ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS

ON RIEMANNIAN SUBMANIFOLDS OF Rn

JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

arXiv:2506.13238v1 [math.DG] 16 Jun 2025

Abstract. The Gaussian integral operator arises naturally as a local Euclidean approximation
of the heat semigroup on a Riemannian manifold and plays a pivotal role in the analysis of graph
Laplacians, particularly within the frameworks of manifold learning and spectral graph theory.
In this paper, we study the asymptotic behavior of the Gaussian integral operator on a smooth
Riemannian submanifold M ⊂ Rn , focusing on its expansion as ε → 0+ . Under the assumption
that the input function is real analytic near a fixed point x ∈ M , we derive a full asymptotic
expansion of the operator and compute the first-order correction term explicitly in terms of the
mean curvature vector and the scalar curvature of the submanifold. In particular, we apply our
results to hypersurfaces in Euclidean space and investigate geometric conditions under which points
exhibit equicurvature.

Keywords: Gaussian integral operator, Gaussian kernel, Laplace–Beltrami operator,

asymptotic expansion, mean curvature, scalar curvature, graph Laplacian, manifold
learning, diffusion maps.

1. Introduction

The Gaussian integral operator plays a crucial role in the analysis of graph Lapla-
cians, particularly in the fields of manifold learning and spectral graph theory. By
examining the spectral properties of the graph Laplacian, one can effectively per-
form dimensionality reduction by embedding data into a lower-dimensional Euclidean
space. For example, diffusion maps utilize the graph Laplacian to uncover intrinsic
low-dimensional structures within high-dimensional datasets. For related works, see
[2], [7], and [8].
In the context of manifold learning, the normalized graph Laplacian serves as a
discrete analogue of the Laplace–Beltrami operator on a Riemannian manifold. A
crucial step in establishing the convergence of normalized graph Laplacians to their
continuous counterparts involves a detailed analysis of the Gaussian integral oper-
ator. This investigation is foundational to both the theoretical understanding and
the practical implementation of these operators in analyzing the geometric structure
of high-dimensional data. For studies on the convergence of the normalized graph
1
2 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

Laplacian when data points lie on a submanifold M ⊂ Rn , see, for example, [9], [12],
[13], [14], [11], [3], [5], [15], [7], etc.
Before introducing the definition of the Gaussian integral operator, we briefly recall
its motivation in a broader analytical context. The origin of this operator lies in the
theory of heat operators. The heat operator Ht = et∆ on a d-dimensional compact,
oriented Riemannian manifold1 M , is defined as the one-parameter semigroup gener-
ated by the Laplace–Beltrami operator ∆. It can be expressed as a family of integral
operators parameterized by time t > 0:
Z
(Ht f )(x) = H(x, y, t)f (y) dV (y),
M
where dV denotes the Riemannian volume form on M , and H : M × M × [0, ∞) → R
is the heat kernel, which satisfies the heat equation
∂H
= ∆x H.
∂t
When both the geodesic distance d(x, y) and the time t are sufficiently small, the heat
kernel admits the following classical asymptotic expansion:
d(x, y)2
 
1
u0 (x, y) + tu1 (x, y) + O(t2 ) ,


H(x, y, t) = d/2
exp −
(4πt) 4t
where u0 (x, y) = 1 + O(d(x, y)2 ).
In practical applications, particularly when M is viewed as a Riemannian submani-
fold of Rn , it is often more convenient to use the ambient Euclidean distance ∥x−y∥Rn
instead of the intrinsic geodesic distance d(x, y). This leads naturally to the formula-
tion of the Gaussian integral operator, constructed using a Gaussian kernel based on
Euclidean distance. Specifically, the Gaussian kernel kε : M × M → R is defined by
∥y − x∥2Rn
 
1
kε (x, y) = exp − ,
(4πε)d/2 4ε
where ε > 0 is a small positive parameter.
For each ε > 0, we define the associated Gaussian integral operator

Kε : C ∞ (M ) → C ∞ (M ),

by Z
(Kε f )(x) = kε (x, y)f (y) dV (y),
M
where dV is the Riemannian volume form on M .
1Throughout this paper, all manifolds are assumed to be connected.
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF Rn3

Our investigation into the asymptotic behavior of Kε is motivated by its role in

the convergence analysis of the normalized graph Laplacian to the Laplace–Beltrami
operator on Riemannian submanifolds of Euclidean space. According to [13], it was
claimed that in unpublished notes, G. Tokarev was the first to derive the leading
two terms in the asymptotic expansion of the Gaussian integral operator. The works
concerning the convergence of graph Laplacians can be viewed as early steps toward a
systematic understanding of the asymptotic behavior of Gaussian integral operators.
In this paper, we aim to derive the asymptotic expansion of (Kε f )(x) as ε → 0+ ,
assuming that f is analytic in a neighborhood of x ∈ M . Specifically, we seek a
sequence of smooth functions {an : M → R}∞
n=0 such that

∞
X
(Kε f )(x) ∼ an (x) εn ,
n=0

meaning that for each N ∈ N, the remainder satisfies

N
X
(Kε f )(x) − an (x) εn = o(εN ) as ε → 0+ .
n=0

The coefficients an (x) can be defined inductively by the recurrence

(Kε f )(x) − nk=0 ak (x) εk

P
an+1 (x) := lim+ ,
ε→0 εn+1

provided that the limit exists.

In Section 2, we analyze the relationship between Euclidean and geodesic distances
for points on a submanifold of Euclidean space and revisit the Taylor expansion of
the Riemannian volume form, following the work of A. Gray [1]. In Section 3, we
compute the asymptotic expansion of the Gaussian integral operator and express the
coefficient a1 explicitly in terms of the mean curvature vector and scalar curvature. In
Section 4, we apply the Gaussian integral operator to hypersurfaces in Euclidean space
to investigate conditions under which points on a hypersurface exhibit equicurvature.
In particular, we classify the equicurved condition in the case where the hypersurface
is a surface in R3 . To the best of our knowledge, no explicit examples of compact
equicurved hypersurfaces are known when d ≥ 3. This observation suggests a natural
and intriguing open problem in differential geometry.
4 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

2. Geometric Expansions in Normal Coordinates

In this section, we present the Taylor expansions of two essential quantities in

normal coordinates required for computing the asymptotic expansion of the Gauss-
ian integral operator. Specifically, we derive the Taylor expansion of the squared
Euclidean distance function composed with the exponential map, as well as the ex-
pansion of the Riemannian volume form expressed in geodesic normal coordinates.
We also obtain the corresponding expansions in geodesic polar coordinates, where the
local geometry is characterized in terms of radial distance and directions on the unit
sphere.
In [12] and [13], O. G. Smolyanov, H. v. Weizsäcker, and O. Wittich prove the
following result:

Theorem 2.1 ([12], [13]). Consider points x, y ∈ M ⊂ N , where M is an isomet-

rically embedded submanifold via the map ϕ. Let p ∈ U (x), where U (x) ⊂ L is
M
sufficiently small so that x and y are joined by a unique minimizing geodesic γxy in
M
M starting at x. Assume γxy is parametrized by arc length. Then
dN (ϕ(x), ϕ(y))2 − dM (x, y)2 1 M M 2
lim 4
=− Bϕ (γ̇xy , γ̇xy ) ,
dM (x,y)→0 dM (x, y) 12
where Bϕ denotes the second fundamental form of the embedding ϕ : M → N .

This identity is equivalent to the expansion:

dM (x, y)4 2
dN (ϕ(x), ϕ(y))2 = dM (x, y)2 − M
Bϕ (γ̇xy M
, γ̇xy ) + O(dM (x, y)5 ).
12
Their proof uses the Taylor expansion of the Riemannian metric in geodesic normal
coordinates to expand dN (·, ϕ(x))2 near x. In this section, we provide a much simpler
proof of this identity in the special case where the ambient manifold is Rn . To this
end, we begin by reviewing some basic facts about the Taylor expansion of scalar
functions.
Recall that a function h : U ⊂ Rd → R, defined on an open neighborhood U of the
origin, is said to be analytic at the origin if there exists δ > 0 such that its Taylor
series expansion
∞
X 1
(2.1) h(s) = Hj (h)(s)
j=0
j!
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF Rn5

converges on the open ball B0 (δ) ⊂ Rd . Here, s = (s1 , . . . , sd ) denotes the standard
coordinates on Rd , and each term Hj (h) is given by
X j 
Hj (h)(s) = (Dα h)(0) sα ,
α
|α|=j

where α = (α1 , . . . , αd ) ∈ Zd+ is a multi-index of total degree |α| = α1 + · · · + αd , and

∂ |α| h
 
j j!
= , sα = sα1 1 · · · sαd d , Dα h(0) = .
α α1 ! · · · αd ! ∂sα1 1 · · · ∂sαd d s=0
Each Hj (h) is either zero or a homogeneous polynomial of degree j. For notational
convenience, we write Hj (h) = hj .
Now let s = sv, where 0 < s < δ and v ∈ S d−1 . Then the analyticity of h, together
with the homogeneity of each hj , implies that
∞
X sj
(2.2) h(sv) = hj (v).
j=0
j!

Consequently, the function h, when restricted to the punctured ball B0′ (δ) := B0 (δ) \
{0}, admits the expansion (2.2) in polar coordinates (s, v).
Let M be a Riemannian submanifold of dimension d in Rn , and let x ∈ M be
a point. Denote by injx (M ) the injectivity radius of M at x. Choose an ordered
orthonormal basis {(e1 )x , . . . , (ed )x } of the tangent space Tx M . For any δ > 0 such
that δ < injx (M ), define the map x : B0 (δ) → M by
d
!
X
x(s1 , . . . , sd ) = expx si (ei )x ,
i=1

where B0 (δ) ⊂ Rd denotes the open Euclidean ball of radius δ centered at the origin.
The map x is a diffeomorphism from B0 (δ) onto its image Bx (δ) ⊂ M , which is an
open neighborhood of x. We refer to the map x, or equivalently to the coordinate
functions (s1 , . . . , sd ), as a system of geodesic normal coordinates centered at x. The
image Bx (δ) is called the geodesic ball of radius δ centered at x. Note that the system
of geodesic normal coordinates depends on the choice of the ordered orthonormal
basis of Tx M .
Let g : B0 (δ) → R be the function defined by

g(s) = ∥x(s) − x∥2Rn ,

6 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

where x : B0 (δ) → M ⊂ Rn is the exponential coordinate chart centered at x ∈ M .

The function g is evidently smooth, and we further assume it is analytic on the
open ball B0 (δ). In what follows, we briefly outline the computation of the Taylor
expansion of g about the origin. It is immediate that g(0) = 0, and the first-order
partial derivatives are given by
 
∂g ∂x
=2 , x(s) − x ,
∂si ∂si
∂g
so in particular ∂si
(0) = 0 for all i. Applying the product rule, we compute the
second-order derivatives:
∂ 2g ∂ 2x
   
∂x ∂x
=2 , x(s) − x + 2 ,
∂si ∂sj ∂si ∂sj ∂si ∂sj
 2 
∂ x
=2 , x(s) − x + 2gij (s).
∂si ∂sj
Evaluating at s = 0, we find
∂ 2g
(0) = 2gij (0) = 2δij .
∂si ∂sj
Furthermore, since the Christoffel symbols satisfy Γkij (x) = 0 at the center of normal
coordinates, it follows that
∂ 3x
(0) = 0.
∂si ∂sj ∂sk
More generally, higher-order derivatives Dα g(0) can be computed by repeated ap-
plication of the product rule, taking into account the symmetries of the exponential
map and the vanishing of certain derivatives at the origin. Consequently, the Taylor
expansion of g(s) near the origin begins as

g(s) = s2 + O(s4 ),
p
where s = s21 + · · · + s2d . This expression reflects the fact that

∥y − x∥2Rn = d(x, y)2 + O(d(x, y)4 ),

for any point y ∈ Bx (δ), where d(x, y) denotes the intrinsic (geodesic) distance be-
tween x and y. The fourth-order term can be expressed in terms of the second
fundamental form via the Gauss equation. Details can be found in [12] and [13].
We now present an alternative approach to derive the geometric expansion of the
squared Euclidean norm. By restricting the function g to the punctured ball B0′ (δ),
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF Rn7

and evaluating its Taylor expansion along each ray of the form s = sv, we obtain the
corresponding expansion in polar coordinates (s, v):
∞
X sj
g(s) = gj (v),
j=0
j!

where each gj (v) is the homogeneous degree-j component of the Taylor series in the
direction v ∈ S d−1 . From previous computations, we have

g0 (v) = 0, g1 (v) = 0, g2 (v) = 2, g3 (v) = 0.

To compute the coefficients gj more directly, we use the series expansion of g in polar
coordinates. Specifically, for each fixed direction v ∈ S d−1 , we define the one-variable
function
gv : (−δ, δ) → R, gv (t) := g(tv),
which corresponds to the restriction of g along the radial line in the direction v.
Then, the coefficient gj (v) in the polar expansion of g is given by the j-th derivative
of gv evaluated at the origin:
dj
gj (v) = gv (t) .
dtj t=0

Let v = (v1 , . . . , vd ) ∈ S d−1 be a unit vector. With respect to the chosen orthonor-
mal basis {(ei )x }di=1 of Tx M , we associate the corresponding tangent vector vx ∈ Tx M
by vx = di=1 vi (ei )x . Let
P

γvx : (−δ, δ) → Bx (δ)

denote the unique geodesic satisfying the initial conditions

γvx (0) = x, γv′ x (0) = vx .

Then, for all |t| < δ, we have

γvx (t) = expx (tvx ).

It follows from the definition that

gv (t) = ∥γvx (t) − x∥2Rn .

We now proceed to compute the Taylor expansion of gv (t) about t = 0. To express

the derivatives of gv (t) at t = 0, we require the notion of the second fundamental
form of a submanifold M ⊂ Rn . We briefly review its definition below.
8 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

Let ∇ be the Riemannian connection on Rn . For each point x ∈ M , we have the

orthogonal direct sum decomposition

Tx Rn = Tx M ⊕ Tx M ⊥ .

We denote the orthogonal projections from Tx Rn onto Tx M and Tx M ⊥ by Tx and ⊥x ,

respectively.
Let X and Y be vector fields defined on an open neighborhood U ⊂ M of a point
x ∈ M . Suppose X and Y are smooth vector fields defined on an open neighborhood
U ⊂ Rn of x, with U ∩ M = U , such that X = X and Y = Y on U . Then, for any
y ∈ U , the Levi-Civita connection ∇ on M satisfies


(∇X Y )(y) = Ty ∇X Y (y) .

We define the second fundamental form

Bx : Tx M × Tx M → Tx M ⊥

by Bx (X(x), Y (x)) = (∇X(x) Y )⊥x . The following result provides the foundation for
the expansion of gv .

Lemma 2.1. Let v ∈ S d−1 be a unit vector. The function gv defined above admits
a Taylor expansion around t = 0 of the form
∞
X gj (v) j
gv (t) = t, for |t| < δ.
j=0
j!
Each coefficient gj (v) is either zero or a homogeneous polynomial of degree j in
v = (v1 , . . . , vd ). The first few coefficients are given by

g0 (v) = 0, g1 (v) = 0, g2 (v) = 2, g3 (v) = 0, g4 (v) = −2∥Bx (vx , vx )∥2 ,

where Bx (vx , vx ) denotes the second fundamental form of the submanifold M ⊂ Rn

at the point x, evaluated on the vector vx ∈ Tx M .

(k) (k)
Proof. By definition, gk (v) = gv (0), so it suffices to compute the derivatives gv (0).
Recall that
gv (t) = ⟨γvx (t) − x, γvx (t) − x⟩Rn .
(k) (k)
Define ξvx (t) := γvx (t) − x. Then ξvx (0) = 0, and ξvx (t) = γvx (t) for all k ≥ 1. Since

gvx (t) = ⟨ξvx (t), ξvx (t)⟩,

ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF Rn9

we apply the Leibniz rule to obtain

k  
X k
gv(k) (0) = ⟨ξv(j)x (0), ξv(k−j)
x
(0)⟩.
j=0
j

The first few derivatives are:

gv′ (0) = 2⟨ξvx (0), ξv′ x (0)⟩ = 2⟨0, vx ⟩ = 0,

gv′′ (0) = 2∥γv′ x (0)∥2 = 2∥vx ∥2 = 2,
gv(3) (0) = 6⟨γv(2)
x
(0), γv′ x (0)⟩ = 6⟨γv(2)
x
(0), vx ⟩.

Since γvx is a geodesic in M , it satisfies

⟨γv(2)
x
(t), γv′ x (t)⟩ = 0 for all t,

so in particular,
gv(3) (0) = 6⟨γv(2)
x
(0), vx ⟩ = 0.

Differentiating the identity above with respect to t, we obtain

⟨γv(3)
x
(t), γv′ x (t)⟩ + ∥γv(2)
x
(t)∥2 = 0.

Evaluating at t = 0, we find

⟨γv(3)
x
(0), vx ⟩ = −∥γv(2)
x
(0)∥2 .

Now we compute the fourth derivative:

gv(4) (0) = 8⟨γv(3)

x
(0), γv′ x (0)⟩ + 6∥γv(2)
x
(0)∥2
= 8 · −∥γv(2) (0)∥2 + 6∥γv(2) (0)∥2


x x

= −2∥γv(2)
x
(0)∥2 .
(2)
Finally, using the fact that γvx (0) = Bx (vx , vx ), we obtain

gv(4) (0) = −2∥Bx (vx , vx )∥2 ,

as claimed. □

Let Bx′ (δ) represent the punctured geodesic ball centered at x, defined as follows:

Bx′ (δ) = Bx (δ) \ {x}.

Building on the previous result, we now present the following corollary, which provides
the expansion we aimed to demonstrate:
10 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

Corollary 2.1. For every y ∈ Bx′ (δ), the squared Euclidean distance can be expressed
as:
s4
∥Bx (vx , vx )∥2 + O(s5 ),
∥y − x∥2Rn = s2 −
12
where s = d(x, y) and vx ∈ Tx M is the unique unit tangent vector corresponding to
the point y = expx (svx ).

We now recall a result from [1] concerning the Taylor expansion of the volume
form on a Riemannian manifold. To state this result precisely, we first introduce the
necessary terminology.
Let ∇ denote the Levi-Civita connection on the Riemannian manifold M . For
smooth vector fields X and Y , the Riemann curvature operator RXY is defined by

RXY := ∇[X,Y ] − [∇X , ∇Y ],

where [X, Y ] denotes the Lie bracket of X and Y , and [∇X , ∇Y ] := ∇X ∇Y − ∇Y ∇X

is the commutator of covariant derivatives.
Let Y1 , . . . , Ym be smooth vector fields on M , and let Z denote a tensor field on
M . The m-th covariant derivative of Z along the vector fields Y1 , . . . , Ym is defined
inductively by:
 
∇m
Y1 ···Ym Z = ∇ Ym ∇m−1
Y1 ···Ym−1 Z ,

with the base case given by ∇1Y Z = ∇Y Z.

We now fix normal coordinates x : B0 (δ) → Bx (δ), as previously described, and we
define the associated local frame fields:
∂
Xi = , for 1 ≤ i ≤ d.
∂si
The collection {Xi }di=1 forms a smooth local frame in the geodesic ball Bx (δ). For
convenience, we adopt the following notation:
d
X
∇ki1 ···ik = ∇kXi1 ···Xi , Rijkl = ⟨RXi Xj Xk , Xl ⟩, Rij = Rikjk ,
k
k=1

where 1 ≤ i, j ≤ d and 1 ≤ i1 , . . . , ik ≤ d.
The pullback of the Riemannian volume form dV under the normal coordinate
parametrization x : B0 (δ) → Bx (δ) can be expressed as

x∗ dV = ρ(s) ds1 ∧ · · · ∧ dsd ,

 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R11

where ρ(s) is a smooth positive function representing the volume density in normal
coordinates. The following result is attributed to Alfred Gray:

Theorem 2.2. In a neighborhood of 0 ∈ Rd , the volume density function ρ(s1 , . . . , sd ),

defined by the pullback of the Riemannian volume form under normal coordinates
centered at x ∈ M , admits the Taylor expansion:
∞
X 1
ρ(s) = ρk (s),
k=0
k!
where each ρk is either zero or a homogeneous polynomial of degree k in s = (s1 , . . . , sd ).
The first few coefficients are given by:
d
1X
ρ0 (s) = 1, ρ1 (s) = 0, ρ2 (s) = − Rij (x)si sj ,
3 i,j=1
d
1 X
ρ3 (s) = − (∇i Rjk )(x)si sj sk ,
2 i,j,k=1
d d
!
X 3 2 1 2 X
ρ4 (s) = − ∇ij Rkl (x) + Rij (x)Rkl (x) − Riajb (x)Rkalb (x) si sj sk sl .
i,j,k,l=1
5 3 15 a,b=1
Here, Rij (x) denotes the Ricci curvature at x, ∇i Rjk (x) denotes the covariant deriv-
ative of the Ricci tensor, and Riajb (x) are the components of the Riemann curvature
tensor at x in normal coordinates.

Let P : (0, δ) × S d−1 → B0 (δ) denote the standard polar coordinate map in Rd ,
defined by P (s, v) = sv, and

y : (0, δ) × S d−1 → Bx′ (δ)

be the map defined by y := x ◦ P. We refer to y as the system of geodesic polar

coordinates on the punctured geodesic ball Bx′ (δ). Since y∗ dV = P ∗ (x∗ dV ), and
P (s, v) = sv is the standard polar coordinate map on Rd , we can express the pullback
of the Riemannian volume form dV under y using the Jacobian determinant of polar
coordinates. This yields:

y∗ dV = ρ(sv) sd−1 ds ∧ dσ(v),

where dσ(v) denotes the standard volume form on the unit sphere S d−1 .Using the
Taylor expansion of ρ and its corresponding representation in polar coordinates, we
obtain the following result (assuming ρ is analytic):
12 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

Corollary 2.2. The pullback of the volume form y∗ dV under geodesic polar coordi-
nates admits the expansion
∞
!
X sj
y∗ dV = ρj (v) sd−1 ds ∧ dσ(v),
j=0
j!

where each function ρj (v) denotes the restriction of the homogeneous polynomial
ρj (s) of degree j to the unit sphere S d−1 .

3. Asymptotic Expansion of the Gaussian integral operator

Let M be a d-dimensional compact oriented Riemannian submanifold of Rn , and

let x ∈ M . Choose δ > 0 such that δ < injx (M ). Denote by Bx (δ) the geodesic ball
centered at x with radius δ, as defined in the previous section.
Since M \ Bx (δ) is compact, there exists a point yδ ∈ M \ Bx (δ) such that

∥yδ − x∥ = inf{∥y − x∥ : y ∈ M \ Bx (δ)}.

We define mδ := ∥yδ − x∥, which satisfies mδ > 0 by construction.

For each function f ∈ C ∞ (M ), the Gaussian kernel operator Kε can be decomposed
as
Z Z
Kε f (x) = kε (x, y)f (y) dVy + kε (x, y)f (y) dVy .
Bx (δ) M \Bx (δ)

To estimate the second term, observe that

m2
Z
d δ
kε (x, y)f (y) dVy ≤ (4πε)− 2 e− 4ε Vol(M )∥f ∥∞ ,
M \Bx (δ)

where Vol(M ) denotes the total volume of M . Since

d m2
δ
(4πε)− 2 e− 4ε
lim+ = 0 for all k ∈ N,
ε→0 εk
it then follows that
Z
kε (x, y)f (y) dVy = o(εk ) for all k ∈ N.
M \Bx (δ)

To estimate the first term, we rewrite the integral using geodesic normal coordinates
x : B0 (δ) → Bx (δ). This gives
Z Z
∥x(s)−x∥2
− d2
kε (x, y)f (y) dVy = (4πε) e− 4ε f (x(s)) x∗ dVy .
Bx (δ) B0 (δ)
 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R13

We now further assume that both the function fe = f ◦ x and the density function ρ,
associated with the pullback volume form

x∗ dVy = ρ(s) ds1 ∧ · · · ∧ dsd ,

are real analytic on B0 (δ), and that their Taylor series converge uniformly on this
domain.
To begin, we compute the Taylor expansion of the product fe(s) x∗ dVy in geodesic
polar coordinates. Since
∞ ∞
!
X sk e X sj
f (s) =
e fk (v), x∗ dVy = ρj (v) sd−1 ds ∧ dσ(v),
k=0
k! j=0
j!
their product admits the expansion
∞
!
ℓ
∗
X s
fe(s) x dVy = αℓ (v) sd−1 ds ∧ dσ(v),
ℓ=0
ℓ!
where each coefficient αℓ (v) is either zero or a homogeneous polynomial of degree ℓ,
and is given by the convolution formula
ℓ  
X ℓ e
αℓ (v) = fj (v) ρℓ−j (v) for each ℓ ≥ 0.
j=0
j
Now, we define the function q : B0 (δ) → R by

q(s) = g(s) − s2 ,

where g : B0 (δ) → R is the squared Euclidean distance function defined in the

p
previous section, and s = s21 + · · · + s2d . Then q is analytic at the origin and admits
a Taylor expansion of the form
∞
X 1
q(s) = qj (s),
j=0
j!
which converges on the open ball B0 (δ). Here, each qj (s) is the degree-j homogeneous
component in the Taylor expansion of q. By Lemma 2.1, the first few terms vanish:

q0 (s) = 0, q1 (s) = 0, q2 (s) = 0, q3 (s) = 0.

Since q(s) = g(s) − s2 , we may write

∥x(s)−x∥2 s2 q(s)
e− 4ε = e− 4ε · e− 4ε .
q(s)
Hence, it suffices to compute the Taylor expansion of the function e− 4ε about s = 0.
To carry out this computation systematically, we introduce the partial exponential
14 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

Bell polynomials, which provide a convenient framework for expressing the Taylor
q(s)
expansion of composite functions such as e− 4ε .

Definition 3.1. The partial exponential Bell polynomial Bm,k (x1 , x2 , . . . , xm−k+1 ) is
defined for integers m ≥ 1 and 1 ≤ k ≤ m by
 jm−k+1
X m!  x  j1  x  j2
1 2 xm−k+1
Bm,k (x1 , x2 , . . . , xm−k+1 ) = ··· ,
j1 !j2 ! · · · jm−k+1 ! 1! 2! (m − k + 1)!
where the sum is taken over all sequences of non-negative integers (j1 , j2 , . . . , jm−k+1 )
satisfying the constraints
m−k+1
X m−k+1
X
ji = k and iji = m.
i=1 i=1

It is established that the generating function for the partial exponential Bell poly-
nomials can be expressed as follows:
∞
! ∞ m
X tj X tm X k
(3.1) exp u xj =1+ u Bm,k (x1 , x2 , . . . , xm−k+1 ).
j=1
j! m=1
m! k=1

By substituting xj = qj (v) and t = s into equation (3.1), we derive

∞ m
−
q(s) X sm X k
e 4ϵ =1+ u bm,k (v),
m=1
m! k=1

where u = (−4ϵ)−1 , and bm,k (v) = Bm,k q1 (v), q2 (v), . . . , qm−k+1 (v) . Since each qi

is either zero or a homogeneous polynomial of degree i, it follows from the definition

of the partial exponential Bell polynomials that each bm,k (v) is also either zero or
a homogeneous polynomial of degree m. In particular, given that q1 (v) = q2 (v) =
q3 (v) = 0, any monomial involving these terms will vanish. Thus, a necessary and
sufficient condition for bm,k (v) to be a nonzero homogeneous polynomial is that all
terms in its expansion must involve only qj with j ≥ 4. This requirement translates
to the inequality: m ≥ 4k. Let us define β0 (v, u) = 1, and for m ≥ 1,
m
X
βm (v, u) = uk bm,k (v).
k=1

Notice that bm,k (v) = 0 when k > m/4. Then the exponential factor admits the
following expansion
∞
q(s) X sm
e− 4ε = βm (v, u).
m=0
m!
 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R15

Define
ℓ  
X ℓ
aℓ (v, u) = αℓ−m (v) βm (v, u).
m=0
m
Since each αj (v) is either zero or a homogeneous polynomial of degree j, and each
βk (v, u) is either zero or a homogeneous polynomial of degree k, their product yields
a finite sum of homogeneous polynomials of total degree ℓ. Moreover, by symmetry,
the integral of any homogeneous polynomial of odd degree over the unit sphere S d−1
vanishes. Therefore,
Z
aℓ (v, u) dσ(v) = 0 whenever ℓ is odd.
S d−1
It thus suffices to consider the case ℓ = 2p, where p ∈ N0 . In that case, we have
Z Z 2p   Z
X 2p
a2p (v, u) dσ(v) = α2p (v) dσ(v) + α2p−m (v) βm (v, u) dσ(v).
S d−1 S d−1 m=1
m S d−1

We now define the scalar coefficients

Z Z
1 1
ηp = α2p (v) dσ(v), wp,m,k = α2p−m (v) bm,k (v) dσ(v),
ωd S d−1 ωd S d−1
for 1 ≤ k ≤ m4 , where ωd = Vol(S d−1 ) denotes the volume of the unit sphere in Rd .
 

Then we obtain
Z 2p m  
1 X X 2p
a2p (v, u) dσ(v) = ηp + (−4ε)−k wp,m,k .
ωd S d−1 m=1 k=1
m
Since
Z ∞ Z Z δ 
−
∥x(s)−x∥2
∗
X 1 2
− s4ε ℓ+d−1
e 4ε f (x(s)) x dVy = e s ds aℓ (v, u) dσ(v),
B0 (δ) ℓ=0
ℓ! S d−1 0

we obtain the expansion

∞ 2p m  
Z !
X ωd cp (ε) X X 2p 1
kε (x, y)f (y) dVy = ηp + wp,m,k ,
Bx (δ) p=0
(2p)! m=1 k=1
m (−4ε)k
where Z δ
1 s2
cp (ε) = d e− 4ε s2p+d−1 ds.
(4πε) 0 2

This power series can be reorganized as

∞ q 2q+2k
Z !
X ωd cq (ε) X X ωd cq+k (ε)
kε (x, y)f (y) dVy = ηq + (−4ε)−k wq+k,m,k .
Bx (δ) q=0
(2q)! k=1 m=4k
m!(2q + 2k − m)!
Next, we will use a sneaky trick to compute the asymptotic expansion.
16 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

Lemma 3.1. For each nonnegative integer p, we have

(4ε)p (4ε)p
   
d 2
p+ d2 − δ8ε d
cp (ε) − d Γ p+ ≤2 e · d Γ p+ .
2(π) 2 2 2(π) 2 2

Proof. Using a change of variables, we have

2
Z δ p+ d2 Z δ4ε
s2 (4ε) d
e− 4ε s2p+d−1 ds = e−t tp+ 2 −1 dt.
0 2 0

Hence,
δ2
!
∞
(4ε)p (4ε)p
Z   Z
4ε
p+ d2 −1 d p+ d2 −1
cp (ε) = d e−t t dt = d Γ p+ − e−t t dt .
2(π) 2 0 2(π) 2 2 δ2
4ε

Therefore,
(4ε)p (4ε)p ∞ −t p+ d −1
  Z
d
cp (ε) − d Γ p+ = d e t 2 dt
2π 2 2 2π 2 δ4ε2
(4ε)p − δ2 ∞ −t/2 p+ d −1
Z
≤ d e
8ε e t 2 dt
2π 2 δ2
Z 4ε
(4ε)p − δ2 ∞ −t/2 p+ d −1
≤ d e
8ε e t 2 dt
2π 2 0
p
 
p+ d2 − δ8ε (4ε) d
2
=2 e · d Γ p+ .
2π 2 2
Simplifying the constants gives the desired bound. □

As a consequence, we obtain the estimate

   
d p+ d2 d δ2
ωd cp (ε) − p
(4ε) ≤ 2 (4ε)p e− 8ε ,
2 p 2 p

where (q)n = n−1

Q
j=0 (q + j) for n ≥ 1, and (q)0 = 1 denotes the Pochhammer symbol.

Therefore,
∞
" q d
q 2q+2k k q d

#
Z X 4 2 q
X X (−1) 4 2 q+k
kε (x, y)f (y) dVy = εq ηq + wq+k,m,k
Bx (δ) q=0
(2q)! k=1 m=4k
m!(2q + 2k − m)!

+ Rε (δ),

where the remainder term Rε (δ) is given by

ωd cq (ε) − d2 q (4ε)q X ωd cq+k (ε) − d2 q+k (4ε)q+k
∞
"
q 2q+2k

#
X X
Rε (δ) = ηq + (−4ε)−k wq+k,m,k .
q=0
(2q)! k=1 m=4k
m!(2q + 2k − m)!
 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R17

We estimate the remainder as follows:

d
ωd cq (ε) − d2 q (4ε)q d



2
q − δ8ε
23q+ 2 2 q
ηq ≤ ε e · |ηq |,
(2q)! (2q)!
d

d d


ωd cq+k (ε) − 2 q+k
(4ε)q+k 2
− δ8ε
23q+k+ 2 2 q+k
(−4ε)−k wq+k,m,k ≤ εq e · |wq+k,m,k | ,
m!(2q + 2k − m)! m!(2q + 2k − m)!
and hence
δ2
|Rε (δ)| ≤ e− 8ε ξ(ε),
P∞
where ξ(ε) = q=0εq ξq is a convergent power series with
d
q 2q+2k 3q+k+ d2 d
23q+ 2 d2 q



X X 2 2 q+k 2q + 2k

ξq = |ηq | + |wq+k,m,k | .
(2q)! k=1 m=4k
(2q + 2k)! m

This also implies that

Rε (δ)
lim+
=0
ε→0 εk
for all k ∈ N. We conclude that the asymptotic expansion of Kε f as ε → 0+ is given
by:

Theorem 3.2. For each q ≥ 0, define

4q d2 q X (−1)k 4q d2 q+k

q 2q+2k


X
aq (x) = ηq + wq+k,m,k .
(2q)! k=1 m=4k
m!(2q + 2k − m)!

Then the operator Kε admits the asymptotic expansion

∞
X
(Kε f )(x) ∼ aq (x) εq as ε → 0+ .
q=0

Proof. We sketch the idea of the proof. Fix x ∈ M , and let x : B0 (δ) ⊂ Rd → M be
a system of geodesic normal coordinates centered at x, where δ > 0 is smaller than
the injectivity radius at x. We decompose the integral defining Kε f (x) as
Z Z 
1 ∥x−y∥2
− 4ε
∥x−y∥2
− 4ε
(Kε f )(x) = d e f (y) dVy + e f (y) dVy .
(4πε) 2 Bx (δ) M \Bx (δ)

The second term is exponentially small: there exists a constant C > 0 such that
Z
∥x−y∥2
e− 4ε f (y) dVy = O(e−C/ε ) = o(εN +1 ),
M \Bx (δ)

for every N , since the integrand decays faster than any polynomial in ε.
18 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

We now focus on the integral over Bx (δ), transforming it into Euclidean coordinates
via the normal chart:
Z
1 ∥x(s)−x∥2
Iε (x) = d e− 4ε fe(s) ρ(s) ds,
(4πε) 2 B0 (δ)

where fe = f ◦ x, ρ(s) is the density of the volume form x∗ dV , and ds = ds1 · · · dsd .
Since ∥x(s) − x∥2 = s2 + q(s), where q(s) consists of quartic and higher-order terms
in s, we expand the exponential as
N k !
(−1)k q(s)

∥x(s)−x∥2 2
− s4ε
X
−
e 4ε =e 1+ + RN (s, ε) ,
k=1
k! 4ε
 
s4N +4
where RN (s, ε) = O εN +1
.
Likewise, we Taylor expand fe(s)ρ(s) at s = 0 up to order 2N :
X Hj (feρ)
fe(s)ρ(s) = (0)sj + O(s2N +1 ).
j≤2N
j!

Multiplying the expansions and collecting terms of total degree 2q, we integrate
term by term over s ∈ B0 (δ). Using spherical coordinates s = sv, with v ∈ S d−1 , and
the fact that
Z
s2
e− 4ε s2q+d−1 ds dσ(v) = cq (ε) ωd ,
B0 (δ)

we find that the coefficient of εq arises from the terms ηq involving s2q in the expansion
of fe(s)ρ(s), as well as interaction terms involving q(s), contributing correction terms
wq+k,m,k from higher-order interactions. Since cq (ε) ∼ εq up to a constant multiple,
we obtain, up to order N ,
q 2q+2k
N
!
X ωd cq (ε) X X ωd cq+k (ε)
Iε (x) = ηq + (−4ε)−k wq+k,m,k + o(εN +1 ).
q=0
(2q)! k=1 m=4k
m!(2q + 2k − m)!

By the previous lemma, this expression can be written as

N
X
aq (x) εq + o(εN +1 ).
q=0

Therefore, we conclude that

N
X
N +1
(Kε f )(x) = Iε (x) + o(ε )= aq (x)εq + o(εN +1 ),
q=0

as required. Since this holds for arbitrary N , the result follows. □

 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R19

We remark that the quantities ηq and wq+k,m,k are averages of homogeneous polyno-
mials over the unit sphere S d−1 , and hence can be computed using spherical moment
integrals. For example, see G. B. Folland, [10].

Theorem 3.3. For each multi-index α = (α1 , . . . , αd ), the average of the monomial
vα = v1α1 · · · vdαd over the unit sphere S d−1 ⊂ Rd is given by

0 Q if any αi is odd,


Z 
1 d
vα dσ(v) = αi +1


i=1 Γ 2
ωd S d−1  2 ·   if all αi are even.
|α|+d

Γ


2

We will compute a0 (x) and a1 (x) using the above formulas. First, let us compute
a0 (x) = η0 . Since α0 (v) = fe0 (v)ρ0 (v), and

fe0 (v) = fe(0), ρ0 (v) = 1,

we have
Z Z
1 1
η0 = α0 (v) dσ(v) = fe(0) dσ(v) = fe(0) = f (x).
ωd S d−1 ωd S d−1
We conclude that
a0 (x) = f (x).
Now, let us compute a1 (x). By the previous result,
d(d + 2)
a1 (x) = dη1 − w2,4,1 .
4!
It therefore suffices to compute η1 and w2,4,1 :
Z Z
1 1
η1 = α2 (v) dσ(v), w2,4,1 = α0 (v) b4,1 (v) dσ(v).
ωd S d−1 ωd S d−1
Since
α2 (v) = fe2 (v)ρ0 (v) + 2fe1 (v)ρ1 (v) + fe0 (v)ρ2 (v),
and fe0 (v) = f (x), by Theorem 2.2, we obtain
Z Z 
1
η1 = f2 (v) dσ(v) + f (x)
e ρ2 (v) dσ(v) .
ωd S d−1 S d−1
Since
d
X ∂ 2 fe 2
X ∂ 2 fe
f2 (v) =
e
2
(0) vi + 2 (0) vi vj ,
i=1
∂s i 1≤i<j≤d
∂s i ∂s j

d
1X 1 X
ρ2 (v) = − Rii (x) vi2 − Rij (x) vi vj ,
3 i=1 3 1≤i<j≤d
20 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

and using the identity Z

1 1
vi2 dσ(v) = ,
ωd S d−1 d
we obtain
Z
1 1
fe2 (v) dσ(v) = − (∆f )(x),
ωd S d−1 d
Z
1 1
ρ2 (v) dσ(v) = − R(x),
ωd S d−1 3d
where
d d
X ∂ 2 fe X
(∆f )(x) = − 2
(0), R(x) = Rii (x),
i=1
∂s i i=1
and R(x) denotes the scalar curvature of M at the point x. Therefore, we conclude
that  
1 1
η1 = −(∆f )(x) − R(x)f (x) .
d 3
Now, let us compute w2,4,1 . Since α0 (v) = f (x), we have
Z
f (x)
w2,4,1 = b4,1 (v) dσ(v).
ωd S d−1
Since Bm,1 (x1 , . . . , xm ) = xm , it follows that

b4,1 (v) = q4 (v) = −2∥Bx (vx , vx )∥2 .

Thus, Z
f (x)
w2,4,1 = −2 ∥Bx (vx , vx )∥2 dσ(v).
ωd S d−1
For v = (v1 , . . . , vd ) ∈ S d−1 , we write vx = di=1 vi (ei )x . Then
P

d
X
∥Bx (vx , vx )∥2 = bijkl (x) vi vj vk vl ,
i,j,k,l=1

where
bijkl (x) = ⟨Bx ((ei )x , (ej )x ), Bx ((ek )x , (el )x )⟩ .
Using the known spherical moment integrals:

1
Z  1

if i ̸= j,
2 2 d(d+1)
vi vj dσ(v) =
ωd S d−1  3

if i = j,
d(d+2)

we obtain
Z d d
1 X 1 X
bijkl (x) vi vj vk vl dσ(v) = (biijj (x) + bijij (x) + bijji (x)) .
ωd S d−1 i,j,k,l=1 d(d + 2) i,j=1
 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R21

Finally, by the Gauss equation, we conclude that

 2

Z d
1 2 1 X
∥Bx (vx , vx )∥ dσ(v) =  3 Bx ((ei )x , (ei )x ) − 2R(x) ,
ωd S d−1 d(d + 2) i=1

where R(x) denotes the scalar curvature at x. Since the mean curvature vector of M
at x is defined by
d
1X
H(x) = Bx ((ei )x , (ei )x ),
d i=1
it follows that
f (x) 2
d ∥H(x)∥2 − 2R(x) ,


a1 (x) = −(∆f )(x) +
4
where ∆ denotes the Laplace–Beltrami operator on M , Bx is the second fundamental
form, and R(x) is the scalar curvature at x. This expression agrees with the result
established in [11].
For a2 (x), the result involves the derivatives of various curvature quantities, in-
cluding the full Riemannian curvature tensor, Ricci curvature, and scalar curvature.
Additionally, it incorporates the derivatives of the second fundamental form, which
in turn affects the derivatives of the mean curvature. To derive the expression for
a2 (x), one can employ the Taylor expansion of the volume density as articulated in
Gray’s work, alongside the spherical moment integrals.

4. Hypersurfaces in Euclidean Spaces and their Equicurved Points

In this section, we employ the asymptotic expansion of the Gaussian integral op-
erator defined on a compact oriented hypersurface M ⊂ Rd+1 , which is equipped
with a global unit normal vector field ν. This analysis facilitates the classification
of points that meet specific curvature criteria, termed equicurved points. To lay the
groundwork for this exploration, we will begin by reviewing essential concepts related
to hypersurfaces in Euclidean space.
The Gauss map associated with the oriented hypersurface (M, ν) is defined as the
smooth map G : M → S d given by

G(x) = Tx,0 ν(x),

where Tx,0 : Tx Rd+1 → Rd+1 denotes the parallel translation (i.e., vector translation)
from the point x to the origin. Explicitly, the map Tx,0 sends a tangent vector
22 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

vx ∈ Tx Rd+1 to the corresponding vector v ∈ Rd+1 , via the identification T0 Rd+1 ∼

=
Rd+1 . Accordingly, for each x ∈ M , the unit normal vector ν(x) ∈ Tx Rd+1 satisfies
ν(x) = (G(x))x , where G(x) ∈ Rd+1 ∼ = T0 Rd+1 and (G(x))x ∈ Tx Rd+1 denotes the
corresponding tangent vector at the point x.
The shape operator Sx at a point x ∈ M is the linear map Sx : Tx M → Tx Rd+1
defined by sending a tangent vector vx ∈ Tx M to

Sx (vx ) = −(TG(x),x ◦ dGx )(vx ).

Here, dGx Tx M → TG(x) S d ⊂ Rd+1 is the differential of the Gauss map G, and TG(x),x
is the translation map that identifies the tangent space TG(x) S d with Tx M .
The image of Sx lies in Tx M , and Sx is a self-adjoint linear operator with respect
to the Riemannian metric induced on M . By the spectral theorem, Sx admits a
complete set of eigenpairs {(κi (x), (ei )x ) 1 ≤ i ≤ d}, where κi (x) is an eigenvalue of
Sx with corresponding eigenvector (ei )x . The eigenvalues are ordered as

κ1 (x) ≥ κ2 (x) ≥ · · · ≥ κd (x).

The vectors {(ei )x }di=1 form an orthonormal basis of Tx M .

We refer to κi (x) as the i-th principal curvature of M at x, and to (ei )x as the
corresponding i-th principal direction.
We define the bilinear form Bx : Tx M × Tx M → R by

Bx (vx , wx ) = ⟨Sx (vx ), wx ⟩.

Then we have
d
X
Bx (vx , wx ) = κi (x)⟨vx , (ei )x ⟩⟨wx , (ei )x ⟩
i=1
for all vx , wx ∈ Tx M . In this setting, the second fundamental form Bx at the point
x ∈ M is given by
Bx (vx , wx ) = Bx (vx , wx ) ν(x)

where ν(x) is the unit normal vector at x. In particular,

∥Bx (vx , wx )∥2 = |Bx (vx , wx )|2 .

Definition 4.1. Let (M, ν) be a smooth oriented hypersurface in Rd+1 , and let κ1 (x), . . . , κd (x)
denote the principal curvatures of M at a point x ∈ M . The i-th mean curvature of
 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R23

M at x is defined as
1

Hi (x) = d

ei κ1 (x), . . . , κd (x)
i

where ei denotes the i-th elementary symmetric polynomial in d variables. Specifically,

ei (κ1 , . . . , κd ) is the sum of all products of i distinct principal curvatures.
In particular, the (first) mean curvature of M at x, denoted by H(x), is defined as

H(x) = H1 (x).

Note that the mean curvature vector of M at x is given by

H(x) = H(x) ν(x),

where H(x) is the mean curvature and ν(x) is the unit normal vector at x.
Note that any hypersurface is locally orientable, meaning that there exists a local
unit normal vector field ν defined on a neighborhood of each point in M . Conse-
quently, one can define a local Gauss map by associating to each point the direction
of the corresponding unit normal vector. In particular, the principal curvatures of
M can be defined locally as the eigenvalues of the shape operator determined by this
local Gauss map.
Let (ei )x represent the i-th principal direction of the manifold M at the point
x ∈ M . The set {(ei )x 1 ≤ i ≤ d} constitutes an ordered orthonormal basis of the
tangent space Tx M . This basis induces a canonical system of normal coordinates
centered at the point x. For any 0 < δ < injx (M ), this basis provides a standard
choice for the normal coordinate parametrization

x : B0 (δ) → Bx (δ)
Pd
as previously defined. For any vx = i=1 vi (ei )x , the second fundamental form at x
satisfies
d
X
Bx (vx , vx ) = κi (x)vi2
i=1

where κi (x) denotes the i-th principal curvature of M at x. It follows that

d
!2
X
∥Bx (vx , vx )∥2 = κi (x)vi2 .
i=1
24 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

Utilizing the spherical moment integrals and the identity R(x) = 2e2 (κ1 (x), . . . , κd (x)),
where e2 denotes the second elementary symmetric polynomial of the principal cur-
vatures, the integral of b4,1 (v) over the unit sphere S d−1 ⊂ Rd is expressed as follows
d
!
−2
Z Z
1 X
2 4
X
2 2
b4,1 (v) dσ(v) = κi (x) vi + 2 κi (x)κj (x)vi vj dσ(v)
ωd S d−1 ωd S d−1 i=1 1≤i<j≤d
d
!
2 X X
=− 3 κi (x)2 + 2 κi (x)κj (x)
d(d + 2) i=1 1≤i<j≤d
2
3e1 (κ1 (x), . . . , κd (x))2 − 4e2 (κ1 (x), . . . , κd (x))


=−
d(d + 2)
2
3d2 H 2 (x) − 2R(x) .


=−
d(d + 2)
Here,H(x) represents the mean curvature of M at x, and R(x) is the scalar curvature
at x. Inserting this expression into the asymptotic expansion results in:
 
1 2 2

(Kε f )(x) ∼ f (x) + −∆f (x) + d H (x) − 2R(x) f (x) ε + · · ·
4
where ∆ represents the Laplace–Beltrami operator on M . This outcome aligns with
the expression derived in the previous section when we apply the identity

∥H(x)∥2 = H(x)2

directly.

Definition 4.2. Let x be a point on a hypersurface M ⊂ Rd+1 . We say that M

is equicurved at x (or that x is an equicurved point of M ) if the following identity
holds:
d2 H(x)2 = 2R(x),
where H(x) denotes the mean curvature and R(x) the scalar curvature at the point x.
Moreover, M is called an equicurved hypersurface if it is equicurved at every point
x ∈ M.

It follows from this definition that:

Theorem 4.3. Let M ⊂ Rd+1 be a compact hypersurface, and let x ∈ M . Then x is

an equicurved point if and only if the following limit holds:
f (x) − (Kε f )(x)
lim+ = ∆f (x)
ε→0 ε
 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R25

for every smooth function f : M → R, where ∆ denotes the Laplace–Beltrami operator

on M .

This theorem offers a practical and computable criterion for identifying equicurved
points on a hypersurface by analyzing the asymptotic behavior of the Gaussian in-
tegral operator. Specifically, at an equicurved point, the Gaussian integral opera-
tor asymptotically behaves similarly to the heat operator. These insights naturally
prompt several fundamental questions: Can a compact hypersurface in Euclidean
space possess an equicurved point? If so, is there a compact hypersurface that is
equicurved at every point? To explore these questions, we will begin by examining
some fundamental properties of equicurved points on hypersurfaces.
Note that

d2 H 2 − 2R = (e1 (κ1 , . . . , κd ))2 − 4e2 (κ1 , . . . , κd ),

where e1 and e2 denote the first and second elementary symmetric polynomials of the
principal curvatures, respectively. It follows that the identity d2 H 2 − 2R = 0 holds if
and only if

(4.1) (e1 (κ1 , . . . , κd ))2 = 4e2 (κ1 , . . . , κd ).

This identity allows us to derive a useful geometric consequence concerning minimal

or scalar-flat hypersurfaces:

Proposition 4.1. Let M ⊂ Rd+1 be a hypersurface, and let x ∈ M be an equicurved

point. If M is minimal at x, i.e., H(x) = 0, or if M is scalar flat at x, i.e. R(x) = 0,
then all principal curvatures vanish at x.

Proof. Suppose M is equicurved at x, so that the identity (4.1) holds.

If H(x) = 0, then e1 = 0, and hence (4.1) implies e2 = 0.
If R(x) = 0, then e2 = 0, and again by (4.1), we conclude that e1 = 0.
In either case, we compute:
d
X
κ2i = e21 − 2e2 = 0,
i=1

which implies κi = 0 for all 1 ≤ i ≤ d. Thus, all principal curvatures vanish at x. □

26 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

The condition (4.1) is also equivalent to the algebraic identity:

d
X X
(4.2) κ2i − 2 κi κj = 0.
i=1 1≤i<j≤d

When d = 2, the identity simplifies to (κ2 − κ1 )2 = 0, which implies that κ1 = κ2 .

Therefore, for d = 2, a point x in the manifold M is equicurved if and only if it
is umbilical. According to a classical result in the differential geometry of surfaces
(see, for instance, Do Carmo [6]), any compact umbilical surface in R3 must be a
round sphere. Consequently, any compact equicurved surface in R3 is necessarily a
sphere. We encourage the reader to verify that Theorem 4.3 holds at every point of
M = S 2 . This can be easily confirmed through direct computation using the explicit
formula for geodesics on the unit sphere S 2 . On the other hand, it is possible to
construct compact nonspherical surfaces that contain umbilical points, and therefore
equicurved points. A classical example of this is a spheroid in R3 . While a spheroid
is not a sphere, it has finitely many umbilical (and thus equicurved) points. We have
therefore answered the previous questions in the case where d = 2.
When d ≥ 3, an equicurved point of a hypersurface may not be umbilical. Suppose
x ∈ M is an equicurved point. If x is also umbilical, then all principal curvatures at
x would be equal, i.e.,

κ1 (x) = κ2 (x) = · · · = κd (x) = κ.

Substituting into (4.2), we obtain

 
2 d 2
(dκ) = 4 κ.
2
This simplifies to d(2 − d)κ2 = 0. Since d ≥ 3, it follows that k = 0. Thus, the point
x must be a flat point of the hypersurface M . We obtain the following result:

Proposition 4.2. Let d ≥ 3. Suppose M ⊂ Rd+1 is a hypersurface and x ∈ M is an

equicurved point. If x is also an umbilical point, then all principal curvatures vanish
at x; that is,

κ1 (x) = κ2 (x) = · · · = κd (x) = 0,

and hence M is flat at x.

 n
ASYMPTOTIC EXPANSION OF THE GAUSSIAN INTEGRAL OPERATORS ON RIEMANNIAN SUBMANIFOLDS OF R27

It is also relatively simple to construct hypersurfaces that contain nonflat equi-

curved points. For example, let M ⊂ R4 be the graph of the smooth function
1
f (x1 , x2 , x3 ) = (x21 + x22 + 4x23 ).
2
This defines a smooth hypersurface in R4 . One can verify that the origin x =
(0, 0, 0, 0) is a nonflat equicurved point of M , since the principal curvatures at the
origin are κ1 = 4, κ2 = 1, and κ3 = 1 which satisfy the equicurvature condition.
Currently, we are not aware of any instances of compact equicurved hypersurfaces
in Rd+1 for d ≥ 3. We do not delve into the construction of such examples in this
paper, as it falls outside the scope of this paper. Hence, we leave this as an open
question for future research.

Acknowledgements

The authors gratefully acknowledge the generous support provided by National

Cheng Kung University.

References
[1] Gray, Alfred:, The volume of a small geodesic ball of a Riemannian manifold. Michigan
Math. J. 20 (1973), 329–344 (1974).
[2] Belkin, M., Niyogi, P.: Laplacian eigenmaps for dimensionality reduction and data rep-
resentation. Neural Computation 15(6), 1373 – 1396 (2003).
[3] M. Belkin, P. Niyogi, Laplacian eigenmaps and spectral techniques for embedding and
clustering, in: T.G. Dietterich, S. Becker, Z. Ghahramani (Eds.), Adv. Neural Inform.
Process. Syst., vol. 14, MIT Press, Cambridge, MA, 2002.
[4] M. Belkin, P. Niyogi, Towards a theoretical foundation for Laplacian-based manifold
methods, in: P. Auer, R. Meir (Eds.), Proc. 18th Conf. Learning Theory (COLT), Lecture
Notes Comput. Sci., vol. 3559, Springer-Verlag, Berlin, 2005, pp. 486–500. MR2203282
[5] Mikhail Belkin, P. Niyogi: Convergence of Laplacian eigenmaps, NIPS’06: Proceedings
of the 20th International Conference on Neural Information Processing Systems Pages
129 - 136.
[6] do Carmo, Manfredo P.(BR-IMPA):Differential geometry of curves and surfaces. Revised
and updated second edition of [MR0394451] Dover Publications, Inc., Mineola, NY, 2016.
xvi+510 pp. ISBN:978-0-486-80699-0 ISBN:0-486-80699-5
[7] Coifman, Ronald R.(1-YALE); Lafon, Stéphane(1-YALE).: Diffusion maps.(English
summary) Appl. Comput. Harmon. Anal. 21 (2006), no. 1, 5–30.
[8] R. R. Coifman, S. Lafon, A. B. Lee, +3 , and S. W. Zucker.: Geometric diffusions as a
tool for harmonic analysis and structure definition of data: Diffusion maps, Proceedings
of the National Academy of Sciences of the United States of America , May 24, 2005,
Vol. 102, No. 21 (May 24, 2005), pp. 7432-7437
28 JIA-MING (FRANK) LIOU AND CHI-CHIEN LU

[9] Dunson, David B.(1-DUKE-S); Wu, Hau-Tieng(1-DUKE); Wu, Nan(1-DUKE) Spec-

tral convergence of graph Laplacian and heat kernel reconstruction in L∞ from random
samples.(English summary) Appl. Comput. Harmon. Anal. 55 (2021), 282–336.
[10] Folland, Gerald B.(1-WA): How to integrate a polynomial over a sphere. Amer. Math.
Monthly 108 (2001), no. 5, 446–448.
[11] Hein, Matthias(D-MPI-BC); Audibert, Jean-Yves(F-ENPC-TIS); von Luxburg, Ul-
rike: From graphs to manifolds—weak and strong pointwise consistency of graph Lapla-
cians.(English summary)Learning theory, 470–485. Lecture Notes in Comput. Sci., 3559
Lecture Notes in Artificial Intelligence
[12] Smolyanov, O. G.(RS-MOSCM); v. Weizsäcker, H.(D-KSRL); Wittich, O.: Brow-
nian motion on a manifold as limit of stepwise conditioned standard Brownian mo-
tions.Stochastic processes, physics and geometry: new interplays, II (Leipzig, 1999),
589–602. CMS Conf. Proc., 29 American Mathematical Society, Providence, RI, 2000
ISBN:0-8218-1960-7
[13] Smolyanov, Oleg G.(RS-MOSCM); Weizsäcker, Heinrich v.(D-TUKS); Wittich,
Olaf(D-TBNG-MI): Chernoff’s theorem and discrete time approximations of Brownian
motion on manifolds.(English summary) Potential Anal. 26 (2007), no. 1, 1–29.
[14] Smolyanov, O. G.(RS-MOSCM); v. Weizsäcker, H.(D-KSRL); Wittich, O. Chernoff’s
theorem and the construction of semigroups.Evolution equations: applications to physics,
industry, life sciences and economics (Levico Terme, 2000), 349–358. Progr. Nonlinear
Differential Equations Appl., 55 Birkhäuser Verlag, Basel, 2003 ISBN:3-7643-0374-3
[15] Singer, A.(1-YALE).: From graph to manifold Laplacian: the convergence rate.(English
summary) Appl. Comput. Harmon. Anal. 21 (2006), no. 1, 128–134.
Department of Mathematics, National Cheng Kung University, No.1, University Road, Tainan City
701, Taiwan, [email protected]

Department of Mathematics, National Cheng Kung University, No.1, University Road, Tainan City
701, Taiwan, [email protected]