1. Introduction
Low-rank matrices appear in many applications involving high-dimensional data. Low-rank models are commonly used in statistics, machine learning or data analysis (see [
1] for a recent survey). Furthermore, low-rank approximation of matrices is the cornerstone of many modern numerical methods for high-dimensional problems in computational science, such as model-order-reduction methods for dynamical systems or parameter-dependent or stochastic equations [
2,
3,
4,
5].
These applications yield problems of approximation or optimization in the sets of matrices with fixed rank:
Fixed-rank matrices appear also in the theory of characteristics of Partial Differential Equations and Monge-Ampère equations [
6]. More precisely, it has been proven [
6,
7] that Monge-Ampère equations with
n independent variables and of Goursat-type are in one-to-one correspondence with the set
Thus, the parabolic or hyperbolic nature of the Monge-Ampère equation is related to the rank of such matrices.
In [
8,
9], the authors point out that Algebraic Geometry appears as a natural tool in study of the set
We wish to mention the papers [
10,
11,
12] that raise the natural question of how large these matrix spaces are.
A usual geometric approach is to endow the set
with the structure of a Riemannian manifold [
13,
14], which is seen as an embedded submanifold of
equipped with the topology
given by matrix norms. Standard algorithms then work in the ambient matrix space
and do not rely on an explicit geometric description of the manifold using local charts (see, e.g., [
15,
16,
17,
18]). However, the matrix rank considered as a map is not continuous for the topology
, which can yield undesirable numerical issues.
The purpose of this paper is to propose a new geometric description of the sets of matrices with fixed rank, which is amenable for numerical use, and relies on the natural parametrization of matrices in
given by
where
and
are matrices with full rank
and
is a non singular matrix. The set
is here endowed with the structure of analytic principal bundle with an explicit description of local charts. This results in a description of the matrix space
as an analytic manifold with a topology induced by local charts that is different from
and for which the rank is a continuous map. Note that the representation (
1) of a matrix
Z is not unique because
holds for every invertible matrix
P in
. An argument used to dodge this undesirable property is the possibility to uniquely define a tangent space (see for example Section 2.1 in [
18]), which is a prerequisite for standard algorithms on differentiable manifolds. The geometric description proposed in this paper avoids this undesirable property. Indeed, the system of local charts for the set
is indexed on the set itself. This allows a natural definition of a neighbourhood for a matrix where all matrices admit a unique representation.
The present work opens the route for new numerical methods for optimization and dynamical low-rank approximation with algorithms working in local coordinates and avoiding the use of a Riemannian structure. In [
19], such a framework is introduced for generalising iterative methods in optimization from Euclidean space to manifolds, which ensures that local convergence rates are preserved. Recently, a splitting algorithm relying on the geometric description of the set of fixed rank matrices proposed in this paper has been introduced for dynamical low-rank approximation [
20].
The introduction of a principal bundle representation of matrix manifolds is also motivated by the importance of this geometric structure in the concept of gauge potential in physics [
21].
Note that the proposed geometric description has a natural extension to the case of fixed-rank operators on infinite dimensional spaces and is consistent with the geometric description of manifolds of tensors with fixed rank proposed by Falcó, Hackbush and Nouy [
22] in a tensor Banach space framework.
Before introducing the main results and outline of the paper, we recall some elements of geometry.
1.1. Elements of Geometry
In this paper, we follow the approach of Serge Lang [
23] for the definition of a manifold
. In this framework, a set
is equipped with an atlas which gives
the structure of a topological space, with a topology induced by local charts, and the structure of differentiable manifold compatible with this topology. More precisely, the starting point is the definition of a collection of non-empty subsets
, with
in a set
A, such that
is a covering of
. The next step is the explicit construction for any
of a local chart
which is a bijection from
to an open set
of the finite dimensional space
such that for any pair
such that
, the following properties hold:
- (i)
and are open sets in and respectively, and
- (ii)
the map
is a
differentiable diffeomorphism, with
or
when the map is analytic.
Under the above assumptions, the set
is an atlas which endows
with a structure of
manifold. Then, we can say that
is a
manifold, or an analytic manifold when
. A consequence of condition
is that when
holds for
, then
In the particular case where
for all
, we say that
is a
manifold modelled on
Otherwise, we say that it is a manifold not modelled on a particular finite-dimensional space. A paradigmatic example is the Grassmann manifold
of all linear subspaces of
, such that
where
and
are trivial manifolds and
is a manifold modelled on the linear space
for
Consequently,
is a manifold not modelled on a particular finite-dimensional space.
The atlas also endows
with a topology given by
which makes
a topological space where each local chart
considered as a map between topological spaces is a homeomorphism. (Here
denotes a topological space, and if
, then
denotes the subspace topology.)
1.2. Main Results and Outline
Our first remark is that the matrix space
is an analytic manifold modelled on itself, and its geometric structure is fully compatible with the topology
induced by a matrix norm. In this paper, we define an atlas on
, which gives this set the structure of an analytic manifold, with a topology induced by the atlas fully compatible with the subspace topology
. This implies that
is an embedded submanifold of the matrix manifold
modelled on itself. (Note that the set
is a trivial manifold, which is trivially embedded in
.) For the topology
, the matrix rank considered as a map is not continuous but only lower semi-continuous. However, if
is seen as the disjoint union of sets of matrices with fixed rank,
then
has the structure of an analytic manifold not modelled on a particular finite-dimensional space equipped with a topology
which is not equivalent to
, and for which the matrix rank is a continuous map.
Note that in the case where , the set coincides with the general linear group of invertible matrices in which is an analytic manifold trivially embedded in . In all other cases are addressed in this paper, our geometric description of relies on a geometric description of the Grassmann manifold , with or m.
Therefore, we start in
Section 2 by introducing a geometric description of
. A classical approach consists of describing
as the quotient manifold
of equivalent classes of full-rank matrices
Z in
with the same column space
. Here, we avoid the use of equivalent classes and provide an explicit description of an atlas
for
, with local chart
where
is such that
(see Remark 1 for a practical choice) and
denotes the column space of a matrix
, and we prove that the neighbourhood
has the structure of a Lie group. This parametrization of the Grassmann manifold is introduced in ([
24] Section 2), but the authors do not elaborate on it.
Then, in
Section 3, we consider the particular case of full-rank matrices. We introduce an atlas
for the manifold
of matrices with full rank
, with local chart
and prove that
is an analytic principal bundle with base
and typical fibre
. Moreover, we prove that
is an embedded submanifold of
and that each of the neighbourhoods
have the structure of a Lie group.
Finally, in
Section 4, we provide an analytic atlas
for the set
of matrices
with rank
, with local chart
and we prove that
is an analytic principal bundle with base
and typical fibre
. Then, we prove that
is an embedded submanifold of
and that each of the neighbourhoods
have the structure of a Lie group.
2. The Grassmann Manifold
In this section, we present a geometric description of the Grassmann manifold
of all subspaces of dimension
r in
,
,
with an explicit description of local charts. We first introduce the surjective map
where
is the column space of the matrix
Z, which is the subspace spanned by the column vectors of
Given
there are infinitely many matrices
Z such that
. Given a matrix
, the set of matrices in
with the same column space as
Z is
2.1. An Atlas for
For a given matrix
Z in
, we let
be a matrix such that
, and we introduce an
affine cross section
which has the following equivalent characterization.
Lemma 1. The affine cross section is characterized byand the mapis bijective. Proof. We first observe that for all , which implies that For the other inclusion, we observe that if , then and hence , the orthogonal subspace to in . Since there exists such that Proving that is bijective is straightforward. □
Proposition 1. For each such that , there exists a unique such thatholds, which means that the set of matrices with the same column space as W intersects at the single point Furthermore, if and only if Proof. By Lemma 1, a matrix is such that for a certain and a certain . Then and is uniquely defined by which proves that is the singleton and if and only if □
Corollary 1. For each , the map is injective.
Proof. Let us assume the existence of such that Then by Proposition 1. □
Lemma 1 and Corollary 1 allow us to construct a system of local charts for
by defining for each
a neighbourhood of
by
together with the bijective map
such that
for
. We denote by
the Moore–Penrose pseudo-inverse of the full rank matrix
, defined by
It satisfies and . Moreover, is the projection onto parallel to . Finally, we have the following result.
Theorem 1. The collection is an analytic atlas for and hence is an analytic -dimensional manifold modelled on .
Proof. Clearly is a covering of Now let Z and be such that . Let such that , with . We can write with and . Therefore, , which implies that . Therefore, is an open set. In the same way, we show that and is an open set. Finally, the map from to is given by , with , which is clearly an analytic map. □
Remark 1. A possible choice for satisfying is where is such that its column space is a complement of the column space of Z. In practice, we can determine a set of r linear independent rows of Z (see, e.g., [25,26]), with indices I, and then choose such that if and 0 if , for , . For a given , the computation of does not require and has a complexity . 2.2. Lie Group Structure of Neighbourhoods
Here we prove that each neighbourhood
of
is a Lie group. For that, we first note that a neighbourhood
of
can be identified with the set
through the application
. The next step is to identify
with a closed Lie subgroup of
denoted by
with associated Lie algebra
isomorphic to
, and such that the exponential map
is a diffeomorphism. (We recall that the matrix exponential
is defined by
) To this end, for a given
, we introduce the vector space
The following proposition proves that is a commutative subalgebra of
Proposition 2. For all ,holds, and is a commutative subalgebra of Moreover,andhold for all Proof. Since
holds for all
the vector space
is a closed subalgebra of the matrix unitary algebra
As a consequence,
holds for all
and all
, which proves (
6). We directly deduce (
7) using
and (
8) using
. □
From Proposition 2 and the definition of , we obtain the following results.
Corollary 2. The affine cross section satisfiesandfor all , where the brackets are used for matrix concatenation. Proof. From Proposition 2 and (
4), we obtain (
9) and we can write
Since
, (
10) follows. □
Now we need to introduce the following definition and proposition (see ([
27] p. 80)).
Definition 1. Let be a ring and let be its additive group. A subset is called a two-sided ideal (or simply an ideal) of if it is an additive subgroup of such that and
Proposition 3. If is a two-sided ideal of the Lie algebra of a group , then the subgroup generated by is normal and closed, with Lie algebra
From the above proposition, we deduce the following result.
Lemma 2. Let and be such that Then is a two-sided ideal of the Lie algebra and henceis a closed Lie group with Lie algebra Furthermore, the map is bijective. Proof. Consider
and
. Noting that
and
, we have that
which proves that
Similarly, we have that
which proves that
This proves that
is a two-sided ideal. The map exp is clearly surjective. To prove that it is injective, we assume
for
. Then, from (
6), we obtain
and hence
, i.e.,
in
□
Finally, we can prove the following result.
Theorem 2. The set together with the group operation defined byfor is a Lie group. Proof. To prove that it is a Lie group, we simply note that the multiplication and inversion maps
and
are analytic. □
It follows that can be identified with a Lie group through the map .
Theorem 3. Each neighbourhood of together with the group operation defined byfor , is a Lie group, and the map given byis a Lie group isomorphism. 3. The Non-Compact Stiefel Principal Bundle
In this section, we give a new geometric description of the set
of matrices with full rank
, which is based on the geometric description of the Grassmann manifold given in
Section 2.
3.1. Principal Bundle Structure of
For
, we define a neighbourhood of
Z as
From Proposition 1, we know that for a given matrix
, there exists a unique pair of matrices
such that
Therefore,
It allows us to introduce a parametrisation
(see
Figure 1) defined through the bijection
such that
for
and
for
. In particular,
Theorem 4. The collection is an analytic atlas for , and hence is an analytic -dimensional manifold modelled on
Proof. is clearly a covering of . Moreover, since is bijective from to we claim that if for then the following statements hold:
- (i)
and are open sets in and
- (ii)
the map is analytic from to .
In this proof, we equip
with the topology
induced by matrix norms. For any
,
is the inverse image of the open set
by the continuous map
from
to
, and therefore,
is an open set of
. Since
and
are open sets in
,
is also an open set in
and since
is a continuous map from
to
, the set
, as the inverse image of an open set by a continuous map, is an open set in
. Similarly,
is an open set. Now let
such that
. From the expressions of
and
, the map
is defined by
with
, which is clearly an analytic map. □
Before stating the next result, we recall the definition of a morphism between manifolds and of a fibre bundle. We introduce notions of maps and manifolds, with or . In the latter case, means analytic.
Definition 2. Let and be two manifolds. Let be a map. We say that F is a morphism between and if given , there exists a chart such that and a chart such that where and the mapis a map of class . If it is a diffeomorphism, then we say that F is a diffeomorphism between manifolds.
We say that is a representation of F using a system of local coordinates given by the charts and
Definition 3. Let be a manifold with atlas , and let be a manifold. A fibre bundle with base and typical fibre is a manifold which is locally a product manifold; that is, there exists a surjective morphism such that for each there is a diffeomorphism between manifoldssuch thatwhereis the projection. For eachis called the fibre overThediffeomorphismsare called fibre bundle charts. Ifandare only required to be topological spaces andan open covering ofIn the case whereis a Lie group, we say thatis a principal bundle,
and if is a vector space, we say that it is a vector bundle.
Theorem 5. The set is an analytic principal bundle with typical fibre and base , with a surjective morphism between and given by the map .
Proof. To show that it is an analytic principal bundle, we first observe that
is a surjective morphism. Indeed, let
and
and
. Noting that
for all
, we obtain that
. Moreover, a representation of
by using a system of local coordinates given by the charts is
which is clearly an analytic map from
to
such that
Now, a representation of the morphism
using the system of local coordinates given by the charts is
defined by
which is clearly an analytic diffeomorphism. To conclude, consider the projection
and observe that
holds for all
□
3.2. as a Submanifold and Its Tangent Space
Here, we prove that the non-compact Stiefel manifold
equipped with the topology given by the atlas
is an embedded submanifold in
. For that, we have to prove that the standard inclusion map
as a morphism is an embedding. To see this, we need to recall some definitions and results.
Definition 4. Let be a morphism between manifolds and let We say that F is an immersion at mif there exists an open neighbourhood of m in such that the restriction of F to induces an isomorphism from onto a submanifold of We say that F is an immersion if it is an immersion at each point of
The next step is to recall the definition of the differential as a morphism which gives a linear map between the tangent spaces of the manifolds (in local coordinates) involved with the morphism. Let us recall that for any , we denote by the tangent space of at m (in local coordinates).
Definition 5. Let and be two manifolds. Let be a morphism of class ; i.e., for any ,is a map of class , where is a chart in containing m and is a chart in containing . Then we define For finite dimensional manifolds we have the following criterion for immersions (see Theorem 3.5.7 in [
28]).
Proposition 4. Let and be manifolds. Letbe a morphism and Then F is an immersion at m if and only if is injective. A concept related to an immersion between manifolds is given in the following definition.
Definition 6. Let and be manifolds and let be a morphism. If f is an injective immersion, then is called an immersed submanifold of .
Finally, we give the definition of embedding.
Definition 7. Let and be manifolds and let be a morphism. If f is an injective immersion, and is a topological homeomorphism, then we say that f is an embedding and is called an embedded submanifold of .
We first note that the representation of the inclusion map
i using the system of local coordinates given by the charts
in
and
in
is
Then the tangent map
at
, defined by
, is
Proposition 5. The tangent map at is a linear isomorphism, with inverse given byfor . Furthermore, the standard inclusion map i is an embedding from to Proof. Let us assume that Multiplying this equality by and on the left, we obtain and , respectively, which implies that is injective. To prove that it is also surjective, we consider a matrix and observe that and is such that . Since is injective, the inclusion map i is an immersion.
To prove that it is an embedding, we equip
with the topology
given by the atlas and we equip
with the topology
induced by matrix norms. We need to check that
is a topological homeomorphism. Since the topology in
has the property that each local chart
is indeed a homeomorphism from
in
to
(see
Section 1.1), we only need to show that the bijection
given by
is a topological homeomorphism for all
Observe that
is given by
Assume that Multiplying this equality by on the left we obtain and hence Multiplying by on the left, we obtain Thus, and as a consequence is a linear isomorphism for each The inverse function theorem says us that is a diffeomorphism, in particular a homeomorphism,, and hence i is an embedding. □
The tangent space to
at
Z is the image through
of the tangent space at
Z in local coordinates
, i.e.,
and can be decomposed into a vertical tangent space
and a horizontal tangent space
3.3. Lie Group Structure of Neighbourhoods
We here prove that each neighbourhood
of
has the structure of a Lie group. For that, we first note that
can be identified with
, with
given by (
9). Noting that
can be identified with the Lie group
defined in (
11), we then have that
can be identified with a product of two Lie groups
, which is a Lie group with the group operation
given by
for
and
. This allows us to define a group operation
over
defined for
and
by
and to state the following result.
Theorem 6. The set together with the group operation defined by (15) is a Lie group and the map given byis a Lie group isomorphism.