1.1 About spectral methods
When doing simulations and solving PDEs, one faces the problem of representing and deriving functions
on a computer, which deals only with (finite) integers. Let us take a simple example of a function
. The most straightforward way to approximate its derivative is through finite-difference
methods: first one must setup a grid
points in the interval, and represent by its values on these grid points
Then, the (approximate) representation of the derivative shall be, for instance,
If we suppose an equidistant grid, so that , the error in the
approximation (1) will decay as (first-order scheme). One can imagine higher-order schemes, with
more points involved for the computation of each derivative and, for a scheme of order , the accuracy
can vary as .
Spectral methods represent an alternate way: the function is no longer represented through its values
on a finite number of grid points, but using its coefficients (coordinates) in a finite basis of
A relatively simple case is, for instance, when is a periodic function of period two, and the
are trigonometric functions. Equation (2) is then nothing but the
truncated Fourier decomposition of . In general, derivatives can be computed from the ’s, with the
knowledge of the expression for each derivative as a function of . The
decomposition (2) is approximate in the sense that represent a complete basis of some
finite-dimensional functional space, whereas usually belongs to some other infinite-dimensional space.
Moreover, the coefficients are computed with finite accuracy. Among the major advantages of using
spectral methods is the rapid decay of the error (faster than any power of , and in practice often
exponential ), for well-behaved functions (see Section 2.4.4); one, therefore, has an infinite-order
In a more formal and mathematical way, it is useful to work within the methods of weighted residuals
(MWR, see also Section 2.5). Let us consider the PDE
where is a linear operator, the operator defining the boundary conditions and is a source term.
A function is said to be a numerical solution of this PDE if it satisfies the boundary conditions (4) and
makes “small” the residual
If the solution is searched for in a finite-dimensional subspace of some given Hilbert space (any relevant
space) in terms of the expansion (2), then the functions are called trial functions
and, in addition, the choice of a set of test functions defines the notion of smallness for the
residual by means of the Hilbert space scalar product
Within this framework, various numerical methods can be classified according to the choice of the trial
- Finite differences: the trial functions are overlapping local polynomials of fixed order (lower
- Finite elements: the trial functions are local smooth functions, which are nonzero, only on
subdomains of .
- Spectral methods: the trial functions are global smooth functions on .
Various choices of the test functions define different types of spectral methods, as detailed in
Section 2.5. Usual choices for the trial functions are (truncated) Fourier series, spherical harmonics or
orthogonal families of polynomials.