Appendix A.1. Differential Geometry Remarks
In this appendix subsection, we will briefly review some well-known concepts that can be further explored in classical references on differential geometry, data analysis, and statistics, such as [
8,
9,
13], among many others.
Consider an
m-dimensional smooth real manifold
, and let
r and
s be non-negative integers.
A is a
-tensor field,
r times
contravariant and
s times
covariant, in an open set
U of a manifold
, and, roughly speaking, is a multilinear map defined at
depending
smoothly on
. Taking into account that a coordinate system
induces a basis vector field in each tangent space,
, denoted as
, and also a basis vector field in each dual,
, denoted as
, the components of the tensor field
A, corresponding to these basis fields, will be denoted by
, and when we change to another coordinate system
, these components are going to change
accordingly to
In (
A2), we have used the summation convention of repeated indices; all these quantities are evaluated at
, and we have used classical notation. If
, then
A is a smooth function on
, which is just an
invariant with respect to coordinate changes. If
, we just say that
A is an
s-
covariant tensor field, while if
, we just say that
A is an
r-
contravariant tensor field. We shall call
intrinsic any object or property independent of the coordinate system in
.
Let us recall that if and are smooth real manifolds, and q is a smooth mapping on an open subset to values in , then this map induces, for each , a linear map between and , which we will call the differential of q at and denote by , such that if , then we define , for all smooth function h on a neighborhood of .
Also, we recall that a curve σ in is a smooth map of an open interval in . When we talk about a curve of a closed interval in , we will assume that the domain of the curve is an open containing .
Given a m-dimensional real manifold , let denote the tangent bundle (the set of all the pairs such that with and having the structure of a -dimensional manifold), and let be the projection function, where if . Let I be an open interval in and be a smooth curve in . A vector field over is a map such that , i.e.: for all .
For each point in the domain, , we define the tangent vector to σ at t as the vector defined from , where is the basis vector field corresponding to the identity chart in . The tangent field can also be denoted by . For simplicity, we often overuse notation and identify the tangent fields on curves with their images, that is, we will write as .
A Riemannian manifold will be a manifold equipped with a Riemannian metric, that is, a second-order covariant smooth tensor field, which is additionally positive definite at each tangent space. In simple words, we have a scalar product defined in each tangent space that varies smoothly when we move to different points in the parametric space, which will be denoted by
. The local version of the Riemannian metric in
will commonly be expressed, using the summation convention of repeated indices, as
, as in (
8), where
.
Corresponding to this fundamental tensor, whose components are given by , there are the well-known operations of raising or lowering indices, which in fact allow us to identify covariant with contravariant tensor fields, or the opposite. For a particular index and for every , these operations are isomorphisms between the tangent bundle and its dual, the cotangent bundle, , the flat map (lowering indexes), symbolized by ♭ and defined through , or the opposite, between the dual of the tangent bundle and its dual, the bidual bundle of canonically identifiable with the tangent bundle itself, , the sharp map (raising indexes), symbolized by ♯ and defined through . In component notation, this is achieved simply by multiplying the components of a specific tensor by the fundamental covariant tensor field at or its inverse , the fundamental contravariant tensor field at the dual of the tangent bundle, the cotangent bundle .
The gradient of a smooth function q in an open neighborhood of a point of a Riemannian manifold is a vector field such that for all vector field , where is the directional derivative of q in the direction given by at . Observe that in the previous notation , it is well known that for , taking into account the definition of the gradient and the Cauchy–Schwarz inequality, we have , and the maximum is reached by the unitary vector . On the other hand, it will be possible to extend these mathematical objects to the complex case. If is a complex-valued function, and if and are the standard real and imaginary parts of a complex number, if we let and , we may define , and we can extend the scalar product in each tangent space to complexified vector fields and , defining the scalar product as , at each point . The corresponding square of the form will be .
The differentiation of vector fields involves the choice of a
connection, i.e., a rule which associates to each point
a tangent vector
and a smooth tensor field
defined at least in a neighborhood of
a vector
, such that
and
,
being smooth vector fields and
q being a smooth function in the neighborhood of
. It is also required that
be linear in
and that
be a smooth vector field. The vector
is commonly called the
covariant derivative of with respect to . At this point, let us recall that we can define the
divergence of a smooth vector field
as the smooth real-valued function defined, at
, as
, and the
Laplacian of a real smooth function
q as
; for details and properties see [
21]. If we add the requirements that
where
is the Lie bracket of both vector fields,
and
, which is another vector field such that their action, over a smooth function
q, is given by
and, additionally,
, we obtain the Levi–Civita connection, which is uniquely determined through the Christoffel symbols of the second kind. These are defined, again using repeated index summation convention, in terms of the metric tensor as
symbols which encode how the basis vectors change from point to point due to curvature, which is properly quantified through several objects, such as the curvature operator
for vector fields
,
, and
in
and, also, the Riemann–Christoffel tensor defined as
with
being another vector field in
, the
Riemannian sectional curvatures corresponding to the linearly independent vector fields
and
being defined as
and the
Ricci tensor being defined as
. Observe that, if
and
, if we define
, we find that the components of the vector field
are given by
where
is the ordinary partial derivative. This formula illustrates how the covariant derivative modifies the ordinary derivative by incorporating terms that account for the twisting and turning of the coordinate system in curved space. The covariant derivative concept is a powerful generalization of ordinary derivatives, extending the concept from flat, Euclidean spaces to the more complex realm of curved manifolds. This extension is crucial for understanding how geometric properties evolve in curved spaces. The covariant derivative incorporates additional terms that account for the curvature of the manifold, effectively correcting for the way basis vectors change as one moves from point to point. This enables us to analyze and describe the behavior of vector fields and other geometric objects in a way that respects the underlying curvature.
It is convenient to generalize the covariant derivative to any tensor field. For this purpose, given
and
q, a real smooth function on an open subset of
we denote by
, and if
and
, then define
. Moreover, given an
-tensor field
, the covariant derivative may be extended to obtain an
-tensor field. If
and
, then
and, finally,
Given a connection, the corresponding
geodesics, the generalization of straight lines, are the curves whose tangent vector field,
T, does not change, i.e.,
. In components, under a coordinate system
, if we denote, with certain overuse of notation, by
the coordinates of a geodesic, we shall have, using the summation convention of repeated indices,
In the Riemannian case, with the Levi–Civita connection, the geodesics are also, locally, the minimum-length curves.
Next, we review the definition of the
exponential map corresponding to a connection that is defined through geodesics as follows. Let
be a point of the manifold
,
,
be the tangent space in
, and let
be a geodesic such that
Then, the exponential map is given by , and it is well defined for all in an open star-shaped neighborhood of .
Hereafter, we will consider the Riemannian case with the Levi–Civita connection. We now define
as
, and for each
, we let
, where
is the Riemannian distance, and
is a geodesic defined in an open interval containing zero, such that
with a tangent vector equal to
at the origin. Then, if we set
it is well known that
maps
diffeomorphically onto
. If the manifold is also complete, then the boundary of
,
is mapped by the exponential map onto
, called the
cut locus of θ in Θ, which in this case has a zero Riemannian measure. Moreover, if the manifold is simply connected, and the Riemannian curvature is non-positive or positive but with a sufficiently small diameter, the cut locus is empty. Additionally, in this case, the inverse of the exponential map is considered a map between two metric spaces:
, a Riemannian manifold, and
, the tangent space with Euclidean structure, which preserves the distance between any point to
, although it does not preserve distances between arbitrary points in general. For additional details, see [
9,
21,
22].
Figure A1.
The figure illustrates the exponential map corresponding to the Levi–Civita connection in a Riemannian manifold. The lengths of the vectors in in the tangent space are equal to the Riemannian distance . This map preserves the radial distances but not the distances in general; the norm is not equal to the Riemannian distance ).
Figure A1.
The figure illustrates the exponential map corresponding to the Levi–Civita connection in a Riemannian manifold. The lengths of the vectors in in the tangent space are equal to the Riemannian distance . This map preserves the radial distances but not the distances in general; the norm is not equal to the Riemannian distance ).
We now briefly review the concept of
Jacobi field along a geodesic. Consider a geodesic
in
and
, the tangent vector field over
. A Jacobi field
along
is a field which satisfies the Jacobi equation
where
is the covariant derivative in the direction given by
along
c.
Appendix A.3. Additional Remarks
As a reminder, it has been proven that well-known indices, such as the Kullback–Leibler divergence, locally induce, up to a proportionality constant, the information metric on the parameter space of the statistical model. Further characterizations of the information metric, in terms of invariance under Markov kernels’ transformations, are given in [
23]. With the Riemannian distance, say
, such that
, the manifold
or
becomes a
length space, unlike these manifolds considering the Hilbert metric structure. The information metric has been studied for several parametric regular statistical models; for instance, see [
11,
24]. Moreover, though the information metric is possible to develop, in a natural way, an
intrinsic approach to statistical estimation, invariant under reparametrizations, see [
25].
Concerning the Hermitian kernel notion, observe that the Hermitian property is a consequence of the positive-definiteness, and it is not necessary to make it explicit in the definition. If, for mutually distinct , the equality only holds for , it is often said that K is a strictly positive definite Hermitian kernel. It is also possible to consider real-valued kernels satisfying such that and for any positive integer , , .
Moreover, observe that the natural kernel suggested, corresponding to a regular parametric family, satisfies , since clearly , and for any scalars , where n is an arbitrary positive integer, ; therefore, is a complex-valued Hermitian kernel.
With additional assumptions about the parameter space from measure and topological theory, many other properties of kernels can be established, such as Mercer’s theorem and relationships between square-integrable function spaces and RKHS via linear operators. Additionally, to clarify the relationship between the fundamental tensor and the kernel, we can utilize several results made explicit in Chapter 4 of [
13].