Transport, exponential weights, and local uniqueness

An exponential weight can separate a region where a solution might survive from the region where a cutoff creates an error. This lesson proves that mechanism for a real transport operator crossing an initial surface. It includes the passage from smooth test functions to distributional solutions and a precise cutoff argument.

The prerequisites are integration by parts, local mollification in $L^2$, and the distinction between an estimate and its domain. Finite type and the sign of the symbol explains why orientation matters in more degenerate first-order models. Basic references are Daniel Tataru's Carleman estimates, unique continuation and applications [T], Nicolas Lerner's Carleman Inequalities [L], and Laurent and Léautaud's Unique continuation and applications [LL].

1. A transverse real direction

Let $U\subset\mathbb R^{1+d}$ be open, with coordinates $(t,z)$, and let $V=\partial_t+\sum_{j=1}^d b_j(t,z)\partial_{z_j}, \qquad P=V+c.$ The $b_j$ are real and $C^1$; $c$ is complex and locally bounded. Set $\operatorname{div}V=\sum_j\partial_{z_j}b_j$. For a real $C^1$ function $\phi$, $e^{\tau\phi}Pe^{-\tau\phi}=P-\tau V\phi. \tag{1.1}$ No symbolic expansion is needed. This is the product rule.

The surface $t=0$ is noncharacteristic because $Vt=1$. A weight will be useful when its derivative along $V$ is positive. That condition is stronger than simply requiring $d\phi\neq0$: the derivative must have the correct direction relative to the operator.

2. The weighted estimate

Theorem 2.1. Let $K\Subset U$, and suppose $V\phi\geq a>0\quad\text{on a neighborhood of }K, \qquad \operatorname{Re}c-\tfrac12\operatorname{div}V\leq M \quad\text{almost everywhere there}.$ For $\tau\geq0$ with $\tau a>M$ and $w\in C_c^\infty(K^\circ)$, $(\tau a-M)\|e^{\tau\phi}w\|_2 \leq\|e^{\tau\phi}Pw\|_2. \tag{2.1}$ The same estimate holds for compactly supported $w\in L^2(U)$ with $Pw\in L^2(U)$, provided its support lies in the region where these bounds hold.

Proof for smooth functions. Put $v=e^{\tau\phi}w$. Integration by parts gives $\operatorname{Re}(Vv,v) =-\tfrac12\int (\operatorname{div}V)|v|^2.$ Consequently $\operatorname{Re}((P-\tau V\phi)v,v) =\int\left(\operatorname{Re}c-\tfrac12\operatorname{div}V -\tau V\phi\right)|v|^2 \leq-(\tau a-M)\|v\|_2^2.$ Cauchy–Schwarz and (1.1) prove (2.1). A $C^1$ weight suffices: $v$ is $C^1$ and compactly supported, so this integration by parts is legitimate. The assertion for the larger domain follows from Lemma 3.1 below. $\square$

The square of (2.1) has a factor of order $\tau^2$. Curvature is not needed to obtain this first-order estimate. We will use curvature to keep the cutoff error away from the largest weight values.

3. Keeping the distributional domain

The following approximation result supplies the domain passage used above.

Lemma 3.1. Suppose $V$ has real $C^1$ coefficients and $c\in L^\infty_{\mathrm{loc}}$. If $w\in L^2(U)$ has compact support in $U$ and $Pw\in L^2(U)$ distributionally, then there are smooth compactly supported $w_\delta$ in a common compact subset of $U$ with $w_\delta\longrightarrow w, \qquad Pw_\delta\longrightarrow Pw \quad\text{in }L^2(U). \tag{3.1}$

Proof. On a neighborhood of the support, extend the coefficients to all of Euclidean space with bounded first derivatives for $V$, and with $c$ bounded. This can be done by multiplying by a smooth cutoff after extending over a slightly larger compact neighborhood. Extend $w$ by zero. Since its support is separated from $\partial U$, this extension creates no boundary term near its support.

Take a smooth compactly supported approximate identity $\rho_\delta$, and set $J_\delta w=\rho_\delta*w$. We prove $VJ_\delta w-J_\delta Vw\longrightarrow0\quad\text{in }L^2. \tag{3.2}$ Write the full vector of coefficients of $V$ as $a(x)$. For smooth $w$, integration by parts in the convolution gives $(VJ_\delta-J_\delta V)w(x) =\int (a(x)-a(y))\cdot\nabla\rho_\delta(x-y)w(y)\,dy +\int\rho_\delta(x-y)(\operatorname{div}a)(y)w(y)\,dy. \tag{3.3}$ The first kernel is bounded in absolute value by $\|Da\|_\infty|x-y||\nabla\rho_\delta(x-y)|$, whose integral is bounded independently of $\delta$. The second term is convolution of $(\operatorname{div}a)w$ with $\rho_\delta$. Young's inequality therefore bounds the operator in (3.3) on $L^2$, uniformly in $\delta$.

For a smooth compactly supported $w$, both $VJ_\delta w$ and $J_\delta Vw$ converge to $Vw$ in $L^2$. The uniform operator bound and density of such functions in $L^2$ establish (3.2) for every $w\in L^2$; formula (3.3) also defines the commutator distributionally.

For the zeroth-order term, $cJ_\delta w-J_\delta(cw) =c(J_\delta w-w)+(cw-J_\delta(cw))\longrightarrow0$ in $L^2$, using only boundedness of $c$. Thus $PJ_\delta w=J_\delta Pw+(PJ_\delta-J_\delta P)w\longrightarrow Pw.$ For small $\delta$, the convolution supports remain in a common compact subset of $U$. Take $w_\delta=J_\delta w$. $\square$

For each fixed $\tau$, the functions $e^{\tau\phi}$ are bounded on that common compact set. Applying (2.1) to $w_\delta$ and passing to the limit proves Theorem 2.1 on the claimed graph domain. No uniform approximation in $\tau$ is asserted or needed: first establish the estimate for each fixed parameter, then let the parameter grow in the uniqueness argument.

4. A curved weight creates a strict separation

Theorem 4.1. Let $P$ satisfy the assumptions of Section 1 near $(0,0)$. If $u\in L^2_{\mathrm{loc}}$ satisfies $Pu=0$ distributionally and vanishes almost everywhere in the positive side $t>0$, then $u$ vanishes in a full neighborhood of $(0,0)$.

Proof. Choose a closed cylinder $K_0=[-T,T]\times\overline{B(0,r)}\Subset U$, with $T,r>0$, small enough that all coefficient bounds are finite. Fix $\mu>0$, and shrink $r$ further so that $\phi(t,z)=t-\mu|z|^2, \qquad V\phi=1-2\mu b(t,z)\cdot z\geq\tfrac12$ on $K_0$. This is possible because $b$ is bounded on an initial fixed cylinder.

Choose real smooth cutoffs $\chi(t)$ and $\psi(z)$, compactly supported in the interior of this cylinder, equal to one on $|t|\leq T/2$ and $|z|\leq r/2$, respectively. Set $w=\chi\psi u$. Distributional product rules and $Pu=0$ give $Pw=V(\chi\psi)u. \tag{4.1}$ Both sides are $L^2$, so Theorem 2.1 applies.

On the portion of the error support where $u$ can be nonzero, $t\leq0$. If $\chi'\neq0$, then either $t<-T/2$ or $t>T/2$; the second portion has $u=0$. On the first portion $\phi\leq-T/2$. If $\nabla\psi\neq0$, then $|z|>r/2$, so $\phi\leq-\mu r^2/4$. Hence, with $d_0=\min(T/2,\mu r^2/4)>0,$ the nonzero error in (4.1) lies entirely in $\phi\leq-d_0$. Its weighted norm is at most $C e^{-\tau d_0}\|u\|_{L^2(K_0)}$.

Let $N=\{|t|<T/2,\ |z|<r/2,\ \phi(t,z)>-d_0/2\}.$ This is an open neighborhood of $(0,0)$, and both cutoffs are one there. Since $e^{\tau\phi}\geq e^{-\tau d_0/2}$ on $N$, the weighted estimate gives $(\tau/2-M)\|u\|_{L^2(N)} \leq C e^{-\tau d_0/2}\|u\|_{L^2(K_0)}.$ Let $\tau\to\infty$. It follows that $\|u\|_{L^2(N)}=0$. $\square$

For example, take $V=\partial_t+(t+z)\partial_z$. On $|t|,|z|<1/4$, the weight $\phi=t-z^2$ satisfies $V\phi=1-2z(t+z)\geq3/4$. The function $I(t,z)=e^{-t}(z+t+1)$ satisfies $VI=0$, so solutions of $Vu=0$ can be written locally as $F(I)$ for smooth $F$. The theorem explains why such a solution cannot be zero on the entire positive side near the origin while remaining nonzero immediately across it.

5. Why the surface hypothesis cannot disappear

For $V=\partial_z$, the surface $t=0$ is characteristic: $Vt=0$. Define $h(t)= \begin{cases} e^{-1/t^2},&t<0,\\ 0,&t\geq0. \end{cases}$ This is smooth, zero on the positive side, and nonzero arbitrarily close to the negative side. The function $u(t,z)=h(t)$ satisfies $Vu=0$. Thus noncharacteristic crossing has real mathematical content.

The proof also distinguishes one-sided local uniqueness from vanishing to infinite order at a point. The former supplies an open side on which the function is zero. A flat jet at one point supplies no such region. The singular weights needed for the latter problem are developed in Detecting a solution from infinite-order silence at one point.

6. A factorized higher-order class

Proposition 6.1. Let $P_j=\partial_t+b_j(t,z)\cdot\partial_z+c_j(t,z)$, $1\leq j\leq m$, have smooth coefficients, with $b_j$ real. If $u$ is smooth, zero for $t>0$, and $P_1P_2\cdots P_mu=0$ near $(0,0)$, then $u=0$ in a neighborhood of that point.

Proof. The function $w=P_2\cdots P_mu$ is smooth and zero on the positive side. It satisfies $P_1w=0$, so Theorem 4.1 makes it zero on a smaller neighborhood. There, apply the same theorem to $P_3\cdots P_mu$ with operator $P_2$. Continue through all $m$ factors, shrinking neighborhoods finitely many times. The last step proves $u=0$. $\square$

This proposition assumes the displayed differential factorization. A polynomial factorization of a principal symbol alone is not such a factorization of the operator. Terms created by composition, pseudodifferential factors, complex roots, and repeated roots require further estimates. The proposition therefore supplies a concrete real factorized class without asserting the full arbitrary-order Cauchy uniqueness theorem.

7. Exercises

Exercise 7.1 — the weight direction, 6 points. For $V=\partial_t+2\partial_z$, decide whether $\phi=t-z$ and $\widetilde\phi=t+z$ satisfy the positive derivative condition. Explain why nonzero gradients do not answer the question.

Solution. $V\phi=1-2=-1$, while $V\widetilde\phi=1+2=3$. Only the latter meets Theorem 2.1 with the displayed orientation. Both gradients are nonzero. The relevant pairing is with the vector field, so the sign of that pairing must also be checked. Negating the first weight produces a positive derivative but reverses which weight levels are large; the cutoff geometry must then be checked anew.

Exercise 7.2 — a distributional solution, 8 points. Suppose $u\in L^2_{\mathrm{loc}}$, $Vu=-cu$ distributionally, and $c$ is locally bounded. Verify all domain conditions needed to apply Theorem 2.1 to $w=\chi\psi u$.

Solution. Compact support of the smooth cutoffs makes $w\in L^2$ with support separated from the boundary. The distributional product rule gives $Pw=V(\chi\psi)u$, which belongs to $L^2$ because $V(\chi\psi)$ is bounded on a compact set. Thus $w$ lies in the graph domain of Lemma 3.1. The coefficient $c$ needs no derivative, since its commutator with mollification converges through the two strong $L^2$ approximation terms displayed there.

Exercise 7.3 — a measurable perturbation, 8 points. Let $Pu=\partial_tu+q(t,z)u$, with $q\in L^\infty_{\mathrm{loc}}$. Prove local uniqueness across $t=0$, even if $q$ jumps across a measurable set.

Solution. This is Theorem 4.1 with $b=0$ and $c=q$. In particular $V\phi=1$ for $\phi=t-\mu|z|^2$. The weighted estimate uses only an upper bound for $\operatorname{Re}q$, and the graph-domain approximation uses only boundedness. A jump in $q$ does not create a derivative of $q$ in this operator. Therefore the proof applies unchanged.

Exercise 7.4 — factor order, 10 points. Explain why Proposition 6.1 does not need its factors to commute. Apply it to $(\partial_t+z\partial_z)(\partial_t+\partial_z)u=0.$ Compute the commutator of the two factors.

Solution. Set $w=(\partial_t+\partial_z)u$. The first factor annihilates $w$, which is zero on the positive side, so it is zero near the origin. The second factor then annihilates $u$, and the same argument makes $u$ zero near the origin. No interchange of factors occurs. Direct calculation gives $[\partial_t+z\partial_z,\partial_t+\partial_z]=-\partial_z$, so these factors do not commute. Changing their order would change the displayed equation by a first-order term.

References

[T] Daniel Tataru, Carleman estimates, unique continuation and applications, 1998 lecture notes, author's version.
[L] Nicolas Lerner, Carleman Inequalities, author's lecture notes.
[LL] Camille Laurent and Matthieu Léautaud, Unique continuation and applications, lecture notes.

Written by GPT-6.1 Sol (OpenAI), at Ultra reasoning effort, September 2026. Self-checked by the writing AI. Public domain (CC0).