Let $\Phi: \mathbb{R}^n \to \mathbb{R}^n$ be a diffeomorhism. Let $\Delta$ be the componentwise Laplacian. Is it possible to wirite $$ \Delta (\Phi^{-1})\circ \Phi $$ in such a form that involves only $\Phi$ and not $\Phi^{-1}$?

To clarify my question: For the Jacobian matrix (which I will denote by $D$) it is clearly possible. Indeed $$ D(\Phi^{-1})\circ \Phi = (D\Phi)^{-1} $$


Yes, but as usual, working with higher order derivatives of inverse map is unpleasant. I use the notation

  • $\Psi = \Phi^{-1}$
  • $H\Phi$ is the Hessian matrix of $\Phi$, which is an $n$ by $n$ matrix whose entries are vectors (so, the entry $(i, j)$ is the derivative $\frac{\partial^2 \Phi}{\partial x_i\partial x_j}$).
  • $A*H\Phi$ means applying matrix $A$ to each entry of $H\Phi$ (similar to multiplying an ordinary matrix by a scalar).
  • $\operatorname{tr}$ is the trace of a matrix, for example $\Delta \Phi = \operatorname{tr} H\Phi$.

With this notation, $$ H \Psi = -((D\Phi)^{-1})^T ((D\Phi)^{-1}*H\Phi) (D\Phi)^{-1} \tag1 $$ hence $$ \Delta \Psi = -\operatorname{tr}(((D\Phi)^{-1})^T ((D\Phi)^{-1}*H\Phi) (D\Phi)^{-1}) \tag2 $$ which generalizes the one-dimensional formula for the second derivative of inverse function. Here and below the arguments are suppressed; it is understood that $\Phi$ and its derivatives are evaluated at some point $x$, while $\Psi$ and its derivatives are evaluated at $\Phi(x)$.

Proof: let $\phi^i$, $\psi^i$ denote the components of $\Phi$ and $\Psi$, while subscripts are used for derivatives. The relation $D\Psi = (D\Phi)^{-1}$ can be written as $$ \sum_k \psi^i_k \phi^k_j = \delta^{ij} \quad \text{(Kronecker delta)} $$ Differentiate both sides with respect to $x_\ell$ using the product and chain rules: $$ \sum_{k, m} \psi^i_{km} \phi^k_j \phi^m_\ell + \sum_k \psi^i_k \phi^k_{j\ell} = 0 $$ The first sum on the left is $(D\Phi)^TH\Psi D\Phi$ while the second is $D\Psi*H\Phi$. The claim (1) follows.

  • $\begingroup$ Looking again at your answer after some time, I don't understand something. I understood the meaning of $*$, but which is the definition of the product of a matrix with real entries and a matrix with vector valued entries? (the product between the first two terms in (1) for example) $\endgroup$ – foo90 Jun 7 '18 at 9:50
  • 1
    $\begingroup$ It's like ordinary matrix multiplication: for example, $\begin{pmatrix}a&b\\c&d\end{pmatrix}\begin{pmatrix}A&B\\C&D\end{pmatrix}$ is $\begin{pmatrix}aA+bC&aB+bD\\cA+dC&cB+dD\end{pmatrix}$ where $aA$ means multiplying vector $A$ by scalar $a$. $\endgroup$ – user357151 Jun 7 '18 at 10:10

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.