betaFeedback

WorksMathematics

Rigidity of the Peres-Virág class

Abstract

Consider the hyperbolic Gaussian analytic function (GAF) \begin{equation*} F_{\mathbb{D}}(z)=\sum_{n=0}^{\infty }ξ_{n}z^{n},\qquad |z|<1, \end{equation*} where the ξnξ_{n} are independent standard complex Gaussian variables. Peres-Virág theorem tells that the distribution of the zeros of FD(z)F_{\mathbb{D}}(z) is provided by a Determinantal Point Process (DPP). We prove that FD(z)F_{\mathbb{D}}(z) is the only GAF whose zeros are determinantal with a Hermitian kernel, up to multiplication by a zero-free holomorphic function and a one-to-one holomorphic change of variable, with a \ non-degeneracy condition imposed to exclude trivial cases. The proof identifies a geometric property of GAF's with (local) determinantal zeros: the metric defined by the first intensity must have constant curvature −4-4.

Introduction

Consider the Gaussian power series

FD(z)=∑n=0∞ξnzn,∣z∣<1,(1)F_{\mathbb{D}}(z)=\sum_{n=0}^{\infty}\xi_n z^n,\qquad|z|<1, \tag*{(1)}

where the ξn\xi_n are independent standard complex Gaussian variables. This is a Gaussian analytic function (GAF) whose covariance kernel is the reproducing kernel of the Hardy space H2(D)H^2(\mathbb{D}):

KD(z,w)=E[FD(z)FD(w)‾]=11−zw‾,z,w∈D.K_{\mathbb{D}}(z,w)=\mathbb{E}[F_{\mathbb{D}}(z)\overline{F_{\mathbb{D}}(w)}]=\frac{1}{1-z\overline{w}},\qquad z,w\in\mathbb{D}.

A celebrated result of Peres and Virág [14 Theorem 1] shows that the zeros of FDF_{\mathbb{D}} form a determinantal point process (DPP) whose correlation kernel is the reproducing kernel of the unweighted Bergman space A2(D)A^2(\mathbb{D}) [5 Chapter 1]:

KD(z,w)=1(1−zw‾)2,(2)K_{\mathbb{D}}(z,w)=\frac{1}{(1-z\overline{w})^2}, \tag*{(2)}

relative to normalized area measure dμ=π−1dx dyd\mu=\pi^{-1}dx\,dy. It is the purpose of this paper to show that (1.1) is the only nontrivial scalar example of a GAF whose zeros are determinantal, beyond the restrictions and transformations of (1.1) described in the definition below of the Peres–Virág class.

To state our main definition and result, we will need some definitions. Throughout, Ω\Omega is a nonempty connected open subset of C\mathbb{C}, dμ=π−1dx dyd\mu=\pi^{-1}dx\,dy, and

∂=12(∂x−i∂y),∂‾=12(∂x+i∂y).\partial=\frac{1}{2}(\partial_x-i\partial_y),\qquad\overline{\partial}=\frac{1}{2}(\partial_x+i\partial_y).

A Gaussian analytic function (GAF) is a random holomorphic function FF whose finite collections of values are centered complex Gaussian vectors and satisfy

E[F(z)F(w)]=0.\mathbb{E}[F(z)F(w)] = 0.

Its covariance kernel is denoted by

K(z,w)=E[F(z)F(w)‾].(3)K(z,w) = \mathbb{E}[F(z)\overline{F(w)}]. \tag*{(3)}

Thus KK is holomorphic in zz and antiholomorphic in ww. Multiplication of (1.1) by a zero-free holomorphic function preserves its zeros [6 proof of Theorem 2.5.2], while a one-to-one holomorphic change of variable ϕ:Ω⟶D\phi:\Omega\longrightarrow\mathbb{D} preserves the determinantal zero property in the domain Ω\Omega (see [6 Section 4.2] and Section 2 below). For the classical extension to bounded simply connected domains with smooth boundary see [6 Corollary 5.4.1].

Definition 1.1. Let Ω⊆C\Omega\subseteq\mathbb{C} be a nonempty connected open set, possibly Ω=C\Omega=\mathbb{C}. A Gaussian analytic function FF on Ω\Omega belongs to the Peres–Virág class if

Cov⁡(F(z),F′(z),F′′(z))>0for every z∈Ω,(4)\operatorname{Cov}(F(z),F'(z),F''(z))>0 \qquad\text{for every } z\in\Omega, \tag*{(4)}

and its zero process is a Hermitian DPP relative to dμd\mu.

The inequality in (1.4) means that the covariance matrix is positive definite. It implies that K(z,z)>0K(z,z)>0 and the first zero intensity is positive. This condition excludes the trivial case of a Gaussian linear polynomial, which has one random zero and gives a DPP (see Section 4.2). Under (1.4), the zeros are almost surely isolated and simple [6 Lemma 2.4.1].

Theorem 1.1. A Gaussian analytic function FF on Ω\Omega belongs to the Peres–Virág class if and only if there are a zero-free holomorphic function hh on Ω\Omega and a one-to-one holomorphic map ϕ:Ω⟶D\phi:\Omega\longrightarrow\mathbb{D} such that

K(z,w)=h(z)h(w)‾1−ϕ(z)ϕ(w)‾.(5)K(z,w)=\frac{h(z)\overline{h(w)}}{1-\phi(z)\overline{\phi(w)}}. \tag*{(5)}

In this case, the zero process admits a correlation kernel relative to dμd\mu, holomorphic in the first variable and antiholomorphic in the second, given by

Kϕ(z,w)=ϕ′(z)ϕ′(w)‾(1−ϕ(z)ϕ(w)‾)2.(6)K_{\phi}(z,w)=\frac{\phi'(z)\overline{\phi'(w)}}{(1-\phi(z)\overline{\phi(w)})^{2}}. \tag*{(6)}

The proof is organized along Section 3 and will be outlined after the following statement of Liouville’s theorem, which can be found in [8 Theorem C] or in [1 Theorem 9.8].

Lemma 1.2 (Liouville’s theorem). Let U⊂CU\subset\mathbb{C} be simply connected, and let g>0g>0 be smooth with

∂∂‾log⁡g(z)=2g(z).(7)\partial\overline{\partial}\log g(z)=2g(z). \tag*{(7)}

There is a holomorphic map ϕ:U→D\phi:U\to\mathbb{D} with ϕ′(z)≠0\phi'(z)\ne0 such that

g(z)=∣ϕ′(z)∣2(1−∣ϕ(z)∣2)2.(8)g(z)=\frac{|\phi'(z)|^{2}}{(1-|\phi(z)|^{2})^{2}}. \tag*{(8)}

The proof of Theorem 1.1 relies on a clear mechanism that we proceed to describe. Edelman–Kostlan’s formula [3 Section 8] for the 1-point zero intensity of a GAF with covariance kernel K(z,w)K(z,w):

ρ1(z)=∂∂ˉlog⁡K(z,z).(9)\rho_1(z)=\partial\bar{\partial}\log K(z,z). \tag*{(9)}

First recall that Sodin’s Calabi rigidity theorem [15 Theorem 2],[6, Theorem 2.5.2] determines the GAF from ρ1(z)\rho_1(z), up to multiplication by a zero-free holomorphic function. Consequently, if we show that ρ1(z)\rho_1(z) satisfies (1.7) in Ω\Omega (not assumed to be simply connected), Liouville’s theorem gives (1.8), which is the diagonal of the kernel (1.6), proving a restriction of Theorem 1.1, on an open disc U⊂ΩU\subset\Omega, with a holomorphic map ϕ:U→D\phi:U\to\mathbb{D} with ϕ′(z)≠0\phi'(z)\ne0. Somehow surprisingly, the identities obtained by imposing the determinantal structure on diagonal perturbations of the 2- and 3-point Kac–Rice intensities are enough to prove (1.7), after showing that they impose a holomorphic structure on the determinantal kernel, which is first proved to be smooth. The local version of Theorem 1.1 is finally extended to Ω\Omega by analytic continuation and ϕ\phi is shown to be globally one-to-one (this only uses the global repulsiveness of the 2-point intensity).

Our results therefore identify the local geometric mechanism for membership in the Peres–Virág class: the local determinantal assumption forces the metric ds2=∂∂ˉlog⁡K(z,z)∣dz∣2ds^2=\partial\bar{\partial}\log K(z,z)|dz|^2 to have constant curvature −4-4 (this is equivalent to (1.7)). Moreover, Definition 1.1 can be relaxed to hold locally, provided the GAF zeros have a global two-point intensity such that (2.3) satisfies G(z,w)≤1G(z,w)\le1 in the whole domain. In Section 4.4 we construct examples of GAF’s satisfying the curvature condition, with determinantal zeros locally, but not globally. This shows that global injectivity in Theorem 1.1 cannot be replaced by local injectivity. A limit case of Katori–Shirai’s GAF [7] is globally repulsive but fails the curvature condition (see § 5.3).

There is a considerable body of work motivated by the work of Peres and Virág. We list a few developments to settings beyond the scope of our results. A real-coefficient counterpart is provided by the limiting Kac series, with independent standard real Gaussian coefficients, whose zero correlations admit Pfaffian formulas. Forrester [4] obtained these through truncated random orthogonal matrices, while Matsumoto and Shirai [10 Theorems 2.1 and 2.7] gave a direct proof for real zeros on (−1,1)(-1,1) and nonreal zeros in the upper half-disk. Matrix constructions with determinantal zero processes have been developed by Krishnapur [9 Theorem 4].

The outline of the paper is straightforward. The preliminary Section 2 recalls the facts about DPPs and GAFs used in the proof. The proof of Theorem 1.1 is given in Section 3: two preparation lemmas are followed by the curvature lemma and a two-step proof of the theorem. We conclude with examples and comments in Section 4.

Preliminaries on DPPs and GAFs

Hermitian determinantal point processes

Let Z\mathcal{Z} be a random subset of Ω\Omega, with no repeated points and finitely many points in every compact subset. Its nn-point intensity ρn\rho_n, relative to the normalized area measure dμ=π−1dx dyd\mu=\pi^{-1}dx\,dy, is defined by

E∑z1,…,zn∈Z≠f(z1,…,zn)=∫Ωnf(z1,…,zn)ρn(z1,…,zn) dμ(z1)⋯dμ(zn)\mathbb{E}\sum_{\substack{z_1,\ldots,z_n\in\mathcal{Z}\\ \ne}}f(z_1,\ldots,z_n)=\int_{\Omega^n}f(z_1,\ldots,z_n)\rho_n(z_1,\ldots,z_n)\,d\mu(z_1)\cdots d\mu(z_n)

for nonnegative measurable ff. The sum is over ordered tuples of distinct points. Integrating ρ1\rho_1 over a set gives its expected number of points, and integrating ρ2\rho_2 over a subset of Ω×Ω\Omega\times\Omega gives its expected number of ordered pairs of distinct points.

For a Hermitian DPP we use the usual operator definition. The correlation operator L\mathcal{L} on L2(Ω,dμ)L^2(\Omega,d\mu) is selfadjoint and satisfies 0≤L≤I0 \le\mathcal{L} \le I. For every relatively compact set UU, the operator 1UL1U\mathbf{1}_U\mathcal{L}\mathbf{1}_U has finite trace. Here 1U\mathbf{1}_U denotes multiplication by the indicator of UU. We denote the measurable Hermitian integral kernel of L\mathcal{L} by K(z,w)\mathcal{K}(z,w), with the diagonal interpreted through the local spectral expansion used in the proof of Lemma 3.2. These are the standard operator assumptions for Hermitian DPPs; see [6, 2, 16]. The intensities are then given by

ρn(z1,…,zn)=det⁡[K(zi,zj)]i,j=1nalmost everywhere.(10)\rho_n(z_1,\ldots,z_n)=\det[\mathcal{K}(z_i,z_j)]_{i,j=1}^n \quad\text{almost everywhere}. \tag*{(10)}

The determinant for two points immediately gives

∣K(z,w)∣2=ρ1(z)ρ1(w)−ρ2(z,w).(11)|\mathcal{K}(z,w)|^2=\rho_1(z)\rho_1(w)-\rho_2(z,w). \tag*{(11)}

Multiplying the kernel by the factors

K(z,w)⟼eiθ(z)K(z,w)e−iθ(w)\mathcal{K}(z,w)\longmapsto e^{i\theta(z)}\mathcal{K}(z,w)e^{-i\theta(w)}

and taking the complex conjugate of the kernel, leaves every determinant unchanged. Restriction to a domain U⊆ΩU\subseteq\Omega preserves the DPP property [6 Remark 4.2.5]. If ψ:V⟶U\psi:V\longrightarrow U is a one-to-one holomorphic map from a planar domain VV onto UU, a kernel relative to dμd\mu for the process ψ−1(Z∩U)\psi^{-1}(Z\cap U) on VV is

ψ′(z)ψ′(w)‾K(ψ(z),ψ(w)),z,w∈V.\psi'(z)\overline{\psi'(w)}\mathcal{K}(\psi(z),\psi(w)),\qquad z,w\in V.

Consequently the normalized pair correlation function, defined as the ratio

G(z,w)=ρ2(z,w)ρ1(z)ρ1(w),(12)G(z,w)=\frac{\rho_2(z,w)}{\rho_1(z)\rho_1(w)}, \tag*{(12)}

is the same in either coordinate, after composition with the coordinate map.

Gaussian analytic functions and Kac–Rice intensities

A standard complex Gaussian variable ξ\xi has density e−∣z∣2e^{-|z|^2} relative to dμd\mu on C\mathbb{C}. Its real and imaginary parts are independent, each with variance 1/21/2. Derivatives of a GAF are again jointly Gaussian with its values. In (1.9), the derivatives of K(z,z)K(z,z) are taken after restriction to the diagonal. For any jointly centered circular complex Gaussian vectors Y,ξY,\xi, with Cov⁡(ξ)\operatorname{Cov}(\xi) invertible, the Gaussian conditioning identity is

Cov⁡(Y∣ξ=0)=Cov⁡(Y)−Cov⁡(Y,ξ)Cov⁡(ξ)−1Cov⁡(ξ,Y).(13)\operatorname{Cov}(Y\mid\xi=0)=\operatorname{Cov}(Y)-\operatorname{Cov}(Y,\xi)\operatorname{Cov}(\xi)^{-1}\operatorname{Cov}(\xi,Y). \tag*{(13)}

The first intensity is provided by Edelman–Kostlan formula (1.9). We will follow [6 Section 3.4, Corollary 3.4.2] in the presentation of the Kac–Rice formulas providing the higher intensities. For distinct points z1,…,znz_1,\ldots,z_n at which the covariance of the values is nonsingular, put

Aij=K(zi,zj),Bij=E[F′(zi)F(zj)‾],Cij=E[F′(zi)F′(zj)‾].A_{ij}=K(z_i,z_j),\qquad B_{ij}=\mathbb{E}[F'(z_i)\overline{F(z_j)}],\qquad C_{ij}=\mathbb{E}[F'(z_i)\overline{F'(z_j)}].

By (13), conditioning on the values being zero gives a centered Gaussian vector of derivatives, whose covariance,

Λ=Cov⁡((F′(zi))i=1n∣F(z1)=⋯=F(zn)=0)=C−BA−1B∗,(14)\Lambda=\operatorname{Cov}\left((F'(z_i))_{i=1}^n\mid F(z_1)=\cdots=F(z_n)=0\right)=C-BA^{-1}B^*, \tag*{(14)}

is called the conditional covariance matrix. The Kac–Rice formula providing the nn-point intensities of the zeros of a GAF is

ρn(z1,…,zn)=E[∏j=1n∣F′(zj)∣2 | F(z1)=⋯=F(zn)=0]det⁡A=per⁡Λdet⁡A.(15)\rho_n(z_1,\ldots,z_n)=\frac{\mathbb{E}\left[\prod_{j=1}^{n}|F'(z_j)|^2 \,\middle|\, F(z_1)=\cdots=F(z_n)=0\right]}{\det A}=\frac{\operatorname{per}\Lambda}{\det A}. \tag*{(15)}

Here per⁡M=∑σ∏iMi,σ(i)\operatorname{per} M=\sum_{\sigma}\prod_i M_{i,\sigma(i)}, where the sum is over all permutations of {1,…,n}\{1,\ldots,n\}. The second equality follows by computing the Gaussian moment in the numerator [14 Proposition 8].

For a GAF in the Peres–Virág class, the determinantal assumption implies that the same intensity has two expressions:

ρn(z1,…,zn)=per⁡Λdet⁡A=det⁡[K(zi,zj)]i,j=1n,(16)\rho_n(z_1,\ldots,z_n)=\frac{\operatorname{per}\Lambda}{\det A}=\det[K(z_i,z_j)]_{i,j=1}^{n}, \tag*{(16)}

where Kac–Rice applies, with the determinantal equality initially holding almost everywhere.

Proof of Theorem 1.1

Perturbation of diagonal Kac–Rice intensities

To study the intensities near the diagonal, we need some identities involving the covariance matrices of the values and derivatives and their determinants, evaluated at the same point zz:

Jm(z)=[E(F(i)(z)F(j)(z)‾)]i,j=0m,Dm(z)=det⁡Jm(z).J_m(z)=\left[\mathbb{E}\left(F^{(i)}(z)\overline{F^{(j)}(z)}\right)\right]_{i,j=0}^{m},\qquad D_m(z)=\det J_m(z).

The required determinant identities needed below are

D0(z)=K(z,z),D1(z)=K(z,z)2ρ1(z),D2(z)=K(z,z)3ρ1(z)2(2ρ1(z)+∂∂‾log⁡ρ1(z)).(17)\begin{aligned} D_0(z)&=K(z,z),\\ D_1(z)&=K(z,z)^2\rho_1(z),\\ D_2(z)&=K(z,z)^3\rho_1(z)^2\left(2\rho_1(z)+\partial\overline{\partial}\log\rho_1(z)\right). \tag*{(17)} \end{aligned}

For the last identity in (3.1) combine log⁡D1(z)=2log⁡K(z,z)+log⁡ρ1(z)\log D_1(z)=2\log K(z,z)+\log\rho_1(z) with Jacobi’s determinant identity,

∂∂‾log⁡D1(z)=D2(z)D0(z)D1(z)2.\partial\overline{\partial}\log D_1(z)=\frac{D_2(z)D_0(z)}{D_1(z)^2}.

The conditional variance of F′′(z)F''(z) is obtained by conditioning on the two Gaussian variables F(z),F′(z)F(z),F'(z) at the same point. Write

J2(z)=(J1(z)b(z)∗b(z)c(z)),c(z)=E∣F′′(z)∣2,J_2(z)=\begin{pmatrix}J_1(z)&b(z)^*\\b(z)&c(z)\end{pmatrix},\qquad c(z)=\mathbb{E}|F''(z)|^2,
b(z)=(E[F′′(z)F(z)‾],E[F′′(z)F′(z)‾]).b(z)=\left(\mathbb{E}[F''(z)\overline{F(z)}],\mathbb{E}[F''(z)\overline{F'(z)}]\right).

The Gaussian conditioning formula (2.4) gives the variance c(z)−b(z)J1(z)−1b(z)∗c(z)-b(z)J_1(z)^{-1}b(z)^*, while the block determinant formula gives

D2(z)=D1(z)(c(z)−b(z)J1(z)−1b(z)∗).D_2(z)=D_1(z)\left(c(z)-b(z)J_1(z)^{-1}b(z)^*\right).

Together with (3.1), this yields

V2(z):=Var⁡(F′′(z)∣F(z)=F′(z)=0)=D2(z)D1(z)=K(z,z)ρ1(z)(2ρ1(z)+∂∂‾log⁡ρ1(z)).(18)\begin{aligned} V_2(z)&:=\operatorname{Var}(F''(z)\mid F(z)=F'(z)=0)=\frac{D_2(z)}{D_1(z)}\\ &=K(z,z)\rho_1(z)\left(2\rho_1(z)+\partial\overline{\partial}\log\rho_1(z)\right). \tag*{(18)} \end{aligned}

Assumption (1.4) makes both ρ1(z)\rho_1(z) and 2ρ1(z)+∂∂‾log⁡ρ1(z)2\rho_1(z)+\partial\overline{\partial}\log\rho_1(z) strictly positive.

Lemma 3.1. Let FF be a GAFGAF on Ω\Omega satisfying (4). Then ρ2\rho_2 extends real-analytically across z=wz=w, with

ρ2(z,z+δ)=ρ1(z)(2ρ1(z)+∂∂‾log⁡ρ1(z))28∣δ∣2+O(∣δ∣3).(19)\rho_2(z,z+\delta)=\frac{\rho_1(z)\left(2\rho_1(z)+\partial\overline{\partial}\log\rho_1(z)\right)^2}{8}|\delta|^2+O(|\delta|^3). \tag*{(19)}

Moreover, ρ3\rho_3 is continuous at nearby distinct triples and, for fixed distinct u1,u2,u3u_1,u_2,u_3,

ρ3(z+ϵu1,z+ϵu2,z+ϵu3)=O(∣ϵ∣6).(20)\rho_3(z+\epsilon u_1,z+\epsilon u_2,z+\epsilon u_3)=O(|\epsilon|^6). \tag*{(20)}

Proof. The mechanism of the proof is inspired by the near-diagonal Kac–Rice argument of Peres-Virág [14 Lemma 32] and Michelen-Yakir [11 Section 3.2]. We use divided differences for both estimates, replacing nearly identical values by quantities whose covariance remains nonsingular as the points approach one another. Conditioning on the Kac-Rice conditions F(z1)=⋯=F(zn)=0F(z_1)=\cdots=F(z_n)=0, we express each derivative as a divided difference multiplied by the product of the corresponding differences between the points (this product also appears in the two-sided bounds of [12 Theorem 1.1], under stronger nondegeneracy assumptions). Let F[z1]=F(z1)F[z_1]=F(z_1) and

F[z1,…,zm]=F[z2,…,zm]−F[z1,…,zm−1]zm−z1.F[z_1,\ldots,z_m]=\frac{F[z_2,\ldots,z_m]-F[z_1,\ldots,z_{m-1}]}{z_m-z_1}.

When nodes coincide, the value is the analytic limit. In particular,

F[z,w]=F(w)−F(z)w−z,F[z,z]=F′(z),F[z,z,z]=F′′(z)2.F[z,w]=\frac{F(w)-F(z)}{w-z},\qquad F[z,z]=F'(z),\qquad F[z,z,z]=\frac{F''(z)}{2}.

Now replace

(F(z),F(w))  by  (F[z],F[z,w]).(F(z),F(w))\ \text{ by }\ (F[z],F[z,w]).

The new pair tends to (F(z),F′(z))(F(z),F'(z)) as w→zw\to z, whose covariance is nonsingular by (4). For w≠zw\ne z, the conditions F(z)=F(w)=0F(z)=F(w)=0 and F[z]=F[z,w]=0F[z]=F[z,w]=0 are equivalent. The linear change of variables has determinant (w−z)−1(w-z)^{-1}, so the Kac–Rice formula (15) gives

ρ2(z,w)=E[∣F′(z)F′(w)∣2∣F[z]=F[z,w]=0]∣w−z∣2det⁡Cov⁡(F[z],F[z,w]).(21)\rho_2(z,w)=\frac{\mathbb{E}\left[|F'(z)F'(w)|^2\mid F[z]=F[z,w]=0\right]}{|w-z|^2\det\operatorname{Cov}(F[z],F[z,w])}. \tag*{(21)}

The divided-difference identities give

F′(z)=F[z,w]−(w−z)F[z,z,w],F′(w)=F[z,w]+(w−z)F[z,w,w].(22)\begin{aligned} F'(z)&=F[z,w]-(w-z)F[z,z,w],\\ F'(w)&=F[z,w]+(w-z)F[z,w,w]. \tag*{(22)} \end{aligned}

Under the conditioning F[z]=F[z,w]=0F[z]=F[z,w]=0, the scaled derivatives therefore satisfy

(F′(z)w−z,F′(w)w−z)=(−F[z,z,w],F[z,w,w]).(23)\left(\frac{F'(z)}{w-z},\frac{F'(w)}{w-z}\right)=\left(-F[z,z,w],F[z,w,w]\right). \tag*{(23)}

Writing δ=w−z\delta=w-z, Taylor’s formula gives

F[z,w]=F′(z)+δ2F′′(z)+O(∣δ∣2),F′(w)=F′(z)+δF′′(z)+O(∣δ∣2).\begin{aligned} F[z,w]&=F'(z)+\frac{\delta}{2}F''(z)+O(|\delta|^2),\\ F'(w)&=F'(z)+\delta F''(z)+O(|\delta|^2). \end{aligned}

We next justify the regularity of their conditional moments at the diagonal. Joint covariances of divided differences are obtained by taking divided differences of KK in its first variable and conjugate divided differences in its second. For example,

E[F[z]F[z,w]‾]=K(z,w)−K(z,z)w‾−z‾.\mathbb{E}[F[z]\overline{F[z,w]}]=\frac{K(z,w)-K(z,z)}{\overline{w}-\overline{z}}.

Since KK is holomorphic–antiholomorphic, the apparent singularity at w=zw=z is removable. The same argument applies to all the entries of the covariance of (F[z],F[z,w],F[z,z,w],F[z,w,w])(F[z], F[z,w], F[z,z,w], F[z,w,w]), which extend jointly as real-analytic functions of z,wz,w near the diagonal. Apply the Gaussian conditioning identity (13) with

A(z,w)=Cov⁡(F[z],F[z,w]),C(z,w)=Cov⁡(F[z,z,w],F[z,w,w]),\begin{aligned} A(z,w)&=\operatorname{Cov}(F[z],F[z,w]),\\ C(z,w)&=\operatorname{Cov}(F[z,z,w],F[z,w,w]), \end{aligned}

and with B(z,w)B(z,w) the cross-covariance of the second pair with the first. On the diagonal A(z,z)=J1(z)>0A(z,z)=J_1(z)>0. Thus det⁡A(z,w)\det A(z,w) stays positive nearby, and the cofactor formula shows that A(z,w)−1A(z,w)^{-1} is real-analytic there. As a result, the conditional covariance matrix of (F[z,z,w],F[z,w,w])(F[z,z,w],F[z,w,w]) is

Λ(z,w)=C(z,w)−B(z,w)A(z,w)−1B(z,w)∗,\Lambda(z,w)=C(z,w)-B(z,w)A(z,w)^{-1}B(z,w)^*,

with real-analytic entries. The corresponding conditional fourth moment is

E[∣F[z,z,w]F[z,w,w]∣2∣F[z]=F[z,w]=0]=Λ11(z,w)Λ22(z,w)+∣Λ12(z,w)∣2.\begin{aligned} \mathbb{E}\left[|F[z,z,w]F[z,w,w]|^2\mid F[z]=F[z,w]=0\right] &=\Lambda_{11}(z,w)\Lambda_{22}(z,w)+|\Lambda_{12}(z,w)|^2. \end{aligned}

Since this is a polynomial in the covariance entries, it is real-analytic as well. Each conditioned derivative contributes a factor w−zw-z. Consequently, the product of their squared moduli contributes the factor ∣w−z∣4|w-z|^4, and the numerator in (21) is ∣w−z∣4|w-z|^4 times a real-analytic function. The denominator contains ∣w−z∣2|w-z|^2 times a real-analytic covariance determinant whose value at w=zw=z is D1(z)>0D_1(z)>0. Therefore ρ2(z,w)/∣w−z∣2\rho_2(z,w)/|w-z|^2 extends real-analytically across the diagonal.

We now compute the leading quadratic coefficient of ρ2(z,w)\rho_2(z,w) (its exact value will be later used to identify p(z)p(z) in (29)). At w=zw=z, the conditioning pair becomes (F(z),F′(z))(F(z),F'(z)). Substitution into the divided-difference identities (22) shows that both F[z,z,w]F[z,z,w] and F[z,w,w]F[z,w,w] tend to F′′(z)/2F''(z)/2. By (23), the two scaled derivatives have limits −F′′(z)/2-F''(z)/2 and F′′(z)/2F''(z)/2, respectively. The product of the squared moduli of the two limiting scaled derivatives is ∣F′′(z)∣416\frac{|F''(z)|^4}{16}. Thus the value of the remaining real-analytic function on the diagonal is

116E[∣F′′(z)∣4∣F(z)=F′(z)=0].\frac{1}{16}\mathbb{E}\left[|F''(z)|^4\mid F(z)=F'(z)=0\right].

Under this conditioning, F′′(z)F''(z) is a centered circular complex Gaussian variable with variance V2(z)V_2(z), defined in (18). Its fourth moment is twice the square of its variance. Hence the last expression equals

116 2V2(z)2=V2(z)28.\frac{1}{16}\,2V_2(z)^2=\frac{V_2(z)^2}{8}.

Dividing by the limiting covariance determinant D1(z)D_1(z) gives the diagonal value V2(z)2/(8D1(z))V_2(z)^2/(8D_1(z)). As two zeros approach one another, their vanishing conditions approach F(z)=F′(z)=0F(z)=F'(z)=0. Accordingly,

ρ2(z,z+δ)=V2(z)28D1(z)∣δ∣2+O(∣δ∣3).\rho_2(z,z+\delta)=\frac{V_2(z)^2}{8D_1(z)}|\delta|^2+O(|\delta|^3).

Substituting (17) and (18) proves (19), with a strictly positive leading coefficient. It remains to prove (20). We again use divided differences to keep the value covariance nonsingular and isolate the factors arising from the distances between the points. Here only the order of vanishing is needed, so it suffices to bound the remaining conditional moment without computing its value. Let Δ=∏i<j(zj−zi)\Delta= \prod_{i<j}(z_j-z_i) and set

ξ=(F[z1],F[z1,z2],F[z1,z2,z3]).\xi= (F[z_1], F[z_1,z_2], F[z_1,z_2,z_3]).

The transformation from the three values to ξ\xi has determinant Δ−1\Delta^{-1}, so

det⁡Cov⁡(F(z1),F(z2),F(z3))=∣Δ∣2det⁡Cov⁡(ξ).\det\operatorname{Cov}(F(z_1), F(z_2), F(z_3)) = |\Delta|^2 \det\operatorname{Cov}(\xi).

As z1,z2,z3→zz_1,z_2,z_3 \to z, the covariance of ξ\xi tends to that of (F(z),F′(z),F′′(z)/2)(F(z),F'(z),F''(z)/2), whose determinant is D2(z)/4>0D_2(z)/4 > 0. After restricting the points to a sufficiently small disk, this covariance remains positive definite and its inverse is bounded. If the three values are conditioned to be zero, the interpolation identity in [11 Lemma 3.5] gives

F′(zi)=∏j≠i(zi−zj)F[z1,z2,z3,zi],i=1,2,3.F'(z_i) = \prod_{j\ne i}(z_i-z_j) F[z_1,z_2,z_3,z_i], \qquad i=1,2,3.

The covariance entries of these last divided differences extend real-analytically when the nodes coincide. Applying the Gaussian conditional covariance formula (13) to the vector

(F[z1,z2,z3,zi])i=13,(F[z_1,z_2,z_3,z_i])_{i=1}^{3},

with conditioning vector ξ\xi, therefore gives bounded conditional covariances for these divided differences. Since

∏i=13∏j≠i∣zi−zj∣2=∣Δ∣4,\prod_{i=1}^{3}\prod_{j\ne i}|z_i-z_j|^2 = |\Delta|^4,

the numerator in Kac–Rice contains ∣Δ∣4|\Delta|^4 times a bounded continuous function. Its denominator contains ∣Δ∣2|\Delta|^2 times a positive continuous function. Thus ρ3\rho_3 is ∣Δ∣2|\Delta|^2 times a bounded continuous function. In particular it is continuous for nearby distinct triples. For zi=z+ϵuiz_i = z+\epsilon u_i, with the uiu_i fixed and distinct,

∣Δ∣2=∣ϵ∣6∏i<j∣uj−ui∣2,|\Delta|^2 = |\epsilon|^6 \prod_{i<j}|u_j-u_i|^2,

which proves (20). □

A holomorphic factorization for the kernel of the Peres–Virág class

The identities (2.7) already involve a determinantal kernel K\mathcal{K}, whose existence is part of the Peres–Virág assumption. At this stage, however, we only know that it is measurable. The next result shows that the process admits a correlation kernel with the smoothness and analytic structure required for the proof of Theorem 1.1. This structure follows by expanding the 2- and 3-point determinantal intensities near the diagonal and comparing these expansions with the Kac–Rice intensities obtained in the previous section.

Lemma 3.2. Let FF belong to the Peres–Virág class. On a sufficiently small disk around each point of Ω\Omega, its zero process admits a correlation kernel of the form

K(z,w)=a(z)a(w)‾H(z,w),(24)K(z,w) = a(z)\overline{a(w)}H(z,w), \tag*{(24)}

where aa is smooth and never zero, and HH is a positive kernel holomorphic in its first variable and antiholomorphic in its second.

Proof. We first obtain a smooth correlation kernel in Step 1, and then compare its two- and three-point determinants with the estimates of Lemma 3.1 in Step 2, leading to (3.8).

Step 1

Recall that the DPP property is preserved by restriction to a domain [6 Remark 4.2.5]. The local trace-class assumption [6 Assumption 4.2.3], gives

tr⁡(1UL1U)<∞for every relatively compact U⊂Ω.\operatorname{tr}(\mathbf{1}_{U}\mathcal{L}\mathbf{1}_{U}) < \infty\qquad\text{for every relatively compact } U \subset\Omega.

Fix a point z0z_0 and take a sufficiently small relatively compact disk UU around it. We can choose orthonormal eigenfunctions eje_j with eigenvalues λj≥0\lambda_j \ge0 for the positive trace-class operator 1UL1U\mathbf{1}_{U}\mathcal{L}\mathbf{1}_{U} on L2(U,dμ)L^2(U,d\mu). The kernel and its diagonal are given, almost everywhere, by the expansions

K(z,w)=∑jλjej(z)ej(w)‾,ρ1(z)=∑jλj∣ej(z)∣2.K(z,w)=\sum_j \lambda_j e_j(z)\overline{e_j(w)}, \qquad\rho_1(z)=\sum_j \lambda_j |e_j(z)|^2.

Since ∑jλj<∞\sum_j\lambda_j<\infty, the vector

v(z)=(λjej(z))j≥1(25)v(z)=\bigl(\sqrt{\lambda_j}e_j(z)\bigr)_{j\ge1} \tag*{(25)}

belongs to ℓ2\ell^2 for almost every zz. Thus

K(z,w)=⟨v(z),v(w)⟩ℓ2,∥v(z)∥ℓ22=ρ1(z),K(z,w)=\langle v(z),v(w)\rangle_{\ell^2}, \qquad\|v(z)\|_{\ell^2}^2=\rho_1(z),

with the inner products linear in the first variable. Equation (11) gives

∣K(z,w)∣2=ρ1(z)ρ1(w)−ρ2(z,w)|K(z,w)|^2=\rho_1(z)\rho_1(w)-\rho_2(z,w)

almost everywhere. By the Edelman–Kostlan formula and Lemma 3.1, the right-hand side is smooth for every pair in a sufficiently small disk, including the diagonal. Now let

X(z)=v(z)v(z)∗,X(z)=v(z)v(z)^*,

where v(z)v(z)∗v(z)v(z)^* is the operator x↦⟨x,v(z)⟩ℓ2v(z)x\mapsto\langle x,v(z)\rangle_{\ell^2}v(z). In the real Hilbert space of selfadjoint Hilbert–Schmidt operators, whose norm satisfies ∥X∥HS2=tr⁡(X2)\|X\|_{\mathrm{HS}}^2=\operatorname{tr}(X^2), we have

∥X(z)∥HS=ρ1(z),\|X(z)\|_{\mathrm{HS}}=\rho_1(z),
⟨X(z),X(w)⟩HS=tr⁡(X(z)X(w))=ρ1(z)ρ1(w)−ρ2(z,w)\langle X(z),X(w)\rangle_{\mathrm{HS}}=\operatorname{tr}(X(z)X(w))=\rho_1(z)\rho_1(w)-\rho_2(z,w)

almost everywhere. Thus XX is locally square-integrable and its inner products have a smooth representative. The smoothness criterion for Hilbert-space-valued maps therefore gives a smooth function equal to XX pointwise [13 Theorem 7.1]. To obtain an almost-everywhere identity, set

M(z,w)=⟨X(z),X(w)⟩HS a.e.,Xε=ηε∗X,M(z,w)=\langle X(z),X(w)\rangle_{\mathrm{HS}}\ \text{a.e.}, \qquad X_\varepsilon=\eta_\varepsilon* X,

where MM denotes the smooth representative of the scalar kernel and ηε\eta_\varepsilon is a standard real mollifier with unit integral. For real-coordinate multi-indices α,β\alpha,\beta, Fubini’s theorem and integration by parts give

⟨DαXε(z),DβXδ(w)⟩HS=∬ηε(z−u)ηδ(w−v)DuαDvβM(u,v) du dv.\langle D^\alpha X_\varepsilon(z),D^\beta X_\delta(w)\rangle_{\mathrm{HS}} =\iint\eta_\varepsilon(z-u)\eta_\delta(w-v)D_u^\alpha D_v^\beta M(u,v)\,du\,dv.

Here z,wz,w lie in a smaller disk and the supports of the shifted mollifiers are contained in the original disk. The right-hand side converges locally uniformly to DzαDwβM(z,w)D_z^\alpha D_w^\beta M(z,w). Taking $\beta=\alpha and w=zw=z, and expanding the squared norm of DαXε(z)−DαXδ(z)D^\alpha X_\varepsilon(z)-D^\alpha X_\delta(z), shows that the mollified maps and all their derivatives are locally uniformly Cauchy. Since Xε→XX_\varepsilon\to X in Lloc2L^2_{\mathrm{loc}} and equals XX almost everywhere, the limit is a smooth representative of XX, which we continue to denote by XX. By continuity, the smooth representative is positive and has rank at most one everywhere, since ∥X(z)∥HS=ρ1(z)>0\lVert X(z)\rVert_{\mathrm{HS}}=\rho_1(z)>0. Since X(z0)X(z_0) is positive and nonzero, choose a fixed vector e∈ℓ2e\in\ell^2 such that ⟨X(z0)e,e⟩ℓ2>0\langle X(z_0)e,e\rangle_{\ell^2}>0. By continuity, there is a disk centered at z0z_0, contained in the preceding disk, where ⟨X(z)e,e⟩ℓ2\langle X(z)e,e\rangle_{\ell^2} is positive. Define

v~(z)=X(z)e⟨X(z)e,e⟩ℓ2.(26)\widetilde{v}(z)=\frac{X(z)e}{\sqrt{\langle X(z)e,e\rangle_{\ell^2}}}. \tag*{(26)}

The denominator is smooth and positive, so v~\widetilde{v} is smooth. For a positive operator of rank one, this formula gives X(z)=v~(z)v~(z)∗X(z)=\widetilde{v}(z)\widetilde{v}(z)^*. Since we also have X(z)=v(z)v(z)∗X(z)=v(z)v(z)^* almost everywhere, v(z)v(z) and v~(z)\widetilde{v}(z) differ almost everywhere by a scalar of modulus one. Consequently, replacing K(z,w)K(z,w) by ⟨v~(z),v~(w)⟩ℓ2\langle\widetilde{v}(z),\widetilde{v}(w)\rangle_{\ell^2}, multiplies each row and the corresponding column of every correlation determinant by reciprocal phase factors, leaving it unchanged almost everywhere. We have obtained the smooth kernel K~(z,w)=⟨v~(z),v~(w)⟩ℓ2\widetilde{K}(z,w)=\langle\widetilde{v}(z),\widetilde{v}(w)\rangle_{\ell^2}.

Step 2. The determinants formed from the smooth correlation kernel obtained in Step 1, K~(z,w)=⟨v~(z),v~(w)⟩ℓ2\widetilde{K}(z,w)=\langle\widetilde{v}(z),\widetilde{v}(w)\rangle_{\ell^2}, agree pointwise with the continuous two- and three-point intensities of Lemma 3.1 at nearby distinct pairs and triples. Indeed, they agree almost everywhere and both sides are continuous. We can therefore Taylor-expand these determinants and compare their coefficients with (3.3) and (3.4). From now on, write vv for v~\widetilde{v} and KK for K~\widetilde{K}. We now prove the local factorization (3.8). At a fixed point zz, let P⊥P^\perp be the orthogonal projection onto the subspace perpendicular to v(z)v(z), and set

u(z)=P⊥∂v(z),w(z)=P⊥∂‾v(z).u(z)=P^\perp\partial v(z),\qquad w(z)=P^\perp\overline{\partial}v(z).

The two-point intensity is the Gram determinant of v(z)v(z) and v(z+δ)v(z+\delta). Removing the component of the second vector along v(z)v(z) gives

ρ2(z,z+δ)=∥v(z)∥2∥P⊥v(z+δ)∥2.\rho_2(z,z+\delta)=\lVert v(z)\rVert^2\lVert P^\perp v(z+\delta)\rVert^2.

Since P⊥v(z)=0P^\perp v(z)=0, Taylor’s formula gives

ρ2(z,z+δ)=∥v(z)∥2∥u(z)δ+w(z)δ‾∥2+O(∣δ∣3)=∥v(z)∥2((∥u(z)∥2+∥w(z)∥2)∣δ∣2+2Re⁡(⟨u(z),w(z)⟩δ2))+O(∣δ∣3).\begin{aligned} \rho_2(z,z+\delta) &=\lVert v(z)\rVert^2\lVert u(z)\delta+w(z)\overline{\delta}\rVert^2+O(|\delta|^3)\\ &=\lVert v(z)\rVert^2\left((\lVert u(z)\rVert^2+\lVert w(z)\rVert^2)|\delta|^2+2\operatorname{Re}(\langle u(z),w(z)\rangle\delta^2)\right)+O(|\delta|^3). \end{aligned}

The term involving δ2\delta^2 vanishes because the coefficient in (3.3) is strictly positive and independent of the direction of δ\delta. Thus, from the 2-point identities, we have obtained

⟨u(z),w(z)⟩=0,∥v(z)∥2(∥u(z)∥2+∥w(z)∥2)>0.(27)\langle u(z),w(z)\rangle=0,\qquad\lVert v(z)\rVert^2(\lVert u(z)\rVert^2+\lVert w(z)\rVert^2)>0. \tag*{(27)}

To factor a holomorphic factor out of v(z)v(z) we now use the 3-point identities. For real ε\varepsilon, set

r1=P⊥v(z+ε),r2=P⊥v(z+iε).r_1=P^\perp v(z+\varepsilon),\qquad r_2=P^\perp v(z+i\varepsilon).

The three-point intensity is the Gram determinant of v(z),v(z+ε),v(z+iε)v(z),v(z+\varepsilon),v(z+i\varepsilon). Removing the components of the last two vectors along v(z)v(z) leaves the determinant unchanged. The remaining vectors r1,r2r_1,r_2 are perpendicular to v(z)v(z), so the Gram matrix becomes block diagonal:

ρ3(z,z+ε,z+iε)=∥v(z)∥2(∥r1∥2∥r2∥2−∣⟨r1,r2⟩∣2).\rho_3(z,z+\varepsilon,z+i\varepsilon)=\lVert v(z)\rVert^2\left(\lVert r_1\rVert^2\lVert r_2\rVert^2-\lvert\langle r_1,r_2\rangle\rvert^2\right).

Taylor’s formula gives

r1=ϵ(u+w)+O(ϵ2),r2=iϵ(u−w)+O(ϵ2).r_1=\epsilon(u+w)+O(\epsilon^2),\qquad r_2=i\epsilon(u-w)+O(\epsilon^2).

Substituting these expressions into the determinant gives

ρ3(z,z+ϵ,z+iϵ)=4ϵ4∥v(z)∥2(∥u(z)∥2∥w(z)∥2−∣⟨u(z),w(z)⟩∣2)+O(∣ϵ∣5).\rho_3(z,z+\epsilon,z+i\epsilon)=4\epsilon^4\lVert v(z)\rVert^2\left(\lVert u(z)\rVert^2\lVert w(z)\rVert^2-\lvert\langle u(z),w(z)\rangle\rvert^2\right)+O(\lvert\epsilon\rvert^5).

On the other hand, (3.4) bounds this intensity by O(∣ϵ∣6)O(\lvert\epsilon\rvert^6). Divide by ϵ4\epsilon^4 and let ϵ→0\epsilon\to0. Using (3.11), we obtain

∥u(z)∥2∥w(z)∥2=0.\lVert u(z)\rVert^2\lVert w(z)\rVert^2=0.

Thus exactly one of u(z),w(z)u(z),w(z) vanishes at each point. The sets where ∥u(z)∥>0\lVert u(z)\rVert>0 and where ∥w(z)∥>0\lVert w(z)\rVert>0 are disjoint open sets whose union is the disk. Connectedness shows that the same choice holds throughout the disk. Suppose w=0w=0. Then ∂ˉv(z)=b(z)v(z)\bar{\partial}v(z)=b(z)v(z), where

b(z)=⟨∂ˉv(z),v(z)⟩ℓ2∥v(z)∥ℓ22b(z)=\frac{\langle\bar{\partial}v(z),v(z)\rangle_{\ell^2}}{\lVert v(z)\rVert_{\ell^2}^2}

is smooth. On a smaller disk the Cauchy–Green formula gives a smooth solution B(z)B(z) of ∂ˉB=b(z)\bar{\partial}B=b(z). Let f(z)=e−B(z)v(z)f(z)=e^{-B(z)}v(z) and a(z)=eB(z)a(z)=e^{B(z)}. Direct differentiation gives

∂ˉf(z)=e−B(z)(∂ˉv(z)−(b(z))v(z))=0.\bar{\partial}f(z)=e^{-B(z)}\left(\bar{\partial}v(z)-(b(z))v(z)\right)=0.

Hence f(z)f(z) is holomorphic and a(z)a(z) is smooth and never zero. Taking

H(z,w)=⟨f(z),f(w)⟩ℓ2=⟨v(z),v(w)⟩ℓ2a(z)a(w)H(z,w)=\langle f(z),f(w)\rangle_{\ell^2}=\frac{\langle v(z),v(w)\rangle_{\ell^2}}{a(z)a(w)}

gives a positive Gram kernel, holomorphic in the first variable and antiholomorphic in the second, and the smooth kernel obtained in the previous step satisfies K(z,w)=a(z)a(w)H(z,w)K(z,w)=a(z)a(w)H(z,w). If u=0u=0, simply replace KK by its complex conjugate. This preserves all correlation determinants and interchanges the roles of uu and ww. We have completed the proof of (3.8). □

The curvature Lemma

Lemma 3.3. Let FF belong to the Peres–Virág class, and let ρ1(z)\rho_1(z) be its first zero intensity relative to dμd\mu. Then

∂ˉ∂log⁡ρ1(z)=2ρ1(z),z∈Ω.(28)\bar{\partial}\partial\log\rho_1(z)=2\rho_1(z),\qquad z\in\Omega. \tag*{(28)}

Thus the conformal metric ds2=ρ1(z)∣dz∣2ds^2=\rho_1(z)\lvert dz\rvert^2 has constant Gaussian curvature −4-4.

Proof.

Step 1. In this step we establish the compatibility identity, valid near the diagonal,

−∂z∂ˉzlog⁡(1−G(z,w))=p(z),p(z)=(2ρ1(z)+∂ˉ∂log⁡ρ1(z))28ρ1(z)>0.(29)-\partial_z\bar{\partial}_z\log(1-G(z,w))=p(z),\qquad p(z)=\frac{\left(2\rho_1(z)+\bar{\partial}\partial\log\rho_1(z)\right)^2}{8\rho_1(z)}>0. \tag*{(29)}

We start by observing that, when computing the normalized pair correlation function G(z,w)G(z,w) with the kernel (3.8) obtained in Lemma 3.2, the scalar factors cancel, giving:

G(z,w)=1−∣H(z,w)∣2H(z,z)H(w,w).(30)G(z,w)=1-\frac{\lvert H(z,w)\rvert^2}{H(z,z)H(w,w)}. \tag*{(30)}

Since H(z,z)>0H(z,z)>0, we may shrink the disk so that H(z,w)≠0H(z,w)\ne0 for every pair in it. For fixed ww, log⁡∣H(z,w)∣2\log|H(z,w)|^2 is harmonic in zz. Taking logarithms of 1−G(z,w)1-G(z,w) in (30) and then differentiating gives

−∂z∂zˉlog⁡(1−G(z,w))=∂∂ˉlog⁡H(z,z).-\partial_z\partial_{\bar z}\log(1-G(z,w))=\partial\bar\partial\log H(z,z).

To identify the right-hand side, expand the logarithm of (3.14 at w=z+δw=z+\delta. Hermitian symmetry of HH and its holomorphic–antiholomorphic dependence cancel the constant, linear and pure quadratic terms. The mixed term gives

−log⁡(1−G(z,z+δ))=(∂∂ˉlog⁡H(z,z))∣δ∣2+O(∣δ∣3).-\log(1-G(z,z+\delta))=(\partial\bar\partial\log H(z,z))|\delta|^2+O(|\delta|^3).

Since G(z,z+δ)=O(∣δ∣2)G(z,z+\delta)=O(|\delta|^2), the quadratic coefficient of −log⁡(1−G)-\log(1-G) is also that of GG. Dividing (19) by ρ1(z)ρ1(z+δ)\rho_1(z)\rho_1(z+\delta) gives

G(z,z+δ)=(2ρ1(z)+∂∂ˉlog⁡ρ1(z))28ρ1(z)∣δ∣2+O(∣δ∣3).G(z,z+\delta)=\frac{(2\rho_1(z)+\partial\bar\partial\log\rho_1(z))^2}{8\rho_1(z)}|\delta|^2+O(|\delta|^3).

Comparison identifies ∂∂ˉlog⁡H(z,z)\partial\bar\partial\log H(z,z) with p(z)p(z) and proves (29). Its positivity follows from (18).

Step 2. In this step we prove (28) throughout Ω\Omega by first computing the coefficient of z2zˉ2z^2\bar z^2 in

G(0,z)=ρ2(0,z)ρ1(0)ρ1(z),(31)G(0,z)=\frac{\rho_2(0,z)}{\rho_1(0)\rho_1(z)}, \tag*{(31)}

using the Kac–Rice formula (15) for the two-point intensity ρ2\rho_2, and then comparing it with the coefficient of z2zˉ2z^2\bar z^2 obtained from the case w=0w=0 of the differential identity (29).

Fix a point and translate it to zero. Since K(z,0)K(z,0) is nonzero nearby, replace FF by

z⟼K(0,0)F(z)K(z,0).(32)z\longmapsto\frac{\sqrt{K(0,0)}F(z)}{K(z,0)}. \tag*{(32)}

The new covariance satisfies K(z,0)=1K(z,0)=1. If c=∂z∂wˉK(0,0)>0c=\partial_z\partial_{\bar w}K(0,0)>0, use the local coordinate

ζ(z)=c−1/2∂wˉK(z,w)∣w=0.\zeta(z)=c^{-1/2}\left.\partial_{\bar w}K(z,w)\right|_{w=0}.

Its derivative at zero is c>0\sqrt{c}>0, so it has a holomorphic local inverse. The chain rule then gives the following identities for the covariance in the new coordinate, again denoted by zz:

K(z,0)=1,∂wˉK(z,w)∣w=0=z.(33)K(z,0)=1,\qquad\left.\partial_{\bar w}K(z,w)\right|_{w=0}=z. \tag*{(33)}

Taking complex conjugates also gives ∂zK(0,w)=wˉ\partial_zK(0,w)=\bar w. These identities determine every coefficient for which one exponent is zero or one. The expansion through total degree six is therefore

K(z,w)=1+zwˉ+βz2wˉ2+ηz3wˉ2+ηˉz2wˉ3+γz3wˉ3+χz4wˉ2+χˉz2wˉ4+Otot(7),(34)\begin{aligned} K(z,w)={}&1+z\bar w+\beta z^2\bar w^2+\eta z^3\bar w^2+\bar\eta z^2\bar w^3\\ &+\gamma z^3\bar w^3+\chi z^4\bar w^2+\bar\chi z^2\bar w^4+O_{\mathrm{tot}}(7), \tag*{(34)} \end{aligned}

with γ\gamma real and β>0\beta>0. In fact, the constant and linear coefficients of FF are independent standard complex Gaussians, orthogonal to every higher coefficient, and β\beta is the variance of the quadratic coefficient. Its positivity follows from (4). The notation Otot(m)O_{\mathrm{tot}}(m) means terms of total degree at least mm. Thus (34) assures that expanding KK through total degree six is sufficient to compute the coefficients A4A_4 and C2C_2 below.

The replacement (32) preserves the zeros of F(z)F(z), since the multiplier is holomorphic and never zero on the chosen neighborhood. The ratio GG is unchanged by the coordinate change, and the kernel still has the form (24). We may make the comparison in these coordinates. Write t=∣z∣2t = |z|^2 and

K(z,z)=1+tU(z),U(z)=1+βt+ηz2zˉ+ηˉzzˉ2+γt2+χz3zˉ+χˉzzˉ3+Otot(5).(35)\begin{aligned} K(z,z) &= 1+tU(z),\\ U(z) &= 1+\beta t+\eta z^2\bar{z}+\bar{\eta}z\bar{z}^2+\gamma t^2+\chi z^3\bar{z}+\bar{\chi}z\bar{z}^3+O_{\mathrm{tot}}(5). \tag*{(35)} \end{aligned}

For the pair (0,z)(0,z), the three matrices in Kac–Rice are

A=(111K(z,z)),B=(0zˉ0∂K(z,z)),C=(111∂∂ˉK(z,z)).A=\begin{pmatrix}1&1\\1&K(z,z)\end{pmatrix},\qquad B=\begin{pmatrix}0&\bar{z}\\0&\partial K(z,z)\end{pmatrix},\qquad C=\begin{pmatrix}1&1\\1&\partial\bar{\partial}K(z,z)\end{pmatrix}.

Since det⁡A=tU(z)\det A=tU(z), computing the conditional covariance matrix (14) gives

Λ=((U(z)−1)/U(z)−zˉ∂ˉlog⁡U(z)−z∂log⁡U(z)tU(z)∂∂ˉlog⁡U(z)).(36)\Lambda= \begin{pmatrix} (U(z)-1)/U(z)&-\bar{z}\bar{\partial}\log U(z)\\ -z\partial\log U(z)&tU(z)\partial\bar{\partial}\log U(z) \end{pmatrix}. \tag*{(36)}

It follows that

ρ2(0,z)=(U(z)−1)∂∂ˉlog⁡U(z)+∣∂log⁡U(z)∣2U(z).(37)\rho_2(0,z)=\frac{(U(z)-1)\partial\bar{\partial}\log U(z)+|\partial\log U(z)|^2}{U(z)}. \tag*{(37)}

To compute the required coefficients, expand

log⁡U(z)=βt+ηz2zˉ+ηˉzzˉ2+(γ−β2/2)t2+χz3zˉ+χˉzzˉ3+Otot(5).\log U(z)=\beta t+\eta z^2\bar{z}+\bar{\eta}z\bar{z}^2+(\gamma-\beta^2/2)t^2+\chi z^3\bar{z}+\bar{\chi}z\bar{z}^3+O_{\mathrm{tot}}(5).

The coefficients of t2t^2 in (U(z)−1)∂∂ˉlog⁡U(z)(U(z)-1)\partial\bar{\partial}\log U(z) and ∣∂log⁡U(z)∣2|\partial\log U(z)|^2 are, respectively,

5βγ−2β3+4∣η∣2,4βγ−2β3+5∣η∣2.5\beta\gamma-2\beta^3+4|\eta|^2,\qquad4\beta\gamma-2\beta^3+5|\eta|^2.

The leading term in (37) is 2β2t2\beta^2t. Dividing the sum of the two fourth-order coefficients by U(z)=1+βt+O(∣z∣3)U(z)=1+\beta t+O(|z|^3) therefore subtracts a further 2β32\beta^3. Thus the coefficient of t2t^2 in ρ2(0,z)\rho_2(0,z) is 9βγ+9∣η∣2−6β39\beta\gamma+9|\eta|^2-6\beta^3. The Edelman–Kostlan formula applied to K(z,z)=1+tU(z)K(z,z)=1+tU(z) gives

ρ1(z)=1+(4β−2)t+6ηz2zˉ+6ηˉzzˉ2+9(γ−β+1/3)t2+8χz3zˉ+8χˉzzˉ3+Otot(5).(38)\begin{aligned} \rho_1(z)={}&1+(4\beta-2)t+6\eta z^2\bar{z}+6\bar{\eta}z\bar{z}^2\\ &+9(\gamma-\beta+1/3)t^2+8\chi z^3\bar{z}+8\bar{\chi}z\bar{z}^3+O_{\mathrm{tot}}(5). \tag*{(38)} \end{aligned}

In particular, ρ1(0)=1\rho_1(0)=1. This allows us to conclude, from (31) and (37), that

G(0,z)=ρ2(0,z)ρ1(z)=2β2∣z∣2+O(∣z∣3).G(0,z)=\frac{\rho_2(0,z)}{\rho_1(z)}=2\beta^2|z|^2+O(|z|^3).

To find its fourth-order coefficient, divide by ρ1(z)=1+(4β−2)t+O(∣z∣3)\rho_1(z)=1+(4\beta-2)t+O(|z|^3). This subtracts 2β2(4β−2)2\beta^2(4\beta-2) from the coefficient of t2t^2 in ρ2\rho_2. Consequently the coefficient of z2zˉ2z^2\bar{z}^2 in G(0,z)G(0,z) is

A4=9βγ+9∣η∣2+4β2−14β3.(39)A_4=9\beta\gamma+9|\eta|^2+4\beta^2-14\beta^3. \tag*{(39)}

We next compute the coefficient of zzˉz\bar{z} in p(z)p(z). Taking the logarithm in (38) and differentiating gives

2ρ1(z)+∂∂ˉlog⁡ρ1(z)=4β+12ηz+12ηˉzˉ+(36γ+4β−32β2)t+24χz2+24χˉzˉ2+Otot(3).\begin{aligned} 2\rho_1(z)+\partial\bar{\partial}\log\rho_1(z) ={}&4\beta+12\eta z+12\bar{\eta}\bar{z}\\ &+(36\gamma+4\beta-32\beta^2)t+24\chi z^2+24\bar{\chi}\bar{z}^2+O_{\mathrm{tot}}(3). \end{aligned}

The coefficient of tt in the square of this expression is 8β(36γ+4β−32β2)+288∣η∣28\beta(36\gamma+4\beta-32\beta^2)+288|\eta|^2. First divide this coefficient by 8. The factor 1/ρ1(z)1/\rho_1(z) then subtracts 2β2(4β−2)2\beta^2(4\beta-2). Therefore the coefficient of zzˉz\bar z in p(z)p(z) is

C2=36βγ+36∣η∣2+8β2−40β3.(40)C_2=36\beta\gamma+36|\eta|^2+8\beta^2-40\beta^3. \tag*{(40)}

Since GG is symmetric, (3.13) with w=0w=0 applies to G(0,z)G(0,z) as well. Now

−log⁡(1−G(0,z))=G(0,z)+12G(0,z)2+O(∣z∣6).-\log(1-G(0,z))=G(0,z)+\frac{1}{2}G(0,z)^2+O(|z|^6).

Its coefficient of z2zˉ2z^2\bar z^2 is A4+2β4A_4+2\beta^4. Applying ∂∂ˉ\partial\bar\partial multiplies this coefficient by four, so (3.13) implies

C2=4(A4+2β4).C_2=4(A_4+2\beta^4).

Substituting (3.23)–(3.24) yields

0=4A4+8β4−C2=8β2(β−1)2.(41)0=4A_4+8\beta^4-C_2=8\beta^2(\beta-1)^2. \tag*{(41)}

Nondegeneracy gives β>0\beta>0, so β=1\beta=1. The terms containing γ\gamma and ∣η∣2|\eta|^2 cancel. The terms containing χ\chi do not enter the indicated coefficients.

At the origin, (3.22) gives (∂∂ˉlog⁡ρ1)(0)=4β2(\partial\bar\partial\log\rho_1)(0)=4\beta^2 and ρ1(0)=1\rho_1(0)=1. It follows that (∂∂ˉlog⁡ρ1)(0)=2ρ1(0)(\partial\bar\partial\log\rho_1)(0)=2\rho_1(0). The point was arbitrary. Under a holomorphic change of variable ψ\psi,

ρ1,ψ(z)=∣ψ′(z)∣2ρ1(ψ(z)),\rho_{1,\psi}(z)=|\psi'(z)|^2\rho_1(\psi(z)),
∂∂ˉlog⁡ρ1,ψ(z)=∣ψ′(z)∣2(∂∂ˉlog⁡ρ1)(ψ(z)),\partial\bar\partial\log\rho_{1,\psi}(z)=|\psi'(z)|^2(\partial\bar\partial\log\rho_1)(\psi(z)),

because log⁡∣ψ′(z)∣2\log|\psi'(z)|^2 is harmonic. The equality therefore holds in the original coordinate, proving (3.12) throughout Ω\Omega.

Proof of Theorem 1.1

The converse follows from Peres–Virág [14]. Indeed, a GAF with covariance (1.5) has the same law as h(FD∘ϕ)h(F_{\mathbb D}\circ\phi). Multiplication by hh does not change its zeros. Restricting the Peres–Virág process to ϕ(Ω)\phi(\Omega) and changing variables by ϕ\phi therefore gives the DPP kernel (1.6). To verify (1.4), consider the independent Gaussian coefficients multiplying h(z)h(z), h(z)ϕ(z)h(z)\phi(z), h(z)ϕ2(z)h(z)\phi^2(z). At every point zz their values and first two derivatives form a matrix with determinant

det⁡[(hϕj)(i)(z)]i,j=02=2h(z)3ϕ′(z)3≠0.\det\left[(h\phi^j)^{(i)}(z)\right]_{i,j=0}^{2}=2h(z)^3\phi'(z)^3\ne0.

These three independent coefficients give a positive definite contribution to the covariance matrix in (1.4). The remaining coefficients add a positive semidefinite matrix, so the full covariance is positive definite.

We now prove necessity. Assume that FF belongs to the Peres–Virág class. By Lemma 3.3, its first intensity satisfies (3.12). We first obtain the global covariance representation and then prove that its map ϕ\phi is one-to-one.

Step 1. We first obtain the covariance kernel on a sufficiently small disk U⊆ΩU\subseteq\Omega. Liouville’s theorem, Lemma 1.2, applied with g=ρ1g=\rho_1 to (3.12), gives a holomorphic map ϕ0:U→D\phi_0:U\to\mathbb{D} with ϕ0′≠0\phi_0'\ne0 and

ρ1(z)=∣ϕ0′(z)∣2(1−∣ϕ0(z)∣2)2.\rho_1(z)=\frac{|\phi_0'(z)|^2}{(1-|\phi_0(z)|^2)^2}.

The right-hand side is the first zero intensity of FD∘ϕ0F_D \circ\phi_0, by the Edelman–Kostlan formula. Thus this GAF and F∣UF|_U have the same first zero intensity. Sodin’s Calabi rigidity theorem [15 Theorem 2], in the form stated in [6 Theorem 2.5.2, pp. 30–31], gives a zero-free holomorphic function h0h_0 on UU such that

K(z,w)=h0(z)h0(w)‾1−ϕ0(z)ϕ0(w)‾,z,w∈U.(42)K(z,w)=\frac{h_0(z)\overline{h_0(w)}}{1-\phi_0(z)\overline{\phi_0(w)}}, \qquad z,w\in U. \tag*{(42)}

It remains to extend this representation to the whole domain Ω\Omega. Fix z0∈Uz_0\in U. By Möbius invariance of the disk, we may assume ϕ0(z0)=0\phi_0(z_0)=0, after a corresponding adjustment of h0h_0 in (42). Multiplying h0h_0 by a constant of modulus one, we may also assume

h0(z0)=K(z0,z0)>0.h_0(z_0)=\sqrt{K(z_0,z_0)}>0.

Evaluating (42) at w=z0w=z_0 now expresses h0(z)h_0(z) as a constant multiple of the globally defined holomorphic function K(z,z0)K(z,z_0). Hence h0h_0 extends holomorphically to Ω\Omega. Now differentiate (42) in w‾\overline{w} and set w=z0w=z_0. Since ϕ0(z0)=0\phi_0(z_0)=0, this identity expresses h0(z)ϕ0(z)h_0(z)\phi_0(z) as a constant linear combination of K(z,z0)K(z,z_0) and ∂w‾K(z,w)∣w=z0\left.\partial_{\overline{w}}K(z,w)\right|_{w=z_0}. The coefficient by which we divide is h0(z0)ϕ0′(z0)‾≠0h_0(z_0)\overline{\phi_0'(z_0)}\ne0. Thus h0ϕ0h_0\phi_0 also extends holomorphically to Ω\Omega. Denote the two extensions by hh and jj, respectively. The diagonal of (42) gives, on UU, the identity

K(z,z)(∣h(z)∣2−∣j(z)∣2)=∣h(z)∣4.(43)K(z,z)\left(|h(z)|^2-|j(z)|^2\right)=|h(z)|^4. \tag*{(43)}

Both sides are real-analytic on the connected domain Ω\Omega, including points where hh vanishes. Since their difference vanishes on UU, real-analytic continuation proves (43) throughout Ω\Omega. Since h(z0)>0h(z_0)>0, the zeros of hh are isolated. At a point where h(z)≠0h(z)\ne0, equation (43) and K(z,z)>0K(z,z)>0 imply

∣j(z)h(z)∣<1.\left|\frac{j(z)}{h(z)}\right|<1.

The quotient j(z)/h(z)j(z)/h(z) is therefore bounded near every zero of h(z)h(z). The removable singularity theorem extends it to a holomorphic function ϕ(z)\phi(z) on Ω\Omega, which is nonconstant on UU, where ϕ(z)=ϕ0(z)\phi(z)=\phi_0(z) and ϕ′(z)≠0\phi'(z)\ne0. By continuity ∣ϕ(z)∣≤1|\phi(z)|\le1 everywhere and, by the maximum modulus principle, ∣ϕ(z)∣<1|\phi(z)|<1 throughout Ω\Omega. On {z:h(z)≠0}\{z:h(z)\ne0\}, equation (43) can now be written as

∣h(z)∣2=K(z,z)(1−∣ϕ(z)∣2).|h(z)|^2=K(z,z)(1-|\phi(z)|^2).

Both sides are continuous, so this equality extends across the isolated zeros of hh. Its right-hand side is strictly positive. Consequently hh has no zeros. Taking logarithms gives

log⁡K(z,z)=log⁡∣h(z)∣2−log⁡(1−∣ϕ(z)∣2).\log K(z,z)=\log|h(z)|^2-\log(1-|\phi(z)|^2).

The first term is harmonic because hh is holomorphic and zero-free. Applying the Edelman–Kostlan formula therefore gives

ρ1(z)=∣ϕ′(z)∣2(1−∣ϕ(z)∣2)2.\rho_1(z)=\frac{|\phi'(z)|^2}{(1-|\phi(z)|^2)^2}.

Since ρ1(z)>0\rho_1(z)>0, the derivative ϕ′(z)\phi'(z) never vanishes. The kernel K(z,w)K(z,w) and the right-hand side of (42), with the extensions h,ϕh,\phi, are now holomorphic in zz and antiholomorphic in ww on Ω×Ω\Omega\times\Omega. For each fixed w∈Uw\in U, the identity theorem in zz extends their equality from UU to Ω\Omega. For each fixed z∈Ωz\in\Omega, the same theorem in the second variable then extends the equality from UU to Ω\Omega. Thus (42) holds on the whole of Ω×Ω\Omega\times\Omega, proving (1.5).

Step 2. We show that the map ϕ\phi obtained in Step 1 is one-to-one, arguing by contradiction. Suppose ϕ(z1)=ϕ(z2)\phi(z_1)=\phi(z_2) with z1≠z2z_1\ne z_2. Since ϕ′(z)\phi'(z) never vanishes, the inverse function theorem gives disjoint neighborhoods on which the two inverse branches are defined. Restrict their images to the same sufficiently small disk. We obtain disjoint relatively compact neighborhoods U,VU,V such that ϕ∣U\phi|_U and ϕ∣V\phi|_V are one-to-one and have that disk as their common image. Set

T=(ϕ∣V)−1∘(ϕ∣U):U⟶V.T=(\phi|_V)^{-1}\circ(\phi|_U):U\longrightarrow V.

Then TT is a holomorphic bijection T:U→VT:U\to V and ϕ(Tz)=ϕ(z)\phi(Tz)=\phi(z). By (5), for every fixed z∈Uz\in U,

F(z)h(z)=F(Tz)h(Tz)\frac{F(z)}{h(z)}=\frac{F(Tz)}{h(Tz)}

almost surely.

Indeed, the covariance identity (5) gives, for z∈Uz\in U,

E∣F(z)h(z)−F(Tz)h(Tz)∣2=21−∣ϕ(z)∣2−2Re⁡11−ϕ(z)ϕ(Tz)=0.\mathbb{E}\left|\frac{F(z)}{h(z)}-\frac{F(Tz)}{h(Tz)}\right|^2=\frac{2}{1-|\phi(z)|^2}-2\operatorname{Re}\frac{1}{1-\phi(z)\phi(Tz)}=0.

Apply the almost sure equality on a countable dense subset of UU and intersect the resulting events of probability one. By continuity, the equality holds simultaneously for every z∈Uz\in U on the intersection event. Since h(z)h(z) never vanishes and T=(ϕ∣V)−1∘(ϕ∣U)T=(\phi|_V)^{-1}\circ(\phi|_U) is a biholomorphic map from UU onto VV, the zeros in UU and VV correspond bijectively. Let NF(E)N_F(E) denote the number of zeros in EE. Consequently,

NF(U)=NF(V)(44)N_F(U)=N_F(V) \tag*{(44)}

almost surely.

By shrinking the common image disk and the corresponding neighborhoods U,VU,V, we may assume that

m:=E[NF(U)]=E[NF(V)]=∫Uρ1(z) dμ(z),0<m<1.m:=\mathbb{E}[N_F(U)]=\mathbb{E}[N_F(V)]=\int_U\rho_1(z)\,d\mu(z),\qquad0<m<1.

Indeed, ρ1\rho_1 is continuous and strictly positive, and the neighborhoods shrink to z1,z2z_1,z_2 as the radius of the common image disk tends to zero. Since UU and VV are disjoint, integration of the (global repulsive) two-point determinantal identity (11), gives

E[NF(U)NF(V)]=∫U×Vρ2(z,w) dμ(z) dμ(w)≤(∫Uρ1(z) dμ(z))(∫Vρ1(w) dμ(w))=m2.\begin{aligned} \mathbb{E}[N_F(U)N_F(V)]=\int_{U\times V}\rho_2(z,w)\,d\mu(z)\,d\mu(w) \\ &\le\left(\int_U\rho_1(z)\,d\mu(z)\right)\left(\int_V\rho_1(w)\,d\mu(w)\right)=m^2. \end{aligned}

On the other hand, since NF(U)N_F(U) is a nonnegative integer, NF(U)2≥NF(U)N_F(U)^2\ge N_F(U). Together with the almost sure equality (44), this gives

E[NF(U)NF(V)]=E[NF(U)2]≥E[NF(U)]=m>m2.\mathbb{E}[N_F(U)N_F(V)]=\mathbb{E}[N_F(U)^2]\ge\mathbb{E}[N_F(U)]=m>m^2.

This contradiction proves that ϕ\phi is one-to-one and completes the proof of Theorem 1.1.

Examples and comments

Restrictions and changes of variable

Every one-to-one holomorphic map ϕ:Ω→D\phi:\Omega\to\mathbb{D} gives an example through FD∘ϕF_{\mathbb{D}}\circ\phi. Taking ϕ(z)=z\phi(z)=z on a strict subdomain of the disk shows why ϕ(Ω)=D\phi(\Omega)=\mathbb{D} is not required. In particular, we may restrict the disk process to an annulus. The domain in the theorem need not be simply connected.

A function with one random zero

Let F(z)=ξ0+ξ1zF(z)=\xi_0+\xi_1z on C\mathbb{C}. The zeros of F(z)F(z) are determinantal, with kernel

K(z,w)=1(1+∣z∣2)(1+∣w∣2).\mathcal{K}(z,w)=\frac{1}{(1+|z|^2)(1+|w|^2)}.

The corresponding operator projects onto the one-dimensional space spanned by (1+∣z∣2)−1(1+|z|^2)^{-1}. Here K(z,w)=1+zw‾K(z,w)=1+z\overline{w} and ∂∂‾log⁡ρ1(z)=−2ρ1(z)\partial\overline{\partial}\log\rho_1(z)=-2\rho_1(z). The two-point intensity vanishes identically. This example is outside the Peres–Virág class, since its covariance matrix is

Cov⁡(F(z),F′(z),F′′(z))=(1+∣z∣2z‾0z10000).\operatorname{Cov}(F(z),F'(z),F''(z))= \begin{pmatrix} 1+|z|^2 & \overline{z} & 0\\ z & 1 & 0\\ 0 & 0 & 0 \end{pmatrix}.

This matrix has rank two and determinant zero at every point. It is therefore not positive definite, so (4) fails.

Perturbing the constant coefficient of the Hardy GAF

Consider the disk limit of the weighted Szegő family studied by Katori and Shirai [7],

Fr(z)=ξ01+r+∑n=1∞ξnzn,r>0.(45)F_r(z)=\frac{\xi_0}{\sqrt{1+r}}+\sum_{n=1}^{\infty}\xi_nz^n,\qquad r>0. \tag*{(45)}

One can directly check that the covariance of (Fr,Fr′,Fr′′)(F_r,F'_r,F''_r) is positive definite. At the origin,

(∂∂‾log⁡ρ1,r)(0)−2ρ1,r(0)=−4r≠0.(\partial\overline{\partial}\log\rho_{1,r})(0)-2\rho_{1,r}(0)=-4r\ne0.

Thus ρ1,r\rho_{1,r} does not satisfy the necessary equation (28) established in Lemma 3.3. Its zero process is therefore not a Hermitian DPP for any r>0r>0. [7 Proposition 1.7] gives a nonempty interval of positive rr for which zeros are negatively correlated. This shows that perturbing the constant coefficient can preserve repulsion, but destroys the Hermitian determinantal property.

Constant curvature without global injectivity

We give an example satisfying (4) whose zeros are determinantal on sufficiently small disks, but fail to be determinantal globally. Let Ω={z∈C:ℜz<0}\Omega=\{z\in\mathbb{C}:\Re z<0\}, ϕ(z)=ez\phi(z)=e^z and define

F(z)=FD(ez)=∑n=0∞ξnenz,z∈Ω.(46)F(z)=F_{\mathbb{D}}(e^z)=\sum_{n=0}^{\infty}\xi_ne^{nz},\qquad z\in\Omega. \tag*{(46)}

Since ∣ez∣<1|e^z|<1 on Ω\Omega, this is a centered circular GAF with covariance

K(z,w)=11−ez+w‾,z,w∈Ω.(47)K(z,w)=\frac{1}{1-e^{z+\overline{w}}},\qquad z,w\in\Omega. \tag*{(47)}

The map ϕ\phi is locally one-to-one, since ϕ′(z)=ez≠0\phi'(z)=e^z\ne0. The covariance of (F(z),F′(z),F′′(z))(F(z),F'(z),F''(z)) is positive definite at every point. Writing x=ℜzx=\Re z, direct differentiation of the Edelman–Kostlan formula applied to (47) shows that

∂∂‾log⁡ρ1(z)=2ρ1(z).(48)\partial\overline{\partial}\log\rho_1(z)=2\rho_1(z). \tag*{(48)}

Thus the metric ds2=ρ1(z)∣dz∣2ds^2 = \rho_1(z)|dz|^2 has Gaussian curvature −4-4 throughout Ω\Omega. On every sufficiently small disk U⊂ΩU \subset\Omega, the exponential map is one-to-one. Restricting the Peres–Virág process to eUe^U and changing variables by ϕ\phi therefore shows that the zeros of F∣UF|_U form a DPP, by [14 Theorem 1]. Its correlation kernel relative to dμd\mu is

KU(z,w)=ez+w‾(1−ez+w‾)2,z,w∈U.K_U(z,w) = \frac{e^{z+\overline{w}}}{(1-e^{z+\overline{w}})^2}, \qquad z,w \in U.

However, on the whole domain, ϕ(z+2πi)=ϕ(z)\phi(z+2\pi i)=\phi(z), so ϕ\phi is not globally one-to-one. Correspondingly, (46) gives

F(z+2πi)=F(z),z∈Ω.(49)F(z+2\pi i)=F(z), \qquad z \in\Omega. \tag*{(49)}

By the argument of Step 2 in the proof of Theorem 1.1, these forced pairs of zeros prevent a Hermitian determinantal representation on Ω\Omega.

Acknowledgement

The author is thankful to OpenAI for the award of an ‘Academic Researchers Account’, providing free access to models aligned with scientific research. We share information of potential methodological interest for mathematical research. A substantial use of the AI tool was made during the experimental part of this work, where a large database of kernels was tested, initially with the goal of finding a new GAF with determinantal zeros. After several frustrated attempts, ‘the curvature hypothesis’ was tested on the same database, which included constant curvature examples such as Bergman kernels, as well as exotic qq-analogues. The experimental results were then organized in a knowledge graph, which was subsequently used in the theoretical investigation, to retrieve connected examples, verify computations and develop some technical arguments (which were then rederived and simplified by the author), and also in the final revisions, to enhance clarity of exposition. This research was funded by the Austrian Science Fund (FWF) through the project 10.55776/PAT8205923.

Faculty of Mathematics, University of Vienna, Oskar-Morgenstern-Platz 1, A-1090, Vienna, Austria

E-mail address: [email protected]

References

  1. [1]A. F. Beardon and D. Minda, The hyperbolic metric and geometric function theory, in Quasiconformal Mappings and Their Applications, S. Ponnusamy, T. Sugawa and M. Vuorinen (eds.), Narosa, New Delhi, 2007, pp. 9–56.
  2. [2]A. Borodin, Determinantal point processes, in The Oxford Handbook of Random Matrix Theory, G. Akemann, J. Baik and P. Di Francesco (eds.), Oxford University Press, Oxford, 2011, pp. 231–249; arXiv:0911.1153.arxiv.org/abs/0911.1153
  3. [3]A. Edelman and E. Kostlan, How many zeros of a random polynomial are real?, Bull. Amer. Math. Soc. (N.S.) 32 (1995), no. 1, 1–37.arxiv.org/abs/math/9501224
  4. [4]P. J. Forrester, The limiting Kac random polynomial and truncated random orthogonal matrices, J. Stat. Mech. Theory Exp. (2010), no. 12, P12018.DOI
  5. [5]H. Hedenmalm, B. Korenblum and K. Zhu, Theory of Bergman Spaces, Graduate Texts in Mathematics 199, Springer, New York, 2000.DOI
  6. [6]J. B. Hough, M. Krishnapur, Y. Peres and B. Virág, Zeros of Gaussian Analytic Functions and Determinantal Point Processes, University Lecture Series 51, American Mathematical Society, Providence, RI, 2009.DOI
  7. [7]M. Katori and T. Shirai, Zeros of the i.i.d. Gaussian Laurent series on an annulus: weighted Szegő kernels and permanental-determinantal point processes, Commun. Math. Phys. 392 (2022), 1099–1151.DOI
  8. [8]D. Kraus and O. Roth, Critical points, the Gauss curvature equation and Blaschke products, in Blaschke Products and Their Applications, J. Mashreghi and E. Fricain (eds.), Fields Institute Communications 65, Springer, New York, 2013, pp. 133–157.arxiv.org/abs/1109.4711
  9. [9]M. Krishnapur, From random matrices to random analytic functions, Ann. Probab. 37 (2009), no. 1, 314–346.arxiv.org/abs/0711.1378
  10. [10]S. Matsumoto and T. Shirai, Correlation functions for zeros of a Gaussian power series and Pfaffians, Electron. J. Probab. 18 (2013), no. 49, 1–18.DOI
  11. [11]M. Michelen and O. Yakir, Fluctuations in the logarithmic energy for zeros of random polynomials on the sphere, Probab. Theory Related Fields 191 (2025), 569–626.arxiv.org/abs/2304.02898
  12. [12]F. Nazarov and M. Sodin, Correlation functions for random complex zeroes: strong clustering and local universality, Commun. Math. Phys. 310 (2012), 75–98.arxiv.org/abs/1005.4113
  13. [13]K.-H. Neeb, On differentiable vectors for representations of infinite dimensional Lie groups, J. Funct. Anal. 259 (2010), 2814–2855.arxiv.org/abs/1002.1602
  14. [14]Y. Peres and B. Virág, Zeros of the i.i.d. Gaussian power series: a conformally invariant determinantal process, Acta Math. 194 (2005), 1–35.
  15. [15]M. Sodin, Zeros of Gaussian analytic functions, Math. Res. Lett. 7 (2000), 371–381.
  16. [16]A. Soshnikov, Determinantal random point fields, Russian Math. Surveys 55 (2000), no. 5, 923–975; arXiv:math/0002099.DOI

Paper details

vlnc:W_p5t3v7z5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

the AI tool

View statement
Looking for related papers…
Contents