Brouwer and Galerkin

S.Kesavan
Adjunct Faculty, Indian Institute of Technology, Madras
email: [email protected]

Abstract
We look at an equivalent form of the Brouwer fixed point theorem and explain
how it can be used in the Galerkin method for solving equations in Hilbert
spaces.

1 The Brouwer fixed point theorem

The two well-known fixed point theorems are the contraction mapping theo-
rem and the Brouwer fixed point theorem. While the former is fairly easy to
prove, the latter needs a sophisticated topological tool, viz. the topological
degree.

Let Ω ⊂ RN be a bounded domain. We denote its closure by Ω and its

boundary by ∂Ω. If f : Ω → RN is a given mapping, a vector b ∈ RN is said
to be a boundary value if there exists x ∈ ∂Ω such that f (x) = b.

Let Ω be as stated above and let f : Ω → RN be a continuous mapping.

Let b ∈ RN . If b is not a boundary value, then the (Brouwer) degree,
denoted d(f, Ω, b) can be defined (cf. Kesavan [2]. Amongst the properties
of the degree, the following are most useful.
• If d(f, Ω, b) 6= 0, then there exists x ∈ Ω such that f (x) = b. Thus, this
helps in proving the existence of a solution to the given equation.
• (Homotopy invariance) Let H : Ω × [0, 1] → RN be a continuous map-
ping such that for every t ∈ [0, 1], the point b ∈ RN is not a boundary
value of the mapping H(·, t), then
d(H(·, 0), Ω, b) = d(H(·, 1), Ω, b).

• If I stands for the identity mapping in RN , then


1, if b ∈ Ω,
d(I, Ω, b) = ,
0, if b ∈ RN \Ω.

1
Using these properties, the Brouwer fixed point theorem is proved in two
steps, as indicated below.
Let B denote the open unit ball, centered at the origin, in RN and let S N −1
be its boundary, i.e. the unit sphere.
Step 1. First we prove the ‘no retraction theorem’: there is no continuous
map from from B onto S N −1 which is the identity map when resticted to
S N −1 .
Step 2: (Brouwer fixed point theorem): Let f be a continuous map of B into
itself. Then f has a fixed point.

To see this, we assume the contrary. Then, for each x ∈ B, the points x
and f (x) are distinct and so the line segment joining them is well defined.
We produce the line segment starting from f (x) and ending at x to meet the
boundary S N −1 and call this point, say, P (x). Then one shows that the map-
ping x 7→ P (x) is a continuous map (this needs checking) which is, obviously,
the identity map when restricted to S N −1 , thereby arriving at a contradiction.

In one-dimension, the intermediate value theorem states that if f : [−1, 1] →

R is a continuous function such that f (−1) and f (1) have opposite signs, then
f has to vanish in the interval (0, 1). If, now, f : [−1, 1] → [−1, 1] is contin-
uous, we can apply the intermediate value theorem to the function f (x) − x
and immediately deduce Brouwer’s theorem in this case.

We now extend this idea to higher dimensions and deduce Brouwer’s

theorem. We denote by (·, ·), the usual euclidean inner-product, and by | · |,
the euclidean norm in RN .
Proposition 1.1 Let g : RN → RN be continuous. Let R > 0. Assume that
(g(x), x) ≥ 0, for all x ∈ RN such that |x| = R. Then there exists x0 ∈ RN
with |x0 | ≤ R such that g(x0 ) = 0.
Proof: Assume that g(x) does not vanish for |x| = R, for, otherwise, we are
done. Define, for |x| ≤ R and for t ∈ [0, 1],
H(x, t) = tg(x) + (1 − t)x.
Then, clearly, H(x, t) does not vanish for |x| = R when t = 0 and when
t = 1. Let t ∈ (0, 1). If |x| = R and if H(x, t) = 0, we hve g(x) = − 1−t
t
x and
so
1−t 2
0 ≤ (g(x), x) = − R < 0,
t
2
which is a contradiction. Thus H(·, t) does not vanish on the boundary of
the ball BR , centred at the origin and of radius R, for all t ∈ [0, 1]. Hence,
by the properties of the degree listed earlier, we get

d(g, BR , 0) = d(I, BR , 0) = 1,

and so, again by a property of the degree, there exists x0 ∈ BR such that
g(x0 ) = 0. 

Remark 1.1 We have not only proved the existence of a solution to the
equation g(x) = 0, but we also have an estimate for the norm of the solution.


Remark 1.2 It is clear that the proposition is true if we have the condition
(g(x), x) ≤ 0 for all x ∈ RN such that |x| = R. 

Corollary 1.1 (Brouwer fixed point theorem) Let BR be the open ball in RN ,
centred at the origin and of radius R > 0. Let f : BR → BR be continuous.
Then f has a fixed point.

Proof: Set g(x) = x − f (x). Then, for |x| = R, by the Cauchy-Schwarz

inequality, we have

(g(x), x) = R2 − (f (x), x) ≥ 0,

since |f (x)| ≤ R. The result now follows from the preceding proposition. 

In fact, all the three results - the Brouwer fixed point theorem the above
proposition and the no retraction theorem - are all equivalent, i.e. each
implies the other two.

Proposition 1.2 Let R > 0 and let BR denote the open ball in RN centred
at the origin and of radius R. Let SR denote its boundary, the sphere with
centre at the origin and of radius R. The following statements are equivalent.
(i) Let f : BR → BR be continuous. Then f has a fixed point.
(ii) Let g : RN → RN be such that (g(x), x) ≥ 0 for all x ∈ SR . Then there
exists x0 ∈ BR such that g(x0 ) = 0.
(iii) There is no continuous mapping from BR onto SR which is the identity
when restricted to SR .

3
Proof: (i) ⇒ (ii): Assume that g does not vanish in BR . Then the mapping
g(x)
f (x) = −R
|g(x)|

is well-defined, continuous and maps BR into itself. In fact, the image is

contained in SR . Thus, by (i), there exists a fixed point, say, x0 , of f . Then
it follows that x0 ∈ SR . But then,
(g(x0 ), x0 )
0 < R2 = |x0 |2 = (f (x0 ), x0 ) = −R ≤ 0,
|g(x0 )|
a contradiction.

(ii) ⇒ (iii): If g : BR → SR is a continuous mapping which, when restricted

to SR , is the identity, we have, for all x ∈ SR ,

(g(x), x) = |x|2 ≥ 0.

Thus g must vanish somewhere, which is impossible since it takes values only
in SR .

(iii) ⇒ (i): This is the standard proof of Brouwer’s theorem outlined at the
beginning of this section. See Kesavan [2] for details. 

2 The Galerkin method

The Galerkin method is a useful approximation method to solve linear and
nonlinear equations in Hilbert spaces. In this section, we will illustrate the
method by using it to prove the famous Lax-Milgram lemma. In the next
section, we will illustrate its use in solving a nonlinear equation and we will
use the equivalent form of Brouwer’s theorem proved in the previous section.

Let A be an N × N matrix with real entries. It is said to be positive

definite if, for every x ∈ RN , x 6= 0, we have

(Ax, x) > 0.

Remark 2.1 If A is a matrix with complex entries and if we still denote the
usual inner-product in CN by (·, ·), then (Az, z) = 0 for all z ∈ CN implies

4
that A = 0. Similarly if (Az, z) ≥ 0 for all z ∈ CN , it follows that A = A∗ ,
i.e. A is self-adjoint. Neither of these results is true in the real case as seen
from the matrix  
0 1
.
−1 0
Since the unit ball is compact, we then deduce that, if A is positive definite,
there exists α > 0 such that
(Ax, x) ≥ α|x|2 , for every x ∈ RN .
It also follows that the linear map associated to A is injective and so A is
invertible. The Lax-Milgram lemma is an infinite-dimensional version of this
result.

Let H be a real Hilbert space whose norm is denoted by k · k and the

inner-product by (·, ·) and let a : H × H → R be a bilinear form. We say
that a is continuous if there exists a constant M > 0 such that
|a(x, y)| ≤ M kxk kyk,
for every x and y in H. The bilinear form is said to be elliptic if there exists
a constant α > 0 such that
a(x, x) ≥ αkxk2 ,
for every x ∈ H.
Theorem 2.1 (Lax-Milgram Lemma) Let H be a real Hilbert space which is
separable and let a : H × H → R be a continuous and elliptic bilinear form.
Given any f ∈ H, there exists a unique vector u ∈ H such that
a(u, v) = (f, v),
for every v ∈ H.
Proof: Since H is separable, we can find an orthonormal basis {wi }∞ i=1 . Let
Wm be the span of {w1 , · · · , wm }, where m is a positive integer.

Step 1. Let m be a positive integer. Using the notation established above, we

consider the following problem: find um ∈ Wm such that, for every v ∈ Wm ,
a(um , v) = (f, v). (2.1)

5
By linearity in the second variable of a, it is enough for us to find um ∈ Wm
such that
a(um , wi ) = (f, wi ), for 1 ≤ i ≤ m.
Since we can write m
X
um = um
j wj ,
j=1

we get the system of m linear equations in the m unknowns {um m

1 , · · · , um },

m
X
a(wj , wi )um
j = (f, wi ), for 1 ≤ i ≤ m.
j=1

The matrix A = (aij ) where aij = a(wj , wi ) is positive definite. For,

m X
X m
(Ax, x) = a(wj , wi )xi xj = a(w, w) ≥ αkwk2 ,
i=1 j=1

Pm
where w = i=1 xi wi . Hence, there exists a unique solution um to (2.1).

Step 2. (a priori estimate) Using the ellipticity of the bilinear form a, we get

αkum k2 ≤ a(um , um ) = (f, um ) ≤ kf kkum k,

whence we get that for all positive integers m,

kf k
kum k ≤ .
α
Since {um } is uniformly bounded in the Hilbert space H, we can extract a
weakly convergent subsequence, say, {unk }. Let the weak limit be u. Given
v ∈ H, let
Xm
(m)
v = vi w i ,
i=1

where, vi = (v, wi ). Then v (m) → v in H, i.e. kv (m) − vk → 0. Since

v (mk ) ∈ Wmk , we get from (2.1) that

a(umk , v (mk ) ) = (f, v (mk ) ).

6
The right-hand side of this equation obviously converges to (f, v). By the
continuity of the bilinear form, for any fixed z ∈ H, the mapping x 7→ a(x, z)
defines a continuous linear functional on H. Hence a(umk , z) → a(u, z), by
the definition of weak convergence. Now we have

a(umk , v (mk ) ) = a(umk , v) + a(umk , v (mk ) − v).

The first term on the right-hand side converges to a(u, v). Since

M
|a(umk , v (mk ) − v)| ≤ M kumk kkv (mk ) − vk ≤ kf kkv (mk ) − vk,
α
it follows that the second term converges to zero. Thus we get that, for all
v ∈ H,
a(u, v) = (f, v), (2.2)
whch proves the existence of a solution.

Step 3. (Uniqueness) If we have two solutions u1 and u2 , then, for all v ∈ H,

we get that a(u1 − u2 , v) = 0. Set v = u1 − u2 . Thus,

αku1 − u2 k2 ≤ a(u1 − u2 , u1 − u2 ) = 0,

which shows that u1 = u2 . This completes the proof. 

Remark 2.1 In fact {umk } converges to u in norm. For,

αku − umk k2 ≤ a(u − umk , u − umk )

= a(u − umk , u − u(mk ) ) + a(u − umk , u(mk ) − umk ).

The second term on the right-hand side vanishes since u(mk ) − umk ∈ Wmk ,
in view of (2.1) and (2.2). Thus,

M
ku − umk k ≤ ku − u(mk ) k,
α
and the conclusion follows since the right-hand side of the above inequality
tends to zero. 

Remark 2.2 Given any subsequence of {um }, there is a further subsequence

which is weakly convergent. The weak limit satisfies (2.2) which has a unique

7
solution. Therefore, it follows that, in fact, the entire sequence {um } con-
verges to the unique solution u of (2.2) in norm. 

For a proof of this theorem, through the context of variational inequal-

ities, see Kesavan [1]. Incidentally, the proof of existence to a variational
inequality uses the contraction mapping theorem. Here, we have proved the
Lax-Milgram lemma, directly, in the case of a separable Hilbert space. This is
not a serious restriction, since all Hilbert spaces which occur in applications
are generally separable.

The Lax-Milgram lemma is the corner stone of the exitence theory for
weak solutions of elliptic boundary value problems. For several examples of
its use in this direction, see Kesavan [1].

Another example of the Galerkin method to solve a linear problem, with-

out the use of the Lax-Milgram lemma, is the solution of the Schrödinger
equation. See Kesavan[1] for details.

3 A nonlinear equation
We will now illustrate the application of the Galerkin method to solve a
nonlinear equation (cf. Kesavan [2]). Let H be a separable Hilbert space
whose norm and inner-product are denoted by k · k and (·, ·), respectively. A
mapping A : H → H is said to be monotone if, for every x, y ∈ H, we have
(Ax − Ay, x − y) ≥ 0.
We now establish some notation. Let {wi }∞ i=1 be an orthonormal basis of
H. We denote by Wm , the span of {w1 , · · · , wm }, which is an m-dimensional
subspace of H. Let v ∈ Wm . Then
m
X
v = (v, wi )wi .
i=1

Set vi = (v, wi ). Consider the vector v ∈ Rm given by v = (v1 , · · · , vm ).

Thus we have a linear bijection between Wm and Rm . In fact, since we are
working with an orthonormal set, it is immediate to see that this bijection
is an isometry, i.e.
kvk2 = |v|2 .

8
Theorem 3.1 Let Hbe separable Hilbert space. Let A : H → H be a con-
tinuous map which is monotone and which maps bounded sets onto bounded
sets. Then, given any f ∈ H, there exists a unique solution u ∈ H of the
equation
u + Au = f. (3.3)
Further,
kuk ≤ kA(0) − f k. (3.4)

Proof: Step 1. (Uniqueness) If u1 and u2 were two solutions of (3.3), we

have
u1 − u2 + Au1 − Au2 = 0.
Thus,
ku1 − u2 k2 + (Au1 − Au2 , u1 − u2 ) = 0,
and hence, by virtue of the monotonicity of A, we have u1 − u2 = 0.

Step 2. (a priori estimate) If u ∈ H is a solution of (3.3), then

kuk2 + (Au − A0, u) = (f − A0, u),

and the estimate (3.4) follows, again thanks to the monotonicity of A.

Step 3. Let {wi }∞

i=1 be an orthonormal basis of H. Let Wm be as defined
above and for v ∈ Wm we associate the vector v ∈ Rm , as explained above.
Define T : Rm → Rm by

(T v)i = (v, wi ) + (Av, wi ) − (f, wi ), 1 ≤ i ≤ m.

Then, T is continuous and (denoting the usual inner product in Rm by (·, ·)m )
we have
(T v, v)m = kvk2 + (Av, v) − (f, v)
= kvk2 + (Av − A0, v) − (f − A0, v)
≥ kvk2 − kA0 − f kkvk.
Setting R = kA0 − f k, we have that (T v, v)m ≥ 0 for all |v| = kvk = R.
Hence, by Proposition 1.1, there exists a um ∈ Rm such that |um | ≤ R and
T um = 0. Thus, um ∈ Wm satisfies

(um , wi ) + (Aum , wi ) = (f, wi ), 1 ≤ i ≤ m,

9
i.e.
(um , v) + (Aum , v) = (f, v) for all v ∈ Wm
and kum k ≤ kA0 − f k.

Step 4. Then, there exists a weakly convergent subsequence of {um }. As

{Aun } is bounded (since A maps bounded sets into bounded sets), we can
also assume (after taking a further subsequence if necessary) that unk *
u, Aunk * χ weakly in H, for a subsequence indexed by nk .

Step 5. Given v ∈ H, the sequence {v (n) } defined, as before, by

n
X
(n)
v = (v, wi )wi ,
i=1

is such that v (n) ∈ Wn for each n and kv (n) − vk → 0. Now,

(unk , v (nk ) ) + (Aunk , v (nk ) ) = (f, v (nk ) ). (3.5)

Passing to the limit in (3.5), we get

(u, v) + (χ, v) = (f, v) for all v ∈ H. (3.6)

Step 6. By the monotonicity of A, we have for any v ∈ H,

0 ≤ Xnk = (Aunk − Av, unk − v)

= (Aunk , unk ) − (Aunk , v) − (Av, unk − v)
= (f, unk ) − kunk k2 − (Aunk , v) − (Av, unk − v),

using (3.5). Thus, X = lim supk→∞ Xnk ≥ 0 and

X = (f, u) − lim inf k→∞ kunk k2 − (χ, v) − (Av, u − v)

≤ (f, u) − kuk2 − (f, v) + (u, v) − (Av, u − v),

using (3.6). Thus,

(f − u − Au, u − v) + (Au − Av, u − v) ≥ 0. (3.7)

Let λ > 0 and w ∈ H. Set v = u − λw in (3.7) to get

(f − u − Au, w) + (Au − A(u − λw), w) ≥ 0.

10
As λ → 0, by the continuity of A, the second term on the left-hand side
tends to zero. Thus (f − u − Au, w) ≥ 0 for all w ∈ H and, by considering
−w in place of w, we conclude that u satisfies (3.3). 

For an example of the Galerkin method applied to a semilinear elliptic

boundary value problem, see Kesavan [1]. The method is always the same.
We study the problem in a finite dimensional subspace. The existence of a
solution in that space and the uniform bound for that solution follow from
Brouwer’s theorem (Proposition 1.1). The sequence of approximate solutions
will have a weakly convergent subsequence and the weak limit will turn out
to be a solution of the original problem. See also Lions [3] for more examples
of this technique.

Thus, while the only serious application of the no retraction theorem

seems to be the proof of Brouwer’s theorem, Proposition 1.1 has many ap-
plications, via the Galerkin method, in the theory of existence of solutions
to linear and nonlinear boundary value problems.

References
[1] Kesavan, S. Topics in Functional Analysis and Applicatios, Third Edi-
tion, New Age International Publishers, 2019.

[2] Kesavan, S. Nonlinear Functional Analysis: A First Course, TRIM

Series 28, Hindustan Book Agency, 2004.

[3] Lions, J. L. Quelques Mt́hodes de Résolution des Problèmes aux Lim-

ites Nonlinéaires, Dunod Gauthier-Villars, 1969.

11