betaFeedback

WorksMathematics

On some stochastic evolutions on Laguerre-Polya functions

Abstract

We prove uniqueness in law for three stochastic evolutions on normalised Laguerre-Pólya functions arising as scaling limits of matrix characteristic polynomial dynamics. In particular, this implies that Ahn's universal Brownian Gibbs line ensemble arising from random matrix products has the law of the logarithms of the reciprocal zeros of the solution, unique in law, to a simple stochastic equation on entire functions started from e−z\mathrm{e}^{-z}. The method of proof relies on some integrable structure of the problem in the special form of a new duality with finite-dimensional diffusions which is closely related to intertwinings.

Introduction

Setting and motivation

Certain infinite-dimensional particle dynamics arise naturally as microscopic limits of eigenvalue and singular-value evolutions of random matrices. Their description by infinite systems of stochastic differential equations (ISDEs) with logarithmic interaction raises basic questions of existence and uniqueness given the choice of initial state. Important approaches to these dynamics include the determinantal constructions of Katori–Tanemura [20], the Dirichlet-form and ISDE theory of Osada and Osada–Tanemura [25, 26], and Tsai’s construction of strong solutions to the infinite Dyson system [32]; see also the authors’ paper [8].

In recent work [6], the infinite-particle limit of the so-called Rider–Valkó dynamics from [29] was constructed and shown that its logarithmic-interaction ISDE,

dxi(t)=xi(t) dBi(t)−ν2xi(t) dt+∑j=1j≠i∞xi(t)xj(t)xi(t)−xj(t) dt,i∈N,(1)\mathrm{d}x_i(t)=x_i(t)\,\mathrm{d}B_i(t)-\frac{\nu}{2}x_i(t)\,\mathrm{d}t+\sum_{\substack{j=1 \\ j\ne i}}^{\infty}\frac{x_i(t)x_j(t)}{x_i(t)-x_j(t)}\,\mathrm{d}t,\qquad i\in\mathbb{N}, \tag*{(1)}

with (Bi)i≥1(B_i)_{i\ge1} independent standard Brownian motions and x1(t)>x2(t)>⋯>0x_1(t)>x_2(t)>\cdots>0, admits distinct (weak) solution laws from the same initial particle configuration. This provided the first concrete examples of such non-uniqueness for an ISDE with logarithmic interaction. To retain additional information that one may hope to need for a characterisation, the full evolution of the limiting reverse characteristic polynomial was subsequently considered in [7], also for two other models: the limiting singular values of multiplicative Brownian motion on the general linear group GLN(C)\mathcal{G L}_N(\mathbb{C}) [23, 18, 1] and the dynamical Cauchy or Hua–Pickrell model [4, 3, 2]. The resulting state spaces consist of (subsets of) normalised Laguerre–Pólya functions, which arise as locally uniform limits of polynomials with real zeros and have the canonical form

E(z)=e−γz−qz2/2∏j=1∞eyjz(1−yjz),E(z)=e^{-\gamma z-qz^2/2}\prod_{j=1}^{\infty}e^{y_jz}(1-y_jz),
γ∈R,q≥0,yj∈R,∑j=1∞yj2<∞.\gamma\in\mathbb{R},\quad q\geq0,\quad y_j\in\mathbb{R},\quad\sum_{j=1}^{\infty}y_j^2<\infty.

The exponential factors record information that is invisible to the zero set alone. This enlarged description distinguishes evolutions whose particle coordinates may solve the same ISDE.

Here, as predicted in [7], we prove that the three entire-function valued equations of [7] do characterise their solutions in law. Such an intrinsic characterisation gives a way to identify subsequential limits and is therefore useful for convergence and universality problems. As a corollary, we obtain a stochastic analytic characterisation of Ahn’s Brownian Gibbs line ensemble [1]: its curves are the logarithms of the reciprocal zeros of a solution started from e−ze^{-z}. This ensemble is expected to describe a broad universality class for products of many large random matrices: in particular Ahn [1] proves convergence of the leading curves in finite-dimensional distributions for a class of products of independent right-unitarily invariant matrices, after centering and a time change, when the number of factors is comparable to the matrix size.

The proof here relies in an essential way on the integrable structure of the problem in terms of a new duality with finite-dimensional diffusions. Based on this duality the arguments are not very complicated and rather elegant. Nevertheless, it would be desirable to have a more robust/less integrable approach to problems of this kind. Finally, a different way to restore uniqueness to the ISDEs has been suggested in [6, 7] by considering so-called "rigid-path" solutions, see [6, 7] for a more detailed discussion. Some preliminary progress with this viewpoint in mind is made in [22]. It would be interesting to understand this other type of characterisation.

Main results

We now move to state our main results precisely. Two of the stochastic evolutions live in the following positive class of Laguerre-Pólya functions

E+={e−dz∏j=1∞(1−xjz):d≥0, x1≥x2≥⋯≥0, ∑j=1∞xj<∞}.(2)\mathcal{E}_{+}=\left\{e^{-dz}\prod_{j=1}^{\infty}(1-x_jz):d\geq0,\ x_1\geq x_2\geq\cdots\geq0,\ \sum_{j=1}^{\infty}x_j<\infty\right\}. \tag*{(2)}

They are the edge-limit equation associated with Brownian motion on GLN(C)\mathcal{G}\mathcal{L}_N(\mathbb{C}) in [7] and the limiting dynamics of the Rider–Valkó equation from the matrix Dufresne identity framework [29]. The third evolves on the full normalised two-sided class of Laguerre-Pólya functions

E={e−γz−qz2/2∏j=1∞eyjz(1−yjz):γ∈R, q≥0, y=(yj)∈ℓ2(N;R)}.(3)\mathcal{E}=\left\{e^{-\gamma z-qz^2/2}\prod_{j=1}^{\infty}e^{y_jz}(1-y_jz):\gamma\in\mathbb{R},\ q\geq0,\ y=(y_j)\in\ell^2(\mathbb{N};\mathbb{R})\right\}. \tag*{(3)}

where ℓ2(N;R)\ell^2(\mathbb{N};\mathbb{R}) is the space of square-summable real sequences, and the nonzero yjy_j’s are put in the signed canonical order fixed below. Both spaces carry the topology of locally uniform convergence. For g∈Eg \in\mathcal{E}, put γ(g)=−g′(0)=γ\gamma(g)=-g'(0)=\gamma and δ(g)=q+∑j=1∞yj2\delta(g)=q+\sum_{j=1}^{\infty}y_j^2; for g∈E+g\in\mathcal{E}_+, put γ(g)=−g′(0)=d+∑j=1∞xj\gamma(g)=-g'(0)=d+\sum_{j=1}^{\infty}x_j.

For every normalised entire function gg, define the analytic "covariance" kernel

Qg(z,w)=zwz−w(g(z)g′(w)−g(w)g′(z)),QgHP=2Qg,(4)Q_g(z,w)=\frac{zw}{z-w}\left(g(z)g'(w)-g(w)g'(z)\right),\qquad Q_g^{\mathrm{HP}}=2Q_g, \tag*{(4)}

with the diagonal values given by analytic continuation. The positive equations have drift

(Lθ,κg)(z)=θ2zg′(z)−12z2g′′(z)−κ2zg(z),θ∈R,κ∈{0,1}.(5)(\mathcal{L}_{\theta,\kappa}g)(z)=\frac{\theta}{2}zg'(z)-\frac{1}{2}z^2g''(z)-\frac{\kappa}{2}zg(z),\qquad\theta\in\mathbb{R},\quad\kappa\in\{0,1\}. \tag*{(5)}

The LLN(C)\mathcal{L}\mathcal{L}_{\mathbb{N}}(\mathbb{C}) edge-limit equation has κ=0\kappa=0; the Rider–Valkó equation with parameter ν\nu has (θ,κ)=(−ν,1)(\theta,\kappa)=(-\nu,1). For the Hua–Pickrell parameter ξ=a+ib\xi=a+ib, the drift is given by,

(La,bHPg)(z)=−(2bz+z2)g(z)−2azg′(z)−z2g′′(z).(6)(\mathcal{L}_{a,b}^{\mathrm{HP}}g)(z)=-(2bz+z^2)g(z)-2azg'(z)-z^2g''(z). \tag*{(6)}

Definition 1.1. Set (L+,Q+)=(Lθ,κ,Q)(\mathcal{L}^+,\mathcal{Q}^+)=(\mathcal{L}_{\theta,\kappa},\mathcal{Q}) and LHP=La,bHP\mathcal{L}^{\mathrm{HP}}=\mathcal{L}_{a,b}^{\mathrm{HP}}. For ⋄∈{+,HP}\diamond\in\{+,\mathrm{HP}\}, a weak solution Φ\Phi is defined on a filtered probability space

(Ω,F,(Ft)t≥0,P)(\Omega,\mathcal{F},(\mathcal{F}_t)_{t\geq0},\mathbb{P})

satisfying the usual conditions, and satisfies

MtΦ(z)=Φt(z)−Φ0(z)−∫0t(L⋄Φr)(z) dr,z∈C,(7)M_t^\Phi(z)=\Phi_t(z)-\Phi_0(z)-\int_0^t(\mathcal{L}^{\diamond}\Phi_r)(z)\,\mathrm{d}r,\qquad z\in\mathbb{C}, \tag*{(7)}

where the processes MΦ(z)M^\Phi(z) are continuous local martingales on one filtration and have complex-bilinear brackets

[MΦ(z),MΦ(w)]t=∫0tQΦr⋄(z,w) dr.(8)[M^\Phi(z),M^\Phi(w)]_t=\int_0^tQ_{\Phi_r}^{\diamond}(z,w)\,\mathrm{d}r. \tag*{(8)}

The above martingale formulation can be written in equivalent form as an evolution equation with respect to a cylindrical Brownian motion. More precisely, on a possibly enlarged probability space, (7), (8) it can be written as

dΦt=L⋄Φtdt+Σ⋄(Φt)dWt,\mathrm{d}\Phi_t=\mathcal{L}_{\diamond}\Phi_t\mathrm{d}t+\Sigma_{\diamond}(\Phi_t)\mathrm{d}W_t,

where WW is a real cylindrical Brownian motion on ℓ2(N0;R)\ell^2(\mathbb{N}_0;\mathbb{R}), with independent standard real coordinate Brownian motions (Bj)j≥0(B_j)_{j\geq0}. Writing

Φt(z)=e−dtz∏j=1∞(1−xj(t)z),if Φt∈E+,\Phi_t(z)=e^{-d_tz}\prod_{j=1}^{\infty}(1-x_j(t)z),\qquad\text{if }\Phi_t\in\mathcal{E}_+,
Φt(z)=e−γtz−qtz2/2∏j=1∞eyj(t)z1−yj(t)zif Φt∈E,\Phi_t(z)=e^{-\gamma_tz-q_tz^2/2}\prod_{j=1}^{\infty}\frac{e^{y_j(t)z}}{1-y_j(t)z}\qquad\text{if }\Phi_t\in\mathcal{E},

respectively, the noise coefficient Σ⋄\Sigma_{\diamond} is specified through evaluations by

(Σ+(Φt)dWt)(z)=−∑j≥1zxj(t)Φt(z)1−xj(t)z dBj(t),(\Sigma_+(\Phi_t)\mathrm{d}W_t)(z)=-\sum_{j\geq1}\frac{zx_j(t)\Phi_t(z)}{1-x_j(t)z}\,\mathrm{d}B_j(t),
(ΣHP(Φt)dWt)(z)=−2qt zΦt(z) dB0(t)−2∑j≥1zyj(t)Φt(z)1−yj(t)z dBj(t).(\Sigma_{\mathrm{HP}}(\Phi_t)\mathrm{d}W_t)(z)=-\sqrt{2q_t}\,z\Phi_t(z)\,\mathrm{d}B_0(t)-\sqrt{2}\sum_{j\geq1}\frac{zy_j(t)\Phi_t(z)}{1-y_j(t)z}\,\mathrm{d}B_j(t).

Note that, the xj(t)x_j(t) and yj(t)y_j(t) are the canonical positive and signed reciprocal-root coordinates of Φt\Phi_t, respectively, and γt=γ(Φt)\gamma_t=\gamma(\Phi_t) and qt=q(Φt)q_t=q(\Phi_t). The apparent poles are removable, and the series are interpreted as localised stochastic integrals. For details on cylindrical stochastic integration and Brownian representation, see [17 Section 4.4] and [24 Theorem 2]. This formulation is closer to the stochastic partial differential equations literature and would be interesting to develop further, but we will not need this here.

The so-called boundary processes of [4, 6] are constructed by the method of intertwiners, which builds a Markov semigroup on a projective boundary from finite-dimensional semigroups consistent under Markov kernels; see [11, 10, 12]. Their entire-function limits in [7 Theorems 1.4 and 1.6] provide weak solutions Φt\Phi_t to the martingale problems we defined. Our main theorem then identifies their laws uniquely from the closed martingale problems above.

Theorem 1.2. For every θ∈R\theta\in\mathbb{R}, κ∈{0,1}\kappa\in\{0,1\}, and deterministic g∈E+g \in\mathcal{E}_+, the martingale problem (7)–(8) with L⋄=Lθ,κ\mathcal{L}^{\diamond}=\mathcal{L}_{\theta,\kappa} and Q⋄=Q\mathcal{Q}^{\diamond}=\mathcal{Q} has a unique in law weak solution in C(R+;E+)C(\mathbb{R}_+;\mathcal{E}_+). The analogous problem has a unique in law weak solution on C(R+;E)C(\mathbb{R}_+;\mathcal{E}), for every a,b∈Ra,b \in\mathbb{R} and deterministic g∈Eg \in\mathcal{E}, with L⋄=La,bHP\mathcal{L}^{\diamond}=\mathcal{L}^{\mathrm{HP}}_{a,b} and Q⋄=QHP\mathcal{Q}^{\diamond}=\mathcal{Q}^{\mathrm{HP}}.

The uniqueness theorem above also characterises Ahn’s GLN(C)\mathcal{G}\mathcal{L}_N(\mathbb{C}) Brownian motion singular-value limit through the zeros of one stochastic entire function. We first fix its scaling conventions.

Definition 1.3. Let YNY_N be Brownian motion on GLN(C)\mathcal{G}\mathcal{L}_N(\mathbb{C}) given by the Itô equation

dYN(t)=YN(t) dBN(t),YN(0)=IN,\mathrm{d}Y_N(t)=Y_N(t)\,\mathrm{d}B_N(t), \qquad Y_N(0)=I_N,

where BN=BN(1)+iBN(2)B_N=B_N^{(1)}+iB_N^{(2)} and all entries of the real matrices BN(1),BN(2)B_N^{(1)},B_N^{(2)} are independent standard Brownian motions. If σ1(N)(t)≥⋯≥σN(N)(t)>0\sigma_1^{(N)}(t)\geq\cdots\geq\sigma_N^{(N)}(t)>0 are its singular values, put

ξi(N)(t)=2log⁡σi(N)(t4)−Nt2−log⁡N,1≤i≤N.(9)\xi_i^{(N)}(t)=2\log\sigma_i^{(N)}\left(\frac{t}{4}\right)-\frac{Nt}{2}-\log N,\qquad1\leq i\leq N. \tag*{(9)}

Ahn’s line ensemble ξ=(ξi(t))i∈N,t>0\xi=(\xi_i(t))_{i\in\mathbb{N},t>0} is their limit: for every k≥1k\geq1 and 0<ε<T0<\varepsilon<T, the first kk curves converge in distribution in C([ε,T];Rk)C([\varepsilon,T];\mathbb{R}^k) [1 Theorem 1.3].

Our general weak uniqueness result identifies this ensemble from the initial state entire function e−ze^{-z}.

Corollary 1.4. Let Φ\Phi solve the positive martingale problem of Theorem 1.2 with (θ,κ)=(0,0)(\theta,\kappa)=(0,0) and Φ0(z)=e−z\Phi_0(z)=e^{-z}. Its ordered positive zeros r1(t)≤r2(t)≤⋯r_1(t)\leq r_2(t)\leq\cdots, counted with multiplicity, form an infinite continuous family for t>0t>0, and

Law⁡((−log⁡ri(t))i∈N,t>0)=Law⁡(ξ).(10)\operatorname{Law}\left((- \log r_i(t))_{i\in\mathbb{N},t>0}\right)=\operatorname{Law}(\xi). \tag*{(10)}

Proof. Set xi(N)(t)=eξi(N)(t)x_i^{(N)}(t)=e^{\xi_i^{(N)}(t)}. The reverse polynomials ∏i=1N(1−zxi(N)(t))\prod_{i=1}^N(1-zx_i^{(N)}(t)) start from (1−z/N)N→e−z(1-z/N)^N\to e^{-z}. By [7 Theorem 1.4 and Eq. (12)], these processes converge jointly in path space with the scaled squared singular values. The limiting entire-function-valued process solves the present martingale problem, and its canonical reciprocal roots coincide with the limiting particle coordinates. By (9) and Ahn’s convergence, those coordinate paths have law (eξi(t))i∈N,t>0(e^{\xi_i(t)})_{i\in\mathbb{N},t>0}. Theorem 1.2 identifies the entire-function path law with that of Φ\Phi, proving the assertion.

On the proof

The whole proof rests on a single idea which exploits the integrable structure of the problem. Our use of integrability is through finite-dimensional duality. The reciprocal products Hn+(g,s)=∏j=1ng(−sj)−1H_n^+(\mathbf{g},\mathbf{s})=\prod_{j=1}^{n}g(-s_j)^{-1} and HnHP(g,u)=∏j=1ng(iuj)−1H_n^{\mathrm{HP}}(\mathbf{g},\mathbf{u})=\prod_{j=1}^{n}g(iu_j)^{-1} are bounded in modulus by one and determine probability laws on the respective spaces of entire functions. Applying Itô’s formula to these probes turns the nonlinear covariance into a finite-dimensional diffusion generator and a multiplication operator. For each n∈Nn\in\mathbb{N} the dual diffusions S=(Si)i=1n\mathbf{S}=(S_i)_{i=1}^{n} have the common form

dSi=cSi dBi+(βSi+c∑j=1j≠inSiSjSi−Sj)dt,1≤i≤n.\mathrm{d}S_i=\sqrt{c}S_i\,\mathrm{d}B_i+\left(\beta S_i+c\sum_{\substack{j=1\\j\ne i}}^{n}\frac{S_iS_j}{S_i-S_j}\right)\mathrm{d}t,\qquad1\le i\le n.

where the Brownian motions are independent and (c,β)=(1,θ/2)(c,\beta)=(1,\theta/2) in the positive case and (c,β)=(2,−2a)(c,\beta)=(2,-2a) in the Hua–Pickrell case. The first diffusion lives in 0<s1<⋯<sn0<s_1<\cdots<s_n; the second lives in a signed chamber u1<⋯<uℓ<0<uℓ+1<⋯<unu_1<\cdots<u_\ell<0<u_{\ell+1}<\cdots<u_n. Both are conservative, avoid collisions, and preserve signs.

This essentially gives the main step of the proof which is the conditional duality of Proposition 3.3: for ⋄=+\diamond=+ or HP,

E[Hn⋄(Φt,v)∣Fr]=Ev[e∫0t−rC⋄(Sq) dqHn⋄(Φr,St−r)],0≤r≤t.\mathbb{E}\left[H_n^\diamond(\boldsymbol{\Phi}_t,\mathbf{v})\mid\mathcal{F}_r\right]=\mathbb{E}_{\mathbf{v}}\left[e^{\int_0^{t-r}C^\diamond(S_q)\,\mathrm{d}q}H_n^\diamond(\boldsymbol{\Phi}_r,\mathbf{S}_{t-r})\right],\qquad0\le r\le t.

where C+(s)=−κ2∑i=1nsiC_+(\mathbf{s})=-\frac{\kappa}{2}\sum_{i=1}^{n}s_i, CHP(u)=2ib∑i=1nui−∑i=1nui2C_{\mathrm{HP}}(\mathbf{u})=2ib\sum_{i=1}^{n}u_i-\sum_{i=1}^{n}u_i^2, and the expectation on the right is over an independent dual diffusion S\mathbf{S}. Then, localisation, together with conditioning on bounded initial-coordinate sets in the Hua–Pickrell case, proves this identity for arbitrary initial laws without moment assumptions. Since the probes determine laws, iteration of the conditional identity gives uniqueness of the entire path law and hence the Markov semigroups.

Quantitative convergence to equilibrium

The duality methods developed here also provide a starting point for studying quantitative convergence to equilibrium towards stochastic zeta functions; see [13, 33, 21, 5] for related constructions and limits of characteristic polynomials. In particular, it is natural to ask whether convergence from fixed initial states is exponentially fast in Wasserstein distances induced by bounded metrics compatible with the locally uniform topology. Combining the duality viewpoint with coupling arguments should provide a route to such estimates. In the Rider–Valkó and real-parameter Hua–Pickrell cases, the equilibrium limits have the form

Φt→t→∞lawΦ∞,Φ∞(z)={∏ξ∈Besν(1−4zξ),Rider–Valkoˊ,ν>−1,pv⁡∏ξ∈Za(1−zξ),Hua–Pickrell,a>−12, b=0.\Phi_t\xrightarrow[t\to\infty]{\mathrm{law}}\Phi_\infty,\qquad \Phi_\infty(z)= \begin{cases} \prod_{\xi\in\mathrm{Bes}_\nu}\left(1-\frac{4z}{\xi}\right), & \text{Rider--Valkó},\quad\nu>-1,\\[6pt] \operatorname{pv}\prod_{\xi\in\mathcal{Z}_a}\left(1-\frac{z}{\xi}\right), & \text{Hua--Pickrell},\quad a>-\frac{1}{2},\ b=0. \end{cases}

with convergence in the locally uniform topology [6, 7]. Here Besν\mathrm{Bes}_\nu is the Bessel determinantal point process in the Tracy–Widom normalization, with hard-edge variable 4Nx4Nx for the Laguerre weight xνe−xx^\nu e^{-x} [31, 15], and Za\mathcal{Z}_a is the reciprocal image of the real-parameter Hua–Pickrell boundary point process [9, 27]. The principal value is taken through symmetric truncations ∣ξ∣<N2|\xi| < N^{2}. At a=b=0a=b=0, the zeros form the sine determinantal point process, up to a deterministic spatial rescaling [4]. Quantitative relaxation to this equilibrium is particularly interesting in view of Suzuki’s result [30]: the unlabelled infinite Dyson Brownian motion at β=2\beta=2, which also has sine equilibrium, has no spectral gap and hence admits no uniform exponential convergence estimate in L2L^{2} of its invariant measure.

Acknowledgements Fengyi Li is grateful to the School of Mathematics of the University of Edinburgh for financial support through a PhD studentship.

AI use statement Motivated by the use of duality to prove uniqueness for martingale formulations of stochastic partial differential equations, and the fact that dualities are known to be intimately related to intertwinings, we used AI tools (ChatGPT 5.6 Sol) to search for and find the precise form of the duality in Proposition 3.3 a few months ago. Once this duality is found the rest of the argument is routine. We were also using AI tools (ChatGPT 5.5, 5.6 Sol) prior to this to investigate other approaches (which led to results) but the arguments were much less elegant to the duality route taken here. We used ChatGPT 6 to assist the writeup with rough drafts following instructions and outlines from the authors. We then revised the proofs and prepared the final manuscript and we are solely responsible for its contents.

State spaces and martingale problems

State-space preliminaries

Throughout, we use the state spaces, drift operators and covariance kernels introduced in § 1. A weak solution with initial law μ\mu is an adapted process with continuous paths in the corresponding state space, defined on a filtered probability space satisfying the usual conditions, with initial distribution μ\mu, such that the evaluation processes in (7) are continuous local martingales with the covariations (8), all relative to the same filtration. Covariations are extended complex-bilinearly. A solution from gg means one with initial law δg\delta_g.

For g∈E+g \in\mathcal{E}_{+}, write x=(xj)j≥1\mathbf{x}=(x_j)_{j\geq1} for its ordered reciprocal-root coordinates. For g∈Eg \in\mathcal{E}, let α+\alpha^{+} and α−\alpha^{-} be the decreasing lists of positive and negative reciprocal-root magnitudes, padded with zeros, and fix the signed enumeration

y2j−1=αj+,y2j=−αj−,j≥1.(11)y_{2j-1}=\alpha_j^{+}, \qquad y_{2j}=-\alpha_j^{-}, \qquad j\geq1. \tag*{(11)}

The coordinates satisfy

γ(g)=−g′(0),δ(g)=γ(g)2−g′′(0)(g∈E).(12)\gamma(g)=-g'(0), \qquad\delta(g)=\gamma(g)^2-g''(0) \qquad(g\in\mathcal{E}). \tag*{(12)}

The canonical products identify E+\mathcal{E}_{+} and E\mathcal{E} homeomorphically with their boundary spaces in the coordinates (x,γ)(\mathbf{x},\gamma) and (α+,α−,γ,δ)(\alpha^{+},\alpha^{-},\gamma,\delta), respectively, equipped with coordinate-wise convergence; see [5], Remarks 2.6–2.7 and the proof of Proposition 5.6. Both spaces are Polish, and the ordered reciprocal-root coordinates are continuous. The "dust" coordinates d=γ−∑j=1∞xjd=\gamma-\sum_{j=1}^{\infty}x_j and η=δ−∑j=1∞yj2\eta=\delta-\sum_{j=1}^{\infty}y_j^2 are Borel measurable. For every R<∞R<\infty, the sets

KR={g∈E+:γ(g)≤R},KR={g∈E:γ(g)2+δ(g)≤R}(13)K_R=\{g\in\mathcal{E}_{+}:\gamma(g)\leq R\}, \qquad \mathcal{K}_R=\{g\in\mathcal{E}:\gamma(g)^2+\delta(g)\leq R\} \tag*{(13)}

are compact, since the corresponding coordinate sublevels are compact. Continuous weak solutions from every deterministic state are supplied by [7 Theorems 1.4 and 1.6]. Theorem 1.4 covers (θ,κ)=(θ,0)(\theta,\kappa)=(\theta,0); Theorem 1.6 covers (θ,κ)=(−ν,1)(\theta,\kappa)=(-\nu,1) and Hua–Pickrell with s=a+ibs=a+ib, and taking ν=−θ\nu=-\theta gives the remaining positive cases. Their initial-state hypotheses follow from the density of normalised polynomials with positive, respectively real, zeros and the equivalence of locally uniform convergence and convergence of the boundary coordinates; see [5 Proposition 2.5 and Remark 2.7].

Hua–Pickrell covariance and bounded probes

We first introduce reciprocal evaluations on the imaginary axis, which provide bounded observables for the duality and tightness arguments.

Definition 2.1. For g∈Eg\in\mathcal{E}, define

hgHP(u)≔1g(iu),u∈R.(14)h_g^{\mathrm{HP}}(u)\coloneqq\frac{1}{g(iu)},\qquad u\in\mathbb{R}. \tag*{(14)}

For fixed uu, this is continuous in gg for locally uniform convergence. If u≠0u\ne0, the denominator is nonzero because every zero of gg is real; the case u=0u=0 is immediate.

The characteristic-function representation supplies the derivative and modulus bounds needed to justify duality and control state-space tails.

Lemma 2.2. Let ZZ be standard normal and let E1,E2,…E_1,E_2,\ldots be independent mean-one exponential variables, independent of ZZ. Then

Yg≔γ+qZ+∑j=1∞yj(Ej−1)(15)Y_g\coloneqq\gamma+\sqrt{q}Z+\sum_{j=1}^{\infty}y_j(E_j-1) \tag*{(15)}

is well defined by L2L^2 convergence of the series, and

hgHP(u)=E[eiuYg],E[Yg]=γ,Var⁡(Yg)=δ.(16)h_g^{\mathrm{HP}}(u)=\mathbb{E}\left[e^{iuY_g}\right],\qquad\mathbb{E}[Y_g]=\gamma,\qquad\operatorname{Var}(Y_g)=\delta. \tag*{(16)}

In particular,

∣hgHP(u)∣≤1,hgHP(−u)=hgHP(u)‾,(17)\left|h_g^{\mathrm{HP}}(u)\right|\le1,\qquad h_g^{\mathrm{HP}}(-u)=\overline{h_g^{\mathrm{HP}}(u)}, \tag*{(17)}

and, for k=1,2k=1,2,

∣∂ukhgHP(u)∣≤E[∣Yg∣k],E[∣Yg∣2]=γ2+δ.(18)\left|\partial_u^k h_g^{\mathrm{HP}}(u)\right|\le\mathbb{E}\left[|Y_g|^k\right],\qquad\mathbb{E}\left[|Y_g|^2\right]=\gamma^2+\delta. \tag*{(18)}

Moreover,

∣hgHP(u)∣2=e−qu2∏j=1∞(1+yj2u2)−1≤(1+δu2)−1.(19)\left|h_g^{\mathrm{HP}}(u)\right|^2=e^{-qu^2}\prod_{j=1}^{\infty}(1+y_j^2u^2)^{-1}\le(1+\delta u^2)^{-1}. \tag*{(19)}

Proof. The centred exponential series in (15) is an orthogonal series in L2L^2, with sum of variances ∑j=1∞yj2<∞\sum_{j=1}^{\infty}y_j^2<\infty. Since

E[eiuy(E−1)]=e−iuy1−iuy,\mathbb{E}\left[e^{iuy(E-1)}\right]=\frac{e^{-iuy}}{1-iuy},

independence and L2L^2 convergence give

E[eiuYg]=eiγu−qu2/2∏j=1∞e−iuyj1−iuyj=1g(iu).(20)\mathbb{E}\left[e^{iuY_g}\right]=e^{i\gamma u-qu^2/2}\prod_{j=1}^{\infty}\frac{e^{-iuy_j}}{1-iuy_j}=\frac{1}{g(iu)}. \tag*{(20)}

This proves (16)–(18). The identity in (19) follows by taking absolute values. Finally,

equ2∏j=1∞(1+yj2u2)≥1+(q+∑j=1∞yj2)u2,e^{qu^2}\prod_{j=1}^{\infty}(1+y_j^2u^2)\ge1+\left(q+\sum_{j=1}^{\infty}y_j^2\right)u^2,

which proves the last inequality. □

Differentiating evaluation martingales recovers the state coordinates; local covariance bounds justify this operation without global moment assumptions. This is the content of the next technical lemma.

Lemma 2.3. Let Φ\Phi solve the Hua–Pickrell martingale problem. For every z0,w0∈Cz_0,w_0\in\mathbb{C} and k,ℓ≥0k,\ell\ge0,

∂zkMtΦ(z0)=∂zkΦt(z0)−∂zkΦ0(z0)−∫0t∂zk(La,bHPΦs)(z0) ds(21)\partial_z^kM_t^\Phi(z_0)=\partial_z^k\Phi_t(z_0)-\partial_z^k\Phi_0(z_0)-\int_0^t\partial_z^k\left(\mathcal{L}_{a,b}^{\mathrm{HP}}\Phi_s\right)(z_0)\,\mathrm{d}s \tag*{(21)}

is a continuous local martingale, and

[∂zkMΦ(z0),∂wℓMΦ(w0)]t=∫0t∂zk∂wℓQsHP(z0,w0) ds.(22)\left[\partial_z^kM^\Phi(z_0),\partial_w^\ell M^\Phi(w_0)\right]_t=\int_0^t\partial_z^k\partial_w^\ell Q_s^{\mathrm{HP}}(z_0,w_0)\,\mathrm{d}s. \tag*{(22)}

Proof. We first abbreviate for simplicity

Mt(z)=MtΦ(z),Qt(z,w)=QtHP(z,w),L=La,bHP.M_t(z)=M_t^\Phi(z),\quad Q_t(z,w)=Q_t^{\mathrm{HP}}(z,w),\quad\mathcal{L}=\mathcal{L}_{a,b}^{\mathrm{HP}}.

Compact-open continuity of Φ\Phi and Cauchy’s estimates imply that t↦LΦtt\mapsto\mathcal{L}\Phi_t is compact-open continuous, hence bounded on compact space–time sets, pathwise. Cauchy’s formula and Fubini therefore give

∂zk(∫0t(LΦs)(z) ds)=∫0t∂zk(LΦs)(z) ds.\partial_z^k\left(\int_0^t(\mathcal{L}\Phi_s)(z)\,\mathrm{d}s\right)=\int_0^t\partial_z^k(\mathcal{L}\Phi_s)(z)\,\mathrm{d}s.

Thus MtM_t is entire in zz, its spatial derivatives are continuous in tt, and (20) holds pathwise. It remains to establish the local-martingale property and the bracket identity (21).

Choose R>1+max⁡{∣z0∣,∣w0∣}R>1+\max\{|z_0|,|w_0|\}, and put

D={z:∣z∣≤R},D′={z:∣z∣≤R+2}.D=\{z:|z|\le R\},\qquad D'=\{z:|z|\le R+2\}.

We define the stopped process as following,

σL:=L∧inf⁡{t≥0:sup⁡z∈D′∣Φt(z)∣+sup⁡z,w∈D′∣Qt(z,w)∣≥L},MtL(z)=Mt∧σL(z).\sigma_L\mathrel{\mathop:}=L\wedge\inf\left\{t\ge0:\sup_{z\in D'}|\Phi_t(z)|+\sup_{z,w\in D'}|Q_t(z,w)|\ge L\right\},\qquad M_t^L(z)=M_{t\wedge\sigma_L}(z).

The quantities defining σL\sigma_L are adapted and continuous. Since the stopping time is common to all evaluation points, MtLM_t^L remains entire in zz. Moreover, ∣Qs(z,w)∣≤L|Q_s(z,w)|\le L on D′×D′D'\times D' for s<σLs<\sigma_L.

Real-type symmetry and the stopped covariance bound imply

[ML(z),ML(z)‾]T≤LT,z∈D′.[M^L(z),\overline{M^L(z)}]_T\le LT,\qquad z\in D'.

Thus each stopped evaluation is a square-integrable martingale. For any complex valued process XX, we define

∥X∥2,T:=(E[sup⁡t≤T∣Xt∣2])1/2.\|X\|_{2,T}\mathrel{\mathop:}=\left(\mathbb{E}\left[\sup_{t\le T}|X_t|^2\right]\right)^{1/2}.

The martingale isometry and Doob’s inequality yield

E[∣MTL(z)∣2]≤LT,∥ML(z)∥2,T≤def2LT,\mathbb{E}\left[|M_T^L(z)|^2\right] \le LT,\qquad\|M^L(z)\|_{2,T} \mathrel{\mathop{\le}^{\mathrm{def}}} 2\sqrt{LT},

These estimates require no initial moment assumption: if the initial supremum already reaches LL, then σL=0\sigma_L=0 and ML≡0M^L\equiv0.

We next obtain spatial continuity in this norm. Cauchy’s estimates and integration of ∂1∂2Qs\partial_1\partial_2Q_s along the segments w→zw\to z and w‾→z‾\overline{w}\to\overline{z} bound the Hermitian bracket density of ML(z)−ML(w)M^L(z)-M^L(w) by L∣z−w∣2L|z-w|^2 before stopping. The martingale isometry and Doob’s inequality therefore give

∥ML(z)−ML(w)∥2,T≤2LT∣z−w∣,z,w∈D.\|M^L(z)-M^L(w)\|_{2,T}\le2\sqrt{LT}|z-w|,\qquad z,w\in D.

This estimate will allow the Cauchy contour sums to converge in ∥⋅∥2,T\|\cdot\|_{2,T}. Put zθ=z0+eiθz_\theta=z_0+e^{i\theta} and Fθ(t)=e−ikθMtL(zθ)F_\theta(t)=e^{-ik\theta}M_t^L(z_\theta). The preceding estimates imply

∥Fθ−Fφ∥2,T≤2(k+1)LT∣θ−φ∣.\|F_\theta-F_\varphi\|_{2,T}\le2(k+1)\sqrt{LT}|\theta-\varphi|.

By Cauchy’s formula,

At=def∂zkMtL(z0)=k!2π∫02πFθ(t) dθ.A_t\mathrel{\mathop{=}^{\mathrm{def}}}\partial_z^kM_t^L(z_0)=\frac{k!}{2\pi}\int_0^{2\pi}F_\theta(t)\,d\theta.

For δN=2π/N\delta_N=2\pi/N and θj=jδN\theta_j=j\delta_N, we have

At(N)=defk!N∑j=0N−1Fθj(t),sup⁡N∈N∥A(N)∥2,T≤2k!LT.A_t^{(N)}\mathrel{\mathop{=}^{\mathrm{def}}}\frac{k!}{N}\sum_{j=0}^{N-1}F_{\theta_j}(t),\qquad\sup_{N\in\mathbb{N}}\|A^{(N)}\|_{2,T}\le2k!\sqrt{LT}.

Each A(N)A^{(N)} is a continuous square-integrable martingale. Minkowski’s inequality and the preceding Lipschitz estimate give

∥A(N)−A∥2,T≤k!2π∑j=0N−1∫θjθj+1∥Fθj−Fθ∥2,T dθ≤2k!(k+1)LTδN⟶0.\|A^{(N)}-A\|_{2,T}\le\frac{k!}{2\pi}\sum_{j=0}^{N-1}\int_{\theta_j}^{\theta_{j+1}}\|F_{\theta_j}-F_\theta\|_{2,T}\,d\theta\le2k!(k+1)\sqrt{LT}\delta_N\longrightarrow0.

Thus AA is adapted and square-integrable. Let ∥⋅∥2\|\cdot\|_2 be the L2L^2 norm. For 0≤s≤t≤T0\le s\le t\le T, the martingale property of A(N)A^{(N)} and the L2L^2-contraction of conditional expectation yield

∥E[At∣Fs]−As∥2≤∥At−At(N)∥2+∥As(N)−As∥2≤2∥A−A(N)∥2,T⟶0.\|\mathbb{E}[A_t\mid\mathcal{F}_s]-A_s\|_2\le\|A_t-A_t^{(N)}\|_2+\|A_s^{(N)}-A_s\|_2\le2\|A-A^{(N)}\|_{2,T}\longrightarrow0.

Hence AA is a continuous square-integrable martingale. Similarly, Bt=def∂wℓMtL(w0)B_t\mathrel{\mathop{=}^{\mathrm{def}}}\partial_w^\ell M_t^L(w_0) is a continuous square-integrable martingale.

Set C(N)=[A(N),B(N)]C^{(N)}=[A^{(N)},B^{(N)}]. By the bilinear covariance identity, C(N)C^{(N)} is the corresponding double-contour Riemann sum involving QsQ_s. Cauchy’s formula and joint continuity identify its limit as

Ct=∫0t∧σL∂zk∂wℓQs(z0,w0) ds,C_t=\int_0^{t\wedge\sigma_L}\partial_z^k\partial_w^\ell Q_s(z_0,w_0)\,ds,

with convergence uniform on bounded time intervals, almost surely. The localized covariance bound gives sup⁡t≤T∣Ct(N)∣≤k!ℓ!LT\sup_{t\le T}|C_t^{(N)}|\le k!\ell!LT, so dominated convergence yields

E[sup⁡t≤T∣Ct(N)−Ct∣]⟶0.\mathbb{E}\left[\sup_{t\le T}|C_t^{(N)}-C_t|\right]\longrightarrow0.

Similarly, the preceding ∥⋅∥2,T\lVert\cdot\rVert_{2,T}-convergence and Cauchy–Schwarz imply

E[sup⁡t≤T∣At(N)Bt(N)−AtBt∣]⟶0.\mathbb{E}\left[\sup_{t\le T}\left|A_t^{(N)}B_t^{(N)}-A_tB_t\right|\right]\longrightarrow0.

By Itô’s product formula, A(N)B(N)−C(N)A^{(N)}B^{(N)}-C^{(N)} is a local martingale. The preceding bounds give it an integrable supremum on every finite interval, hence it is a true martingale. Passing its conditional-expectation identities to the limit in L1L^1 shows that AB−CAB-C is a martingale. Since AB−[A,B]AB-[A,B] is also a local martingale, [A,B]−C[A,B]-C is a continuous local martingale of finite variation, starting at zero. It therefore vanishes, proving

[A,B]t=∫0t∧σL∂z∂wQs(z0,w0) ds.[A,B]_t=\int_0^{t\wedge\sigma_L}\partial_z\partial_w\mathcal{Q}_s(z_0,w_0)\,ds.

Finally, compact-open path continuity and continuity of the covariance kernel imply that the quantities defining σL\sigma_L are bounded on every finite time interval, pathwise. Hence σL↑∞\sigma_L\uparrow\infty. Because the stopping time is independent of the spatial variables, AA and BB are the stopped versions of the corresponding derivatives of MΦM^\Phi. These derivatives are therefore continuous local martingales, and removing the stopping in the preceding identity gives (22). Together with the pathwise identity (20), this proves the lemma. □ □

To estimate the martingale part of the coordinate process in the next lemma, we first record a factorisation of the Hua–Pickrell covariance. Retain the measurable signed enumeration (11). Logarithmic differentiation of (3) gives

g′(z)g(z)=−γ−qz−∑j=1∞yj2z1−yjz.(23)\frac{g'(z)}{g(z)}=-\gamma-qz-\sum_{j=1}^{\infty}\frac{y_j^2z}{1-y_jz}. \tag*{(23)}

Consequently,

2zwz−w(g(z)g′(w)−g(w)g′(z))=2zwg(z)g(w)(q+∑j=1∞yj2(1−yjz)(1−yjw)).(24)2\frac{zw}{z-w}\left(g(z)g'(w)-g(w)g'(z)\right)=2zwg(z)g(w)\left(q+\sum_{j=1}^{\infty}\frac{y_j^2}{(1-y_jz)(1-y_jw)}\right). \tag*{(24)}

The following algebraic square root of the covariance will be used below:

b0(g;z):def=−2q zg(z),bj(g;z):def=−2zyjg(z)1−yjz,j≥1.(25)b_0(g;z)\mathrel{\mathop:^{\mathrm{def}}}=-\sqrt{2q}\,zg(z),\qquad b_j(g;z)\mathrel{\mathop:^{\mathrm{def}}}=-\sqrt{2}\frac{zy_jg(z)}{1-y_jz},\qquad j\ge1. \tag*{(25)}

The apparent poles are removable, and (23) shows that the products of the coefficients in (24) sum to the prescribed covariance.

Set R(g)=γ(g)2+δ(g)R(g)=\gamma(g)^2+\delta(g).

The first two coordinates satisfy a closed growth estimate that controls probe derivatives, including after truncation on an initial event.

Lemma 2.4. Let Φ\Phi be a continuous weak solution of the Hua–Pickrell martingale problem (7)–(8) with parameters a,b∈Ra,b\in\mathbb{R}. Write

Γt:def=γ(Φt),Δt:def=δ(Φt),Rt:def=R(Φt).\Gamma_t\mathrel{\mathop:^{\mathrm{def}}}=\gamma(\Phi_t),\qquad\Delta_t\mathrel{\mathop:^{\mathrm{def}}}=\delta(\Phi_t),\qquad R_t\mathrel{\mathop:^{\mathrm{def}}}=R(\Phi_t).

Then

dΓt=dNt+2(b−aΓt) dt,d[N]t=2Δt dt.(26)d\Gamma_t=dN_t+2(b-a\Gamma_t)\,dt,\qquad d[N]_t=2\Delta_t\,dt. \tag*{(26)}
drift⁡(dΔt)=(2+2Γt2−4aΔt) dt,drift⁡(dRt)=(2+4bΓt+(2−4a)Rt) dt.(27)\operatorname{drift}(\mathrm{d}\Delta_t) = (2+2\Gamma_t^2-4a\Delta_t)\,\mathrm{d}t,\qquad\operatorname{drift}(\mathrm{d}R_t) = (2+4b\Gamma_t+(2-4a)R_t)\,\mathrm{d}t. \tag*{(27)}

The martingale part NRN^R of RR satisfies

d[NR]t≤16Rt2 dt.(28)\mathrm{d}[N^R]_t \le16R_t^2\,\mathrm{d}t. \tag*{(28)}

Consequently, for every m≥1m\ge1, T<∞T<\infty, and A∈F0\mathcal{A}\in\mathcal{F}_0 such that E[1A(1+R0)m]<∞\mathbb{E}[1_{\mathcal{A}}(1+R_0)^m]<\infty,

E[1Asup⁡t≤T(1+Rt)m]≤Cm,T,a,bE[1A(1+R0)m].(29)\mathbb{E}\left[1_{\mathcal{A}}\sup_{t\le T}(1+R_t)^m\right]\le C_{m,T,a,b}\mathbb{E}\left[1_{\mathcal{A}}(1+R_0)^m\right]. \tag*{(29)}

Proof. Lemma 2.3 allows differentiation of the martingale identities. The first two derivatives of the drift at zero are

−2b−2ag′(0),4bΓ−2−2(2a+1)Φ′′(0),-2b-2ag'(0),\qquad4b\Gamma-2-2(2a+1)\Phi''(0),

and the first mixed derivative of the bracket is 2δ2\delta. These give (25); Itô’s formula for Δ=Γ2−Φ′′(0)\Delta=\Gamma^2-\Phi''(0) and R=Γ2+ΔR=\Gamma^2+\Delta gives (26). The covariance factorisation (24) gives

d[NR]tdt=8qΓ2+8∑j=1∞(Γyj+yj2)2≤16(Γ2+Δ)2,\frac{\mathrm{d}[N^R]_t}{\mathrm{d}t}=8q\Gamma^2+8\sum_{j=1}^{\infty}(\Gamma y_j+y_j^2)^2\le16(\Gamma^2+\Delta)^2,

using ∑j=1∞yj4≤Δ2\sum_{j=1}^{\infty}y_j^4\le\Delta^2. Put X=1+RX=1+R and τL=inf⁡{t:Rt≥L}\tau_L=\inf\{t:R_t\ge L\}. Its drift is bounded in absolute value by Ca,bXC_{a,b}X, and d[NR]≤16X2dt\mathrm{d}[N^R]\le16X^2\mathrm{d}t. Itô’s formula for XmX^m, followed by the Burkholder–Davis–Gundy and Young inequalities, gives

E[sup⁡s≤t1A∣∫0s∧τLXvm−1 dNvR∣]≤12E[1Asup⁡s≤tXs∧τLm]+Cm∫0tE[1AXv∧τLm] dv.\mathbb{E}\left[\sup_{s\le t}1_{\mathcal{A}}\left|\int_0^{s\wedge\tau_L}X_{v}^{m-1}\,\mathrm{d}N_v^R\right|\right]\le\frac{1}{2}\mathbb{E}\left[1_{\mathcal{A}}\sup_{s\le t}X_{s\wedge\tau_L}^m\right]+C_m\int_0^t\mathbb{E}\left[1_{\mathcal{A}}X_{v\wedge\tau_L}^m\right]\,\mathrm{d}v.

Here A∈F0\mathcal{A}\in\mathcal{F}_0 permits multiplication of the martingale integrand by 1A1_{\mathcal{A}}. Absorbing the supremum term and applying Gronwall gives (28) up to τL\tau_L, uniformly in LL. Path continuity gives τL↑∞\tau_L\uparrow\infty, and Fatou removes the stopping. □

Positive bounded probes

Similarly to the Hua–Pickrell case, the reciprocals on the negative axis form bounded observables in the positive case.

Definition 2.5. For g∈E+g\in\mathcal{E}_+, and s>0s>0, define

hg+(s):=1g(−s),s>0.h_g^+(s)\mathrel{\mathop:}= \frac{1}{g(-s)},\qquad s>0.

For fixed ss, this is continuous in gg for locally uniform convergence.

We also have the following lemma.

Lemma 2.6. If (Ej)(E_j) are independent mean-one exponential variables, then

hg+(s)=E[e−sYg+],Yg+:=d(g)+∑j=1∞xjEj.(30)h_g^+(s)=\mathbb{E}\left[e^{-sY_g^+}\right],\qquad Y_g^+\mathrel{\mathop:}=d(g)+\sum_{j=1}^{\infty}x_jE_j. \tag*{(30)}

Moreover, for all m≥0m \ge0, we have

0<hg+(s)≤1,sup⁡g∈E+∣∂smhg+(s)∣≤sup⁡y≥0yme−sy=(mes)m,(31)0 < h_g^{+}(s) \le1,\qquad\sup_{g\in\mathcal{E}_{+}} \left|\partial_s^m h_g^{+}(s)\right| \le\sup_{y\ge0} y^m e^{-sy} = \left(\frac{m}{\mathrm{e}s}\right)^m, \tag*{(31)}

with value 11 when m=0m=0.

Proof. Indeed, the nonnegative random variable Yg+Y_g^{+} has finite expectation γ(g)\gamma(g), and independence gives the formula first for partial products and then by bounded convergence in general. For all m≥0m \ge0, differentiation under the expectation gives (30), hence completing the proof. □

Reciprocal dualities and weak uniqueness

Positive dual

The product of reciprocals on the negative axis determine the positive state law. Letting hg+h_g^{+} be the probe defined in Definition 2.5, put

Hn+(g,s):=∏i=1nhg+(si),s∈(0,∞)n.(32)H_n^{+}(g,s) \mathrel{\mathop:}= \prod_{i=1}^{n} h_g^{+}(s_i),\qquad s\in(0,\infty)^n. \tag*{(32)}

On Wn={0<s1<⋯<sn}\mathcal{W}_n=\{0<s_1<\cdots<s_n\}, define

Gnθ:=12∑i=1nsi2∂ii+θ2∑i=1nsi∂i+∑1≤i<j≤nsisjsi−sj(∂i−∂j),Vn+(s):=12∑i=1nsi.(33)G_n^\theta\mathrel{\mathop:}= \frac{1}{2}\sum_{i=1}^{n}s_i^2\partial_{ii}+\frac{\theta}{2}\sum_{i=1}^{n}s_i\partial_i+\sum_{1\le i<j\le n}\frac{s_i s_j}{s_i-s_j}(\partial_i-\partial_j),\qquad V_n^{+}(s)\mathrel{\mathop:}= \frac{1}{2}\sum_{i=1}^{n}s_i. \tag*{(33)}

The positive dual evolves finitely many probe arguments s\mathbf{s}.

Lemma 3.1. For every θ∈R\theta\in\mathbb{R}, GnθG_n^\theta generates a conservative continuous diffusion S=(Si)i=1n\mathbf{S}=(S_i)_{i=1}^{n} in Wn\mathcal{W}_n, satisfying

dSi=Si dBi+(θ2Si+∑j=1j≠inSiSjSi−Sj)dt.(34)\mathrm{d}S_i=S_i\,\mathrm{d}B_i+\left(\frac{\theta}{2}S_i+\sum_{\substack{j=1\\j\ne i}}^{n}\frac{S_iS_j}{S_i-S_j}\right)\mathrm{d}t. \tag*{(34)}

Moreover, for m≥1m\ge1, define

ρm:=inf⁡{t≥0:S(t)∉Wn or Sn(t)≥m},ρ:=lim⁡m→∞ρm,\rho_m\mathrel{\mathop:}= \inf\{t\ge0:\mathbf{S}(t)\notin\mathcal{W}_n\text{ or }S_n(t)\ge m\},\qquad\rho\mathrel{\mathop:}= \lim_{m\to\infty}\rho_m,

with inf⁡∅=∞\inf\varnothing=\infty. Then Ps(ρ=∞)=1\mathbb{P}_{\mathbf{s}}(\rho=\infty)=1 for every s∈Wn\mathbf{s}\in\mathcal{W}_n.

Proof. We will do the construction in logarithmic coordinates. Put yi=log⁡siy_i=\log s_i, μ=(θ−n)/2\mu=(\theta-n)/2, and

Vn(y):=∏1≤i<j≤n2sinh⁡(yj−yi2)=det⁡[eηiyj]i,j=1n,ηi=i−n+12.\mathcal{V}_n(\mathbf{y})\mathrel{\mathop:}= \prod_{1\le i<j\le n}2\sinh\left(\frac{y_j-y_i}{2}\right)=\det\left[e^{\eta_i y_j}\right]_{i,j=1}^{n},\qquad\eta_i=i-\frac{n+1}{2}.

The identity sj/(si−sj)=12coth⁡((yi−yj)/2)−12s_j/(s_i-s_j)=\frac{1}{2}\coth((y_i-y_j)/2)-\frac{1}{2} and Itô’s formula give the target generator in logarithmic coordinates

G~nθ=12Δy+μ∑i=1n∂yi+∇log⁡Vn⋅∇.(35)\widetilde{G}_n^\theta=\frac{1}{2}\Delta_{\mathbf{y}}+\mu\sum_{i=1}^{n}\partial_{y_i}+\nabla\log\mathcal{V}_n\cdot\nabla. \tag*{(35)}

Every determinant monomial has the exponents ηi\eta_i in some order, so

12ΔyVn=κnVn,∑i=1n∂yiVn=0,κn=12∑i=1nηi2=n(n2−1)24.\frac{1}{2}\Delta^y\mathcal{V}_n=\kappa_n\mathcal{V}_n,\qquad\sum_{i=1}^{n}\partial_{y_i}\mathcal{V}_n=0,\qquad\kappa_n=\frac{1}{2}\sum_{i=1}^{n}\eta_i^2=\frac{n(n^2-1)}{24}.

Let Z\mathbf{Z} be independent Brownian motions with common drift μ\mu, and define the first collision time

τ∂:=inf⁡{t≥0:Zi(t)=Zj(t) for some 1≤i<j≤n}.\tau_{\partial}\mathrel{\mathop:}=\inf\{t\ge0: Z_i(t)=Z_j(t)\text{ for some }1\le i<j\le n\}.

Karlin–McGregor [19] and Andréief [16] identities give

Ey[Vn(Zt);t<τ∂]=det⁡[eηiyj+μηit+ηi2t/2]i,j=1n=eκntVn(y).\mathbb{E}_y\left[\mathcal{V}_n(\mathbf{Z}_t);t<\tau_{\partial}\right]=\det\left[e^{\eta_i y_j+\mu\eta_i t+\eta_i^2t/2}\right]_{i,j=1}^{n}=e^{\kappa_nt}\mathcal{V}_n(y).

Here the product of the two determinants is symmetric, so the factor 1/n!1/n! from chamber integration cancels the factor in Andréief’s identity; all integrals are finite Gaussian exponential moments. Consequently,

Qtf(y)=e−κntVn(y)Ey[Vn(Zt)f(Zt);t<τ∂].Q_tf(y)=\frac{e^{-\kappa_nt}}{\mathcal{V}_n(y)}\mathbb{E}_y\left[\mathcal{V}_n(\mathbf{Z}_t)f(\mathbf{Z}_t);t<\tau_{\partial}\right].

is a probability semigroup. Its conservative Doob hh-transform [14, 28] has continuous paths staying in the open chamber at every finite time and generator (34). Exponentiation proves the claim. □

The Hua–Pickrell signed dual

Using the probes from Lemma 2.2, put

HnHP(g,u):=∏i=1nhgHP(ui),VnHP(u):=∑i=1nui2,(36)H_n^{\mathrm{HP}}(\mathbf{g},\mathbf{u})\mathrel{\mathop:}=\prod_{i=1}^{n}h_g^{\mathrm{HP}}(u_i),\qquad V_n^{\mathrm{HP}}(\mathbf{u})\mathrel{\mathop:}=\sum_{i=1}^{n}u_i^2, \tag*{(36)}
Cb,n(u):=2ib∑i=1nui−VnHP(u).C_{b,n}(\mathbf{u})\mathrel{\mathop:}=2ib\sum_{i=1}^{n}u_i-V_n^{\mathrm{HP}}(\mathbf{u}).

For pairwise distinct nonzero coordinates define

Hna:=∑i=1nui2∂ii−2a∑i=1nui∂i+2∑1≤i<j≤nuiujui−uj(∂i−∂j).(37)\mathcal{H}_n^a\mathrel{\mathop:}=\sum_{i=1}^{n}u_i^2\partial_{ii}-2a\sum_{i=1}^{n}u_i\partial_i+2\sum_{1\le i<j\le n}\frac{u_iu_j}{u_i-u_j}(\partial_i-\partial_j). \tag*{(37)}

The signed chambers are

Cℓ,n:={u1<⋯<uℓ<0<uℓ+1<⋯<un},0≤ℓ≤n.(38)\mathcal{C}_{\ell,n}\mathrel{\mathop:}=\{u_1<\cdots<u_\ell<0<u_{\ell+1}<\cdots<u_n\},\qquad0\le\ell\le n. \tag*{(38)}

Any vector u∈Rn\mathbf{u}\in\mathbb{R}^n with pairwise distinct nonzero coordinates has a unique increasing rearrangement in Cℓ,n\mathcal{C}_{\ell,n}, where ℓ=#{i:ui<0}\ell=\#\{i:u_i<0\}. All the duality formulas below are invariant under this relabelling.

The signed dual preserves each argument’s sign and avoids collisions, providing the corresponding Feynman–Kac representation for Hua–Pickrell probes.

Lemma 3.2. On every Cℓ,n\mathcal{C}_{\ell,n}, Ana\mathcal{A}^{a}_{n} generates a conservative continuous diffusion satisfying

dSi=2Si dBi+[−2aSi+2∑j=1j≠inSiSjSi−Sj]dt.(39)\mathrm{d}S_i=\sqrt{2}S_i\,\mathrm{d}B_i+\left[-2aS_i+2\sum_{\substack{j=1\\j\ne i}}^{n}\frac{S_iS_j}{S_i-S_j}\right]\mathrm{d}t. \tag*{(39)}

Moreover, for m≥1m\ge1, define

ρm≔inf⁡{t≥0:S(t)∉Cℓ,n or max⁡1≤i≤n∣Si(t)∣≥m},ρ≔lim⁡m→∞ρm,\rho_m\coloneqq\inf\left\{t\ge0:S(t)\notin\mathcal{C}_{\ell,n}\ \text{or}\ \max_{1\le i\le n}|S_i(t)|\ge m\right\},\qquad\rho\coloneqq\lim_{m\to\infty}\rho_m,

with inf⁡∅=∞\inf\varnothing=\infty. Then Pu(ρ=∞)=1\mathbb{P}_u(\rho=\infty)=1 for every u∈Cℓ,nu\in\mathcal{C}_{\ell,n}.

Proof. Set

V(u)≔det⁡[uij−1]i,j=1n=∏1≤i<j≤n(uj−ui),β≔−2a−2(n−1),L0≔∑i=1n(ui2∂ii+βui∂i).\mathcal{V}(\mathbf{u})\coloneqq\det\left[u_i^{j-1}\right]_{i,j=1}^{n}=\prod_{1\le i<j\le n}(u_j-u_i),\qquad\beta\coloneqq-2a-2(n-1),\qquad\mathcal{L}_0\coloneqq\sum_{i=1}^{n}\left(u_i^2\partial_{ii}+\beta u_i\partial_i\right).

and Λn=∑r=0n−1r(r−1+β)\Lambda_n=\sum_{r=0}^{n-1}r(r-1+\beta). The monomial expansion gives L0V=ΛnV\mathcal{L}_0\mathcal{V}=\Lambda_n\mathcal{V}; the product rule, ∂ilog⁡V=∑j≠i(ui−uj)−1\partial_i\log\mathcal{V}=\sum_{j\ne i}(u_i-u_j)^{-1}, and ui2/(ui−uj)=ui+uiuj/(ui−uj)u_i^2/(u_i-u_j)=u_i+u_iu_j/(u_i-u_j) give

Anaf=V−1L0(Vf)−Λnf.(40)\mathcal{A}^{a}_{n}f=\mathcal{V}^{-1}\mathcal{L}_0(\mathcal{V}f)-\Lambda_nf. \tag*{(40)}

Let ptβ(x,y)p_t^\beta(x,y) be the transition density of signed geometric Brownian motion with generator x2∂xx+βx∂xx^2\partial_{xx}+\beta x\partial_x on R∗=R∖{0}\mathbb{R}^{*}=\mathbb{R}\setminus\{0\}. It vanishes when xy<0xy<0; hence Karlin–McGregor [19] applied separately to the negative and positive blocks gives the killed density det⁡[ptβ(ui,vj)]\det[p_t^\beta(u_i,v_j)] on Cℓ,n\mathcal{C}_{\ell,n}. Its product with V(v)\mathcal{V}(v) is symmetric in vv and vanishes unless precisely ℓ\ell terminal coordinates are negative. That labelled region is the union of the n!n! permutations of the chamber. The moment identity

∫R∗ptβ(u,v)vr dv=er(r−1+β)tur\int_{\mathbb{R}^{*}}p_t^\beta(u,v)v^r\,\mathrm{d}v=e^{r(r-1+\beta)t}u^r

and Andréief identity [16] therefore give

∫Cℓ,ndet⁡[ptβ(ui,vj)]V(v) dv=det⁡[∫R∗ptβ(ui,v)vj−1 dv]i,j=1n=eΛntV(u).(41)\int_{\mathcal{C}_{\ell,n}}\det\left[p_t^\beta(u_i,v_j)\right]\mathcal{V}(v)\,\mathrm{d}v =\det\left[\int_{\mathbb{R}^{*}}p_t^\beta(u_i,v)v^{j-1}\,\mathrm{d}v\right]_{i,j=1}^{n} =e^{\Lambda_nt}\mathcal{V}(u). \tag*{(41)}

The integrals are absolutely convergent because geometric Brownian motion has every finite absolute moment. Since V>0\mathcal{V}>0 in the chamber, (39)–(40) define a conservative Doob hh-transform. Its continuous paths remain in the chamber at every finite time, and its generator gives the displayed SDE. □

Conditional duality

The following identities express conditional probe expectations through the finite duals, without assuming a Markov property for the original solution. This is the heart of the argument.

Proposition 3.3. For every initial law and 0≤r≤t0 \le r \le t, every solution of the positive (θ,κ)(\theta,\kappa) martingale problem satisfies

E[Hn+(Φt,s)∣Fr]=Es[e−κ∫0t−rVn+(Sv) dvHn+(Φr,St−r)].(42)\mathbb{E}\left[H_n^+(\Phi_t,\mathbf{s}) \mid\mathcal{F}_r\right] = \mathbb{E}_{\mathbf{s}}\left[ e^{-\kappa\int_0^{t-r} V_n^+(\mathbf{S}_v)\,\mathrm{d}v} H_n^+(\Phi_r,\mathbf{S}_{t-r}) \right]. \tag*{(42)}

where the expectation on the right is taken only with respect to the (independent) dual diffusion with generator GnθG_n^\theta. Furthermore, every solution of the Hua–Pickrell martingale problem satisfies

E[HnHP(Φt,u)∣Fr]=Eu[e∫0t−rCb,n(Sv) dvHnHP(Φr,St−r)].(43)\mathbb{E}\left[H_n^{\mathrm{HP}}(\Phi_t,\mathbf{u}) \mid\mathcal{F}_r\right] = \mathbb{E}_{\mathbf{u}}\left[ e^{\int_0^{t-r} C_{b,n}(\mathbf{S}_v)\,\mathrm{d}v} H_n^{\mathrm{HP}}(\Phi_r,\mathbf{S}_{t-r}) \right]. \tag*{(43)}

for pairwise distinct nonzero u\mathbf{u}, where the expectation on the right is over the (independent) signed dual diffusion with generator Anu\mathcal{A}_n^u.

Proof. Write rt(z)=Φt(z)−1r_t(z)=\Phi_t(z)^{-1} on C∖[0,∞)\mathbb{C}\setminus[0,\infty) in the positive case and on C∖R\mathbb{C}\setminus\mathbb{R} in the Hua–Pickrell case. These are fixed zero-free domains. For each evaluation, continuity and nonvanishing make the stopping times before the denominator approaches zero tend to infinity. Holomorphic Itô calculus, with dNtΦ(z)=−Φt(z)−2dMtΦ(z)\mathrm{d}N_t^\Phi(z)=-\Phi_t(z)^{-2}\mathrm{d}M_t^\Phi(z), gives, in the positive and Hua–Pickrell cases respectively,

drt(z)=dNtΦ(z)+(κ2zrt(z)+θ2zrt′(z)+12z2rt′′(z))dt,(44)\mathrm{d}r_t(z)=\mathrm{d}N_t^\Phi(z)+\left(\frac{\kappa}{2}zr_t(z)+\frac{\theta}{2}zr_t'(z)+\frac{1}{2}z^2r_t''(z)\right)\mathrm{d}t, \tag*{(44)}
drt(z)=dNtΦ(z)+((2bz+z2)rt(z)−2azrt′(z)+z2rt′′(z))dt.(45)\mathrm{d}r_t(z)=\mathrm{d}N_t^\Phi(z)+\left((2bz+z^2)r_t(z)-2azr_t'(z)+z^2r_t''(z)\right)\mathrm{d}t. \tag*{(45)}

Indeed, r′=−Φ′/Φ2r'=-\Phi'/\Phi^2 and r′′=2(Φ′)2/Φ3−Φ′′/Φ2r''=2(\Phi')^2/\Phi^3-\Phi''/\Phi^2; the diagonal bracket supplies the second derivative term in each formula. For distinct arguments, the cross-bracket is

d[NΦ(z),NΦ(w)]t=c⋄zwz−w(rt′(z)rt(w)−rt(z)rt′(w))dt,c+=1,cHP=2.(46)\mathrm{d}\left[N^\Phi(z),N^\Phi(w)\right]_t = c_\diamond\frac{zw}{z-w}\left(r_t'(z)r_t(w)-r_t(z)r_t'(w)\right)\mathrm{d}t, \qquad c_+=1,\quad c_{\mathrm{HP}}=2. \tag*{(46)}

Put

ht+(s)≔hΦt+(s)≔rt(−s),htHP(u)≔hΦtHP(u)≔rt(iu),h_t^+(s)\coloneqq h_{\Phi_t}^+(s)\coloneqq r_t(-s), \qquad h_t^{\mathrm{HP}}(u)\coloneqq h_{\Phi_t}^{\mathrm{HP}}(u)\coloneqq r_t(\mathrm{i}u),

and let

Lt+(s)≔NtΦ(−s),LtHP(u)≔NtΦ(iu).L_t^+(s)\coloneqq N_t^\Phi(-s), \qquad L_t^{\mathrm{HP}}(u)\coloneqq N_t^\Phi(\mathrm{i}u).

The chain rule gives

(ht+)′(s)=−rt′(−s),(ht+)′′(s)=rt′′(−s),(h_t^+)'(s)=-r_t'(-s), \qquad (h_t^+)''(s)=r_t''(-s),
(htHP)′(u)=irt′(iu),(htHP)′′(u)=−rt′′(iu).(h_t^{\mathrm{HP}})'(u)=\mathrm{i}r_t'(\mathrm{i}u), \qquad (h_t^{\mathrm{HP}})''(u)=-r_t''(\mathrm{i}u).

Thus the two reciprocal Itô identities become

dht+(s)=dLt+(s)+(12s2∂s2+θ2s∂s−κ2s)ht+(s) dt,\mathrm{d}h_t^+(s)=\mathrm{d}L_t^+(s)+\left(\frac{1}{2}s^2\partial_s^2+\frac{\theta}{2}s\partial_s-\frac{\kappa}{2}s\right)h_t^+(s)\,\mathrm{d}t,
dhtHP(u)=dLtHP(u)+(u2∂u2−2au∂u+2ibu−u2)htHP(u) dt.\mathrm{d}h_t^{\mathrm{HP}}(u)=\mathrm{d}L_t^{\mathrm{HP}}(u)+\left(u^2\partial_u^2-2au\partial_u+2\mathrm{i}bu-u^2\right)h_t^{\mathrm{HP}}(u)\,\mathrm{d}t.

Likewise, substituting into (45) yields, for distinct arguments,

d[L⋄(vi),L⋄(vj)]t=c⋄vivjvi−vj((ht⋄)′(vi)ht⋄(vj)−ht⋄(vi)(ht⋄)′(vj))dt.\mathrm{d}\left[L^\diamond(v_i),L^\diamond(v_j)\right]_t = c_\diamond\frac{v_iv_j}{v_i-v_j} \left((h_t^\diamond)'(v_i)h_t^\diamond(v_j)-h_t^\diamond(v_i)(h_t^\diamond)'(v_j)\right)\mathrm{d}t.

Here primes on ht⋄h_t^\diamond denote derivatives in its real argument, and c+=1c_+=1, cHP=2c_{\mathrm{HP}}=2.

Fix v\mathbf{v} in the corresponding chamber and put

Ht⋄(v)≔Hn⋄(Φt,v)=∏i=1nht⋄(vi).H_t^{\diamond}(\mathbf{v}) \coloneqq H_n^{\diamond}(\mathbf{\Phi}_t,\mathbf{v}) = \prod_{i=1}^{n} h_t^{\diamond}(v_i).

The product rule for continuous semimartingales gives

dHt⋄(v)=∑i=1n(∏k=1k≠inht⋄(vk))dht⋄(vi)+∑1≤i<j≤n(∏k=1k∉{i,j}nht⋄(vk))d[L⋄(vi),L⋄(vj)]t.\mathrm{d}H_t^{\diamond}(\mathbf{v}) = \sum_{i=1}^{n} \left(\prod_{\substack{k=1\\k\ne i}}^{n} h_t^{\diamond}(v_k)\right) \mathrm{d}h_t^{\diamond}(v_i) + \sum_{1\le i<j\le n} \left(\prod_{\substack{k=1\\k\notin\{i,j\}}}^{n} h_t^{\diamond}(v_k)\right) \mathrm{d}[L^{\diamond}(v_i),L^{\diamond}(v_j)]_t.

Empty products are understood to equal one. Since

∂iHt⋄=(ht⋄)′(vi)∏k=1k≠inht⋄(vk),\partial_i H_t^{\diamond} = (h_t^{\diamond})'(v_i)\prod_{\substack{k=1\\k\ne i}}^{n} h_t^{\diamond}(v_k),
(∂i−∂j)Ht⋄=((ht⋄)′(vi)ht⋄(vj)−ht⋄(vi)(ht⋄)′(vj))∏k=1k∉{i,j}nht⋄(vk),(\partial_i-\partial_j)H_t^{\diamond} = \left((h_t^{\diamond})'(v_i)h_t^{\diamond}(v_j)-h_t^{\diamond}(v_i)(h_t^{\diamond})'(v_j)\right)\prod_{\substack{k=1\\k\notin\{i,j\}}}^{n} h_t^{\diamond}(v_k),

the pairwise covariations give exactly the interaction terms in the dual generators. The remaining drift terms give the one-coordinate operators and the multiplication terms. Hence, with

(B+,C+)=(Gnθ,−κVn+),(BHP,CHP)=(Ana,Cb,n),(\mathcal{B}_{+},\mathcal{C}_{+}) = (\mathcal{G}^{\theta}_{n},- \kappa V^{+}_{n}), \qquad(\mathcal{B}_{\mathrm{HP}},\mathcal{C}_{\mathrm{HP}}) = (\mathcal{A}^{a}_{n},C_{b,n}),

the process

Zt⋄≔Hn⋄(Φt,v)−Hn⋄(Φ0,v)−∫0t(B⋄+C⋄)Hn⋄(Φs,⋅)(v) ds(47)Z_t^{\diamond} \coloneqq H_n^{\diamond}(\mathbf{\Phi}_t,\mathbf{v})-H_n^{\diamond}(\mathbf{\Phi}_0,\mathbf{v})-\int_0^t(\mathcal{B}_{\diamond}+\mathcal{C}_{\diamond})H_n^{\diamond}(\mathbf{\Phi}_s,\cdot)(\mathbf{v})\,\mathrm{d}s \tag*{(47)}

has the representation

Zt⋄=∑i=1n∫0t(∏k=1k≠inhs⋄(vk))dLs⋄(vi).Z_t^{\diamond} = \sum_{i=1}^{n}\int_0^t\left(\prod_{\substack{k=1\\k\ne i}}^{n}h_s^{\diamond}(v_k)\right)\mathrm{d}L_s^{\diamond}(v_i).

The integrands are continuous, adapted and bounded in modulus by one, so this is a continuous local martingale for every fixed v\mathbf{v} in the relevant chamber.

For the positive probes, (30) bounds all spatial derivatives uniformly over gg on every compact subchamber. Thus Z+Z^+ is a true martingale on bounded time intervals. For Hua–Pickrell probes, Lemma 2.2 gives

∣∂iHnHP∣≤R1/2,∣∂iiHnHP∣≤R,∣∂ijHnHP∣≤R(i≠j),|\partial_i H_n^{\mathrm{HP}}| \le R^{1/2}, \qquad|\partial_{ii}H_n^{\mathrm{HP}}| \le R, \qquad|\partial_{ij}H_n^{\mathrm{HP}}| \le R \quad(i\ne j),

and hence, for every compact subchamber D\mathcal{D},

sup⁡u∈D∣(Ana+Cb,n)HnHP(g,⋅)(u)∣≤CD(1+R(g)).(48)\sup_{\mathbf{u}\in\mathcal{D}}|(\mathcal{A}_n^a+C_{b,n})H_n^{\mathrm{HP}}(g,\cdot)(\mathbf{u})| \le C_{\mathcal{D}}(1+R(g)). \tag*{(48)}

If B∈F0\mathcal{B}\in\mathcal{F}_0 and E[1B(1+R(Φ0))2]<∞\mathbb{E}[1_{\mathcal{B}}(1+R(\Phi_0))^2]<\infty, Lemma 2.4 shows that 1BZHP1_{\mathcal{B}}Z^{\mathrm{HP}} is a true martingale: stop at τL=inf⁡{t ⁣:R(Φt)≥L}\tau_L=\inf\{t\colon R(\Phi_t)\ge L\}, where the stopped process is bounded on each finite time interval, and then use the integrable bound

sup⁡t≤T∣1BZt∧τLHP∣≤1B(2+CDTsup⁡s≤T(1+R(Φs)))\sup_{t\le T}|1_{\mathcal{B}}Z_{t\wedge\tau_L}^{\mathrm{HP}}|\le1_{\mathcal{B}}\left(2+C_{\mathcal{D}}T\sup_{s\le T}(1+R(\Phi_s))\right)

to remove the stopping by dominated convergence. In particular, ZHPZ^{\mathrm{HP}} is a true martingale when the unweighted initial second moment is finite. We now prove both conditional identities at once. Take B\mathcal{B} to be the whole sample space in the positive case and B=BK={R(Φ0)≤K}\mathcal{B}=\mathcal{B}_{K}=\{R(\Phi_{0})\le K\} in the Hua–Pickrell case. For A∈Fr\mathcal{A}\in\mathcal{F}_{r} and 0≤q≤T0\le q\le T, put

U(q,v):=E[1A1BHnH(Φr+q,v)].U(q,\mathbf{v}) \mathrel{\mathop:}= \mathbb{E}\left[1_{\mathcal{A}}1_{\mathcal{B}}H_{n}^{\mathsf{H}}(\boldsymbol{\Phi}_{r+q},\mathbf{v})\right].

The first two spatial derivatives of the probes are continuous in (g,v)(g,\mathbf{v}) by locally uniform convergence and Cauchy’s formula. The bounds above, and the event-weighted moment bound of Lemma 2.4 in the Hua–Pickrell case, justify differentiation under the expectation and continuity of these derivatives in (q,v)(q,\mathbf{v}) on compact subchambers. Taking expectations of the increment of the true-martingale Zt⋄Z_{t}^{\diamond} between rr and r+qr+q gives

U(q,v)=U(0,v)+∫0q(B⋄+C⋄)U(v,v) dv.(49)U(q,\mathbf{v})=U(0,\mathbf{v})+\int_{0}^{q}(\mathcal{B}_{\diamond}+\mathcal{C}_{\diamond})U(\mathbf{v},\mathbf{v})\,\mathrm{d}v. \tag*{(49)}

The integrand is continuous, so

U∈C1,2,∂qU=(B⋄+C⋄)U,∣U∣≤P(A∩B),(50)U\in C^{1,2},\qquad\partial_{q}U=(\mathcal{B}_{\diamond}+\mathcal{C}_{\diamond})U,\qquad|U|\le\mathbb{P}(\mathcal{A}\cap\mathcal{B}), \tag*{(50)}

locally up to the time endpoints. For fixed qq, backward Itô calculus applied to

e∫0vC⋄(Ss) dsU(q−v,Sv),0≤v≤q,e^{\int_{0}^{v}\mathcal{C}_{\diamond}(S_{s})\,\mathrm{d}s}U(q-v,S_{v}),\qquad0\le v\le q,

has zero drift. The displayed process is therefore a continuous local martingale on [0,q][0,q]. By (49) and ℜC⋄≤0\Re\mathcal{C}_{\diamond}\le0, its modulus is bounded by P(A∩B)=1\mathbb{P}(\mathcal{A}\cap\mathcal{B})=1, so it is a true martingale. Taking expectations at time qq gives

U(q,v)=Ev[e∫0qC⋄(Ss) dsU(0,Sq)].U(q,\mathbf{v})=\mathbb{E}_{\mathbf{v}}\left[e^{\int_{0}^{q}\mathcal{C}_{\diamond}(S_{s})\,\mathrm{d}s}U(0,S_{q})\right].

Fubini gives the desired identity tested against 1A1B1_{\mathcal{A}}1_{\mathcal{B}}. In the positive case this removes B\mathcal{B} immediately. For Hua–Pickrell, 1BK↑11_{\mathcal{B}_{K}}\uparrow1 and both tested integrands are bounded by one, so bounded convergence removes BK\mathcal{B}_{K}. Arbitrariness of A∈Fr\mathcal{A}\in\mathcal{F}_{r} proves both claims for every initial law. □\square

Remark 3.4 (On the choice and use of probes). The form of the reciprocal probes can be understood by considering, in the positive case, observables of the more general form ∏i=1nf(Φt(−si))\prod_{i=1}^{n}f(\Phi_{t}(-s_{i})). For n=1n=1, write

Gt(s)=Φt(−s),Ft(s)=f(Gt(s)),s>0,G_{t}(s)=\Phi_{t}(-s),\qquad F_{t}(s)=f(G_{t}(s)),\qquad s>0,

where f∈C2((0,∞))f\in C^{2}((0,\infty)). Using Qg(z,z)=z2(g′(z)2−g(z)g′′(z))Q_{g}(z,z)=z^{2}(g'(z)^{2}-g(z)g''(z)), the drift and covariance prescribed by the martingale problem give, by Itô’s formula,

drift⁡(dFt(s))=[12s2∂ssFt(s)+θ2s∂sFt(s)+κ2sGt(s)f′(Gt(s))−12s2(2f′(Gt(s))+Gt(s)f′′(Gt(s)))∂ssGt(s)]dt.\operatorname{drift}(\mathrm{d}F_{t}(s))=\left[\frac{1}{2}s^{2}\partial_{ss}F_{t}(s)+\frac{\theta}{2}s\partial_{s}F_{t}(s)+\frac{\kappa}{2}sG_{t}(s)f'(G_{t}(s))-\frac{1}{2}s^{2}\left(2f'(G_{t}(s))+G_{t}(s)f''(G_{t}(s))\right)\partial_{ss}G_{t}(s)\right]\mathrm{d}t.

Here the derivatives of ff are with respect to its scalar argument. For a general ff, taking expectations (assuming the required integrability is valid) introduces additional quantities through the last two terms. For example, f(u)=u−2f(u)=u^{-2} leaves the remainder −s2Gt(s)−3∂ssGt(s)-s^{2}G_{t}(s)^{-3}\partial_{ss}G_{t}(s), whose expectation must also be determined to describe the evolution of E[Ft(s)]\mathbb{E}[F_{t}(s)].

To obtain closure with the spatial operator G1θ=12s2∂ss+θ2s∂sG_1^\theta=\frac{1}{2}s^2\partial_{ss}+\frac{\theta}{2}s\partial_s, the derivative remainder vanishes when

uf′′(u)+2f′(u)=0.u f''(u)+2f'(u)=0.

The solutions are f(u)=c0+c1/uf(u)=c_0+c_1/u. Choosing the reciprocal part f(u)=u−1f(u)=u^{-1} also turns the term involving κ\kappa into multiplication by −κs/2-\kappa s/2. Thus, if we choose ht(s)=Gt(s)−1h_t(s)=G_t(s)^{-1}, we have

drift⁡(dht(s))=(G1θ−κ2s)ht(s) dt.\operatorname{drift}\bigl(\mathrm{d}h_t(s)\bigr)=\left(G_1^\theta-\frac{\kappa}{2}s\right)h_t(s)\,\mathrm{d}t.

The resulting operator acts only on the probe argument ss, with coefficients independent of the state Φt\boldsymbol{\Phi}_t. The bounds and martingale argument in the preceding proof therefore give a closed linear equation for the corresponding event-weighted expectations.

The diffusion interpretation provides a Feynman–Kac representation of this equation. In particular, for a deterministic state gg, put

Ψτ,s(g)=Es[exp⁡(−κ2∫0τSv dv)1g(−Sτ)],\Psi_{\tau,s}(g)=\mathbb{E}_s\left[\exp\left(-\frac{\kappa}{2}\int_0^\tau S_v\,\mathrm{d}v\right)\frac{1}{g(-S_\tau)}\right],

where SS has generator G1θG_1^\theta. The conditional duality identity is then (recall the notation hg+h_g^+)

E[hΦt+(s)∣Fr]=Ψt−r,s(Φr).\mathbb{E}\left[h_{\Phi_t}^+(s)\mid\mathcal{F}_r\right]=\Psi_{t-r,s}(\Phi_r).

The dual law and its weight are fixed independently of the original solution. Thus Ψτ,s\Psi_{\tau,s} gives a solution-independent prescription for the conditional expectation of each one-point probe, depending on the past only through Φr\Phi_r.

To determine the full conditional law, we use all finite products of reciprocal probes. In Itô’s product rule, the cross-covariations become the pairwise first-order interaction terms in the finite-dimensional dual generators, so the product observables also satisfy closed equations. Analogous arguments apply in the Hua–Pickrell case. For each model, taking all finite products at rational probe arguments, together with the constant function 11, gives the bounded determining class used in Proposition 3.5. The conditional dualities for these classes therefore yield uniqueness in law.

Determining classes and uniqueness

Conditional identities on a countable determining class are enough to construct a Borel transition kernel and identify the entire continuous path law.

Proposition 3.5. Let E\mathcal{E} be Polish and let M⊂Cb(E;C)\mathcal{M}\subset C_b(\mathcal{E};\mathbb{C}) be countable, unital, multiplicatively closed, self-adjoint, and separating. For each t≥0t\geq0 and F∈MF\in\mathcal{M}, fix a bounded Borel function PtF\mathcal{P}_tF on E\mathcal{E}. Suppose that, for every x∈Ex\in\mathcal{E}, there exists a continuous adapted E\mathcal{E}-valued process XX on a filtered probability space, with X0=xX_0=x almost surely, satisfying

E[F(Xt)∣Fr]=(Pt−rF)(Xr),F∈M,0≤r≤t.(51)\mathbb{E}\left[F(X_t)\mid\mathcal{F}_r\right]=(\mathcal{P}_{t-r}F)(X_r),\qquad F\in\mathcal{M},\quad0\leq r\leq t. \tag*{(51)}

Then there exists a Borel Markov semigroup such that every continuous adapted E\mathcal{E}-valued process satisfying (50) is Markov with this semigroup. Any two such processes with the same initial law have the same law on C(R+;E)C(\mathbb{R}_+;\mathcal{E}).

Proof. Enumerate M=(Fj)j≥1\mathcal{M}=(F_j)_{j\geq1}. The continuous injection x↦(Fj(x))jx\mapsto(F_j(x))_j into CN\mathbb{C}^{\mathbb{N}} and Lusin–Souslin give σ(M)=B(E)\sigma(\mathcal{M})=\mathcal{B}(\mathcal{E}). The map

ι:P(E)⟶CN,ι(ρ)=(∫EFj dρ)j\iota:\mathcal{P}(\mathcal{E})\longrightarrow\mathbb{C}^{\mathbb{N}},\qquad\iota(\rho)=\left(\int_{\mathcal{E}}F_j\,\mathrm{d}\rho\right)_j

is Borel and injective: equality of its coordinates gives equality on the unital self-adjoint algebra spanned by M\mathcal{M}, and then on bounded Borel functions by the functional monotone-class theorem. Lusin–Souslin makes its range and inverse Borel. Existence and (51) at r=0r=0 show that vt(x)=((PtFj)(x))jv_t(x)=((\mathcal{P}_tF_j)(x))_j lies in this range, so Pt(x,⋅)=ι−1(vt(x))P_t(x,\cdot)=\iota^{-1}(v_t(x)) is a Borel probability kernel. The conditional identity extends to every bounded Borel function by linear combinations and bounded pointwise convergence, using conditional dominated convergence and the functional monotone-class theorem. Successive conditioning gives the semigroup law and all finite-dimensional distributions. Rational evaluations generate the Borel sigma-field of C(R+;E)C(\mathbb{R}_+;\mathcal{E}), proving path-law uniqueness. □

Proof of Theorem 1.2. Existence follows from [7 Theorems 1.4 and 1.6]. We verify the hypotheses of Proposition 3.5. Set Q+=Q>0\mathbb{Q}_+=\mathbb{Q}_{>0}, QHP=Q∖{0}\mathbb{Q}_{\mathrm{HP}}=\mathbb{Q}\setminus\{0\}, and, for ⋄∈{+,HP}\diamond\in\{+,\mathrm{HP}\}, let

M⋄:def={1}∪{Hn⋄(⋅,q):n≥1, q∈Q⋄n}.\mathcal{M}_{\diamond}\mathrel{\mathop:}^{\mathrm{def}}=\{1\}\cup\{H_n^{\diamond}(\cdot,\mathbf{q}):n\geq1,\ \mathbf{q}\in\mathbb{Q}_{\diamond}^{n}\}.

First suppose that q\mathbf{q} has pairwise distinct coordinates and write F=Hn⋄(⋅,q)F=H_n^{\diamond}(\cdot,\mathbf{q}). Define (PtF)(g)(\mathcal{P}_tF)(g) by the corresponding dual expectation in (41) or (42), with r=0r=0 and Φ0\boldsymbol{\Phi}_0 replaced by the deterministic state gg. Thus the expectation defining PtF\mathcal{P}_tF is over the dual alone. Unordered tuples are treated by increasing rearrangement, since the probes and the dual weights are invariant under relabelling. Put Pt1=1\mathcal{P}_t1=1. The probes and the exponential weights have modulus at most one, so ∣PtF∣≤1|\mathcal{P}_tF|\leq1; dominated convergence under the fixed dual law also shows that PtF\mathcal{P}_tF is continuous in gg.

Proposition 3.3 now gives, for every solution Φ\boldsymbol{\Phi} with any initial law,

E[F(Φt)∣Fr]=(Pt−rF)(Φr),0≤r≤t,\mathbb{E}\left[F(\boldsymbol{\Phi}_t)\mid\mathcal{F}_r\right]=(\mathcal{P}_{t-r}F)(\boldsymbol{\Phi}_r),\qquad0\leq r\leq t,

whenever FF is a probe with distinct arguments. This is precisely the conditional identity required in (50), initially for this subclass.

We next extend the definition and the identity to repeated arguments. Given q∈Q⋄n\mathbf{q}\in\mathbb{Q}_{\diamond}^{n}, choose q(m)→q\mathbf{q}^{(m)}\to\mathbf{q} with pairwise distinct coordinates and put

Fm=Hn⋄(⋅,q(m)),F=Hn⋄(⋅,q).F_m=H_n^{\diamond}(\cdot,\mathbf{q}^{(m)}),\qquad F=H_n^{\diamond}(\cdot,\mathbf{q}).

Then Fm(g)→F(g)F_m(g)\to F(g) for every state gg, and ∣Fm∣≤1|F_m|\leq1. Fix gg and choose any solution started from gg. Applying Proposition 3.3 at r=0r=0 and taking expectations gives

(PtFm)(g)=Eg[Fm(Φt)]⟶Eg[F(Φt)],(\mathcal{P}_tF_m)(g)=\mathbb{E}_g[F_m(\boldsymbol{\Phi}_t)]\longrightarrow\mathbb{E}_g[F(\boldsymbol{\Phi}_t)],

where Eg\mathbb{E}_g denotes expectation for the chosen solution and the convergence follows by bounded convergence. Consequently, we may define

(PtF)(g):def=lim⁡m→∞(PtFm)(g).(\mathcal{P}_tF)(g)\mathrel{\mathop:}^{\mathrm{def}}=\lim_{m\to\infty}(\mathcal{P}_tF_m)(g).

The preceding argument shows that this limit exists and is independent of the approximation, and that every solution from gg gives the same value. Only existence of a solution from gg is used here. The same argument identifies different tuples representing the same function. Hence PtF\mathcal{P}_tF is unambiguous, bounded by one in modulus, and Borel as a pointwise limit of Borel functions.

For an arbitrary solution and 0≤r≤t0 \le r \le t, bounded convergence gives

Fm(Φt)⟶F(Φt),(Pt−rFm)(Φr)⟶(Pt−rF)(Φr)in L1.F_m(\Phi_t) \longrightarrow F(\Phi_t), \qquad(\mathcal{P}_{t-r}F_m)(\Phi_r) \longrightarrow(\mathcal{P}_{t-r}F)(\Phi_r) \quad\text{in } L^1.

The L1L^1 contraction of conditional expectation therefore allows us to pass to the limit in the conditional identities supplied by Proposition 3.3. We obtain

E[F(Φt)∣Fr]=(Pt−rF)(Φr),F∈M⋄,0≤r≤t.\mathbb{E}[F(\Phi_t)\mid\mathcal{F}_r] = (\mathcal{P}_{t-r}F)(\Phi_r), \qquad F \in\mathcal{M}_{\diamond}, \quad0 \le r \le t.

Each class M⋄\mathcal{M}_{\diamond} consists of bounded continuous functions and is countable and unital. It is multiplicatively closed because multiplication concatenates the argument tuples, with repetitions allowed. It is self-adjoint because positive probes are real and hgHP(−q)=hgHP(q)h_g^{\mathrm{HP}}(-q)=h_g^{\mathrm{HP}}(q). If two states have equal one-point probes, their entire functions agree on the negative rational axis or the nonzero imaginary rational axis; the identity theorem therefore gives separation. Consequently,

σ(M+)=B(E+),σ(MHP)=B(E).(52)\sigma(\mathcal{M}_{+})=\mathcal{B}(\mathcal{E}_{+}), \qquad\sigma(\mathcal{M}_{\mathrm{HP}})=\mathcal{B}(\mathcal{E}). \tag*{(52)}

Since the state spaces are Polish and solutions exist from every deterministic state, all the hypotheses of Proposition 3.5 are satisfied. That proposition proves uniqueness in law, in fact for every initial law, and supplies the Borel Markov semigroups. □

Relation to the boundary intertwinings

The reciprocal dualities are related directly to the intertwinings used to construct the boundary processes in [4, 6, 7] in the first place. We now explain how.

Identify the boundaries with the entire-function spaces of Section 2. For either model, write Pt⋄\mathsf{P}^{\diamond}_t for its boundary semigroup, Pn,t⋄\mathsf{P}^{\diamond}_{n,t} for the original finite-dimensional semigroup, and Λn∞\Lambda_n^\infty for the boundary-to-corner eigenvalue kernel [4, 6]. Let Wn⋄\mathcal{W}^{\diamond}_n denote the corresponding closed positive or real Weyl chamber, with increasing coordinates. The boundary construction gives Pt⋄Λn∞=Λn∞Pn,t⋄\mathsf{P}^{\diamond}_t\Lambda_n^\infty=\Lambda_n^\infty\mathsf{P}^{\diamond}_{n,t}, for every nn and t≥0t\ge0.

Let Dv=diag⁡(v1,…,vn)D_v=\operatorname{diag}(v_1,\ldots,v_n), put ε+=−1\varepsilon_+=-1, εHP=i\varepsilon_{\mathrm{HP}}=i, and define the orbital Laplace and Fourier kernels, with normalized Haar measure, by

Kn⋄(x,v)=∫U(n)eε⋄Tr⁡(DvUDxU∗) dU.\mathcal{K}^{\diamond}_n(x,v)=\int_{\mathcal{U}(n)}e^{\varepsilon_{\diamond}\operatorname{Tr}(D_vUD_xU^*)}\,dU.

Here v=s∈Wnv=s\in\mathcal{W}_n in the positive case and v=uv=u belongs to a signed chamber in the Hua–Pickrell case. For fixed gg, the Gaussian matrix representation of Λn∞\Lambda_n^\infty [6 Definitions 2.17–2.18] has independent diagonal entries in its corner, with common law that of Yg+Y_g^+ in (29), or YgY_g in Lemma 2.2. Their transforms and unitary invariance therefore give

Hn⋄(g,v)=∫Wn⋄Kn⋄(x,v) Λn∞(g,dx).H_n^\diamond(g,v)=\int_{\mathcal{W}^{\diamond}_n}\mathcal{K}^{\diamond}_n(x,v)\,\Lambda_n^\infty(g,dx).

Thus the reciprocal product is the joint transform of the diagonal entries of the corner. The orbital average is essential when working with its eigenvalues. Both kernels have modulus at most one, so the identity holds under arbitrary boundary laws without moment assumptions.

Write Dn⋄D_{n}^{\diamond} for the generator of Pn,t⋄P_{n,t}^{\diamond}. Conjugating it by the Vandermonde reduces it, up to a constant, to a sum of the scalar operators

x22∂x2+[(θ2−n+1)x+κ2]∂x,⋄=+,\frac{x^{2}}{2}\partial_{x}^{2}+\left[\left(\frac{\theta}{2}-n+1\right)x+\frac{\kappa}{2}\right]\partial_{x}, \qquad\diamond=+,
(1+x2)∂x2+[(2−2n−2a)x+2b]∂x,⋄=HP.(1+x^{2})\partial_{x}^{2}+\left[(2-2n-2a)x+2b\right]\partial_{x}, \qquad\diamond=\mathrm{HP}.

The HCIZ formula expresses Kn⋄K_{n}^{\diamond} as a constant times det⁡[eε⋄xivj]/[Δ(x)Δ(v)]\det[e^{\varepsilon_{\diamond}x_{i}v_{j}}]/[\Delta(x)\Delta(v)]. Expanding the determinant, the quadratic and linear homogeneous terms exchange xx and vv. The remaining terms give −κ∑i=1nsi/2-\kappa\sum_{i=1}^{n}s_{i}/2 and 2ib∑i=1nui−∑i=1nui22ib\sum_{i=1}^{n}u_{i}-\sum_{i=1}^{n}u_{i}^{2}, respectively. Undoing the Vandermonde conjugation cancels the dimension shifts and gives

Dn,x⋄Kn⋄=(Bn,v⋄+C⋄(v))Kn⋄,Bn+=Gnθ,C+=−κVn+,BnHP=Ana,CHP=Cb,n.D_{n,x}^{\diamond}K_{n}^{\diamond}=\left(B_{n,v}^{\diamond}+C_{\diamond}(v)\right)K_{n}^{\diamond}, \qquad \begin{aligned} B_{n}^{+}&=G_{n}^{\theta}, & C_{+}&=-\kappa V_{n}^{+},\\ B_{n}^{\mathrm{HP}}&=A_{n}^{a}, & C_{\mathrm{HP}}&=C_{b,n}. \end{aligned}

This generator identity lifts to the finite semigroups. From deterministic finite initial states the positive trace and the Hua–Pickrell squared norm have all polynomial moments on bounded time intervals. Derivatives of the orbital kernels have polynomial growth, and symmetry removes the apparent collision poles. Localized Itô formulas and these moment bounds justify the finite duality. Feller continuity extends it to closed-chamber initial states. Writing Tn,t⋄T_{n,t}^{\diamond} for the dual Feynman–Kac operator, this gives

Pn,t⋄Kn⋄(⋅,v)(x)=Tn,t⋄Kn⋄(x,⋅)(v).P_{n,t}^{\diamond}K_{n}^{\diamond}(\cdot,v)(x)=T_{n,t}^{\diamond}K_{n}^{\diamond}(x,\cdot)(v).

Consequently, composing with the boundary intertwining yields the duality

Pt⋄Hn⋄(⋅,v)(g)=Λn∞Pn,t⋄Kn⋄(⋅,v)(g)=Tn,t⋄Hn⋄(g,⋅)(v).P_{t}^{\diamond}H_{n}^{\diamond}(\cdot,v)(g)=\Lambda_{n}^{\infty}P_{n,t}^{\diamond}K_{n}^{\diamond}(\cdot,v)(g)=T_{n,t}^{\diamond}H_{n}^{\diamond}(g,\cdot)(v).

Conversely, the orbital transforms determine finite eigenvalue laws: their unitary-invariant matrix lifts are determined by their Laplace or Fourier transforms. Hence the reciprocal dualities, for all nn, recover the boundary intertwinings for any specified boundary Markov semigroup.

School of Mathematics, University of Edinburgh, James Clerk Maxwell Building, Peter Guthrie Tait Rd, Edinburgh EH9 3FD, U.K. [email protected]

School of Mathematics, University of Edinburgh, James Clerk Maxwell Building, Peter Guthrie Tait Rd, Edinburgh EH9 3FD, U.K. [email protected]

References

  1. [1]A. Ahn, Extremal singular values of random matrix products and Brownian motion on ℳℲ(N,C), Probability Theory and Related Fields 187 (2023), 949–997, doi:10.1007/s00440-023-01217-5.DOI
  2. [2]J. Arista and N. Demni, Explicit expressions of the Hua–Pickrell semigroup, Theory of Probability and Its Applications 67 (2022), no. 2, 208–228, doi:10.1137/S0040585X97T990885.DOI
  3. [3]T. Assiotis, A matrix Bougerol identity and the Hua–Pickrell measures, Electronic Communications in Probability 23 (2018), 1–11, Paper no. 7. doi:10.1214/18-ECP107.DOI
  4. [4]—————, Hua–Pickrell diffusions and Feller processes on the boundary of the graph of spectra, Annales de l’Institut Henri Poincaré, Probabilités et Statistiques 56 (2020), 1251–1283.arxiv.org/abs/1703.01813
  5. [5]—————, Random entire functions from random polynomials with real zeros, Advances in Mathematics 410 (2022), 108701, doi:10.1016/j.aim.2022.108701.DOI
  6. [6]T. Assiotis and Z. S. Mirsajjadi, ISDE with logarithmic interaction and characteristic polynomials, arXiv:2408.00717, 2024.arxiv.org/abs/2408.00717
  7. [7]———, 𝒢 ℒ_N(ℂ) Brownian motion and stochastic PDE on entire functions, arXiv:2605.06429, 2026.arxiv.org/abs/2605.06429
  8. [8]Theodoros Assiotis and Fengyi Li, Infinite-dimensional dyson brownian motion (s), arXiv preprint arXiv:2609.26386 (2026).arxiv.org/abs/2609.26386
  9. [9]A. Borodin and G. Olshanski, Infinite random matrices and ergodic measures, Communications in Mathematical Physics 223 (2001), 87–123.
  10. [10]———, Markov processes on the path space of the Gelfand–Tsetlin graph and on its boundary, Journal of Functional Analysis 263 (2012), no. 1, 248–303, doi:10.1016/j.jfa.2012.03.018.DOI
  11. [11]———, The boundary of the Gelfand–Tsetlin graph: a new approach, Advances in Mathematics 230 (2012), no. 4–6, 1738–1779, doi:10.1016/j.aim.2012.04.005.DOI
  12. [12]———, Markov dynamics on the Thoma cone: a model of time-dependent determinantal processes with infinitely many particles, Electronic Journal of Probability 18 (2013), 1–43, Paper no. 75. doi:10.1214/EJP.v18-2729.DOI
  13. [13]R. Chhaibi, J. Najnudel, and A. Nikeghbali, The Circular Unitary Ensemble and the Riemann zeta function: the microscopic landscape and a new approach to ratios, Inventiones Mathematicae 207 (2017), no. 1, 23–113, doi:10.1007/s00222-016-0669-1.DOI
  14. [14]J. L. Doob, Classical Potential Theory and Its Probabilistic Counterpart, Grundlehren der mathematischen Wissenschaften, vol. 262, Springer-Verlag, New York, 1984, doi:10.1007/978-1-4612-5208-5.DOI
  15. [15]P. J. Forrester, Log-Gases and Random Matrices, London Mathematical Society Monographs Series, vol. 34, Princeton University Press, Princeton, NJ, 2010.DOI
  16. [16]Peter J. Forrester, Meet Andréief, Bordeaux 1886, and Andreev, Kharkov 1882–1883, Random Matrices: Theory and Applications 8 (2019), no. 2, 1930001.DOI
  17. [17]Martin Hairer, An introduction to stochastic PDEs, Lecture notes, version of November 10, 2025, 2025, https://www.hairer.org/notes/SPDEs.pdf.
  18. [18]L. Jones and N. O’Connell, Weyl chambers, symmetric spaces and number variance saturation, ALEA. Latin American Journal of Probability and Mathematical Statistics 2 (2006), 91–118.
  19. [19]Samuel Karlin and James McGregor, Coincidence probabilities, Pacific Journal of Mathematics 9 (1959), no. 4, 1141–1164.DOI
  20. [20]M. Katori and H. Tanemura, Non-equilibrium dynamics of Dyson’s model with an infinite number of particles, Communications in Mathematical Physics 293 (2010), 469–497, doi:10.1007/s00220-009-0912-3.DOI
  21. [21]Y. Li and B. Valkó, Operator level limit of the circular Jacobi β-ensemble, Random Matrices: Theory and Applications 11 (2022), no. 4, 2250043, doi:10.1142/S2010326322500435.
  22. [22]Zahra Sadat Mirsajjadi, Infinite-Particle Determinantal Dynamics from Brownian Motion on GL_N(ℂ), arXiv:2609.35383, 2026.arxiv.org/abs/2609.35383
  23. [23]J. R. Norris, L. C. G. Rogers, and D. Williams, Brownian motions of ellipsoids, Transactions of the American Mathematical Society 294 (1986), no. 2, 757–765, doi:10.2307/2000214.DOI
  24. [24]Martin Ondrejat, Brownian representations of cylindrical local martingales, martingale problem and strong Markov property of weak solutions of SPDEs in Banach spaces, Czechoslovak Mathematical Journal 55 (2005), no. 4, 1003–1039, doi:10.1007/s10587-005-0084-z.
  25. [25]H. Osada, Infinite-dimensional stochastic differential equations related to random matrices, Probability Theory and Related Fields 153 (2012), 471–509, doi:10.1007/s00440-011-0352-9.
  26. [26]H. Osada and H. Tanemura, Infinite-dimensional stochastic differential equations and tail σ-fields, Probability Theory and Related Fields 177 (2020), 1137–1242, doi:10.1007/s00440-020-00981-y.
  27. [27]Y. Qiu, Infinite random matrices and ergodic decomposition of finite or infinite Hua–Pickrell measures, Advances in Mathematics 308 (2017), 1209–1268.
  28. [28]D. Revuz and M. Yor, Continuous Martingales and Brownian Motion, third ed., Grundlehren der mathematischen Wissenschaften, vol. 293, Springer-Verlag, Berlin, 1999, doi:10.1007/978-3-662-06400-9.DOI
  29. [29]B. Rider and B. Valkó, Matrix Dufresne identity, International Mathematics Research Notices 2016 (2016), no. 1, 174–218, doi:10.1093/imrn/rnv127.
  30. [30]K. Suzuki, The infinite Dyson Brownian motion with β = 2 does not have a spectral gap, Bulletin of the London Mathematical Society 57 (2025), no. 2, 426–431, doi:10.1112/blms.13204.
  31. [31]C. A. Tracy and H. Widom, Level-spacing distributions and the Bessel kernel, Communications in Mathematical Physics 161 (1994), 289–309.
  32. [32]L.-C. Tsai, Infinite dimensional stochastic differential equations for Dyson’s model, Probability Theory and Related Fields 166 (2016), 801–850, doi:10.1007/s00440-015-0672-2.
  33. [33]B. Valkó and B. Virág, The many faces of the stochastic zeta function, Geometric and Functional Analysis 32 (2022), no. 5, 1160–1231, doi:10.1007/s00039-022-00613-8.

Paper details

vlnc:W_urrx27nt
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 1

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents