Random Bergman Metrics

• Random Bergman metrics are probability-theoretic models on finite-dimensional Bergman spaces that approximate the infinite-dimensional space of Kähler metrics.
• They are constructed using eigenvalue ensembles, heat kernel measures, and action-weighted measures to capture geometric fluctuations on compact Kähler manifolds.
• This framework connects Kähler quantization, random matrix theory, and large deviations, offering explicit solvable models and critical stability thresholds.

Searching arXiv for recent and foundational papers on random Bergman metrics and closely related Bergman-kernel probabilistic geometry. Random Bergman metrics are probability-theoretic models on the finite-dimensional spaces of Bergman metrics that approximate the infinite-dimensional space of Kähler metrics in a fixed Kähler class. On a compact Kähler manifold $(M,\omega_0)$ with a positive holomorphic line bundle $L\to M$, the basic construction replaces formal path integrals on

$$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$

by matrix integrals on symmetric spaces of Bergman metrics, typically

$$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$

Randomness is introduced by choosing probability measures on the positive Hermitian matrices that parametrize $\mathcal B_k$, or by weighting the invariant Haar measure with geometric action functionals. This framework connects Kähler quantization, Bergman kernel asymptotics, random matrix models, large deviations, and geometric stability (Ferrari et al., 2011, Ferrari et al., 2011, Klevtsov et al., 2014, Shiffman et al., 2023).

1. Geometric definition and finite-dimensional model

Let $(M,\omega_0)$ be a compact Kähler manifold of complex dimension $n$, with $[\omega_0]\in H^{1,1}(M,\mathbb{Z})$. Then there is a holomorphic line bundle $L\to M$ with $c_1(L)=[\omega_0]$, and for each $L\to M$0 one considers the finite-dimensional space of holomorphic sections

$L\to M$1

Given an orthonormal basis of sections $L\to M$2, a Bergman metric is defined from a positive Hermitian matrix $L\to M$3 by the Kähler potential

$L\to M$4

and the associated Kähler form is

$L\to M$5

Since only the projective class of $L\to M$6 matters, one modds out by overall scalars, and the space of Bergman metrics is identified with the symmetric space $L\to M$7. In the survey formulation, the same space is described as

$L\to M$8

where $L\to M$9 is the space of positive definite Hermitian $$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$0 matrices with determinant one (Ferrari et al., 2011, Shiffman et al., 2023).

The parameterization is basis-independent in the precise sense that, for a fixed orthonormal basis $$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$1, any other basis can be written as

$$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$2

and the corresponding metric depends only on

$$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$3

Equivalently, Bergman metrics arise as pullbacks of the Fubini–Study form under Kodaira embeddings, for instance

$$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$4

The finite-dimensional spaces $$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$5 or $$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$6 serve as approximants to the full space of Kähler metrics. The key approximation statements are

$$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$7

and

$$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$8

By the Tian–Yau–Zelditch expansion, the canonical Bergman metric corresponding to the identity matrix satisfies

$$K([\omega_0])=\{\phi\in C^\infty(M)/\mathbb{R}\mid \omega_\phi:=\omega_0+i\partial\bar\partial \phi>0\}$$9

so Bergman metrics approximate arbitrary Kähler metrics increasingly well as $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$0 (Ferrari et al., 2011, Ferrari et al., 2011, Shiffman et al., 2023).

2. Measures on Bergman metric spaces

A random Bergman metric is obtained by choosing a probability measure on the positive Hermitian matrix $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$1. One class consists of unitarily invariant or eigenvalue-type measures

$$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$2

where $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$3, $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$4, and $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$5 is the invariant measure on positive Hermitian matrices. The ensemble is gauge-fixed to the Bergman metric space by enforcing $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$6, giving a measure $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$7 on $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$8 (Ferrari et al., 2011).

A second construction uses heat kernel measures on the symmetric space of Bergman metrics: $$\mathcal B_k \simeq SL(N_k,\mathbb C)/SU(N_k),\qquad N_k=\dim H^0(M,L^k).$$9 where $\mathcal B_k$0 is the heat kernel on $\mathcal B_k$1 and $\mathcal B_k$2 is Haar measure. This measure is invariant under $\mathcal B_k$3, independent of the choice of orthonormal basis, and interpreted as Brownian motion on $\mathcal B_k$4 starting at the identity (Shiffman et al., 2023).

A third construction weights the invariant Haar measure by a geometric action functional. The basic partition function is

$\mathcal B_k$5

where $\mathcal B_k$6 is the invariant Haar measure on $\mathcal B_k$7, $\mathcal B_k$8 is a coupling constant, and $\mathcal B_k$9 is a geometric action functional. Because the Haar volume is infinite and grows rapidly along noncompact geodesic directions, convergence of $(M,\omega_0)$0 becomes a central issue (Klevtsov et al., 2014).

Construction	Measure	Structural feature
Eigenvalue-type ensemble	$(M,\omega_0)$1	Unitary invariance
Wishart ensemble	$(M,\omega_0)$2	Exactly solvable positive-matrix model
Heat kernel ensemble	$(M,\omega_0)$3	Brownian motion on $(M,\omega_0)$4
Action-weighted ensemble	$(M,\omega_0)$5	Geometric partition function

These constructions are complementary rather than competing. The matrix-model viewpoint is used both as a regularization of path integrals over Kähler metrics and as a source of explicit probabilistic models on $(M,\omega_0)$6 (Ferrari et al., 2011, Ferrari et al., 2011, Klevtsov et al., 2014).

3. Observables, correlation functions, and solvable models

The basic observables are correlators of the Bergman potential

$(M,\omega_0)$7

and of the metric

$(M,\omega_0)$8

The $(M,\omega_0)$9-point functions are defined schematically by

$n$0

Because $n$1 depends on the directions of the section vectors $n$2, these correlators require nontrivial angular integration over $n$3; the Harish-Chandra–Itzykson–Zuber formula is used to perform this integration exactly (Ferrari et al., 2011).

For any eigenvalue-type measure, the one-point function is universal: $n$4 Thus the mean random Bergman metric converges to the background metric. The two-point function carries the nontrivial fluctuation data. After angular integration, all dependence on the points $n$5 enters through the normalized off-diagonal Bergman kernel

$n$6

where $n$7 is the Calabi diastatic function of the background Bergman metric. In the large-$n$8 limit, the general structure is

$n$9

Away from the diagonal the correlator factorizes, while near the diagonal there is a contact term reflecting short-distance metric fluctuations (Ferrari et al., 2011).

The Wishart model is the main solvable example: $[\omega_0]\in H^{1,1}(M,\mathbb{Z})$0 It allows explicit finite-$[\omega_0]\in H^{1,1}(M,\mathbb{Z})$1 formulas for correlators. In particular, the two-point function can be written in terms of a hypergeometric function involving

$[\omega_0]\in H^{1,1}(M,\mathbb{Z})$2

and the final finite-$[\omega_0]\in H^{1,1}(M,\mathbb{Z})$3 expression depends on the geometry only through $[\omega_0]\in H^{1,1}(M,\mathbb{Z})$4. In the off-diagonal regime $[\omega_0]\in H^{1,1}(M,\mathbb{Z})$5,

$[\omega_0]\in H^{1,1}(M,\mathbb{Z})$6

with corrections suppressed as $[\omega_0]\in H^{1,1}(M,\mathbb{Z})$7. For the standard scaling of the Wishart parameter $[\omega_0]\in H^{1,1}(M,\mathbb{Z})$8, the contact terms vanish in the large-$[\omega_0]\in H^{1,1}(M,\mathbb{Z})$9 limit, and the ensemble becomes sharply concentrated around the background metric (Ferrari et al., 2011).

A recurrent limitation is that, because the random variable is $L\to M$0 itself, the resulting measures may become too sharply concentrated at large $L\to M$1. The suggestion that a smoother and more physically interesting geometry may arise if one uses $L\to M$2 as the random variable identifies a technical issue in simple matrix-model ensembles rather than a contradiction in the framework (Ferrari et al., 2011).

4. Stability functionals and the critical coupling constant

A distinct line of development studies partition functions on $L\to M$3 with actions drawn from geometric functionals called stability functions. In the GIT setting, a stability function is a function on $L\to M$4 whose gradient is the moment map, written abstractly as

$L\to M$5

When restricted to a geodesic $L\to M$6 in $L\to M$7, a stability function is asymptotically linear: $L\to M$8 where $L\to M$9 is the asymptotic slope and $c_1(L)=[\omega_0]$0 is the $c_1(L)=[\omega_0]$1-intercept. In the stable case, the slope is positive along all geodesic rays, so the action gives a “V-shaped” confining potential. The worst geodesic rays are those with the smallest asymptotic slope, and these least stable directions determine the threshold for convergence of the partition function (Klevtsov et al., 2014).

The corresponding invariant is the critical coupling constant

$c_1(L)=[\omega_0]$2

If $c_1(L)=[\omega_0]$3, the partition function diverges; if $c_1(L)=[\omega_0]$4, the Boltzmann damping dominates the Haar growth and the integral converges. This invariant measures the minimal degree of stability of geodesic rays in $c_1(L)=[\omega_0]$5 relative to the chosen action. A rescaled version is also defined when the symmetric-space metric is dilated by a factor $c_1(L)=[\omega_0]$6: $c_1(L)=[\omega_0]$7

The main explicit computation is for the $c_1(L)=[\omega_0]$8-balancing energy $c_1(L)=[\omega_0]$9, the normalized restriction to $L\to M$00 of

$L\to M$01

with matrix form

$L\to M$02

This is a convex stability functional on $L\to M$03, bounded below by $L\to M$04, and its critical point is the $L\to M$05-balanced metric characterized by

$L\to M$06

Along a geodesic

$L\to M$07

the asymptotics are

$L\to M$08

with uniformly bounded $L\to M$09-intercept and positive slope

$L\to M$10

The exact threshold for the standard Haar measure is

$L\to M$11

equivalently, the partition function converges iff

$L\to M$12

For the scaled measure one obtains

$L\to M$13

On a compact Riemann surface of genus $L\to M$14, Riemann–Roch gives

$L\to M$15

hence

$L\to M$16

The proof identifies a precise competition between the linear growth of $L\to M$17 and the exponential growth of the Haar volume form, whose eigenvalue-coordinate asymptotics involve the Vandermonde-type factor

$L\to M$18

The partition function has a pole precisely at $L\to M$19, and the threshold is sharp (Klevtsov et al., 2014).

5. Asymptotics, fluctuations, and large deviations

Large-$L\to M$20 analysis is central to the subject because Bergman metrics approximate the background Kähler metric while retaining nontrivial fluctuation theory. For heat kernel measures on $L\to M$21, the covariance of the metric potential is given exactly by

$L\to M$22

where $L\to M$23 is the normalized Bergman kernel, also called the Berezin kernel in this context. Differentiating yields the variance current

$L\to M$24

For fixed $L\to M$25, as $L\to M$26,

$L\to M$27

so in the paper’s notation

$L\to M$28

In words, large-time random Bergman metrics degenerate toward random zero divisors viewed as singular metrics, and their fluctuation theory matches that of random zero currents (Shiffman et al., 2023).

The same survey describes a coupled scaling limit obtained by rescaling the Cartan–Killing metric on $L\to M$29 by

$L\to M$30

so that

$L\to M$31

This expresses a universal local limit for metric fluctuations on the $L\to M$32 scale (Shiffman et al., 2023).

From the path-integral perspective, convergence of measures on $L\to M$33 is formulated באמצעות a large deviations principle. A sequence $L\to M$34 satisfies an LDP with speed $L\to M$35 and rate function $L\to M$36 if

$L\to M$37

A useful construction is the contraction-principle formula

$L\to M$38

which is presented as a natural approximation of a target geometric action on $L\to M$39. Explicit matrix ensembles then yield explicit rate functions. For the Wishart model, the large deviation rate function for the empirical eigenvalue measure $L\to M$40 is

$L\to M$41

while a Gaussian model in the logarithmic variable

$L\to M$42

leads to

$L\to M$43

These are presented as finite-dimensional actions whose pullbacks suggest geometric actions on the infinite-dimensional symmetric space of Kähler metrics (Ferrari et al., 2011).

Related probabilistic geometry appears in determinantal point processes associated with Bergman kernels. For an orthonormal basis of $L\to M$44, the $L\to M$45-particle density

$L\to M$46

defines a DPP with kernel equal to the Bergman kernel $L\to M$47. After $L\to M$48-rescaling, the $L\to M$49-point functions converge to determinants of the universal kernel

$L\to M$50

the multidimensional analogue of the infinite Ginibre kernel. The empirical measures converge in probability to the equilibrium Monge–Ampère measure, and weighted variants satisfy an LDP with speed

$L\to M$51

and good rate function given by the Legendre–Fenchel transform of the Mabuchi functional (Lemoine, 2022). This probabilistic geometry is explicitly linked to earlier work of Douglas–Klevtsov, Klevtsov, and Klevtsov–Ma–Marinescu–Wiegmann (Lemoine, 2022).

6. Related frameworks, analogues, and scope

Random Bergman metrics sit within a broader web of Bergman-kernel constructions. One important bridge is to random normal matrix theory and projective embeddings. For a fermionic system on a compact Kähler manifold, the partition function is

$L\to M$52

In complex dimension one, its large-$L\to M$53 expansion has the form

$L\to M$54

so the first terms are the Aubin–Yau, Mabuchi, and Liouville functionals. Restricting this to the Bergman subspace produces the balancing energy

$L\to M$55

and the Liouville balancing energy

$L\to M$56

This establishes a precise relation between matrix-model free energies, the determinant of the Hilb map, and geometric functionals on the space of Bergman metrics (Klevtsov, 2013).

A real-variable analogue is provided by Riemannian Bergman metrics. There, one replaces holomorphic sections by eigenspaces of the Laplace–Beltrami operator and defines

$L\to M$57

These spaces form finite-dimensional symmetric-space approximations to $L\to M$58, and every smooth Riemannian metric can be approximated in $L\to M$59 by such metrics. Randomness is not developed as a central theme in that setting; the connection to random Bergman metrics is explicitly described as conceptual rather than carried out (Potash, 2013).

A different, information-geometric direction concerns Bergman metrics on bounded domains. For a bounded domain $L\to M$60, the Bergman statistical model

$L\to M$61

satisfies

$L\to M$62

so the Bergman metric is the pullback of the Fisher information metric of this statistical model. The resulting Schwarz lemma is proved by a covariance estimate based on the probabilistic Cauchy–Schwarz inequality. That work is not mainly about random Bergman metrics, but it treats Bergman geometry in explicitly probabilistic terms and is therefore methodologically adjacent to stochastic Bergman geometry (Seo et al., 30 Dec 2025).

These neighboring frameworks clarify a common misconception. Not every probabilistic construction involving Bergman kernels is itself a random Bergman metric ensemble. Determinantal point processes built from Bergman kernels describe random point configurations rather than random sampling of metrics, although they are described as the kind of probabilistic geometry that underlies random Kähler geometry (Lemoine, 2022). Likewise, the bounded-domain Fisher-metric approach studies canonical probability densities associated with Bergman kernels rather than ensembles on $L\to M$63 (Seo et al., 30 Dec 2025). The core meaning of random Bergman metrics remains the assignment of probability measures to finite-dimensional spaces of Bergman metrics, viewed as approximations to the space of Kähler metrics (Ferrari et al., 2011, Ferrari et al., 2011, Klevtsov et al., 2014).