Numéricamente, se obtiene unos resultados muy interesantes para la matriz $M_{ij}=(i-1)^{j-1}$ si se expresa en la base de Fourier con la alternancia de la fila de signos,
$$A := \begin{bmatrix}1 \\ & -1 \\ && 1 \\&&& -1 \\ &&&& \ddots \end{bmatrix}\mathcal{F}^{-1} \left[\begin{matrix}
0^0 & 0^1 & 0^2 & 0^3 & \ldots\\
1^0 & 1^1 & 1^2 & 1^3 & \ldots\\
2^0 & 2^1 & 2^2 & 2^3 & \ldots\\
3^0 & 3^1 & 3^2 & 3^3 & \ldots\\
\vdots & \vdots & \vdots & \vdots &\ddots
\end{de la matriz}\right] \mathcal{F}.$$
As the discretization size of $M$ goes up, $M$ becomes extremely ill-conditioned and the entries blow up, so numerical calculations using standard doubles (15 digits of accuracy) will fail when it becomes much larger than 10-by-10. However, there are toolboxes that let you do computation with greater precision, and I used one with 512 digits of accuracy to produce the following images of the real and complex entries of the matrices $$ up to size 128-by-128:
The color of the $(i,j)$'th pixel in each plot represents $A_{ij}$, the value of the $(i,j)$'th entry of $$. The first row is the real part of $$, y la segunda fila es la parte imaginaria. El rojo significa gran valor positivo, azul significa el gran valor negativo, y el máximo valor real en la trama se muestra en el medio. La imagen es grande, pero a escala para la visualización en matemáticas.stackexchange - usted puede abrir en una pestaña nueva para ver en más detalle.
Parece que en este Fourier base normalizada en las matrices están convergiendo a un operador integral con un buen kernel,
$$\frac{1}{N}A v \rightarrow \int_0^1 (K(x,y) + iJ(x,y)) v(y) dy,$$
where $K$ and $J$ are the smooth functions in the pictures, and $N$ is some renormalization factor.
Since the kernel is smooth, the action of $$ will annihilate highly oscillatory functions. Recalling the definition $ = \text{diag}(1,-1,\dots) \mathcal{F}^{-1} M \mathcal{F}$, we see exactly how $M$ is ill-conditioned and what functions are in its numerical null space - functions that are the Fourier transform of a highly oscillatory functions.
Turning this around, there should exist an inverse $^{-1}$ for the renormalized limit, acting on a space of functions that are sufficiently smooth. Indeed, the following is a plot of the spectrum of the renormalized $$ en el de 128 por 128 caso (arriba a la derecha), y es dominante vectores singulares (izquierda real en la parte superior imaginario en la parte inferior):
Aquí está el código de Matlab he utilizado:
%Using Advanpix multiprecision computing toolbox, http://www.advanpix.com/
mp.Digits(512+9);
mp.GuardDigits(9);
jjmin = 2;
jjmax = 7;
jjrange = jjmax - jjmin + 1;
for jj=jjmin:jjmax
N = 2^jj;
%Generate original matrix M, where M_nm = (n-1)^(m-1)
v = mp((0:N-1)');
M = mp(zeros(N,N));
for kk=0:(N-1)
M(:,kk+1)=v.^kk;
end
%Generate matrix D F^(-1) M F, where F is the fft, and
%D is the diagonal matrix with diagonal [1, -1, 1, -1, ...]
FMF = mp(zeros(N,N));
for k=1:N
ek = mp(zeros(N,1));
ek(k)=1;
FMF(:,k) = ifft(ifftshift(M*fft(fftshift(ek))));
end
for k=1:N
FMF(k,:) = (-1)^(k-1)*FMF(k,:);
end
%plot it
subplot(2, jjrange,jj - jjmin+1)
imagesc(real(double(FMF)))
title(['N=', num2str(N)]);
subplot(2, jjrange, jjrange + jj -jjmin+1)
imagesc(imag(double(FMF)))
format short
title(['max= ',num2str(max(real(FMF(:))),3)])
end
subplot(2,jjrange,1)
ylabel('real(D F^{-1} M F)')
subplot(2,jjrange,jjrange + 1)
ylabel('imag(D F^{-1} M F)')