f on
K. In the univariate case, a characterisation of such points is very well-know ([7], [8], [18]). In
contrast, in the multivariate case, no characterisation is known and very few examples of such
good points are currently available. The only general example is given by arrays of Fekete points
(Fekete points for K ⊂ Cp are points which maximises the (modulus) of the Vandermonde de-
terminant of a given order, see below). Unfortunately, these points are impossible to determine
explicitly and even the numerical computation of their discrete version is complicated, so that
they are essentially of theoretical interest. Apart from the Fekete points, the sole examples of
good points are furnished by the intertwining of univariate good points ([5], [4], [23]) and special
remarkable configurations such as the famous Padua points in the unit square of R2 ( [10], [13],
[14]). The interest of other configurations is supported by numerical experiments (see e.g. [1],
[12]).
In the univariate case, while the nature of good points is completely understood from the
theoretical point of view, the computation of such points for an arbitrary regular compact set is
far from being easy. A way of constructing univariate good points were introduced in [5] with
the notion of pseudo Leja sequences. Such a sequence is defined recursively as follows. The first
point ξ0 ∈ K is chosen arbitrarily and, if the first j points ξ0 , . . . , ξj−1 are already obtained, then
ξj is chosen in K to ensure that
j−1
Y j−1
Y
(1) Mj |ξj − ξi | ≥ max |z − ξi |.
z∈K
i=0 i=0

Date: December 2023.

2000 Mathematics Subject Classification. Primary 32E30, 41A05, 41A63.
Key words and phrases. Lagrange interpolation, Leja sequences, transfinite diameter, Green function, admis-
sible mesh, intertwining sequence.
1
2 DIMITRI JORDAN KENNE

Here (Mj ) is a sequence of real numbers Mj ≥ 1, j ≥ 1 which is required to be of sub-exponential

growth. The classical Leja points correspond to the special case, Mj = 1, j ≥ 1. The point is
that the possibility Mj > 1 alleviates the computation of the points without weakening the
approximation quality of the Leja points : it is shown in [5, Theorem 1, page 56] that they are
good points (for reasonable compact sets) in the sense above. The strong result that, on a set
of positive logarithmic capacity, the Lebesgue constants of one-dimensional pseudo Leja points
with bounded sequence (Mj ) are subexponential has been recently established in [24].
It is readily seen that the product in (1) can be replaced by the (modulus) of a Vandermonde
determinant and this gives the way to the extension of the definition to the higher dimensional
case. The study of this generalisation is the subject of the present paper. The precise definition
of a multivariate pseudo Leja sequence is given in Section 2 . Although we are not able to show
that our multivariate pseudo Leja points are good points (in the multivariate case, this is not
even known for ordinary Leja sequence), we prove that they permit to recapture the multivariate
transfinite diameter of K, which, in turn, by deep results due to Berman, Boucksom and Nyström
([3, Theorem A, page 3 ]) and Bloom ([9, Corollary 4.5, page 1563]), implies that they permit to
recover the multivariate equilibrium measure and the pluri-complex Green function for a regular
compact set K. This is usually regarded as an indicator that the points should actually be good
in the above sense, especially because these conditions are equivalent in the univariate case (see
[18, Theorem 1, page 65] or [8, Theorem 1.5, page 446]).
In Section5, we will briefly explain how to compute pseudo-Leja points, showing that discrete
Leja points can be seen as examples of pseudo-Leja points.
We will also prove an interesting stability property of multivariate pseudo Leja sequences.
Roughly speaking, by intertwining, in some specific way, the points of a pseudo Leja sequence for
K1 ∈ Cp1 and that of a pseudo Leja sequence for K2 ∈ Cp2 , we obtain a pseudo Leja sequence
for the product K1 × K2 , see 4.
For the convenience of the reader, we will now recall standard notations and basic definitions
on multivariate interpolation. Multi-indices will be ordered according to the graded lexicographic
order, so that α ≺ β if |α| < |β| or |α| = |β| and the left-most non zero element of α − β is
negative. We arrange the elements of Np in an increasing sequence κ = κ(p) (with respect to ≺),
(2) κ = κ(p) : N → Np .
Thus, when p = 2, we have
κ(0) = (0, 0) ≺ κ(1) = (0, 1) ≺ κ(2) = (1, 0) ≺ κ(3) = (0, 2)
≺ κ(4) = (1, 1) ≺ κ(5) = (2, 0) ≺ κ(6) = (0, 3) ≺ . . . .
The monomials of p complex variables will be denoted as follows
(3) eN (z) = z κ(N ) = z1α1 · · · zpαp , κ(N ) = (α1 , . . . , αp ).
We will use the space of polynomials spanned by the first N monomials :
PN = span{ei : i = 0, . . . , N − 1}, N ≥ 1.
The dimension of the classical space of polynomial of total degree not bigger than d is of dimension
 
p+d
hd = ,
d
thus
N = hd − 1 =⇒ PN = span{z → z α : |α| ≤ d}.
The Vandermonde determinant of a set of the points ξ0 , . . . , ξq in Cp is defined as
(4) vdm(ξ0 , . . . , ξq ) = det [eN (ξM )]0≤M,N ≤q .
 MULTIDIMENSIONAL PSEUDO LEJA SEQUENCES 3

In particular, vdm(ξ0 ) = 1. Observe that its modulus is independent of the ordering of the
points. The Vandermonde determinant is a polynomial in the coordinates of the points ξj , thus
a polynomial in p(q + 1) variables. As such, its degree for q = hd − 1 is given by
d  
X p+d
(5) ld = i(hi − hi−1 ) = p .
i=1
p+1

Given a set of N points, ΩN = {ζN 0 , . . . , ζN N } ⊂ Cp , with a non-zero Vandermonde determi-

nant, we can form the Fundamental Lagrange Interpolation Polynomials (FLIP)
(N ) vdm(ζN 0 , . . . , ζN (j−1) , z, ζN (j+1) . . . , ζN N )
(6) lj (z) = , j = 0, . . . , N.
vdm(ζN 0 , . . . , ζN N )
(N )
Observe that lj is an element of PN +1 . For a function f defined at the points in ΩN ,
N
(N )
X
(7) LΩN f (z) = f (ζN j )lj (z)
j=0

is the Lagrange Interpolation Polynomial (LIP) of f in PN +1 at the points of ΩN . In

particular,
(8) f ∈ PN +1 =⇒ f = LΩN f (z).
In the whole paper the compact sets K considered will be assumed to be (polynomially) deter-
mining. This means that no nonzero polynomial vanishes on K. Equivalently, K is not included
in an algebraic set. This property can also be expressed in terms of Vandermonde determinants.
Namely, setting
(9) Vj (K) = max |vdm(ξ0 , . . . , ξj−1 )|, j ≥ 1,
ξ0 ,...,ξj−1 ∈K

then K is determining if and only if no Vj (K) vanishes. The sufficiency follows from an ap-
plication of (8). To see the converse, suppose VN −1 (K) 6= 0 for 1 ≤ j < N but VN (K) =
0 and let {ξ0 , . . . ξN −1 } ⊂ K such that vdm(ξ0 , . . . ξN −1 ) 6= 0. The polynomial QN (z) =
vdm(ξ0 , . . . , ξN −1 , z) vanishes on K but is not the zero polynomial since its z κ(N ) coefficient
is not null so that K is not determining.
If it is necessary to specify the dimension of the ambient space, we will add a superscript to
(p)
the above objects thus writing for instance k (p) (N ), PN , and so on.
For survey on the actual state of knowledge on multivariate polynomial interpolation we refer
to [7].

2. Pseudo Leja sequences

Definition 2.1. Let K be a determining compact subset in Cp . On says that (ξj ) is a pseudo
Leja sequence in K if there exists a sequence of real numbers (Mj )j≥1 ⊂ [1, +∞) satisfying:
(10) Mj |vdm(ξ0 , . . . , ξj−1 , ξj )| ≥ max |vdm(ξ0 , . . . , ξj−1 , z)| for any j ≥ 1,
z∈K

with the growth condition,

(11) lim ( max Mj )1/d = 1.
d→+∞ hd−1 ≤j<hd

We say that L is a pseudo-Leja sequence of Edrei growth (Mj )j≥1 .

4 DIMITRI JORDAN KENNE

In the case p = 1, this definition coincides with that given in [5]. A Leja sequence is a
pseudo Leja sequence of Edrei growth 1. The fact that K is determining immediately implies
the existence of pseudo Leja sequence with non vanishing Vandermonde determinant. Observe
that we may with very limited loss of generality, assume that (Mj ) is non-decreasing and, in that
case, (11) reduces to a somewhat less clumsy condition.
Observe that if (ξj ) is a pseudo Leja sequence for K then
(12) |vdm(ξ0 , . . . , ξj−1 , ξj )| > 0, for all j.
This follows from the same reasoning sketched in the introduction : if j was the smallest integer
for which |vdm(ξ0 , . . . , ξj−1 , z)| = 0 on K then P (z) = vdm(ξ0 , . . . , ξj−1 , z) would be a non zero
polynomial vanishing on K thus contradicting the fact that it is determining. The non vanishing
of the Vandermonde determinant, implies, see (7), the following proposition.
Proposition 2.1. Let (ξi ) a pseudo Leja sequence as above. Lagrange interpolation at ΩN =
{ξ0 , . . . , ξN }, N ∈ N, is always possible.
We immediately draw of a consequence of the growth condition that will be used in the proof
of our main theorem below.
Lemma 2.2. If {Mj }j≥1 is a sequence of real numbers satisfying Property (11) from the defini-
tion above, then
 hY
d −1
1/ld
lim Mj = 1,
d→+∞
j=1

where ld is defined in (5).

Proof. Notice that
hY
d −1
 
hd hd 1 p+1
1≤ Mj ≤ ( max Mj ) and = .
j=1
j=1,...,hd −1 ld d p

Choose the smallest j(d) such that

max Mj = max Mj .
j=1,...,hd −1 hj(d)−1 ≤j<hj(d)

Since it is a non-decreasing sequence of integers, (j(d))d≥1 tends to +∞ or is eventually constant.

We examine both cases.
If j(d) → ∞, then, by (11),
lim ( max Mj )1/j(d) = 1,
d→+∞ hj(d)−1 ≤j<hj(d)

hence, since 0 < j(d) ≤ d,

lim ( max Mj )1/d = 1.
d→+∞ hj(d)−1 ≤j<hj(d)

If j(d) is eventually constant, the conclusion is even simpler since (Mj ) is then bounded and
hd /ld → 0. 

We now state our main theoretical result and a consequence. A similar result was obtained
in [20] and [8] for Leja sequences. The method of proving Theorem 2.3 is in fact not different
from that of Leja sequences except that here we take into consideration the Edrei growth of
the pseudo Leja sequence. Its proof is postponed to the next section where we will recall the
technical points that we will need.
 MULTIDIMENSIONAL PSEUDO LEJA SEQUENCES 5

Theorem 2.3. Let K ⊂ Cp be a determining compact set. If (ξj )j≥0 ⊂ K is a pseudo-Leja

sequence of Edrei growth (Mi )j≥1 , then
(13) lim |vdm(ξ0 , . . . , ξhd −1 )|1/ld = D(K),
d→+∞

where D(K) denotes the multivariate transfinite diameter.

Thanks to R. Berman and S. Boucksom outstanding result which first appeared in the preprint
[2, Theorem 1.1] and later was presented as a consequence of a more general convergence result
in [3, Theorem A] with D. W. Nyström, we can derive the following.
Theorem 2.4. Let K ⊂ Cp be a determining compact set. If (ξj )j≥0 ⊂ K is a pseudo-Leja se-
Phd
quence, then the normalised counting measures h−1
d j=0 [ξj ] converge to the equilibrium measure
of K in the weak-* sense.
Here [ξ] refers to the Dirac measure at ξ of total mass 1.
Setting
vdm(ξ0 , . . . , ξk−1 , ξ)
(14) Pk (ξ) = ,
vdm(ξ0 , . . . , ξk−1 )
we will see that
hY
d −1

1/ld
(15) lim kPk kK = D(K),
d→∞
k=0

and, by a strong result of Bloom ([6, Corollary 4.5]), we will obtain the following.
Theorem 2.5. Let K be a polynomially convex regular compact subset of Cp . Then
 
1 |PN (z)|
(16) GK (z) = lim sup log , z ∈ Cp \ K,
N →+∞ |κ(N )| kPN kK
where GK denotes the pluri-complex Green function of K.
For the notions of pluricomplex potential theory used in the above statement (regularity,
Green function, equilibrium measure), we refer the reader to [21].
Observe that in the case where K would be non determining, a pseudo Leja sequence would
be eventually arbitrary (as soon as Vj (K) = 0), and Theorem 2.3 would be trivially true (with
D(K) = 0).

3. Transfinite diameter and proof of Theorem 2.3

We will recall in details the objects and results related to the transfinite diameter. We follow
the presentation given in [8]. As usual, K denotes a determining compact set in Cp . We will write
Vj instead of Vj (K), see (9), when there is no ambiguity. We define the d-th order transfinite
diameter of K, denoted by Dd (K) = Dd , as
1/l
Dd = Vhd d .
Fekete proved in [17] that the limit (as d → ∞) of the sequence Dd , denoted as D(K), exists
for any compact set K ⊂ C. Later in [22], Leja introduced the name transfinite diameter and
posed the problem of its existence in the multivariate case. A positive answer to his problem
was given by Zaharjuta in the celebrated paper [25] where it also shown a remarkable connection
with some certain polynomials of minimal norm that we will now describe.
6 DIMITRI JORDAN KENNE

A monic polynomial is a polynomial whose leading term with respect to ≺ is equal to 1. Thus
a monic polynomial of leading term ei is of the form
X
(17) P (z) = ei (z) + cj ej (z) : cj ∈ C}, i ∈ N.
0≤j<i

We will indicate such a form in writing lead(P ) = ei . We define the i-th Chebyshev constant
τi of K as
n o
1/|κ(i)|
(18) τi = inf kP kK : lead(P ) = ei .
Given θ = (θ1 , . . . , θp ) in the standard p-simplex Σ,
p
X
Σ = θ ∈ Rn :

θi = 1, θi ≥ 0, i = 1, . . . , p ,
i=1
the limit
(19) τ (K, θ) = lim sup τj
κ(j)
j→+∞, |κ(j)| →θ

is called the directional Chebyshev constant of K in the θ-direction. Now, if we denote by

Td , the geometric mean of Chebyshev constants for a given degree,
 Y 1/(hd −hd−1 )
(20) Td = τi ,
|κ(i)|=d

the fundamental result of Zaharjuta can be stated as follows.

Theorem 3.1. [25, Theorem 1] The sequence Td converges as d → ∞ and we have
 
1
Z
(21) lim Td = exp log(τ (K, θ)) dθ = D(K)
d→∞ meas(Σ) Σ
where meas(Σ) denotes the standard volume of the simplex.
We will need a “mean modification” of the above.
Corollary 3.2. We have
d
Y r 1/ld
(22) lim Tj j = D(K)
d→∞
j=1

where rj = j(hj − hj−1 ) for j ≥ 1.

The corollary follows from Theorem 3.1 by applying the following classical lemma of Cesaro
type (for which we omit the proof) with ud = log(Td ) and θd,j = rj /ld .
Lemma 3.3. Let (ud ) be a sequence of real numbers converging to A. Consider an array of non
negative numbers θd,j , d, j ∈ N⋆ verifying
d
X
lim max θd,j = 0 and θd,j = 1, d ∈ N.
d→+∞ j=1....,d
j=1

Then the sequence

d
X
sd = θd,j uj , d ∈ N,
j=1
also converges to A.
 MULTIDIMENSIONAL PSEUDO LEJA SEQUENCES 7

We have now assembled the material required in our proof of Theorem 2.3.
Proof of Theorem 2.3. Set Lk = |vdm(ξ0 , . . . , ξk−1 )| for all k ≥ 1. In view of (9) and (12), we
have
(23) 0 < Lk ≤ Vk , k ∈ N.
Recall that the polynomial Pk defined in (14) is given by
vdm(ξ0 , . . . , ξk−1 , ξ) X
(24) Pk (ξ) = = ek (ξ) + cj ej (ξ),
vdm(ξ0 , . . . , ξk−1 )
0≤j<k

Observe that
Pk ∈ P k , leadPk = ek , deg(Pk ) = deg(ek ) = κ(k).
By definition of a pseudo Leja sequence, we have
Lk+1 M −1 maxξ∈K vdm(ξ0 , . . . , ξk−1 , ξ)
|Pk (ξk )| = ≥ k ≥ Mk−1 kPk kK .
Lk vdm(ξ0 , . . . , ξk−1 )
Hence by definition of τk , see (18), we have
Lk+1 |κ(k)|
(25) ≥ Mk−1 kPk kK ≥ Mk−1 τk .
Lk
By using (23) and iterating the above inequality, we deduce that, for d ≥ 0
L hd Lh −1 L2
(26) Vhd ≥ Lhd = × d × ··· ×
Lhd −1 Lhd −2 L1
h hY
d −1 i−1 h hY
d −1 i
|κ(j)|
(27) ≥ Mj τj .
j=1 j=1

We may re-arrange the terms in the last product above taking into account the definition of Ti
in (20) as follows :
hY
d −1 d h hY
i −1 ii Yd
[κ(j)| i(h −h )
Y
τj = τj = Ti i i−1 .
j=1 i=1 j=hi−1 i=1

Now, taking the ld -roots in (27), we obtain

h hY
d −1 d
i−1/ld h Y i1/ld
i(hi −hi−1 )
(28) (Vhd )1/ld ≥ (Lhd )1/ld ≥ Mj Ti .
j=1 i=1

Now, passing to the limit and using Corollary 3.2 and Lemma 2.2, we obtain
lim (Lhd )1/ld = D(K). 
d→∞

As to prove Relation (15), we proceed as follows. From Inequalities (25) and (28), we get
h hY
d −1 d
ih Y i hY
d −1
i(hi −hi−1 )
Mj Ti ≤ kPk kK ≤ Vhd .
j=1 i=1 k=0

Then, we take the ld −roots and pass to the limit as d → ∞ which leads to (15).
Remark 1. Due to inequalities (23) and (25) from the above proof, we obtain a stronger version
of Theorem 2.3, namely
1/ld(k)
(29) D(K) = lim Lk , where d(k) = |κ(k)|.
k→+∞
8 DIMITRI JORDAN KENNE

4. Intertwining pseudo Leja sequences

We show how to construct a pseudo Leja sequence on Cp = Cp1 × Cp2 out of two pseudo
Leja sequences on Cpi , i = 1, 2. To this aim, we use a natural process introduced by Calvi that
generalises a method introduced by Biermann in the case pi = 1, see [15]. As indicated in the
introduction, in this section, when necessary, we use a superscript to indicate the ground spaces
(p)
for the related object, so that ej denotes the j-th monomial on Cp and so on.
We will use
ϕ : N ∋ j 7−→ ϕ(j) = ϕ1 (j), ϕ2 (j) ∈ N2


(30)
satisfying, see the introduction for the definition of κ,


κ(p) (j) = κ(p1 ) ϕ1 (j) , κ(p2 ) ϕ2 (j) ,


(31) j ∈ N.

Definition 4.1. Let A = (aj )j∈N ⊂ Cp1 and B = (bj )j∈N ⊂ Cp2 two sequences of points. The
intertwining sequence of A and B is the sequence Ω = (ωj )j∈N of points in Cp , p = p1 + p2 ,
denoted by A ⊕ B and defined by


(32) ωj = aϕ1 (j) , bϕ2 (j) ) , j ∈ N.
The important point for our purpose is that we can easily express the vandermondian for an
intertwining sequence Ω = A ⊕ B in terms of the factor sequences A and B. This is specified
below.
We say that X = (x0 , . . . , xN −1 ) is completely ordered unisolvent if
(33) vdm(x0 , . . . , xi−1 ) 6= 0, for all i = 1, . . . , N.
Every section (ξj : j = 0, . . . , N − 1) of a pseudo Leja sequence (ξj )j∈N is completely ordered
unisolvent (the compact sets being assumed determining). In view of [15, Theorem 3.3], the
notion of completely ordered unisolvent array is stable by intertwining.
Given a sequence X = (xj )j∈N , we set Xj = (x0 , . . . , xj−1 ). The following result is a conse-
quence of [15, Proposition 2.5].
Proposition 4.1. Let X = (x0 , . . . , xN −1 ) be a completely ordered unisolvent array. We have
N
Y −1
(34) vdm(x0 , . . . , xN −1 ) = (ej − LXj ej )(xj )
j=0

where LXj is the Lagrange interpolation operator at the points Xj = (x0 , . . . , xj−1 ) with the
convention that LX0 f = 0 and ej is the j-th monomial, see (3).
Proof. The proof follows the reasoning used to prove [15, Proposition 2.5]. 
The next proposition follows from [15, Theorem 4.2].
Proposition 4.2. Let A, B and Ω = A ⊕ B as above. We have
 
(p) (p) (p ) (p )
(p ) (p )

(35) ej − LΩj ej (z) = eϕ11(j) − LAϕ1 (j) eϕ11(j) (z 1 ) · eϕ22(j) − LBϕ2 (j) eϕ22(j) (z 2 )

for all z = (z 1 , z 2 ) ∈ Cp1 × Cp2 and for all j ∈ N.

We now combine Proposition 4.1 and Proposition 4.2 to deduce the following result. Obverse
that it involves three forms of vandermondians, with variables ωi in Cp , ai in Cp1 , and bi in Cp2 .
Theorem 4.3. Let A, B and Ω = A ⊕ B as above. If, for all j ∈ N⋆ , Aj and Bj are completely
ordered unisolvent arrays, then for all j ∈ N⋆ , we have
(36) vdm(ω0 , . . . , ωj−1 , z) = Pj (z) · vdm(ω0 , . . . , ωj−1 ),
 MULTIDIMENSIONAL PSEUDO LEJA SEQUENCES 9

where, for z = (z 1 , z 2 ),
(p) (p)
(37) Pj (z) = (ej − LΩj ej )(z)
(p ) (p )
(p ) (p )

(38) = eϕ11(j) − LAϕ1 (j) eϕ11(j) (z 1 ) · eϕ22(j) − LBϕ2 (j) eϕ22(j) (z 2 )
Hence, if ϕi (j) > 0 for i = 1, 2,
vdm(a0 , . . . , aϕ1 (j)−1 , z 1 ) vdm(b0 , . . . , bϕ2 (j)−1 , z 2 )
(39) Pj (z) = · .
vdm(a0 , . . . , aϕ1 (j)−1 ) vdm(b0 , . . . , bϕ2 (j)−1 )
While if, say, ϕ1 (j) > 0 and ϕ2 (j) = 0,
vdm(a0 , . . . , aϕ1 (j)−1 , z 1 )
(40) Pj (z) = .
vdm(a0 , . . . , aϕ1 (j)−1 )
Now, we state the main result of this section.
Theorem 4.4. Let Ki , i = 1, 2, be a determining compact subset in Cpi . Let A = (aj )j∈N ⊂ K1 ,
B = (bj )j∈N ⊂ K2 and Ω = A ⊕ B. The following assertions are equivalent:
(1) Ω is a pseudo Leja sequence for K = K1 × K2 .
(2) A and B are pseudo Leja sequences for K1 and K2 respectively.
Proof. We prove (1) =⇒ (2). Supposing that Ω is a pseudo Leja sequence for K of Edrei growth
(Mj )j≥1 . We prove that A is a pseudo Leja sequence for K1 . Let i ∈ N and N ∈ N such that
ϕ(N ) = (i, 0). From the definition of a pseudo Leja sequence, we have
MN |vdm(ω0 , . . . , ωN −1 , ωN )| ≥ max |vdm(ω0 , . . . , ωN −1 , z)|.
z∈K

Applying Theorem 4.3 to both sides of the above inequality, we obtain

MN |PN (ωN )| ≥ max |PN (z)|.
z∈K

Since ωN = (ai , b0 ), in view of (40), the above inequality reduces to

MN |vdm(a0 , . . . , ai−1 , ai )| ≥ max
1
|vdm(a0 , . . . , ai−1 , z 1 )|.
z ∈K1

Therefore, the sequence A will satisfy the inequality property of a pseudo Leja sequence on
condition that the growth sequence
Mi′ = MN , ϕ(N ) = (i, 0)
satisfies the Edrei growth condition in (11), that is
lim max (Mj′ )1/d = 1,
d→+∞ hd−1 ≤j<hd
(p )
where hs = hs 1 is the dimension of the space of polynomial of degree s on the ground space of
A, that is Cp1 . Let d ≥ 1 and let
Mi = Mi′ = max Mj′ .
1 (p ) (p1 )
hd−1 ≤j<hd

As previously, we consider the number N = N (i) such that ϕ(N ) = (i, 0). Therefore, κ(p) (N ) =
(p) (p)
(κ(p1 ) (i), 0) and |κ(p) (N )| = |κ(p1 ) (i)| = d. It follows that N ∈ {hd−1 , . . . , hd − 1} and hence
Mi ≤ max Mj .
(p) (p)
hd−1 ≤j<hd

Taking the 1/d root and passing to the limit as d → ∞, we obtain the result, taking into account
that Mi ≥ 1, thanks to the growth condition of pseudo Leja sequences. Thus, A is a pseudo
10 DIMITRI JORDAN KENNE

Leja sequence of Edrei growth (Mi′ )i≥1 . Of course, the same reasoning shows that B is also a
pseudo Leja sequence for K2 .
We now prove (2) =⇒ (1). We assume that A is a pseudo Leja sequence for K1 of Edrei
growth (Mj′ )j≥1 and B is a pseudo Leja sequence for K2 of Edrei growth (Mj′′ )j≥1 . We show
that Ω = A ⊕ B is a pseudo Leja sequence for K.
Fix j ∈ N⋆ . If ϕi (j) > 0, i = 1, 2, from Theorem 4.3, we have
|vdm(ω0 , . . . , ωj−1 , ωj )| = |Pj (ωj )| · |vdm(ω0 , . . . , ωj−1 )|,
where
vdm(a0 , . . . , aϕ1 (j)−1 , aϕ1 (j) ) vdm(b0 , . . . , bϕ2 (j)−1 , bϕ2 (j) )
Pj (ωj ) = · .
vdm(a0 , . . . , aϕ1 (j)−1 ) vdm(b0 , . . . , bϕ2 (j)−1 )
In view of (40), one readily check that the above formula remains true when ϕi (j) = 0, i = 1 or
2. Using the hypothesis that the factor sequences are pseudo Leja sequences, we have
vdm(a0 , . . . , aϕ1 (j)−1 , z 1 ) vdm(b0 , . . . , bϕ2 (j)−1 , z 2 )
Mϕ′ 1 (j) Mϕ′′2 (j) |Pj (ωj )| ≥ max | | · max | |
1z ∈K1 vdm(a0 , . . . , aϕ1 (j)−1 ) z 2 ∈K2 vdm(b0 , . . . , bϕ2 (j)−1 )
= max |Pj (z)|.
z∈K

Therefore,
Mj |vdm(ω0 , . . . , ωj )| ≥ max |Pj (z)vdm(ω0 , . . . , ωj−1 )| = max |vdm(ω0 , . . . , ωj−1 , z)|
z∈K z∈K

where Mj = Mϕ′ 1 (j) Mϕ′′2 (j) with the convention that M0′ = M0′′ = 1. It remains to prove that
1/d
(41) lim max Mj = 1.
d→+∞ h(p) ≤j<h(p)
d−1 d

(p) (p)
Fix d ≥ 1. Let jd ∈ {hd−1 , . . . , hd − 1} such that
Mjd = max Mj .
(p) (p)
hd−1 ≤i<hd

From the definition of ϕ, see (2), we have

κ(p) (j) = κ(p1 ) (ϕ1 (j)), κ(p2 ) (ϕ2 (j))


(42)
(43) =⇒ d = |κ(p) (j)| = |κ(p1 ) (ϕ1 (j))| + |κ(p2 ) (ϕ2 (j))|.
Let δi (d) = |κ(pi ) (ϕi (jd ))| for i = 1, 2. Now we have
 |δ1 (d)|/d  |δ2 (d)|/d
1/d
Mjd = (Mϕ′ 1 (jd ) )1/|δ1 (d)| × (Mϕ′′2 (jd ) )1/|δ2 (d)| .

We consider two cases. If say δ1 (d) is bounded then, in view of (43), δ2 (d)/d → 1, and (41)
follows from the fact that δ1 (d)/d → 0 and that, B being a pseudo Leja sequence, M ′′ satisfies
the required growth condition. Since, by (43), both δi (d)/d cannot be bounded, it remains only
to the study the case for which they are unbounded hence, since they are increasing, tending to
∞. In that case, (41) follows from the fact that M ′ and M ′′ satisfies the Edrei growth condition
and that δi (d)/d is bounded by 1, i = 1, 2. 
Theorem 4.4 is a generalisation of Irigoyen’s recent result in [19, Theorem 1 ] about the
intertwining of Leja sequences when p1 = p2 = 1. It is well known that the intertwining of two
good univariate sequences of interpolation points (good in the sense given in the introduction)
is also good for the interpolation of holomorphic functions in C2 . This property was first proved
by Siciak in [23] but can also be found in [4], [5] where different proofs are proposed.
 MULTIDIMENSIONAL PSEUDO LEJA SEQUENCES 11

5. Computing pseudo-Leja points

In practice, we compute finitely many points, even the first few points, see below, of a pseudo-
Leja sequences. Such a computation is based on the following observation which shows how to
obtain pseudo-Leja points on K as Leja points for a finite subset of K.
Proposition 5.1. Let A a finite subset of K ∈ Cp that is determining for the space of polynomial
(p)
of degree ≤ d (hence Card(A) ≥ hd = hd ). We define inductively hd points as follows.
(1) We choose arbitrarily ξ0 ∈ A and,
(2) for each d ∈ N, we compute recursively the points ξhd−1 , . . . , ξhd −2 and ξhd −1 from A
such that
(44) |vdm(ξ0 , . . . , ξN −1 , ξN )| = max |vdm(ξ0 , . . . , ξN −1 , ξ)|,
ξ∈A

for hd−1 ≤ N < hd .

Then the points ξj forms the first hd points of a pseudo-Leja sequences for K of (constant) Edrei
growth MA where
MA = max{1/kP kA, kP kK = 1, deg(P ) ≤ d}.
These points are called discrete Leja points for A.
Proof. Observe that the finiteness of MA follows from the facts that A is determining for the
space of polynomial of degree ≤ d and that the set of polynomials of bounded degree d and sup
norm not bigger than 1 on K is compact. This being said, the claim follows immediately from
the observation that for every hd−1 ≤ N < hd , the function
ξ 7−→ |vdm(ξ1 , . . . , ξN −1 , ξ)|
is a polynomial of n complex variables and of degree at most d so that, by definition of MA ,

max |vdm(ξ1 , . . . , ξN −1 , ξ)| ≤ MA max |vdm(ξ1 , . . . , ξN −1 , ξ). 

ξ∈K ξ∈A

In practice, the smallest the constant MA the better the points. In general, it is not easy
to select a mesh A with low constant MA and low cardinality. The best algorithm currently
available to compute discrete Leja points is provided in [11]. In short, it is as follows. If we
write A = {a1 , . . . , aM } (M > hd ) and if P = {P1 , . . . , Phd } is an ordered polynomial basis
then applying the standard LU factorisation with row pivoting to the following rectangular
Vandermonde matrix
(45) V (a1 , . . . , aM ; P) = [Pj (ai )]1≤i≤M,1≤j≤hd .
produces a permutation vector σ whose first hd components are the indices of the first hd Leja
points in A. During this process, the matrix V (a1 , . . . , aM ; P) is progressively modified by op-
erations on the rows and the pivot row is always selected from the candidates with the largest
absolute value in the first column of the submatrix of interest. Observe, due to the large di-
mension of the determinant involved, one can currently compute points only for low degrees, say
d = 10 for p = 2.
Finally, we mention a (theoretical) method, based on the same idea as the above proposition,
using the concept of (weakly) admissible mesh (for K) introduced in [16] which consists of a
sequence of sets (Ad )d ∈ N, such that Ad ⊂ K is determining for the space of polynomials of
degree ≤ d and both MAd and the cardinality of Ad grows subexponentially as d → ∞. The
proof is immediate and is omitted.
12 DIMITRI JORDAN KENNE

Proposition 5.2. Let K a determining compact set in Cp . and (Ad )d∈N an admissible mesh for
K. A pseuso-Leja sequence (ξN )N ∈N for K of Edrei growth M N with
M N = M Ad , hd−1 ≤ N < hd ,
is obtained as follows :
(1) ξ0 ∈ A0 and,
(2) for each d ∈ N, we compute recursively the points ξhd−1 , . . . , ξhd −2 and ξhd −1 from Ad
such that
(46) |vdm(ξ0 , . . . , ξN −1 , ξN )| = max |vdm(ξ0 , . . . , ξN −1 , ξ)|.
ξ∈Ad

Acknowledgement. I would like to thank Professor Leokadia Bialas-Cież, my research super-

visor, for her patient guidance, enthusiastic encouragement and useful critiques of this work. I
would also like to express my sincere gratitude to Professor Jean-Paul Calvi for his unconditional
help in structuring and writing this paper.
Funding: This work was partially supported by the National Science Center, Poland, grant
Preludium Bis 1 N. 2019/35/O/ST1/02245.

References
[1] J. Baglama, D. Calvetti, and L. Reichel. Fast Leja points. Electron. Trans. Numer. Anal, 7(124-140):119–120,
1998.
[2] R. Berman and S. Boucksom. Equidistribution of Fekete points on complex manifolds. arXiv preprint
arXiv:0807.0035, 2008.
[3] R. Berman, S. Boucksom, and D. W. Nyström. Fekete points and convergence towards equilibrium measures
on complex manifolds. Acta Math., 207(1):1–27, 2011.
[4] F. Bertrand and J.-P. Calvi. The Newton product of polynomial projectors. Part 2: approximation properties.
Rend. Circ. Mat. Palermo (2), pages 1–34, 2022.
[5] L. Bialas-Cież and J.-P. Calvi. Pseudo Leja sequences. Ann. Mat. Pura Appl., 91(1):53–75, 2012.
[6] T. Bloom. On families of polynomials which approximate the pluricomplex Green function. Indiana Univ.
Math. J., pages 1545–1566, 2001.
[7] T. Bloom, L. Bos, J-P Calvi, and N. Levenberg. Polynomial interpolation and approximation in Cd . In Ann.
Pol. Math., volume 1(106), pages 53–81, 2012.
[8] T. Bloom, L. Bos, C. Christensen, and N. Levenberg. Polynomial interpolation of holomorphic functions in
C and Cn . Rocky Mountain J. Math., 22(2):441–470, 1992.
[9] T. Bloom and J.-P. Calvi. The distribution of extremal points for Kergin interpolations: real case. In Ann.
Inst. Fourier, volume 48(1), pages 205–222, 1998.
[10] L. Bos, M. Caliari, S. De Marchi, M. Vianello, and Y. Xu. Bivariate Lagrange interpolation at the Padua
points: the generating curve approach. J. Approx. Theory, 143(1):15–25, 2006.
[11] L. Bos, S. De Marchi, A. Sommariva, and M. Vianello. Computing multivariate Fekete and Leja points by
numerical linear algebra. SIAM J Numer Anal, 48(5):1984–1999, 2010.
[12] L. Bos, S. De Marchi, A. Sommariva, and M. Vianello. On multivariate Newton interpolation at discrete Leja
points. Dolomites Res. Notes Approx, 4(Proceedings of Robert Schaback’s 65th birthday):15–20, 2011.
[13] L. Bos, S. De Marchi, M. Vianello, and Y. Xu. Bivariate Lagrange interpolation at the Padua points: the
ideal theory approach. Numer Math, 108:43–57, 2007.
[14] M. Caliari, S. De Marchi, and M. Vianello. Bivariate polynomial interpolation on the square at new nodal
sets. Appl. Math. Comput, 165(2):261–274, 2005.
[15] J.-P. Calvi. Intertwining unisolvent arrays for multivariate Lagrange interpolation. Adv. Comput. Math.,
23(4):393–414, 2005.
[16] J.-P. Calvi and N. Levenberg. Uniform approximation by discrete least squares polynomials. J. Approx.
Theory, 152(1):82–100, 2008.
[17] M. Fekete. Über die verteilung der Wurzeln bei gewissen algebraischen Gleichungen mit ganzzahligen Koef-
fizienten. Math. Zeitschrift, 17(1):228–249, 1923.
[18] D. Gaier. Lectures on complex approximation. Birkhauser Boston Inc., 1987.
[19] A. Irigoyen. Multidimensional intertwining Leja sequences and applications in bidimensional Lagrange inter-
polation. J. Approx. Theory, 264:105540, 2021.
 MULTIDIMENSIONAL PSEUDO LEJA SEQUENCES 13

[20] M. Jedrzejowski. Transfinite diameter and extremal points for a compact subset of Cn . Univ. Iagell. Acta.
Math., 29:65–70, 1992.
[21] M. Klimek. Pluripotential theory. London Math. Soc. Monogr.(NS), 1991.
[22] F. Leja. Problèmes à résoudre posés à la conférence. In Colloq. Math, volume 7(1), page 153, 1959.
[23] J. Siciak. On some extremal functions and their applications in the theory of analytic functions of several
complex variables. Trans. Amer. Math. Soc., 105(2):322–357, 1962.
[24] V. Totik. The Lebesgue constants for Leja points are subexponential. J. Approx. Theory, page 105863, 2023.
[25] V.P. Zaharjuta. Transfinite diameter, Chebyshev constants and capacity for compacts in Cn . Math. USSR-
Sb., 25(3):350, 1975.

Doctoral school of Exact and Natural Sciences, Jagiellonian University, Lojasiewicza 11, Kraków,
30-348, Lesser Poland, Poland.
Email address: [email protected]