A new framework for the stability analysis of perturbed saddle-point problems and applications in poromechanics

This paper provides a new abstract stability result for perturbed saddle-point problems which is based on a proper norm fitting. We derive the stability condition according to Babuška's theory from a small inf-sup condition, similar to the famous Ladyzhenskaya-Babuška-Brezzi (LBB) condition, and the other standard assumptions in Brezzi's theory under the resulting combined norm. The proposed framework allows to split the norms into proper seminorms and not only results in simpler (shorter) proofs of many stability results but also guides the construction of parameter robust norm-equivalent preconditioners. These benefits are demonstrated with several examples arising from different formulations of Biot's model of consolidation.

READ FULL TEXT VIEW PDF

Authors

page 1

page 2

page 3

page 4

10/18/2018

Concentration of the Frobenius norms of pseudoinverses

In many applications it is useful to replace the Moore-Penrose pseudoinv...
10/18/2018

Concentration of the Frobenius norms of generalized matrix inverses

In many applications it is useful to replace the Moore-Penrose pseudoinv...
01/09/2021

Analysis of an abstract mixed formulation for viscoelastic problems

This study provides an abstract framework to analyze mixed formulations ...
11/17/2015

Understanding Adversarial Training: Increasing Local Stability of Neural Nets through Robust Optimization

We propose a general framework for increasing local stability of Artific...
05/01/2021

A Framework for Automatic Monitoring of Norms that regulate Time Constrained Actions

This paper addresses the problem of proposing a model of norms and a fra...
03/30/2021

Structural backward stability in rational eigenvalue problems solved via block Kronecker linearizations

We study the backward stability of running a backward stable eigenstruct...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1. Introduction

Saddle-point problems (SPPs) arise in various areas of computational science and engineering ranging from computational fluid dynamics [22, 48, 23], elasticity [2, 20, 13], and electromagnetics [39, 10] to computational finance [31]. Moreover, SPPs play a vital role in the context of image reconstruction [25], model order reduction [47], constrained optimization [21], optimal control [6], and parameter identification [16], to mention only a few but important applications.

In the mathematical modeling of multiphysics phenomena described by (initial-) boundary-value problems for systems of partial differential equations, SPPs often naturally arise and are frequently posed in a variational formulation. Mixed finite element methods and other discretization techniques can and have been successfully used for their discretization and numerical solution, see, e.g. 

[15, 19, 7, 10] and the references therein.

The pioneering works laying the foundations of the solution theory for SPPs have been conducted by Jindřich Nečas, Olga Ladyzhenskaya, Ivo Babuška, and Franco Brezzi [42, 33, 3, 14], see also the contributions [4, 32].

Designing and analyzing discretizations and solvers for SPPs require a careful study of the mapping properties of the underlying operators. Of particular interest are their continuity and stability, which not only guarantee the well-posedness of (continuous and discrete) mathematical models but also provide the basis for error estimates and a convergence analysis of iterative methods and preconditioners, see, e.g. 

[37, 19, 38, 10], for a review see also [7].

Saddle-point problems/systems are of a two-by-two block form and characterized by an operator/matrix of the form

(1)

where and denote positive semidefinite operators/matrices and the adjoint/transpose of an operator/matrix . We consider the symmetric case in this paper where , , and . Problems in which are often referred to as perturbed saddle-point problems. They are in the focus of this paper.

More general SPPs in which (and ) are allowed to be nonsymmetric and have also been studied by many authors, see, e.g., [43, 15, 10], and the references therein. Their analysis, in general is more complicated and is mostly done following the monolithic approach, i.e., imposing conditions on rather than on , , , and separately.

Our work is motivated by the stability analysis of variational problems occurring in poromechanics (cf. [17]), a subarea of continuum mechanics which originates from the early works of Terzaghi and Biot [49, 8]. Various formulations of Biot’s consolidation model have been considered and analyzed since it had been introduced in [8, 9], including two-field ([40, 41]), three-field ([45, 46, 44, 34, 26]), and four-field-formulations ([51, 35, 30]), for generalizations to several fluid networks as considered in multiple network poroelastic theory (MPET), see also [5, 24, 27, 28, 29, 36, 50].

Although they typically relate more than two physical fields, or quantities of interest (except for the two-field formulation of Biot’s model), the variational problems arising from the above-mentioned formulations–subject to a proper grouping or rather aggregation of variables–result in symmetric two-by-two block systems of saddle point form characterized by a self-adjoint operator .

The abstract framework presented in the next section of this paper applies to such saddle-point operators. After introducing some notation, we recall the classical stability results of Babuška and Brezzi for classical (unperturbed) SPPs. Next, we focus on perturbed (symmetric) SPPs, initially summarizing some of the additional conditions which, together with the Ladyzhenskaya-Babuška-Brezzi (LBB) condition (small inf-sup condition), imply the necessary and sufficient stability condition of Babuška (big inf-sup condition). Our main theoretical result then follows in Section 2.3 where we propose a generalization of the classical Brezzi conditions for the analysis of perturbed SPPs with . These new conditions imply the Babuška condition. The fitted norms on which they are based provide a constructive tool for designing norm-equivalent preconditioners.

This paper does not discuss discretizations and discrete variants of inf-sup conditions, although the proposed framework directly translates to discrete settings where it would also allow for shorter and simpler proofs of the well-posedness of discrete models and error-estimates for stable discretizations.

2. Abstract framework

2.1. Notation and problem formulation

Consider two Hilbert spaces and equipped with the norms and induced by the scalar products and , respectively. We denote their product space by and endow it with the norm defined by

(2)

Next, we introduce an abstract bilinear form on defined by

(3)

for some symmetric positive semidefinite bilinear forms on , on , i.e.,

(4)
(5)
(6)
(7)

and a bilinear form on .

We assume that , and are continuous with respect to the norms and , i.e.,

(8)
(9)
(10)

Then each of these bilinear forms defines a bounded linear operator as follows:

(11a)
(11b)
(11c)
(11d)

Here and denote the dual spaces of and , respectively, and the corresponding duality pairing.

Associated with the bilinear form defined in (3) we consider the following abstract perturbed saddle-point problem

(12)

which can also be written as

thereby using the definitions and , or, in operator form

(13)

where

(14)

and , i.e., for all

The operator can also be represented in block form by

(15)

Problem (12) (and (13)) is called a perturbed saddle-point problem (in operator form) when and a classical saddle-point problem in the case .

2.2. Babuška’s and Brezzi’s conditions for stability of saddle-point problems

As is well-known from [3], the abstract variational problem (12) is well-posed under the following necessary and sufficient conditions (i) and (ii) given in the following theorem.

Theorem 1 (Babuška [3]).

Let be a bounded linear functional. Then the saddle-point problem (12) is well-posed if and only if there exist positive constants and for which the conditions

(16)
(17)

hold. The solution then satisfies the stability estimate

Remark 1.

Estimate (16) ensures continuity, that is, boundedness of the operator from above, whereas (17) is a stability condition, sometimes referred to as Babuška condition, which grants boundedness of from below.

For the classical saddle-point problem, i.e., , the following theorem which we formulate under the conditions that is symmetric positive semidefinite and

(18)

has been proven in [14], see also [15, 10].

Theorem 2.

[Brezzi [14]] Assume that the bilinear forms and are continuous on and on , respectively, is symmetric positive semidefinite, and also that

(19)
(20)

hold. Then the classical saddle-point problem (problem (12) with ) is well-posed.

Remark 2.

Note that if , the statement of the above theorem (Theorem 2) remains valid if we identify any two elements , for which is an element of , i.e., replacing the space with the quotient space and also the norm with , the latter being defined by

In this case, the solution is only unique up to an arbitrary element .

For the classical saddle-point problem Brezzi’s stability condition (20) and the continuity of imply Babuška’s stability condition (17), see [18], where it has also been shown that from (17) it follows (20) and the inf-sup condition for in the kernel of , the latter being equivalent to the coercivity estimate (19) if is symmetric positive semidefinite.

Obviously, the stability condition (17) directly applies to perturbed saddle-point problems, a reason why they can be studied using Babuška’s theory. However, conditions (19) and (20) together with the continuity of , and in general are not sufficient to guarantee the stability condition (17) when . Additional conditions to ensure (17) have been studied, for example in [15, 12, 10].

In [15] it has been shown that a condition on the kernel of can be used as an additional assumption to ensure well-posedness of the perturbed saddle-point problem, that is, in particular, for Babuška’s inf-sup condition (17) to hold. This condition is expressed in terms of the following auxiliary problem

(21)

and requires the following general assumption:

Assumption 1.

There exists a such that for every and for every it holds that the solution of (21) is bounded by .

The theorem then reads as follows:

Theorem 3 (Brezzi and Fortin [15]).

Assume that , and are continuous bilinear forms on , on , and on , respectively. Further assume that and are symmetric positive semidefinite. Finally, let (19), (20) (conditions (i) and (ii) from Theorem 2) and Assumption 1 be satisfied. Then for every and every problem (12) with as defined in (3) and has a unique solution in where

Moreover, the estimate

holds with a constant only depending on , , , and .

Remark 3.

The result in [15] is more general than Theorem 3 in that it applies also to non-symmetric but positive semidefinite . We are considering only the case of symmetric positive semidefinite in this paper.

In order to ensure the boundedness (continuity) of the symmetric positive semidefinite bilinear form with respect to the norm , and more generally the boundedness of with respect to the combined norm , it is natural to include the contribution of in the norm , e.g., by defining via

(22)

for a proper seminorm or norm and a parameter .

As has been shown in [12] the stability of the perturbed saddle-point problem then can be proven under Brezzi’s conditions for the classical saddle-point problem and the additional condition

(23)

where is defined by

(24)

and provides a specific choice of , i.e., . The corresponding theorem then reads as:

Theorem 4 (Braess [12]).

Assume that the classical saddle-point problem is stable, i.e., Brezzi’s conditions (19) and (20) are fulfilled. If in addition condition (23) holds with and we choose in (24) for , then the perturbed saddle-point problem (12) is stable under the norm and the stability constant in (17) depends only on , and and the choice of .

Remark 4.

Note that as it can easily be seen if is symmetric positive semidefinite, condition (23) is equivalent to the condition that there exists a constant such that

(25)

Moreover, as shown in [12], then (25) is also equivalent to the condition that there exists a constant such that

(26)

Since (23) is an inf-sup condtion for which can be interpreted as a big inf-sup condition on for under the specific norm , see (26), Theorem 4 still does not provide us with the desired stability result in terms of conditions on , and separately. On the other hand, Theorem 3 requires the solution of the auxiliary problem (21) on for which one has to verify Assumption 1 which, in some situations, is a difficult task.

Our aim is to avoid the latter and still impose Brezzi-type conditions, in particular a small inf-sup condition on . In the next section, we will prove a theorem (Theorem 5) which ensures the stability of the perturbed saddle-point problem (12) under conditions which are equivalent to the conditions in Brezzi’s theorem (Theorem 2) when the perturbation term vanishes. Moreover, our approach provides a framework suited for finding norms in which stability can be shown and allows for simplifying and shortening proofs based on the result of Babuška.

2.3. A new framework for the stability analysis of perturbed saddle-point problems

The key idea for studying and verifying the stability of perturbed saddle-point problems we follow in this paper is to construct proper norms as part of an abstract framework which applies to a variational formulation of various multiphysics models. As we have already observed in the previous subsection, a norm-splitting of the form (22) is quite natural if the symmetric positive semidefinite perturbation form is not identical to zero. For fixed , the norm defined in (22) is equivalent to the norm defined by

(27)

where is a linear operator, which follows from the fact that the norm is induced by the scalar product on the Hilbert space . Now we introduce the following splitting of the norm defined by

(28)

where is a proper seminorm, which is a norm on satisfying

and is defined by

(29)

Here, is an isometric isomorphism since is an isometric isomorphism (Riesz isomorphism), i.e.,

Remark 5.

Note that both and can be seminorms as long as they add up to a full norm. Likewise, only the sum of the seminorms and has to define a norm. In some particular situations, it is also useful to identify certain of the involved seminorms with , in which case the corresponding splitting becomes a trivial splitting. The splitting (28) is closely related to a Schur complement type operator, corresponding to the modified (regularized) bilinear form resulting from by replacing with .

In order to present our main theoretical result, we give the following definition.

Definition 1.

Two norms and on the Hilbert spaces and are called fitted if they satisfy the splittings (27) and (28), respectively, where is a seminorm on and and are seminorms on the latter defined according to (29).

Theorem 5.

Let and be fitted norms according to Definition 1, which immediately implies the continuity of and in these norms with and , cf. (9)–(10). Consider the bilinear form defined in (3) where is continuous, i.e., (8) holds, and and are symmetric positive semidefinite. Assume, further, that satisfies the coercivity estimate

(30)

and that there exists a constant such that

(31)

Then the bilinear form is continuous and inf-sup stable under the combined norm defined in (2), i.e., the conditions (16) and (17) hold.

Remark 6.

The continuity of readily follows from

If is induced by the bilinear form then the continuity of also follows directly from the definition of the fitted norms.

Remark 7.

Theorem 5 is a generalization of Theorem 2 in the sense that given two norms and under which the conditions of Theorem 2 are satisfied, one can always find two fitted equivalent norms and such that the conditions of Theorem 5 are satisfied in these fitted norms when .

More specifically, for , we have and . If we define the fitted norm by choosing

(32)

then (30) obviously holds. In addition, there exits a constant such that (see [10, Proposition 4.3.4])

(33)

At the same time, under the conditions of Theorem 2, the continuity of and in the norm and , we have

(34)

Thus the fitted norm is equivalent to the norm , and (31) is induced by (20).

Remark 8.

Note that under the conditions of the theorem the coercivity of on in the (semi-) norms and are equivalent since for all . The inf-sup condition (31), however, uses the seminorm instead of as in Brezzi’s condition (20).

Proof of Theorem 5.

Demonstrating (16) is straightforward since

with .

In order to prove (17) for a positive constant , which will be selected later, we choose

(36)

where is such that

(37a)
(37b)

and

(38)

where

(39)

Note that the existence of an element satisfying (37) follows from (31).

Then we have

and, consequently,

Hence, it follows

(42)

Moreover, for the same choice of and , we obtain

and, hence, for and , we have

(43)

Together,  (42) and (43) imply the inf-sup condition (17) which can equivalently be formulated as

(44)

because the supremum on the left-hand side of (44) is bounded from below by

if we insert any fixed , in particular the choice we made and for which we proved

Remark 9.

The statement of Theorem 5 remains valid if the norms and , as defined in (27) and (28), are replaced with equivalent norms and , hence using in (31) in this case. The proof remains unchanged and the only difference in the final result is that the inf-sup constant in (44) with respect to the (new) equivalent combined norm has to be scaled by the quotient of the constants in the norm equivalence relation for the combined norms. For that reason, without loss of generality, we can use the fitted norms defined by (27) and (28) directly in the formulation of the theorem.

3. Application to poromechanics

In this section we consider several mixed variational problems arising in the context of modeling flow in and deformation of porous media. All examples except the first two are based on different formulations of Biot’s consolidation model. Here, we use bold letters to denote vector-valued functions and the spaces to which they belong which means that we identify certain non-bold symbols from the abstract framework in the previous section with bold symbols, e.g.,

. To prove stability of these formulations, we assume that proper boundary conditions are imposed and the following classical inf-sup conditions, see [14], also [15, 10], for the pairs of spaces hold, i.e., there exist constants and such that

(45)
(46)

where the norms , and denote the standard , and norms and is the -inner product.

Example 1.

The first example, taken from [11], is the mixed variational problem resulting from a weak formulation of a generalized Poisson equation

where is a parameter. The bilinear forms generating are given by

where , . Using the norm fitting technique, we define , by

Obviously, , and (27) and (28) take the form

respectively. Since for all , condition (30) in Theorem 5 is satisfied with .

Finally, we have to verify condition (31) in Theorem 5, namely

which follows directly from the classical inf-sup condition (45) on the spaces since .

Example 2.

The second example we consider is taken from [34]. Its mixed variational formulation is given by

where is a parameter. The bilinear forms defining here are given by

In this example, we set , , and , to be

Then the operator is defined by , and the norm splittings (27) and (28) are given by

(49)