Notes on weak convergence (MAT4380 - Spring 2006)

Kenneth H. Karlsen (CMA)

February 2, 2006

1 Weak convergence
In what follows, let U denote an open, bounded, smooth subset of RN with N ≥ 2. We
assume 1 ≤ p < ∞ and let p0 be the conjugate exponent, i.e.,
1 1
+ 0 =1
p p
(p0 := ∞ when p = 1).
A sequence {un }n≥1 ⊂ Lp (U ) converges weakly to u ∈ Lp (U ), in which case we write
un * u in Lp (U ),
if Z Z
0
un v dx → uv dx, ∀v ∈ Lp (U ).
U U
When p = ∞, we say that a sequence {un }n≥1 ⊂ Lp (U ) converges weakly - ? to
u ∈ L∞ (U ), and we write
?
un * u in L∞ (U ),
if Z Z
un v dx → uv dx, ∀v ∈ L1 (U ).
U U
We remark that when Ω is bounded the weak - ? convergence of un in L∞ (Ω) to some
u ∈ L∞ (Ω) implies weak convergence of un to u in any Lp (Ω), 1 ≤ p < ∞.
Theorem 1.1 (boundedness of weakly converging sequences). Suppose 1 ≤ p < ∞
and
?
un * u in Lp (Ω) (* in L∞ (Ω) if p = ∞).
Then
un is bounded in Lp (Ω)
and
kukLp (Ω) ≤ lim inf kun kLp (Ω)
n↑∞

1
 We have the following compactness theorem:
Theorem 1.2 (Weak convergence in Lp ). Suppose 1 < p < ∞ and the sequence
{un }n≥1 is bounded in Lp (U ). Then there is a subsequence, still denoted by {un }n≥1 , and
a function u ∈ Lp (U ) such that

un * u in Lp (U ).
?
If p = ∞, the result still holds with * replaced by *.
Theorem 1.2 is false for p = 1 since L1 (U ) is not the dual of L∞ (U ). But a good
substitute result exists by regarding L1 (U ) as a subset1 of the space of (signed) Radon
measures on U with finite mass, a space which we denote by M(U ), and using the weak -
? topology on M(U ). Let Cc (U ) the denote the space of continuous functions on U with
compact support. If µ ∈ M(U ), then
Z


µ, v = v dµ, ∀v ∈ Cc (U ).
U

Recall that µ ∈ M(U ) if and only if



 µ, v  ≤ C kvk ∞ , ∀v ∈ C0 (U ).
L (U )

We define n
 o
kµkM(U ) = sup  µ, v  : v ∈ Cc (U ), kvkL∞ (U ) ≤ 1 .


The space M(U ), k·kM(U ) is a Banach space and it is isometrically isomorphic to the


dual space of Cc (U ), k·kL∞ (U ) .
A sequence {µn }n≥1 ⊂ M(U ) converges weakly - ? to µ ∈ M(U ), in which case we
write
?
µn * µ in M(U ),
if Z Z
v dµn → v dµ, ∀v ∈ Cc (U ).
U U

Theorem 1.3. Suppose

?
µn * µ in M(Ω).
Then
lim sup µn (K) ≤ µ(K),
n↑∞

1
We recall that each u ∈ L1 (U ) defines a linear functional on Cc (U ) via
Z
v 7→ uv dx, v ∈ Cc (U ).
U

2
for each compact set K ⊂ Ω, and
µ(O) ≤ lim inf µn (O),
n↑∞

for each open set O ⊂ Ω.

We have the following compactness theorem for measures:
Theorem 1.4 (Weak compactness in M(U )). Suppose the sequence {µn }n≥1 is bounded
in M(U ). Then there is a subsequence, still denoted by {µn }n≥1 , and a measure µ ∈ M(U )
such that
?
µn * µ in M(U ).
Theorem 1.5 (Characterizations of weak convergence in Lp ). Let un : Ω → R be a
sequence in Lp (Ω) and u ∈ Lp (Ω), 1 < p < ∞. Suppose the sequence un is equibounded in
Lp (Ω) Then the following statements are equivalent:
1. un * u in Lp (Ω).
?
2. un * u in M(Ω).
3. un → u in D0 (Ω).
4. For any Borel set E ⊂ Ω, |E| > 0,
Z Z
1 1
(un )E := un dx → (u)E := u dx.
|E| E |E| E
5. For any cube Q ⊂ Ω, |Q| > 0,
Z Z
1 1
(un )E := un dx → (u)E := u dx.
|E| E |E| E

Remark 1.1. Similar equivalences hold also if p = ∞, replacing in (1) weak convergence
with weak - ? convergence in L∞ (Ω).
Remark 1.2. Condition (5) expresses the intuitive idea of weak convergence as convergence
of mean values.
The next lemma is simple but quite useful in a number of situations.
Lemma 1.1 (products of weak-strong converging sequences). Let 1 < p < ∞,
un : Ω → R be a sequence in Lp (Ω), and u ∈ Lp (Ω). Let vn : Ω → R be a sequence in
0 p
Lp (Ω), p1 + p10 = 1 (or p0 = p−1 ). Suppose

un * u in Lp (Ω),
0
vn → u in Lp (Ω).
Then
un vn * uv in L1 (Ω).

3
2 Some typical behaviors of weakly converging sequences
Relevant for our study of nonlinear partial differential equations and calculus of variation,
we will illustrate some typical behaviors of sequences that converge weakly but not strongly.

2.1 Oscillations
Sequences of rapidly oscillating functions provide examples of weakly – but not strongly –
converging sequences.
Letting
un (x) = sin(nx), x ∈ (0, 2π), n = 1, 2, . . . ,
one can easily check that
?
un * u := 0 in Lp (0, 2π) ∀p ≥ 1 (* in L∞ (0, 2π) if p = ∞),

but
kun kLp (0,2π) = C(p) > 0.
Hence un does not converge strongly in Lp (0, 2π) for any 1 ≤ p ≤ ∞. Furthermore, we
have
kukLp (0,2π) < lim inf kun kLp (0,2π) .
n↑∞

Recall that if un * u in Lp , then by the weak lower semicontinuity of the Lp norms we

have always
kukLp (0,2π) ≤ lim inf kun kLp (0,2π) . (1)
n↑∞

If un → u in Lp , then we have instead (trivially) equality in (1), but be aware that we can
have this equality under mere weak convergence, as the the next examples shows.
Let
un (x) = 1 + sin(nx), x ∈ (0, 2π), n = 1, 2, . . . ,
then one can easily check that
?
un * u := 1 in Lp (0, 2π) ∀p ≥ 1 (* in L∞ (0, 2π) if p = ∞)

and Z 2π Z 2π
|un | dx = |u| dx = 2π, ∀n,



Z0 2π 0 Z
2π
|un − u| dx = |sin(nx)| dx = 4, ∀n,



0 0
so that
un converges weakly – but strongly – to u in L1 (0, 2π).

4
On the other hand, we have
Z 2π
|u|p dx < lim inf |un |p dx, ∀p > 1,
0 n↑∞

since otherwise, passing to a subsequence if necessary, we would have for all p > 1

un * u in Lp (0, 2π) and kun kLp (0,2π) → kukLp (0,2π) ,

which implies (via Brezis-Lieb’s refinement of Fatous’s lemma)

un → u in Lp (0, 2π).

More generally, the following theorem about weak limits of rapidly oscillating periodic
functions is well known:

Theorem 2.1. Let 1 ≤ p ≤ ∞ and u be a Y periodic function in Lp (Y ). For n = 1, 2, . . . ,

set
un (x) := u(nx).
Then, if 1 ≤ p < ∞, as n ↑ ∞,
Z
1
un * u(y) dy in Lp (O) for any bounded open set O ⊂ RN .
|Y | Y

If p = ∞, as n ↑ ∞, Z
? 1
un * u(y) dy in L∞ (RN ).
|Y | Y

2.2 Concentrations
For example, Dirac masses arise by concentration. But concentration phenomena arise also
in the context of functions.
Define un : (−1, 1) → R by
(
n, if x ∈ [0, 1/n],
un (x) =
0, otherwise

Then once can easily check that

?
un * δ0 in M(−1, 1). (2)
?
One can also smooth out un and (2) still holds. Observe here also that while un * 0 in
M(0, 1), we do not have that un converges weakly to 0 in L1 (0, 1) (in view of Dunford-Pettis
there is a lack of equiintegrability).

5
 Here is another example. Define un : (−1, 1) → R by

−n, if x ∈ (−1/n, 0),

un (x) = n, if x ∈ (0, +1/n),

0, otherwise.


Then once can check that

?
un * u := 0 in M(−1, 1),
but no subsequence of un converges weakly in L1 (−1, 1) (again there is a lack of equiinte-
grability). Indeed, for ϕ = χ(0,1) ∈ L∞ (−1, 1), we have
Z 1 Z 1 Z 1
un ϕ dx = un dx = 1, while uϕ dx = 0.
−1 0 −1

In other words, the sequence is not precompact in weak topology of L1 (−1, 1), although
Z 1
|un | dx = 2 ∀n,
−1

i.e., the sequence un is uniformly bounded (equibounded) in L1 (−1, 1). Observe also that
there is a “loss of energy”:

0 = kukL1 (−1,1) < lim inf kun kL1 (−1,1) = 2.

n↑∞

Here the energy disappears as a measure. Finally, note that kun kLp (−1,1) → ∞ for any
p > 1.
Of course, concentration phenomena show up also in Lp spaces with p > 1. For p > 1,
define un : (−1, 1) → R by
( 1
n p , if x ∈ [0, 1/n],
un (x) =
0, otherwise.

Then we have
kun kLp (−1,1) = 1 for all n and un * 0 in Lp (−1, 1),
but un does not converge strongly in Lp (−1, 1) since the sequence kun kLp (−1,1) concentrates.
But we have
un * 0 in Lq (−1, 1) for any 1 ≤ q < p.
Indeed, we have Z 1 q
−1 q
|un |q dx = n p → as < 1.
−1 p
Hence the sequence the sequence kun kLp (−1,1) does not concentrate.

6
2.3 Nonlinearity destroys weak convergence
If un converges weakly – but not strongly – to u, then for a generic nonlinear function the
sequence f (un ) does NOT converge weakly to f (u). Here is an example. Let

un (x) = sin(nx), x ∈ (0, 2π), f (ξ) = ξ 2 for ξ ∈ R.

Then
?
un * u := 0 in Lp (0, 2π) ∀p ≥ 1 (* in L∞ (0, 2π) if p = ∞),
but
1 1
f (un ) = sin2 (nx) =
(1 − cos(2nx)) * 6= f (u) = 0.
2 2
Observe here also that f (un ) is uniformly bounded (equibounded) in L∞ (0, 2π).
Another example is
f (ξ) = max(0, ξ), ξ ∈ R.
Then
1
f (un ) * 6= f (u) = 0 in L1 (0, 2π).
π
If
f (ξ) = |ξ| , ξ ∈ R,
then
2
6= f (u) = 0 in L1 (0, 2π).
f (un ) *
π
Here is a final example. Let f be any continuous (nonlinear) function and select two
number a < b such that  
a+b f (a) + f (b)
f 6= .
2 2
Define a sequence of functions un : (0, 1) → R by
(
a, if x ∈ i/n, (i + 12 )/n , i =, 0, . . . , n − 1,
 
un (x) =
b, otherwise.

Then one can easily check that

? a+b
un * u := in L∞ (0, 1) and hence weakly in any Lp (0, 1), 1 ≤ p < ∞,
2
but  
? f (a) + f (b) a+b
f (un ) * f := 6= f (u) = f .
2 2

7
3 Weak compactness in L1
Let us now turn to the more delicate issue of weak compactness in L1 . We start with a
definition.

Definition 3.1 (equiintegrability). Let Ω ⊂ RN and U ⊂ L1 (Ω) be a family of integrable

functions. We say that U is an equiintegrable family if the following two conditions hold:

1. For any ε > 0 there exists a measurable set A with |A| < ∞ such that
Z
|u| < ε,
Ω\A

for all u ∈ U. (This condition is trivially satisfied if |Ω| < ∞, since then we can take
A = Ω.)

2. For any ε > 0 there exists δ > 0 such that for every measurable set E with |E| < δ
there holds Z
|u| < ε
E
for all u ∈ U.

We have the following three equivalent formulations of the equiintegrability property.

Lemma 3.1. Let Ω ⊂ RN and U ⊂ L1 (Ω) be a family of integrable functions.

1. Then U is equiintegrable if and only if for any sequence of measurable sets En with
En ↓ ∅ there holds Z
lim sup |u| dx = 0.
n↑∞ u∈U En

2. If |Ω| < ∞ and U is bounded in L1 (Ω), then U is equiintegrable if and only if

 Z 
1
U ⊂ u ∈ L (Ω) : Ψ(|u|) dx ≤ 1 ,
Ω

for some increasing function Ψ : [0, ∞) → [0, ∞] satisfying

Ψ(ξ)
lim → ∞.
ξ↑∞ ξ

3. If |Ω| < ∞ and U is bounded in L1 (Ω), then U is equiintegrable if and only if

Z
lim sup |u| dx = 0.
ξ↑∞ u∈U {|u|>ξ}

8
 Here is a restatement of (2). If |Ω| < ∞ and U is bounded in L1 (Ω), then the family U
is equiintegrable if and only if
Z
sup Ψ(|u|) dx < ∞,
u∈U Ω

for some increasing function Ψ : [0, ∞) → [0, ∞] satisfying

Ψ(ξ)
lim → ∞.
ξ↑∞ ξ
Let us give an example illustrating how to use this restatement in the context of sequences
of functions. Let |Ω| < ∞ and un : Ω → R be a sequence of functions that are equibounded
(uniformly bounded) in L1 (Ω) , i.e.,
Z
|un | dx ≤ C, ∀n.
Ω

Then a sufficient condition for the sequence un to be equiintegrable is that there exists a
constant C, independent of n, such that
Z
|un |1+θ dx ≤ C,
Ω

for some θ > 0.

The next theorem gives a necessary and sufficient condition (namely equiboundedness
and equiintegrability) for compactness with respect to the weak convergence in L1 .

Theorem 3.1 (Dunford-Pettis). Let un : Ω → R be a sequence in L1 (Ω). Suppose

1. the sequence un is equibounded in L1 (Ω), i.e.,

sup kun kL1 (Ω) < ∞,

n

2. the sequence un is equiintegrable.

Then there exists a subsequence of un that converges weakly in L1 (Ω). Conversely, if un

converges weakly in L1 (Ω), then 1. and 2. hold.

The following theorem is the L1 analog of Theorem 1.5.

Theorem 3.2 (Characterizations of weak convergence in L1 ). Let un : Ω → R be a

sequence in L1 (Ω) and u ∈ L1 (Ω). Suppose the sequence un is equibounded in L1 (Ω) and
equiintegrable. Then the following statements are equivalent:

1. un * u in L1 (Ω).

9
 ?
2. un * u in M(Ω).

3. un → u in D0 (Ω).

4. For any Borel set E ⊂ Ω, |E| > 0,

(un )E → (u)E .

5. For any cube Q ⊂ Ω, |Q| > 0,

(un )Q → (u)Q .

The next lemma is simple but still very useful.

Lemma 3.2 (products of weak-strong converging sequences). Let un , u, vn , v : Ω →
R be measurable functions.
1. If un → u a.e. in Ω, kun kL∞ (Ω) ≤ C for all n, and vn * v in L1 (Ω), then

un vn * uv in L1 (Ω).

2. If un → u in L1 (Ω), kun kL∞ (Ω) ≤ C for all n, and vn * v in L1 (Ω), then

un vn * uv in L1 (Ω).

The next lemma is useful in many applications.

Lemma 3.3 (Vitali). Let Ω ⊂ RN be bounded and un : Ω → R be a sequence in L1 (Ω).
Suppose
1. lim un (x) exists and is finite for a.e. x ∈ Ω,
n↑∞

2. the sequence un is equiintegrable.

Then Z Z
lim un (x) dx = lim un (x) dx.
n↑∞ Ω Ω n↑∞

A typical application of Vitali’s lemma is provided by the next simple lemma.

Lemma 3.4. Let Ω ⊂ RN be bounded and un : Ω → R be a sequence in L1 (Ω). Suppose
1. un → u a.e. in Ω,

2. the sequence un is bounded in Lp (Ω) for some p > 1.

Then
un → u in Lr (Ω) for all 1 ≤ r < p.

10
Proof. By a previous theorem,

u ∈ Lp (Ω) (and also un * u in Lp (Ω)).

Define
vn = |u − un |r , r < p.
Then
vn → 0 a.e. in Ω
and p
vn is bounded in L r (Ω) and p/r > 1.
Hence the sequence vn is equiintegrable, so that by Vitali’s lemma
Z
lim vn dx = 0,
n↑∞

that is un → u in Lr (Ω).

The next lemma recall the well known fact that convex (concave) functions are lower
(upper) semicontinuous with respect to the weak convergence.

Lemma 3.5 (weak lower semicontinuity of convex functions). If F : R → R is

convex and
un * u in L1 ,
then Z Z
F (u) dx ≤ lim inf F (un ) dx.
n↑∞

If F : R → R is concave and
un * u in L1 ,
then Z Z
F (u) dx ≥ lim sup F (un ) dx.
n↑∞

4 Additional reading - collection of some results

Definition 4.1 (convergence in measure). Let un , u : Ω → R be measurable functions.
We say that
un → u in measure
if
lim |{x ∈ Ω : |un (x) − u(x)| > ε}| = 0, ∀ε > 0.
n↑∞

11
 The a.e. convergence and convergence in measure can be easily compared. Indeed, the
following statements hold:

1. If |Ω| < ∞ and un → u a.e., then un → u in measure.

2. If un → u in measure, then a subsequence of un converges a.e. to u.

We also remark that if

un → u in Lp (Ω),
then a subsequence of un converges a.e. to u. This is trivial for p = ∞, and follows for
1 ≤ p < ∞ by Chebyshev’s inequality, stating that
Z
1
|{x ∈ Ω : g(x) > ε}| ≤ g dx, ∀ε > 0, 0 ≤ g ∈ L1 (Ω),
ε Ω

which implies immediately that

un → u in measure,

and hence (2) applies.

The following lemma exploits convergence of the Lp norm to get strong convergence
from a.e. convergence.

Lemma 4.1. Let un : Ω → R be a sequence in Lp (Ω), 1 ≤ p < ∞, and suppose

1. un → u a.e. in Ω,

2. kun kLp (Ω) → kun kLp (Ω) .

Then
un → u in Lp (Ω).

Let us now look at a refinement of Fatou’s lemma. Suppose

(
un * u in Lp (Ω),
(3)
un → u a.e. in Ω.

As explained before, concentration phenomena will arise when we have the weak con-
vergence in (3) simultaneously with the a.e. convergence in (3). From the a.e. convergence
and Fatou’s lemma,
kukLp (Ω) ≤ lim inf kun kLp (Ω) .
n↑∞

But we know already this from the weak convergence. Brezis and Lieb have analyzed this
situation more carefully and proved the following sharpened assertion:

12
Theorem 4.1 (Brezis-Lieb). Suppose (3) holds with 1 ≤ p < ∞. Then
 
lim kun kpLp (Ω) − kun − ukpLp (Ω) = kukLp (Ω) .
n↑∞

The message is that un in the limit (as measured in the Lp norm) decouples into un − u
and u. Note that the case p = 2 is immediate and does need the a.e. convergence in (3).
Indeed, if
un * u in L2 (Ω),
then we have
Z Z 
2 2
lim |un | dx − |un − u| dx
n↑∞ Ω Ω
Z 
2 2 2
= lim |un | − |un | + 2un u − |u| dx
n↑∞ Ω
Z  Z
= lim 2un u − |u|2 dx = |u|2 dx.
n↑∞ Ω Ω

If |Ω| < ∞, then a.e. convergence is equivalent to uniform convergence, up to arbitrarily

small sets. This is the content of Egoroff’s theorem.
Theorem 4.2 (Egoroff ). Suppose |Ω| < ∞ and that a sequence of measurable functions
un : Ω → R converges to u a.e. in Ω. Then for any ε > 0 there exists a measurable set Ωε
such that |Ω \ Ωε | < ε and
un → u uniformly in on Ωε .
We recall next some deeper results showing that under additional assumptions the
convergence of a sequence in Lp can be improved. The first result says that convergence
a.e. or in measure implies weak convergence if the norms are uniformly bounded.
Theorem 4.3. Let un : Ω → R be a sequence in Lp (Ω), 1 < p < ∞, converging a.e. or in
measure to u, with
kun kLp (Ω) ≤ C, ∀n.
Then u ∈ Lp (Ω) and
un * u in Lp (Ω).
Using the uniform convexity of the Lp spaces, 1 < p < ∞, it is not hard to prove the
following result:
Theorem 4.4 (Radon-Riesz). With 1 < p < ∞, let un : Ω → R be a sequence in Lp (Ω)
converging weakly to u ∈ Lp (Ω) and
kun kLp (Ω) → kukLp (Ω) .
Then
un → u in Lp (Ω).

13
 This theorem shows that if a weakly converging sequence in Lp , 1 < p < ∞, does not
converge strongly in Lp , then there is must be a loss in the p - norm energy.
As a corollary of the previous theorem or by a direct proof, here is a weaker form of
the Radon-Riesz theorem.

Corollary 4.1. Let un : Ω → R be a sequence in Lp (Ω) and u ∈ Lp (Ω), 1 < p < ∞.

Suppose

un → u a.e. in Ω,
kun kLp (Ω) → kukLp (Ω) .

Then
un → u in Lp (Ω).

Final remarks
A large part of this section is based on Evans’ lecture notes [1], see also Evans’ book [2].

References
[1] L. C. Evans. Weak convergence methods for nonlinear partial differential equations,
volume 74 of CBMS Regional Conference Series in Mathematics. Published for the
Conference Board of the Mathematical Sciences, Washington, DC, 1990.

[2] L. C. Evans. Partial differential equations. American Mathematical Society, Providence,

RI, 1998.

14