Compact Subsets of Lp Spaces

S. Kumaresan
School of Math. and Stat.
University of Hyderabad
Hyderabad 500046
[email protected]

It is very useful to characterize the compact subsets of concrete metric spaces using in-
trinsic properties of elements of the spaces. For instance, Heine-Borel theorem gives such a
characterization for compact subsets of Rn and Arzela-Ascoli theorem achieves this for the
space (C(X), k ksup ) of continuous functions on a compact space X. Our aim in this article
is to give a characterization of compact subsets of Lp [0, 1], 1 ≤ p < ∞ and later in Lp (R).
Let f ∈ Lp [0, 1]. We can consider it as a function on all of R by setting f (x) = 0 for
x∈/ [0, 1].
R1
Ex. 1. Prove that Lp [0, 1] ⊂ L1 [0, 1]. Hint: Apply Hölder’s inequality to 0 f (x) · 1 dx where
1 stands for the constant function 1.

For f ∈ Lp [0, 1] and h > 0, define

Z t+h
1
fh (t) := f (x) dx.
2h t−h

The integral exists thanks to Ex. 1.

Lemma 2. For f ∈ Lp [0, 1] and h > 0, fh is continuous and fh ∈ Lp [0, 1].

Proof. Let tn → t. Let gn := f χ(tn −h,tn +h) . Then gn → f χ(t−h,t+h) pointwise. Also, we
observe that |gn | ≤ |f |. We apply Lebesgue’s dominated convergence theorem to conclude
Z tn +h Z t+h
1 1
lim fh (tn ) = lim f (x) dx = f (x) dx = fh (t).
2h tn −h 2h t−h

Since C[0, 1] ⊂ Lp [0, 1], the lemma is proved.

Lemma 3. Let 1 ≤ p < ∞, and f ∈ Lp [0, 1]. Then for each h > 0, we have

|fh (t)| ≤ (2h)−1/p kf kp , for all t ∈ [0, 1]. (1)

1
Proof. Assume p > 1. Let q be the conjugate index such that 1/p + 1/q = 1. We apply
Hölder’s inequality to get
Z t+h
1
|fh (t)|p = | 1 · f (x) dx|p
(2h)p t−h
Z t+h p/q Z t+h
1
≤ 1 dx · |f (x)|p dx
(2h)p t−h t−h
Z t+h
1
= |f (x)|p dx.
2h t−h

Thus we have shown Z t+h

1
|fh (t)|p ≤ |f (x)|p dx (2)
2h t−h

holds for 1 < p < ∞ and for all t ∈ [0, 1]. When p = 1, (2) is obviously true. The inequality
(1) follows from (2).

Lemma 4. With the notation of the last lemma, we have

kfh kp ≤ kf kp . (3)

Proof. Using the inequality (2), it follows that

Z 1 Z 1 Z t+h 
p 1 p
|fh (t)| dt ≤ |f (x)| dx dt
0 2h 0 t−h
Z 1 Z h 
1 p
= |f (t + y)| dy dt. (4)
2h 0 −h

Observe that (t, y) 7→ f (t + y) is Lebesgue measurable. Since the integrand is nonnegative,

we can apply Fubini-Toneli’s theorem to get
Z 1 Z h  Z h Z 1 
|f (t + y)|p dy dt = |f (t + y)|p dt dy
0 −h −h 0
Z 1
≤ 2h |f (x)|p dx.
0

Thus (4) implies

Z 1 Z 1
p
|fh (t)| dt ≤ |f (x)|p dx,
0 0
and the result follows.

Theorem 5. Let 1 ≤ p < ∞. Let K be a closed and bounded subset of Lp [0, 1]. Then K is
compact iff for each ε > 0 there exists a δ > 0 such that kf − fh kp < ε for all f ∈ K and
0 < h < δ.

Proof. We prove that the condition is necessary. Let ε > 0 be given. Since C[0, 1] is dense
in Lp [0, 1] and K is compact there exists continuous functions fj , 1 ≤ j ≤ n, such that
K ⊂ ∪nj=1 B(fj , ε).

2
 By uniform continuity of each fj , there exists some δ > 0 such that |f (x) − f (t)| < ε holds
for each 1 ≤ j ≤ n for all x, t ∈ [0, 1] with |x − t| < δ. In particular,, for 0 < h < δ, we have
Z t+h
1
|fj (t) − (fj )h (t)| = | [fj (t) − fj (x)] dx| ≤ ε.
2h t−h

Thus, kfj − (fj )h kp ≤ ε.

Now if f ∈ K, we choose j so that f ∈ B(fj , ε). By Lemma 4 we have kfh − (fj )h kp ≤
kf − fj kp < ε. Therefore,

kf − fh kp ≤ kf − fj kp + kfj − (fj )h kp + k(fj )h − fh kp < 3ε

holds for all f ∈ K and 0 < h < δ.

We now prove that the condition is sufficient. Since Lp [0, 1] is complete, it is enough to
prove that K is totally bounded. Towards this end, let ε > 0 be given. Fix an h > 0 such
that kf − fh kp < ε for all f ∈ K. Let M > 0 be such that kf kp < M for all f ∈ K. By
Lemma 4, it follows that
|fh (t)| ≤ M (2h)−1/p =: C,
holds for all t ∈ [0, 1] and f ∈ K. Set Kh := {fhh : f ∈ K}. Here
Z t+h
1
fhh := fh (x) dx.
2h t−h

Clearly, |fhh (t)| ≤ C for t ∈ [0, 1] and f ∈ K. Hence Kh is a bounded set in (C[0, 1] < k ksup ).
We now show that Kh is equicontinuous. To see this, note that if s < t and f ∈ K, then
Z t+h Z s+h
1
|fhh (t) − fhh (s)| = | fh (x) dx − fh (x) dx|
2h t−h s−h
Z t+h Z t+h
1
= | fh (x) dx − fh (x) dx|
2h s+h s−h
Z t+h Z t−h 
1
≤ |fh (x)| dx + |fh (x)| dx
2h s+h s−h
1 C
≤ [2C(t − s)] = (t − s).
2h h
The equicontinuity of Kh follows from this.
By Arzela-Ascoli, Kh is totally bounded and hence we can choose f1 , . . . , fn ∈ K such that
for each f ∈ K, there exists j such that kfhh − (fj )hh ksup < ε. We then have

kf − fj kp ≤ kf − fh kp + kfh − fhh kp + kfhh − fj kp

< 2ε + kfhh − fj kp
< 2ε + kfhh − (fj )hh kp + k(fj )hh − (fj )h kp + k(fj )h − (fj )h kp
< 5ε.

This shows that K is totally bounded subset of Lp [0, 1] and completes the proof of the theorem.

3
 We now characterize compact subsets of Lp (R). We need to impose one more condition
on the family, viz. that of ‘uniform integrability’ (see condition (iv) in the theorem below).
We use kf k in place of kf kp to simplify our typing.

Theorem 6. Let K ⊂ Lp (R), 1 ≤ p < ∞. Then K is compact iff K satisfies the following
conditions:
(i) K is closed.
(ii) K is bounded.
(iii) Given ε > 0, there exists δ > 0 such that
Z
|f (t + s) − f (s)|p ds < ε f orallf ∈ K, 0 < |t| < δ.
R

(iv) Given ε > 0, there exists α > 0 such that

Z
|f (s)|p ds < ε, for all f ∈ K.
|s|>α

Proof. Let K be compact. Then it is closed and bounded. In fact. it is totally bounded.
Therefore, given ε > 0, there exist f1 , . . . , fn ∈ Lp such that K ⊂ ∪j B(fj , ε). Since the set
of all finite linear combinations of characteristic functions of bounded intervals is dense in
Lp , there exist such functions gj such that kfj − gj k < ε. Now for α > 0 sufficiently large,
support of gj will be contained in [−α, α] for 1 ≤ j ≤ n. Hence for all such large α, we have
Z −α Z ∞ 1/p Z −α Z ∞ 1/p
p p p p
|f (s)| + |f (s)| ≤ |f (s) − gj (s)| + |f (s) − gj (s)|
−∞ α −∞ α
Z −α Z ∞ 1/p
p p
+ |gj (s)| + |gj (s)|
−∞ α
Z −α Z ∞ 1/p
p p
≤ kf − gj k + |gj (s)| + |gj (s)| .
−∞ α

Hence (iv) follows from

kf − gj k ≤ kf − fj k + kfj − gj k ≤ 2ε.

To prove (iii), we start by observing that it holds for the characteristic function χJ of any
finite interval J: Z
lim |χJ (s + t) − χJ (s)|p ds = 0.
t→0 R

Thus (iii) holds for finite linear combinations of such functions, in particular, for gj ’s as above,

4
Hence we have for any f ∈ K,
Z 1/p Z 1/p
p p
lim sup |f (s + t) − f (s)| ds ≤ lim sup |f (s + t) − fj (s + t)| ds
t→0 R t→0 R
Z 1/p
p
+ lim sup |fj (s + t) − gj (s + t)| ds
t→0 R
Z 1/p
p
+ lim sup |gj (s + t) − gj (s)| ds
t→0 R
Z 1/p
p
+ |gj (s) − fj (s)| ds
R
Z 1/p
p
+ |fj (s) − f (s)| ds
R
≤ ε + ε + 0 + ε + ε,

if j is so chosen that kf − fj k < ε. This completes the proof of the necessary part of the
conditions.
To prove the sufficiency part, we need only show that if K satisfies the conditions, then
K is totally bounded. We define the translation operator Tt by Tt f (s) := f (t + s). Condition
(ii) means that Tt f → f in Lp as t → 0 uniformly for f ∈ K. We also define the mean value
Z a Z a
1 1
Ma f (s) := Tt f (s) ds ≡ f (t + s) ds.
2a −a 2a −a

Using Hölder’s inequality and Fubini-Tonelli as in the last theorem, we get

Z Z a p 1/p
1
kMa f − f k ≤ |f (t + s) − f (s)| dt ds
R −a 2a
Z Z a 1/p
1
≤ |f (t + s) − f (s)|p dt(2a)p/q ds
2a R −a
 Z a Z 1/p
1 p
≤ dt |f (t + s) − f (s)| ds .
2a −a R

Thus we have shown

kMa f − f k ≤ sup kTt f − f k . (5)
|t|≤a

Hence it is enough to establish the total boundedness of Ka := {Ma f : f ∈ K} for any fixed
but sufficiently small a.
We shall show that for any fixed a > 0, the set Ka is bounded in L∞ -norm and equicon-
tinuous. In fact, we have
Z a
1
|Ma f (s1 ) − Ma f (s2 )| ≤ |f (s1 + t) − f (s2 + t)| dt
2a −a
 Z a 1/p
1
≤ |f (s1 + t) − f (s2 + t)|p dt .
2a −a

5
This along with (ii) shows that Ka is equicontinuous. A similar proof establishes that Ka is
bounded in L∞ -norm.
Let α > 0 be given. By Arzela-Ascoli applied to C[−α, α] for any given ε > 0, there exist
Ma fj , 1 ≤ j ≤ n with fj ∈ K such that for any given f ∈ K, there exists j such that

sup |Ma f (s) − Ma fj (s)| ≤ ε. (6)

|s|≤α

We use this to show that Ka is totally bounded in Lp (R).

We have
Z α
kMa f − Ma fj kp = |Ma f (s) − Ma fj (s)|p ds
−α
Z
+ |Ma f (s) − Ma fj (s)|p ds. (7)
|s|>α

The first term on the right side is ≤ 2αεp for an appropriate choice of j. We estimate the
second term on the right of (7):
Z Z !1/p
p p
|Ma f (s) − Ma fj (s)| ds ≤ kMa f − f k + |f (s) − fj (s)| ds
|s|>α |s|>α
Z !1/p
p
+ |fj (s) − Ma fj (s)| ds .
|s|>α

The term kMa f − f k → 0 as a → 0+. By virtue of (iii), the other two terms go to 0 as
α % ∞ if a remains bounded. This completes the proof of the fact that Ka is totally bounded
in Lp (R) and hence the proof of the theorem.