Abstract
This paper proposes an algorithm for the numerical reconstruction of the coefficient (also called ‘potential’) function in the canonical Sturm–Liouville differential operator from three known sequences of its eigenvalues. These correspond respectively to three sets of mixed boundary conditions: Dirichlet type for the interval , Dirichlet–Robin type for the interval and Robin–Dirichlet type for the interval , where is an interior point of the interval . This problem originates in an engineering application involving a string and its frequencies of oscillations when set into vibration. The method will be illustrated with numerical examples including both continuous and discontinuous potential functions.
Introduction
This paper proposes an algorithm to numerically reconstruct the unknown potential of the canonical Sturm–Liouville operator from three spectra, corresponding to three sets of mixed boundary conditions. This problem originates in the following vibrating string problem: an elastic, inaccessible string of negligible weight fixed at the end points is set into infinitesimal vertical vibrations. Its frequencies of oscillation are measured. Next, two more sets of frequencies of oscillation are measured, one set for each piece of the string, when the string has now an interior node attached to a spring with a known stiffness constant. Each piece of the string is set into infinitesimal, vertical vibrations independently of the other. From these measurements, one would like to determine the density of the string. Some details about problems of this type can be found in [Citation1, Section 10.1].
Physical and mathematical arguments turn this problem into the following inverse spectral problem: obtain the coefficient (also called ‘potential’) function of the canonical Sturm–Liouville differential operator from three known sequences of real numbers with specific properties , , and . These sequences will be the eigenvalues of the Sturm–Liouville operator mentioned above corresponding respectively to Dirichlet type boundary conditions on , Dirichlet–Robin type boundary conditions on , and Robin–Dirichlet type boundary conditions on , where is an interior point of , and the boundary parameter at the interior point being a given real number. Specifically, , and are respectively eigenvalues of such type, if there exist some non-identically zero functions (called eigenfunctions) , , such that:(1.1) (1.1) (1.2) (1.2) (1.3) (1.3) Here , for some constant and some .
The overview of the method proposed is as follows (the meaning of the notations and concepts, and the details of the method will be provided in the subsequent sections). If is the unknown coefficient function of the canonical Sturm–Liouville operator to be recovered from the three types of eigenvalue sequences , and , then these sequences will be respectively the zeros of the characteristic function for each type of eigenvalue problem (), (), or (). This information along with the following identity:(1.4) (1.4) will suggest a way of obtaining two pairs of Cauchy data. One pair will correspond to and the other to which is the ‘reflection’ of about the line . Since the recovery of the potential function from one pair of Cauchy data is already known in the literature (see [Citation2]), our efforts will be focused towards determining two pairs of Cauchy data for the interval , by following the guidelines suggested by the characteristic functions and the identity (Equation1.4(1.4) (1.4) ). Once the Cauchy data are obtained, two coefficient functions for will be produced. Pasting together the first with the ‘reflection’ of the second, the coefficient function for the entire interval will be obtained.
The inverse Sturm–Liouville problem by three Dirichlet spectra (i.e. non-mixed boundary conditions) was discussed in [Citation3]. The present work distinguishes itself from [Citation3] in the following aspects: at the interior node the boundary condition is now of Robin type as opposed to Dirichlet type discussed in [Citation3]; the fundamental identity (2.7) in [Citation3] now becomes (Equation1.4(1.4) (1.4) ) to highlight the characteristic function of the Sturm–Liouville problem with Dirichlet–Robin boundary conditions on the interval . Another difference between [Citation3] and the present work resides in the fact that the kernels in the integral representations of the characteristic functions for the Dirichlet eigenvalue problem and for the Dirichlet–Robin eigenvalue problem on , respectively are different. Also, in [Citation3] the kernel in the integral representation of the characteristic functionfor the Dirichlet eigenvalue problem on the interval , and the kernel in the integral representation of the functionwere used to infer information about the Cauchy data. By contrast, in the present paper, information about the Cauchy data was obtained from the kernel in the integral representation of the characteristic functionfor the Dirichlet–Robin eigenvalue problem on the interval with boundary parameter , and the kernel in the integral representation of the functionVarious aspects on inverse three spectra problems were discussed by many other authors, and we mention here only a few: [Citation4–Citation9].
Preliminaries
Some notations are in order. Let and . We denote by and the weak (i.e. in ) solutions to the initial value problems:(2.1) (2.1) and respectively(2.2) (2.2) The notations were chosen suggestively, to remind of the solutionsto (Equation2.1(2.1) (2.1) ) and respectively (), when .
Note that is a Dirichlet–Robin eigenvalue on the interval of the canonical Sturm–Liouville operator (recall ) with boundary parameter if and only ifThis can be seen as follows. If is such an eigenvalue, let be its corresponding eigenfunction. So such that , and . Hence, , since otherwise the Dirichlet boundary condition along with the ODE would imply that . (The initial value problem that consists of this ODE and the initial conditions has only one solution, and the identically zero function is a solution to it). So is well defined. Moreover, it satisfies the above ODE since does. It also satisfies the initial conditions and . Therefore, , for all , because is the only solution of this initial value problem. It follows then thatConversely, if is such that , then by the definition of above, will be the eigenfunction corresponding to for the Dirichlet–Robin boundary conditions. This means that is a Dirichlet–Robin eigenvalue on of the Sturm–Liouville operator .
So, the Dirichlet–Robin eigenvalues on of with boundary parameter are exactly the zeros of the functionFor this reason this function is called the characteristic function of the canonical Sturm–Liouville operator on with Dirichlet–Robin boundary conditions and boundary parameter , by analogy with the matrix case. (A linear operator between two finite dimensional spaces can be represented by a matrix).
By similar arguments, the Dirichlet eigenvalues on of are exactly the zeros of the functioncalled the characteristic function of the Sturm–Liouville operator on with Dirichlet boundary conditions.
Another important fact to be used later is that is a Robin-Dirichlet eigenpair of the Sturm–Liouville operator on with boundary parameter if and only if is a Dirichlet–Robin eigenpair of the Sturm–Liouville operator on with boundary parameter , where , , , , with . (Here and ). This statement can be easily verified by the definition of an eigenpair corresponding to Robin-Dirichlet boundary conditions and respectively to Dirichlet–Robin boundary conditions, and the changes of variables indicated above.
It is known in the literature (see [Citation10, p.255] and rescale to by the change of variables , for ) that the spectrum of the canonical Sturm–Liouville operator on with Dirichlet–Robin boundary conditions and boundary parameter is a sequence of real, simple eigenvalues, increasing and satisfying the asymptotic formula:(2.3) (2.3) where is an sequence of real numbers (i.e. ). Here we denote the mean of by (i.e. ).
It is also known (see [Citation10, p.256], or [Citation11, Theorem 1 (p.29), Theorem 2 (p.30), Theorem 4 (p.35)], or [Citation12, Lemma 4.7 (p.135-136) and formula (4.25) on p.139], and rescale to ) that the spectrum of the canonical Sturm–Liouville operator on with Dirichlet boundary conditions is a sequence of real, simple eigenvalues, increasing and satisfying the asymptotic formula:(2.4) (2.4) where is an sequence of real numbers (i.e. ). Here, as before, .
Therefore, if and , , are respectively the Dirichlet eigenvalues on , the Dirichlet–Robin eigenvalues on , and the Robin-Dirichlet eigenvalues on of the Sturm–Liouville operator with boundary parameter at the interior node , then letting , we can write by using () and () with the proper choices of and , and the paragraph preceding the paragraph of ():(2.5) (2.5) (2.6) (2.6) (2.7) (2.7) for some real-valued sequences , , and in . Here, , for given by , , and . Dividing (Equation2.7(2.7) (2.7) ) by we obtain:(2.8) (2.8) for .
Other important facts here are the integral representations with Gelfand–Levitan–Marchenko kernel of the function and of its derivative . They are (see [Citation12, Theorem 4.18 and Example 4.19, p.154–157] with similar calculations for ):(2.9) (2.9) which by differentiating with respect to gives:(2.10) (2.10) where the Gelfand–Levitan–Marchenko kernel is , the week solution to the Goursat boundary value problem:(2.11) (2.11) We introduce some notations. For , the subscript or attached to and is meant to refer to the first subinterval of , or to the second subinterval of . That is for instance, means , and means . With these being said and , and , as described in the paragraph of Equations (Equation2.5(2.5) (2.5) ) –(Equation2.8(2.8) (2.8) ), we recall Equation (2.7) of [Citation3]:Since at the interior node we have now a Robin type boundary condition, it is to our benefit to put the above equation into one that highlights the characteristic function of the canonical Sturm–Liouville operator with Dirichlet–Robin boundary conditions on . So we subtract and add in the right-hand side of the above equation, and then suitably combine the terms and factor out:which is Equation (Equation1.4(1.4) (1.4) ) mentioned in Section 1. Here .
The method
Consider the canonical Sturm–Liouville differential operator having , , as its Dirichlet, Dirichlet–Robin and Robin–Dirichlet eigenvalues on , , and where the boundary parameter at the interior node is , a known number. Then, as noted in Section Section12 each sequence is a sequence of real numbers, increasing, the eigenvalues are simple and satisfy (2.5), (Equation2.6(2.6) (2.6) ), (Equation2.8(2.8) (2.8) ). It also follows that are the Dirichlet eigenvalues on corresponding to , are the Dirichlet–Robin eigenvalues on corresponding to and boundary parameter , and are the Dirichlet–Robin eigenvalues on corresponding to and boundary parameter (see paragraph 5 in Section Section12). These further imply that (see paragraphs 2–4 in Section Section12):(3.1) (3.1) (3.2) (3.2) (3.3) (3.3) Equation (Equation1.4(1.4) (1.4) ) holds as well, and making in (Equation1.4(1.4) (1.4) ) and using (Equation3.2(3.2) (3.2) ) we obtain:(3.4) (3.4) Then making in (Equation1.4(1.4) (1.4) ) and using (Equation3.3(3.3) (3.3) ) we obtain:(3.5) (3.5) Note that if and only if is a Dirichlet–Robin eigenvalue corresponding to and boundary parameter (see paragraphs 2–3 in Section Section12), which further happens if and only if is a Robin–Dirichlet eigenvalue corresponding to and boundary parameter (see paragraph 5 in Section Section12). That means, if and only if. But if so, then by () we would have , which means (see paragraph 4 in Section Section12) that is a Dirichlet eigenvalue corresponding to (i.e. ).
Also, if and only if is a Dirichlet–Robin eigenvalue corresponding to and boundary parameter (see paragraphs 2–3 in Section Section12). That means, if and only if. But if so, then by () we would have , which means (see paragraph 4 in Section Section12) that is a Dirichlet eigenvalue corresponding to (i.e. ).
For these reasons we shall assume further that no two of the sets , , overlap. It follows then by (Equation3.4(3.4) (3.4) ) and (Equation3.5(3.5) (3.5) ) that the quantities and are well defined. Specifically, they are given by:(3.6) (3.6) and respectively by:(3.7) (3.7) Let , for , , for , and , for be respectively the solutions to the Goursat problems of type () with the appropriate choices of and . Then using (Equation2.9(2.9) (2.9) ), (Equation2.10(2.10) (2.10) ) with and the appropriate choices of and , and Equations (Equation3.1(3.1) (3.1) ), (Equation3.2(3.2) (3.2) ), (Equation3.3(3.3) (3.3) ) we obtain:(3.8) (3.8) (3.9) (3.9) (3.10) (3.10) Multiplying (Equation3.8(3.8) (3.8) ) by , (Equation3.9(3.9) (3.9) ) by , and (Equation3.10(3.10) (3.10) ) by , and solving for the integral terms we obtain:(3.11) (3.11) (3.12) (3.12) (3.13) (3.13) where stands for the mean of the function .
Letting denote the right-hand side of (Equation3.6(3.6) (3.6) ), and denote the right-hand side of (Equation3.7(3.7) (3.7) ), and in (Equation3.6(3.6) (3.6) ) and (Equation3.7(3.7) (3.7) ) using again (Equation2.9(2.9) (2.9) ) with and the appropriate choices of and , and solving for the integral terms we obtain:(3.14) (3.14) (3.15) (3.15) We note now that solving the system of equations () for , and the system of equations (Equation3.14(3.14) (3.14) ) for , we get in the possession of one pair of Cauchy data, namely , because is prescribed. We can even get by direct differentiation of . It is actually the pair that we need in order to use the algorithm in [Citation2] to produce .
Also, solving the system of equations () for , and the system of equations (Equation3.15(3.15) (3.15) ) for we obtain a second pair of Cauchy data, namely from which the pair is immediately obtained and then used in the algorithm of [Citation2] to produce .
To solve numerically (Equation3.11(3.11) (3.11) ) – (Equation3.15(3.15) (3.15) ), we shall use a truncated Fourier series expansion to represent each of the functions , , , and . What would be an appropriate orthogonal basis of the space in each case? It depends on the known (or otherwise inferred) property of the function and of the eigenvalue sequences. The sequences, , and we use for the reconstruction are real-valued, positive, strictly increasing and satisfying the asymptotic formulas:(3.16) (3.16) (3.17) (3.17) (3.18) (3.18) So, because , the asymptotic formula:(3.19) (3.19) will hold as well. In (Equation3.16(3.16) (3.16) ) – (Equation3.19(3.19) (3.19) ), , , and are some real-valued sequences, and , , , are real-valued constants related in a specific way (to be discussed at the end of this section).
About we know that:and thatThe first equality is due to the first boundary condition of () for and , and the last equality follows by comparing the given asymptotics () with the asymptotics () the ’s are supposed to satisfy, since they are expected to be the Dirichlet eigenvalues on corresponding to . From (Equation3.16(3.16) (3.16) ), we have thatand so is close to , for large, making the matrixalmost diagonal, so easy to invert. All of these suggest we use as an orthogonal basis of to represent not but rather , because both - the basis functions, and this latter function are zero at and . Note thatis an orthonormal basis of . (See [Citation13, Example 2(d1), p.308-309].) So we write:(3.20) (3.20) Here is the number of available ’s. Inserting (Equation3.20(3.20) (3.20) ) into (Equation3.11(3.11) (3.11) ), we obtain a linear system to be solved for the Fourier coefficients ’s:(3.21) (3.21) A similar discussion pertains to and also to . So we write:(3.22) (3.22) where , and is the number of available ’s. Also,(3.23) (3.23) where (because , with given by – see the paragraph of () - ()), so , where . Here is the number of available ’s (and so of ’s). Inserting () into (Equation3.14(3.14) (3.14) ), and (Equation3.23(3.23) (3.23) ) into (Equation3.15(3.15) (3.15) ), we obtain:(3.24) (3.24) and(3.25) (3.25) From (Equation3.14(3.14) (3.14) ) we only kept the first equations, and from () only the first equations to make square systems for the unknowns , and respectively . Note that the coefficient matrices in () and ()and respectivelyare no longer almost diagonal, because by () and () we have:andand so and are close to , for large.
Next, the appropriate orthogonal basis for the Fourier series representation of and of is , because these functions are integrated against , and respectively (see () and ()), which we showed above are close toThe appropriateness of as an orthogonal basis in is strengthen by the fact that each of and is zero when (by the second condition of the Goursat problem () with the appropriate choices of and ), as the functions are. Note that nothing can be said about and when . The fact that is an orthogonal basis of follows from the Auxiliary result 1 with (see Section Section17). So we write:(3.26) (3.26) and(3.27) (3.27) Inserting () into (), and () into (), and keeping only the first equations from (), and respectively only the first equations from () we obtain two square linear systems to be solved for the Fourier coefficients and , respectively:(3.28) (3.28) (3.29) (3.29) (Recall that , and .) The coefficient matrices in () and (), respectivelyare almost diagonal because and are close to , as discussed previously.
A great attention must be paid to and . The linear systems () and () can be solved for and , respectively only if ’s and ’s are known. Recall the definitions of and in the paragraph of (Equation3.14(3.14) (3.14) ) – (Equation3.15(3.15) (3.15) ). Hence, using also () and () with and the appropriate choices of and we can write:(3.30) (3.30) and respectively(3.31) (3.31) Due to (Equation3.20(3.20) (3.20) ), () and (), Equations () and () become:and so(3.32) (3.32) and respectivelyand so(3.33) (3.33) because . In () and (), the Fourier coefficients , , are the solutions to (), () and (), respectively. So and can be calculated this way, and we can proceed with the calculations of and via () and (). Hence, approximations (via truncated Fourier series) for each of , , , and will be available. Therefore, as mentioned in the two paragraphs immediately following (), and are the potential functions on that correspond to the pair of Cauchy data and respectively to the pair , andwhere , with given by .
The information gathered here will serve as a model in Section Section15. We need to say a few words about the relationship between the constants , , , mentioned in ()– (Equation3.19(3.19) (3.19) ), and how to estimate , , , from the given asymptotics of ’s, ’s, ’s. We shall work backwards from (Equation3.16(3.16) (3.16) ) – (Equation3.18(3.18) (3.18) ), since these will be assumed known in the inverse problem. It follows immediately from (Equation3.18(3.18) (3.18) ) and (Equation3.19(3.19) (3.19) ) and the definition of (i.e. , for ) that . Next, comparing (Equation3.16(3.16) (3.16) ) - (Equation3.18(3.18) (3.18) ) with the asymptotics (Equation2.5(2.5) (2.5) ), (Equation2.6(2.6) (2.6) ) and (), the ’s, ’s, ’s are supposed to satisfy, since they are expected to be the Dirichlet, Dirichlet–Robin and Robin-Dirichlet eigenvalues corresponding to the potential function and boundary parameter on , and , respectively, we can write:(3.34) (3.34) Since , it follows by (Equation3.34(3.34) (3.34) ) that:(3.35) (3.35) Next, from (Equation3.16(3.16) (3.16) ) – (Equation3.18(3.18) (3.18) ) we infer that:It follows from these and (Equation3.34(3.34) (3.34) ) that if only , , are available, then:
Some remarks on (Equation3.21(3.21) (3.21) ), (), (), (), ()
The reason we chose to solve for the first Fourier coefficients ’s of , and not for any other coefficients is that the sequence of the partial sums converges to , and not any other sequence of finite sums. As another remark, from (Equation3.11(3.11) (3.11) ) we kept only equations in order to make a square system for the unknowns . If less than equations were kept, say , then the corresponding linear system for the ’s would have had more than one solution since at least of the ’s would have become parameters. On the other hand, if more than equations are kept, say , then it would be possible for some of the equations to be incompatible with each other, thus making the linear system not have a solution. Finally, we kept the first equations from (Equation3.11(3.11) (3.11) ), i.e. the equations that correspond to , and not any other equations. Had we done otherwise, say by keeping the equations that correspond to with some not all in , then certainly would have been pushed outside the set . This would implyThat means that the last row of the matrix associated with the linear system for the ’s would have had all entries near zero, making the numerical computations of the ’s difficult.
By similar arguments, the first Fourier coefficients of and the first equations from () must be kept. Also, the first Fourier coefficients of and the first equations from () must be kept. That means that , and must be involved. This implies that the first equations from (), and the first equations from (Equation3.15(3.15) (3.15) ) must be kept, and so the first Fourier coefficients of , and the first Fourier coefficients of must be kept.
Therefore, the need for the first of the ’s, ’s, ’s, respectively, is justified.
The numerical algorithm
Let be a given real number. Let , , be three sequences of positive numbers, strictly increasing, satisfying the asymptotics (Equation3.16(3.16) (3.16) ) – (Equation3.18(3.18) (3.18) ), for some real-valued sequences , , in , and some real-valued constants , , such that () holds. We assume further that no overlap holds between , , , and they interlace in the following sense:where arranged increasingly. Then , , are the Dirichlet, Dirichlet–Robin and Robin-Dirichlet eigenvalues on , and , respectively, with boundary parameter at , for some potential function .
Note that in practice only a few eigenvalues of each type are available: , , . But this is enough information for us to estimate the constants in () – (Equation3.18(3.18) (3.18) ). Now we can present the reconstruction algorithm.
Introduce the constants , , , by:
Define
Solve the linear system (Equation3.21(3.21) (3.21) ) for .
Solve the linear system () for .
Solve the linear system () for .
Construct by ().
Construct by ().
Solve the linear system () for .
Solve the linear system () for .
Defineand(The pair is supposed to be the pair .)
Defineand(The pair is supposed to be the pair .)
Defineand
Use a divided difference scheme to obtain from the known and .
Use a divided difference scheme to obtain from the known and .
Use in the algorithm of [Citation2] to produce a function on .
Use in the algorithm of [Citation2] to produce a function on .
’Reflect’ about the line to produce a new function on . Precisely,
Paste together and to produce the function on .
The best numbers of eigenvalues
In this section we are concerned with finding the best numbers , , such that the sequences , , assure a good reconstruction of the potential function . We start by noting that the quantities and are needed to solve the linear systems () and () for and , respectively. We see in () that the quantityand the numerator of in (),are bounded (uniformly in ). Hence, in order for the right-hand side of () to be well defined, we need the denominator of in () to be non-zero, for all . Theoretically, they are so due to the non-overlap of with . (See the three paragraphs immediately following ().) Numerically, they should not even be close to zero. So we need to require:(6.1) (6.1) Similarly, in order for the right-hand side of () to be well defined, we need to require that the denominator of in () stay away from zero, for all . That is, we need to require:(6.2) (6.2) We argue below that () and () will impose a constraint on the numbers and . To this end, we shall take advantage of the orthogonality in of the set . We claim that if we hope () and () to hold, then:(6.3) (6.3) For if , then . Let be the positive integer such that . Then , so andIt follows that(because and imply ) from which we have that(6.4) (6.4) Also,(6.5) (6.5) And(6.6) (6.6) Formulas (Equation6.4(6.4) (6.4) ), (Equation6.5(6.5) (6.5) ) and (Equation6.6(6.6) (6.6) ) show that () does not hold for .
On the other hand, if , then letting be the positive integer such that we have that , so . It follows thatfrom which we can write:(because and imply ). This gives(6.7) (6.7) Also,(6.8) (6.8) And(6.9) (6.9) Formulas (Equation6.7(6.7) (6.7) ), (Equation6.8(6.8) (6.8) ), and (Equation6.9(6.9) (6.9) ) show that () will not hold for .
Thus, these calculations justify the constraint imposed by () and (): the numbers and must be taken such that (Equation6.3(6.3) (6.3) ) holds. Bearing in mind the interlacing of , , (see the beginning of Section Section15), we also write:(6.10) (6.10)
Numerical examples
To illustrate the effectiveness of the algorithm presented in Section Section15, specific ’s were chosen and used in the software MATSLISE (see [Citation14]) to produce three sequences of eigenvalues: Dirichlet on , Dirichlet–Robin on , and Robin-Dirichlet on , with boundary parameter at the interior node . These sequences were then used as the input data in the reconstruction algorithm of Section Section15. The comparison between the true potential (i.e. the one with which we generated the eigenvalues) and the numerically constructed potential is featured in Figure . In each picture, we indicate the three sets of boundary conditions for each inverse eigenvalue problem. Also the numbers , , of the ’s, ’s, ’s that we used are indicated. The potential functions we chose to reconstruct are (from left to right, top to bottom):with ,with ,with ,with ,with ,with .
Acknowledgments
I would like to thank Dr. Paul E. Sacks of Iowa State University for helpful discussions on the general topic of numerical reconstructions for inverse Sturm–Liouville problems.
References
- GladwellGML. Inverse problems in vibration. Dordrecht: Kluwer Academic; 2004.
- RundellW, SacksPE. Reconstruction techniques for classical inverse Sturm-Liouville problems. Math. Comput. 1992;58:161–183.
- Drignei MC. Constructibility of an solution to an inverse Sturm Liouville problem using three Dirichlet spectra. Inverse Prob. 2010;26. 025003. doi:10.1088/0266-5611/26/2/025003
- Boyko O, Pivovarchik V. The inverse three-spectral problem for a Stieltjes string and the inverse problem with one-dimensional damping. Inverse Prob. 2008;24. 015019. doi:10.1088/0266-5611/24/1/015019
- CoxS, EmbreeM, HokansonJ. One can hear the composition of a string: experiments with an inverse eigenvalue problem. SIAM Rev. 2012;54:157–178.
- Gesztesy F, Simon B. On the determination of a potential from three spectra. Differential operators and spectral theory, 85–92, Amer. Math. Soc. Transl. Ser. 2, 189, Amer. Math. Soc., Providence, RI; 1999.
- Hryniv RO, Mykytyuk YaV. Inverse spectral problems for Sturm-Liouville operators with singular potentials, III. Reconstruction by three spectra, J. Math. Anal. Appl. 2003;284:626–646.
- MichorJ, TeschlG. Reconstructing Jacobi matrices from three spectra, Spectral methods for operators of mathematical physics. Oper. Theory Adv. Appl. 2004;154:151–154.
- PivovarchikVN. An inverse Sturm-Liouville problem by three spectra. Integr. Equat. Operator Theory. 1999;34:234–243.
- DahlbergBEJ, TrubowitzE. The inverse Sturm-Liouville problem III. Communications on Pure and Applied Mathematics. 1984;37:255–267.
- Pöschel J, Trubowitz E. Inverse spectral theory, pure and applied mathematics. Vol. 130. Orlando, FL: Academic Press; 1987.
- KirschA. An introduction to the mathematical theory of inverse problems, applied mathematical sciences 120. New York: Springer-Verlag; 1996.
- StakgoldI. Green’s functions and boundary value problems. Pure and Applied Mathematics. New York: John Wiley and Sons; 1998.
- Matslise, http://users.ugent.be/vledoux/MATSLISE/
Appendix
Auxiliary result 1 Let . Then is an orthogonal basis of .
Proof: We shall make use of the known statement: ‘Let be an orthogonal sequence in a separable Hilbert space . Then is an orthogonal basis of if and only if is complete (i.e. the zero vector alone is perpendicular to all ’s).’ So, in order to establish the conclusion it is enough to prove that is an orthogonal sequence in the Hilbert space , and it is complete in . We check orthogonality first. Let . Then using the trigonometric formulawe obtain:Now we prove the completeness in of . We shall make use of the completeness in of . This latter sequence of functions is complete in due to the statement mentioned above (about the equivalence between orthogonal basis and complete sequence), and because it is an orthogonal basis of (see [Citation13, Example 2(d1), p.308–309] with ). Let be such that(8.1) (8.1) Consider the even reflection of about the line . That is, define(8.2) (8.2) Then for all :(8.3) (8.3) () and the completeness in of imply that , for all . Hence, by () we obtain , for . Thus, the completeness in of is proved.