# Bounded solutions of self-adjoint second order linear difference equations with periodic coeffients

A.M. Encinas and M.J. Jiménez
From the journal Open Mathematics

# Abstract

In this work we obtain easy characterizations for the boundedness of the solutions of the discrete, self–adjoint, second order and linear unidimensional equations with periodic coefficients, including the analysis of the so-called discrete Mathieu equations as particular cases.

MSC 2010: 39A12; 39A70

## 1 Introduction

Discrete Schrödinger operators over finite or infinite paths have been subject of an intensive research over the last four decades. They represent the discrete analogs of one–dimensional self–adjoint operators on a bounded or unbounded interval on the real line, see for instance [1]. In addition, those operators are in relation with Jacobi matrices and hence with the classical theory of orthogonal polynomials.

The particular case of the so–called almost Mathieu operator has deserved special attention not only by its connections to physics but for its rich spectral theory. In fact, one of the main problems in this area, related to the topological structure of the spectra and popularized as the Ten Martini Problem, has been recently solved by concatenating the work of many outstanding researchers, see [2,3,4]. The problem is closely related to the determination of those energies for which the corresponding Schrödinger equation has non trivial bounded eigenfunctions.

The aim of this communication is by far much more modest. We use recent advances in the study of linear difference equations with periodic coefficients, see [5], to provide easy characterizations for the boundedness of the solutions of the Mathieu equations, that correspond to some specific Schrödinger equations with periodic potential, see [6]. Moreover, we also extend the results to general second order linear difference equations with periodic coefficients.

## 2 Preliminaries

Throughout the paper, (ℤ) denotes the vector space of real sequences; that is, (ℤ) = {z:ℤ → ℝ}, whereas (ℤ) is the set of sequences z(ℤ) such that z(k) ≠ 0 for all k ∈ ℤ. The null sequence, also called the trivial sequence, is denoted by 0.

Given z(ℤ) and p ∈ ℕ, for any m ∈ ℤ we denote by zp,m(ℤ) the subsequence of z defined as

zp,m(k)=z(kp+m),kZ.

Clearly, any sequence z(ℤ) is completely determined by the values of the sequences zp,j, for 0 ≤ jp – 1. In particular, z1,0 = z, whereas z2,0 and z2,1 are the subsequences of z formed by the even or odd indexes, respectively. Moreover, the sequences z1,m are the shift subsequences of z, since z1,m(k) = z(k + m) for any k ∈ ℤ. Notice that if we also allow p = – 1, then z–1,m are the flipped shift subsequences of z, since z–1,m(k) = z(mk) for any k ∈ ℤ.

The sequence z(ℤ) is called periodic with periodp ∈ ℕ if it satisfies that

z(p+k)=z(k),kZ,

which also implies that z(kp + m) = z(m) for any k,m ∈ ℤ.

The set of periodic sequences with period p is denoted by (ℤ;p) and we define (ℤ;p) = (ℤ;p)∩ (ℤ). In particular (ℤ;1) consists of all constant sequences and then, it is identified with ℝ.

## Lemma 2.1

Givenz(ℤ), thenz is bounded iff there exists p ∈ ℕsuch thatzp,jis bounded, for 0 ≤ jp – 1 and then, zr,mis bounded for anyr ∈ ℕand anym ∈ ℤ.

Given p ∈ ℕ, then z(ℤ;p) iff zp,m(ℤ;1) for any m ∈ ℤ. Moreover, all periodic sequence is also bounded.

Given p ∈ ℕ, a(ℤ;p) and c(ℤ;p), consider the associated self–adjoint operator Δa,c(ℤ)→(ℤ), defined as

Δa,c(z)(k)=c(k)z(k+1)+c(k1)z(k1)a(k)z(k),kZ(1)

and the corresponding (irreductible) homogeneous equation

Δa,c(z)=0.(2)

The sequences a and c are called the coefficients of the Equation (2) and any sequence z(ℤ) satisfying the Identity (2) is called a solution of the equation. It is well-known that for any z0,z1 ∈ ℝ and any m ∈ ℤ, there exists a unique solution of Equation (2) satisfying z(m) = z0 and z(m + 1) = z1.

The problem we are interested in, can be formulated as follows:

For which coefficientsa, c(ℤ;p) has the equation Δa,c(z) = 0 bounded solutions, other than the trivial one?

Operator (1), and hence Equation (2), encompasses many specific examples that have been widely considered in the literature. For instance, when c(k) > 0 for any k ∈ ℤ, then Δa,c = –𝓛q, the Schrödinger operator on the infinite path with conductance c and potencial q(k) = a(k) – c(k) – c(k – 1); that is,

Lq(z)(k)=c(k)(z(k)z(k+1))+c(k1)(z(k)z(k1))+q(k)z(k).

In particular, when c(k) = 1, for any k ∈ ℤ, then Δa,c is known as the Harper operator and denoted by 𝓗a. More specifically when, in addition, the coefficient a is given by a(k) = E – λ cos(2πωk + θ), k ∈ ℤ, then the operator 𝓗a is called Mathieu operator and the parameters E,λ ∈ ℝ, ω ∈ ℚ, θ∈ [0, 2π), are called the energy, coupling, frequency, and phase, respectively. In this case the operator 𝓗a is usually represented as 𝓗E,λ,ω,θ. If we permit the frequency not to be a rational number; that is, ω ∈ ℝ, then 𝓗E,λ,ω,θ is called almost Mathieu operator, but it does not have periodic coefficients. Therefore, in this work we are only interested in Mathieu operators; that is, in rational frequencies. We must bear in mind that when ω=mp, where m ∈ ℤ and p ∈ ℕ are relative primes, then a(ℤ;p).

Observe that the equation 𝓗E,λ,ω,θ(u) = 0 is equivalent to the equation 𝓗λ,ω,θ(u) = Eu; where 𝓗λ,ω,θ denotes the operator 𝓗E,λ,ω,θ when E = 0. Therefore, the energy E is an eigenvalue and u a corresponding eigenfunction of the operator 𝓗λ,ω,θ.

The interested reader can find the physics meaning of these parameters and the physics background of these kind of operators in [1,2,3] and also in [7].

The paper [5] was devoted to the Floquet Theory for the equation Δa,c = 0; that is, to the condition under which the above equation has periodic solutions. Since any periodic solution is bounded, this characterization gives us only a partial answer to the main question. However, we can follow the same techniques as in [5] to completely solve the question.

We end this preliminary section by remarking that when only a finite interval in ℤ is considered, namely when k = 0, 1, …,n for some n ∈ ℕ, then Equation (2) must be supplied with some boundary conditions and it is related with the inversion of finite and symmetric Jacobi matrices, see for instance [8]. Another interesting application of these boundary value problems falls in the ambit of Organic Chemistry, see Examples 1 and 2 in page 364 of [5]. In this case, all the eigenfuncions are bounded, so the main problem is nothing else that the consideration of the eigenvalue problem. For the Mathieu equation with null frequency, this analysis in the finite interval case can be found in [9].

## 3 The easiest case

The most simple case of the proposed problem corresponds to a,c(ℤ;1); that is, when the coefficients of Δa,c are constant; i.e. a ∈ ℝ and c ∈ ℝ. Self–adjoint linear difference equations with constant coefficients can be characterized as those satisfying that z(ℤ) is a solution iff any shift and any flipped shift of z is also a solution. Moreover, in this case, Equation (2) is equivalent, in the sense that both have the same solutions, to the Chebyshev equation with parameter q

z(k+1)2qz(k)+z(k1)=0,kZ,(3)

where q=a2c. So, we can say that the most simple case to analyze corresponds to both the uncoupled Harper equation and the coupled Harper equation with null frequency. Moreover, these two kinds of equations can be viewed in an unified manner as Chebyshev equations. Any solution of a Chebyshev equation with parameter q is called Chebyshev sequence with parameter q.

Recall that a polynomial sequence {Pk(x)}k∈ℤ ⊂ ℝ[x] is a sequence of Chebyshev polynomials if it satisfies the following three-term recurrence, see [10],

Pk+1(x)=2xPk(x)Pk1(x),kZ.(4)

Therefore, any Chebyshev sequence with parameter q is of the form {Pk(q)}k∈ℤ, where {Pk(x)}k∈ℤ is a sequence of Chebyshev polynomials. So, many properties of Chebyshev sequences are the consequence of properties of Chebyshev polynomials and conversely.

As a by–product of the Proposition 2.1 in [5], we have the following basic result about periodic and bounded Chebyshev sequences.

## Proposition 3.1

Givena ∈ ℝ, c ≠ 0 we have the following results:

1. The equation Δa,c(z) = 0 has bounded solutions iff |a| ≤ 2|c|. Moreover, when |a|≤2|c| all the solutions are bounded.

2. When a = 2c the unique bounded solutions are the constant ones, whereas when a = –2 c the bounded solutions are all multiple ofz(k) = (–1)k.

3. The equation Δa,c(z) = 0 has periodic solutions with periodpiffa=2ccos(2πjp),j=0,,p12.

4. The equation Δa,c(z) = 0 has constant solutions iff a = 2c

The main result about the boundedness of solutions for difference equations with constant coefficients has the following particularization for Mathieu operators.

## Corollary 3.2

GivenE, λ ∈ ℝ andθ ∈ [0, 2π) the Mathieu equation with null frequency

z(k+1)+z(k1)λcos(θ)z(k)=Ez(k),kZ

has bounded solutions iff |E – λ cos(θ)| ≤ 2 and all its solutions are bounded when the inequality is strict.

The equation has constant solutions iff E = 2 + λ cos(θ) in which case the constant ones are the unique bounded solutions, whereas when E = λ cos(θ) – 2 the bounded solutions are multiple of z(k) = (–1)k. Finally, the equation has has periodic solutions with period p iffE=λcos(θ)+2cos(2πjp),j=0,,p12.

## 4 The general case

Back to the general case, consider p ∈ ℕ, a(ℤ;p), c(ℤ;p) and the associate self–adjoint operator Δa,c. Although this scenario seems to be far away from the easiest one analyzed in the previous section, we will show that in fact Chebyshev equations contain all the information needed to conclude the existence of bounded solutions for the difference equation Δa,c(z) = 0. This is true because the main result in [5] establishes that (irreductible) second order difference equations (not necessarily self-adjoint) with periodic coefficients are basically equivalent to some Chebyshev equation. For the setting concerning to this paper we have the following facts.

## Lemma 4.1

([5, Theorem 3.3]). Givenp ∈ ℕ, a(ℤ;p) andc(ℤ;p); there exists q(a, c;p) ∈ ℝ, depending only on the coefficients a and c and on the period p, such that z(ℤ) is a solution of the equationΔa,c(z) = 0 iff for anym ∈ ℤ, zp,mis a solution of the Chebyshev equation with parameter q(a, c;p); that is

v(k+1)2q(a,c;p)v(k)+v(k1)=0,kZ.

As the boundedness of z is equivalent to the boundedness of the sequences zp,m, m = 0, …, p – 1, we can conclude that existence of bounded solutions for the equation Δa,c(z) = 0, depends only on the knowledge of the specific value q(a, c;p). Since in [5, Theorem 3.3] the existence of this parameter was proved by induction the above result is not useful in practice. For this reason, most of the above mentioned paper was devoted to the explicit computation of the so–called Floquet function; that is, the function assigning the value q(a, c;p) to any a(ℤ;p) and c(ℤ;p). Notice that, in fact, the value q(a, c;p) only depends on a(j),c(j), j = 0, …,p – 1. Once this function was obtained, the characterization of the existence of periodic solutions for the equation Δa,c(z) = 0 appears as a simple by–product, since from Lemma 2.1, they are characterized as being constant the sequences zp,m, 0 ≤ mp – 1, see [5, Corollary 4.8]. So, the main novelty of this paper is to derive the characterization of the existence of bounded solutions for the equation Δa,c(z) = 0, from the value q(a, c;p). To do this, we need to introduce some notations and concepts.

A binary multi-index of order p is a p-tuple α = (α0, …, αp – 1) ∈ {0,1}p and its length is defined as |α|=j=0p1αjp. So |α| = m iff exactly m components of α are equal to 1 and exactly pm components of α are equal to 0.

Given a binary multi-index of order p, α ∈ {0, 1}p such that |α| = m ≥ 1, we consider 0 ≤ i1 < … < imp – 1 such that αi1 = … = αim = 1. Given p ∈ ℕ, we define the following subsets of the set {0, 1}p of binary multi-indexes of order p:

1. Λp0={(0,,0)}, for p ≥ 1.

2. Λp1={α:|α|=1}, for p ≥ 2.

3. Λpm={α:|α|=m,ij+1ij2,1jm1andimp2ifi1=0} for p ≥ 4, and m=2,,p2, where 0 ≤ i1 < … < imp – 1 are the indexes such that αi1 = … = αim = 1.

In addition, if p ≥ 2, m = 1,…, p2 and αΛpm, let 0 ≤ i1 < … < imp – 1 be the indexes such that αi1 = … = αim = 1. Then, we define the binary multi-index α of order p as

α¯ij=α¯ij+1=0,j=1,,m,andα¯i=1otherwise,

where if im = p – 1, then αp – 1 = α0 = 0. Moreover, if α =(0,…,0); that is, if αΛp0, then we define α = (1,…,1). It is clear that, in any case, |α| = p – 2m.

We are now ready to show the expression for the value of q(a, c; p). In the sequel, we always assume that 00 = 1 and also the usual convention that empty sums and empty products are defined as 0 and 1, respectively.

## Lemma 4.2

([5, Theorem 4.4]). Given p ∈ ℕ*, a(ℤ; p) and c* (ℤ; p), then

q(a,c;p)=12(i=0p1c(i))1j=0p2(1)jαΛpji=0p1c(i)2αia(i)α¯i.

Observe that when p = 1, the above identity becomes q(a,c;1)=a2c; that is, the value corresponding to the case in which the coefficients a and c are constant; or equivalent both have period p = 1.

Our main result appears now as a consequence of the Proposition 3.1 together with Lemma 2.1 and also the above Lemma.

## Theorem 4.3

Given p ∈ ℕ*, a(ℤ; p) and c*(ℤ; p), then the equation

c(k)z(k+1)+c(k1)z(k1)a(k)z(k)=0,kZ

has bounded solutions iff

|j=0p2(1)jαΛpji=0p1c(i)2αia(i)α¯i|2i=0p1|c(i)|

and when the inequality is strict, all the solutions are bounded. Moreover, if

j=0p2(1)jαΛpji=0p1c(i)2αia(i)α¯i=2i=0p1c(i)

then the equation has periodic solutions with period p and these are the unique bounded solutions.

## Corollary 4.4

Given E, λ ∈ ℝ, θ ∈ [0, 2π) andω=mp,where p ∈ ℕ*, m ∈ ℤ and (p, m) = 1, then the Mathieu equation

z(k+1)+z(k1)+λcos(2πωk+θ)z(k)=Ez(k),kZ

has bounded solutions iff

|j=0p2(1)jαΛpji=0p1(Eλcos(2πωi+θ))α¯i|2

and when the inequality is strict, all the solutions are bounded. Moreover, if

j=0p2(1)jαΛpji=0p1(Eλcos(2πωi+θ))α¯i=2

then the Mathieu equation has periodic solutions with period p and these are the unique bounded solutions.

Clearly, the main difficulty to apply the above characterizations is to obtain the binary multi–indexes involved in them. In general, this is a difficult task and, in fact, the number of multi–indexes in Λpj,0jp2, grows dramatically with p. Specifically, we have |Λpj|=ppjpjjfor any0jp2, which for any m ∈ ℕ* implies that j=0m|Λ2mj|=2Tm(32)and thatj=0m|Λ2m+1j|=Wm(32),, see [5, Proposition 4.2].

We end this paper with some specific examples using the given characterization for the existence of bounded solutions for difference equations with periodic coefficients with period up to 4. Remember that the case p = 1, the easiest case, has been analyzed in the previous sections.

## Corollary 4.5

(Period p = 2) Given a(ℤ; 2) and c*(ℤ; 2), then the equation

c(k)z(k+1)+c(k1)z(k1)a(k)z(k)=0,kZ

has bounded solutions iff

(|c(0)||c(1)|)2a(0)a(1)(|c(0)|+|c(1)|)2

and when both inequalities are strict, then all the solutions are bounded. In particular, given E, λ∈ ℝ, θ ∈ [0, 2π), then the Mathieu equation

z(k+1)+z(k1)+λcos(πk+θ)z(k)=Ez(k),kZ

has bounded solutions iff

0E2λ2cos2(θ)4.

## Proof

In this case we have Λ20 = {(0, 0)}, Λ21 = {(1, 0),(0, 1)}, Λ¯20 = {(1, 1)} and Λ¯21 = {(0, 0),(0, 0)}, which implies that q(a,c;2)=a(0)a(1)c(0)2c(1)22c(0)c(1).

In particular, for the Mathieu which coefficient has period 2, the frequency is ω=m2, where m ∈ ℤ is odd and hence ω=n+12 with n ∈ ℤ. Therefore, the coefficient is a(k) = Eλ cos(πk + θ), which implies that a(0) = Eλ cos(θ), whereas a(1) = Eλ cos(π + θ) = E + λ cos(θ).

□

## Corollary 4.6

(Period p = 3). Given a(ℤ; 3) and c*(ℤ; 3), then the equation

c(k)z(k+1)+c(k1)z(k1)a(k)z(k)=0,kZ

has bounded solutions iff

|a(0)a(1)a(2)c(0)2a(2)c(1)2a(0)c(2)2a(1)|2|c(0)c(1)c(2)|

and when the inequality is strict, all the solutions are bounded. In particular, given E, λ ∈ ℝ, θ ∈ [0, 2π) andω=m3,where m ∈ ℤ and (3, m) = 1, then the Mathieu equation

z(k+1)+z(k1)+λcos(2πωk+θ)z(k)=Ez(k),kZ

has bounded solutions iff

|(Eλcos(θ))(E234λ2+λ2cos2(θ)+Eλcos(θ))3E|2.

## Proof

In this case we have Λ30 = {(0, 0, 0)} and Λ31 = {(1, 0, 0),(0, 1, 0),(0, 0, 1)}, which implies that Λ¯30 = {(1, 1, 1)} and Λ¯31 = {(0, 0, 1),(1, 0, 0),(0, 1, 0)} and hence,

q(a,c;3)=a(0)a(1)a(2)c(0)2a(2)c(1)2a(0)c(2)2a(1)2c(0)c(1)c(2).

In particular, for the Mathieu equation the condition for the existence of bounded solutions becomes

|a(0)a(1)a(2)a(0)a(1)a(2)|2.

On the other hand, the frequency is ω=m3 where (m, 3) = 1 which implies that ω=n+r3, where n ∈ ℤ and r = 1, 2. Therefore, the coefficient is given by ar(k)=Eλcos(π2r3k+θ),, and hence

a1(0)=a2(0)=Eλcos(θ),a1(1)=a2(2)=Eλcos(π23+θ)=E+λ2[cos(θ)+3sin(θ)],.a1(2)=a2(1)=Eλcos(π43+θ)=E+λ2[cos(θ)3sin(θ)]

□

## Corollary 4.7

(Period p = 4). Given a(ℤ; 4) and c*(ℤ; 4), then the equation

c(k)z(k+1)+c(k1)z(k1)a(k)z(k)=0,kZ

has bounded solutions iff

|a(0)a(1)a(2)a(3)c(0)2a(2)a(3)c(1)2a(0)a(3)c(2)2a(0)a(1)c(3)2a(1)a(2)+c(0)2c(2)2+c(1)2c(3)2|2|c(0)c(1)c(2)c(3)|

and when the inequality is strict, all the solutions are bounded. In particular, given E, λ ∈ ℝ, θ ∈ [0, 2π) andω=m4,where m ∈ ℤ and (4, m) = 1, then the Mathieu equation

z(k+1)+z(k1)+λcos(2πωk+θ)z(k)=Ez(k),kZ

has bounded solutions iff

4(E21)(E2λ2cos2(θ))(E2λ2sin2(θ))4E2.

## Proof

In this case we have Λ40 = {(0, 0, 0, 0)},

Λ41={(1,0,0,0),(0,1,0,0),(0,0,1,0),(0,0,0,1)},Λ42={(1,0,1,0),(0,1,0,1)}

which implies that Λ¯40 = {(1, 1, 1, 1)} and

Λ¯41={(0,0,1,1),(1,0,0,1),(1,1,0,0),(0,1,1,0)},Λ¯42={(0,0,0,0),(0,0,0,0)}

and hence,

q(a,c;4)=12c(0)c(1)c(2)c(3)[a(0)a(1)a(2)a(3)c(0)2a(2)a(3)c(1)2a(0)a(3)c(2)2a(0)a(1)c(3)2a(1)a(2)+c(0)2c(2)2+c(1)2c(3)2]

In particular, for the Mathieu equation the condition for the existence of bounded solutions becomes

4a(0)a(1)a(2)a(3)a(2)a(3)a(0)a(3)a(0)a(1)a(1)a(2)0.

On the other hand, the frequency is ω=m4 where (m, 4) = 1 which implies that ω=n+r4, where n ∈ ℤ and r = 1, 3. Therefore, the coefficient is given by ar(k) = Eλcos(πr2k+θ), and hence

a1(0)=a3(0)=Eλcos(θ),a1(1)=a3(3)=E+λsin(θ),a1(2)=a3(2)=E+λcos(θ),a1(3)=a3(1)=Eλsin(θ)

□

# Acknowledgement

This work has been partly supported by the Spanish Research Council (Comisión Interministerial de Ciencia y Tecnología,) under project MTM2014-60450-R.

### References

[1] Last, Y., Spectral theory of Sturm-Liouville operators on infinite intervals: A review of recent developments, in Sturm–Liouville Theory: Past and Present (W. O. Amrein, A. M. Hinz, D. B. Pearson, eds.), Birkhäuser, Basel, (2005), 99–120.10.1007/3-7643-7359-8_5Search in Google Scholar

[2] Avila, A., Jitomirskaya, S., The Ten Martini Problem. Ann. of Math. (2)170 (2009), 303–342.10.4007/annals.2009.170.303Search in Google Scholar

[3] Avila, A., Jitomirskaya, S., Solving the ten martini problem, in Mathematical physics of quantum mechanics, 5–16, Lecture Notes in Phys., 690, Springer, Berlin, 2006.10.1007/3-540-34273-7_2Search in Google Scholar

[4] Puig, J., Cantor spectrum for the almost Mathieu operator, Comm. Math. Phys. 244 (2004), 297–309.10.1007/s00220-003-0977-3Search in Google Scholar

[5] Encinas, A.M., Jiménez, M.J., Floquet Theory for second order linear difference equations, J. Diff. Eq. and App. 22 (2016), 353–375.10.1080/10236198.2015.1100609Search in Google Scholar

[6] Mouche, P. van, The coexistence problem for the discrete Mathieu operator, Comm. Math. Phys., 122 (1989), 23–33.10.1007/BF01221406Search in Google Scholar

[7] Simon, B., Schrödinger operators in the twenty-first century, in Mathematical Physics 2000 (A. Fokas, A. Grigoryan, T. Kibble, and B. Zegarlinski, eds.), Imp. Coll. Press, London, 2000, 283–288.10.1142/9781848160224_0014Search in Google Scholar

[8] Fonseca, C.M. da, Petronilho, J., Explicit inverse of a tridiagonal k-Toeplitz matrix, Numer.Math. 100 (2005), 457–482.10.1007/s00211-005-0596-3Search in Google Scholar

[9] Bendito, E., Encinas, A.M., Carmona, Á., Eigenvalues, Eigenfunctions and Green’s Functions on a Path via Chebyshev Polynomials, Appl. Anal. Discrete Math. 3 (2009), 282–302.10.2298/AADM0902282BSearch in Google Scholar

[10] Mason, J.C., Handscomb, D.C. Chebyshev Polynomials. Chapman & Hall/CRC, 2003.10.1201/9781420036114Search in Google Scholar