Skip to main content

ORIGINAL RESEARCH article

Front. Phys., 10 May 2022
Sec. Statistical and Computational Physics
This article is part of the Research Topic Data-Driven Modeling and Optimization in Fluid Dynamics: From Physics-Based to Machine Learning Approaches View all 11 articles

Low-Rank Approximations for Parametric Non-Symmetric Elliptic Problems

Toms Chacn Rebollo
Tomás Chacón Rebollo1*Macarena Gmez MrmolMacarena Gómez Mármol2Isabel Snchez MuozIsabel Sánchez Muñoz3
  • 1Departamento EDAN & IMUS, Universidad de Sevilla, Sevilla, Spain
  • 2Departamento EDAN, Universidad de Sevilla, Sevilla, Spain
  • 3Departamento Matemática Aplicada I, Universidad de Sevilla, Sevilla, Spain

In this study, we obtained low-rank approximations for the solution of parametric non-symmetric elliptic partial differential equations. We proved the existence of optimal approximation subspaces that minimize the error between the solution and an approximation on this subspace, with respect to the mean parametric quadratic norm associated with any preset norm in the space of solutions. Using a low-rank tensorized decomposition, we built an expansion of approximating solutions with summands on finite-dimensional optimal subspaces and proved the strong convergence of the truncated expansion. For rank-one approximations, similar to the PGD expansion, we proved the linear convergence of the power iteration method to compute the modes of the series for data small enough. We presented some numerical results in good agreement with this theoretical analysis.

1 Introduction

This study deals with the low-rank tensorized decomposition (LRTD) of parametric non-symmetric linear elliptic problems. The basic objective in model reduction is to approximate large-scale parametric systems by low-dimension systems, which are able to accurately reproduce the behavior of the original systems. This allows tackling in affordable computing times, control, optimization, and uncertainty quantification problems related to the systems modeled, among other problems requiring multiple computations of the system response.

The PGD method, introduced by Ladévèze in the framework of the LATIN method (LArge Time INcrement method [22]) and extended to multidimensional problems by Ammar et al [2], has experienced an impressive development with extensive applications in engineering problems. This method is an a priori model reduction technique that provides a separate approximation of the solution of parametric PDEs. A compilation of advances in the PGD method may be found in [11].

Among the literature studying the convergence and numerical properties of the PGD, we can highlight [1], where the convergence of the PGD for linear systems of finite dimension is proved. In [8], the convergence of the PGD algorithm applied to the Laplace problem is proven, in a tensorized framework. The study [9] proves the convergence of the PGD for an optimization problem, where the functional framework is strongly convex and has a Lipschitz gradient in bounded sets. In [17], the authors prove the convergence of an algorithm similar to a PGD for the resolution of an optimization problem for a convex functional defined on a reflective Banach space. In [16], the authors prove the convergence of the PGD for multidimensional elliptic PDEs. The convergence is achieved because of the generalization of Eckart and Young’s theorem.

The present study is motivated by [5], where the authors present and analyze a generalization of the previous study [16] when operator A depends on a parameter. The least-squares LRTD is introduced to solve parametric symmetric elliptic equations. The modes of the expansion are characterized in terms of optimal subspaces of a finite dimension that minimize the residual in the mean quadratic norm generated by the parametric elliptic operator. As a by-product, this study proves the strong convergence in the natural mean quadratic norm of the PGD expansion.

A review of low-rank approximation methods (including PGD) may be found in the studies [18, 19]. In particular, minimal residual formulations with a freely chosen norm for Petrov–Galerkin approximations are presented. In addition, the study [10] gives an overview of numerical methods based on the greedy iterative approach for non-symmetric linear problems.

In [3, 4], a numerical analysis of the computation of modes for the PGD to parametric symmetric elliptic problems is reported. The nonlinear coupled system satisfied by the PGD modes is solved by the power iteration (PI) algorithm, with normalization. This method is proved to be linearly convergent, and several numerical tests in good agreement with the theoretical expectations are presented.

Actually, for symmetric problems, the PI algorithm to solve the PGD modes turns out to be the adaptation of the alternating least-squares (ALS) method thoroughly used to compute tensorized low-order approximations of high-order tensors. The ALS method was used in the late 20th century within the principal components analysis (see [6, 7, 20, 21]) and extended in [27] to the LRTD approximation of high-order tensors. Its convergence properties were subsequently analyzed by several authors; local and global convergence proofs within several approximation frameworks are given in [14, 25, 26]. Several generalizations were reported; as mentioned, without intending to be exhaustive, in the studies [1215, 23].

The convergence proofs within the studies [14, 25, 26] cannot be applied to our context as we are dealing with least-squares with respect to the parametric norm intrinsic to the elliptic operator, even for symmetric problems. Comparatively, the difference is similar to that between proving the convergence of POD or PGD approximations to elliptic PDEs. The POD spaces are optimal with respect to a user-given fixed norm, while the PGD spaces are optimal with respect to the intrinsic norm associated to the elliptic operator (see [5]). This use of the intrinsic parametric norm does not make it necessary the previous knowledge of the tensor object to be approximated (in our study, the parametric solution of the targeted PDE), as is needed in standard ALS algorithms.

In the present study, we propose an LRTD to approximate the solution of parametric non-symmetric elliptic problems based upon symmetrization of the problem (Section 2). Each mode of the series is characterized in terms of optimal subspaces of finite dimension that minimize the error between the parametric solution and its approximation on this subspace, but now with respect to a preset mean quadratic norm as the mean quadratic norm associated to the operator in the non-symmetric case is not well-defined (Section 3). We prove that the truncated LRTD expansion strongly converges to the parametric solution in the mean quadratic norm (Section 4).

The minimization problems to compute the rank-one optimal modes are solved by a PI algorithm (Section 5). We prove that this method is locally linearly convergent and identifies an optimal symmetrization that provides the best convergence rates of the PI algorithm, with respect to any preset mean quadratic norm (Section 6).

We finally report some numerical tests for 1D convection–diffusion problems that confirm the theoretical results on the convergence of the LRTD expansion and the convergence of the PI algorithm. Moreover, the computing times required by the optimal symmetrization are compared advantageously to those required by the PGD expansion (Section 7).

2 Parametric Non-Symmetric Elliptic Problems

Let us consider the mathematical formulation for parametric elliptic problems introduced in [5] that we shall extend to non-symmetric problems.

Let (Γ,B,μ) be a measure space, where we assume that the measure μ is σ-finite. Let H be a separable Hilbert space endowed with the scalar product (⋅, ⋅) and associated norm ‖ ⋅‖, denote by H′ the dual space of H and by ⟨⋅, ⋅⟩ the duality pairing between H′ and H. We will consider the Lebesgue space Lμ2(Γ) and the Bochner space Lμ2(Γ;H) and its dual space Lμ2(Γ;H), denoting ⋅, ⋅ as the duality between Lμ2(Γ;H) and Lμ2(Γ;H). We are interested in solving the parametric family of variational problems:

 Find u:ΓH such that auγ,v;γ=fγ,v,vH,μ-a.e. γΓ,(1)

where a(,;γ):H×HR is a parameter-dependent, possibly non-symmetric, bilinear form and f(γ) ∈ H′ is a parameter-dependent continuous linear form.

It is assumed that a (⋅, ⋅; γ) is uniformly continuous and uniformly coercive on H μ-a. e. γ ∈ Γ and there exist positive constants α and β independent of γ such that,

aw,v;γβwv,w,vH,μ-a.e. γΓ,(2)
αw2aw,w;γ,wH,μ-a.e. γΓ.(3)

By the Lax–Milgram theorem, problem (1) admits a unique solution μ-a. e. γ ∈ Γ. To treat the measurability of u with respect to γ, let us consider the problem:

Let f be a function that belongs to Lμ2(Γ;H) such that f(γ) = f(γ) μ-a. e. γ ∈ Γ,

 Find uLμ2(Γ;H) such that āu,v=f,v,vLμ2(Γ;H),(4)

where ā:Lμ2(Γ;H)×Lμ2(Γ;H)R is defined by

āw,v=Γawγ,v(γ);γdμγ.(5)

By the Lax–Milgram theorem, owing to (2) and (3), problem (4) admits a unique solution. Problems (1) and (4) are equivalent in the sense that (see [5])

uγ=uγ,μ-a.e. γΓ.

We shall consider a symmetrized reformulation of the formulation (4). Let us consider a family of inner products in H, {(,)γ}γΓ which generate the norms ‖ ⋅‖γ uniformly equivalent to the ‖ ⋅‖ norm and there exist αH > 0 and βH > 0 such that,

αHvvγβHv for all γΓ,(6)

considering the associated scalar product in Lμ2(Γ;H) and ∫Γ(⋅,⋅)γ (γ). As ā is continuous and coercive on Lμ2(Γ;H), there exists a unique isomorphism in Lμ2(Γ;H) that we denote A, such that

ΓAwγ,vγγdμγ=āw,v,w,vLμ2(Γ;H).(7)

Let us define a new bilinear form b̄:Lμ2(Γ;H)×Lμ2(Γ;H)R by

b̄w,vāw,Av.(8)

It is to be noted that b̄ is symmetric, as it can be written as

b̄w,v=ΓAwγ,Avγγdμγ.(9)

Thus, as a consequence of (2) and (3), the form b̄ defines a scalar product in Lμ2(Γ;H) and generates a norm equivalent to the usual norm is this space.

Now, problem (4) is equivalent to

 Find uLμ2(Γ;H) such that b̄u,v=f,Av,vLμ2(Γ;H)..(10)

We shall use this formulation to build optimal approximations of problem (1) on subspaces of finite-dimension, when the form ā(,) is not symmetric. For a given integer k ≥ 1, we denote by Sk the Grassmannian variety of H formed by its subspaces of dimension smaller than or equal to k and consider the problem:

minZSkb̄uuZ,uuZ,(11)

where u is the solution of problem (4) and uZ is its approximation given by the Galerkin approximation of problem (10) in Lμ2(Γ;Z):

Find uZLμ2(Γ;Z) such thatb̄uZ,v=f,Av,vLμ2(Γ;Z).(12)

Then, for any kN, an optimal subspace of dimension smaller than or equal to k is the best subspace of the family Sk that minimizes the error between the solution u and its Galerkin approximation uZ on this subspace, with respect to the mean quadratic norm generated by b̄. Theorem 4.1 of [5] proves the following result:

Theorem 2.1 For any k ≥ 1, problem (11) admits at least one solution.

3 Targeted-Norm Optimal Subspaces

It is assumed that we give a family of inner products {(,)H,γ}γΓ on H, which generate norms ‖ ⋅‖H,γ uniformly equivalent to the reference norm in H. Eventually, we may set (⋅,⋅)H,γ = (⋅, ⋅). Our purpose is to determine the inner products (⋅,⋅)γ (introduced in Section 2 and which we will call (⋅,⋅)γ,⋆) in such a way that the corresponding bilinear form b̄ defined by (8) actually is

b̄w,v=Γwγ,vγH,γdμγ.

In this way, the optimal subspaces are the solution of the problem

minZSkΓuγuZγH,γ2dμγ.(13)

Let us consider the operators Aγ,⋆: HH and the bilinear forms (⋅,⋅)γ,⋆ on H × H defined by

aw,Aγ,v;γ=w,vH,γ,w,vH,(14)
w,vγ,=Aγ,1w,Aγ,1vH,γ,w,vH.(15)

It is to be noted that Aγ,⋆ is an isomorphism on H and consequently (⋅,⋅)γ,⋆ is an inner product on H. Due to (2) and (3), the norms generated by these inner products are uniformly equivalent to the reference norm in H. Moreover, by (14) and (15),

Aγ,w,vγ,=w,Aγ,1vH,γ=aw,v;γ.(16)

Let us consider now the inner product in Lμ2(Γ;H) given by ∫Γ(⋅,⋅)γ,⋆ and the isomorphism A:Lμ2(Γ;H)Lμ2(Γ;H) defined by

ΓAwγ,vγγ,dμγ=āw,v,w,vLμ2(Γ;H).(17)

Then, it holds.

Lemma 1 Let Aγ : HH be the continuous linear operators defined by

Aγw,vγ=aw,v;γ,w,vH,μ-a.e. γΓ.(18)

Then, it holds

Awγ=Aγwγ,wH,μ-a.e. γΓ.(19)

This result follows from a standard argument using the separability of space H that we omit for brevity. Then, by Lemma 1, we have

Awγ=Aγ,wγ,wLμ2(Γ;H),μ-a.e. γΓ.(20)

Let us denote by b̄ the bilinear form on Lμ2(Γ;H) given by

b̄w,v=āw,Av,w,vLμ2(Γ;H).(21)

Then, by (20) and (14),

b̄w,v=Γawγ,Aγ,vγ;γdμγ=Γwγ,vγH,γdμγ.(22)

As a consequence, the optimal subspaces obtained by the least-squares problem (11) when b̄=b̄ satisfy (13).

Remark 1 When the forms a (⋅, ⋅; γ) are symmetric, if we choose

w,vH,γ=aw,v;γ,w,vH,μ-a.e. γΓ,(23)

then Aγ,⋆ defined by (14) is the identity operator in H. From (21), it follows that b̄=ā. We, thus, recover the intrinsic norm in the symmetric case to determine the optimal subspaces.

4 A Deflation Algorithm to Approximate the Solution

Following the PGD procedure, we approximate the solution of problem (1) by a tensorized expansion with summands of rank k, obtained by deflation. For all N ≥ 1, we approximate

uuNi=1Nsi, with siLμ2(Γ;H),(24)

computed by the following algorithm:

Initialization: let be u0 = 0.

Iteration: assuming ui−1, known for any i ≥ 1, set ei−1uui−1. Then,

ui=ui1+si, where si=ei1W(25)

with (ei1)W the approximation of ei−1 given by the problem (12) on an optimal approximation subspace W solution of the problem (11),

WargminZSkb̄ei1ei1Z,ei1ei1Z.(26)

It is to be noted that this algorithm does not need to know the solution u of problem (4) since ei−1 is defined in terms of the current residual fi=fAui1 by

ei1Lμ2(Γ;H) such that āei1,v=f,vāui1,vfi,vvLμ2(Γ;H)

The convergence of the uN to u is stated in Theorem 5.3 of [5] as the form b̄ is an inner product in Lμ2(Γ;H), with the generated norm equivalent to the standard one.

Theorem 4.1 The truncated series uN determined by the deflation algorithm (24)(26) satisfying

limNb̄uuN,uuN=0.

Consequently, uN strongly converges in Lμ2(Γ;H) to the solution u of problem (4).

5 Rank-One Approximations

An interesting case from the application point of view arises when we consider rank-one approximations. Indeed, when k = 1, the solution of (12) uZ can be obtained as

uZ(γ)=φγw,μ-a.e. γΓ, for some φLμ2(Γ),wH.

Then, the problem (11) can be written as (see [5], Sect. 6)

minvH,ψLμ2ΓJv,ψ,(27)

where

Jv,ψ=b̄uψv,uψv.(28)

Any solution of problem (27) has to verify the following conditions:

Proposition 1 If (w,φ)H×Lμ2(Γ) is a solution of problem (27), then it is also a solution of the following coupled nonlinear problem:

āφw,φAγv=f,φAγvvH,(29)
āφw,ψAγw=f,ψAγwψLμ2(Γ).(30)

We omit the proof of this result for brevity; let us just mention that conditions (29) and (30) are the first-order optimality conditions of the problem (27) that take place as the functional J:H×Lμ2(Γ)R and is Gateaux-differentiable. It is to be noted that the PGD method corresponds to replacing Aγ by the identity operator in (29)(30). From Theorem 2.1 and Proposition 1, there exists at least a solution to problem (27) and then to problem (29)(30). However, as functional J is not convex, there is no warranty that it admits a unique minimum. Then, a solution to problem (29)(30) could not be a solution to the problem (27).

Relations (29)(30) suggest an alternative way to compute the modes in the PGD expansion to solve (1). Indeed, we propose a LRTD expansion for u given by

uuNi=1Nφiwi,(31)

where the modes (wi,φi)H×Lμ2(Γ) are recursively computed as a solution of the problem (fifAui1) :

āφiwi,φiAγv=fi,φiAγvvH,āφiwi,ψAγwi=fi,ψAγwiψLμ2(Γ).(32)

If this problem is solved in such a way that its solution is also a solution of

minvH,ψLμ2ΓJiv,ψ,with Jiv,ψ=b̄uui1ψv,uui1ψv,(33)

then expansion (31) will be optimal in the sense of the expansion (24), where each mode of the series is computed in an optimal finite-dimensional subspace that minimizes the error.

6 Computation of Low-Rank Tensorized Decomposition Modes

In this section, we analyze the solution of the nonlinear problem (32) by the power iteration (PI) method. As the operator Aγ appears in the test functions, a specific treatment is needed, in particular, to compute targeted-norm subspaces. We also introduce a simplified PI algorithm that does not need to compute Aγ. Here, we report both.

We focus on solving the model problem (29)(30), which we assume to admit a nonzero solution. For simplicity, the notation f will stand either for the r. h. s. of the problem (4) or for any residual fi.

It is to be observed that (30) is equivalent to

Γφγaw,Aγw;γψγdμγ=Γfγ,Aγwψγdμγ,ψLμ2(Γ),

and then

φγ=φw,γfγ,Aγwaw,Aγw;γμ-a.e. γΓ.(34)

Thus, problem (29)(30) consists in

 Find wH such thatΓφw,γ2aw,Aγv;γdμγ=Γφw,γfγ,Aγvdμγ,vH.(35)

with φ(w,)Lμ2(Γ) defined by (34). For simplicity, we shall denote by φ(w) the function φ(w, ⋅).

Let us define the operator T : HH that transforms wH into T(w) ∈ H solution to the problem

Γφw,γ2aTw,Aγv;γdμγ=Γφw,γfγ,Aγvdμγ,vH.(36)

T is well-defined from (2) and (3), and a solution of (35) is a fixed point of this operator. Moreover, as Aγ is linear,

φλw=λ1φw and then Tλw=λTw,λR.

Thus, if (w, φ) is a solution to (29)(30), then (λ w, λ−1φ) is also a solution to this problem. So, we propose to find a solution to problem (35) with unit norm. For that, we apply the following PI algorithm with normalization:

Initialization:Given any nonzero w0H such that φ0 = φ(w0) is not zero in Lμ2(Γ).

Iteration: Knowing wnH, the following is computed :

iw̃n+1=TwnHiiwn+1=w̃n+1w̃n+1Hiiiφn+1=φwn+1Lμ2(Γ).(37)

The next result states that this iterative procedure is well-defined.

Lemma 2 It is assumed that for some nonzero wH, it holds that φ(w)Lμ2(Γ) is not zero. Then w̃=T(w) is not zero in H and φ(w̃) is not zero in Lμ2(Γ).

Proof First, by reduction to the absurd, it is assumed that T(w) = 0. From (36), we have

Γφw,γfγ,Aγvdμγ=0,vH.

In particular, for v = w and taking into account (34), we deduce that

Γ|fγ,Aγw|2aw,Aγw;γdμγ=0.

Then,

fγ,Aγw=0,μ-a.e. γΓ,

and thus, φ(w) = 0 is in contradiction with the initial hypothesis. This proves that T(w) is not zero. Second, arguing again by reduction to the absurd, it is assumed that φ(w̃)=0. From (34), we have

fγ,Aγw̃=0,μ-a.e. γΓ.

Then, setting v=w̃ in (36) and using (19), we obtain

b̄φww̃,φww̃=Γaφw,γw̃,φw,γAγw̃dμγ=0.

As b̄ is a scalar product in Lμ2(Γ;H), this implies that

φww̃Lμ2(Γ;H)=φwLμ2(Γ)w̃=0.

We have already proven that w̃0. So, φ(w) has to be equal to zero, in contradiction with the initial hypothesis. Thus, our assumption is false and φ(w̃) is not zero.

This result proves that if w0 and φ0 are not zero, then the algorithm (37) is well-defined.

6.1 Computation of Power Iteration Algorithm for Targeted-Norm Optimal Subspaces

From a practical point of view, in general, the algorithm (37) is computationally expensive. Indeed, in practice, H is a space of large finite-dimension and the integral in Γ is approximated by some quadrature rule with nodes {γi}i=1M. The method requires the computation of Aγiv for all the elements v on a basis of H and all the γi.

It is to be noted that when targeted subspaces are searched for, in the way considered in Section 3, the expression of algorithm (37) simplifies. Indeed, as a (w, Aγ,⋆v; γ) = (w,v)H,γ, then

φw,γ=fγ,Aγ,wwH,γ2μ-a.e. γΓ.(38)

In addition, the problem (36) that defines the operator T simplifies to problem:

Γφw,γ2Tw,vH,γdμγ=Γφw,γfγ,Aγ,vdμγ,vH.(39)

We shall refer to the method (38)(39) as the TN (targeted-norm) method.

6.2 A Simplified Power Iteration Algorithm

An approximate, but less expensive, method is derived from the observation that Aγ is an isomorphism in H. If we approximate the first equation in (32) by

āφiwi,φiAγ0v=fi,φiAγ0vvH,

for some γ0 ∈ Γ, then this equation is equivalent to

āφiwi,φiv=fi,φivvH.(40)

We then consider the following adaptation of the PI method (37) to compute an approximation of the solution of the optimality conditions (32):

Iteration: Known wnH, the following is computed:

iw̃n+1=T̂wnHiiwn+1=w̃n+1w̃n+1Hiiiφn+1=φwn+1Lμ2(Γ).(41)

where T̂(w) is computed by

Γφw,γ2aT̂w,v;γdμγ=Γφw,γfγ,vdμγ,vH,,(42)

where φ(w, γ) is defined by (34). We shall refer to method (41)(42) as the STN (simplified targeted-norm) method. The difference between the STN method and the standard PGD one is only the definition of the function φ that in this case is given by

φγfγ,waw,w;γμ-a.e. γΓ.

6.3 Convergence of the Power Iteration Algorithms

In this section, we analyze the convergence of the PI algorithms (37) and (41) (Theorem 6.1). We prove that the method with optimal convergence rate corresponds to the operator Aγ,⋆ introduced in Section 3, choosing all the inner products ‖ ⋅‖H,γ equal to the reference inner product ‖ ⋅‖.

As in [5], we shall assume that the iterates φn remain in the exterior of some ball of positive radius, say ɛ > 0, of Lμ2(Γ), that is,

φnLμ2Γε,n=0,1,2,.(43)

This is a working hypothesis that makes sense whenever the mode that we intend to compute is not zero, considering that the wn is normalized.

From the definition of Aγ and (6), it holds

α̃vAγvγβ̃v for all γΓ,with α̃=αβH,β̃=βαH,(44)

where α and β are given by (2) and (3), respectively, and αH and βH are defined in (6).

Let us define the function:

δr,s=2λk21+2r1rk21+2r1rλk2ss2,
where k=β̃α̃,andλ=k2 for method (37)βα for method (41)

It holds.

Theorem 6.1 It is assumed that (43) holds, and

Δ=δr,s̄<1,for some r0,1 with s̄=uLμ2(Γ;H)ε.(45)

Then, there exists a unique solution with norm 1 w of problem (35); the sequence {wn}n1 computed by either method (37) or method (41) is contained in the ball BH(w, r) if w0BH(w, r), and

wwn+1Δwwn,n0.(46)

As a consequence, the sequence {wn}n1 that is defined by either method (37) or method (41) is strongly convergent to w with linear rate and the following error estimate holds:

wwnΔnww0,n1,whenever w0BHw,r.(47)

Proof. Let us consider at first the method (37). Let xBH(w, r) such that φ(x)Lμ2(Γ)ε. Denote x̃=T(x) which by the definition of operator T in (36) is the solution to the problem

Γφx,γ2ax̃,Aγv;γdμγ=Γφx,γfγ,Aγvdμγ,vH.(48)

We aim to estimate wx̃x̃. To do that, from problems (35) and (48), we obtain

Γφw,γ2awx̃,Aγv;γdμγ=Γφw,γ2φx,γ2ax̃,Aγv;γdμγ+Γφw,γφx,γfγ,Aγvdμγ,vH.(49)

It holds

ay,Aγz;γ=Aγy,Aγzγβ̃2yz,y,zH,(50)

using (44). Thus,

fγ,Aγv=auγ,Aγv;γβ̃2uγv,vH.(51)

Moreover,

ay,Aγy;γ=Aγy,Aγyγα̃2y2yH.(52)

Setting v=wx̃ in (62) and using (50)-(52), we have

α̃2φwLμ2Γ2wx̃2β̃2φ2wφ2xLμ1Γx̃+φwφxLμ2(Γ)uLμ2(Γ;H)wx̃.(53)

To bound the second term in the r. h s. of (53), from (34), it holds

φw,γφx,γ=auγ,Aγwx;γaw,Aγw;γ+ax,Aγxw;γ+axw,Aγw;γaw,Aγw;γax,Aγx;γauγ,Aγx;γ.

Then, using (50) and (52),

|φw,γφx,γ|k21+k21x+1uγwxk21+2r1rk2uγwx,(54)

where k=β̃α̃. In the last estimate, we have used that as xBH(w, r),

w‖ − r ≤ ‖x‖ and then 1x11r. Therefore,

φwφxLμ2(Γ)ϕ1ruLμ2(Γ;H)wx,(55)

where

ϕ1r=k21+2r1rk2.

It is to be noted that from (34),

|φz,γ|=auγ,Aγz;γaz,Aγz;γk2uγz,zH.(56)

Then, from (54) and (56)

|φw,γ2φx,γ2||φw,γφx,γ||φw,γ+φx,γ|k2ϕ1r1+1xuγ2wx.

Hence,

φw2φx2Lμ1(Γ)ϕ2ruLμ2(Γ;H)2wx,(57)

where

ϕ2r=k2ϕ1r2r1r.

Combining (53) with (55) and (57), we deduce

wx̃uLμ2(Γ;H)φwLμ2Γ2k2ϕ1r+ϕ2rx̃wx.(58)

Setting v=x̃ in (48) and using (52) and (51), we obtain

α̃2φxLμ2Γ2x̃2β̃2φxLμ2ΓuLμ2(Γ;H)x̃.

Thus,

x̃k2uLμ2(Γ;H)φxLμ2Γk2uLμ2(Γ;H)ε=k2s̄.(59)

It holds wwx̃x̃2wx̃. Then, using (58) and (59), we deduce

wwx̃x̃2k2s̄2ϕ1r+ϕ2rk2s̄wx.

That is,

wx̃x̃Δwx,with Δgiven by (45).(60)

Estimate (46) follows from this last inequality for x = wn, assuming that wnBH(w, r). Assuming w0BH(w, r) this recursively proves that all the wn are in BH(w, r). Furthermore, suppose that there exists another solution to (35) with norm one in the ball BH(w, r), w*. In this case, estimating (60) for x = w* implies

ww*Δww*

because x̃=T(w*)=w*. Then w = w*, and there is uniqueness of solution with norm one in the ball BH(w, r). Finally, (47) follows from (46) by recurrence.

Let us now consider method (41). In this case x̃=T̂(x), by (42), is the solution of the problem

Γφx,γ2ax̃,v;γdμγ=Γφx,γfγ,vdμγ,vH.(61)

To estimate wx̃x̃, from problems (35) and (61), we obtain

Γφw,γ2awx̃,v;γdμγ=Γφw,γ2φx,γ2ax̃,v;γdμγ+Γφw,γφx,γfγ,vdμγ,vH.(62)

As f(γ), v = a (u(γ), v; γ) ≤ βu(γ)‖ ‖v‖, ∀ vH, setting v=wx̃, we have

αφwLμ2Γ2wx̃2Γφw,γ2awx̃,wx̃;γdμγβφ2wφ2xLμ1Γx̃+φwφxLμ2(Γ)uLμ2(Γ;H)wx̃.(63)

The functions φ(w) and φ(x) have the same expressions for methods (37) and (41).

Moreover, setting v=x̃ in (61)

αφwLμ2Γ2x̃2Γφw,γ2ax̃,x̃;γdμγ=Γφw,γfγ,x̃dμγ=Γauγ,φw,γx̃;γdμγβuLμ2(Γ;H)φwLμ2Γx̃.(64)

Hence, x̃βαs̄. Then, similarly to (58), we obtain

wx̃βαs̄2ϕ1r+ϕ2rx̃wxβαs̄2ϕ1r+ϕ2rβαs̄wx.(65)

As wx̃x̃2wx̃, the conclusion follows as for method (37).

Remark 2 The optimal convergence rate Δ corresponds to k = 1 and λ = 1, that is, α̃=β̃ and α = β. As α and β are predetermined, the optimal convergence rate can only be obtained with method (37). When the inner products (⋅,⋅)γ = (⋅,⋅)γ,⋆ and thus the operator Aγ = Aγ,⋆, introduced in Section 3 are used to construct the optimal targeted subspaces, it satisfies, by (15),

Aγ,vγ,2=vH,γ2,vH.

Then, from (44), choosing all the inner products ‖ ⋅‖H,γ equal to the reference inner product ‖ ⋅‖, we obtain α̃=β̃=1. Therefore, the convergence rate is optimal for method (37) with this choice. It can be interpreted as preconditioning of the problem to solve, similar to classical preconditioning to accelerate convergence in solving linear systems (For example, see [24]).

Remark 3 If we intend to compute a mode of order i ≥ 2, Theorem 6.1 and Remark 2 also hold, replacing f by the residual fi=fAui1 and u by the error uui−1, where ui−1 is defined by (31).

7 Numerical Tests

In this section, we discuss the numerical results obtained with the methods TN (38)(39) and STN (41)(42) to solve some non-symmetric second-order PDEs. Our purpose, on the one hand, is to confirm the theoretical results stated in Theorem 4.1 and Theorem 6.1, and on the other hand, to compare the practical performances of these methods with the standard PGD.

We consider a parametric 1D advection–diffusion problem with fixed constant advection velocity β,

γβuu=γfinΩ=0,10u0=1,u10=0(66)

We assume that β = 1, and then γ is the Péclet number. The source term is f = 1/500. We have set Γ = [2.5, 50]; then, there is a large asymmetry of the advection–diffusion operator.

Once the nonhomogenous boundary condition at x = 0 is lifted, problem (66) is formulated under the general framework (1) when the space H and the bilinear form a (⋅, ⋅) are given by H = {vH1(Ω) |v (0) = 0 }, and

au,v;γ=γβu,v+u,v.(67)

We endow space H with the H01(Ω) norm (that we still denote ‖ ⋅‖), which is equivalent to the H1(Ω) norm on H.

In practice, we replace the continuous problem (1) by an approximated one on a finite element space Hh formed by piecewise affine elements. In addition, the integrals on Γ are approximated by a quadrature formula constructed on a subdivision of Γ into M subintervals,

ΓψγdμγIMψ=i=1Mωiψγi.

This is equivalent to approximating the Lebesgue measure μ by a discrete measure μΔ located at the nodes of the discrete set ΓΔ = {γi, i = 1, … , M} with weights ωi, i = 1, … , M. Consequently, all the theoretical results obtained in the previous sections apply, by replacing the Lμ2(Γ,H) space by LμΔ2(ΓΔ,Hh). It is to be noted that

vLμΔ2ΓΔ,Hh2=IMv2.(68)

In our computations, we have used the midpoint quadrature formula with M = 100 equally spaced subintervals of Γ to construct IM and constructed Hh with 300 subintervals of Ω of the same length.

Test 1:

This first experiment is intended to check the theoretical results on the convergence rate of the PI algorithm, stated in Theorem 6.1, for the TN and STN methods: we consider optimal targeted subspaces, in the sense of the standard H01(Ω) norm. That is, using

w,vH,γ=w,vH01Ω=w,v,w,vH1Ω

to define the mappings Aγ,⋆ by (14) and the form b̄* by (22).

For each mode wi of the LRTD expansion (31), we have estimated the numerical convergence rate of the PI algorithm by

rin+1=winwin1win+1win.(69)

Tables 1 and 2 show the norm of the difference between two consecutive approximations and the ratios rin. We display the results for the first three modes.

TABLE 1
www.frontiersin.org

TABLE 1. Convergence rates of the PI algorithm for the first TN modes.

TABLE 2
www.frontiersin.org

TABLE 2. Convergence rates of the PI algorithm for the first STN modes.

We observe that the PI method converges with a nearly constant rate for each mode, in agreement with Theorem 6.1. The convergence rate is larger for the TN method, also as expected from this theorem. It is also noted that the convergence rates are smaller for higher-order modes.

In Figure 1, we present the comparison between the solution obtained by finite elements for γ = 2.7375 and γ = 49.7625 and the truncated series sum for the TN, the results for the STN are similar.

FIGURE 1
www.frontiersin.org

FIGURE 1. Solutions of problem (66) by the LRTD (called FLSTD within the figure) expansion computed with the TN method. Left γ = 2.7375 Right γ = 49.7624.

Test 2:

In this test, we compare the convergence rates of PGD, TN, and STN methods to obtain the LRTD expansion (31) for the problem (66).

Figure 2 displays the errors of the truncated series with respect to the number of modes, in norm Lμ2(Γ,H01(Ω)). A spectral convergence may be observed for the three expansions. We observe that the convergence of the TN expansion, in terms of the number of modes needed to achieve an error level, is much faster than the convergence of the STN expansion, while this one is faster than the PGD one. This is clarified if we consider the number of modes required to achieve an error smaller than a given level. We display these numbers for an error level of 10–6 in Table 3, where much more modes are needed by the PGD expansion. The TN and STN methods, thus, appear to be well-adapted to fit the asymmetry of the operator.

FIGURE 2
www.frontiersin.org

FIGURE 2. Convergence history of the PGD, TN, and STN series for Test 2.

TABLE 3
www.frontiersin.org

TABLE 3. Numerical behavior of the PGD, STN, and TN methods for Test 2.

Finally, we compare the CPU times required by the three methods. By construction, it is clear that to compute every single iteration of the PI method, the TN method is much more expensive since it involves the calculation of the Aγ,* operator for each finite element base function. However, due to the fast convergence of the associated LRTD expansion, it is less expensive than PGD to compute the expansion. Figure 3 displays the CPU times for the TN, STN, and PGD methods as a function of the number of subintervals M considered in the partition of Γ. The STN method is more expensive than the PGD method; this arises due to the small convergence rate of the PI algorithm with the STN method. However, the TN method is less expensive than the PGD one, requiring approximately half the CPU time.

FIGURE 3
www.frontiersin.org

FIGURE 3. Comparison of CPU times to compute the TN, STN, and PGD expansions.

8 Conclusion

In this study, we have proposed a new low-rank tensorized decomposition (LRTD) to approximate the solution of parametric non-symmetric elliptic problems, based on symmetrization of the problem.

Each mode of the series is characterized as a solution to a calculus of variation problem that yields an optimal finite-dimensional subspace, in the sense that it minimizes the error between the parametric solution and its approximation on this subspace, with respect to a preset mean quadratic norm. We have proven that the truncated expansion given by the deflation algorithm strongly converges to the parametric solution in the mean quadratic norm.

The minimization problems to compute the rank-one optimal modes are solved by the power iteration algorithm. We have proven that this method is locally linearly convergent when the initial data are close enough to an optimal mode. We also have identified an optimal symmetrization that provides the best convergence rates of the PI algorithm, with respect to the preset mean quadratic norm.

Furthermore, we have presented some numerical tests for 1D convection–diffusion problems that confirm the theoretical results on the convergence of the LRTD expansion and the convergence of the PI algorithm. Moreover, the computing times required by the optimal symmetrization compare advantageously to those required by the PGD expansion.

In this study, we have focused on rank-one tensorized decompositions. In our forthcoming research, we intend to extend the analysis to ranks k ≥ 2. This requires solving minimization problems on a Grassmann variety to compute the LRTD modes. We will also work on the solution of higher-dimensional non-symmetric elliptic problems by the method introduced in order to reduce the computation times as these increase with the dimension of the approximation spaces.

Data Availability Statement

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.

Author Contributions

TC and IS performed the theoretical derivations while MG did the computations.

Funding

The research of TC and IS has been partially funded by PROGRAMA OPERATIVO FEDER ANDALUCIA 2014-2020 grant US-1254587, which in addition has covered the cost of this publication. The research of MG is partially funded by the Spanish Research Agency - EU FEDER Fund grant RTI2018-093521-B-C31.

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s Note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors, and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. Ammar A, Chinesta F, Falcó A. On the Convergence of a Greedy Rank-One Update Algorithm for a Class of Linear Systems. Arch Computat Methods Eng (2010) 17:473–86. Number 4. doi:10.1007/s11831-010-9048-z

CrossRef Full Text | Google Scholar

2. Ammar A, Mokdad B, Chinesta F, Keunings R. A New Family of Solvers for Some Classes of Multidimensional Partial Differential Equations Encountered in Kinetic Theory Modeling of Complex Fluids. J Non-Newtonian Fluid Mech (2006) 139:153–76. doi:10.1016/j.jnnfm.2006.07.007

CrossRef Full Text | Google Scholar

3. Azaïez M, Chacón Rebollo T, Gómez Mármol M. On the Computation of Proper Generalized Decomposition Modes of Parametric Elliptic Problems. SeMA (2020) 77:59–72. doi:10.1007/s40324-019-00198-7

CrossRef Full Text | Google Scholar

4. Azaïez M, Belgacem FB, Rebollo TC. Error Bounds for POD Expansions of Parameterized Transient Temperatures. Comp Methods Appl Mech Eng (2016) 305:501–11. doi:10.1016/j.cma.2016.02.016

CrossRef Full Text | Google Scholar

5. Azaïez M, Belgacem FB, Casado-Díaz J, Rebollo TC, Murat F. A New Algorithm of Proper Generalized Decomposition for Parametric Symmetric Elliptic Problems. SIAM J Math Anal (2018) 50(5):5426–45. doi:10.1137/17m1137164

CrossRef Full Text | Google Scholar

6. ten Berge JMF, de Leeuw J, Kroonenberg PM. Some Additional Results on Principal Components Analysis of Three-Mode Data by Means of Alternating Least Squares Algorithms. Psychometrika (1987) 52:183–91. doi:10.1007/bf02294233

CrossRef Full Text | Google Scholar

7. Bulut H, Akkilic AN, Khalid BJ. Soliton Solutions of Hirota Equation and Hirota-Maccari System by the (m+1/G’)-expansion Method. Adv Math Models Appl (2021) 6(1):22–30.

Google Scholar

8. Le Bris C, Lelièvre T, Maday Y. Results and Questions on a Nonlinear Approximation Approach for Solving High-Dimensional Partial Differential Equations. Constr Approx (2009) 30(3):621–51. doi:10.1007/s00365-009-9071-1

CrossRef Full Text | Google Scholar

9. Cancès E, Lelièvre T, Ehrlacher V. Convergence of a Greedy Algorithm for High-Dimensional Convex Nonlinear Problems. Math Models Methods Appl Sci (2011) 21(12):2433–67. doi:10.1142/s0218202511005799

CrossRef Full Text | Google Scholar

10. Cancès E, Lelièvre T, Ehrlacher V. Greedy Algorithms for High-Dimensional Non-symmetric Linear Problems. Esaim: Proc (2013) 41:95–131. doi:10.1051/proc/201341005

CrossRef Full Text | Google Scholar

11. Chinesta F, Ammar A, Cueto E. Recent Advances and New Challenges in the Use of the Proper Generalized Decomposition for Solving Multidimensional Models. Arch Computat Methods Eng (2010) 17(4):327–50. doi:10.1007/s11831-010-9049-y

CrossRef Full Text | Google Scholar

12. Espig M, Hackbusch W, Rohwedder T, Schneider R. Variational Calculus with Sums of Elementary Tensors of Fixed Rank. Numer Math (2012) 122:469–88. doi:10.1007/s00211-012-0464-x

CrossRef Full Text | Google Scholar

13. Espig M, Hackbusch W. A Regularized Newton Method for the Efficient Approximation of Tensors Represented in the Canonical Tensor Format. Numer Math (2012) 122:489–525. doi:10.1007/s00211-012-0465-9

CrossRef Full Text | Google Scholar

14. Espig M, Hackbusch W, Khachatryan A. On the Convergence of Alternating Least Squares Optimisation in Tensor Format Representations. arXiv:1506.00062v1 [math.NA] (2015).

Google Scholar

15. Espig M, Hackbusch W, Litvinenko A, Matthies HG, Zander E. Iterative Algorithms for the post-processing of High-Dimensional Data. J Comput Phys (2020) 410:109–396. doi:10.1016/j.jcp.2020.109396

CrossRef Full Text | Google Scholar

16. Falcó A, Nouy A. A Proper Generalized Decomposition for the Solution of Elliptic Problems in Abstract Form by Using a Functional Eckart-Young Approach. J Math Anal Appl (2011) 376:469–80. doi:10.1016/j.jmaa.2010.12.003

CrossRef Full Text | Google Scholar

17. Falcó A, Nouy A. Proper Generalized Decomposition for Nonlinear Convex Problems in Tensor Banach Spaces. Numer Math (2012) 121:503–30. doi:10.1007/s00211-011-0437-5

CrossRef Full Text | Google Scholar

18. Nouy A. Low-rank Tensor Methods for Model Order Reduction. In: Handbook of Uncertainty Quantification (Roger GhanemDavid HigdonHouman Owhadi. Eds. Philadelphia, PA: Springer (2017). doi:10.1007/978-3-319-12385-1_21

CrossRef Full Text | Google Scholar

19. Nouy A. Low-rank Methods for High-Dimensional Approximation and Model Order Reduction. In: P Benner, A Cohen, M Ohlberger, and K Willcox, editors. Model Reduction and Approximations. Philadelphia, PA: SIAM (2017).

Google Scholar

20. Kiers HAL. An Alternating Least Squares Algorithm for PARAFAC2 and Three-Way DEDICOM. Comput Stat Data Anal (1993) 16:103–18. doi:10.1016/0167-9473(93)90247-q

CrossRef Full Text | Google Scholar

21. Kroonenberg PM, de Leeuw J. Principal Component Analysis of Three-Mode Data of using Alternating Least Squares Algorithms. Psychometrika (1980) 45:69–97. doi:10.1007/bf02293599

CrossRef Full Text | Google Scholar

22. Ladévèze P. Nonlinear Computational Structural Mechanics. New Approaches and Non-incremental Methods of Calculation. Berlin: Springer (1999).

Google Scholar

23. Mohlenkamp MJ. Musings on Multilinear Fitting. Linear algebra and its applications (2013) 438(2): 834–52. doi:10.1016/j.laa.2011.04.019

CrossRef Full Text | Google Scholar

24. Rasheed SM, Nachaoui A, Hama MF, Jabbar AK. Regularized and Preconditioned Conjugate Gradient Like-Methods Methods for Polynomial Approximation of an Inverse Cauchy Problem. Adv Math Models Appl (2021) 6(2):89–105.

Google Scholar

25. Uschmajew A. Local Convergence of the Alternating Least Squares Algorithm for Canonical Tensor Approximation. SIAM J Matrix Anal Appl (2012) 33(2):639–52. doi:10.1137/110843587

CrossRef Full Text | Google Scholar

26. Wang L, Chu MT. On the Global Convergence of the Alternating Least Squares Method for Rank-One Approximation to Generic Tensors. SIAM J Matrix Anal Appl (2012) 35(3):1058–72.

Google Scholar

27. Zhang T, Golub GH. Rank-one Approximation to High Order Tensors. SIAM J Matrix Anal Appl (2001) 23(2):534–50. doi:10.1137/s0895479899352045

CrossRef Full Text | Google Scholar

Keywords: low-rank tensor approximations, non-symmetric problems, PGD mode computation, alternate least squares, power iteration algorithm

Citation: Chacón Rebollo T, Gómez Mármol M and Sánchez Muñoz I (2022) Low-Rank Approximations for Parametric Non-Symmetric Elliptic Problems. Front. Phys. 10:869681. doi: 10.3389/fphy.2022.869681

Received: 04 February 2022; Accepted: 28 March 2022;
Published: 10 May 2022.

Edited by:

Traian Iliescu, Virginia Tech, United States

Reviewed by:

Yusif Gasimov, Azerbaijan University, Azerbaijan
Francesco Ballarin, Catholic University of the Sacred Heart, Brescia, Italy

Copyright © 2022 Chacón Rebollo, Gómez Mármol and Sánchez Muñoz. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Tomás Chacón Rebollo, chacon@us.es

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.