Subellipticity and unique continuation · Self-checked by the writing AI

Finite type and the sign of the symbol

A complex symbol can fail ellipticity and still control a fractional derivative. Finite order of vanishing determines the possible exponent. The direction of a sign change decides whether that exponent is available at all. This lesson proves both facts for explicit first-order models and connects the calculation to the Poisson bracket.

The prerequisites are Degenerate energy and sharp gains of regularity, the Fourier transform, and the Cauchy–Schwarz inequality. Basic references are Nicolas Lerner's Semi-classical estimates for non-selfadjoint operators [L], Camille Laurent and Matthieu Léautaud's Unique continuation and applications [LL], and Lars Hörmander's The Analysis of Linear Partial Differential Operators IV [H]. The proof below uses one-dimensional integral kernels rather than the general microlocal localization theory.

1. Fixing the direction before fixing the exponent

Set \(D_t=-i\partial_t\) and \(D_y=-i\partial_y\). For an integer \(k\geq1\), consider \[ P_k^\sigma=D_t+i\sigma t^kD_y,\qquad \sigma\in\{1,-1\}. \] Its symbol is \(p(t,y;\tau,\eta)=\tau+i\sigma t^k\eta\). We restrict attention to the frequency cone \(\eta>0\). We use \[ \{a,b\}=\partial_\tau a\,\partial_t b+ \partial_\eta a\,\partial_y b- \partial_t a\,\partial_\tau b- \partial_y a\,\partial_\eta b, \qquad H_a b=\{a,b\}. \] Hence \(H_{\operatorname{Re}p}=\partial_t\), and at \(t=0\), \[ H_{\operatorname{Re}p}^j\operatorname{Im}p=0\quad(j<k), \qquad H_{\operatorname{Re}p}^k\operatorname{Im}p=\sigma k!\eta. \tag{1.1} \] There are \(k+1\) symbol factors in the last iterated bracket: one imaginary part and \(k\) real parts.

If \(k\) is even, the imaginary part has constant sign on either side of \(t=0\). Either \(\sigma\) is allowed. If \(k\) is odd, \(\sigma=1\) gives a change from negative to positive; \(\sigma=-1\) gives a change from positive to negative. These two orientations have different analytic behavior.

For a precise test-function space, let \(u(t,y)\) be smooth, compactly supported in \(t\), and Schwartz in \(y\), with partial Fourier transform supported in \(\eta\geq1\). Estimates below refer to this space. Band restriction is explicit: a sign condition proved in \(\eta>0\) does not silently become a statement about \(\eta<0\).

2. A one-sided kernel for even order

Lemma 2.1. If \(k\) is a positive even integer and \(h>0\), then for every real \(t\), \[ \int_t^{t+h}r^k\,dr\geq c_kh^{k+1}, \qquad c_k=\frac1{2^k(k+1)}. \tag{2.1} \]

Proof. For fixed \(h\), differentiate this integral with respect to \(t\). Its derivative is \((t+h)^k-t^k\), which vanishes only at \(t=-h/2\), is negative before that point, and positive after it. The minimum is therefore \[ \int_{-h/2}^{h/2}r^k\,dr =\frac{h^{k+1}}{2^k(k+1)}. \] \(\square\)

Lemma 2.2. Let \(k\geq2\) be even, \(\lambda>0\), and \(Q_\lambda=\partial_t-\lambda t^k\). Then \[ \|f\|_2\leq A_k\lambda^{-1/(k+1)}\|Q_\lambda f\|_2, \tag{2.2} \] for \(f\in C_c^\infty(\mathbb R)\), where \(A_k=\int_0^\infty e^{-c_kr^{k+1}}\,dr<\infty\).

Proof. Set \(g=Q_\lambda f\). Multiplication by an integrating factor and the condition that \(f\) vanishes for sufficiently large \(t\) give \[ f(t)=-\int_t^\infty \exp\left(-\lambda\int_t^s r^k\,dr\right)g(s)\,ds. \tag{2.3} \] Let \(K(t,s)\) be the nonnegative kernel on the right, including the indicator of \(s\geq t\). By Lemma 2.1, \[ \sup_t\int K(t,s)\,ds\leq A_k\lambda^{-1/(k+1)}, \qquad \sup_s\int K(t,s)\,dt\leq A_k\lambda^{-1/(k+1)}. \] For completeness, if both suprema are at most \(M\), Cauchy–Schwarz with measure \(K(t,s)ds\) gives \[ \left|\int K(t,s)g(s)ds\right|^2 \leq M\int K(t,s)|g(s)|^2ds. \] Integrating in \(t\) and using the second supremum bounds the squared norm by \(M^2\|g\|_2^2\). This proves (2.2). Finiteness of \(A_k\) follows, for example, by splitting at \(1\) and using \(r^{k+1}\geq r\) for \(r\geq1\). \(\square\)

Replacing \(t\) by \(-t\) turns \(\partial_t+\lambda t^k\), for even \(k\), into the negative of the preceding operator. Thus the same estimate holds for the other constant sign.

3. Odd order with the favorable orientation

Lemma 3.1. If \(k\geq1\) is odd and \(\lambda>0\), then \[ \|Q_\lambda f\|_2^2 =\|f'\|_2^2+\lambda^2\|t^kf\|_2^2 +\lambda k\int t^{k-1}|f|^2. \tag{3.1} \] In particular, with \(\varepsilon=1/(k+1)\), \[ \lambda^\varepsilon\|f\|_2\leq\sqrt{12}\|Q_\lambda f\|_2, \qquad \|f'\|_2\leq\|Q_\lambda f\|_2. \tag{3.2} \]

Proof. Expand the squared norm. The mixed term is \[ -2\lambda\operatorname{Re}\int f'\overline f\,t^k =-\lambda\int (|f|^2)'t^k =\lambda k\int t^{k-1}|f|^2. \] For odd \(k\), \(t^{k-1}\geq0\). Drop this nonnegative term and apply Corollary 2.2 of the preceding lesson with \(\eta=\lambda\). Dropping the potential term as well proves the derivative bound. \(\square\)

For \(k=1\), the identity shows exactly where the bracket contributes: the extra term is \(\lambda\|f\|_2^2\). Reversing the sign reverses this contribution and permits a rapidly decaying kernel.

4. A complete fractional estimate

For even \(k\), the kernel estimate must be supplemented by a derivative estimate.

Lemma 4.1. For even \(k\geq2\), \[ \|f'\|_2\leq B_k\|Q_\lambda f\|_2 \tag{4.1} \] for a constant \(B_k\) independent of \(\lambda>0\).

Proof. First let \(\lambda=1\). The identity (3.1), valid also for even \(k\), gives \[ \|f'\|_2^2+\|t^kf\|_2^2 \leq\|Q_1f\|_2^2+k\int |t|^{k-1}|f|^2. \] Choose \(M_k<\infty\) such that \(k|t|^{k-1}\leq\tfrac12|t|^{2k}+M_k\) everywhere. Then \[ \|f'\|_2^2+\tfrac12\|t^kf\|_2^2 \leq\|Q_1f\|_2^2+M_k\|f\|_2^2 \leq(1+M_kA_k^2)\|Q_1f\|_2^2 \] by Lemma 2.2. The unitary dilation \(t\mapsto\lambda^\varepsilon t\) conjugates \(Q_\lambda\) to \(\lambda^\varepsilon Q_1\); it gives the same factor for the derivative. These factors cancel in the asserted estimate. \(\square\)

Theorem 4.2. If \(k\) is even with either \(\sigma\), or \(k\) is odd with \(\sigma=1\), then on the test-function space of Section 1, \[ \|D_tu\|_2+\big\||D_y|^\varepsilon u\big\|_2 \leq C_k\|P_k^\sigma u\|_2, \tag{4.2} \] and consequently \[ \|u\|_{H^\varepsilon} \leq C_k'\big(\|P_k^\sigma u\|_2+\|u\|_2\big). \tag{4.3} \] The isotropic exponent \(\varepsilon\) is optimal.

Proof. At positive \(y\)-frequency \(\eta\), multiplication of \(P_k^1\) by \(i\) gives \(Q_\eta=\partial_t-\eta t^k\). Apply the preceding lemmas at each frequency, square their estimates, and integrate. Reflection treats the even negative-sign case. This proves (4.2). The inequality \[ (1+\tau^2+\eta^2)^\varepsilon \leq1+|\tau|^{2\varepsilon}+|\eta|^{2\varepsilon}, \qquad |\tau|^{2\varepsilon}\leq1+\tau^2, \] proves (4.3).

To check sharpness, take nonzero \(f\in C_c^\infty(\mathbb R)\), and a nonzero Schwartz function \(g\) whose Fourier transform is smooth and supported in \([-1/2,1/2]\). For \(\lambda\geq2\), put \[ u_\lambda(t,y)=\lambda^{\varepsilon/2}f(\lambda^\varepsilon t) g(y)e^{i\lambda y}. \] It belongs to the stated test-function space and has fixed \(L^2\) norm. Direct differentiation gives \(\|P_k^\sigma u_\lambda\|_2=O(\lambda^\varepsilon)\). Its Fourier support in \(y\) lies in \([\lambda-1/2,\lambda+1/2]\), so \(\|u_\lambda\|_{H^s}\geq c_s\lambda^s\) for \(s>0\). An estimate with \(s>\varepsilon\) is impossible. \(\square\)

The operator has order one. Thus its equation estimate loses \[ \delta=1-\varepsilon=\frac{k}{k+1} \] derivatives relative to ellipticity. In the second-order model of the previous lesson, the energy had the same exponent \(\varepsilon\), but the equation had exponent \(2\varepsilon\). Matching only an exponent, while ignoring the operator order and the norm on the right, would give a false comparison.

5. The forbidden sign creates quasimodes

Theorem 5.1. If \(k\) is odd and \(\sigma=-1\), no estimate \[ \big\||D_y|^s u\big\|_2 \leq C\big(\|P_k^{-1}u\|_2+\|u\|_2\big) \tag{5.1} \] holds for any \(s>0\), even when the \(t\)-support is restricted to a fixed small neighborhood of zero.

Proof. The one-dimensional operator at frequency \(\eta>0\) is, up to multiplication by a complex number of modulus one, \[ \partial_t+\eta t^k. \] It annihilates \[ h_\eta(t)=\exp\left(-\frac{\eta t^{k+1}}{k+1}\right). \] Since \(k+1\) is even, this function decays at both ends. Fix \(0<r<R\), and choose \(\chi\in C_c^\infty((-R,R))\) equal to one on \([-r,r]\). On the support of \(\chi'\), \(|t|\geq r\). Therefore \[ (\partial_t+\eta t^k)(\chi h_\eta)=\chi'h_\eta, \qquad \|\chi'h_\eta\|_2\leq C e^{-c\eta}. \tag{5.2} \] The substitution \(z=\eta^\varepsilon t\) shows \(\|\chi h_\eta\|_2\asymp\eta^{-\varepsilon/2}\) for \(\eta\geq1\). Define \(F_\eta=\chi h_\eta/\|\chi h_\eta\|_2\). Its norm is one, and its equation norm is at most \(C\eta^{\varepsilon/2}e^{-c\eta}\).

Choose \(a\in C_c^\infty((-1/2,1/2))\) with \(\|a\|_2=1\). Define \(u_\lambda\) by its partial Fourier transform \[ \widehat u_\lambda(t,\eta)=a(\eta-\lambda)F_\eta(t),\qquad \lambda\geq2. \] This is smooth, has the required \(t\)-support, and is Schwartz in \(y\). Plancherel gives \(\|u_\lambda\|_2=1\). Uniformly over its frequency window, (5.2) gives \(\|P_k^{-1}u_\lambda\|_2\leq C\lambda^{\varepsilon/2}e^{-c'\lambda}\). But \(\||D_y|^su_\lambda\|_2\geq(\lambda-1/2)^s\). This contradicts (5.1) as \(\lambda\to\infty\). \(\square\)

The nonzero bracket in (1.1) still has order \(k\). Finite type alone therefore does not suffice. For odd order, the orientation of the first nonzero bracket matters.

6. Exercises

Exercise 6.1 — sign and cone, 6 points. For \(P=D_t+it^3D_y\), determine the favorable frequency cone and compute the loss there. Do the same for its complex conjugate differential expression, written using \(D_t,D_y\).

Solution. In \(\eta>0\), the imaginary part \(t^3\eta\) changes from negative to positive. Theorem 4.2 gives gain \(1/4\) and loss \(3/4\). In \(\eta<0\), the direction is reversed and Theorem 5.1 applies after reflection in \(y\). Complex conjugation of the coefficients gives \(D_t-it^3D_y\); this has the favorable cone \(\eta<0\) and the same gain and loss. Here “complex conjugate differential expression” means conjugating the coefficients in the \(D\)-notation; conjugating the action on functions also changes \(D\) to \(-D\), giving an additional overall minus sign, which does not change norm estimates.

Exercise 6.2 — a semiclassical parameter, 8 points. Let \(Q_h=h\partial_t-t^k\), \(0<h\leq1\), with a favorable orientation as above. Deduce the exponent in \(\|Q_hf\|_2\geq c h^\alpha\|f\|_2\), and prove it is optimal.

Solution. Write \(Q_h=hQ_{1/h}\). The preceding coercivity bounds give \(\alpha=1-\varepsilon=k/(k+1)\). For optimality use the normalized dilation \(f_h(t)=h^{-\varepsilon/2}f(t/h^\varepsilon)\). Its norm is fixed. Both terms in \(Q_hf_h\) have size \(h^{1-\varepsilon}=h^{k\varepsilon}\), so a stronger bound with \(\alpha<k/(k+1)\) is impossible as \(h\to0\). A smaller exponent would give a larger lower bound for small \(h\).

Exercise 6.3 — an adjoint kernel, 8 points. For \(k=1\), prove that \(Q_\lambda=\partial_t-\lambda t\) obeys \(\|Q_\lambda f\|_2^2\geq2\lambda\|f\|_2^2\). Explain why a kernel of its adjoint does not contradict this inequality.

Solution. Direct expansion gives \[ \|Q_\lambda f\|_2^2- \|(\partial_t+\lambda t)f\|_2^2=2\lambda\|f\|_2^2. \] Drop the nonnegative second norm. The adjoint is \(Q_\lambda^*=-\partial_t-\lambda t\), which annihilates \(e^{-\lambda t^2/2}\). A lower bound for \(Q_\lambda\) says that its own kernel is zero and its range is closed in the appropriate graph domain. It does not say that its range is all of \(L^2\). The adjoint kernel is orthogonal to that range. The Gaussian may be approximated by compact cutoffs to show the constant \(2\lambda\) is optimal.

Exercise 6.4 — a coefficient that vanishes on an interval, 10 points. Suppose \(a(t)\) is smooth and zero on a nonempty open interval. Show that finite-type gain fails for \(D_t+ia(t)D_y\), even if \(a\) never changes sign.

Solution. Choose \(f\neq0\) compactly supported in that interval, and choose the band-limited \(g\) used in Theorem 4.2. Put \(u_\lambda=f(t)g(y)e^{i\lambda y}\). Its norm is fixed and \(Pu_\lambda=D_tf(t)g(y)e^{i\lambda y}\), whose norm is also fixed. Every positive \(y\)-Sobolev norm grows like \(\lambda^s\). Thus no positive gain is possible. Constant sign controls orientation but does not replace finite order of vanishing.

References

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