recently in my researches I've come across the following operator

$$L\left(\begin{array}{c} a_1\\ \vdots\\ a_n \end{array}\right)=M_1\left(\begin{array}{c} \partial_{z_1}a_1\\ \vdots\\ \partial_{z_1}a_n \end{array}\right)+\dotso+M_n\left(\begin{array}{c} \partial_{z_n}a_1\\ \vdots\\ \partial_{z_n}a_n \end{array}\right),$$ where $a_1,\dotso,a_n:\mathbb R^n\to\mathbb R$ are smooth functions of the independent variables $z_1,\dotso,z_n$ and $M_1,\dotso M_n$ are real square $n\times n$ matrices with constant coefficients (though none of them is invertible).

Is there a general theory to treat this simple system of linear first order pdes? Is it possible to write explicitly a solution if we were to solve

$$L\left(\begin{array}{c} a_1\\ \vdots\\ a_n \end{array}\right)=\left(\begin{array}{c} b_1\\ \vdots\\ b_n \end{array}\right),$$ for some smooth functions $b_1,\dotso,b_n:\mathbb R^n\to\mathbb R$?

Thank you for the references and the patience.

Kind regards,



Following the suggestions, I report here the $3\times 3$ system which is the toy model I'm currently studying

$$\left(\begin{array}{ccc} B_{11} & B_{12} & 0\\ -B_{11} & 0 & 0\\ 0 & -B_{12} & 0 \end{array} \right)\left(\begin{array}{c} \partial_{z_1}a_1\\ \partial_{z_1}a_2\\ \partial_{z_1}a_3\\ \end{array}\right)+ \left(\begin{array}{ccc} 0 & 0 & -B_{23}\\ 0 & B_{22} & B_{23}\\ 0 & -B_{22} & 0 \end{array} \right)\left(\begin{array}{c} \partial_{z_2}a_1\\ \partial_{z_2}a_2\\ \partial_{z_2}a_3\\ \end{array}\right)+ \left(\begin{array}{ccc} 0 & 0 & -B_{33}\\ -B_{31} & 0 & 0\\ B_{31} & 0 & B_{33} \end{array} \right)\left(\begin{array}{c} \partial_{z_3}a_1\\ \partial_{z_3}a_2\\ \partial_{z_3}a_3\\ \end{array}\right)= \left(\begin{array}{c} b_1\\ b_2\\ -b_1-b_2 \end{array}\right), $$

$a_i$ and $b_i$ are smooth functions, the $b_i$'s are compactly supported. I'm interested in solving locally the above system. I think it has something to do with the usual theory for the hyperbolic systems of differential equations, however I'm to unexperienced in the field to be sure about the steps to follow. I understand the condition to be imposed on the functions $b_i$ as a necessary condition for the system to be possibly solvable, but it is not clear to me why there should exist a solution at all! This is the most I can say up to now. Thanks again for the patience and the help

  • $\begingroup$ I apologize. I thought your equation was an evoluation equation. That is, I thought the left-hand side was $L\partial_t a$ for some matrix $L$. Forget it. $\endgroup$ Dec 16, 2014 at 12:24
  • $\begingroup$ Could you say on what domain you want a solution? A bounded domain, all of $\mathbb{R}^n$, in a neighborhood of a point, or something else? $\endgroup$
    – Deane Yang
    Dec 16, 2014 at 14:11
  • 1
    $\begingroup$ You might try the answers to mathoverflow.net/questions/186779/… $\endgroup$
    – Deane Yang
    Dec 16, 2014 at 14:14
  • $\begingroup$ @DeaneYang Let us say tat for my actual purposes this could be included as a part of the answer I am looking for. Do there appear substantial differences between the two cases? $\endgroup$ Dec 16, 2014 at 15:00
  • $\begingroup$ @DeaneYang And also it seems that the reference you gave me treates a casae which is considerable more difficult than mine. I mean I have at disposal an $n$-tuple of scalar functions, not just one, but maybe I am misreading something. Thank you btw $\endgroup$ Dec 16, 2014 at 15:01

2 Answers 2


Since you wrote the system explicitly, here's an explicit solution using Fourier transforms. Let $a(z) = \int e^{ik\cdot z} a(k)\, d^3k $ and $b(z) = \int e^{ik\cdot z} b(k)\, d^3k$. The $a(k)$ and $b(k)$ vectors will satisfy the Fourier transformed equation $iM(k) a(k) = b(k)$, where $M(k) = (M_1 k_1 + M_2 k_2 + M_3 k_3)$. It can be checked explicitly that the following formula gives a complete solution for $a(k)$: $$\begin{gathered} \begin{pmatrix} a_1(k) \\ a_2(k) \\ a_3(k) \end{pmatrix} = \frac{i}{\Delta(k)} \begin{pmatrix} B_{23} k_2 & B_{23} k_2 + B_{33} k_3 \\ 0 & 0 \\ B_{11} k_1 + B_{31} k_3 & B_{11} k_1 \end{pmatrix} \begin{pmatrix} b_1(k) \\ b_2(k) \end{pmatrix} + \begin{pmatrix} \Delta_1(k) \\ \Delta_2(k) \\ \Delta_3(k) \end{pmatrix} c(k), \\ \Delta(k) = B_{11} B_{33} k_1 k_3 + B_{23} B_{31} k_2 k_3 + B_{31} B_{33} k_3^2 , \\ \Delta_1(k) = B_{12} B_{23} k_1 k_2 + B_{22} B_{23} k_2^2 + B_{22} B_{33} k_2 k_3 , \\ \Delta_2(k) = B_{11} B_{33} k_1 k_3 + B_{23} B_{31} k_2 k_3 + B_{31} B_{33} k_3^2 , \\ \Delta_3(k) = B_{11} B_{12} k_1^2 + B_{11} B_{22} k_1 k_2 + B_{12} B_{31} k_1 k_3 , \end{gathered} $$ where $c(k)$ is an arbitrary scalar function.

Of course, because the denominator $\Delta(k)$ vanishes for some $k$, the Fourier coefficients $a(k)$ must be interpreted as distributions.

  • $\begingroup$ I'm glad to see the same procedure I have applied in my calculations. I'm just a little bit confused about the Inverse Fourier transform.. What is that allows us to invert our solutions? $\endgroup$ Jan 2, 2015 at 17:55
  • $\begingroup$ I mean, in the hyperbolic theory there is the hypothesis concerning the reality or the bound on the imaginary part of the eigenvalues so that the fourier anti-transform will be indeed well-defined.. What applies in our case? $\endgroup$ Jan 2, 2015 at 17:56
  • 1
    $\begingroup$ I'm not sure I understand and I'm not sure that your intuition about "imaginary parts of eigenvalues" is applicable here. Are you concerned about the convergence of the integral $a(z) = \int e^{ik\cdot z} a(k)\, d^3k$? If so, then of course the integral must be regularized near the singularities of $1/\Delta(k)$. Are you familiar with distributions and their Fourier transforms? $\endgroup$ Jan 2, 2015 at 18:02
  • $\begingroup$ Yes yes... I was just wondering if we can be so lighthearted in taking the inverse fourier transform.. But now that I realize it.. the hypothesis on the compact support of the $b_i$ should take care of that. In any case if you don't mind just expanding your comment about regularizing near the singularities of $\Delta(k)^{-1}$ then I will be very pleased to consider your advices... $\endgroup$ Jan 2, 2015 at 18:23
  • $\begingroup$ Sorry to bother you again.. I was just wondering what can we say about the regularity of the solutions we have found? what is the best regularity we can expect? $\endgroup$ Jan 5, 2015 at 9:53

Perhaps this is too naive an answer, but since this problem has constant coefficients, it seems perfectly set up for separation of variables. Look for complex solutions of the form $e^{i\vec{\lambda}\cdot\vec{x}} \vec{A}(\vec{\lambda})$. The first equation then takes the form $0=(-L+\sum M_i \lambda_i)\vec{A}$. This has a nonzero solution iff $det(-L+\sum M_i\lambda_i)=0$. The determinant is a polynomial in $\vec{\lambda}$, so there is a well established theory for finding the solution surfaces. (Of course, in high dimensions, well established does not necessarily mean easy.) Once the frequencies $\vec{\lambda}$ have been found, it is an eigenvalue problem to find the associated solutions $\vec{A}$. Since the original matrices were real valued, the frequencies $\vec{\lambda}$ will either be real or come in complex pairs, allowing for the construction of real solutions.

Be aware that series solutions of this type conceal a lot of information. By appropriate choices of $M_i$, each component of $\vec{a}$ can be made to satisfy the Laplace equation or to satisfy the wave equation. These have radically different behaviours. The Laplace equation has no nonzero solutions that vanish at infinity, whereas the wave equation has an infinite dimensional space of such solutions (for $n>1+1$).

As the comment with curl above shows, introducing $\vec{b}=L\vec{a}$ does not seem to help the analysis.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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