The treatment and the visualization of the
multidimensional data is extremely difficult problem due to the “curse of
dimensionality (exponential growth of the required computer memory with the
expansion of the problem dimensionality). By this reason we consider the
opportunities provided by the tensor form of the multidimensional problems and
their approximation by tensor decompositions for the operations with the
multidimensional data.
As examples we consider the function
defined in the domain
and corresponding to the probability
density in Boltzmann equation and the set of functions
, where
corresponds to the flow variables (density,
velocity components, inner energy), and
correspond to the parameters of
problem (Mach, Reynolds numbers, angle of attack, etc.).
Herein we consider the tensor as
a multi-way array [1,2] without discussion of the physical sense of tensors. In
our case it corresponds to the grid function, defined on the regular (it is
important) grid in the multidimensional space. By the gasdynamical variables we
always imply their discrete form
in this work. Thus, we consider
the variables corresponding an unsteady flow-field (
is the number of time step,
is the number of the gasdynamical
variable
)
as a tensor
.
|
(1)
|
Also, we consider the ensemble of the flow-fields as
the tensor
,
|
(2)
|
obtained when solving the considered problem in the
space of parameters
, that corresponds to the statements typical for the
generalized computational experiment [3].
Accordingly, the operator of the solution
evolution (propagator) is also tensor. In the simplest case (1), the propagator
is the tensor of the order 8 that acts on the gasdynamical variables.
.
|
(3)
|
Herein we imply the summation over repeating indexes that is not
standard for the operations with tensors, however, it may be convenient for our
purposes.
So, the discretizations of both the gasdynamical variables and the
corresponding propagators have the form of tensors. Nevertheless, this
circumstance is not usually stressed and is not used in applications due to the
huge needs for computer memory. The propagators are implicitly applied at the
numerical solution of PDE (partial differential equations). They may be easily
stated in explicit form for the single time step.
It should be noted that the tensor form of the numerical solutions
enables to compress and analyze the results for numerical solutions of the
multiparameter problems (defined in the spaces of great (more than three)
dimensionality). The compression of the data is performed using the tensor
decompositions that are the main topic of present paper.
It is important that the tensor form enables to find nontrivial
inner structures both in the solution and in the propagator. The simplest
examples concern tensors written in the vectorized and matricized forms. Both
vectorization and matrization provide more common and lucid forms of tensor
representations and are formally valid. Symmetric matrization (natural for the
propagator structure) has the form
for the 6D space. The corresponding vectorization
. However, one should
remember that the transition from vectors and matrices to tensor (inverse
transformation, tensorization) is not always feasible. For example, for the
length of vectors equal to the simple number this operation is impossible.
The eigenvectors and the eigenvalues
(specific for the matrix algebra) appear in the result of the vectorization and
matrization. These objects are not defined in the tensor form. However, they
can reflect some inner (hidden) structure of solution and can have the
nontrivial physical sense.
For example, in the paper [4] (concerning
the atmosphere dynamics) such eigenvectors correspond to the flow disturbances,
which maximally grow at the selected time interval (singular vectors). They may
be related with the eigenvectors of the operator generated by the product of
the forward and adjoint propagators. For this purpose, the gasdynamical
variables are vectorized as
and the flow evolution is described by the propagator
.
|
(4)
|
The norm of the solution has the appearance
.
|
(5)
|
The search for the maximally (in the selected
norm) growing linear disturbances
at time interval
is reduced to the search of the eigenvectors of the problem
corresponding to the maximum eigenvalue
.
The dynamic mode decomposition (DMD) (for the
unsteady Euler equations presented by [5,6]) may be provided as another example
for implicit vectorization of the solution and the matrization of the
propagator. In the DMD frame, the numerical solution of the aerogasdynamics
problem is vectorized at time step
and
is written as
(snapshot).
The linear operator
is assumed to exist such that
.
Then the snapshots form the
Krylov sequence
. Two sequences
and
,
are selected from it. Generally (in simplified form) these data enable
to construct the approximation of the propagator
(
is the Moore-Penrose pseudoinverse), which is written in compressed
form as the product of the rectangular matrices
.
|
(6)
|
This form enables the radical reduction of the
memory necessary for storage of the operator
. This circumstance enables to use the
propagator for the resolution of the set of interesting problems, such as the
search for the singular vectors, approximation of the Perron-Frobenius and
Koopman operators [5,6].
The applicability of the tensor statements of the computational
fluid dynamics problems is severely restricted by the curse of dimensionality (storage
of the tensor with the number of indices above three requires nonrealistic
memory) despite their straightforward form.
However, at present, the significant progress may be observed at
overcoming these difficulties that is connected with the application of tensor
decompositions [7,8,9,10].
In the present paper we consider the feasibility of the
approximation of the six-dimensional tensors using such tensor decompositions
as the canonical decomposition [7,8] and the tensor train [9,10] and illustrate
it by the numerical experiments. The choice of the tensor order is related to
the three-dimensional Boltzmann equation approximation, is not principal and
does not prevent from the use of the considered algorithms for the parametric
problems of aerogasdynamics.
For the further exposition we
need to use rather high number of rarely used (if not to use term “exotic”)
notations [1,11] that we present below for convenience.
The tensor space
in accordance with [1]
is
the tensor (outer) product
of vector spaces
.
Herein, the symbol
notes
the tensor (outer) product of vectors
(sometimes, the outer product is noted by symbol
, in order to distinguish it from the Kronecker product).
Tensor
(
(d-way array)) is the element of the tensor
space.
The tensor is described by following parameters:
Order
of the tensor is equal to the
number of it’s indices
(number
of modes, dimensions). There is
nodes over every
index (mode)
. The order of the tensor is
equal to the dimensionality of the space at approximation of functions.
Size
of the tensor is equal to the product of the number of
nodes over all dimensions
, corresponds to the number of
memory required for the tensor storage, and grows exponentially
(
) in dependence on
the tensor order
.
Rank
of tensor
is defined as the minimum number
of the layers of cores
(at fixed
are the normed
vectors), which is necessary for the tensor approximation in the following form
(canonical decomposition)):
.
|
(7)
|
Fiber of the tensor
is the vector that is obtained at varying
of one index at others
fixed .
The following fibers exist for three-dimensional
tensor:
(m
ode-1 fiber,
column),
(m
ode-2 fiber,
row), and
(m
ode-3 fiber, “tunnel”
).
The following tensor operations we use or discuss.
n –mode product of the tensor and the vector
is noted using symbol
as
. Every mode-n fiber is scalarly multiplied by the vector as
. The tensor of the less order is obtained in result.
Product of two tensors
is noted by the symbol
and has the form
.
Kronecker product
(
) of the arbitrary size matrices is the generalization of the outer product
from vectors to matrices. The element of the first matrix is multiplied by the
second matrix in the form that follows:
.
The Kronecker product
of matrices
,
,
may be written in the index form as follows:
Unfortunately, as can be seen from (8), the index notation of such
operations does not provide the lucidity that is common in the physical
applications [12].
Khatri-Rao product
is usually noted by
⊙
,
herein, it is more convenient to use the symbol
. It is used for matrices with the coinciding number of the columns. Every
element of the first matrix column is multiplied by the total column of the
second one. The
result
is
formed
as
the
column
.
For
,
A
⊙
B
=
.
|
(9)
|
Hadamard
product
(
)
.
|
(10)
|
(no summation over repeating indices) is
usually noted as
and corresponds the element-wise multiplication
of same dimension matrices.
Very often it is convenient to roll out the tensor into pancake or
stretch into a fiber. The corresponding matricization or vectorization (unfold)
are performed by the following operations.
Mode-n matricization
of the tensor
is denoted by
and sets the mode-n fibers in
the matrix column.
In general case, the matrization of the tensor
has the form of the transformation
,
. For
,
, the matrization, for example, may has
the appearance
,
,
,
.
Vectorization
unfolds the matrix
into the vector
,
.
The matrization and the vectorization of tensors are very popular since
they enable to use all spectrum of the linear algebra algorithms. However, their
practical application at the tensor order above three is restricted by the
curse of dimensionality.
We are interested in the
tensor decompositions by the tensors of the less order or size, such as
Tucker decomposition
, in the index form
Canonical decomposition
,
in the
index form
.
Tensor train
,
in the index form
The above notations,
as a rule, are cumbersome, not transparent from the intuitive viewpoint, are
used in very narrow domain and are unknown for the most specialists.
Unfortunately, it is impossible to describe the current state of affairs in the
tensor decomposition without these notations. However, we shall try to use more
common index notations where it is possible or, in some events, duplicating both
approaches.
The above mentioned
c
anonical decomposition
is written in the index form (for non-normed cores) as
|
(11)
|
This expression is
unique if not account for the permutation or scaling. The problem for the
determination of the set of cores
has the
appearance (in the variational form)
.
|
(12)
|
The rank of the tensor
is
the key parameter at the application of the canonical decomposition. In
accordance with [1,13] it is not computable due to the ill-posedness of the
problem
.
|
(13)
|
The canonical decomposition suffers from instabilities
and requires a regularization [1,13]. However, in accordance with [14] the
approximation of the positive functions (such as the probability density) by
the canonical decomposition engenders the well-posed stable statement. The
oscillating behaviour of the cores was observed in the present work, also.
However, it is not principal from the standpoint of the multidimensional
functions’ approximation.
The canonical decomposition implies the compression
, where
is the space dimension,
is the number of nodes
over single direction,
is the
tensor rank.
The canonical decomposition is equivalent to DMD
[5,6]
in the
two-dimensional case.
The canonical decomposition is some extension of
the
Principal Components Analysis (PCA)) at the tensor
order expansion over two (transition from matrices to the multi-way arrays) and
also enables to reduce the dimensionality of the problem.
The canonical decomposition
rather
often suffers from instabilities [13]. By this reason, the attempts to find
alternative decompositions are natural. The tensor train (TT) format [8] is one
of such attempts. The works exist ([9]) that state that the tensor train is more
stable if compare with the canonical decomposition.
The Tensor train
(TT)
enables to write the
-way
tensor
in the form
,
|
(14)
|
where
are the cores of the size
,
,
,
.
The tensor train is not unique transformation since it is invariant
regarding the transformation
,
.
The tensor train provides less compression
, if
compared with the canonical decomposition
.
The tensor train is an interesting
alternative to the canonical decomposition
.
The comparison of the
tensor train and the
canonical decomposition is interesting both from the viewpoint of the stability
of results and the computational efficiency
.
We hope to perform the corresponding comparison in future works.
The methods for the tensor decomposition calculation
may be divided into two subclasses: the methods which are based on the linear
algebra, for example [9] and the variational methods
[6,7].
The linear algebra based methods significantly apply tensor
matrization, singular decomposition and contain a lot of interesting and
original algorithms that enable to execute operations on cores without
appealing to approximated functions.
The variational statements, as a rule, are based on alternating
least squares (ALS)) [15,16], however, the matrization of tensors is also used.
The utilization of the tensor matrization is, by our
opinion, the weak point of both approaches, since it requires a huge memory.
However, we believe that the variational methods may be
relieved from this drawback. By this reason, herein we use certain combination
of alternating least squares and the stochastic gradient descent (SGD) [17,18,19],
which will be described below.
Roughly speaking, we minimize the discrepancy on the single randomly
selected
fiber using ALS, that enables us to resolve
one-dimensional problem with the moderate requirements to memory at every step.
The
alternating least
squares (ALS) method [15,16] is commonly used at the search for the tensor
decompositions and enables optimization of the single parameter while others
are fixed. For the canonical decomposition ALS is realized by
Khatri-Rao product,
usually, for
three-dimensional problems. For the case of our interest (multidimensional) [19,20]
cores
are determined by the consequent solution of the following
problem
.
|
(15)
|
Herein
is the mode-
k
matrization
of the tensor,
is the auxiliary matrix of the same dimensionality. One
may obtain the elegant expression for the minimum of (15)
,
|
(16)
|
which enables the estimation of the core
by the matrix algebra methods.
In the general case (at a variation of all parameters), the
convexity is not guaranteed and the gradient descent is not obliged to converge.
The fixation of the main part of parameters and the variation of the single
core enable to obtain the convex goal functional and to optimize it with
success. The low rate of the ALS convergence is the cost of the relative
universality of the method.
The approach to the canonical decomposition calculation using
expressions of Eq. (16) kind dominates at present. Unfortunately, this approach
is not applicable for our purposes, since it uses the tensor matrization, which
requires the same memory as the tensor itself. The memory, which is necessary for
the problems of considered class, is above the range of modern computers parameters
(in this paper we use tensors, formally containing
numbers), that excludes the application of the tensor
matrization.
The stochastic gradient descent
(SGD) is widely used in the problems of high dimensionality [17,18,19]. Paradoxically,
it enables to find the point of the functional minimum in the space of the
control parameters for less time than it takes to fully calculate this
functional to . This occurs since the local functional (on single random point
or small set of points (minibatch) is computed instead the global (batch) functional
(12) that requires huge time for computation. Rather often SGD is used in a combination
with ALS in order to overcome difficulties caused by huge memory requirements
at the
Khatri-Rao product
application
[17,18].
In our case we use the
functional computed on the single randomly selected fiber
(
at other fixed indices) or a small set of such fibers.
Corresponding algorithm is described in the Section that follows in details.
As we stated above, our
approach corresponds to some combination of the stochastic gradient descent (in
minibatch variant) and
alternating least squares method. We present it in details for
the single fiber case (the case for the set of fibers may be obtained by simple
summation of the discrepancy and the gradient) since we failed to find the
description of this algorithm in publications. We consider the six-dimensional
case and the following approximation
.
|
(17)
|
At the beginning, we determine the first core
related with the coordinate
. Other cores are determined consequently and
the corresponding expressions may be found by the cyclic change.
We select a fibre along
(
) by
the random uniformly distributed choice of other indices
.
Let’s consider the discrepancy along this fibre obtained
by summation over
of the local (point-wise) discrepancies
.
|
(18)
|
Here
is the exact magnitude of the function at the point
that is known beforehand (in the present work from the analytic
expressions for the test functions). In accordance with the ALS approach, the
discrepancy is considered to depend on the single core
. Let’s
disturb this core by
. The corresponding disturbance of the discrepancy has
the form
.
|
(19)
|
Let’s choose
, which is not equal to zero only at single
point
, that
free us from summation over
in (19). Then, one may extract the corresponding
value of the gradient in form:
.
|
(20)
|
The defect of expression (20) if compared with
ALS methods, using the Khatri-Rao product (16), is the impossibility of the direct
use of condition
(since the sought value
is summated over
) for
calculation of cores. The merit of expression (20) if compared with (16) is the
economy of the memory (matrization is not used).
The regularized term
should be added to (20) if the zero order
Tikhonov regularization (
) is used.
The steepest descent iterations that minimize the
functional (18) over the core
element at point
have the form
,
|
(21)
|
where
is
the iteration step.
Iterations on the single core and selected fibre
are performed until relaxation. The discrepancy over selected fibre (18) was
used as the stopping criterion. Iterations terminated at
.
We proceed to the next core using a new
randomly selected fibre past stopping iterations on the current core. The form
of the gradient is obtained by permutations in the second term of (19) while the
sum in (18) depends on the chosen fibre.
We proceed to the next step of the global
iteration (from the new values of cores) and again start the search for the
optimal cores from
past all cores are locally optimized. Formally, the
quality of the function approximation by the canonical decomposition at every
step of the global iteration may be estimated via the discrepancy that follows
.
|
(22)
|
Unfortunately, this functional is nod directly
computable (at least on usual personal computers) due to the problems with dimensionality
and the great demands for computer time. We numerically estimated its value using
the Monte-Carlo method used in the form
,
|
(23)
|
where at every step of summation
every index from
was chosen as the random uniformly distributed
number. In result, we computed the averaged over the ensemble sum of the
approximation error squares. The number of trials in the ensemble was in the range
. The results varied rather weakly at the change of
.
The optimization of all cores (and the
total process of the canonical decomposition generation) stopped at
,
.
In general, the combinations of ALS and SGD
are rather widespread [17,18], however, in all known to authors events they use
the Khatri-Rao product. The approach, presented here, does not use the Khatri-Rao
product in any form, but is based on a direct numerical differentiation of the
discrepancy, which is defined on the single fibre, and the gradient descent. By
this reason, neither tensor matrization nor cores product in the Khatri-Rao form
are not used, that enables the radical reduction of the memory requirements. The
increase of the number of iterations necessary for the problem solution is the
cost of this success. Fortunately, it does not lead to the sensible consequences
for the considered problems, since the computation time for the considered
problems on the personal computer (Intel I5, 2.66 GHz) remains in the limits of
the several minutes.
In the results of calculations we obtain the
approximation of the six-dimensional function
(more correctly, the tensor
, corresponding
the values of the function in the nodes of the regular grid) using the
canonical decomposition and the corresponding set of cores
.
The grid containing 100 nodes on every
coordinate was used in numerical experiments. Formally, the storage of
on
such grid requires
cells of memory that is not realistic neither from
the viewpoint of storage nor from the viewpoint of the visualization. We mark
that the memory necessary for cores with the rank 100 requires
cells, which illustrates super high compression of the information (
)
at application of the canonical decomposition.
The comparison of the numerical (obtained by
the direct numerical differentiation) and analytical gradients (obtained by
expression (20)) was performed during debugging and demonstrate their practically
complete coincidence.
Formally, the quality of the approximation
of the function by the canonical decomposition may be estimated using
discrepancy (22), but it was estimated using Monte-Carlo method (23).
The results of computations provide
sufficiently stable and reproducible error estimations.
The numerical tests were performed using the
authors’ codes written in Fortran-95 specifically for the considered problems.
The tests of the approximation of different
functions by the canonical decomposition is performed. The quality of the
approximation (17) is of interest at testing both from the viewpoint of visual
presentation and from the value of the discrepancy (23). The determination of
the real rank of the function and the convergence rate are of interest. The
corresponding data are presented in this Section. The six-dimensional functions
of the different rank of the tensor presentation are selected. This enables to
estimate not only the quality of the approximation but also the possibilities
for the estimation of the rank for the functions under consideration. The
following multidimensional functions are considered that are situated in the
order of the complexity (canonical decomposition rank) increasing.
1.
The product of vectors
|
(25)
|
It is the simplest function with the rank of the
tensor equal unit. Single fibre, approximating rank of cores in the range from
1 to 10, 30 iterations are used. The dependence of discrepancy on the rank is
provided in Table 1.
Table 1. The dependence of discrepancy (23) on the
rank in (17) for the function (25)
rank
|
1
|
2
|
3
|
4
|
5
|
10
|
discrepancy
(23)
|
|
|
|
|
|
|
These results show that the magnitude of the
discrepancy may serve as the indicator of the true rank of the tensor. The
noise in the results increases as the rank rises, obviously, this reflects the
instabilities occurring at the rank estimation arising due to the ill-posedness
of this problem for the canonical decomposition [1,13]. The results of the computations
(
) are
presented in Fig. 1 (exact function) and Fig. 2 (approximation, rank 5).
|
|
Fig. 1
Exact function
(25)
|
Fig. 2. Approximation of function (25), rank 5
|
2.
The sum of vectors
|
(26)
|
It is also very simple function, but it’s rank a priori
is unknown and it would be desirable to estimate it in computations. Single fibre,
rank in the
range
from 1 to 10, 30 iterations
are used. The dependence of the discrepancy on the rank is presented in Table
2.
Table 2. The dependence of discrepancy (23) on the
rank in (17) for the function (26)
rank
|
1
|
2
|
3
|
4
|
5
|
10
|
discrepancy
(23)
|
|
|
|
|
|
|
This function has the rank 3÷4 if the minimum
of the discrepancy is analyzed. Fig. 3 presents the exact function, Fig. 4
presents its approximation (rank 5), one may mark the sufficient coincidence.
|
|
Fig. 3. The exact function (26)
|
Fig. 4. The approximation of function (26)
|
3.
The sum of sines
|
(27)
|
It is the two-dimensional function in the six-dimensional
space that is visually significantly more complex if compare with (25) and (26).
The calculations demonstrates the rank of this function to be about 10. The
results of computations are presented in Fig. 5 (exact function) and Fig. 6 (approximation,
rank 10, 1 fibre,
, 13 iterations).
|
|
Fig. 5. Exact function (27)
|
Fig. 6. The approximation of function (27)
|
4.
Two-dimensional product of sines
|
(28)
|
It is also two-dimensional functions in the six-dimensional
space, it is the multiplicative analogue of (27). The rank of this function is
about 10. Results of computations are provided in Fig. 7 (exact function) and
Fig. 8 (approximation, rank 10, 1 fibre,
, 30 iterations).
|
|
Fig. 7.
The exact function
(28)
|
Fig. 8. The approximation of function (28)
|
The previous tests were performed in the
six-dimensional space over the two-dimensional functions and demonstrated
enough fast convergence (10-30 iterations) and low values of the discrepancy. Further
we consider truly multidimensional problems that require greater number of
iterations and demonstrate greater discrepancies.
5.
The Gaussian in the multidimensional space
Let’s consider the function in the multidimensional
space that is described by the following equation
|
(29)
|
Formally, this function is defined in the six-dimensional
space, but, really, it is one-dimensional (depends only on the radius) and is determined
by the product of vectors, so it’s rank equals unit. Table 3 presents the
dependence of discrepancy on the rank for the function (29)
Table 3. The dependence of discrepancy (23) on the
rank in (17) for the function (29)
rank
|
1
|
2
|
3
|
4
|
5
|
10
|
discrepancy
(23)
|
|
|
|
|
|
|
30 iterations and 1 fibre are used. The complete
coincidence of function and its approximation is observed for the rank 1. The noise
in results grows and the accuracy decreases (discrepancy grows) as the rank
increases.
Thus, the estimation of the exact rank of function
is necessary and the calculations for the greater rank (“with the reserve”) may
not provide the necessary quality.
|
|
Fig. 9. The exact function (29)
|
Fig. 10. The approximation of function (29), rank 5.
|
6.
The sum of sines
|
(30)
|
It is the most difficult for calculations variant that
is truly six-dimensional. The rank of this function is about 200 as will be
demonstrated in the next Section. The calculations for this variant converge
rather slow (about 300 iterations) and require a great enough rank. Figs. 11 and
12 demonstrate results for 10 fibres and rank 200 in the plane
, the
discrepancy
. Other variables correspond to the centres of the intervals on the
grid 100 (
).
|
|
Fig. 11. Exact function (30)
|
Fig. 12. The approximation of function (30), rank 200
|
Fig. 13 presents the behaviour of different convergence
criteria in the dependence on the number of iterations for the function (27). The
discrepancy over the fibre (eps_fibre), global discrepancy estimated by
Monte-Carlo method (eps_MC), and the norm of the gradient of discrepancy (grad norm)
are provided. The increasing of the discrepancy at transition to the next
global step (rank values refreshing) was permitted in the variant of the optimization,
illustrated by Fig. 13
|
Fig. 13. Different criteria of convergence in
dependence on the number of iterations.
|
The present variant of the optimization does
not provide the monotonic convergence, although, the minimization is observed
for the simple enough functions. The local convergence (summation on the fiber)
occurs at every step of the global iteration. The convergence of the gradient
norm is not observed. Slow and nonmonotonic convergence of the global discrepancy
(the difference between the exact and approximate solutions in
norm,
estimated by Monte-Carlo method (23)) is observed.
The transition to the variant of the
minimization, which prohibits the increase of discrepancy at the transition to
the next (random) set of fibres (at the next step of the global iteration) was
used for more complex functions. Actually, the gradient optimization at certain
moment is replaced by the stochastic search over fibres in this version of the
algorithm.
One may conclude from the intuitive
viewpoint that the utilization of several fibres instead single one (minibatch)
should improve the monotonicity of the convergence. However, the numerical experiments
demonstrated that, starting from certain moment, it spoils the convergence rate
and the achievable value of discrepancy. Fig. 14 presents the results of the
convergence for 1 and 5 fibres for function (27).
|
Fig. 14. The logarithm of the discrepancy (23) in
dependence on number of iterations for 1 and 5 fibres.
|
The numerical estimation of the tensor rank was
obtained by viewing the results over the rang magnitude at the solution of the
variational problem (23). It was assumed, that the discrepancy should decrease
as the rank expanses as it is usually observed in calculations. However, for some
simple functions the opposite behavoiur is demonstrated. For example, the
minimum of the discrepancy occurs at unit rank for the multidimensional
Gaussian (29). The expansion of the rank increases the discrepancy.
The dependence of discrepancy logarithm on
the rank is presented in Fig. 15 (1 fibre, 300 iterations) for the function described
by the Eq. (30).
|
Fig. 15. The dependence of the discrepancy on the
rank for function (30)
|
The dependence of the discrepancy on the
rank value is enough nonmonotonic for this function. Nevertheless, the rank
expansion decreases the discrepancy. Thus, the rank should be adapted for
discrepancy diminishing at calculations. The simplest variant (the run over
expanding number of the ranks) is used herein. However, the solution of the
problem in the consequently expanding space is rather costly. So, the search
for more suitable algorithm for the rank estimation is necessary despite the
available principal difficulties [13].
Fig. 16 presents the core
for function (30), rank is equal 100. Fig. 17 presents core
for function (30) for rank 250. The oscillating character of the
core is observed. No damping at the expansion of the rank is observed,
although, the accuracy of the approximation increases.
The difficulties with the determination of the
rank in canonical decomposition stimulate the investigation of other tensor decompositions,
the tensor train in particular.
|
|
Fig. 16,
rank
100
|
Fig. 17,
rank
250
|
If the initial approximation of the cores is
chosen as constants, for example
, all magnitudes of gradients for different
automatically
coincides (19) and the optimization process fails. So, the expression
was
used for the initial approximation of cores (the random normally distributed
value with dispersion
was used). Naturally, there is no convergence at
.
The instabilities were observed at the great
. The optimal magnitude is
,
which was used in all above described tests.
The obtained results depend on the random
selection of the fibres and on the noise in the initial approximation of cores
that partly make difficult comparison at debugging and the search for the
optimal parameters for the algorithm installation. The fixation of the start point
of the random number generator and storing the sequences of fibre coordinates
(used at optimization) enables the stabilization of results.
Due to the problem ill-posedness [13] the
zero order Tikhonov [21] quadratic regularization is provided. However, the
dumping of the oscillations at calculation of cores caused the break of the
approximation in numerical tests. So, the positive influence of the
regularization was not observed and the above provided numerical tests
correspond to the absence of the regularization. The ill-posedness of the
canonical decomposition does not have an influence in the frame of
approximation of functions. In the frame of the solution of the evolutional problems
in partial differential equations (PDE) [10] the stability of the tensor
decomposition is useful since it enables to perform some part of evolution in
the space of cores without returning to the space of functions. The possibility
of the evolution in the space of cores is provided in the frame of tensor train
[9]. There exists the set of operations including special TT-rounding
operation, which enables to reduce the rank of the approximation past several
steps. It explains the significant interest to the tensor train format. So, the
instability at the level of the determination of cores is not insurmountable
obstacle at the simulation of PDE at spaces of the high dimensionality.
Nevertheless, the transition to the more stable statements (tensor train) may enable
to obtain more fast algorithms.
There exists the coincidence of numerical
and analytical gradients, but the analytical calculation is faster about two
order of the magnitude at the considered problem. It is related with the circumstance,
that disturbance of the discrepancy (18) should be computed for any element of the
core at the numerical calculation of the gradient. The need for this operation
(and in summation over the coordinate) is absent at the analytical computation
(20).
The plausible idea that the transition from
the single fibre to several ones provides more stable optimization failed in
calculations. This transition slows down the convergence rate and deteriorates
the quality of optimization (increases the value of the reachable discrepancy).
The numerical tests show that the tensor
rank strongly depends on the kind of the approximated function.
As far the rank increases the noise in the
result rises, so the choice of the rank value “from above”, which partly
simplifies the algorithm, may cause the deterioration of the results.
The tensor decompositions are actively used
for the visualization purposes since they enable to build the easily computable
model, which approximate the difficult data set in the space of parameters. For
example, papers [22,23] use tensor train format and the cross approximation for
these purposes.
The tensor decompositions (canonical decomposition, tensor train,
hierarchical Tucker) are used for the economic solution of the multidimensional
problems of the Boltzmann equation type [7,8,10].
The canonical decomposition enables to efficiently
approximate and store the multidimensional functions. The computer time cost
for the operations with the functions in the six-dimensional space (at using
100 nodes along each coordinate that formally requires storage and operations
with
numbers)
takes 2-3 minutes of the personal computer (processor Intel I5, 2.66 GHz) at
memory requirements for cores store about
numbers.
The applicability of the tensor statements of the computational
fluid dynamics problems is restricted by the “curse of dimensionality”.
Fortunately, the application of the tensor decompositions provides some hope
for its overcoming.
The algorithm is offered that unifies the alternate least squares
and the stochastic gradient descent and requires the memory, which is far less
if compared with the methods using Khatri-Rao product.
The numerical experiments demonstrate that the application of this
algorithm for the canonical decomposition enables storing and visualization of functions
in the multidimensional space with the very moderate costs from the standpoint
of the memory and the time of computation. The results of the visualization of
the performed numerical experiments are provided.
1. W. Hackbusch.
Tensor Spaces and Numerical Tensor Calculus. Springer, 2012.
2. H. Yorick, S.
Willi-Hans, Matrix Calculus, Kronecker Product and Tensor Product: A
Practical Approach to Linear Algebra, Multilinear Algebra and Tensor
Calculus with Software Implementation, Singapore: World Scientific
Publishing Co. Pte. Ltd., 2019.
3. Alekseev, A.K., Bondarev,
A.E., Galaktionov, V.A., Kuvshinnikov A.E. Generalized Computational Experiment
and Verification Problems. Program Comput Soft 47, 177–184 (2021). https://doi.org/10.1134/S0361768821030026
4. Farrell B.F. and A.M. Moore, An
adjoint method for obtaining the most rapidly growing perturbation to
oceanic flows, J. Phys. Oceanogr, 22 338-349, 1992
5. Alekseev A.K., Bondarev A.E., On the application
of the dynamic mode decomposition in problems of computational fluid
dynamics, KIAM Preprints. 2018, N 154. p. 30
6. A.K. Alekseev, D.A. Bistrian, A.E. Bondarev, I.M.
Navon,
On Linear and Nonlinear Aspects
of Dynamic Mode Decomposition, Int. J. Numer. Meth. Fluids,
2016,
V. 82, Issue 6, p. 348–371
7.
A. M. P. Boelens, D. Venturi, D. M. Tartakovsky, Parallel tensor methods
for high-dimensional linear PDEs, J. Computat. Phys. 375 (2018) 519-539.
8.
Arnout M. P. Boelens, Daniele Venturi, Daniel M. Tartakovsky, Tensor
methods for the Boltzmann-BGK equation, arXiv:1911.04904v2 2020
9.
Oseledets I. V., Tensor- train decomposition,
SIAM J. Sci. Comput. , 33 (2011), pp. 2295–2317
10.
A.V.
Chikitkin, E.K. Kornev, V.A. Titarev, Numerical solution of the Boltzmann
equation with S-model collision integral using tensor decompositions,
arXiv:1912.04582v1 2019
11.
Kolda
T. G. and Bader B. W., Tensor Decompositions and Applications,
SIAM
Review, 51(3):455–500, 2009.
12.
Korenev
G.V., Tensor calculus, M. MIPT, 1995
13.
V.
D. Silva and L -H. Lim, Tensor rank and the ill-posedness of the best low rank
approximation problem, SIAM J. Matrix Anal. Appl., 30(3) 1084-1127, 2008.
14.
L-H.
Lim and Pierre Comon, Nonnegative approximations of nonnegative tensors,
Chemometrics 2009; 23: 432–441.
15.
P. Comon, X. Luciani, and A. L.F. De
Almeida, Tensor decompositions, alternating least squares and other tales,
Journal of Chemometrics, vol. 23, no. 7-8, pp. 393–405, 2009.
16.
A.
Uschmajew, Local convergence of the alternating least squares algorithm for
canonical tensor approximation, SIAM J. Matrix Anal. Appl. 33 (2) (2012)
639–652.
17.
C. Battaglino, G. Ballard, T.G. Kolda, A practical
randomized CP tensor decomposition, arXiv:1701.06600, 2017.
18. X. Fu, S. Ibrahim, H.-T. Wai, C. Gao, and K. Huang,
Block-randomized
stochastic proximal gradient for
low-rank tensor factorization,
IEEE Transactions on Signal Processing
, 2020.
19. Ioanna Siaminou,
Athanasios P. Liavas, An Accelerated Stochastic Gradient for Canonical
Polyadic Decomposition, arXiv:2109.13964v1 2021.
20.
Guoxu
Zhou, Andrzej Cichocki, and Shengli Xie, Accelerated Canonical Polyadic
Decomposition by Using Mode Reduction, arXiv:1211.3500v2, 2013
21.
Tikhonov, A.N.,
Arsenin, V.Y.: Solutions of Ill-Posed Problems. Winston and Sons, Washington
DC (1977).
22.
R.
Ballester-Ripoll, E. G. Paredes, and R. Pajarola, A Surrogate Visualization
Model Using the Tensor Train Format, SA '16: SIGGRAPH ASIA 2016 Symposium on
Visualization November 2016 Article No. 13 Pages 1–8 https://doi.org/10.1145/3002151.3002167
23.
Susanne
K. Suter, Tensor Approximation in Visualization and Graphics: Background
Theory, PhD thesis, University of Zurich, Switzerland, April 2013
24.
Oseledets I., Tyrtyshnikov E., TT-cross approximation for
multidimensional arrays, Linear Algebra Appl., 432 (2010), pp. 70–88.