Available online at www.sciencedirect.

com

ScienceDirect
Nuclear Physics B 973 (2021) 115582
www.elsevier.com/locate/nuclphysb

Riemannian Gaussian distributions, random matrix

ensembles and diffusion kernels
Leonardo Santilli a , Miguel Tierz b,a,∗
a Departamento de Matemática, Grupo de Física Matemática, Faculdade de Ciências, Universidade de Lisboa,
Campo Grande, Edifício C6, 1749-016 Lisboa, Portugal
b Departamento de Matemática, ISCTE - Instituto Universitário de Lisboa, Avenida das Forças Armadas,
1649-026 Lisboa, Portugal
Received 8 July 2021; received in revised form 27 September 2021; accepted 21 October 2021
Available online 26 October 2021
Editor: Hubert Saleur

Abstract
We show that the Riemannian Gaussian distributions on symmetric spaces, introduced in recent years,
are of standard random matrix type. We exploit this to compute analytically marginals of the probability
density functions. This can be done fully, using Stieltjes-Wigert orthogonal polynomials, for the case of the
space of Hermitian matrices, where the distributions have already appeared in the physics literature. For the
case when the symmetric space is the space of m × m symmetric positive definite matrices, we show how
to efficiently compute densities of eigenvalues by evaluating Pfaffians at specific values of m. Equivalently,
we can obtain the same result by constructing specific skew orthogonal polynomials with regards to the log-
normal weight function (skew Stieltjes-Wigert polynomials). Other symmetric spaces are studied and the
same type of result is obtained for the quaternionic case. Moreover, we show how the probability density
functions are a particular case of diffusion reproducing kernels of the Karlin-McGregor type, describing
non-intersecting Brownian motions, which are also diffusion processes in the Weyl chamber of Lie groups.
© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/). Funded by SCOAP3 .

* Corresponding author.
E-mail addresses: [email protected] (L. Santilli), [email protected], [email protected] (M. Tierz).

https://doi.org/10.1016/j.nuclphysb.2021.115582
0550-3213/© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/). Funded by SCOAP3 .
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

1. Introduction

The statistical analysis of a probability measure Q on a differentiable manifold is a topic

attracting much attention and has developed considerably due to the wide range of interesting
applications that require the analysis of data not living in ordinary Euclidean spaces. There is
great interest, for example, in trying to process datasets which lie in the space Pm of m × m
symmetric positive definite matrices.
This space can be equipped with a Riemannian metric, giving it the structure of a Riemannian
homogeneous space of negative curvature. Such Riemannian manifold has been thoroughly stud-
ied from many points of view, corresponding to different disciplines, such as harmonic analysis
and number theory [1,2], matrix analysis [3] and information geometry and multivariate statistics
[4,5].
This and other symmetric spaces have been studied endowed with metrics such as the Rao-
Fisher metric or the log-Euclidean metric [6], in order to develop the tools to properly carry
out statistical inference of data on such spaces. The landscape of applications is very vast as it
includes, for example, medical imaging [7,8], continuum mechanics [9], radar signal process-
ing [10,11] and computer vision [12–16], among many other references that could be quoted
here.
In the present paper, the Riemannian metric will be the Rao-Fisher metric, also known as
affine-invariant metric, and is the subject of Section 1.1. A specific probabilistic model to repre-
sent the statistical variability of data in non-Euclidean spaces, such as Pm , has been developed
by introducing the so-called Gaussian probability distributions on Riemannian symmetric spaces
[17–20].
The Riemannian Gaussian distribution G(Ȳ , σ ) will depend on two parameters, Ȳ ∈ Pm and
σ > 0. The form of its probability density function, generalising that of a Gaussian distribution
on the Euclidean space Rp , is given by
 
1 d 2 (Y, Ȳ )
p(Y | Ȳ , σ ) = exp − , (1)
ζ (σ ) 2σ 2
where d : Pm × Pm → R+ is Rao’s Riemannian distance (defined below), and the density is
with respect to the Riemannian volume element of Pm , denoted by dv(Y ). In comparison √ to a
Gaussian distribution on R, the normalising factor ζ (σ ) plays the role of the factor 2π σ 2 .
Among other possible applications, the distributions provide a statistical foundation for the
concept of Riemannian centre of mass, which plays a major role in many contexts [21,22]. They
also play a role in machine learning when the dataset is inherently structured [23–25].
In addition to the relationship between Riemannian Gaussian distributions and the concept
of Riemannian centre of mass and the potential for applications, discussed in [17–20], we shall
show here that one of the hallmarks of these distributions is analytical tractability. In particular,
we will show how the obtained distributions are of standard random matrix type and how the
ensuing analytical tools associated to the study of random matrices can be used effectively to
further characterize the distributions analytically [26–28].
More specifically, we shall use Pfaffian and determinant descriptions of the random matrix
ensembles and the classical orthogonal polynomial method in random matrix theory [26–28].
We do this in the setting which includes skew-orthogonal polynomials, necessary to solve models
which do not correspond to the case of Hermitian matrices, such as the case of the real symmetric
matrices, which is central for statistical applications.

2
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

We now briefly summarize the derivation of the model for the case of real symmetric ma-
trices, along the lines of [17–20], to set up the context. Further details are in these references,
while we will focus, from Section 2 on, on the analytical characterization of the joint probability
distribution functions using random matrix theory.

1.1. The Rao-Fisher metric: distance, geodesics and volume

We overview the general setup, following closely the presentation in [18].

The Rao-Fisher metric [5,2]
ds 2 (Y ) = tr [Y −1 dY ]2 , Y ∈ Pm (2)
is a Riemannian metric on Pm . It induces a Riemannian distance d : Pm × Pm → R+ , known as
Rao’s distance, in the standard way.
When equipped with the Rao-Fisher metric, the space Pm becomes a Riemannian manifold of
negative sectional curvature [1,2]. Since it is also complete and simply connected, the distance
between any two points in Pm is given by the length of the geodesic connecting them.
The volume form on Pm induced by the Rao-Fisher metric is [2]
m+1 
dv(Y ) = det(Y )− 2 dYij (3)
i≤j

where indices denote matrix entries.

One can then consider the expressions of ds 2 (Y ) and dv(Y ), given by (2) and (3), in terms of
polar coordinates [27]. Then, for any class function f : Pm → R, the angular coordinates can be
integrated out, obtaining
     m
|ri − rj | 
m(m−1)
8 4
f (Y ) dv(Y ) = ω m f (r1 , . . . , rm ) sinh dri , (4)
m! 2m 2
Pm Rm i<j i=1

where (see [29, Page 71]),

2
2m π m /2
ωm ≡ (5)
m (m/2)
with m the multivariate Gamma function, given in [29]. Formula (4) follows directly from [2,
Proposition 3, Page 43] but has been written following the notation in [19].
For any σ > 0, let f (Y |σ ) be given by,
 
d 2 (Y, I )
f (Y |σ ) = exp − (6)
2σ 2
where I ∈ Pm is the m × m identity matrix. The normalization constant ζ (σ ) is then defined as

ζ (σ ) = f (Y |σ ) dv(Y ), (7)
Pm
where dv(Y ) is the volume form (3). Then,
m(m−1) 
8 4   
m
e−(r1 +···+rm )/2σ
2 2 2
ζ (σ ) = ωm · m
sinh |ri − rj |/2 dri (8)
m! 2
Rm 1≤i<j ≤m i=1

3
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

with ωm defined in (5).

We show how to use random matrix tools, in particular the evaluation of Pfaffians, to compute
analytically ζ (σ ) for several values of m and also in the large m limit. This extends some of the
results in [17,18] and the other works discussing Gaussian distributions on Riemannian symmet-
ric spaces, where the case m = 2 is evaluated analytically and the rest is carried out numerically,
with Monte-Carlo methods. Before that, we note that the corresponding model for Hermitian
matrices, also discussed in [19], which is characterized by an analogous expression but with
the sinh |ri − rj |/2 term squared, is fully solvable with orthogonal polynomials. It has already
been studied in the physics literature in detail, but we summarize and develop the results in the
next Section and we will study more difficult cases to treat analytically, such as (8), in Section 3
and in the Appendix. Finally, Section 4 will contain a discussion of these statistical distributions
from the point of view of diffusion processes.

2. Full solution of the case of Hermitian matrices

Let us define the function zβ (σ ) as

1
     m
|ri − rj | β  −cβ ri2 dri
2

zβ (σ ) = 2 sinh e σ √ , (9)
m! 2 πσ 2
m 1≤i<j ≤m i=1
R
that we call the partition function, adopting the language of random matrix theory. The parameter
β is called Dyson’s index. While (9) makes sense for every β > 0, with a suitable choice of
cβ > 0, we will be mainly interested in β ∈ {1, 2, 4}, corresponding to a random matrix ensemble
with orthogonal, unitary and symplectic symmetry, respectively. To avoid cumbersome overall
factors of 2m/2 in what follows, we adopt the normalization of [30] and choose
1
β = 1,
cβ = 2 (10)
1 β ∈ {2, 4} ,

but we stress that other suitable choices exist, as for example cβ = β2 , and the present discussion
σ2
does not depend on such choice. Note that the variance with this normalization is 2cβ , so in
σ2
particular it is σ2 for β = 1 but it is 2 for β = 2. For β = 1 in (9) we find
m(m−1) m(m+1)
2 4 π 2
ζ (σ ) = |σ |m z1 (σ ). (11)
m (m/2)
The crucial fact we want to exploit is that zβ (σ ) can be written in fully standard random
matrix form, meaning in terms of a Vandermonde determinant. This was done in [31], in the
context of the study of vicious walkers1 and, later on, it was also crucial to study Chern-Simons
gauge theories [32]. One can therefore, use all the power of the traditional random matrix tools,
such as determinantal expressions and the method of orthogonal polynomials. Denoting xi = eui
and using
 β
 β
xi − xj = e ui − e uj
1≤i<j ≤m 1≤i<j ≤m

1 We detail further the connection with diffusion in Section 4.

4
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

⎡  ⎤
m    β
ui |ui − uj |
=⎣ e2 (m−1)
2 sinh ⎦ , (12)
2
i=1 1≤i<j ≤m

we see that
  
m cβ
β − (log xi )2 dxi
xi − xj e σ2 √
πσ 2
(0,∞)m 1≤i<j ≤m i=1
    β 
m u2  
|ui − uj | −cβ i +u β m−1+ 2 dui
= 2 sinh e σ2
i 2 β
√ . (13)
2 πσ 2
1≤i<j ≤m i=1
Rm
 
σ 2β
Changing variables ri = ui − 4cβ m − 1 + β2 to complete the square, it follows that:
 2
β2
−σ 2 16c m m−1+ β2   
m cβ
e β
β − (log xi )2 dxi
zβ (σ ) = xi − xj e σ2 √ (14)
m! πσ 2
(0,∞)m 1≤i<j ≤m i=1

which describes the partition function of the Stieltjes-Wigert random matrix model. Note that
this model is of the standard random matrix form but with a weight function of log-normal
cβ
− (log x )2
type: w(x) = e σ 2 i
. Random matrix theory gives formulas2 for any marginal of the joint
probability density function of eigenvalues, including the normalization constant, in terms of the
polynomials orthogonal with regards to the log-normal weight w(x) above. These polynomials
are the Stieltjes-Wigert polynomials [33], which are essentially a type of q-deformed Hermite
polynomial.
A Riemannian log-normal probability distribution was already introduced in [34] (see [35] for
a review) but is not of the random matrix type and not comparable to (14) due to an extra factor
in the Vandermonde part.
Another way to see this q-deformed structure at play is to define the q-parameter

q = e−σ (15)
and introduce the q-number
q−x/2 − qx/2
[x]q = , (16)
q−1/2 − q1/2
which reduces to an ordinary real number, [x]q → x, in the limit q → 1− . Then we have
β   
m
σ 2 m(m−1) 1 2 d r̃i
zβ (σ ) = 2 sinh |[r̃i − r̃j ]q |β e−cβ r̃i √ (17)
2 m! π
1≤i<≤m i=1
Rm

where r̃i = ri /σ . This latter form shows that the partition function zβ (σ ) is, up to the explicitly
known overall σ -dependent factor, the q-deformation of the Gaussian β-ensemble, for every
β > 0, with q as in (15).

2 For the case β = 2. The other cases are more complicated as they require skew-orthogonal polynomials.

5
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Table 1
A-series of the Cartan classification of random matrix ensembles.
Root system Cartan class. Matrix type Dyson index Symmetry
Am−1 A Hermitian β =2 unitary
Am−1 AI real symmetric β =1 orthogonal
Am−1 AII quaternion Hermitian β =4 symplectic

We study the partition function zβ (σ ). Before that, we briefly review the relation between
random matrix ensembles and symmetric spaces in Subsection 2.1. Then we discuss first the
case β = 2 in the rest of the present Section. The other two cases β = 1 and β = 4 are studied in
Section 3.

2.1. Cartan classification: from symmetric spaces to matrix ensembles

A correspondence between symmetric spaces and random matrix ensembles was established
by Altland and Zirnbauer [36] and used to classify the latter ones in terms of the former (see
[37,38] for an overview).3
As we have seen above, when dealing with invariant joint probability distributions, the inte-
grals only depend on the eigenvalues (x1 , . . . , xm ) of the random matrix. The Jacobian coming
from this reduction is the Vandermonde determinant [26]:

J (x) = |xi − xj |β . (18)
1≤i<j ≤m
For matrix ensembles of the transfer matrix type (following the terminology of [38]) the Jacobian
is naturally
  
xi − xj β
J (x) = sinh (19)
2
1≤i<j ≤m
which is recast in the form (18) with an exponential change of variables. For circular ensembles,
on the other hand, one gets [40]
  
θi − θj β
J (x) = 2 sin (20)
2
1≤i<j ≤m

which is related to (18) through x = eiθ . One readily observes that (18), (19) and (20) are pre-
cisely the Jacobians to pass to spherical coordinates in symmetric spaces of zero, negative and
positive curvature, respectively.
We are interested in ensembles of real symmetric, Hermitian or quaternion real Hermitian
matrices, corresponding to β = 1, β = 2 and β = 4, respectively [26,27]. These three matrix
algebras are realised as the tangent space at the origin of coset spaces of Cartan type A, yielding
the classification summarized in Table 1.
In conclusion, we single out three hyperbolic symmetric spaces, which correspond to random
transfer matrix ensembles with β ∈ {1, 2, 4}. All of them are discussed in the following. Besides,
we can go beyond Cartan type A: two examples of type D are provided in Appendix B.

3 These results have been recently extended to include double-coset spaces [39]. That setup seems especially well-
suited for applications in statistical analysis.

6
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

2.2. Moments of the log-normal distribution and the Stieltjes-Wigert orthogonal polynomials

In what follows we exploit the rewriting (14), which allows to directly apply standard results
from random matrix theory. Define the inner product
∞
dx −(log x)2 /σ 2
(f, g)2 = √ e f (x)g(x) (21)
πσ 2
0

for analytic real functions f and g. We keep the dependence on σ implicit in the notation. The
inner product of two monomials is4
∞
dx −(log x)2 /σ 2 k+
(x , x )2 =
k
√ e x
πσ 2
0
+∞
dr −r 2 /σ 2 +r(k+ +1) σ2
+1)2
= √ e =e 4 (k+ . (22)
πσ 2
−∞

For the cases β = 1 and β = 4 we will need to define skew-symmetric products, but for
β = 2, zβ (σ ) can be evaluated exactly [31], thanks to the Stieltjes-Wigert polynomials, which
are orthonormal polynomials with respect to inner product (·, ·)2 defined in (21), that is

(Pk , P )2 = δk . (23)
This family of polynomials is given by
  n  n  
− 1 
1
n+ 12  n 2 ν
q ν + 2 (−x)ν ,
2
Pn (x) = (−1)n q 2 1 − qj (24)
ν q
j =1 ν=0

where the q-parameter is

q = e−σ
2 /2
(25)
(not to be confused with q in (15)), and
  ν
n [n]q 1 − q n−j +1
= = (26)
ν q [n − ν]q [ν]q 1 − qj
j =1

is the q-binomial. Let us denote by an the leading coefficient, that is Pn (x) = an x n + . . . , and
define the monic orthogonal polynomials pn (x), which satisfy the orthogonality relation
1
(pk , p )2 = δk , (27)
an2
with a0−2 = 1/q and

4 The same moments can be obtained working on the unit circle, if the function is instead Jacobi’s third theta function,
upon a formal replacement q → q −1 . This leads to a unitary matrix model with analogous properties to the models here
discussed, see [41–44] for details.

7
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Fig. 1. Plot of log zβ=2 (σ ) as a function of σ 2 /2, for m = 5, 6, . . . , 12.

n
j =1 (1 − q
j)
an−2 = (28)
2n2 +2n+ 12
q
for n = 1, 2, . . . [31].
The main tool to evaluate zβ=2 (σ ) exactly is the celebrated Andréief identity [45,46], which
allows to rewrite the matrix model (14) as a determinant:
σ2 3
  
z2 (σ ) = e 4 m det x i−1 , x j −1 . (29)
1≤i,j ≤m 2

In turn, the inner product of monomials inside the determinant can be replaced by pi−1 , pj −1 2 ,
so that the matrix (of which we are taking the determinant) becomes diagonal in this basis, and
we obtain [31]

σ2 3 
m−1
σ2
 m(m−1)/2 m−1

z2 (σ ) = e− an−2 = e 12 m(m
2 −1)
1 − e−σ
2 /2
4 m q (j + 1), (30)
n=0 j =1

with q defined in (25) and q the q-Gamma function


j
q (j + 1) = [n]q . (31)
n=1
Explicitly:
m 
  2 m−j
σ2 2 −1)− σ 2 m(m−1) σ
z2 (σ ) = e 12 m(m 8 −2 sinh j . (32)
4
j =1

We plot z2 (σ ) as a function of σ for various m in Fig. 1, and as a function of m for various

fixed σ in Fig. 2.
The dependence on m is better captured by log zβ rather than zβ itself. For this reason, while
our discussion will focus on zβ , in the plots and numerical evaluations we will instead consider
log zβ .

2.3. Eigenvalue density, β = 2

Let us introduce the eigenvalue density, which by definition is obtained integrating the joint
probability density over m − 1 of the m eigenvalues [26]:

8
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Fig. 2. Plot of log zβ=2 (σ ) as a function of m, for fixed values of σ 2 = 0.5, 1, 2.

cβ 2
− r
m! e σ2
ρβ (r; σ ) = ·√
(m − 1)!z2 (σ ) πσ 2
     m−1
|ri − rj | β  − cβ2 ri2 dri
· 2 sinh e σ √ . (33)
2 πσ 2
m−1 1≤i<j ≤m i=1
R
Clearly, from the invariance of the integral z2 (σ ) under permutation of the ri ’s, we can integrate
over any m − 1 of the m variables and obtain the same definition of ρβ (r; σ ). Throughout this
σ2
Subsection we consider β = 2. It is convenient to use the change of variables xi = eri − 2 m as
in (14), and exploit the Stieltjes-Wigert polynomials [32]. We can use the Christoffel-Darboux
formula to rewrite the eigenvalue density of any random matrix ensemble as [27, Eq. (5.13)]
 
ρ2 (x) = am−1
2
w(x) pm−1 (x)pm (x) − pm (x)pm−1 (x) , (34)
where w(x) is the weight function, pn are the monic orthogonal polynomials with respect to the
−(log x)2 /σ 2
weight w(x), and the prime means derivative with respect to x. In our case, w(x) = e σ π 1/2
and pn are the monic Stieltjes-Wigert polynomials introduced in the previous Subsection. We
find
pm−1 (x)pm (x) − pm (x)pm−1 (x)
    

m−1 m
m−1 m
2m2 −m+ 12 k+ −1 k 2 + 2 + k+
=q k(−1) k+
x q 2 − (k ↔ ) .
k q q
k=0 =0
(35)
Undoing the change of variables, using ρ2 (x)dx = ρ2 (e )e r+σ 2 m/2
dr and multiplying the r+σ 2 m/2

last expression by the Gaussian prefactor coming from the weight function in the Christoffel-
Darboux formula (34), we find that ρ2 (r; σ ) takes the form
 2

2m−2
− 1 2
r− σ2 (k+1−m)
ρ2 (r; σ ) = ck (q)e σ2 (36)
k=0

with the coefficients ck (q) obtainable from the expressions above. The upshot is that ρ2 (r; σ ) is
the sum of 2m − 1 Gaussian distributions, centered at the points

9
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Fig. 3. Eigenvalue density ρ2 (r; σ ) at m = 20. Left: σ 2 = 0.2. Right: σ 2 = 1.5.

Fig. 4. Eigenvalue density ρ2 (r; σ ). Left: small σ regime, m = 6 and σ 2 = 0.005. Right: large σ regime, m = 10 and
σ 2 = 6.2. Note that the height of the vertical axis in the left picture it twice the height in the other plots.

σ2
rk = (k + 1 − m), k ∈ {−m + 1, . . . , 0, . . . , m − 1} . (37)
2
It is also easy to check that the coefficients ck (q) have alternating sign, thus the Gaussians cen-
tered at rk with k odd interfere constructively with the other Gaussians at odd positions, and
destructively with those at even positions. We therefore expect ρ2 (r; σ ) to be described by m
peaks and m − 1 valleys among them. Moreover, the support of the eigenvalue density grows
linearly with the product t ≡ mσ 2 . We plot the eigenvalue densities for various σ in Fig. 3, and
in the small σ and large σ regime in Fig. 4.

2.4. Limits of zβ (σ )

Before delving into a more detailed analysis of zβ (σ ) for β ∈ {1, 4}, it is instructive to analyze
various limits for generic β > 0.

2.4.1. σ → 0+ limit
First, it is clear from (17) that sending σ → 0+ we recover the usual Gaussian β-ensemble,
whose partition function in known exactly for every β > 0, see [27, Eq. (1.160)]. In particular,
with our normalization we have

σ →0+ 
m/2−1
(2j + 1)
zβ=1 (σ ) −−−−→ σ m(m−1)/2 2m/2 (38)
22j
j =0

for m even (and a similar expression for m odd), and

10
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

σ m(m−1)  (j + 1)
m
σ →0+
zβ=2 (σ ) −−−−→ (39)
m! 2j −1
j =1
2m(m−1)  m
(2j + 1)
σ →0+ σ
zβ=4 (σ ) −−−−→ . (40)
m! 22j +1
j =1
2 /2
Comparing with formula (11), the limit (38) shows in particular that ζ (σ ) goes to zero as ∼ σ m
in the small σ limit, that is, small variance in the dataset.

2.4.2. σ 2 → ∞ limit
Another possible limit is the σ 2 → ∞ limit. We give the result taking first the large σ 2 limit
and then approximating for large m. It is convenient to write zβ (σ ) as
⎛ ⎞
  m 
m 
σm β
zβ (σ ) ≈ m/2 d r̃i exp ⎝−cβ σ 2 r̃i2 + σ 2 r̃i − r̃j ⎠ , (41)
π m! 2
i=1 i=1 i=j
Rm
  
where we have changed variables r̃i = ri /σ 2 and approximated 2 sinh σ 2 |x| ≈ eσ |x|/2 at
2
2
large σ 2 . Making the ansatz that r̃i grows as r̃i = mα si for some α > 0 and si of order 1, we see
that a saddle point exists in the large m limit if the two terms in the exponential in (41) are of the
same order in m. The first term goes as m1+2α and the second as m2+α , meaning that the large m
limit requires m1+2α = m2+α , that is α = 1. This implies that log zβ (σ ) ∝ σ 2 m3 at leading order
at large σ 2 and large m.
We can also foresee the form of the sub-leading correction. Indeed, it will come from inte-
grating over fluctuations around the saddle point. The entries of the Hessian matrix from (41)
are ∼ cβ σ 2 , and performing the m Gaussian integrals and taking into account the coefficient we
expect this sub-leading correction to be of order ∼ m log σ 2 . Note, however, that (41) as it stands
is not suitably written to study the saddle point equation. To get rid of the absolute value, we
should restrict to the principal Weyl chamber in Rm and look for a saddle point therein. We do
not pursue this approach, and instead quote the result found in [47], where a rigorous approach
to this limit has been undertaken, based on work by Baxter [48].5 One gets:
β 2 β
log zβ (σ ) ≈ σ m(m2 − 1) + m log σ 2 . (42)
24 8
2.4.3. Planar limit
Another useful limit is the scaled large m limit with m → ∞ and σ 2 → 0 keeping their
product mσ 2 ≡ t fixed. In this limit, the leading contribution to zβ comes from the saddle point
configuration, and one finds
m β
log zβ (σ ) ≈ − log(m!) − log(πσ 2 ) + Funiv. (t), (43)
2 2
where Funiv. (t) is a β-independent quantity, explicitly found solving the saddle point equation.
To obtain (43) we have used the normalization cβ = β/2, and the normalization (10) is recovered
simply shifting t → 2t in the β = 4 case.

5 The extension of Baxter’s result [48] to arbitrary β, as we need, is straightforward. The statistics of extreme eigen-
values in this regime has been recently studied in [49,50].

11
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Both Funiv. (t) and the limiting eigenvalue density ρ(r; t) are explicitly known: they have been
computed almost twenty years ago in the context of Chern-Simons theory [51]. We refer to [51]
or the review [52] for detailed analysis and formulas in this regime.
In [31] this scaled limit was studied and subdivided in three cases, according to the value of
mσ 2 ≡ t being finite, infinite or 0. In Section 4, we will discuss the relationship between the
probability densities in [17–19] with diffusion processes and how they appear in the study of
Brownian motion on Weyl chambers of Lie groups.

3. The cases of real symmetric and quaternion Hermitian matrices

According to the results in Subsection 2.4, we can extract the behaviour of zβ (σ ) in certain
limits from the knowledge of z2 (σ ). Nevertheless, we can do more, and evaluate exactly z1 (σ )
and z4 (σ ) using standard tools in random matrix theory.

3.1. Skew-symmetric products and β = 1 partition function

We discuss zβ=1 (σ ), which is directly related to ζ (σ ) through (11). In contrast to the case
β = 2, the partition functions zβ=1 (σ ) and zβ=4 (σ ) do not correspond to determinants, but rather
to Pfaffians. It is convenient to discuss separately the two cases with m even or m odd, starting
with the former. Before that, we define the skew-symmetric products [30]
∞
1 dx −(log x)2 /σ 2  
f, g4 = √ e f (x)g (x) − f (x)g(x) , (44)
2 πσ 2
0
df
where f = dx , and

∞ ∞
1 dx −(log x)2 /2σ 2 dy −(log y)2 /2σ 2
f, g1 = √ e f (x) √ e g(y)sign(x − y). (45)
2 πσ 2 πσ 2
0 0

Note that the coefficient in the exponent of the weight function is cβ as defined in (10).
Computing the skew-symmetric products of any pair of monomials gives
+∞
1 dr −r 2 /σ 2 +r  k+ −1 −1

x , x 4 =
k
√ e e − kek+
2 πσ 2
−∞
σ2
)2 ( − k)
=e 4 (k+ (46)
2
and
+∞  r1
1 dr1 −r 2 /2σ 2 +r1 (k+1) dr2 −r 2 /2σ 2 +r2 ( +1)
x k , x 1 = √ e 1 √ e 2
2 πσ 2 πσ 2
−∞ −∞
+∞
dr2 −r 2 /2σ 2 +r2 ( +1)
− √ e 2
πσ 2
r1

12
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Fig. 5. log ζ (σ ) as a function of σ 2 for various fixed values of m.

⎛ ! ⎞
+∞
σ2 dr1 r1 σ2
=e 2 ( +1)2
√ e −r12 /2σ 2 +r1 (k+1)
erf ⎝ √ + ( + 1)⎠
2πσ 2 2σ 2 2
−∞
σ2
  σ 
= e 2 (k+1) +( +1) erf
2 2
( − k) , (47)
2
where we have used the change of variables x = er . In (47), erf(x) is the error function, and in
the last line we have used the integration formula for erf(x) with a Gaussian weight. The error
function is odd, erf(−x) = −erf(x), whence both products are manifestly skew-symmetric.
Assuming m even, we use the de Bruijn identity [53] to write
σ2
 
zβ=1 (σ )|m even = e− 8 m(m+1) Pf 2x i−1 , x j −1 1
2
(48)
1≤i,j ≤m

with the skew-symmetric product x i−1 , x j −1 1 given in (47). For m odd, instead, the de Bruijn
identity [53] leads us to the Pfaffian of a (m + 1) × (m + 1) skew-symmetric matrix:
    
2 2x i−1 , x j −1 1 i,j ≤m 2 1, x i−1 2 i≤m
zβ=1 (σ )|m odd = e − σ8 m(m+1)2
Pf    .
1≤i,j ≤m+1 −2 1, x j −1 2 j ≤m 0
(49)
In both cases, we could expand the Pfaffian and obtain zβ=1 (σ ), and hence ζ (σ ), as a finite sum
of terms, explicitly known from (47). The advantage of this approach is that the Pfaffians can be
obtained explicitly for fixed m with the aid of a computer algebra. We provide numerical results
at various m and σ 2 in Tables 2 and 3, computed in MATHEMATICA using the algorithm of [54].
From zβ=1 (σ ) we immediately obtain ζ (σ ), which we show in Fig. 5 as a function of σ 2 for
various m, and in Fig. 6 as a function of m for various fixed values of σ 2 . We also show the
agreement of log zβ=1 with the theoretical predictions of Subsection 2.4 for small σ and large σ
in Fig. 7.
While the Pfaffian representation of zβ=1 is very useful for computational purposes, the simple
example m = 2 can be easily solved by direct integration:
∞ 2 ∞ 2  
1 dr1 − r12 dr2 − r22 |r1 − r2 |
zβ=1 (σ )|m=2 = √ e 2σ √ e 2σ 2 sinh
2! πσ 2 πσ 2 2
−∞ −∞

13
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Table 2
log ζ (σ ) for m from 2 to 12 and σ 2 from = 0.1 to 2.4. Evaluation time 0.207 s (whole table) on macbook.
m | σ 2 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2
2 -0.680 0.376 1.001 1.449 1.801 2.092 2.340 2.557 2.751 2.926 3.086 3.234
3 -2.776 -0.033 1.561 2.685 3.553 4.260 4.855 5.370 5.824 6.230 6.597 6.933
4 -2.173 1.446 3.628 5.223 6.495 7.566 8.497 9.328 10.080 10.773 11.416 12.020
5 -6.142 0.537 4.492 7.339 9.583 11.449 13.057 14.478 15.758 16.930 18.014 19.028
6 -4.374 3.418 8.200 11.755 14.641 17.108 19.287 21.259 23.073 24.765 26.359 27.874
7 -9.899 2.342 9.725 15.138 19.485 23.169 26.403 29.315 31.987 34.473 36.813 39.035
8 -7.182 6.511 15.056 21.513 26.837 31.456 35.596 39.390 42.927 46.264 49.444 52.496
9 -14.007 5.538 17.541 26.502 33.829 40.142 45.779 50.935 55.734 60.261 64.575 68.720
10 -10.489 10.945 24.546 34.983 43.719 51.401 58.374 64.842 70.936 76.742 82.323 87.725
11 -18.406 10.309 28.260 41.895 53.230 63.156 72.146 80.477 88.324 95.805 103.003 109.975
12 -14.188 16.951 37.033 52.678 65.954 77.778 88.632 98.803 108.472 117.760 126.751 135.506
m | σ 2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4
2 3.372 3.500 3.621 3.735 3.844 3.947 4.046 4.141 4.232 4.319 4.404 4.486
3 7.243 7.530 7.799 8.051 8.290 8.516 8.731 8.936 9.133 9.322 9.504 9.680
4 12.590 13.131 13.649 14.145 14.624 15.086 15.534 15.969 16.393 16.807 17.211 17.607
5 19.984 20.891 21.757 22.588 23.388 24.162 24.914 25.645 26.358 27.055 27.738 28.409
6 29.323 30.716 32.062 33.368 34.638 35.879 37.093 38.283 39.452 40.602 41.735 42.854
7 41.161 43.207 45.186 47.108 48.981 50.813 52.608 54.371 56.107 57.817 59.505 61.173
8 55.443 58.303 61.090 63.815 66.486 69.110 71.694 74.242 76.759 79.248 81.713 84.154
9 72.726 76.619 80.417 84.135 87.785 91.377 94.918 98.4151 101.874 105.298 108.692 112.059
10 92.979 98.112 103.144 108.090 112.964 117.774 122.529 127.237 131.902 136.530 141.126 145.691
11 116.766 123.408 129.927 136.343 142.671 148.924 155.111 161.243 167.325 173.364 179.364 185.329
12 144.071 152.479 160.758 168.927 177.002 184.999 192.926 200.793 208.607 216.374 224.101 231.790

Table 3
− log zβ=1 for even m from 2 to 40 and fixed σ 2 = 0.01. Evaluation time 0.133 s (whole table) on macbook.
m 2 4 6 8 10 12 14 16 18 20
− log z1 4.150 13.822 29.007 49.694 71.487 93.286 113.424 135.129 156.057 175.890
m 22 24 26 28 30 32 34 36 38 40
− log z1 199.322 219.185 244.263 267.156 295.836 319.471 349.711 378.587 410.486 444.106

Fig. 6. log ζ (σ ) as a function of m for σ 2 = 0.5, 1, 2.1. The dashed line shows the theoretical prediction in the large σ
regime at σ = 2.1.

14
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Fig. 7. Left: log zβ=1 in the small σ regime, m = 6. Right: log zβ=1 in the large σ regime, m = 8. “sub.” stands for the
sub-leading corrections obtained in (42).

∞ ∞  r1 r2 r1 r2 
2 2
1 dr1 − r12 dr2 − r22
= √ e 2σ √ e 2σ sgn(r1 − r2 ) e 2 − 2 − e− 2 + 2
2 πσ 2 πσ 2
−∞ −∞
∞ r2
    
σ2 dr1 − 1 r1 r1 + σ 2 /2 r1 r1 − σ 2 /2
=e 8 √ e 2σ 2 e 2 erf √ − e− 2 erf √ .
2πσ 2 2σ 2 2σ 2
−∞
(50)
We now observe that the integral of an error function with a Gaussian weight is again an error
function and, after simplifications, we obtain
 
σ2 |σ |
zβ=1 (σ )|m=2 = 2e 4 erf . (51)
2
This is a check of the result (48) for the smallest value of m. For higher values of m a direct
computation becomes more involved. However, from the Pfaffian representation, we infer that
the result is always a finite sum of products of error functions, weighted by integer powers of
2
eσ /4 . So, for example, de Bruijn’s identity immediately gives
    |σ |  
− 54 σ 2 2σ 2
zβ=1 (σ )|m=3 = 4 e 1+e erf − erf (|σ |) , (52)
2
      
5 2 |σ | 2 |σ | 3|σ |
zβ=1 (σ )|m=4 = 4e 2 σ erf − erf (|σ |)2 + erf erf . (53)
2 2 2
A similar type of expression, involving a sum of error functions, is obtained for  the Gaussian
distribution on the Poincaré ball [23]. In that case, the Vandermonde term sinh |ri − rj |/2 is
replaced by the simpler term sinh (|r|/2).
It would be interesting if the quick numerical evaluation of the normalization constant we
obtained, based on Pfaffians, can be of further exploited, given the relevance of the distribution
to the engineering of algorithms for machine learning and detection of structured collections
of data on graphs [25], as well as in the study of auto-encoders [23,24], just to name a few
applications, in addition to the ones discussed in [17,19,18,20].

3.2. Eigenvalue density, β = 1

The eigenvalue density at β = 2 has been obtained relying on the explicit knowledge of the
Stieltjes-Wigert polynomials. For a closed expression of ρβ=1 , we would need the corresponding

15
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

skew-orthogonal polynomials [26,30], which however are not known. To gain insight, let us first
study the simplest case m = 2. Then the eigenvalue density is obtained by direct integration, and
the computations are identical to the ones at the end previous Subsection. We find:
∞ 2 2  
1 dr2 − (r +r22 ) |r − r2 |
ρβ=1 (r; σ )|m=2 = e 2σ 2 sinh
zβ=1 (σ )|m=2 πσ 2 2
−∞

− σ8
2
− r2     
e e 2σ 2 r r + σ 2 /2 − 2r r − σ 2 /2
= ·√ e erf
2 √ − e erf √ .
erf (|σ |/2) 2πσ 2 2σ 2 2σ 2
(54)
This formula already shows the salient features of the β = 1 eigenvalue density: it is a sum of
products of Gaussians (centered at different, shifted points) multiplied by error functions.
Although without the skew-orthogonal polynomials in closed form, the eigenvalue density
can still be calculated for any fixed m. For this, we need to introduce some notation. Let p "n(x)
be the monic skew-orthogonal polynomials of the Stieltjes-Wigert family, that is, they satisfy

" "2
p2k , p +1 1 ="
hk δk , " "2 1 = 0 = "
p2k , p "2
p2k+1 , p +1 1 (55)
with the skew-symmetric product ·, ·1 defined in (45). These polynomials have an explicit Pfaf-
fian representation [55,56]. To reduce clutter, let us denote throughout this Subsection
(m) σ2 2
z1 =e 8 m(m+1) zβ=1 (σ )|m . (56)
(m)
That is, z1 is just the partition function at β = 1 for a given value of m, with the overall factor
coming from the definition (14) stripped off to avoid cumbersome coefficients.
(2k+2)
The constants "hk in (55) can then be expressed as "
z
hk = 1 (2k) . Moreover we have
z1
   i−1
1 2x i−1 , x j −1 1 i,j ≤2 x
"2 (x) = (2 )
p Pf  j −1 +1 i≤2 +1
(57)
z1 1≤i,j ≤2 +2 −x j ≤2 +1
0

for even degree n = 2 , and

1
"2
p +1 (x) = Pf
(2 +1) 1≤i,j ≤2n+4
[M(x)] , (58)
z1
for odd degree n = 2 +1. In the latter expression we have introduced the skew-symmetric matrix
[56]
⎛ 0 2(1, 1)2 2(1, x)2 2(1, x 2 )2 ··· 2(1, x 2 +1 )2 0
⎞
⎜ −2(1, 1)2 0 21, x1 21, x 2 1 ··· 1 ⎟
⎜ −2(1, x)2 −21, x1 0 2x, x 2 1 ··· x ⎟
⎜ −2(1, x 2 ) ⎟
M(x) = ⎜
⎜
2 −21, x 2 1 −2x, x 2 1 0 ··· x2 ⎟.
⎟
⎜ .. .. .. ..
···
.. ⎟
⎝ . . . . . ⎠
−2(1, x 2 +1 )2 −21, x 2 +1 1 −2x, x 2 +1 1 −2x 2 , x 2 +1 1 ··· x 2 +1
0 −1 −x −x 2 ··· 0
(59)
We also need the functions

16
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

∞
1 dy −(log y)2 /2σ 2
"n (x) =
φ √ e "n (y) sgn(x − y).
p (60)
2 πσ 2
0
Explicit expressions of the first few skew orthogonal polynomials and of the corresponding
"n (x) functions are given in Appendix A.
φ
Putting these definitions at work, we can obtain ρβ=1 (x) as a particular case of [27, Propo-
sition 6.3.3, Page 254]. We focus on the case of even m for simplicity, but a similar expression
exists for odd m. We get
2
− (log x) 
m/2−1
e 2
2σ 1  
ρ1 (x) = √ φ"2k (x)" "2k+1 (x)"
p2k+1 (x) − φ p2k (x) . (61)
πσ 2 "
hk
k=0
This expression allows to characterize the eigenvalue density ρβ=1 (r; σ ), undoing the change of
σ2
variables x = er− 2 (m+1) . Each p "n (x) is a polynomial of degree n in the variable x (and hence
2
in e ), with coefficients ratios of sums in which each term is of the form eσ a/4 erf(|σ |b/2) with
r

a and b integers. In turn, if we write schematically


n
"n (x) =
p αn;k x k (62)
k=1
with αn;n = 1 by definition, we find
 
1 
n
σ 2 (k+1)
2 u − kσ 2
"
φn (e ) = √
u
αn;k e 2 erf √ . (63)
2 k=0 2σ 2
Writing (61) in terms of the exponential variable r instead of x, and putting all together after
some rewriting, we arrive at
2
σ
e− 8 (m+5)
2 2
− r + r (m+3)
 z(2 )
m/2−1
ρβ=1 (r; σ ) = √ · e 2σ 2 2 · 1
(2 +2)
2πσ 2
=0 z1
2 2 +1   
σ2 r − (j + m/2)σ 2
α2 ;j α2 +1;k ekr+ 2 ((j +1) −(m+1)k) erf
2
× √
j =0 k=0 2σ 2
 
σ2 r − (k + m/2)σ 2
−α2 ;k α2 +1;j ej r+ 2 ((k+1) −(m+1)j ) erf
2
√ . (64)
2σ 2
In the latter expression the coefficients αn;k are read off form the skew-orthogonal polynomi-
(n)
"n (x). Since the formulas for both the polynomials and the normalizations z1 are known,
als p
the challenge to%obtain the density & ρβ=1 (r; σ ) is reduced to the task of evaluating Pfaffians.
Moreover, once p "n (eu )
"n (eu ), φ are found, it is possible to go beyond the eigenvalue
n=0,...,m−1
density and compute any marginal distribution of the eigenvalues [26, Chapter 6], [27, Chap-
ter 6].
We conclude that ρβ=1 (r; σ ) is a sum of terms that depend on r through the product of a
Gaussian and an error function in r. Comparing the results at β = 2 and β = 1, we see that
each Gaussian term in ρβ=2 (r; σ ) is corrected by an error function in the β = 1 case. The
2
same happens with the coefficients: the dependence ∼ eσ a/2 in the β = 2 case is replaced by
2
∼ eσ a/2 erf(|σ |b/2) at β = 1.

17
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Fig. 8. log zβ=4 (σ ) as a function of σ 2 for various fixed values of m.

Fig. 9. log zβ=4 (σ ) as a function of m for σ 2 = 0.5, 1, 2.1.

3.3. β = 4 partition function

We can apply the tools of Subsection 3.1 to study the partition function zβ=4 (σ ) as well. This
computes the normalization constant of the density considered in [57]. In this case, we must use
a different de Bruijn identity [53], which allows us to write
 2  
σ 2 m m− 12
zβ=4 (σ ) = e Pf 2x i−1 , x j −1 4 , (65)
1≤i,j ≤2m

in terms of a Pfaffian with entries the skew-symmetric products given in (46). Note that the
Pfaffian is that of a 2m × 2m skew-symmetric matrix. We plot the result as a function of σ 2 in
Fig. 8, and as a function of m for various fixed values of σ 2 is Fig. 9.
Moreover, we evaluate numerically log zβ=4 (σ ) in Tables 4 and 5.
The determination of the density ρβ=4 (r; σ ) for arbitrary m would entail constructing skew-
orthogonal polynomials with respect to the skew-symmetric product ·, ·4 , in analogy with the
β = 1 case. However, for the case m = 2, it can be easily obtained from direct integration:
∞ 2 2   
1 dr2 − c4 (r 2+r2 ) r − r2 4
ρβ=4 (r; σ )|m=2 = e σ 2 sinh
z4 (σ )|m=2 πσ 2 2
−∞

18
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Table 4
log zβ=4 (σ ) for m from 2 to 7 and σ 2 from 0.1 to 2.4. Evaluation time 0.109 s (whole table) on macbook.
m | σ 2 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2
2 -5.460 -4.008 -3.115 -2.454 -1.917 -1.460 -1.057 -0.693 -0.359 -0.04683 0.247 0.526
3 -12.993 -8.527 -5.737 -3.633 -1.902 -0.403 0.940 2.171 3.318 4.402 5.435 6.430
4 -22.443 -13.291 -7.477 -3.021 0.703 3.976 6.947 9.706 12.308 14.793 17.187 19.511
5 -33.163 -17.533 -7.439 0.4189 7.083 13.019 18.475 23.596 28.476 33.176 37.740 42.199
6 -44.620 -20.597 -4.830 7.629 18.342 28.002 36.977 45.480 53.649 61.572 69.319 76.931
7 -56.352 -21.896 1.082 19.502 35.542 50.166 63.881 76.859 91.584 110.199 132.388 152.264
m | σ 2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.4 2.4
2 0.794 1.051 1.301 1.543 1.780 2.011 2.239 2.462 2.683 2.901 3.116 3.330
3 7.392 8.329 9.244 10.142 11.026 11.897 12.758 13.610 14.455 15.293 16.126 16.955
4 21.779 24.002 26.189 28.346 30.478 32.591 34.686 36.767 38.837 40.896 42.947 44.991
5 46.576 50.887 55.146 59.365 63.550 67.705 71.846 76.049 80.407 83.995 94.476 99.749
6 83.662 94.386 108.756 120.963 137.087 134.410 165.131 180.287 171.118 179.404 230.606 211.766
7 163.070 198.697 214.792 202.635 288.985 277.438 344.463 373.588 361.125 359.058 413.542 408.297

Table 5
− log zβ=4 (σ ) for m from 1 to 20 and fixed σ 2 = 0.02. Evaluation time 0.110 s (whole table) on macbook.
m 1 2 3 4 5 6 7 8 9 10
− log z4 0.693 8.789 23.061 42.750 67.319 92.443 117.980 141.695 167.451 191.576
m 11 12 13 14 15 16 17 18 19 20
− log z4 211.572 235.962 256.665 275.833 293.453 308.834 321.895 334.330 343.452 356.662

2
c r
− 4 ∞
dr2 − c4 r22  
2
1 e σ2
= ·√ √ e σ 6 − 4 er−r2 + e−r+r2
z4 (σ )|m=2 πσ 2 πσ 2
−∞
 
+ e2r−2r2 + e−2r+2r2
c r2
e σ 2  σ 2 /c4 
− 4
2 4
= ·' e cosh(2r) − 4eσ /(4c4 ) cosh(r) + 3 . (66)
z4 (σ )|m=2 c4 πσ 2
We recall from (10) that in our conventions cβ=4 = 1, but we have kept it explicitly here. The
normalization zβ=4 (σ ) at m = 2 can be determined likewise, and we find
2
e2σ 2
zβ=4 (σ )|m=2 = − 2eσ /2 + 3. (67)
2

4. Diffusion and kernel interpretation

The first instance where the probability density (1) appeared in physics is seemingly in the
problem of vicious walkers in statistical physics [58]. The formulation of the vicious random
walker problem on a lattice, according to the so-called lock step model, consists in considering
m random walkers, each an even number of lattice spacings apart, on a one-dimensional lattice.
Then, at regular time intervals, each walker must take either a step one lattice site to the left or
a step one lattice site to the right, with equal probability 1/2. Coincidence of two walkers at the

19
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

same site and time results in the annihilation of both walkers and hence the process ends (whence
the name vicious).
In the diffusive limit, this model describes a set of non-intersecting Brownian motions and
one obtains the Gaussian Riemannian density (1) and also the ones corresponding to other sym-
metric spaces, as we shall see, by consideration of the classical, determinantal formalism of
non-intersecting random walkers and non-intersecting Brownian motion, initiated by Karlin and
McGregor in the seminal work [59] (see more recent work and textbooks [60–62] for a clear
description, including a modern explanation of Dyson’s Brownian motion [63]).
Given the relevance of diffusion processes in the statistical analysis of data [64,65], it is worth
to dwell further into the diffusive origins and interpretations of the probability densities discussed
above.
In [58], the probability density that all walkers survive at time t is obtained, giving an expres-
sion in terms of the positions (x1 , . . . , xm ) of the form
  
1 −|x|2
ct (x) = exp [exp(xj /t) − exp(xi /t)], (68)
(2πt)m/2 2t
1≤i<j ≤m

which, again using r.h.s. of (12), is the probability density function of (9), up to a shift of vari-
ables. Therefore, this density is the solution of a diffusion process and the density satisfies a heat
equation [31].
Interestingly, the problem of m vicious walkers on a one-dimensional lattice has an equivalent
description in terms of Brownian motion on the Weyl chamber of a Lie group [60,66]. All the
densities discussed in this work emerge in a natural way in such a context. To see this, we quote
a basic theorem [60] which states that if ct is the density function for a continuous stochastic
process, then for absorbing boundary conditions, it holds

bt (x, η) = sgn(w)ct (w(x) − η). (69)
w∈W

In this formula, η = (η1 , . . . , ηm ) is the vector of initial positions of the m particles, bt (x, η) is
the probability density for the particles to be at x given the initial positions η, and the sum is over
elements w of the Weyl group W .
These expressions are all determinantal and this theorem applies to standard Brownian
motion, in the Weyl chambers of Am−1 , Bm (and thus Cm ), and Dm , with absorbing or re-
flecting  boundary conditions. The measure for unconstrained standard Brownian motion is
ct (x) = m i=1 Nt (xi ), where Nt is the normal distribution function with mean 0 and variance t.
Consider the root system Am−1 , with associated Weyl group the symmetric group Sm . The
principal Weyl chamber is characterized by x1 > x2 > · · · > xm . This models the Brownian mo-
tion of m independent particles in one dimension. With absorbing boundary conditions, collisions
are forbidden and the sum can be written as a determinant, which gives
 
bt (x, η) = det Nt (xi − ηj ) . (70)
1≤i,j ≤m

This determinant gives the probability for n particles that start at positions (η1 , . . . , ηm ) and are in
independent Brownian motion to be at positions (x1 , . . . , xm ) at time t without having collided.
The expression (70) can be written as
    x η 
1 −|x|2 − |η|2 i j
bt (x, η) = exp det exp . (71)
(2πt)m/2 2t 1≤i,j ≤m t

20
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

The ηj are all integers and, if in addition we choose ηj = m − j , the determinant on the right-
hand side is the Vandermonde determinant in the variables exi /t , equal to
 
exj /t − exi /t . (72)
1≤i<j ≤m

Therefore, we have that, for this equal spacing condition on ηj :

   
−|η|2 −(m − 1)(2m − 1)/6
bt (x, η) = exp ct (x) = exp ct (x), (73)
2t 2t
where ct (x) is (68). Therefore, because of (12), the density is that of a Gaussian distribution on
the Riemannian space of real symmetric matrices introduced in [18,17,57], and studied here in
Section 3.
The corresponding Gaussian density for the space of Hermitian matrices appears from the
above when studying the probability of reunion, using for example the extensivity property of
probabilities [66, Eq. (14)], which in turn is related to the reproducing property of the (determi-
nantal) kernel (71).
For a more general setting involving any non-negative integers ηj , the determinant above is
the product of this Vandermonde determinant and the Schur function [27]

sη1 −m+1,η2 −m+2,...,ηm ex1 /t , . . . , exm /t , (74)
giving a more general probability density than the ones discussed here, but relevant (and com-
putable), in other contexts [66–68,42]. See also [69] to see how the Stieltjes-Wigert ensemble
emerges directly from the Karlin-McGregor formula when considering Brownian bridges.
For root system Bm , the corresponding Weyl group is the hyperoctahedral group, which in-
cludes permutations with any number of sign changes. Its Weyl chamber is x1 > x2 > · · · > xm >
0. This also models m independent particles in one dimension, with an additional wall at x = 0
[60]. In the Brownian motion setting at hand, the formalism gives
   x η   
1 −|x|2 − |η|2 i j −xi ηj
bt (x, η) = exp det exp − exp . (75)
(2πt)m/2 2t 1≤i,j ≤m t t
This determinant gives the measure for m particles which start at (η1 , . . . , ηm ) to be at
(x1 , . . . , xm ) at time t, neither having collided nor having touched x = 0. Choosing again
ηj = m − j , the determinant in (75) can be written in a Vandermonde form [60]. Doing so
and with some rewriting, that part gives
     m  
|λi − λj | |λi + λj |  λi
4 sinh sinh 2 sinh . (76)
2 2 2
1≤i<j ≤m i=1

This expression is associated with the Jacobian for matrix integration on o(2m + 1).
These non-intersecting Brownian motions are interwoven with many other areas. One of these
relationships is the connection with the Harish-Chandra-Itzykson-Zuber integral [70,71,61]. For
o(m) and sp(2m), this has been further developed in [72] where, in addition, it is proven that the
density bt (x, η) satisfies a diffusion equation.
Likewise, for Dm , with Weyl group the even hyperoctahedral group, which includes permu-
tations with an even number of sign changes, the principal Weyl chamber is x1 > x2 > · · · > xm ,
with xm−1 > −xm . This does not give a natural model for m particles in one dimension, but one
can still find a determinantal expression leading to the so(2m) case studied in Appendix B, see

21
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Eq. (80), and also appeared in statistical work in [19] (but not analyzed analytically there). In
Appendix B we study that case and the one corresponding to sp(2m), Eq. (93), which is closely
related to the odd orthogonal case quoted above (see also [72]).
It is worth mentioning that bt (x, η) is itself a reproducing kernel, as can be quickly checked
(again, see for example [66, Eq. (14)]), even though this fundamental property does not seem to
have been exploited, in spite of the broad literature on the subject.
It would be interesting if this fact could be used in a statistical context, either in conjunction
with the results described and cited or on its own, given the very well established relevance of
reproducing kernels in traditional statistical analysis [73] and, at the same time, the remarkable
and deep relevance of diffusion, in the statistical analysis of data, at many different levels [64,65].
It should be stressed that this is a different reproducing kernel from the one that immedi-
ately emerges by identifying the probability density in terms of a random matrix ensemble (14).
Random matrix theory [26] associates to (14) a reproducing kernel of the form:

' 
m−1
Km (x, y) = w(x)w(y) pk (x) pk (y), (77)
k=0

where w(x) is the Stieltjes-Wigert weight and pk (x) are the Stieltjes-Wigert polynomials, in-
troduced in Section 2.2. This is a delta sequence type of kernel [73]. Since the polynomials are
q-deformations of Hermite polynomials, the kernel is a one-parameter deformation of a Her-
mite kernel, used in probability density estimations [74,73]. From it, in principle any marginal
of the probability density can be obtained. The result in Section 2.3 is an example of this, corre-
sponding to the diagonal limit of the kernel. See [75] for explicit evaluations, and [76] for recent
developments along these lines.
An interesting open problem is to construct the skew-orthogonal polynomials associated to
the Stieltjes-Wigert (log-normal) weight function. This would put on equal footing the analyti-
cal results for the real-symmetric and quaternionic models with the full solution of the case of
Hermitian matrices. That would be a new result from the random matrix theory point of view
as well, since only classical orthogonal polynomials [30,55] and, more recently, semi-classical
polynomials [77] have been studied, whereas skew-polynomials in the q-deformed setting, to
which the Stieltjes-Wigert polynomials belong, remain unstudied.

CRediT authorship contribution statement

Leonardo Santilli: Conceptualization, Formal analysis, Investigation, Methodology, Visual-

ization, Writing – original draft, Writing – review & editing. Miguel Tierz: Conceptualization,
Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing –
review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal rela-
tionships that could have appeared to influence the work reported in this paper.

22
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

Acknowledgements

We thank Professors Peter Forrester and Gregory Schehr for correspondence. The work is
partially supported by FCT Project PTDC/MAT-PUR/30234/2017. The work of LS is supported
by the FCT through the doctoral grant SFRH/BD/129405/2017.

Appendix A. Skew-orthogonal polynomials

This Appendix collects the explicit expressions of the skew-orthogonal polynomials and the
"n (x), used in Section 3.2 in the computation of the density of states for the case
corresponding φ
of real symmetric matrices (β = 1). The first few skew-orthogonal polynomials are

"0 (x) = 1,
p (78)
"1 (x) = x,
p
5 2 erf(|σ |) 2
"2 (x) = x 2 − xe 2 σ
p + e4σ ,
erf(|σ |/2)
    
7σ 2 |σ | 11σ 2 19σ 2 3|σ |
−e 4 erf 2 − e 2 erf (|σ |) + e 4 erf 2
"3 (x) = x 3 + x 2
p  2   5σ 2
e2σ + 1 erf |σ2 | − e 4 erf (|σ |)
    
17σ 2
e8σ erf |σ2 | + e 4 erf (|σ |) − e6σ erf 3|σ |
2 2
2
+x  2   5σ 2
e2σ + 1 erf |σ2 | − e 4 erf (|σ |)
   
23σ 2 35σ 2 15σ 2
−e 4 − e 4 erf |σ2 | + e 2 erf (|σ |)
+  2   5σ 2
e2σ + 1 erf |σ2 | − e 4 erf (|σ |)

and the associated φ" functions, in the variable x = eu , are

   
√ u − σ2 √ u − 2σ 2
"0 (eu ) = e 2 σ 2 erf √ "1 (eu ) = e2σ 2 erf √
1
2φ , 2φ , (79)
2σ 2 2σ 2
      
√ u − 3σ 2 u − 2σ 2 u − σ2
"
9 2
2φ2 (e ) = e
u 2 σ
erf √ − erf √ + erf √ ,
2σ 2 2σ 2 2σ 2
   
√ u − 4σ 2 9σ 2 u − 3σ 2
"3 (eu ) = e8σ 2 erf √ √
2φ + e 2 erf √
2 σ2 2σ 2
⎡ 2   2   ⎤
19σ 2
−e 4 erf |σ2 | − e 2 erf (|σ |) + e 4 erf 3|σ |
7σ 11σ

⎢ 2 ⎥
×⎢⎣  2   2
⎥
⎦
|σ | 5σ
e + 1 erf 2 − e erf (|σ |)
2σ 4

⎡     ⎤
2 |σ | 17σ 2 2 3|σ |
  e erf 2 + e
8σ 4 erf (|σ |) − e erf 2
6σ
2 u − 2σ 2 ⎢ ⎢
⎥
⎥
+ e2σ erf √ ⎣    2 ⎦
2σ 2 |σ | 5σ
+ 1 erf 2 − e 4 erf (|σ |)
2σ 2
e

23
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

⎡    ⎤
23σ 2 35σ 2 15σ 2
 2 −e 4 − e 4 erf |σ2 | + e 2 erf (|σ |)
σ2 u−σ ⎢ ⎥
+e erf √ ⎢   ⎥.
 2
2
⎣ 2 ⎦
e2σ + 1 erf |σ2 | − e 4 erf (|σ |)
5σ
2σ 2

Appendix B. The case of other symmetric spaces

The methods presented in the main text can be extended to other symmetric spaces. As dis-
cussed previously, switching to other spaces in the Cartan classification corresponds, on the
random matrix theory side, to consider matrix integrals with different integration domain. We
restrict ourselves to hyperbolic spaces and β = 1 for concreteness.
We consider the partition function of the β = 1 ensemble of so(2m) matrices. The choice so
belongs to the D-class in the Cartan classification, and β = 1 (orthogonal symmetry) means that
we are working in the tangent space at the origin to the coset space SO(2m)/U (m), hence in the
DIII class [37]. The model also appears in the latter part of [19] where it is shown to be related
to integration on a Siegel disc.
The partition function is
      m
|ri + rj |  − ri 2 dri
2
so(2m) 1 |ri − rj |
zβ=1 (σ ) = 4 sinh sinh e 2σ √ .
(2m)! 2 2 πσ 2
1≤i<j ≤m i=1
Rm
(80)
We are now working with 2m × 2m matrices, therefore the Weyl group permuting the 2m eigen-
values is of order (2m)!. We then use the invariance of the integral under such permutations to
restrict the integration domain to the principal Weyl chamber
r1 ≥ r2 ≥ · · · ≥ rm ≥ 0 (81)
and obtain
      
m ri2
so(2m) ri − rj ri + rj − dri
z1 (σ ) = 4 sinh sinh e 2σ 2 √ .
2 2 πσ 2
0≤rm ≤···≤r1 <∞ 1≤i<j ≤m i=1

(82)
Using the simple property
   
ri − rj ri + rj  
4 sinh sinh = 2 cosh ri − cosh rj (83)
2 2
and changing variables xi = cosh ri , we arrive at
  
m
so(2m) dxi
z1 (σ ) = 2m(m−1)/2 (xi − xj ) w so (x; σ ) √ , (84)
πσ 2
1≤xm ≤···≤x1 <∞ 1≤i<j ≤m i=1

where we have defined

*   2 +
√
exp − 2σ1 2 log x + x − 1
2

w so (x; σ ) =  √  . (85)
log x + x 2 + 1

24
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

The function w so satisfies

r2
−
w so (x; σ )dx|x=cosh r = e 2σ 2 dr. (86)
The expression (84) is now suitably written to apply the de Bruijn identity [53]. To this aim, we
introduce the symmetric inner product
∞ √
dx
(f, g)so
2 = √ w so (x; σ/ 2)f (x)g(x)
πσ 2
1
∞
dr − r 22
= √ e σ f (cosh(r))g(cosh(r))
πσ 2
0
∞
1 dr − r 22
= √ e σ f (cosh(r))g(cosh(r)). (87)
2 πσ 2
−∞

We will need
 so ∞
j −1 1 dr − r 22
1, x = √ e σ (cosh r)j −1
2 2 πσ 2
−∞
j −1   ∞
1  j −1 dr − r 22 r(j −1−2
= j √ e σ e )
2 πσ 2
=0 −∞
j −1 
 
1 j −1 σ2
2 (j −1−2 )2
= e . (88)
2j
=0
Moreover, we introduce the skew-symmetric product
∞ ∞
1 dx dy
f, gso
1 = √ w so (x; σ )f (x) √ w so (y; σ )g(y) sign(x − y)
2 πσ 2 πσ 2
1 1
∞ r2 ∞ 2
dr1 − 12 dr2 − r22 sign(r1 − r2 )
= √ e 2σ f (cosh(r1 )) √ e 2σ g(cosh(r2 )) , (89)
πσ 2 πσ 2 2
0 0
which for monomials reads
, -so ∞ dr r2 ∞
dr2 − r22
2
sign(r1 − r2 )
i−1 j −1 − 12
e 2σ (cosh r2 )j −1
1
x ,x = √ e 2σ (cosh r1 )i−1
√
1 πσ 2 πσ 2 2
0 0
i−1 
  j −1  
1 i −1  j −1
=
2i+j −2 k
k=0 =0
∞ r2 ∞ 2
dr1 − 12 dr2 − r22 r2 (j −1−2 ) sign(r1 − r2 )
× √ e 2σ e r1 (i−1−2k)
√ e 2σ e
πσ 2 πσ 2 2
0 0

25
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

i−1 
  j −1  
1 i − 1  j − 1 σ22 (j −1−2 )2
= e
2i+j −1 k
k=0 =0
∞ r12
dr1 − +r1 (i−1−2k)
× √ e 2σ 2
2πσ 2
0
     
r1 + σ 2 (j − 1 − 2 ) r1 − σ 2 (j − 1 − 2 )
× erf √ + erf √ −1 .
2σ 2 2σ 2
(90)
We are now ready to apply the de Bruijn identity to (84). We obtain the Pfaffian form:
 
2x i−1 , x j −1 s0
so(2m)
z1 (σ )|m even = 2m(m−1)/2 Pf 1 (91)
1≤i,j ≤m

for even m, and

⎡     ⎤
2x i−1 , x j −1 so
so
2 1, x i−1
i,j ≤m
1 2 i≤m
so(2m)
z1 (σ )|m odd = 2m(m−1)/2 Pf ⎣   ⎦
1≤i,j ≤m+1 −2 1, x j −1 so
0
2 j ≤m
(92)
for odd m.
The procedure extends to the sp(2m) case. For β = 1, this corresponds to the class CI in the
Cartan classification [37]. The partition function is [27]
     
sp(2m) 1 |ri − rj | |ri + rj |
z1 (σ ) = 4 sinh sinh
(2m)! 2 2
1≤i<j ≤m
Rm

m
−
ri2
dri
× 2 sinh(|ri |) e 2σ 2 √ . (93)
i=1 πσ 2
We mimic the steps above, using the invariance to restrict the integration domain to the principal
Weyl chamber and change variables xi = cosh ri . We arrive at
  . 2
  
m 1
− log xi + xi2 −1
sp(2m) 2σ 2
z1 (σ ) = 2m(m+1)/2 (xi − xj ) e
1≤xm ≤···≤x1 <∞ 1≤i<j ≤m i=1

dxi
×√ . (94)
πσ 2
The de Bruijn identity [53] gives
 
2x i−1 , x j −1 1
sp(2m) sp
z1 (σ )|m even = 2m(m+1)/2 Pf (95)
1≤i,j ≤m

for even m, and

⎡    ⎤
sp 
2x i−1 , x j −1 1 i,j ≤m
sp
2 1, x i−1 2 i≤m
sp(2m)
z1 (σ )|m odd = 2m(m+1)/2 Pf ⎣    ⎦
−2 1, x j −1 2
sp
1≤i,j ≤m+1 0
j ≤m

26
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

(96)
sp
for odd m. The inner product (·, ·)2 in (96) is given by

 sp ∞
j −1 dr −r 2 /σ 2
1, x = √ e sinh(r)[cosh(r)]j −1
2 πσ 2
0
∞ j −1  
−j dr −r 2 /σ 2  j − 1  r(j −2 ) 
=2 √ e e − er(j −2 −2)
πσ 2
0 =0
j −1     
1  j −1 2
)2 σ4 σ (j − 2 )
= j +1
e(j −2 1 + erf
2 2
=0
  
2
−2)2 σ4 σ (j − 2 − 2)
− e(j −2 1 + erf , (97)
2
sp
and the skew-symmetric product ·, ·2 for monomials is given by
∞ 2
1 dr1 − r12
, x j −1 1
sp
x i−1
= √ e 2σ [cosh(r1 )]i−1 sinh(r1 )
2 πσ 2
0
∞ 2
dr2 − r22
× √ e 2σ [cosh(r2 )]j −1 sinh(r2 ) sign(r1 − r2 )
πσ 2
0
j −1    ∞
 dr1 − r12  r1 (i−k) 
i−1  2
−i−j i−1 j −1
=2 √ e 2σ e − er1 (i−k−2)
k πσ 2
k=0 =0 0
∞ r2   sign(r − r )
dr2 − 22
er2 (j − ) − er( j − −2) 1 2
× √ e 2σ
πσ 2 2
0

 j −1 
i−1   
1 i −1 j −1
=
2i+j +1 k
k=0 =0
∞
dr1 − r12  r1 (i−k) 
2

× √ e 2σ e − er1 (i−k−2)
πσ 2
0
*     
2 σ2 σ (j − 2 ) σ (j − 2 ) r1
× e(j −2 ) 4 erf − 1 − 2erf −
2 2 σ
  
σ2 σ (j − 2 − 2)
− e(j −2 −2) 4 erf
2

2
 +
σ (j − 2 − 2) r1
− 1 − 2erf − . (98)
2 σ
We note that these models, in the trigonometric version and with squared interaction terms
(that is, β = 2) have been solved fully in [44], using the connection to determinants, instead

27
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

of Pfaffians, of Toeplitz+Hankel matrices. One method that could be applied here would be for
example the use of expansions in Schur polynomials, see [44, Theorem 4] and the proof therein.

References

[1] S. Helgason, Differential Geometry, Lie Groups, and Symmetric Spaces, Graduate Studies in Mathematics, vol. 34,
American Mathematical Society, Providence, RI, 2001.
[2] A. Terras, Harmonic Analysis on Symmetric Spaces and Applications II, Springer-Verlag, New York, NY, 1988.
[3] R. Bhatia, Positive Definite Matrices, Princeton Series in Applied Mathematics, Princeton University Press, 41
William Street, Princeton, NJ, 2007.
[4] L.T. Skovgaard, A Riemannian geometry of the multivariate normal model, Scand. J. Stat. 11 (1984) 211, Available
online.
[5] C. Atkinson, A.F.S. Mitchell, Rao’s distance measure, Indian J. Stat. 43 (1981) 345, Available online.
[6] V. Arsigny, P. Fillard, X. Pennec, N. Ayache, Log-Euclidean metrics for fast and simple calculus on diffusion tensors,
Magn. Reson. Med. 56 (2006) 411.
[7] X. Pennec, P. Fillard, N. Ayache, A Riemannian framework for tensor computing, Int. J. Comput. Vis. 66 (2006) 41.
[8] A. Barachant, S. Bonnet, M. Congedo, C. Jutten, Multiclass Brain–computer interface classification by Riemannian
geometry, IEEE Trans. Biomed. Eng. 59 (2012) 920.
[9] M. Moakher, On the averaging of symmetric positive-definite tensors, J. Elast. 82 (2006) 273.
[10] M. Arnaudon, F. Barbaresco, L. Yang, Riemannian medians and means with applications to radar signal processing,
IEEE J. Sel. Top. Signal Process. 7 (2013) 595.
[11] M. Arnaudon, L. Yang, F. Barbaresco, Stochastic algorithms for computing p-means of probability measures, geom-
etry of radar Toeplitz covariance matrices and applications to HR Doppler processing, in: 12th International Radar
Symposium (IRS), 2011, pp. 651–656.
[12] S. Jayasumana, R. Hartley, M. Salzmann, H. Li, M. Harandi, Kernel methods on the Riemannian manifold of
symmetric positive definite matrices, in: IEEE Conference on Computer Vision and Pattern Recognition, 2013,
pp. 73–80.
[13] L. Zheng, G. Qiu, J. Huang, J. Duan, Fast and accurate nearest neighbor search in the manifolds of symmetric
positive definite matrices, in: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
2014, pp. 3804–3808.
[14] G. Dong, G. Kuang, Target recognition in SAR images via classification on Riemannian manifolds, IEEE Geosci.
Remote Sens. Lett. 12 (2015) 199.
[15] O. Tuzel, F. Porikli, P. Meer, Pedestrian detection via classification on Riemannian manifolds, IEEE Trans. Pattern
Anal. Mach. Intell. 30 (2008) 1713.
[16] R. Caseiro, J.F. Henriques, P. Martins, J. Batista, A nonparametric Riemannian framework on tensor field with
application to foreground segmentation, in: International Conference on Computer Vision, 2011, pp. 1–8.
[17] G. Cheng, B.C. Vemuri, A novel dynamic system in the space of SPD matrices with applications to appearance
tracking, SIAM J. Imaging Sci. 6 (2013) 592.
[18] S. Said, L. Bombrun, Y. Berthoumieu, J.H. Manton, Riemannian Gaussian distributions on the space of symmetric
positive definite matrices, IEEE Trans. Inf. Theory 63 (2017) 2153, arXiv:1507.01760.
[19] S. Said, H. Hajri, L. Bombrun, B.C. Vemuri, Gaussian distributions on Riemannian symmetric spaces: statistical
learning with structured covariance matrices, IEEE Trans. Inf. Theory 64 (2017) 752, arXiv:1607.06929.
[20] S. Said, L. Bombrun, Y. Berthoumieu, Warped Riemannian metrics for location-scale models, in: F. Nielsen (Ed.),
Geometric Structures of Information, Springer, Switzerland, 2019, pp. 251–296.
[21] B. Afsari, Riemannian Lp center of mass: existence, uniqueness, and convexity, Proc. Am. Math. Soc. 139 (2011)
655.
[22] M. Moakher, A differential geometric approach to the geometric mean of symmetric positive-definite matrices,
SIAM J. Matrix Anal. Appl. 26 (2005) 735.
[23] E. Mathieu, C. Le Lan, C.J. Maddison, R. Tomioka, Y. Teh, Continuous hierarchical representations with Poincaré
variational auto-encoders, in: H. Wallach, H. Larochelle, A. Beygelzimer, F. d’ Alché-Buc, E. Fox, R. Garnett
(Eds.), Advances in Neural Information Processing Systems (NeurIPS2019), vol. 32, Curran Associates, Inc., 2019,
pp. 12565–12576, arXiv:1901.06033.
[24] I. Ovinnikov, Poincaré Wasserstein Autoencoder, in: Bayesian Deep Learning Workshop (NeurIPS 2018), 2019,
arXiv:1901.01427.
[25] T. Gerald, H. Zaatiti, H. Hajri, N. Baskiotis, O. Schwander, From node embedding to community embedding: a
hyperbolic approach, arXiv:1907.01662.

28
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

[26] M.L. Mehta, Random Matrices, Pure and Applied Mathematics, Elsevier Science, 2004.
[27] P.J. Forrester, Log-Gases and Random Matrices, London Mathematical Society Monographs Series, vol. 34, Prince-
ton University Press, Princeton, NJ, 2010.
[28] G. Livan, M. Novaes, P. Vivo, Introduction to Random Matrices: Theory and Practice, SpringerBriefs in Mathemat-
ical Physics, vol. 26, Springer, Switzerland, 2018, arXiv:math-ph/0402061.
[29] R.J. Muirhead, Aspects of Multivariate Statistical Theory, Wiley Series in Probability and Statistics, vol. 197, Wiley,
2009.
[30] M. Adler, P.J. Forrester, T. Nagao, P. van Moerbeke, Classical skew orthogonal polynomials and random matrices,
J. Stat. Phys. 99 (2000) 141, arXiv:solv-int/9907001.
[31] P.J. Forrester, Vicious random walkers in the limit of a large number of walkers, J. Stat. Phys. 56 (1989) 767.
[32] M. Tierz, Soft matrix models and Chern-Simons partition functions, Mod. Phys. Lett. A 19 (2004) 1365, arXiv:
hep-th/0212128.
[33] G. Szegő, Orthogonal Polynomials, vol. 23, American Mathematical Soc., 1939.
[34] A. Schwartzman, Random ellipsoids and false discovery rates: Statistics for diffusion tensor imaging data, Ph.D.
thesis, Stanford University, 2006.
[35] A. Schwartzman, Lognormal distributions and geometric averages of symmetric positive definite matrices, Int. Stat.
Rev. 84 (2016) 456.
[36] A. Altland, M.R. Zirnbauer, Nonstandard symmetry classes in mesoscopic normal-superconducting hybrid struc-
tures, Phys. Rev. B 55 (1997) 1142, arXiv:cond-mat/9602137.
[37] M.R. Zirnbauer, Symmetry classes, in: G. Akemann, J. Baik, P. Di Francesco (Eds.), The Oxford Handbook of Ran-
dom Matrix Theory, in: Oxford Handbooks in Mathematics, Ch., vol. 3, Oxford University Press, Great Clarendon
Street, Oxford, UK, 2011, arXiv:1001.0722.
[38] M. Caselle, U. Magnea, Random matrix theory and symmetric spaces, Phys. Rep. 394 (2004) 41, arXiv:cond-mat/
0304363.
[39] A. Edelman, S. Jeong, The generalized Cartan decomposition for classical random matrix ensembles, arXiv:2011.
08087.
[40] F.J. Dyson, The threefold way. Algebraic structure of symmetry groups and ensembles in quantum mechanics, J.
Math. Phys. 3 (1962) 1199.
[41] M. Romo, M. Tierz, Unitary Chern-Simons matrix model and the Villain lattice action, Phys. Rev. D 86 (2012)
045027, arXiv:1103.2421.
[42] G. Giasemidis, M. Tierz, Torus knot polynomials and susy Wilson loops, Lett. Math. Phys. 104 (2014) 1535, arXiv:
1401.8171.
[43] Y. Takahashi, M. Katori, Oscillatory matrix model in Chern-Simons theory and Jacobi-theta determinantal point
process, J. Math. Phys. 55 (2014) 093302, arXiv:1312.5848.
[44] D. Garcia-Garcia, M. Tierz, Matrix models for classical groups and Toeplitz±Hankel minors with applications to
Chern-Simons theory and fermionic models, J. Phys. A 53 (2020), arXiv:1901.08922.
[45] C. Andréief, Note sur une relation entre les intégrales définies des produits des fonctions, Mém. Soc. Sci. Phys. Nat.
Bordeaux 2 (1886) 1.
[46] P.J. Forrester, Meet Andréief, Bordeaux 1886, and Andreev, Kharkov 1882–1883, Random Matrices: Theory Appl.
08 (2019) 1930001, arXiv:1806.10411.
[47] G. Giasemidis, R.J. Szabo, M. Tierz, Supersymmetric gauge theories, Coulomb gases and Chern-Simons matrix
models, Phys. Rev. D 89 (2014) 025016, arXiv:1310.3122.
[48] R.J. Baxter, Statistical mechanics of a one-dimensional Coulomb system with a uniform charge background, Math.
Proc. Camb. Philos. Soc. 59 (1963) 779–787.
[49] A. Dhar, A. Kundu, S.N. Majumdar, S. Sabhapandit, G. Schehr, Exact extremal statistics in the classical 1D
Coulomb gas, Phys. Rev. Lett. 119 (2017) 060601, arXiv:1704.08973.
[50] A. Dhar, A. Kundu, S.N. Majumdar, S. Sabhapandit, G. Schehr, Extreme statistics and index distribution in the
classical 1d Coulomb gas, J. Phys. A 51 (2018) 295001, arXiv:1802.10374.
[51] M. Aganagic, A. Klemm, M. Marino, C. Vafa, Matrix model as a mirror of Chern-Simons theory, J. High Energy
Phys. 02 (2004) 010, arXiv:hep-th/0211098.
[52] M. Marino, Les Houches lectures on matrix models and topological strings, arXiv:hep-th/0410165, 2004.
[53] N.G. de Bruijn, On some multiple integrals involving determinants, J. Indian Math. Soc. 19 (1955) 133.
[54] M. Wimmer, Efficient numerical computation of the Pfaffian for dense and banded skew-symmetric matrices, ACM
Trans. Math. Softw. 38 (2012) 1, arXiv:1102.3440.
[55] M. Adler, E. Horozov, P. van Moerbeke, The Pfaff lattice and skew-orthogonal polynomials, Int. Math. Res. Not. 11
(1999) 569, arXiv:solv-int/9903005.

29
L. Santilli and M. Tierz Nuclear Physics B 973 (2021) 115582

[56] X.-K. Chang, Y. He, X.-B. Hu, S.-H. Li, Partial-skew-orthogonal polynomials and related integrable lattices with
Pfaffian tau-functions, Commun. Math. Phys. 364 (2018) 1069–1119, arXiv:1712.06382.
[57] S. Said, N. Le Bihan, J.H. Manton, Riemannian Gaussian distributions on the space of positive-definite quaternion
matrices, in: B.F.F. Nielsen (Ed.), International Conference on Geometric Science of Information, in: Lecture Notes
in Computer Science, vol. 10589, Springer, 2017, pp. 709–716, arXiv:1703.09940.
[58] M.E. Fisher, Walks, walls, wetting, and melting, J. Stat. Phys. 34 (1984) 667.
[59] S. Karlin, J. McGregor, Coincidence probabilities, Pac. J. Math. 9 (1959) 1141.
[60] D.J. Grabiner, Brownian motion in a Weyl chamber, non-colliding particles, and random matrices, Ann. Inst. Henri
Poincaré B, Probab. Stat. 35 (1999) 177, arXiv:math/9708207.
[61] M. Katori, Bessel Processes, Schramm-Loewner Evolution, and the Dyson Model, SpringerBriefs in Mathematical
Physics, vol. 11, Springer, Singapore, 2016.
[62] J. Baik, P. Deift, T. Suidan, Combinatorics and Random Matrix Theory, Graduate Studies in Mathematics, vol. 172,
American Mathematical Society, Providence, RI, 2016.
[63] F.J. Dyson, A Brownian-motion model for the eigenvalues of a random matrix, J. Math. Phys. 3 (1962) 1191.
[64] R.R. Coifman, S. Lafon, Diffusion maps, Appl. Comput. Harmon. Anal. 21 (2006) 5.
[65] Z.I. Botev, J.F. Grotowski, D.P. Kroese, Kernel density estimation via diffusion, Ann. Stat. 38 (2010) 2916, arXiv:
1011.2602.
[66] S. de Haro, M. Tierz, Brownian motion, Chern-Simons theory, and 2-D Yang-Mills, Phys. Lett. B 601 (2004) 201,
arXiv:hep-th/0406093.
[67] Y. Dolivet, M. Tierz, Chern-Simons matrix models and Stieltjes-Wigert polynomials, J. Math. Phys. 48 (2007)
023507, arXiv:hep-th/0609167.
[68] M. Tierz, Schur polynomials and biorthogonal random matrix ensembles, J. Math. Phys. 51 (2010) 063509.
[69] J. Baik, T.M. Suidan, Random matrix central limit theorems for nonintersecting random walks, Ann. Probab. 35
(2007) 1807, arXiv:math/0605212.
[70] M. Katori, H. Tanemura, Scaling limit of vicious walks and two-matrix model, Phys. Rev. E 66 (2002) 011105,
arXiv:cond-mat/0203549.
[71] M. Katori, H. Tanemura, Symmetry of matrix-valued stochastic processes and noncolliding diffusion particle sys-
tems, J. Math. Phys. 45 (2004) 3058, arXiv:1712.07903.
[72] P.J. Forrester, J.R. Ipsen, D.-Z. Liu, L. Zhang, Orthogonal and symplectic Harish-Chandra integrals and matrix
product ensembles, Random Matrices: Theory Appl. 8 (2019) 1950015, arXiv:1711.10691.
[73] H. Hendriks, Nonparametric estimation of a probability density on a Riemannian manifold using Fourier expansions,
Ann. Stat. 18 (1990) 832.
[74] G.G. Walter, Properties of Hermite series estimation of probability density, Ann. Stat. 5 (1977) 1258.
[75] P.J. Forrester, Properties of an exact crystalline many-body ground state, J. Stat. Phys. 76 (1994) 331.
[76] P.J. Forrester, Global and local scaling limits for the β = 2 Stieltjes-Wigert random matrix ensemble, arXiv:2011.
11783.
[77] A. Mays, A. Ponsaing, G. Schehr, Tracy-Widom distributions for the Gaussian orthogonal and symplectic ensembles
revisited: a skew-orthogonal polynomials approach, arXiv:2007.14597.

30