Fine stability constants and error bounds for

asymmetric approximate saddle point problems

Vitoriano Ruas1 *
1
Institut Jean Le Rond d’Alembert, CNRS UMR 7190, Sorbonne Université, Paris, France.
e-mail: vitoriano.ruas@upmc.fr

Dedicated to the memory of Roland Glowinski

arXiv:2307.03742v1 [math.NA] 7 Jul 2023

Abstract
The theory of mixed finite element methods for solving different types of elliptic partial differ-
ential equations in saddle-point formulation is well established since many decades. However, this
topic was mostly studied for variational formulations defined upon the same finite-element prod-
uct spaces of both shape- and test-pairs of primal variable-multiplier. Whenever these two product
spaces are different the saddle point problem is asymmetric. It turns out that the conditions to be sat-
isfied by the finite elements product spaces stipulated in the few works on this case may be of limited
use in practice. The purpose of this paper is to provide an in-depth analysis of the well-posedness
and the uniform stability of asymmetric approximate saddle point problems, based on the theory of
continuous linear operators on Hilbert spaces. Our approach leads to necessary and sufficient con-
ditions for such properties to hold, expressed in a readily exploitable form with fine constants. In
particular standard interpolation theory suffices to estimate the error of a conforming method.

Keywords: Asymmetric; Mixed finite elements; Saddle point; Uniform stability; Weak coercivity.

AMS Subject Classification: 65N30, 70G75, 74A15, 76M30, 78M30, 80M30.

1 Introduction
The mathematical background for the numerical analysis of mixed finite element methods for solv-
ing saddle point problems is well established since the last quarter of century (cf. Brezzi & Fortin
(1991) [4] and references therein). Most of this theoretical setting applies to symmetric variational for-
mulations. Here the term symmetry refers to formulations with only one bilinear form defined upon
a product space for the pair primal variable-multiplier and also with the same product spaces for both
shape- and test-pairs. In the asymmetric case the two product spaces are different and/or the varia-
tional formulations incorporates two distinct bilinear forms having the multiplier as an argument. To
the best of the author’s knowledge, a handy framework for this case was lacking, even when the saddle
point formulation is perfectly symmetric, as far as the bilinear forms involved therein are concerned,
but a symmetry breaking occurs owing to the difference between both product spaces. In this paper we
provide general and easily exploitable formal grounds for studying the finite element approximation of
saddle point problems based on asymmetric variational formulations. Akin to Nicolaides (1982) [8], the
key to the problem are inf-sup conditions involving three constants. However, in contrast, essentially
nothing else is needed here to establish best possible error estimates.
According to the well known theory set up by Babuška (1971) [1] and Brezzi (1974) [3], a linear
variational problem is well-posed only if the underlying continuous bilinear form is weakly coercive. In
the finite dimensional case this reduces to proving that the vector space of shape-elements as related to
the one of test-elements in the underlying bilinear form satisfies only one inf-sup condition. This fact
* Sorbonne Université, Campus Pierre et Marie Curie, 4 place jussieu, Couloir 55-65, 4ème étage, 75005 Paris, France.

1
was highlighted in Cuminato & Ruas (2015) [6], where the weak coercivity constant was also refined.
The particular case of saddle point problems and their approximation was studied in Babuška (1973)
[2] and Brezzi (1974) [3] for symmetric variational formulations. These studies were extended by Nico-
laides (1982) [8] to non symmetric problems. In the latter work the approximate mixed formulation is
defined upon different product spaces of finite dimension for both shape- and test-pairs. However, the
conditions stipulated therein for uniform stability and in the error bounds are not always so easy to ex-
ploit in practice, since they involve subspaces that may be rather difficult to deal with. In this article we
build up a new theoretical framework for asymmetric saddle point problems, which circumvents such an
issue thanks to some tools of Functional Analysis (cf. Yosida (1980) [9]), among other ingredients.
To a large extent, the material presented here is an adaption to the asymmetric case of the technique
of analysis with fine constants employed by Dupire (1986) [7] for symmetric saddle point problems.

2 Preparatory material
In this section we use the following notations. For a given Hilbert space W the underlying inner product
is denoted by (·, ·)W , the corresponding norm is denoted by k · kW and its null vector is represented by
0W . The standard norm of a given continuous linear operator A between two Hilbert spaces is denoted
by kAk.
To begin with we prove
Lemma 2.1 Let V and U be two Hilbert spaces. If J is a bounded linear operator from V onto U which
is an isomorphism between both spaces, then we have:
kJ (v)kU
inf = kJ −1 k−1 .
v∈V\{0V } kvkV

P ROOF. First we recall that in the case at hand the inverse J −1 of J is a well defined linear operator,
which is an isomorphism between V and U . Moreover from the Banach Inverse Theorem (see e.g. Yosida
(1980) [9]), it is also a bounded operator.
Now, since for every v ∈ V we can define u ∈ U by u = J (v), we may write
kJ (v)kU kukU
inf ≥ inf
v∈V\{0V } kvkV u∈U \{0U } kJ −1 (u)kV
" #−1 (1)
kJ −1 (u)kV
= sup = kJ −1 k−1 .
u∈U \{0U } kuk U

Actually, on the first row of (1) the equality applies, since, conversely, for every u ∈ U ∃v ∈ V given by
v = J −1 (u).

Let now K be a bounded linear operator from Y to X, Y and X being two Hilbert spaces.
We first recall a classical result of Functional Analysis (see e.g. Conway (1990) [5]), namely,
The adjoint operator K∗ from X to Y satisfies kK∗ k = kKk.

Next we prove
Proposition 2.2 The operator K satisfies

kKµkX ≥ κkµkY , ∀µ ∈ Y. (2)

for a suitable constant κ > 0, if and only if its adjoint operator K∗ is a surjection.
P ROOF. Using the usual notation R(A) and Ker(A) for the range and the kernel of a linear operator
A, first we establish that (2) holds if and only if R(K) is closed and Ker(K) = {0Y }.

2
Assume that (2) holds: In this case we obviously have Ker(K) = {0Y }. Moreover R(K) is closed due
to the following argument. Let {ξn }n be a sequence of R(K) converging to ξ ∈ X. In order to prove
that ξ ∈ R(K) we note that, by definition, there exists a sequence {µn }n in Y such that Kµn = ξn
for all n. Since {ξn }n is a Cauchy sequence of X, (2) implies that {µn }n is a Cauchy sequence of Y .
Therefore there exists µ ∈ Y such that µn converges to µ . By the continuity of K, ξn = Kµn converges
to Kµ, which is thus equal to ξ.
Assume that R(K) is closed and Ker(K) = {0Y }: In this case K is an isomorphism between Y and
R(K). Hence the inverse of K is also bounded operator and an isomorphism between R(K) and Y
and we may write, ∀µ ∈ Y , kµkY = kK−1 KµkY ≤ kK−1 kkKµkX . It follows that (2) holds with
κ = kK−1 k−1 .

Now assume that K∗ is a surjection: In this case R(K∗ ) = Y , whose orthogonal space is {0Y }, which
happens to be Ker(K). Moreover since R(K∗ ) is closed, by The Closed Range Theorem (cf. Yosida
(1980) [9]) R(K) is also closed. Therefore, from the first part of the proof, (2) holds true.
Now assume that (2) holds: In this case, the first part of the proof also tells us that R(K) is closed, and
hence so is R(K∗ ) again by The Closed Range Theorem. It also tells us that Ker(K) = {0Y }, and thus
the orthogonal space of Ker(K) is Y . Since the latter space is nothing but the closure of R(K∗ ), the
fact that it is closed implies that it is equal to Y , and the result follows.

Corollary 2.3 The largest possible constant κ for which (2) holds is kK−1 k−1 , K being viewed as a
linear operator from Y onto the complete space W := R(K).

P ROOF. We saw that K is a continuous isomorphism between the Hilbert spaces Y and W . In view
of this the result is a direct consequence of Lemma 2.1.

Corollary 2.4 Assuming that (2) holds, let Z ⊂ X be the orthogonal complement of Ker(K∗ ). Then
the following condition holds:
kK∗ ζkY ≥ κkζkX , ∀ζ ∈ Z. (3)

P ROOF. Clearly enough K∗ is a continuous isomorphism between Z and Y since from Proposition
2.2 it is a surjection. Therefore From Corollary 2.3 we have

kK∗ ζkY ≥ kK∗ −1 k−1 kζkX , ∀ζ ∈ Z. (4)

∗ ∗
On the other hand it is well known that K∗ −1 = K−1 and thus kK∗ −1 k = kK−1 k = kK−1 k. Since
from Corollary 2.3 κ ≤ kK−1 k−1 the result follows.

Before proceeding to the study of asymmetric mixed formulations, we recall a technical lemma due
to Dupire (1986) [7], whose proof we give in full here, in order to make the arguments that follow
self-contained.

Lemma 2.5 (cf. Dupire (1986) [7])

a2 , a3 ) ∈ (ℜ+ )3 and a4 := max[a1 x1 − a2 x2 , a3 x2 ]. Then ∀(x1 ; x2 ) ∈ (ℜ+ )2 we have
Let (a1 ,p
a4 ≥ a5 x21 + x22 , where
a1 a3
a5 := p 2 . (5)
a1 + (a2 + a3 )2

P ROOF. We have x2 ≤ a4 /a3 and x1 ≤ (a4 + a2 x2 )/a1 ≤ a4 (a2 + a3 )/a1 a3 . Thus x21 + x22 ≤
a24 {[(a2 + a3 )/(a1 a3 )]2 + 1/a23 } and the result follows.

3
3 Weak coercivity of asymmetric saddle-point bilinear forms
This paper’s abstract setting deals with bilinear forms defined upon two pairs of product spaces, namely,
W := P × U and Z := Q × V , where P , Q, U and V are four finite-dimensional spaces fulfilling
dimP = dimQ and dimU = dimV equipped with their respective norms and inner products denoted as
indicated in the previous section. The product spaces in turn are equipped with the norms ||(p; u)||W :=
[kpk2P + kuk2U ]1/2 ∀(p; u) ∈ W and ||(q; v)||Z := [kqk2Q + kvk2V ]1/2 ∀(q; v) ∈ Z.
More precisely, we consider bilinear forms given by

c((p; u), (q; v)) := a(p, q) + b(u, q) + d(p, v) ∀(p; u) ∈ W and ∀(q; v) ∈ Z, (6)

where a, b and d are continuous bilinear forms defined upon P × Q, P × V and U × Q, respectively.
Let B : Q → U and D : P → V be the continuous linear operators defined by

(u, Bq)U = b(u, q) ∀(u; q) ∈ U × Q; (Dp, v)V = d(p, v) ∀(p; v) ∈ P × V. (7)

Notice that the norm kBk is bounded above by kbk. Indeed, clearly enough, ∀q ∈ Q
(u, Bq)U b(u, q)
kBqkU = sup = sup ≤ kbkkqkQ .
u∈U \{0U } kukU u∈U \{0U } kukU

Similarly, the norm of the linear operator D is bounded above by kdk. Moreover the norm of the adjoint
operator D ∗ : V → P defined by

(p, D ∗ v)P = d(p, v) ∀p ∈ P and ∀v ∈ V

is also bounded above by kdk.

Next we endeavor to exhibit conditions on a, b and d such that c is continuous and weakly coercive.
First of all, the continuity of c means that there exists a constant kck, viewed as the norm of c, such that

c((p; u), (q; v)) ≤ kckk(p; u)kW k(q; v)kZ ∀(p; u) ∈ W and ∀(q; v) ∈ Z. (8)

Henceforth we assume that there are constants kak, kbk and kdk, viewed as the norms of a, b and d,
such that a(p, q) ≤ kakkpkP kqkQ ∀(p; q) ∈ P × Q, b(u, q) ≤ kbkkukU kqkQ ∀(u; q) ∈ U × Q and
d(p, v) ≤ kdkkpkP kvkV ∀(p; v) ∈ P × V . Under this assumption it is easy to see that c is continuous
with kck := max[kak, kbk, kdk].
On the other hand, according to Cuminato and Ruas (2015) [6], in the context of finite-dimensional
spaces, the weak coercivity of c is ensured solely if there exists a constant γ > 0 such that
c((p; u), (q; v))
∀(p; u) ∈ W sup ≥ γk(p; u)kW . (9)
(q;v)∈Z\{0Z } k(q; v)kZ

Next we prove the following fundamental result.

Theorem 3.1 Let R := Ker(D), S := Ker(B) and M and N be the orthogonal complements of R
and S in P and Q, respectively. We assume that dimR = dimS, and also that the forms a, b and d
satisfy the following inf-sup conditions:
There exist real numbers α > 0, β > 0 and δ > 0 such that

 ∀r ∈ R sup a(r, s) ≥ αkrkP
s∈S s.t. kskQ =1




∀u ∈ U sup b(u, q) ≥ βkukU (10)
 q∈Q s.t. kqkQ =1

 ∀v ∈ V sup d(p, v) ≥ δkvkV .


p∈P s.t. kpkP =1

Then there exists a constant γ > 0 such that (9) holds.

4
P ROOF. The second inf-sup condition in (10) implies that
∀u ∈ U ∃q ∈ Q with kqkQ = 1 s.t. b(u, q) ≥ βkukU . (11)
Analogously, the third inf-sup condition in (10) implies that
kD ∗ vkP := sup d(p, v) ≥ δkvkV ∀v ∈ V. (12)
p∈P s.t. kpkP =1
Now, setting K = D ∗ , Y = U and X = M , since [D ∗ ]∗ = D Corollary 2.4 allows us to assert that,
kDmkV := sup d(m, v) ≥ δkmkP ∀m ∈ M. (13)
v∈V s.t. kvkV =1
It is noticeable that (13) is equivalent to the condition
∀m ∈ M ∃v ∈ V with kvkV = 1 s.t. d(m, v) ≥ δkmkP . (14)
Now we proceed as follows:
Given a pair (p; u) ∈ W , we associate with it three pairs (qj ; vj ) ∈ Z, fulfilling k(qj ; vj )kZ = 1 for
j = 1, 2, 3 defined in accordance to (11), (14) and the first inf-sup condition in (10). More precisely, we
first define the following fields:
• r ∈ R is the orthogonal projection of p onto R;
• m ∈ M is the orthogonal projection of p onto M .
Then we set
1. v1 = 0V and q1 fulfills b(u, q1 ) ≥ βkukU with kq1 kQ = 1.
It is easy to see that c((p; u), (q1 ; v1 )) = a(p, q1 ) + b(u, q1 ) ≥ βkukU − kakkpkP ;
2. q2 = 0Q and v2 ∈ V is the function associated with m fulfilling (14), that is, d(m, v2 ) ≥
δkmkP with kv2 kV = 1.
Since p = r + m, it is also clear that c((p; u), (q2 ; v2 )) = d(m, v2 ) ≥ δkmkP ;
3. v3 = 0V , and q3 ∈ S fulfills a(r, q3 ) ≥ αkrkP with kq3 kQ = 1.
By a straightforward calculation we have c((p; u), (q3 ; v3 )) = a(r, q3 ) + a(m, q3 ) ≥ αkrkP −
kakkmkP .
Taking into account the definitions and properties listed above, we have
sup c((p; u), (q; v)) ≥ max c((p; u), (qi ; vi )) ≥ Φ(krkP , kmkP , kukU ), (15)
(q;v)∈Zs.t. ||(q;v)||Z =1 1≤i≤3
p
where Φ(x, y, z) := max[βz − kakt, δy, αx − kaky) with t = x2 + y 2 , x, y, z being non negative
real numbers.
Using Lemma 2.5 we can write, max[δy, αx − kaky ≥ νt with
p
ν = αδ/ (kak + δ)2 + α2 . (16)
Hence we can write Φ(x, y, z) ≥ max[βz − kakt, νt], which yields
p p
Φ(x, y, z) ≥ γ x2 + y 2 + z 2 with γ := βν/ (ν + kak)2 + β 2 , (17)
by applying again Lemma 2.5.
Plugging (17) into (15), we obtain
q
sup c((p; u), (q; v)) ≥ γ krk2P + kmk2P + kuk2U . (18)
(q;v)∈Z s.t. k(q;v)kZ =1
It follows from (18) and the Pythagorean Theorem that (9) holds true with γ defined by (17)-(16).

5
4 Error bounds and conclusion
It is easy to figure out that if one of the inf-sup conditions in (10) does not hold, the bilinear form
c cannot be weakly coercive. Hence, taking into account Theorem 3.1, if the uniform stability of a
mixed finite element approximate problem based on an asymmetric saddle point formulation is to be
established, the following issues must be overcome:
First of all it is necessary to prove the validity of the inf-sup conditions in (10) for the three bilinear
forms a, b and d involved in the larger bilinear form c defined upon two different finite-element product
spaces. This will guarantee the well-posedness of the approximate problem. Once these results are
obtained, uniform stability will be a consequence of the fact that the constants α, β and δ appearing
in (10) do not depend on discretization parameters such as mesh sizes. In order to clarify this point,
we address a few words on the error bound for the solution of a saddle point problem, viewed as an
approximation of another pair of unknowns (p̃; ũ) in the following framework:
Without loss of essential aspects, we consider that (p̃; ũ) belongs to a Hilbert product space of infinite
dimension Z̃ := Q̃ × Ṽ , where Q̃ (resp. Ṽ ) contains both P and Q (resp. U and V ). F being a bounded
linear functional on Z̃, the approximation of (p̃; ũ) is the solution (p; u) ∈ W of a problem of the form
c((p; u), (q; v)) = F ((q; v)) ∀(q; v) ∈ Z. (19)
Assuming for simplicity that a conforming method is being used, (p̃; ũ) in turn is the solution of a
saddle point problem of the same nature in space Z̃. This means that c is also defined and continuous
on Z̃ × Z̃ with the same norm kck and that the inner product and norm (·, ·)Z̃ and k · kZ̃ restricted to W
(resp. Z) reduces to (·, ·)W and k · kW (resp. (·, ·)Z and k · kZ ).
Resorting to the theoretical setting of Cuminato and Ruas (2015) [6], it holds
k(p̃; ũ) − (p; u)kZ̃ ≤ γ −1 kck inf k(p̃; ũ) − (r; y)kZ̃ . (20)
(r;y)∈W

Inequality (20) tells us that the existence of a lower bound for the weak coercivity constant γ of the
bilinear form c independent of the dimension of approximation spaces is essentially all that is needed to
attain the best possible error estimates for the method in use. According to (17)-(16), this property is a
consequence of the existence of uniform lower bounds for the constants α, β and δ.

References
[1] I. Babuška, Error bound for the finite element method, Numer. Mathematik, 16 (1971), 322–333.
[2] I. Babuška, The finite element method with Lagrangian multipliers, Numerische Mathematik, 20
(1973), 179–192.
[3] F. Brezzi, On the existence, uniqueness and approximation of saddle-point problems arising from
Lagrange multipliers, RAIRO Analyse Numérique, 8-2 (1974), 129-151.
[4] F. Brezzi and M. Fortin, Mixed and Hybrid Finite Element Methods, Springer-Verlag, 1991.
[5] J. B. Conway, A Course in Functional Analysis, Springer-Verlag, 1990.
[6] J.A. Cuminato and V. Ruas, Unification of distance inequalities for linear variational problems,
Computational and Applied Mathematics, 34 (2015), 1009-1033.
[7] B. Dupire, Problemas Variacionais Lineares, Sua Aproximação e Formulações Mistas, doctoral
thesis, Pontifı́cia Universidade Católica do Rio de Janeiro, Brazil, 1985.
[8] R.A. Nicolaides, Existence, Uniqueness and Approximation for Generalized Saddle Point Prob-
lems, SIAM Journal on Numerical Analysis, 2-19 (1982), 349-357.
[9] K. Yosida, Functional Analysis, Fundamental Principles of Mathematical Sciences, vol. 123 (6th
ed.), Springer-Verlag, Berlin, 1980.