HAL Id: hal-01151663
https://hal.archives-ouvertes.fr/hal-01151663
Submitted on 13 May 2015
HAL
is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire
HAL, estdestinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
New characterizations for the eigenvalues of the prolate spheroidal wave equation
Françoise Richard-Jung, Jean-Pierre Ramis, Jean Thomann, Frédéric Fauvet
To cite this version:
Françoise Richard-Jung, Jean-Pierre Ramis, Jean Thomann, Frédéric Fauvet. New characterizations
for the eigenvalues of the prolate spheroidal wave equation. Studies in Applied Mathematics, Wiley-
Blackwell, 2017, 138 (1), pp.3-42. �10.1111/sapm.12134�. �hal-01151663�
equation
By F. Richard-Jung, J.-P. Ramis, J. Thomann, F. Fauvet
In this paper, we give new characterizations for the eigenvalues of the prolate wave equation as limits of the zeros of some families of polynomials: the coefficients of the formal power series appearing in the solutions near 0, 1 or ∞ (in the variables x, x − 1 or 1/x respectively). The result, which seems to be true for all values of the parameter τ , according to our numerical experiments, is here proved for small values of the parameter τ .
1. Introduction
This paper is devoted to the study of the spectral problem D
τy = µy where:
D
τ: y 7→ (x
2− 1)y
′′+ 2xy
′+ τ
2x
2y, τ being a real parameter (τ ≥ 0).
It is equivalent to the study of the spectral problem for the prolate spheroidal wave operators
1but D
τ, being a perturbation of the opposite of the Legendre operator:
D
0= (x
2− 1) d
dx
2+ 2x d
dx = − d dx
(1 − x
2) d dx
, is more adapted to a perturbative approach.
The differential operator D
τis interpreted as defined on the Riemann sphere P
1(
C) :=
C
∪ {∞} . It admits 3 singularities: x = 1, x = − 1, x = ∞ ; 1 and − 1 are regular-singular and ∞ is irregular. The differential equation D
τy − µy = 0 is a confluent Heun equation.
In a preceding paper [1], we made the following observations.
For τ ≥ 0 fixed and µ ∈
R, the following properties are equivalent:
1. µ is not an eigenvalue of D
τ;
Francoise.Jung@imag.fr, ramis@picard.ups-tlse.fr, thomann@math.u-strasbg.fr, fauvet@math.u-strasbg.fr
1Of orderm= 0.
DOI: 10.1002/Pro.late˙fin 1
STUDIES IN APPLIED MATHEMATICS 00:1–35 c Wiley Periodicals, Inc., A Wiley Company
2. a non trivial power series solution of D
τy − µy = 0 at x = 1 admits 2 as convergence radius;
3. a non trivial power series solution of D
τy − µy = 0 at x = − 1 admits 2 as convergence radius;
4. the non trivial power series solution of D
τy − µy = 0 at x = 0 admits 1 as convergence radius;
5. the power series solutions appearing in a fundamental system of formal solutions of D
τy − µy = 0 at x = ∞ are divergent.
For τ ≥ 0 fixed and µ ∈
R, the following properties are equivalent:
1. µ is an eigenvalue of D
τ;
2. a power series solution of D
τy − µy = 0 at x = 1 admits + ∞ as convergence radius;
3. a power series solution of D
τy − µy = 0 at x = − 1 admits + ∞ as convergence radius;
4. there exists a non trivial even or odd power series solution of D
τy − µy = 0 at x = 0 admitting + ∞ as convergence radius;
5. the power series solutions of D
τy − µy = 0 appearing in the formal series solutions at x = ∞ are convergent.
In the above observations all the power series (respectively in the variables x − 1, x + 1, x, 1/x) admits as coefficients some polynomials in µ with coefficients in
C[τ, τ
−1], satisfying some polynomial linear recurrences. For example there exists a unique even power series solution at x = 0: P
n∈N
P
2nx
2n, P
0:= 1. (The P
2nsatisfy a three terms recurrence, they are polynomials of degree n in the variable µ with coefficients in
Q[τ ].)
For a fixed value of the parameter τ and µ ∈
R, the problem to decide if µ is an eigenvalue of D
τis a priori a global problem but it follows from the above observations that this problem can be “solved” locally at x = 1, or x = − 1, or x = 0, or x = ∞ . For example at x = 0 we have to “see” if the radius of convergence of P
n∈N
P
2n(µ)x
2nis + ∞ or finite.
The source of our article is the following experimental observation. (We will explain what is happening at x = 0 but there are similar phenomena for x = 1, x = − 1 or x = ∞ .) Trying to use the above considerations, that is the “jump” in the radius of convergence when µ crosses an eigenvalue, to get an heuristic “quick and efficient” numerical method to compute the eigenvalues
2, we discovered a surprising phenomenon: experimentally, for all j ∈
N, the j-th zero of the polynomial P
2p(p ≥ j) tends to the 2j-th eigenvalue of D
τwhen p tends to + ∞
3.
Our article contains a proof of this result and of its variations at x = 1, x = − 1 or x = ∞ .
4Our proof does not use the classical results on the eigenvalues problem for the prolate spheroidal equations as the convergent power series expansions in τ of the eigenvalues obtained by continued fractions methods in [2]
5(cf. [2] 3.24 (10), page 240, [3], 3, page 16).
2In the line of a question asked in the conclusion of [1].
3The reader can easily check that ifτ:= 0, then the eigenvalue 2j(2j+ 1) is a zero of all the polynomialsP2pforp≥j.
4Most of our methods are perturbative, therefore we have complete results only for “small values” of the parameterτ.
5Therefore, as a byproduct, we get new proofs of such results.
The starting point of our proof was to try to give a rigourous meaning to a method proposed in [4], that is the “computation of the zeroes” of an infinite dimensional determinant, a Hill determinant, associated to the expansion of prolate spheroidal functions in series of Legendre polynomials, as limits of zeroes of finite dimensional determinants obtained by truncation. Our work was nearly finished when we noticed [5]. This paper contains a proof of this result for an arbitrary value of τ . Our proof only works for “small values” of τ , however it is necessary for the proofs of our main results.
2. Solutions in the complex plane
In this section, we precise the formal solutions which build basis of solutions in the neighbor- hood of the points 0, 1, − 1 and ∞ . Indeed, a complete description of these formal solutions can been obtained using the
maplepackage
Desir[6, 7, 8].
In the neighborhood of the origin, we find a basis of power series, whose first terms are the following:
y
1(x) = 1 − µ 2 x
2+
1
24 µ
2+ 1
12 τ
2− 1 4 µ
x
4+ O x
6,
y
2(x) = x
2 + − µ + 2 3 x
2+
1
60 µ
2+ 1
10 τ
2− 7 30 µ + 2
5
x
4+ O x
6.
In the neighborhood of each of the points ± 1, we have a basis of solutions constituted of a regular (holomorphic at the singularity) function f and a solution of the form f(x) log(x ± 1) + g(x), where g is also regular.
For example, in the neighborhood of the point 1, we obtain:
f
1(x) = 1+ − τ
2+ µ
2 (x − 1) + 1
16 τ
4− 1
8 µ τ
2+ 1
16 µ
2− 1 8 τ
2− 1
8 µ
(x − 1)
2+O
(x − 1)
3f
2(x) = ln (x − 1) f
1(x) +
− 1
2 + τ
2− µ
(x − 1)
+ 1
8 τ
2+ 1 8 µ + 1
8 − 3
16 τ
4+ 3
8 µ τ
2− 3 16 µ
2(x − 1)
2+ O
(x − 1)
3.
In general, in the neighborhood of a point x
0, the solutions are computed using the rational Newton algorithm [9], this means that we obtain “generalized formal solutions”:
they are parametrized by a new variable t and have the form (x(t) − x
0= Λt
r, y(t) =
exp(Q(1/t))t
λΦ(t)). In this expression, Λ is a complex number (usefull in order to reduce the
algebraic extension needed), r is a positive integer, called the ramification, Q is a polynomial
without constant term, λ is a complex number, called the exponent, and t
λΦ(t) is the regular
part of the formal solution. Φ(t) is a polynomial in log(t) with power series coefficients.
In our example, as 1 is a regular singularity, the parametrization is only a translation x(t) − 1 = t, and Q = 0.
In the neighborhood of ∞ , we obtain a basis of formal solutions constituted of:
x(u) = 1
u , y(u) = e
−RootOf(τ2 +Z2 )u
u (1 − RootOf(τ
2+ Z
2)(µ − τ
2) u
2 τ
2+ O(u
2))
Remark that we have above a condensed shape representing two solutions, corresponding to the two possible values of the
RootOf:± iτ . This gives raise to a basis of formal solutions constituted of:
ˆ
y
1(x) = e
−iτ xx
1 − i (µ − τ
2) 2τ
1
x + − µ
2+ 2 µ + 2τ
2µ + 2τ
2− τ
48τ
21
x
2+ O( 1 x
4)
ˆ
y
2(x) = e
iτ xx
1 + i (µ − τ
2) 2τ
1
x + − µ
2+ 2 µ + 2τ
2µ + 2τ
2− τ
48τ
21
x
2+ O( 1 x
4)
The parametrization is now given by the change of variable x = 1/u, the ramification is trivial, and the two series which appear in these solutions are a priori divergent, but 1-summable in each direction but ± iτ
R+.
Of course, we give here only the first terms of the series, but the following ones can be generated using a recurrence formula, so (theoretically) we can obtain as much terms as wanted in all the series appearing in the solutions.
3. The eigenvalues as roots of an infinite determinant
Following [[4], chapter 7, paragraph 7.5] (cf. also [5]), we try to expand a solution of the equation
D
τ(y) = µy (1)
as an infinite linear combination of eigenfunctions of the unperturbated problem, i.e. the Legendre polynomials,
y(x) = X
∞n=0
a
nL
n(x).
We consider also x
2L
n(x) as expanded in terms of the Legendre polynomials:
x
2L
n(x) = X
∞m=0
A
nmL
m(x).
Applying the differential equation (1) to y(x) and equating coefficients of L
nfor each n gives an infinite matrix equation satisfied by the coefficients a
n:
M a =
− µ + τ
2A
00τ
2A
10τ
2A
20τ
2A
30. . . τ
2A
012 − µ + τ
2A
11τ
2A
21τ
2A
31. . . τ
2A
02τ
2A
126 − µ − τ
2A
22τ
2A
32. . .
.. . .. . .. . .. . . ..
a
0a
1a
2.. .
= 0.
In our case, it is possible to obtain explicit formulas for the coefficients A
nm. Indeed, A
nm=
m+1/21R
1−1
x
2L
n(x)L
m(x)dx. Using the recurrence relation (2n + 1)xL
n= (n + 1)L
n+1+ nL
n−1and the fact that R
1−1
L
n(x)L
m(x)dx = 2δ
mn2n + 1 , we obtain:
A
nn= 2n
2+ 2n − 1
(2n + 3)(2n − 1) , A
n+2n= (n + 1)(n + 2)
(2n + 1)(2n + 3) , A
nn+2= (n + 1)(n + 2) (2n + 3)(2n + 5) . Then the matrix M is tridiagonal:
M =
−µ+τ2/3 0 2τ2/3 0 0 . . .
0 2−µ+ 3τ2/5 0 2τ2/5 0 . . .
2τ2/15 0 6−µ+ 11τ2/21 0 12τ2/35 . . .
0 6τ2/35 0 12−µ+ 23τ2/45 0 . . .
0 0 4τ2/21 0 20−µ+ 39τ2/77 . . .
... ... ... ... ... . ..
.
Moreover, we can split the odd and even coefficients a
n, by extracting from the matrix M the corresponding columns and rows.
For example, the even coefficients satisfy the following system:
− µ + τ
2/3 2τ
2/3 0 . . .
2τ
2/15 6 − µ + 11τ
2/21 12τ
2/35 . . . 0 4τ
2/21 20 − µ + 39τ
2/77 . . .
.. . .. . .. . . ..
a
0a
2a
4.. .
= 0. (2)
Next we truncate the previous infinite matrix, and define M
(n)the n × n matrix whose entries
are the same as the first n rows and columns of the infinite matrix.
We define D
nas the determinant of M
(n). D
nis a polynomial in the variables µ and τ
2, of nth-order in the variable µ. Considering the special shape of the matrix M
(n), it is easy to derive the recurrence formula:
D
n+1= 2n(2n + 1) − µ + A
2n2nτ
2D
n− A
2n2n−2A
2n2n−2τ
4D
n−1. So we obtain a family of polynomials, defined by the initial values
D
0= 1, D
1= − µ + 1 3 τ
2,
and a recurrence equation, whose roots are good candidates to approach the exact eigenvalues of even index of the differential equation (1).
Proposition
1. The polynomial D
nis a polynomial of degree n in the variable µ. For all n ≥ 1,
D
nmod τ
2= ( − 1)
nn−1
Y
j=0
(µ − 2j(2j + 1))
Then D
nmod τ
2admits n (integer) roots: 0, 6, . . . , 2j(2j + 1), . . . , (2n − 2)(2n − 1), all are simple.
The polynomial D
nadmits n Puiseux series solutions, we note µ
(n)2jthe series, which is equal to 2j(2j + 1) mod τ
2.
Proof: Perform the reduction of D
nmodulo τ
2:
(D
j+1mod τ
2) = (2j(2j + 1) − µ) (D
jmod τ
2).
Proposition
2. Convergence radii of the series µ
(n)2j.
The series µ
(n)2j, j ≥ 0 converge for | τ | < 1.
Proof: The following result can be found in [[10], p. 95]:
Let X be a unitary space, let T (x) = T + xT
(1)a linear operator on X and let T be normal. Then the power series for the eigenvalues λ(x) are convergent if “the magnitude of the perturbation” || xT
(1)|| is smaller than half the isolation distance of the eigenvalue λ of T.
More precisely, if λ(x) = P
p≥0
λ
px
p, the coefficients λ
psatisfy the majorations:
| λ
1| ≤ a, | λ
p| ≤ a
p2
d
p−1, p ≥ 2, where a = || T
(1)|| and d is the isolation distance of the eigenvalues of T . We apply this result here by searching the eigenvalues of:
F
(n)+ τ
2G
(n)where
F
(n)=
0
6 20
. ..
2(n − 1)(2n − 1)
,
and
G
(n)=
1/3 2/3
2/15 11/21 12/35
4/21 39/77 10/33 . .. . .. . ..
A
2n2n−−42A
2n2n−−22
The matrices F
(n)and G
(n)are n × n square matrices, G
(n)is a tridiagonal matrix, whose elements are explicitely known.
It is easy to majorate the 2-norm of G
(n), using the inequality || G
(n)||
2≤ p || G
(n)||
1|| G
(n)||
∞≤ 3.
We obtain the announced result by noting that d = 6 (for all n).
In the following paragraph, we will study the link between the series µ
(n)2j, when j is fixed and n is growing.
4. The eigenvalues as sums of a series
First we compute the beginning of the Puiseux series of the polynomials D
n, for small n.
> algcurves[puiseux](D(1),tau=0,mu,9);
1 3τ2
> algcurves[puiseux](D(2),tau=0,mu,9);
1 3τ2−
2
135τ4+ 4
8505τ6+ 58
2679075τ8,6 +11
21τ2+ 2 135τ4−
4 8505τ6−
58 2679075τ8
> algcurves[puiseux](D(3),tau=0,mu,9);
1 3τ2−
2
135τ4+ 4
8505τ6+ 26
1913625τ8,6 +11
21τ2+ 94 9261τ4−
21388 44925111τ6−
2946550 217931713461τ8, 20 +39
77τ2+ 8
1715τ4+ 16 2773155τ6−
2476 37368263625τ8
> algcurves[puiseux](D(4),tau=0,mu,9);
1 3τ2−
2
135τ4+ 4
8505τ6+ 26
1913625τ8,6 +11
21τ2+ 94 9261τ4−
21388 44925111τ6−
3633830 257555661363τ8, 20 +39
77τ2+ 52674
29674645τ4+ 935076
175940970205τ6+ 37764832611642 74924536936711587125τ8, 42 + 83
165τ2+ 60134
76366125τ4+ 101060636
276516011165625τ6+ 504082644490258 6322139029665881296875τ8
We notice that the first terms of the series stabilize little by little.
More precisely, consider the “first” series solution of all polynomials, the series whose constant term is null. The term of degree 2 is the same for all poynomials, the terms of degree 4 and 6 are the same for all polynomials of index greater than 1, the terms of degrre 6 and 8 are the same for all polynomial of index greater than 2.
This phenomenon is repeated for the “second” series solution, the series whose constant term is 6. The term of degree 2 is the same for all polynomials of index greater than 1, the terms of degree 4 and 6 are the same for all polynomials of index greater than 2, etc...
We can express that in the following way:
if we put µ
0= 0, D
1(µ
0) = 0 mod τ
2, µ
1=
τ32, D
1(µ
1) = 0 mod τ
4; µ
2= µ
1−
1352τ
4+
4
8505
τ
6, D
2(µ
2) = 0 mod τ
8.
Also, if µ
0= 6, D
2(µ
0) = 0 mod τ
2, µ
1= 6 +
1121τ
2, D
2(µ
1) = 0 mod τ
4; µ
2= µ
1+
926194τ
4−
21388
44925111
τ
6, D
3(µ
2) = 0 mod τ
8.
And also, if µ
0= 20, D
3(µ
0) = 0 mod τ
2, µ
1= 20 +
3977τ
2, D
3(µ
1) = 0 mod τ
4. So we can prove the
Proposition
3. Let j ≥ 0. We put µ
0= 2j(2j + 1). Then D
j+1(µ
0) = 0 mod τ
2. We build µ
1= µ
0− D
j+1(µ
0)
D
j+1′(µ
0) mod τ
4and we prove that D
j+1(µ
1) = 0 mod τ
4. More generally, for all i ≥ 1, if µ
i= µ
i−1− D
j+i(µ
i−1)
D
′j+i(µ
i−1) mod τ
4i, then D
j+i(µ
i) = 0 mod τ
4i.
Proof: From above, it is clear that D
j+1(µ
0) = 0 mod τ
2and that (D
j+1mod τ )
′(µ
0) 6 =
0. So D
j+1′(µ
0) is a polynomial in τ with a non nul constant term, which means that it is
invertible in the ring of formal series, then µ
1is well defined. Moreover µ
1= µ
0mod τ
2.
We recall the following Taylor formula [[11], p. 51]: let a be a univariate polynomial over an
arbitrary integral domain A . In the polynomial domain A [x, y], a(x + y) = a(x) + a
′(x)y +
b(x, y)y
2for some polynomial b ∈ A [x, y]. We apply this result with a = D
j+1and A =
Q[τ ] : D
j+1(µ
1) = D
j+1(µ
0) + (µ
1− µ
0)D
j+1′(µ
0) + (µ
1− µ
0)
2b(µ
0, µ
1)
= D
j+1(µ
0) + (µ
1− µ
0)D
j+1′(µ
0) mod τ
4= 0 mod τ
4.
We assume now that we have built the first terms until the index i − 1 (for some i ≥ 2).
By construction: µ
i−1= µ
0mod τ
2, then D
′j+i(µ
i−1) mod τ = (D
j+imod τ )
′(µ
i−1mod τ ) = (D
j+imod τ )
′(µ
0) is not null. Then µ
iis well defined. Moreover, by using the recurrence equation and the fact that D
j+i−1(µ
i−1) = 0 mod τ
4(i−1)and that D
j+i−2(µ
i−1) mod τ
4(i−2)= D
j+i−2(µ
i−2) mod τ
4(i−2)= 0, we obtain that D
j+i(µ
i−1) = 0 mod τ
4(i−1), thus µ
i= µ
i−1mod τ
4(i−1).
The same Taylor formula enables us to prove that D
j+i(µ
i) = 0 mod τ
4i.
In fact, µ
iis built by adding to µ
i−1two monomials, of degree 4i − 4 and 4i − 2, and from a computing point of view µ
i= µ
i−1− D
j+i(µ
i−1)
D
′j+i(µ
1) mod τ
4i.
Proposition
4. For all j ≥ 0, we have built a formal series in the variable τ , µ ˆ
2j, such that
ˆ
µ
2j= 2j(2j + 1) mod τ
2et ∀ n > j, D
n(ˆ µ
2j) = 0 mod τ
4(n−j). To fix the ideas, we give the first terms of the series ˆ µ
0, ˆ µ
2, ˆ µ
4.
ˆ
µ
0=
13τ
2−
1352τ
4+
85054τ
6+
191362526τ
8−
3788977592τ
10+
9050920003125513988τ
12ˆ
µ
2= 6 +
1121τ
2+
926194τ
4−
4492511121388τ
6−
2575556613633633830τ
8+
1622600666586939611204τ
10−
277773545116 4906403669618802351τ
12ˆ
µ
4= 20 +
3977τ
2+
2967464552674τ
4+
175940970205935076τ
6+
37764832611642 74924536936711587125τ
8−
3187867616210148241152114584499914320525
τ
10−
1613990098021782094984490612997487168811993405138946875
τ
12In the following paragraphs, we will apply the following lemma:
Lemma
1. Let P be a polynomial, P ∈
C[τ ][y], µ
∗a series ( ∈
C[[τ ]]) satisfying P (µ
∗) = 0, P
′(µ
∗) 6 = 0 and µ ˆ an other series such that P (ˆ µ) = 0 mod τ
kand µ ˆ = µ
∗mod τ . Then ˆ
µ = µ
∗mod τ
k.
Proof: We use the above Taylor formula [[11], p. 51].
Proposition
5. Link between the series µ
(n)2jand the series µ ˆ
2j∀ j ≥ 0, ∀ n > j, µ
(n)2j= ˆ µ
2jmod τ
4(n−j).
Proof: By construction of the series ˆ µ
2jor by applying the lemma.
Theorem
1. For all j ≥ 0, the series µ ˆ
2jis convergent in the open unit disk, and for all | τ | < 1, the sequence µ
(n)2j(τ ) converges, when n tends to infinity, to µ ˆ
2j(τ ), which is an eigenvalue D
τ.
Proof: Let α
(n)pbe the coefficients of the series µ
(n)2jand α
pthose of ˆ µ
2j. The majoration given by Kato is independent of n, because d is independent of n and || G
(n)||
2is bounded independently of n : | α
(n)p| ≤ 3 (for all n and for all p). So we have also | α
p| ≤ 3 (for all p).
Let | τ | < 1.
µ
(n)0(τ ) − µ ˆ
0(τ ) = X
p>4n
(α
(n)p− α
p)τ
2p, and
| µ
(n)0(τ ) − µ ˆ
0(τ ) | ≤ 6 X
p>4n
| τ
2p| , what we can make arbitrarily small when n tends to infinity.
The last part of the theorem, the fact that ˆ µ
2j(τ ) is an eigenvalue of D
τ, will be proved in the paragraph 9. It is also possible to prove this result using [5] but we propose a self-contained proof for all our results.
5. The eigenvalues and the coordinates of the eigenfunctions in Legendre basis Consider an eigenfunction of even index of the equation D
τ(y) = µy as an infinite linear combination of the Legendre polynomials,
y(x) = X
∞n=0
a
2nL
2n(x), a
0= 1.
We will now see that there exists a simple relation between a
2nand D
n.
Proposition6.
∀ n ≥ 1, D
n= ( − 1)
nn
Y
i=1
A
2i2i−2!
τ
2na
2n. Proof: By recurrence.
For n = 1: a
2is defined by the equation D
1a
0+ A
20τ
2a
2= 0.
Suppose that the relation is satisfied for a fixed integer n ≥ 1. We perform the Gaussian
elimination algorithm, in order to put the matrix M
(n+2)in an echelon form.
We assume that the first steps give the intermediate result:
D1 A2
0τ2 D2 A4
2τ2D1 ...
...
Dj A2j
2j−2τ2Dj−1 A2j−2
2j τ2 2j(2j+ 1)−µ+A2j
2jτ2 A2j+2 2j τ2 ..
.
.. .
.. . A2n−2
2n τ2 2n(2n+ 1)−µ+A2n
2n τ2 A2n+2
2n τ2 A2n
2n+2τ2 (2n+ 2)(2n+ 3)−µ+A2n+2 2n+2τ2
.
The next step is: replace row
j+1by D
jrow
j+1− A
2j2j−2τ
2row
j; using the recurrence relation giving D
j+1in terms of D
jand D
j−1, we obtain:
D1 A20τ2
D2 A42τ2D1
. .. . ..
Dj A2j2j
−2τ2Dj−1
Dj+1 A2j+22j τ2Dj
. .. . .. . ..
A2n2n−2τ2 2n(2n+ 1)−µ+A2n2nτ2 A2n+22n τ2
A2n2n+2τ2 (2n+ 2)(2n+ 3)−µ+A2n+22n+2τ2
.
Then, at the final step, we will have:
D1 A20τ2
D2 A42τ2D1
. .. . .. . ..
Dj A2j2j−2τ2Dj−1
Dj+1 A2j+22j τ2Dj
. .. . ..
Dn+1 A2n+22n τ2Dn
A2n2n+2τ2 (2n+ 2)(2n+ 3)−µ+A2n+22n+2τ2
.
The penultimate row gives:
D
n+1a
2n+ A
2n+22nτ
2D
na
2n+2= 0,
which is the expected relation for n + 1.
Corollary
1: for all n ≥ 1, the coefficient a
2nhas the same zeros as the polynomial D
n. Thus, for all | τ | < 1, the zero of index j of the coefficient a
2nconverges to ˆ µ
2j(τ ), when n tends to infinity.
Proposition
7. For all j ≥ 0, the series µ ˆ
2jsatisfies :
∀ n > 2j, a
2n(ˆ µ
2j) = 0 mod τ
2(n−2j).
6. The eigenvalues and the formal solutions at the origin
We recall that, in the neighborhood of the origin, we know a basis of convergent series solutions, y
1(x) et y
2(x).
In particular:
y
1(x) = 1 − µ 2 x
2+
1
24 µ
2+ 1
12 τ
2− 1 4 µ
x
4+ O x
6. Let P
nbe the coordinates of this function in the monomial basis:
P
0= 1, P
1= 0, P
2= − µ
2 , P
3= 0, P
4= 1
24 µ
2+ 1
12 τ
2− 1 4 µ, the following polynomials satisfying a three terms recurrence:
− n(n − 1)P
n+ (n
2− µ − 3n + 2)P
n−2+ τ
2P
n−4= 0.
Proposition
8. The odd coordinates are null. The even coordinates, P
2nare polynomials of degree n in the variable µ and for all n ≥ 1, P
2nmod τ
2=
((2n)!−1)nQ
n−1j=0
µ − 2j(2j + 1).
The polynomial P
2nadmits n Puiseux series solutions, we will denote δ
j(n), j = 0 . . . n − 1, the series which has 2j(2j + 1) as constant term.
Proof: Reduce the recurrence equation modulo τ
2:
2n(2n − 1)(P
2nmod τ
2) = ((2n − 1)(2n − 2) − µ)(P
2n−2mod τ
2).
Proposition
9. For all j ≥ 0, the series µ ˆ
2jsatisfies:
∀ n > j, P
2n(ˆ µ
2j) = 0 mod τ
2(n−j).
Proof: Taking into account the parity of the functions, the eigenfunction of even index that we consider:
y(x) = X
∞n=0
a
2nL
2n(x), a
0= 1,
is proportional to y
1, that is y(x) = cy
1(x).
We obtain by change of basis:
c cP
2cP
4.. .
= A
1 a
2a
4.. .
,
where we only need to note that A is an infinite upper triangular matrix, whose elements will be noted α
ij.
A =
1 − 1/2 3/8 . . . 3/2 − 15/4 . . . 35/8 . . . . ..
.
Then: c = P
∞j=0
α
1ja
2jand, using the proposition 7, c(ˆ µ
0) = 1 mod τ
2. We deduce:
c(ˆ µ
0)P
2n(ˆ µ
0) = α
nna
2n(ˆ µ
0) mod τ
2n+2= 0 mod τ
2n, ∀ n ≥ 1.
Also τ
2c(ˆ µ
2) = τ
2−
12τ
2a
2(ˆ µ
2) +
38τ
2a
4(ˆ µ
2) mod τ
4=
2A120
D
1(ˆ µ
2) mod τ
26 = 0 mod τ . We deduce: τ
2c(ˆ µ
2)P
4(ˆ µ
2) = −
358τ
2a
4(ˆ µ
2) mod τ
4= −
8A2035A42τ2D
2(ˆ µ
2) mod τ
4= 0 mod τ
2, hence P
4(ˆ µ
2) = 0 mod τ
2.
And ∀ n ≥ 2,
τ
2c(ˆ µ
2)P
2n(ˆ µ
2) = α
nnτ
2a
2n(ˆ µ
2) mod τ
2n= cste
τ
2(n−1)D
n(ˆ µ
2) mod τ
2n= 0 mod τ
2(n−1), then P
2n(ˆ µ
2) = 0 mod τ
2(n−1), ∀ n ≥ 2.
In a general way:
τ
2jc(ˆ µ
2j) = X
i=0
α
1iτ
2ja
2i(ˆ µ
2j) =
2j
X
i=0
α
1iτ
2ja
2i(ˆ µ
2j) mod τ
2j+2= α
1j( − 1)
jQ
jk=1