Bessel Point Process: Hard-Edge Scaling in Random Matrices

• Bessel point process is a determinantal point process on (0, ∞) defined via the Bessel kernel with parameter α > -1, arising as the hard-edge scaling limit in random matrix ensembles.
• It exhibits rich integrable structures, featuring explicit Fredholm determinant formulations, Riemann–Hilbert problems, and Painlevé-type equations.
• Rigidity, conditional measures, and Gaussian asymptotics in counting statistics are rigorously derived, highlighting its deep mathematical and physical implications.

The Bessel point process is the determinantal point process on $(0,\infty)$ governed by the Bessel kernel with parameter $\alpha>-1$; the same parameter is often denoted by $\nu$ or $a$ in the literature. It is the hard-edge scaling limit for eigenvalues near the origin in Laguerre or Wishart ensembles, and it is distinguished by projection-kernel determinantal correlations, rigidity, explicit conditional measures, and a substantial integrable-systems structure involving Fredholm determinants, Riemann–Hilbert problems, and Painlevé-type equations (Bufetov, 2015, Wu, 2021, Ruzza, 2024).

1. Definition and determinantal structure

Let $\mathbb R_+=(0,\infty)$ with Lebesgue measure. A standard form of the Bessel kernel is

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$

where $J_\alpha$ is the Bessel function of the first kind of order $\alpha$. An equivalent Tracy–Widom form is

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$

By the Macchi–Soshnikov theorem, the associated integral operator on $L^2(\mathbb R_+,dx)$ is an orthogonal projection of locally trace-class, and therefore defines a unique determinantal point process on $\alpha>-1$0 (Bufetov, 2015).

Its $\alpha>-1$1-point correlation functions are

$\alpha>-1$2

in the sense that for test functions on $\alpha>-1$3 the joint intensities are given by determinants of the kernel. Equivalent integral representations also appear. One common hard-edge form is

$\alpha>-1$4

while a scaled representation used in line-ensemble and field formulations is

$\alpha>-1$5

up to an innocuous gauge, or

$\alpha>-1$6

in a different normalization (Molag et al., 2019, Wu, 2021, Benigni et al., 2021).

The determinantal formulation immediately yields Fredholm-determinant expressions for gap events. For example, the probability of no points in a bounded set such as $\alpha>-1$7 is expressed as the Fredholm determinant of the cut-down Bessel operator on $\alpha>-1$8 (Gorbunov, 2024).

2. Hard-edge origin and dynamical realizations

The principal random-matrix origin is the Laguerre Unitary Ensemble. If $\alpha>-1$9 is an $\nu$0 complex Brownian matrix at time $\nu$1, then $\nu$2 has the Wishart law, and its ordered eigenvalues $\nu$3 have joint density

$\nu$4

Under hard-edge scaling near the smallest eigenvalues, one lets $\nu$5 and sets $\nu$6, with time scaled as $\nu$7. The resulting point process converges to the Bessel point process of index $\nu$8 (Wu, 2021).

A dynamical realization is given by the Dyson Bessel process: one starts with $\nu$9 independent squared Bessel processes of index $a$0, begun at $a$1 and conditioned never to collide. Under the same hard-edge scaling near time $a$2, the top curves converge to a countable Bessel line ensemble whose one-time marginal is the Bessel point process. Its finite-dimensional distributions are governed by the extended Bessel kernel, and the limiting ensemble is stationary under horizontal translations because the extended kernel depends on time only through $a$3 in the exponential factor (Wu, 2021).

A broader two-parameter object is the Bessel field $a$4. For fixed $a$5 it is the classical Bessel point process. Along any time-like path or space-like path in the $a$6 plane, the joint configuration is determinantal with an explicit extended kernel built from Bessel functions and exponential propagation factors. For fixed $a$7, the field in the $a$8-direction is described as an exponential Gibbsian line ensemble; for fixed $a$9, the $\mathbb R_+=(0,\infty)$0-evolution is a squared Bessel Gibbsian line ensemble (Benigni et al., 2021).

3. Rigidity and conditional measures

A point process on $\mathbb R_+=(0,\infty)$1 is called rigid if for every bounded Borel set $\mathbb R_+=(0,\infty)$2 the counting variable

$\mathbb R_+=(0,\infty)$3

is almost surely determined by the restriction of the configuration to $\mathbb R_+=(0,\infty)$4. For every $\mathbb R_+=(0,\infty)$5, the Bessel point process is rigid. In particular, if $\mathbb R_+=(0,\infty)$6, then the number of points in $\mathbb R_+=(0,\infty)$7 is almost surely a measurable function of the configuration outside $\mathbb R_+=(0,\infty)$8 (Bufetov, 2015).

The proof uses additive statistics approximating the indicator of $\mathbb R_+=(0,\infty)$9. For an increasing sequence $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$0, define cutoff functions

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$1

and then set

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$2

By construction, $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$3 on $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$4, $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$5, and $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$6. For bounded $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$7 one has

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$8

Using the known small-argument and large-argument asymptotics of $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_\alpha'(\sqrt y)-\sqrt x\,J_\alpha'(\sqrt x)\,J_\alpha(\sqrt y)}{2(x-y)}, \qquad \alpha>-1,$$9, together with the facts that for $J_\alpha$0 the Bessel kernel decays like $J_\alpha$1 and that for $J_\alpha$2 or $J_\alpha$3 small the kernel is integrable, one obtains $J_\alpha$4, and the Ghosh–Peres criterion then yields rigidity (Bufetov, 2015).

Rigidity has a stronger conditional consequence. For a realization $J_\alpha$5 of the Bessel process and $J_\alpha$6, write $J_\alpha$7. The conditional distribution of the points inside $J_\alpha$8, given the configuration outside $J_\alpha$9, is almost surely an orthogonal polynomial ensemble on $\alpha$0 with exactly $\alpha$1 points and weight

$\alpha$2

Its joint density for ordered points $\alpha$3 is

$\alpha$4

so the conditional law is again determinantal, now with a Christoffel–Darboux kernel (Molag et al., 2019).

Molag and Stevens proved a universality statement for these conditional measures. If $\alpha$5 is any strictly increasing sequence in $\alpha$6 satisfying

$\alpha$7

then for every compact $\alpha$8,

$\alpha$9

uniformly on $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$0. They also verify, via Soshnikov’s variance estimates, that a random Bessel configuration almost surely satisfies

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$1

so almost every conditional orthogonal polynomial ensemble recovers the Bessel point process as $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$2 (Molag et al., 2019).

4. Fluctuations, counting statistics, and large gaps

For the counting function

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$3

exponential-moment asymptotics are explicit. The moment-generating function

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$4

satisfies

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$5

where

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$6

and $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$7 is Barnes’ $$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$8-function. Consequently,

$$K^{\mathrm{Bessel}_\alpha}(x,y) = \frac{J_\alpha(\sqrt x)\,\sqrt y\,J_{\alpha+1}(\sqrt y)-J_\alpha(\sqrt y)\,\sqrt x\,J_{\alpha+1}(\sqrt x)}{2(x-y)}.$$9

and

$L^2(\mathbb R_+,dx)$0

After normalization by the variance, $L^2(\mathbb R_+,dx)$1 converges in distribution to $L^2(\mathbb R_+,dx)$2 (Charlier, 2018).

More general linear statistics also admit quantitative Gaussian asymptotics. For compactly supported $L^2(\mathbb R_+,dx)$3, define

$L^2(\mathbb R_+,dx)$4

Under the centering condition $L^2(\mathbb R_+,dx)$5, one has

$L^2(\mathbb R_+,dx)$6

and, more precisely, if $L^2(\mathbb R_+,dx)$7 denotes the distribution function of

$L^2(\mathbb R_+,dx)$8

then

$L^2(\mathbb R_+,dx)$9

for all $\alpha>-1$00, where $\alpha>-1$01 is the standard normal distribution function (Gorbunov, 2024).

Large-gap asymptotics are likewise explicit. For disjoint consecutive intervals $\alpha>-1$02, piecewise thinning with retention probabilities $\alpha>-1$03 yields a thinned process whose joint gap probability is again a Fredholm determinant. Writing $\alpha>-1$04 with $\alpha>-1$05, the exponential moment

$\alpha>-1$06

has an asymptotic expansion involving

$\alpha>-1$07

together with Barnes $\alpha>-1$08-factors and a quadratic form in the $\alpha>-1$09 (Charlier, 2018).

For unions of widely separated intervals, large-$\alpha>-1$10 gap probabilities acquire a theta-function structure. If

$\alpha>-1$11

then

$\alpha>-1$12

As $\alpha>-1$13, $\alpha>-1$14 admits an expansion of the form

$\alpha>-1$15

where the coefficients are described through a genus-$\alpha>-1$16 Riemann surface, normalized holomorphic one-forms, and a linear flow on a $\alpha>-1$17-dimensional torus. In the ergodic case, Birkhoff’s ergodic theorem identifies the leading logarithmic term explicitly (Blackstone et al., 2021).

5. Fredholm determinants, Riemann–Hilbert problems, and Painlevé structures

A central integrable observable is the joint generating function of occupancy numbers on disjoint intervals. For

$\alpha>-1$18

set

$\alpha>-1$19

This function is a Fredholm determinant,

$\alpha>-1$20

with deformed kernel

$\alpha>-1$21

Charlier and Doeraene derive a $\alpha>-1$22 Riemann–Hilbert problem with simple poles at $\alpha>-1$23; from its Lax pair they obtain a system of $\alpha>-1$24 coupled Painlevé V equations for functions $\alpha>-1$25. Their main representation writes

$\alpha>-1$26

with hard-edge boundary conditions

$\alpha>-1$27

For $\alpha>-1$28, this reduces to the Tracy–Widom description of the one-interval Bessel gap probability in terms of a single Painlevé V equation (Charlier et al., 2017).

A more general framework treats multiplicative statistics depending on two parameters. For a bounded measurable weight $\alpha>-1$29, define

$\alpha>-1$30

Introduce

$\alpha>-1$31

Then $\alpha>-1$32 satisfies the nonlinear integrable PDE

$\alpha>-1$33

This equation is equivalent to the compatibility of the linear system

$\alpha>-1$34

with

$\alpha>-1$35

or, in Lax form,

$\alpha>-1$36

Thus the $\alpha>-1$37-deformation is isospectral for a Sturm–Liouville problem. When $\alpha>-1$38 is specialized to a step function, the PDE reduces to the coupled Painlevé V system above; in the single-interval case it reduces to the Tracy–Widom Painlevé V equation (Ruzza, 2024).

This integrable description is also compatible with the Its–Izergin–Korepin–Slavnov theory of integrable operators and a Bessel-model Riemann–Hilbert problem. The resulting identities connect Fredholm determinants, boundary-value problems for the associated Sturm–Liouville equation, and nonlocal or coupled Painlevé-type reductions (Ruzza, 2024).

6. Related processes, limits, and variants

The infinite Bessel process arises in the study of infinite Pickrell measures. For $\alpha>-1$39, one defines a modified kernel

$\alpha>-1$40

equivalently written in a Tracy–Widom-type Christoffel–Darboux form. The corresponding determinantal process is obtained as the hard-edge scaling limit of radial parts of finite-dimensional Pickrell measures after the rescaling $\alpha>-1$41. Different parameters give mutually singular laws, and the ergodic decomposition analysis shows that the additional “Gaussian parameter” vanishes almost surely (Bufetov, 2016).

A distinct but related object is the finite-temperature deformation of the discrete Bessel point process on

$\alpha>-1$42

For a decay weight $\alpha>-1$43 and $\alpha>-1$44, the kernel

$\alpha>-1$45

defines a determinantal point process on $\alpha>-1$46. Its largest-particle distribution

$\alpha>-1$47

satisfies a reduction of the $\alpha>-1$48D Toda hierarchy,

$\alpha>-1$49

together with a discrete integro-differential Painlevé II system. Under a formal continuum scaling, this deformation converges to the finite-temperature Airy framework and its KdV/integro-Painlevé II structure (Cafasso et al., 2022).

Another direction replaces the determinantal kernel by a random differential operator. For the stochastic Bessel operator

$\alpha>-1$50

with random weights

$\alpha>-1$51

the low-lying eigenvalues define a point process at the hard edge. In the high-temperature limit $\alpha>-1$52, the rescaled eigenvalues

$\alpha>-1$53

converge to a limiting point process $\alpha>-1$54 characterized through coupled stochastic differential equations and an alternating reflected Brownian-motion construction. No closed-form joint density, Fredholm determinant, or Pfaffian kernel is known for $\alpha>-1$55, and the limiting process is not Poisson (Magaldi, 2024).

These variants clarify a structural boundary. The classical $\alpha>-1$56 hard-edge Bessel point process is determinantal and admits exact Fredholm and Painlevé descriptions; discrete, finite-temperature, infinite-dimensional, and high-temperature analogues preserve some of its hard-edge or Bessel-function features, but they generally modify the kernel structure, the integrable hierarchy, or the determinantal character itself (Cafasso et al., 2022, Magaldi, 2024).