I have found the following claim made very clearly at least once in the published literature (see below):

Let $P$ be a linear partial differential operator defined on an open set $\Omega \subset \mathbb{R}^{n+1}$, strictly hyperbolic with respect to the level surfaces of the first coordinate, which I will denote by $t$. Let $u$ be a distribution in $H^s_\mathrm{loc}(\Omega)$ for some $s \in \mathbb{R}$, such that $Pu=0$. Then the restriction (i.e. pullback) of $u$ to each of the level sets $\Sigma_{\tau} := \Omega \cap \{ t = \tau \}$ is itself in $H^s_\mathrm{loc}(\Sigma_\tau)$.

In fact, a more general statement may be found in Section 2 of:

Bao, Gang; Symes, William W., A trace theorem for solutions of linear partial differential equations, Math. Methods Appl. Sci. 14, No.8, 553-562 (1991). ZBL0754.35023.

There, the authors say that the statement boxed above follows from standard energy estimates, and refer to e.g. Taylor's book on pseudo-differential operators for the arguments. However, I meet the following difficulties when attempting to fill in all details:

  • Ostensibly, the "standard energy estimates" discussed in Taylor's book (and elsewhere, e.g. in Hörmander's) pertain distributional solutions which (in the case $\Omega = \mathbb{R}^{n+1}$) are in spaces $\bigcap_{j=0}^{m-1} C^j([\tau_1,\tau_2]; H^{s-j}(\mathbb{R}^n))$, where $m$ is the order of $P$. I, on the other hand, am interested in distributions (locally) in $H^s$ in both time and space.
  • In any case, as far as I am aware, the usual energy estimate controls the evolution in time of the sum of (appropriate) "spatial" Sobolev norms of (appropriate) $t$-derivatives of $u$ restricted to the level sets of $t$. Specifically, in the case $Pu=0$ the energy estimate says that the supremum of this quantity over a compact interval $[\tau_0,\tau_1]$ is bounded by a constant times its value at $\tau_0$. To prove that the restriction of $u$ to the $t=\tau_0$ surface is (locally) in $H^s$, presumably we would like to control the $H^s$ norm of this restriction and place an upper bound on it, but this does not seem to be the scenario in the energy inequality.

By the way, I am not doubting that $u$ can be pulled back to a spacelike hypersurface, as a distribution. I know that this is the case because $WF(u)$ is disjoint from the normal bundle of the hypersurface whenever $Pu$ is smooth.


I might have the beginning of an argument, but this leads to a further question about Bochner–Sobolev spaces whose answer seems clear in the case of positive Sobolev order $s$, but not so much otherwise. I will explain schematically and in particular assuming for simplicity and definiteness that $P$ is of second order and that $\Omega = \mathbb{R}^{n+1}$ (so that also $\Sigma_\tau \cong \mathbb{R}^n$).

Namely, let $u \in H^s_\mathrm{loc}(\mathbb{R}^{n+1})$ be in the kernel of $P$, and let $\phi \in C_\mathrm{c}^\infty(\mathbb{R}^{n+1})$. Let $\psi \in C_\mathrm{c}^\infty(\mathbb{R}^{n+1})$ be such that $\psi = 1$ on a neighbourhood of $\mathrm{supp}{\phi}$. Then $P(\phi u) = [P,\phi] u = [P,\phi] (\psi u)$, and the commutator is a first-order differential operator. Hence, $\phi u$ solves the PDE with non-zero right-hand side but vanishing Cauchy data on $\Sigma_{T_0}$ for any sufficiently small $T_0$. Now, if $u$ were smooth then the standard energy estimates would say that for any $T_1 > T_0$ there exists a constant $C>0$ such that $$ (*) \quad \mathcal{E}_s(\tau; \phi u) \leq C \int_{T_0}^{T_1}\| [P,\phi] (\psi u) (t,\cdot) \|^2_{H^{s-1}(\mathbb{R}^n)} \, \mathrm{d}t $$ where the energy of order $s \in \mathbb{R}$ and at time $\tau$ is defined as $$\mathcal{E}_s(\tau; \phi u) = \| (\phi u)|_{\Sigma_\tau} \|^2_{H^s(\mathbb{R}^n)} + \| \partial_t (\phi u)|_{\Sigma_\tau} \|^2_{H^{s-1}(\mathbb{R}^n)}.$$ If we could provide a lower bound for the right-hand side of $(*)$, in terms of the norm of $\psi u$ in $H^s(\mathbb{R}^{n+1})$, then presumably we would be close to being done (by a suitable density argument from $C^\infty$). Now I would like to say something along these lines: $$ (**) \quad \int_{T_0}^{T_1}\| [P,\phi] (\psi u) (t,\cdot) \|^2_{H^{s-1}(\mathbb{R}^n)} \, \mathrm{d}t \leq C_1 \| [P,\phi] (\psi u) \|^2_{H^{s-1}(\mathbb{R}^{n+1})} \leq C_2 \| \psi u \|^2_{H^{s}(\mathbb{R}^{n+1})}$$ where the second inequality would be due to $[P,\phi]$ being differential of order $1$. But how can I justify the first inequality? It seems almost obvious in the case of integer $s \geq 1$ – we are simply ignoring those multi-indices with a non-zero entry corresponding to $t$, which would otherwise contribute to the $H^{s-1}(\mathbb{R}^{n+1})$ norm. For general $s \in \mathbb{R}$, I'm not sure. Notice that the left-hand side of $(**)$ is the square of the norm in the Bochner–Sobolev space $L^2([T_0,T_1]; H^{s-1}(\mathbb{R}^{n}))$.

Update 2

Perhaps proving the first inequality in $(**)$ is not so difficult when the Sobolev order $s$ is real and $s \geq 1$. For then it seems we can use the Fourier transform, as follows. Set $\ell := s-1$ and $g := [P,\phi](\psi u) \in C^\infty_\mathrm{c}(\mathbb{R}^{n+1})$, and denote partial Fourier transforms with respect to $t$ and the $x$-variables in $\mathbb{R}^n$ by $\mathcal{F}_t$ and $\mathcal{F}_x$ respectively, while the full Fourier transform of $g$ will be $\hat{g}$. Then, using that $\mathcal{F}_t \mathcal{F}_x g = \hat{g}$, and the Plancherel theorem: \begin{align*} \int_{T_0}^{T_1} \| g(t,\cdot) \|^2_{H^{\ell}(\mathbb{R}^n)} \, \mathrm{d}t &\leq \int_{\mathbb{R}} \| g(t,\cdot) \|^2_{H^{\ell}(\mathbb{R}^n)} \, \mathrm{d}t \\ &= \int_{\mathbb{R}} \left( \int_{\mathbb{R}^n} (1 + |k|^2)^\ell |[\mathcal{F}_x g](t,k)|^2 \, \mathrm{d}^n k\right) \mathrm{d}t \\ &= \int_{\mathbb{R}} \left( \int_{\mathbb{R}^n} (1 + |k|^2)^\ell |[\mathcal{F}_t^{-1} \hat{g}](t,k)|^2 \, \mathrm{d}^n k\right) \mathrm{d}t \\ &= \int_{\mathbb{R}^n} (1 + |k|^2)^\ell \left( \int_{\mathbb{R}} |[\mathcal{F}_t^{-1} \hat{g}](t,k)|^2 \, \mathrm{d}t \right) \mathrm{d}^n k \\ &= \int_{\mathbb{R}^n} (1 + |k|^2)^\ell \left( \int_{\mathbb{R}} |\hat{g}(\tau,k)|^2 \, \mathrm{d}\tau \right) \mathrm{d}^n k \\ &\leq \int_{\mathbb{R}^{n+1}} (1 + |\tau|^2 + |k|^2)^\ell|\hat{g}(\tau,k)|^2 \, \mathrm{d}\tau \, \mathrm{d}^n k \\ &= \| g \|^2_{H^{\ell}(\mathbb{R}^{n+1})}. \end{align*} Of course, I used the fact that $\ell \geq 0$ (i.e. $s \geq 1$) in the final inequality above.


Your Answer

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

Browse other questions tagged or ask your own question.