Counting zeros of holomorphic functions of exponential growth (Q601089)

Starting with the works of M.~Hager, there has been a number of results that establish Weyl asymptotics of the eigenvalues of non-self-adjoint (pseudo)differential operators with small random perturbations, in the semi-classical limit and in the limit of large eigenvalues. A common feature here (as well as in many other non-self-adjoint spectral problems) is that one identifies the eigenvalues with the zeros of a holomorphic function \(u(z) =u(z; h)\), \(0 < h \ll 1\), in a set \(\Gamma \Subset\mathbb C\). The available information is an upper bound \(|u(z; h)| \leq \exp(\phi(z)/h)\) for \(z\) near the boundary \(\gamma=\partial\Gamma\) as well as lower bounds \(|u(z_j; h)| \geq \exp(\phi(z_j) -\epsilon_j)/h\), for finitely many points \(z_ j =z_j(h)\), \(1\leq j\leq N(h)\), that are suitably distributed near the boundary. Hager obtained a result of this type with the conclusion that the number of zeros in \(\Gamma\) is \[ (2\pi h)^{-1}\left(\int_\Lambda\Delta\phi(z)L(dz) + o(1)\right) \] in the limit of small \(h\), when \(\Gamma\) is independent of \(h\) with smooth boundary and \(\phi\) is a \(C^2\) function, also independent of \(h\). In the paper [ \textit{M. Hager} and \textit{J. Sjöstrand}, Math. Ann. 342, No. 1, 177--243 (2008; Zbl 1151.35063)], this result is generalized by weakening the regularity assumptions on \(\phi\). However, due to some logarithmic losses, the authors were not quite able to recover Hager's original result, and we still have a fixed domain \(\Gamma\) with smooth boundary. In many spectral problems of the above type, the domain should be allowed to depend on \(h\), for instance, it could be a long thin rectangle, and the boundary regularity should be relaxed. In the present paper, the author revisits systematically the proof of the counting proposition in the above work and obtains a general and quite natural result allowing an \(h\)-dependent exponent \(\phi\) to be merely continuous and the \(h\)-dependent domain \(\Gamma\) to have Lipschitz boundary. Let \(r :\gamma \to]0,\infty[\) be a Lipschitz function of Lipschitz modulus \(\leq 1/2\). Assume that \(\gamma\) is Lipschitz in the following precise sense, where \(r\) enters: There exists a constant \(C_0\) such that for every \(x\in\gamma\) there exist new affine coordinates \(\widetilde y = ({\widetilde y}_1, {\widetilde y}_2)\) of the form \({\widetilde y} =U(y -x)\), \(y\in\mathbb C\simeq\mathbb R^2\) being the old coordinates, where \(U = U_x\) is orthogonal, such that the intersection of \(\Gamma\) and the rectangle \[ R_x := \big\{y\in\mathbb C;\, |{\widetilde y}_1| < r(x),\, |{\widetilde y}_2| < C_0r(x)\big\} \] takes the form \[ \big\{y\in R_x;\, {\widetilde y}_2 > f_x({\widetilde y}_1),\, |{\widetilde y}_1| < r(x)\big\}, \] where \(f_x({\widetilde y}_1)\) is Lipschitz on \([-r(x), r(x)]\), with Lipschitz modulus \(\leq C_0\). Let \(D(z, r)\) denote the open disc of center \(z\in\mathbb C\) and radius \(r\). The complete version of the result is: Let \(\Gamma\Subset\mathbb C\) be simply connected and have Lipschitz boundary \(\gamma\) with an associated Lipschitz weight \(r\). Put \[ {\widetilde\gamma}_{\alpha r} =\bigcup_{x\in\gamma}D\big(x,\alpha r(x)\big) \] for any constant \(\alpha > 0\). Let \(z^0_j\in\gamma\), \(j\in\mathbb Z/N\mathbb Z\), be distributed along the boundary in the positively oriented sense such that \(r(z^0_j)/4\leq|z^0_{j+1}-z^0_j|\leq r(z^0_j)/2\). Then there exists a constant \(C_1 > 0\) depending only on the constant \(C_0\) such that if \(z_j\in D\big(z^0_j, r(z^0_j)/(2C_1)\big)\), then we have the following: \ Let \(0 < h \leq 1\) and let \(\phi\) be a continuous subharmonic function on \({\widetilde\gamma}_r\) with a distribution extension to \(\Gamma\cup {\widetilde\gamma}_r\). Then there exists a constant \(C_2 > 0\) such that if \(u\) is a holomorphic function on \(\Gamma\cup{\widetilde\gamma}_r\) satisfying \[ h\ln|u|\leq\phi(z) \text{ on }{\widetilde\gamma}_r,\tag{1} \] \[ h\ln|u(z_j)|\geq\phi(z_j)-\epsilon_j\quad \text{for}\quad j=1,2,\dots,N,\tag{2} \] where \(\epsilon_j \geq 0\), then the number of zeros of \(u\) in \(\Gamma\) satisfies \[ \left|\#\big(u^{-1}(0)\cap\Gamma\big)-\frac{1}{2\pi h}\mu(\Gamma)\right| \] \[ \leq\frac{C_2}{h}\left(\mu({\widetilde \gamma}_r)+\sum_1^N\left(\epsilon_j+\int_{D\left(z_j,\frac{r(z_j)}{4C_1}\right)}\left|\ln\frac{|w-z_j|}{r(z_j)}\right|\mu(dw)\right)\right). \] Here \(\mu := \Delta\phi\in\mathcal D'(\Gamma\cup{\widetilde \gamma}_r)\) is a positive measure on \({\widetilde \gamma}_r\) so that \(\mu(\Gamma)\) and \(\mu({\widetilde \gamma}_r)\) are well-defined. Moreover, the constant \(C_2\) only depends on \(C_0\) and on \(C_1\). By observing that the average of \(\left|\ln\frac{|w-z_j|}{r(z_j)}\right|\) with respect to the Lebesgue measure \(L(dz_j)\) over \(D\left(z^0_j,\frac{r(z^0_j)}{2C_1}\right)\) is \(O(1)\), one can get rid of the logarithmic terms to the price of making a suitable choice of \(z_j ={\widetilde z}_j\): There exists \({\widetilde z}_j\in D\left(z^0_j,\frac{r(z^0_j)}{4C_1}\right)\) such that if \(h\) and \(u\) are as in above, satisfying (1) and \[ h\ln\big |u({\widetilde z}_j)\big| \geq\phi({\widetilde z}_j)-\epsilon_j,\quad j=1,2,\dots,N, \] instead of (2), then \[ \left|\#\big(u^{-1}(0)\cap\Gamma\big)-\frac{1}{2\pi h}\mu(\Gamma)\right| \leq \frac{C_2}{h}\left(\mu({\widetilde \gamma}_r)+\sum\epsilon_j\right). \] An application to sums of exponential functions is considered and a connection with classical results on zeros of entire functions is established.