\begin{quotation}
\textit{If the numerical analysis is an art and not merely
a science, it is hoped that this art show will illustrate
the little tricks that are important in applications.}\\
\textbf{J. P. Boyd}, 1987
\end{quotation}
The foundations of spectral methods are not recent. For many years, before the appearance of
contemporary computers, the theoretical studies in mathematical physics, and especially in fluid mechanics,
made an extensive use of series expansions using the so-called "special functions".
The main aim of this work is to show that spectral collocation, based on
Laguerre, Hermite and Sinc functions can efficiently and reliably solve initial and/or boundary
value problems, as well as eigenvalue problems, formulated on unbounded domains. The expected convergence rates are $O\left(e^{-cN}\right)$ or
at least $O\left(e^{-c\sqrt{N}}\right)$ where $N$ is the number of degrees of freedom in the expansion of solution,
i.e., the approximation order, and $c$ is a constant independent of $N$. These preoccupations on spectral methods are
originated in our previous text \cite{CIG07}.
A major advantage in using these basis functions consists in the fact that for such problems boundary conditions are
usually “natural” rather than “essential” in the sense that the singularities of the differential
problem will force the numerical solution to
have the correct behavior at infinity even if no constraints are imposed on these functions.
In this way we definitely refuse the use of empirical \textit{domain truncation} along with \textit{various
shooting techniques}.
We mainly focus on the computational and practical aspects of this type of collocation. We do not
prove theorems concerning analytical aspects of this method but we carry out a lot of numerical
experiments and then analyze and compare their outcomes.\footnote{\label{WMK} And it is
not uncommon for proofs to be wrong. Philip Davis has done a survey in mathematical proofs
and found that something in the order of third of them have serious deficiencies-a tenth of them,
uncorrectable-but, thank goodness, nobody cares about most of those theorems.
Therefore you also have to have \textit{\textbf{testing}}. Testing requires a level of
deviousness and ingenuity and understanding that transcends what is needed to write the
programm in the first place. William M. Kahan, 2005 (see: \texttt{history.siam.org/})}
Accordingly we utilize these numerical investigations in order to find \textit{the goods and the
bads} of spectral collocation applied to various problems formulated on unbounded domains.
Thus, in the first Chapter we carry out a review of the main weighted interpolation
techniques available for unbounded intervals. We consider Hermite and Laguerre
interpolations based on homonymous functions as well as the interpolation
based on Sinc functions. Some mapping techniques, which bring into play the
best approximation polynomials, i.e., the Chebyshev polynomials, are also
reviewed along with some elementary preconditioning. Technical aspects concerning polynomial transforms and boundary condition implementation
are also shortly examined.
Most important, we make use of \textit{a strategy to evaluate the convergence rate} (the accuracy) of our collocation solutions.
Namely, from the physical space, where every collocation works, we transfer the solution into the coefficient (phase) space.
In this aim we use FFT and polynomial transforms. Then, in a log linear plot, we analyse (measure) the rate at which the expansion
coefficients of solution decrease to zero. In this way a rough estimation of the accuracy becomes possible.
This concept is graphically illustrated.
In the second Chapter we use LC and LGRC in order to solve boundary value problems on the half line and
HC and SiC to solve problems on the real line. We successively consider a linear test problem,
a class of nonlinear second order t. p. b. v. p.,
the modified Troesch's and Thomas-Fermi's problems, a semiconductor problem and the Kidder's problem.
The Falkner-Skan equation coupled with the heat transfer equation and the Blasius viscous flow of a kind of a
non-Newtonian fluid are briefly reviewed. In order to assess the \textit{accuracy} of our outcomes we
use the strategy introduced in the first Chapter. For a linear test problem
on the real line we determine the order of convergence of SiC.
In the third Chapter we deal with some singular eigenvalue problems as well as with
eigenvalue problems supplied with boundary conditions depending on the
eigenparameter. As a typical example of eigenvalue problem with eigenparameter dependent boundary conditions we have
chosen the so called Fourier problem. This is a second order
differential problem well studied analytically. This has enabled us some
important validation of our numerical results. The second problem is the
so called Charney's baroclinic stability problem which involves an analytic
singularity.
The Schr\"{o}dinger eigenvalue problems on the half as well as on the real line occupy
a large space in this Chapter. We attach various potentials, bounded or unbounded, to these problems.
A particular attention is paid to the challenging issue of \textit{continuous spectra} and
numerical (discrete) eigenvalues. It is known that some spectral problems for differential operators which are
naturally posed on the whole real line,
often lead to discrete eigenvalues plus a continuous spectrum. Any numerical approximation
typically involves three processes: (a) reduction to a finite interval; (b) discretization;
(c) application of a numerical eigenvalue solver.
Reduction to a finite interval and discretization typically eliminates the continuous
spectrum. Even if we do not truncate, a priori and arbitrary, the domain on which the problem is
formulated, such an inherent reduction can not be avoided.
Two techniques are used in order to eliminate the "bad" eigenvalues as well as to estimate the accuracy
of computations. The first one is the \textit{drift}, with respect to the order of approximation
and the scaling factor, of some eigenvalues of interest. The second one is based on the check of the \textit{eigenvectors orthogonality}.
We have slightly extended the notion of drift introduced by John Boyd
and confirmed its validity in appreciating the correctness of computed eigenvalues.
The algebraic generalized eigenvalue problems involved have been solved by various methods including
some subspace methods of Jacobi-Davidson type.
Actually in this Chapter we carry on the main topic of our recent monograph \textit{Spectral Methods
for Non Standard Eigenvalue Problems. Fluid and Structural Mechanics and Beyond}, Springer, 2014.
In the introductory part of the fourth Chapter we consider some linear multidimensional boundary value problems
which can be reduced to one dimensional t. p. b. v. p. (various self-similar solutions to quasilinear parabolic equations).
The main part of this Chapter is devoted to
\textit{weakly nonlinear p. d. e. of evolution}. The Benjamin Bona Mahony (BBM),
modified BBM and Benjamin Bona Mahony-Burgers (BBM-B),
the K\"{o}rteweg-de Vries (KdV), and the nonlinear Schr\"{o}dinger (NLS) initial value problems on the real axis are important
examples of such problems. All these equations involve linear dispersion.
For BBM and BBM-B cases we have some preliminary results in our recent contributions.
In order to solve such
problems we use the \textit{method of lines}, i.e., we couple various high order
collocation methods for the discretization of spatial variables with some finite difference schemes in order to march in time.
Then, we prove the linear stability of the method. In order to determine the regions of
linear stability we use the pseudospectra of the discrete linearized operators (matrices).
Particular radially symmetric solutions (waveguide solutions) as well as general radially symmetric solutions
to NLS have been computed. In the latter case we take into account Gaussian and Lorentzian initial data. Blow-up
self similar solutions to the cubic NLS and their convergence to Townes solitons is also examined.
Additionally, we pay attention to accurate approximation of various bound states of multiple
"solitons" solutions to cubic NLS.
Whenever there exist invariants of the above problems, a special attention is paid to their conservation.
The importance of such invariants is two folded. First, we can use them in order to optimize the
scaling parameter. Second, they provide useful hints in order to eliminate spurious solutions of
some nonlinear boundary value problems solved by collocation.
Whenever, a boundary value problem exhibits some \textit{first integrals} we can use the \textit{modal persistence} or alternatively
the $\rho$-\textit{criterion} in order to assure the numerical stability and to eliminate the spurious solutions.
In the fifth Chapter we provide some typical MATLAB scripts. They could improve the understanding
of the text and could open the way for various numerical experiments.
We end up with a Concluding Remarks Chapter. The existence of a
scaling parameter in all interpolation techniques considered, enables one to avoid
the empirical domain truncation and to adjust the numerical outcomes. By trial and error we can improve the
accuracy, stability and the conservation of various invariants. Unfortunately a rigorous and
automatic way to perform the tuning of scaling parameters is still missing.
A Section containing the acronyms used throughout the work has been provided in order to facilitate the reading.
Throughout this work the numerical results have been carried out using the MATLAB
environment on a HP xw8400 workstation with clock speed of 3.2 Ghz.
The author acknowledges the friendly and lucrative climate from "Tiberiu Popoviciu"
Institute of Numerical Analysis, Cluj-Napoca, Romania, which has decisively contributed to
the accomplishing of this volume.
Last but not least, I will fairly appreciate any remark, suggestion and correction concerning this text. These will be a relevant indicator
of this open project, and I promise to address them all and incorporate them in the future editions.
\vspace{\baselineskip}
\begin{flushright}\noindent
Cluj-Napoca, Romania \hfill {\it C\u{a}lin-Ioan Gheorghiu}\\
January 2018 \hfill \\
\end{flushright}