In order to find an orthogonal basis of eigenvectors of the Fourier transform operator $F:L_2(\mathbb{R})\to L_2(\mathbb{R})$, $f\mapsto\lim_{N\to\infty}\int_{[-N,N]}f(x)e^{-i\lambda x}d\mu_x$ for Euclidean separable space $L_2(\mathbb{R})$, so that $F$ would be represented by an infinite diagonal matrix, A.N. Kolmogorov and S.V. Fomin, in their Элементы теории функций и функционального анализа (Elements of the theory of functions and functional analysis), uses functions that are, up to a constant factor, the Hermite functions, which I know to constitute an orthogonal basis of $L_2(\mathbb{R})$.

While looking for such an orthogonal basis of eigenvectors, Kolmogorov and Fomin search for Schwartz functions, belonging to $S_\infty\subset L_2(\mathbb{R})$, in the form $w(x)e^{-x^2/2}$ where $w$ is a polynomial, satisfying the equation$$\frac{d^2f}{dx^2}-x^2f=\mu g\quad\quad\text{equation }(3)$$where $\mu$ is a constant. Such an equation change into $\frac{d^2g}{d\lambda^2}-\lambda^2g=\mu f$ when acted upon by operator $F$, $f\mapsto F[f]=g$. It is shown that the polynomials $w$ of such (Hermite up to a constant factor) functions satisfy equation (3) for $\mu=-(2n+1)$ when $\deg w=n$ (let us call such a polynomial $P_n$) and they have non-null coefficients $a_k$ of the variable $x^k$ only for the $k$'s of the same oddity of $n$. It had also previously proved (p. 401) that $P_n$ is, up to a constant, $(-1)^ne^{x^2}\frac{d^n e^{-x^2}}{dx^n}$ (I am writing this in the case it can be used to prove that $P_n(x)e^{-x^2/2}$ defines an eigenvector of $F$).

Kolmogorov-Fomin's says that the fact that the $P_ne^{-x^2/2}$ are eigenvectors of $F$ and their eigenvalues are $\pm\sqrt{2\pi}$, $\pm i\sqrt{2\pi}$ derives from the following fact:

  1. Equation (3) is invariant with respect to transformation $F$.
  2. Equation (3) has got, up to a constant factor, one solution of the form $P_ne^{-x^2/2}$.
  3. The Fourier transform maps $x^ne^{-x^2/2}$ to $i^n\sqrt{2\pi}\frac{d^n}{dx^n}e^{-x^2/2}$ (as I knew, by using the fact that $F[e^{-x^2/2}](\lambda)=\sqrt{2\pi}e^{-\lambda^2/2}$).

The proof of this derivation contained in the book only says that $F^4[P_n e^{-x^2/2}]=4\pi^2 P_n e^{-x^2/2}$, but I cannot see this and, even proved this, I could not see how this is related to the fact that the eigenvalues are the fourth roots of $4\pi^2$ (since I do not think that if the fourth power of matrix $A$ is diagonal $A^4=\text{diag}(4\pi^2,4\pi^2,\ldots)$ then $A=\text{diag}((4\pi^2)^{1/4},(4\pi^2)^{1/4},\ldots)$). I have found some resources talking about the topic on line, but nothing using Kolmogorov-Fomin's argument, which I would like to understand... Does anybody understand and can explain it? I heartily thank you!

P.S.: If it were useful to understand what Kolmogorov-Fomin's says, I know that $\forall f\in S_\infty$, $k\in\mathbb{N}$ $F[f^{(k)}](\lambda)=(i\lambda)^kF[f](\lambda)$ and $\frac{d^k}{d\lambda^k}F[f](\lambda)=(-i)^kF[x^k f](\lambda)$.


The Fourier transform is unitary on $L^{2}(\mathbb{R})$ if you choose the correct normalization: $$ Ff = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}e^{-ist}f(t)\,dt. $$ Some people instead absorb the constant into the exponent, but this is what I'm used to seeing because this $F$ is directly linked to the spectral resolution of the operator $A=\frac{1}{i}\frac{d}{dt}$ on the domain consisting of all $f\in L^{2}$ which are absolutely continuous on $\mathbb{R}$ with $f' \in L^{2}$. This operator $A$ is selfadjoint.

For the $F$ as stated, it is true that $F : L^{2}\rightarrow L^{2}$ is unitary with unitary inverse $$ F^{-1}g = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty}e^{ixs}g(s)\,ds. $$ The fact that $F^{-1}$ is related to $F$ is where you get the general identity that $F^{4}=I$. Indeed, $(F^{2}f)(x) = f(-x)$ is a reflection. That's how you get $F^{4}=I$ on $L^{2}$. That alone is enough to imply that the spectrum of $F$ is $\{ 1,i,-1,-i\}$. To see why, let these roots be identified with $\lambda_{n}=i^{n-1}$ for $n=1,2,3,4$, and form the Lagrange interpolating polynomials $$ p_{k}(\lambda)=\frac{\prod_{n\ne k}(\lambda-\lambda_{n})}{\prod_{n\ne k}(\lambda_{k}-\lambda_{n})},\;\;\; k =1,2,3,4. $$ The polynomial $p_{k}$ is $1$ on $\lambda_{k}$ and $0$ on the other $\lambda_{n}$'s. Hence, $1=p_{1}+p_{2}+p_{3}+p_{4}$ because this is a third order polynomial that is $1$ at each of $4$ distinct points. It follows that $$ I=p_{1}(F)+p_{2}(F)+p_{3}(F)+p_{4}(F). $$ That means that each $P_{k}=p_{k}(F)$ is a projection. To see why, first observe that $p_{j}(F)p_{k}(F)=0$ for $j \ne k$ because $\lambda^{4}-1$ divides $p_{k}p_{j}$ in that case. Applying $p_{k}(F)$ to the above equation gives $$ p_{k}(F) = p_{k}(F)^{2}. $$ Therefore each $P_{k}=p_{k}(F)$ is a projection; these projections are disjoint; and their sum is the identity $I$ on $L^{2}$.

$P_{k}$ is the projection onto the eigenspace of $F$ with eigenvalue $\lambda_{k}$ because $(F-\lambda_{k}I)P_{k}=0$. This is a spectral decomposition for $F$: $$ P_{k}^{2}=P_{k},\;\;\; P_{k}P_{j}=0,\;\; j \ne k,\\ I = P_1+P_2+P_3+P_4,\\ F = \lambda_1 P_1 + \lambda_2 P_2 + \lambda_3 P_3 + \lambda_4 P_4. $$ Notice that these arguments don't require inner products or norms. These arguments work for any linear operator $F$ on a complex linear space for which $F^{4}-I=0$. The same thing works for the unnormalized $F$ that you have, too.

Hermite Functions: The Hermite polynomials, as you have defined them, are $$ P_{n}(x) = e^{x^{2}}\frac{d^{n}}{dx^{n}}e^{-x^{2}}. $$ The Hermite functions $H_{n}(x)=e^{-x^{2}/2}P_{n}$ are solutions of the ODE $$ \left[-\frac{d^{2}}{dx^{2}}+x^{2}\right]H_{n}=\lambda_{n}H_{n}. $$ These are the unique $L^{2}(\mathbb{R})$ eigenfunctions of the differential equation and $\{ H_{n} \}_{n=0}^{\infty}$ forms a complete orthonormal basis of $L^{2}(\mathbb{R})$. Knowing this, when you Fourier transform the Hermite equation, you get $$ s^{2}\widehat{H_{n}}(s)-\frac{d^{2}}{ds^{2}}\widehat{H_{n}}(s)=\lambda_{n}\widehat{H_{n}}(s). $$ Because of the uniqueness of such solutions, it follows that $\widehat{H_{n}}$ must be a constant multiple of $H_{n}$. That constant multiple is an eigenvalue of the Fourier transform, which must be either $1$, $i$, $-1$ or $-i$. (Remember, I use the normalized Fourier transform with $1/\sqrt{2\pi}$ in front of it so that $\|Ff\|=\|f\|$.)

  • 1
    $\begingroup$ A projection is a linear operator satisfying $P^{2}=P$. Notice that $P(P-I)=0$. So the spectrum is contained in $\{ 0, 1\}$. Using the same construction the linear space decomposes into a sum of two subspaces $X=PX\oplus (I-P)X$. On the first subspace, $P=I$ and on the second $P=0$. If you have a matrix with $P^{2}=P$, then there is a basis for $P$ where the matrix is diagonal with $1$'s and $0$'s along the diagonal. $\endgroup$ – Disintegrating By Parts Nov 27 '14 at 9:36
  • 1
    $\begingroup$ If $P(P-1)=0$, or equivalently, $P^{2}=P$, then every $x=(I-P)x+Px$, and $P(Px)=(Px)$, $P((I-P)x)=0$. So every $x$ decomposes into components on which $P$ is either $0$ or $I$. For $F^{4}=I$, I showed you that $x=x_{1}+x_{i}+x_{-1}+x_{-i}$ where $Fx_{1}=x_{1}$, $Fx_{i}=ix_{i}$, $Fx_{-1}=-x_{-1}$ and $Fx_{-i}=-ix_{-i}$. This decomposition is unique. $\endgroup$ – Disintegrating By Parts Nov 27 '14 at 10:52
  • 1
    $\begingroup$ @DavideZena : Notice that these ideas extend to any linear operator $L$ for which $m(L)=0$ for some polynomial $m$ with distinct linear factors. It is this algebraic trick that forces a diagonalization of $L$, regardless of dimension. In finite dimensional spaces, $L$ is similar to a diagonal matrix iff the minimal polynomial has no repeated linear factors. The algebra of this is very strong because it leads to projections through the interpolating polynomials $p_k$, which decomposes the space into a direct sum of eigenspaces. The $P_{k}$ project to the component associated with $\lambda_k$. $\endgroup$ – Disintegrating By Parts Nov 27 '14 at 11:06
  • 1
    $\begingroup$ Because $(\lambda - \lambda_{k})p_{k}(\lambda)$ is a multiple of $\lambda^{4}-1$, then $(F-\lambda_{k}I)p_{k}(F)=0$. That's how the $p_{k}$ were designed: they're missing just one factor so that when you apply that last operator $(F-\lambda_{k}I)$ to $p_{k}(F)$, you get $0$, which means the range of $p_{k}(F)$ is contained in the null space of $F-\lambda_{k}I$. $\endgroup$ – Disintegrating By Parts Nov 27 '14 at 12:07
  • 1
    $\begingroup$ @DavideZena : I've added more about the Hermite functions. $\endgroup$ – Disintegrating By Parts Nov 27 '14 at 14:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.