See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/374115653

Learning the model from the data

Article in Revista de la Unión Matemática Argentina · September 2023

DOI: 10.33044/revuma.4371

CITATIONS READS

0 40

2 authors:

Carlos Cabrelli Ursula Molter

University of Buenos Aires and CONICET (ARGENTINA) National Scientific and Technical Research Council and Universidad de Buenos Aires
122 PUBLICATIONS 1,895 CITATIONS 112 PUBLICATIONS 1,542 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Carlos Cabrelli on 30 October 2023.

The user has requested enhancement of the downloaded file.

REVISTA DE LA UNIÓN MATEMÁTICA ARGENTINA
Vol. 66, No. 1, 2023, Pages 141–152
Published online: September 21, 2023
https://doi.org/10.33044/revuma.4371

LEARNING THE MODEL FROM THE DATA

CARLOS CABRELLI AND URSULA MOLTER

Abstract. The task of approximating data with a concise model comprising

only a few parameters is a key concern in many applications, particularly in
signal processing. These models, typically subspaces belonging to a specific
class, are carefully chosen based on the data at hand. In this survey, we
review the latest research on data approximation using models with few pa-
rameters, with a specific emphasis on scenarios where the data is situated in
finite-dimensional vector spaces, functional spaces such as L2 (Rd ), and other
general situations. We highlight the invariant properties of these subspace-
based models that make them suitable for diverse applications, particularly in
the field of image processing.

Cuando entras en el corazón de un amigo,

no importa el lugar que ocupes,
lo importante es que nunca salgas de ahı́.
Anónimo

1. Introduction
In this note, we will provide an overview of recent developments in the field
of optimal subspaces, which has gained recently significant attention due to its
application in signal and image models. We refer the reader to the references for
more details and proofs.
The proliferation of available data has transformed the process of extracting
meaningful information from it. As each type of data possesses specific characteris-
tics, the design of tailored algorithms can take advantage of these shared attributes,
leading to improved efficiency.
Therefore, it is crucial to construct a model for each type of data that relies
on the fewest possible parameters while capturing their common features. One
potential approach to achieving this is by assuming certain hypotheses about the

2020 Mathematics Subject Classification. 94A20, 42C15, 46N99.

Key words and phrases. Sampling theory, theorem of Eckart–Young, shift invariant spaces,
crystal groups, rotation invariant spaces.
The research of C. Cabrelli and U. Molter is partially supported by Grants PICT 2011-
0436 (ANPCyT), PIP 2008-398 (CONICET) and UBACyT 20020100100502 and 20020100100638
(UBA).

141
142 CARLOS CABRELLI AND URSULA MOLTER

device or phenomenon that generated the data, such as assuming that the signals
under consideration are band-limited.
However, given the vast diversity of data available today, this approach may
not be suitable in many cases, particularly when considering for example, data as
internet traffic or stock market values. Instead, our strategy is to generate the
model from the data itself, using a set of subspaces as models, from which we can
choose the best fit for our data. The subspaces that we select and the data are all
from the same vector space.
In signal and image processing, there are often certain transformations that are
known to leave important features of the data, invariant. For example, in image
processing, translations, rotations, and scaling are common transformations that
preserve the spatial structure of an image.
To build effective models for such data, it is important to incorporate these
known invariances into the model. This can be done by explicitly including trans-
formation parameters and optimizing them along with the other parameters.
Incorporating invariances into the model can lead to more robust and accu-
rate performance on real-world data, as the model is better equipped to handle
variations and changes in the input data.
We will take into account subspaces that are invariant under both translations
and rotations. To simplify the model, we will only consider discrete sets of trans-
lations and rotations.
We want the subspaces in the class to be “small” in a sense that will be specified
in each case. This condition will be essential for the applications.
So the general scheme will be the following: Let H be a Hilbert space and M a
family of subspaces of H. Consider a finite set of data F = {f1 , . . . , fm } and define

m
X
E(F, S) = kfj − PS fj k2 , (1)
j=1

where S ∈ M and PS denote the orthogonal projection into the subspace S. The
functional E will be our gauge that will measure the fitness of the data to the
subspace. We analyze the existence and construction of an optimal subspace in the
class M that minimizes the functional E(F, S) over M.
Section 2 will focus on the case of a finite dimensional Hilbert space H and a
class M, which consists of all the subspaces of H with dimensions smaller than
a fixed positive integer `. Next, in Section 3, we will examine the prototypical
scenario of subspaces that are invariant under integer translations (SIS). We will
consider optimality for the subclass of SIS that exhibits additional invariance in
Section 4. Lastly, in Section 5, we will present the outcomes for models that are
invariant under translation and rotation.

2. Optimality for the class of finite dimensional subspaces.

When the approximation class M is the class of the finite dimensional subspaces,
the problem can be solved using Singular Value Decomposition techniques. The

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

 LEARNING THE MODEL FROM THE DATA 143

next theorem is an adaptation of the Eckart–Young theorem ([12, 18]) and will be
used throughout the paper.
Given a set of vectors F = {f1 , . . . , fm } of a Hilbert space H define the Gramian
matrix of F by [GF ]i,j = hfi , fj iH , X = span {f1 , . . . , fm }, and let r = dim X =
rank GF .
With this notation we have:
Theorem 2.1 ([1, Theorem 4.1]). Let F = {f1 , . . . , fm } ⊆ H, where H is a Hilbert
space, and let n ≤ r be a positive integer. Let λ1 ≥ · · · ≥ λm ∈ R be the eigenvalues
of the matrix GF and y1 , . . . , ym ∈ Cm , with yi = (yi1 , . . . , yim )t the associated left
orthonormal eigenvectors. Define the vectors q1 , . . . , qn ∈ H by
m
X
qi = θi yij fj , i = 1, . . . , `,
j=1

−1/2
where θi = λi if λi 6= 0 and θi = 0 otherwise. Then {q1 , . . . , q` } is a Parseval
frame of W ∗ = span {q1 , . . . , q` } and the subspace W ∗ is optimal in the sense that,
if W is any subspace with dim(W ) ≤ `, we have
m
X m
X
E(F, W ∗ ) = kfi − PW ∗ fi k2 ≤ E(F, W) = kfi − PW fi k2 .
i=1 i=1

Furthermore we have the following formula for the error:

m
X
E(F, W ∗ ) = λi .
i=`+1

3. Optimality for the class of SIS in L2 (Rd )

In [1] the authors give a solution for the case where the approximation class is
the class of shift-invariant spaces (SIS) of L2 (Rd ). A closed subspace V ∈ L2 (Rd )
is shift-invariant if it is invariant under the translations along Zd . A shift invariant
space V always has a set of generators, i.e. a set Φ ⊆ L2 (Rd ) finite or countable
such that
V = S(Φ) = span{Tk φ : φ ∈ Φ, k ∈ Zd }.
Here Tk denotes the translation along k, i.e. (Tk f )(x) = f (x − k), k ∈ Zd . The
length of a SIS is the cardinal of the minimun set of generators.
Shift-invariant spaces can be seen, using a theorem of Helson [15] (and a unitary
transformation from L2 (Rd ) onto L2 ([0, 1]d , `2 (Zd )), see [9]), as a continuous of
subspaces of `2 (Zd ). When the SIS is finitely generated, these subspaces are finite-
dimensional, and Theorem 2.1 can be used to obtain in each of them a solution.
The generators of the optimal SIS are then constructed by measurably gluing the
solution in each component. (see [1] for details and a proof).
For a set of functions F = {f1 , . . . , fm } in L2 (Rd ). we define the Gramian as
X
[GF ]i,j (ω) = fˆi (ω + k)fˆj (ω + k), ω ∈ U.
k∈Zd

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

144 CARLOS CABRELLI AND URSULA MOLTER

Here fˆ denotes the Fourier transform of f and U = [0, 1]d .

Theorem 3.1 ([1, Theorem 2.3]). Let F = {f1 , . . . , fm } be a set of functions in
L2 (Rd ). Let λ1 (ω) ≥ · · · ≥ λm (ω) be the eigenvalues of the Gramian GF (ω). Then,
there exists V ∗ ∈ V ` = {V : V is a SIS of length at most `} such that
m
X m
X
kfi − PV ∗ fi k2 ≤ kfi − PV fi k2 , ∀ V ∈ V `.
i=1 i=1

Moreover, we have that

(1) The eigenvalues λi (ω), 1 ≤ i ≤ m are Zd -periodic, measurable functions in
L2 (U) and
m Z
X
E(F, `) = λi (ω) dω.
i=`+1 U

(2) Let θi (ω) = λ−1

i (ω) if λi (ω) is different from zero, and zero otherwise.
Then, there exists a choice of measurable left eigenvectors Y 1 (ω), . . . , Y ` (ω)
with Y i = (y1i , . . . , ym
i t
) , i = 1, . . . , `, associated with the first ` largest
eigenvalues of GF (ω) such that the functions defined by
m
X
ϕ
bi (ω) = θi (ω) yji (ω)fbj (ω), i = 1, . . . , `, ω ∈ Rd
j=1

are in L2 (Rd ). Furthermore, the corresponding set of functions Φ = {ϕ1 ,

. . . , ϕ` } is a generator set for the optimal subspace V ∗ and the set {ϕi (·−k),
k ∈ Zd , i = 1, . . . , `} is a Parseval frame for V ∗ .

4. Optimality for the class of SIS with extra-invariance

4.1. Sets of invariance and extra invariance. In this section we will use for
our approximation a subclass of the class V ` defined in the previous section. We
will consider the class of the extra-invariant subspaces of length `. We need first
some definitions.
Definition 4.1. Let V ⊆ L2 (Rd ) be a SIS. We define the invariance set as follows:
M := {x ∈ Rd : Tx f ∈ V, ∀f ∈ V }.
In [2] (see also [3]), the authors proved that the invariance set of a shift invariance
space V ⊆ L2 (Rd ) is a closed additive subgroup of Rd that contains Zd . For
instance, in the case of the line the invariant set of a shift invariant space could be
Z, n1 Z for some n ∈ N or R.
Definition 4.2. Let Φ ⊆ L2 (Rd ). We will say that a SIS V is M extra-invariant
if Tm f ∈ V for all m ∈ M and for all f ∈ V . If M = Rd we will say that V has
total extra-invariance.
In other words, a shift invariant space has extra invariance if the the set of
invariance is bigger than Zd . One example of a translation invariant space in R

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

 LEARNING THE MODEL FROM THE DATA 145

is the Paley–Wiener space of functions that are bandlimited to [−1/2, 1/2] defined
by
P W = {f ∈ L2 (R) : supp(fb) ⊆ [−1/2, 1/2]}.
It is easy to prove that for a measurable set Ω ⊆ Rd , the space
VΩ := {f ∈ L2 (Rd ) : supp(fb) ⊆ Ω} (2)
is translation invariant. Moreover, Wiener’s theorem (see [15]) proves that any
closed translation invariant subspace of L2 (Rd ) is of the form (2).
Note that if Φ is a set of generators of V , i.e. V = S(Φ), and V has extra
invariance M then
S(Φ) = span{Tk φ : φ ∈ Φ, k ∈ Zd } = span{Tα φ : φ ∈ Φ, α ∈ M }.
In [2] the authors characterize those shift invariant spaces V ⊆ L2 (R) that have
extra-invariance. They show that either V is translation invariant, or there exists
a maximum positive integer n such that V is n1 Z-invariant.
The d-dimensional case is considered in [3]. There, a characterization of the
extra invariance of V when M is not all Rd is obtained.

4.2. Optimality and extra-invariance. Here we consider the approximation

problem for the class of finitely generated SIS with extra invariance under a given
proper subgroup M of Rd .
For a whole treatment we refer the reader to [10, 19, 11, 6].
Let us start introducing some notation. Let m, ` ∈ N, M be a closed proper
subgroup of Rd containing Zd , M ∗ = {x ∈ Rd : hx, mi ∈ Z ∀m ∈ M }, and
F = {f1 , . . . , fm } ⊆ L2 (Rd ). Define
`
VM = {V : V is a SIS of length at most ` and V is M -invariant}. (3)
d ∗
Let N = {σ1 , . . . , σκ } be a section of the quotient Z /M and {Bσ : σ ∈ N }
the partition defined by
[
Bσ = Ω + σ + M ∗ = (Ω + σ) + m∗ ,
m∗ ∈M ∗

where Ω is a section of the quotient Rd /Zd . We refer the reader to [3] for more
details.
For each σ ∈ N , we consider F σ = {f1σ , . . . , fm
σ
} ⊆ L2 (Rd ) where, fjσ is such
that fcσ = fbj χB for j = 1, . . . , m.
j σ

Also, let Fe = {f1σ1 , . . . , fm

σ1
, . . . , f1σκ , . . . , fm
σκ
}.
For each ω ∈ U, let GFe(ω) be the associated Gramian matrix of the vectors in
Fe with eigenvalues
λ1 (ω) ≥ · · · ≥ λmκ (ω) ≥ 0.
that are measurable functions.
Since fiσs is orthogonal to fiσt if s 6= t, the Gramian GFe(ω) is a diagonal block
matrix with blocks Gσ (ω), σ ∈ N . Here Gσ (ω) is the m × m Gramian associated

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

146 CARLOS CABRELLI AND URSULA MOLTER

to the data F σ . On the other hand we have that

Gσ (ω) = Uσ (ω)Λσ (ω)Uσ∗ (ω) a.e. ω ∈ U,

where Uσ are unitary and Λσ (ω) := diag(λσ1 (ω), . . . , λσm (ω)) ∈ Cm×m and they are
also measurable matrices. We also have λσ1 (ω) ≥ · · · ≥ λσm (ω) for each σ ∈ N .
Using the decompositions of the blocks Gσ we have that

GFe(ω) = U (ω)Λ(ω)U ∗ (ω),

where U has blocks Uσ in the diagonal, and Λ is diagonal with blocks Λσ . We want
to recall here that for almost each ω the matrix Λ(ω) collects all the eigenvalues
of the Gramian GFe(ω) and the columns of the matrix U (ω) are the associated left
eigenvectors. Note that an eigenvector associated to the eigenvalue λσj (ω) has all
the components not corresponding to the block σ equal to zero.
Now for each fixed ω ∈ U, we consider {(i1 (ω), j1 (ω)), . . . , (in (ω), jn (ω))} with
is (ω) ∈ N and js (ω) ∈ {1, . . . , m} and n = mκ such that
i (ω) i (ω)
λj11 (ω) ≥ · · · ≥ λjnn (ω) ≥ 0

are the ordered eigenvalues of GFe(ω), with corresponding left eigenvectors

Y (is (ω),js (ω)) ∈ Cn , for s = 1, . . . , n.
Here is (ω) indicates the block of the matrix GFe(ω) in which the eigenvalue
is (ω)
λjs (ω) (ω) is found and js (ω) indicates the displacement in this block of the matrix
i (ω)
GFe(ω). More precisely, we have that λjss (ω) (ω) coincides with λ(is (ω)−1)m+js (ω) (ω),
the ((is (ω) − 1)m + js (ω))-th eigenvalue of GFe(ω). When ω ∈ U is fixed, we will
write is instead of is (ω) and js instead of js (ω).
i (ω)
It can be proven (see [10]) that γs (ω) := λjss(ω) (ω) is measurable as a function
on ω for each s = 1, . . . , n, and the associated eigenvectors are also measurable.
Finally we define hs : Rd → C, for s = 1, . . . , `
m
(i ,j )
X
hs (ω) := θjiss (ω) y(iss−1)m+k
s
(ω)fbkis (ω), (4)
k=1

where θjiss (ω) = (λijss (ω))−1/2 if λijss (ω) 6= 0 and θjiss (ω) = 0 otherwise.
Now we are ready to state the main result of this section.

Theorem 4.3. Let m, ` ∈ N, and M be a closed proper subgroup of Rd containing

Zd . Assume that F = {f1 , . . . , fm } ⊆ L2 (Rd ) is given data and let VM
`
be the class
∗ `
defined in (3). Then, there exists a shift invariant space V ∈ VM such that
m
X
V ∗ = argmin kfj − PV fj k2 .
V `
∈VM j=1

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

 LEARNING THE MODEL FROM THE DATA 147

Furthermore, with the above notation,

(1) The eigenvalues {λσj (ω) : σ ∈ N , j = 1, . . . , m}, are Zd -periodic, measur-
able functions in L2 (U) and the error of approximation is
Xm Z X mκ
E(F, M, `) := kfj − PV ∗ fj k2 = λijss (ω) dω.
j=1 U s=`+1

(2) The functions {h1 , . . . , h` } defined in (4) are in L2 (Rd ) and if we define
cj = hj , then Φ = {ϕ1 , . . . , ϕ` } is a generator set for the op-
ϕ1 , . . . , ϕ` by ϕ
timal subspace V ∗ and the set {ϕi (· − k), k ∈ Zd , i = 1, . . . , `} is a Parseval
frame for V ∗ .

5. Approximation with translation and rotation invariant subspaces

In the previous sections, we have only considered optimization over subspaces
that are translation invariant and lack other important invariances, such as rota-
tional invariance, which are crucial for applications.
In [5] the authors study the approximation problem for subspaces that are invari-
ant under the action of a discrete locally compact group Γ, not necessarily commu-
tative, with some hypotheses. This class in particular includes the crystallographic
groups that split. So, the spaces become invariant under rigid movements. One re-
cent application of these results to datasets of digital images appeared in [4]. This
approach turns out to be mathematically very challenging and requires many dif-
ferent techniques such as fiberization, grammian analysis, frame theory and group
representation methods. In this survey we will describe the problem using the
straightforward example of crystallographic or crystal groups, which encompasses
all the vital components.

5.1. Crystal groups. Crystal groups (crystallographic groups or space groups)

are groups of isometries of Rd that generalize the notion of translations along
a lattice, allowing to move using different (rigid) movements in Rd following a
bounded pattern that is repeated until it fills up space. Precisely (see [13]):
Definition 5.1. A crystal group is a discrete subgroup Γ ⊆ Isom(Rd ) such that the
quotient Isom(Rd )/Γ is compact, where Isom(Rd ) is endowed with the pointwise
convergence topology.
Equivalently, one can define a crystal group to be a discrete subgroup Γ ⊆
Isom(Rd ) such that there exists a compact fundamental domain P for Γ, i.e. there
exists a bounded closed set P such that
[
γ(P ) = Rd and γ(P ◦ ) ∩ γ 0 (P ◦ ) = ∅ for γ 6= γ 0 ,
γ∈Γ
◦
where P is the interior of P .
Note that the set of translations on a lattice is the simplest of the crystal groups.
It is known that d-dimensional crystal groups are intrinsically related to regular
tessellations of Rd , being Γ = {τk : k ∈ Λ}, the group of translations (τk (x) = x+k)

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

148 CARLOS CABRELLI AND URSULA MOLTER

Figure 1. Two versions of a tiling of the ceiling in the Alhambra,

using a basic tile, its translations and rotations.

on a lattice Λ the simplest example.

We have the fundamental theorem of Bieberbach [7], [20] which states the fol-
lowing:

Theorem 5.2 (Bieberbach). Let Γ be a crystal subgroup of Isom(Rd ). Then

(1) Λ = Γ ∩ Trans(Rd ) is a finitely generated abelian group of rank d which
spans Trans(Rd ), and
(2) the linear parts of the symmetries Γ, the point group of Γ, is finite, and is
isomorphic to Γ/Λ.

(See also [16, IV-4]). Here Trans(Rd ) stands for translations of Rd .

We will denote the point group of Γ by G.

Remark 5.3.
• Note that the set Λ is not empty by Bierberach’s theorem [7] and consists of
translations on the lattice Λ which is isomorphic to Zd , and we will denote
by Tk for k ∈ Λ.
• The Point Group G of Γ is a finite subgroup of O(d), the orthogonal group
of Rd , that preserves the lattice of translations, i.e. GΛ = Λ. The simplest
examples are if G is a group of rotations, so we will abuse notation, and
denote the action of G on L2 (Rd ) by Rg for g ∈ G.

General results on crystal groups can be found for example in [14], [21], [17], [7],
and [8].
Note that the simplest example of a crystal group is the group of translations
on a lattice Λ, i.e. Γ = {Tk : k ∈ Λ}, where Tk (x) = x + k.

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

 LEARNING THE MODEL FROM THE DATA 149

One very important class of crystal groups are the splitting crystal groups:

Definition 5.4. Γ is called a splitting crystal group if it is the semidirect product

of the subgroups Λ and G. In this case Γ = Λ o G, and for each γ, γ e ∈ Γ, we have
γ·γ e = (k + g e
k, ge
g ), for γ = (k, g), γ k ∈ Λ and g, ge ∈ G and
k, ge) with k, e
e = (e
γ(x) = g(x) + k.

Every crystal group is naturally embedded in a splitting group, and very often
arguments for general groups can be relatively easy reduced to the splitting case
and then be proved for that simpler case. This justifies, that from now on Γ will
always be considered to be a splitting crystal group.

5.2. The structure of Γ-invariant spaces. Let us recall the structure of closed
subspaces of L2 (Rd ) that are invariant under the action of Γ = ΛoG, the semidirect
product of a uniform lattice Λ in Rd and a discrete and countable group G that
acts on Rd by continuous invertible automorphisms. We will assume that gΛ = Λ
for all g ∈ G, which implies that the Haar measure of Rd is invariant under the
action of G.
A closed subspace V ⊆ L2 (Rd ) is Γ-invariant if Tk Rg V ⊆ V for all (k, g) ∈
Γ. Here for f ∈ V , Tk f (x) = f (x − k), k ∈ Λ and Rg f (x) = f (g −1 x), g ∈ G.
Equivalently, V is Γ-invariant if
f ∈ V ⇒ Tk f ∈ V ∀ k ∈ Λ and Rg f ∈ V ∀ g ∈ G.

For an at most countable family Φ ⊆ L2 (Rd ), we will write

SΓ (Φ) := span{Tk Rg ϕ : k ∈ Λ, g ∈ G, ϕ ∈ Φ}.
SΓ (Φ) is a Γ-invariant space and the set Φ is called a set of generators. Note that,
since Tk Rg = Rg Tg−1 k , we also have that
SΓ (Φ) = span{Rg Tk ϕ : k ∈ Λ, g ∈ G, ϕ ∈ Φ}.

Since L2 (Rd ) is separable, if V is a Γ-invariant subspace of L2 (Rd ), there always

exists a countable set Φ ⊆ L2 (Rd ) such that V = SΓ (Φ).
Let V be a Γ-invariant subspace of L2 (Rd ). As before, we denote by L(V ), the
length of V , the minimum number of generators of V :
L(V ) = min{n : ∃ Φ = {ϕ1 , . . . , ϕn } : V = SΓ (Φ)}.
If V does not have a finite number of generators we set L(V ) = ∞.
Γ-invariant closed subspaces have been characterized in [5] in terms of a covari-
ance property of the range function associated to its Λ-invariant subspace.

Definition 5.5. Let Ω ⊆ R cd /Λ⊥ ≈ Λ.

cd be a Borel section of R b A range function
is a map
J : Ω → {closed subspaces of `2 (Λ⊥ )}.

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

150 CARLOS CABRELLI AND URSULA MOLTER

Theorem 5.6 ([5, Theorem 3.3]). Let Ω ⊆ R cd /Λ⊥ such

cd be a Borel section of R
∗ 2 d
that for ω ∈ Ω, g ω ∈ Ω ∀g ∈ G. A closed subspace V of L (R ) is Γ-invariant if
and only if it is Λ-invariant (shift-invariant by Λ) and its range function JV = J
satisfies
J (g ∗ ω) = rg−1 J (ω) , a.e. ω ∈ Ω , ∀g ∈ G.

5.3. Approximation by Γ-invariant subspaces. In this subsection we study

the approximation problem mentioned in the introduction. The idea is to find a
low dimensional model (a subspace), among all Γ-invariant subspaces that best fits
a given dataset. The subspace will be optimal for the data in the sense that it
minimizes the gauge function E, defined in (1) The importance of the approach
in this subsection is that our class includes subspaces that are invariant by rigid
movements in Rd , since we are able to include rotations and symmetries.
We will always assume that G is finite, and that a Borel section of R cd /(Λ⊥ o G)
exists.
Using the previously mentioned characterization of these spaces, we can employ a
strategy similar to that used for shift-invariant spaces to obtain the desired theorem
(for proofs of this section, see [5]).
We start with a necessary lemma.

Lemma 5.7 ([5, Lemma 5.1]). Let F g be the family {R(g)fi : (i, g) ∈ Im × G} ⊆
L2 (Rd ) ordered with the lexicographical ordering of Im × G := {1, 2, . . . , m} × G,
and let GF g be its Grammian as before.
1. For ω ∈ Ω, let {σi,g (ω)2 : (i, g) ∈ Im × G} be the eigenvalues of G(ω)
ordered decreasingly with the lexicographical ordering of Im × G, counted
with their multiplicity. Then they are G-invariant, in the sense that

σi,g (g0∗ ω) = σi,g (ω) ∀ (i, g) ∈ Im × G, ∀ g0 ∈ G, a.e. ω ∈ Ω.

2. For ω ∈ Ω0 , let {V i,g (ω) : (i, g) ∈ Im × G} ⊆ Cm|G| be the corresponding

orthonormal eigenvectors of G(ω), and denote the components of the (i, g)-
i,g
th eigenvector by {Vj,q (ω) : (j, q) ∈ Im × G} ⊆ C. Then, it is possible to
obtain a family of orthonormal eigenvectors of G(ω) at a.e. ω ∈ Ω whose
components satisfy
i,g ∗ i,g
Vj,q (g0 ω) = Vj,g 0q
(ω) ∀ g0 ∈ G, a.e. ω ∈ Ω.

Theorem 5.8 ([5, Theorem 5.2]). Let F = {f1 , . . . , fm } be a set of functional data
in L2 (Rd ). Using the same notations as in Lemma 5.7, the following holds:
1. For all κ ∈ {1, . . . , m} there exists a Γ-invariant space W ⊆ L2 (Rd ) gen-
erated by Γ-orbits of a family {ψi }κi=1 ⊆ L2 (Rd ) such that

E(F, W) = min{E(F, W) : V ⊆ L2 (Rd ) , Γ-invariant and L(V) ≤ κ}

and the system {Tk Rg ψi : k ∈ Λ, g ∈ G, i ∈ {1, . . . , κ}} is a Parseval

frame of W.

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

 LEARNING THE MODEL FROM THE DATA 151

2. The approximation error for the minimizing space W is given by

Xm XZ
E(F, W) = σ(i,g) (ω)2 dω.
i=κ+1 g∈G Ω0

3. A family {ψi }κi=1 ⊆ L2 (Rd ) that generates a minimizer W is given by

X 0
T[ψi ](ω) = Cij,g (ω)T[Rg0 fj ](ω),
(j,g 0 )∈Im ×G

where
0 X
Cij,g (ω) = i,g
θi,g (ω)Vj,g 0 (ω)χ ∗
g Ω0
(ω) , i = 1, . . . , κ
g∈G

and θi,g (ω) = (σi,g (ω))−1 if σi,g (ω) 6= 0 and 0 otherwise. All identities
hold for a.e. ω ∈ Ω.

References
[1] A. Aldroubi, C. Cabrelli, D. Hardin, and U. Molter, Optimal shift invariant spaces and
their Parseval frame generators, Appl. Comput. Harmon. Anal. 23 no. 2 (2007), 273–283.
DOI MR Zbl
[2] A. Aldroubi, C. Cabrelli, C. Heil, K. Kornelson, and U. Molter, Invariance of a shift-
invariant space, J. Fourier Anal. Appl. 16 no. 1 (2010), 60–75. DOI MR Zbl
[3] M. Anastasio, C. Cabrelli, and V. Paternostro, Invariance of a shift-invariant space in
several variables, Complex Anal. Oper. Theory 5 no. 4 (2011), 1031–1050. DOI MR Zbl
[4] D. Barbieri, C. Cabrelli, E. Hernández, and U. Molter, Optimal translational-rotational
invariant dictionaries for images, in Proc. SPIE 11138, Wavelets and Sparsity XVIII, 2019.
DOI
[5] D. Barbieri, C. Cabrelli, E. Hernández, and U. Molter, Approximation by group in-
variant subspaces, J. Math. Pures Appl. (9) 142 (2020), 76–100. DOI MR Zbl
[6] D. Barbieri, C. Cabrelli, E. Hernández, and U. Molter, Data approximation with time-
frequency invariant systems, in Landscapes of Time-Frequency Analysis—ATFA 2019, Appl.
Numer. Harmon. Anal., Birkhäuser/Springer, Cham, 2020, pp. 29–42. DOI MR Zbl
[7] L. Bieberbach, Über die Bewegungsgruppen der Euklidischen Räume. (Erste Abh.), Math.
Ann. 70 no. 3 (1911), 297–336. DOI MR Zbl
[8] L. Bieberbach, Über die Bewegungsgruppen der Euklidischen Räume. (Zweite Abh.), Math.
Ann. 72 no. 3 (1912), 400–412. DOI MR Zbl
[9] C. de Boor, R. A. DeVore, and A. Ron, The structure of finitely generated shift-invariant
spaces in L2 (Rd ), J. Funct. Anal. 119 no. 1 (1994), 37–78. DOI MR Zbl
[10] C. Cabrelli and C. A. Mosquera, Subspaces with extra invariance nearest to observed
data, Appl. Comput. Harmon. Anal. 41 no. 2 (2016), 660–676. DOI MR Zbl
[11] C. Cabrelli, C. A. Mosquera, and V. Paternostro, An approximation problem in multi-
plicatively invariant spaces, in Functional Analysis, Harmonic Analysis, and Image Process-
ing: A Collection of Papers in Honor of Björn Jawerth, Contemp. Math. 693, Amer. Math.
Soc., Providence, RI, 2017, pp. 143–165. DOI MR Zbl
[12] C. Eckart and G. Young, The approximation of one matrix by another of lower rank,
Psychometrika 1 (1936), 211–218. DOI Zbl

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

 152 CARLOS CABRELLI AND URSULA MOLTER

[13] D. R. Farkas, Crystallographic groups and their mathematics, Rocky Mountain J. Math.
11 no. 4 (1981), 511–551. DOI MR Zbl
[14] B. Grünbaum and G. C. Shephard, Tilings and Patterns, W. H. Freeman, New York, 1987.
MR Zbl
[15] H. Helson, Lectures on Invariant Subspaces, Academic Press, New York, 1964. MR Zbl
[16] J. S. Lomont, Applications of Finite Groups, Dover, New York, 1993. MR Zbl
[17] G. E. Martin, Transformation Geometry: An Introduction to Symmetry, Undergraduate
Texts in Mathematics, Springer-Verlag, New York-Berlin, 1982. MR Zbl
[18] E. Schmidt, Zur Theorie der linearen und nicht linearen Integralgleichungen. Zweite Ab-
handlung: Auflösung der allgemeinen linearen Integralgleichung, Math. Ann. 64 no. 2 (1907),
161–174. DOI MR Zbl
[19] R. Tessera and H. Wang, Uncertainty principles in finitely generated shift-invariant spaces
with additional invariance, J. Math. Anal. Appl. 410 no. 1 (2014), 134–143. DOI MR Zbl
[20] J. A. Wolf, Spaces of Constant Curvature, McGraw-Hill, New York, 1967. MR Zbl
[21] H. Zassenhaus, Beweis eines Satzes über diskrete Gruppen, Abh. Math. Sem. Univ. Hamburg
12 no. 1 (1938), 289–312. DOI MR Zbl

Carlos Cabrelli, Ursula Molter B

Departamento de Matemática, Facultad de Ciencias Exactas y Naturales, Universidad de
Buenos Aires, Ciudad Universitaria, Pabellón I, 1428 Buenos Aires, Argentina and IMAS,
Instituto de Investigaciones Matemáticas Luis A. Santaló, UBA–CONICET
[email protected], [email protected]

Received: May 1, 2023

Accepted: May 8, 2023

Rev. Un. Mat. Argentina, Vol. 66, No. 1 (2023)

View publication stats