THE KOLMOGOROV–RIESZ COMPACTNESS THEOREM

HARALD HANCHE-OLSEN AND HELGE HOLDEN

Abstract. We show that the Arzelà–Ascoli theorem and Kolmogorov com-

pactness theorem both are consequences of a simple lemma on compactness
in metric spaces. Their relation to Helly’s theorem is discussed. The paper
contains a detailed discussion on the historical background of the Kolmogorov
arXiv:0906.4883v2 [math.CA] 4 Jan 2010

compactness theorem.

1. Introduction
Compactness results in the spaces Lp (Rd ) (1 ≤ p < ∞) are often vital in exis-
tence proofs for nonlinear partial differential equations. A necessary and sufficient
condition for a subset of Lp (Rd ) to be compact is given in what is often called the
Kolmogorov compactness theorem, or Fréchet–Kolmogorov compactness theorem.
Proofs of this theorem are frequently based on the Arzelà–Ascoli theorem. We
here show how one can deduce both the Kolmogorov compactness theorem and the
Arzelà–Ascoli theorem from one common lemma on compactness in metric spaces,
which again is based on the fact that a metric space is compact if and only if it is
complete and totally bounded.
Furthermore, we trace out the historical roots of Kolmogorov’s compactness
theorem, which originated in Kolmogorov’s classical paper [18] from 1931. However,
there were several other approaches to the issue of describing compact subsets
of Lp (Rd ) prior to and after Kolmogorov, and several of these are described in
Section 4. Furthermore, extensions to other spaces, say Lp (Rd ) (0 ≤ p < 1),
Orlicz spaces, or compact groups, are described. Helly’s theorem is often used as
a replacement for Kolmogorov’s compactness theorem, in particular in the context
of nonlinear hyperbolic conservation laws, in spite of being more specialized (e.g.,
in the sense that its classical version requires one spatial dimension). For instance,
Helly’s theorem is an essential ingredient in Glimm’s ground breaking existence
proof for nonlinear hyperbolic systems [14]. We show below that Helly’s theorem
is an easy consequence of Kolmogorov’s compactness theorem.

2. Preliminary results
An ε-cover of a metric space is a cover of the space consisting of sets of diameter
at most ε. A metric space is called totally bounded if it admits a finite ε-cover for
every ε > 0. It is well known that a metric space is compact if and only if it is
complete and totally bounded (see, e.g., [34, p. 13]). Since we are interested in
compactness results for subsets of Banach spaces, we may, and shall, concentrate
our attention on total boundedness.

Date: January 4, 2010.

2000 Mathematics Subject Classification. Primary: 46E30, 46E35; Secondary: 46N20.
Key words and phrases. Kolmogorov–Riesz compactness theorem, compactness in Lp .
Supported in part by the Research Council of Norway. This paper was written as part of
the international research program on Nonlinear Partial Differential Equations at the Centre for
Advanced Study at the Norwegian Academy of Science and Letters in Oslo during the academic
year 2008–09. The authors would like to thank Helge Kristian Jenssen for fruitful discussions.
1
2 HANCHE-OLSEN AND HOLDEN

Here is the key lemma for many compactness results (in this lemma and its proof,
every metric is named d):
Lemma 1. Let X be a metric space. Assume that, for every ε > 0, there exists
some δ > 0, a metric space W , and a mapping Φ : X → W
so that Φ[X] is totally
bounded, and whenever x, y ∈ X are such that d Φ(x), Φ(y) < δ, then d(x, y) < ε.
Then X is totally bounded.
Proof. For any ε > 0, pick δ, W and Φ as in the statement of the lemma. Since
Φ[X] is totally bounded, there exists a finite δ-cover {V1 , . . . , Vn } of Φ[X]. Then
it immediately follows from the assumptions that {Φ−1 (V1 ), . . . , Φ−1 (Vn )} is an
ε-cover of X. Thus X is totally bounded. 

Lemma 1 embodies the main argument in the standard proof of the classical Arzelà–
Ascoli theorem, as we now demonstrate.
Theorem 2 (Arzelà–Ascoli). Let Ω be a compact topological space. Then a subset
of C(Ω) is totally bounded in the supremum norm if, and only if,
(i) it is pointwise bounded, and
(ii) it is equicontinuous.
Recall the definition of equicontinuity: Condition (ii) means that for every x ∈ Ω
and every ε > 0 there is a neighborhood V of x so that |f (y) − f (x)| < ε for all
y ∈ V and all f in the given set of functions.

Proof. Assume F ⊂ C(Ω) is pointwise bounded and equicontinuous. Let ε > 0.

Combining the equicontinuity of F and compactness of Ω, we can find a finite set
of points x1 , . . . , xn ∈ Ω with neighborhoods V1 , . . . , Vn covering all of Ω so that
|f (x) − f (xj )| < ε whenever f ∈ F and x ∈ Vj .
Define Φ : F → Rn by


Φ(f ) = f (x1 ), . . . , f (xn ) .
By the pointwise boundedness of F, the image Φ[F] is bounded, and hence totally
bounded, in Rn .
Furthermore, if f, g ∈ F with kΦ(f ) − Φ(g)k∞ < ε, then since any x ∈ Ω belongs
to some Vj ,
|f (x) − g(x)| ≤ |f (x) − f (xj )| + |f (xj ) − g(xj )| + |g(xj ) − g(x)| < 3ε,
and so kf − gk∞ ≤ 3ε. By Lemma 1, F is totally bounded.
For the converse, assume that F is a totally bounded subset of C(Ω).
The existence of a finite ε-cover for F, for any ε, clearly implies the bound-
edness of F, thus establishing the uniform boundedness and hence also pointwise
boundedness of F.
To prove equicontinuity, let x ∈ Ω and ε > 0 be given. Pick an ε-cover
{U1 , . . . , Un } of F, and chose gj ∈ Uj for j = 1, . . . , n. Pick a neighborhood Vj of x
so that |gj (y) − gj (x)| < ε whenever y ∈ Vj , for j = 1, . . . , n. Let V = V1 ∩ · · · ∩ Vm .
If f ∈ Uj then kf − gj k∞ ≤ ε, and so when y ∈ V ,
|f (y) − f (x)| ≤ |f (y) − gj (y)| + |gj (y) − gj (x)| + |gj (x) − f (x)| < 3ε,
which proves equicontinuity. 

Remark 3. This theorem was first proved by Ascoli [3] for equi-Lipschitz functions
and extended by Arzelà [2] to a general family of equicontinuous functions. See [4,
p. 203].
 THE KOLMOGOROV–RIESZ COMPACTNESS THEOREM 3

We present the following theorem, first proved by Fréchet [12] for the case p = 2,
as a warm-up exercise, as the proof is short and nicely exposes some key ideas for
the proof of Theorem 5.
Theorem 4. A subset of lp , where 1 ≤ p < ∞, is totally bounded if, and only if,
(i) it is pointwise bounded, and
(ii) for every ε > 0 there is some n so that, for every x in the given subset,
X
|xk |p < εp .
k>n
p
Proof. Assume that F ⊂ l satisfies the two conditions. Given ε > 0, pick n as in
the second condition, and define a mapping Φ : F → Rn by
Φ(x) = (x1 , . . . , xn ).
By the pointwise boundedness of F, the image Φ(F) is totally bounded.
p 1/p
Pn

If x, y ∈ F with |Φ(x) − Φ(y)|p = k=1 |xk − yk | < ε, then
Xn 1/p  X 1/p
kx − ykp ≤ |xk − yk |p + |xk − yk |p < ε + 2ε = 3ε.
k=1 k>n
By Lemma 1, F is totally bounded.
We will leave proving the converse as an exercise to the reader. The techniques from
the proof of Theorem 2 are easily adapted. See also the proof of Theorem 5. 

3. The Kolmogorov–Riesz theorem

Theorem 5 (Kolmogorov–Riesz). Let 1 ≤ p < ∞. A subset F of Lp (Rn ) is totally
bounded if, and only if,
(i) F is bounded,
(ii) for every ε > 0 there is some R so that, for every f ∈ F,
Z
|f (x)|p dx < εp ,
|x|>R

(iii) for every ε > 0 there is some ρ > 0 so that, for every f ∈ F and y ∈ Rn
with |y| < ρ,
Z
|f (x + y) − f (x)|p dx < εp .
Rn
p n
Proof. Assume that F ⊂ L (R ) satisfies the three conditions. First, given ε > 0,
pick R as in the second condition, and ρ as in the third condition.
Let Q be an open cube centered at the origin so that |y| < 21 ρ whenever y ∈ Q.
Let Q1 , . . . , QN be mutually non-overlapping translates of Q so that the closure of
S
i Qi contains the ball with radius R centered at the origin. Let P be the projection
map of Lp (Rn ) onto the linear span of the characteristic functions of the cubes Qi
given by
 1
 Z
f (z) dz, x ∈ Qi , i = 1, . . . , N,
P f (x) = |Qi | Qi
0 otherwise.


From (ii) and the definition of P f we find, for f ∈ F,

XN Z
kf − P f kpp < εp + |f (x) − P f (x)|p dx
i=1 Qi
N Z
X  1 Z
p
= εp + f (x) − f (z) dz  dx.

Qi |Qi | Qi

i=1
4 HANCHE-OLSEN AND HOLDEN

Next we use Jensen’s inequality and change a variable of integration, where we note
that x − z ∈ 2Q when x, z ∈ Qi :
N Z
1
Z
f (x) − f (z)p dz dx
X
kf − P f kpp < εp +
 

i=1 Qi
|Qi | Qi
N Z
1
Z
f (x) − f (x + y)p dy dx
X
≤ εp +
 

i=1 Qi
|Qi | 2Q
1
Z Z
≤ εp + f (x) − f (x + y)p dx dy
 
|Q| 2Q Rn
1
Z
p
<ε + εp dy = (2n + 1)εp
|Q| 2Q

by (iii). Thus kf − P f kp < (2n + 1)1/p ε, and kf kp < (2n + 1)1/p ε + kP f kp . By the
linearity of P , if f, g ∈ F and kP f − P gkp < ε then kf − gkp < (2n + 1)1/p + 1 ε.

Moreover, since P is bounded (in fact kP k = 1) and F is bounded by (i), the

image P [F] is bounded. Since the image of P is finite dimensional, P [F] is totally
bounded. Thus F is totally bounded by Lemma 1.
For the converse, assume that F is totally bounded.
The existence of a finite ε-cover for F, for any ε, clearly implies the boundedness
of F, thus establishing Condition (i).
To establish Condition (ii), let ε > 0 be given, let {U1 , . . . , Un } be an ε-cover of
F, and chose gj ∈ Uj for j = 1, . . . , n. Select R so that
Z
|gj (x)|p dx < εp , j = 1, . . . , m.
x>R

If f ∈ Uj then kf − gj kp ≤ ε, and so
Z 1/p Z 1/p Z 1/p
|f (x)|p dx ≤ |f (x) − gj (x)|p dx + |gj (x)|p dx
x>R x>R x>R
Z 1/p
p
≤ kf − gj kp + |gj (x)| dx < 2ε,
x>R

thus establishing Condition (ii).

Condition (iii) is established similarly, by noting that the inequality of the con-
dition is easily established for any single function f ∈ Lp (Rn ), for example using
the fact that Cc∞ (Rn ) is dense in Lp (Rn ). Then, picking an ε-cover {U1 , . . . , Un }
and gj ∈ Uj for each j as in the previous paragraph, given ε > 0 we can find ρ > 0
with Z
|gj (x + y) − gj (x)|p dx < εp , |y| < ρ, j = 1, . . . , m.
Rn
Again, if f ∈ Uj we find
Z 1/p Z 1/p
|f (x + y) − f (x)|p dx ≤ |f (x + y) − gj (x + y)|p dx
Rn Rn
Z 1/p
+ |gj (x + y) − gj (x)|p dx
Rn
Z 1/p
+ |gj (x) − f (x)|p dx
Rn
< 3ε,
and the proof is complete. 
 THE KOLMOGOROV–RIESZ COMPACTNESS THEOREM 5

Remark 6. (I) A singleton set is clearly totally bounded, yet Condition (iii) is not
obvious for a singleton set at first glance. However, it follows easily from the density
of the space of smooth functions with compact support in Lp .
(II) In applications, one sometimes constructs a sequence f1 , f2 , . . . in Lp satis-
fying the first two conditions of Theorem 5 and the condition
Z   1/p
fn (x + y) − fn (x) dx < α(y) + β(n), lim α(y) = 0, lim β(n) = 0.
Rn y→0 n→∞

Then for some N and δ > 0, the right-hand side of the above inequality is less than
ε for all n > N and |y| small enough. By the fact noted in the previous paragraph,
we can choose a smaller upper bound for |y| to make the integral smaller than ε for
n = 1, 2, . . . , N . Thus {f1 , f2 , . . .} satisfies Condition (iii), and hence a convergent
subsequence exists.
An interesting corollary to the Kolmogorov theorem is the following result, see
[22], which also contains a variant using the uniform smoothness of the functions in
F and their Fourier transforms. See also [7], which contains an alternate formulation
based on the short-time Fourier transform, as well as one based on the wavelet
transform.
Corollary 7. Let F ⊆ L2 (Rd ) be such that supf ∈F kf k2 ≤ M < ∞. If
Z Z
lim sup |f (x)|2 dx = 0 and lim sup |fˆ(ξ)|2 dξ = 0,
r→∞ f ∈F |x|≥r ρ→∞ f ∈F |ξ|≥ρ

then F is totally bounded in L2 (Rd ).

Proof. We show that F satisfies the conditions of Theorem 5 for p = 2. Clearly,
Conditions (i) and (ii) are among our assumptions, so we only need to prove (iii).
For f ∈ F we find:
Z Z
|f (x + y) − f (x)|2 dx = |(eiξ·y − 1)fˆ(ξ)|2 dξ
R n n
ZR Z
2
− 1)fˆ(ξ) dξ + 4 fˆ(ξ)2 dξ
 iξ·y  
≤ (e
|ξ|<ρ |ξ|>ρ
2 iξ·y 2
≤ M sup |e − 1| + ε for ρ big enough
|ξ|<ρ

< M 2 ρ2 |y|2 + ε < 2ε

√
if |y| < ε/(ρM ). Here ρ, and hence the upper bound on |y|, can be chosen
independently of f . This shows Condition (iii) of Theorem 5 and finishes the
proof. 
In the following result, Lploc (Ω) is equipped with the topology of Lp convergence
on compact subsets of Ω. Recall that Ω is the countable union of compacts, e.g.,
Ω = K1 ∪K2 ∪. . . with Kk = {x ∈ Ω : |x| ≤ k and dist(x, Rn \Ω) ≥ 1/k}. Moreover
any compact subset of Ω is contained in some Kk , and so the topology on Lploc (Ω)
is given by the countable family of seminorms kf kk = kf |Kk kLp (Kk ) . Lploc (Ω) is
P∞
complete with respect to the metric (f, g) 7→ k=1 min(2−k , kf − gkk ).
Corollary 8. Let Ω ⊆ Rn be an open set. Write fK (x) = f (x) when x ∈ K,
fK (x) = 0 otherwise. A subset F ⊆ Lploc (Ω) is totally bounded if, and only if, the
following holds:
(i) For every compact K ⊂ Ω there is some M so that
Z
|fK (x)|p dx < M, f ∈ F.
6 HANCHE-OLSEN AND HOLDEN

(ii) For every ε > 0 and every compact K ⊂ Ω there is some ρ > 0 so that
Z
|fK (x + y) − fK (x)|p dx < εp , f ∈ F, |y| < ρ.

Proof. Note that F is totally bounded in Lploc (Ω) if and only if Fk = {fKk : f ∈ F}
is totally bounded for every k, with Kk as defined above. 

For the next result, recall that the Sobolev space W k,p (Rn ) is defined to con-
sist of those measurable functions f which, together with all their distributional
derivatives Dα f of order |α| ≤ k, belong to Lp (Rn ). Here α = (α1 , . . . , αn ) is
a multi-index, i.e., each αj is a nonnegative integer, |α| = α1 + · · · + αn , and
Dα = ∂ |α| /(∂xα αn
1 · · · ∂xn ). Finally, W
1 k,p
(Rn ) is equipped with the complete norm
Z X 1/p
kf kk,p = |Dα f (x)|p dx .
Rn |α|≤k

Corollary 9. A subset F ⊆ W k,p (Rn ) is totally bounded if, and only if, the fol-
lowing holds:
(i) F is bounded, i.e., there is some M so that
Z
|Dα f (x)|p dx < M, f ∈ F, |α| ≤ k.

(ii) For every ε > 0 there is some R so that

Z
|Dα f (x)|p dx < εp , f ∈ F, |α| ≤ k.
|x|>R

(iii) For every ε > 0 there is some ρ > 0 so that

Z
|Dα f (x + y) − Dα f (x)|p dx < εp , f ∈ F, |α| ≤ k, |y| < ρ.
Rn

Proof. Note that F is totally bounded in W k,p (Rn ) if and only if Dα [F] = {Dα f : f ∈
F} is totally bounded in Lp (Rn ) for every multi-index α with |α| ≤ k. 

4. A bit of history
In 1931, Kolmogorov [18] proved the first result in this direction. It characterizes
compactness in Lp (Rn ) for 1 < p < ∞, in the case where all functions are supported
in a common bounded set. Condition (iii) of Theorem 5 is replaced by the uniform
convergence in Lp norm of spherical means of each function in the class to the
function itself. (Clearly, our Condition (ii) is automatic in this case.)
Just a year later, Tamarkin [28] expanded this result to the case of unbounded
supports by adding Condition (ii) of Theorem 5.
In 1933, Tulajkov [31] expanded the Kolmogorov–Tamarkin result to the case
p = 1.
In the same year, and probably independently, M. Riesz [25] proved the result
for 1 ≤ p < ∞, essentially in the form of our Theorem 5. Thus we feel somewhat
justified in using the names Kolmogorov and Riesz in referring to the theorem,
though we are perhaps being a bit unfair to Tamarkin and Tulajkov in doing so.
The compactness theorem has also seen generalizations in other directions.
Hanson [15] proved a necessary and sufficient condition for compactness of a
family of measurable functions on a bounded measurable set, with respect to con-
vergence in measure. (Here the measurable functions form a metric space in which
the distance between two functions is the infimum of all ε > 0 so that the two
functions differ by at most ε except on a set of measure ≤ ε.)
 THE KOLMOGOROV–RIESZ COMPACTNESS THEOREM 7

Fréchet [13] replaced Conditions (i) and (ii) of Theorem 5 with a single condition
(“equisummability”), and generalized the theorem to arbitrary positive p.
Phillips [23, Thm 3.7] proved a necessary and sufficient condition for compactness
in Lp on a general measure space (1 ≤ p ≤ ∞), and indeed in any Banach space,
which is however somewhat less suited to applications to PDEs. Nevertheless, our
sufficiency proof for Theorem 5 is based on Phillips’ criterion. (It is more common,
albeit more involved, to use mollifiers in the proof.)
Weil [33] (see also [9, p. 269 ff]) extended the result to Lp (G) where G is a
locally compact group. Tsuji [30] considered the case of Lp (Rd ) with 0 < p < 1,
and Takahashi [27] studied the same problem in Orlicz spaces. A characterization
of compact subsets of Lp ([0, T ]; B) (B a Banach space), which is very convenient
in the context of time-dependent partial differential equations, is given by Simon
[26] (see also [20]). A readable account of some of the historical development can
be found in [8, p. 388]. Helly’s theorem [16], which was published already in 1912,
is easily seen to be a special case of Kolmogorov’s compactness theorem in the
one-dimensional case, see Section 6.
Further references include [32], [17], [5], [6], [11], [21].

5. The Rellich–Kondrachov theorem

In this section we use Kolmogorov’s theorem to prove a simple variant of the
Rellich–Kondrachov theorem [24, 19]. Our simplification consists in avoiding bound-
ary regularity conditions by working on the entire space Rn . The standard Rellich–
Kondrachov theorem requires a bounded region. The present version replaces this
by a uniform decay estimate, specially tailored to fit the framework of the present
paper.
The Sobolev norm kf k1,p on W 1,p (Rn ) is defined by
Z
1/p X ∂f p 1/p
kf k1,p = |f (x)|p + |∇f (x)|pp dx , |∇f |p = .
 
∂x
 
R n
j i

According to the Sobolev embedding theorem, if p < n then W 1,p (Rn ) ⊂ Lq (Rn ),
and the inclusion map is bounded, for any q satisfying p ≤ q ≤ p∗ , where p∗ is the
conjugate Sobolev exponent:
1 1 1
= − .
p∗ p n
To see where this exponent comes from, consider a function f and its scalings
f λ (x) = f (x/λ) where λ > 0, and note that kf λ kp = λn/p kf kp and k∇f λ kp =
λn/p−1 k∇f kp , so the inclusion map W 1,p → Lq can only be bounded if there exists
a constant C with λn/q ≤ C(λn/p + λn/p−1 ) for all λ > 0. In the limits λ → ∞ and
λ → 0 we conclude n/q ≤ n/p and n/q ≥ n/p − 1 respectively.
Theorem 10. Assume p < n and p ≤ q < p∗ , and let F be a bounded subset
of W 1,p (Rn ). Assume that for every ε > 0 there exists some R so that, for every
f ∈ F, Z
|f (x)|p + |∇f (x)|pp dx < εp .


|x|>R
Then F is a totally bounded subset of Lq (Rn ).
Proof. We shall show that F satisfies the hypotheses of Theorem 5, with p replaced
by q. We shall use the Sobolev embedding inequality kf kq ≤ Ckf k1,p , where
the constant C depends only on p, q and n, and which is valid under the stated
assumption, see [1, 4.30 (p. 101) and Theorem 4.12 I C (p. 85) with j = 0,
k = n, m = 1]. Condition (i) of Theorem 5 follows immediately from the Sobolev
embedding inequality. Condition (ii) is almost equally immediate, from applying
8 HANCHE-OLSEN AND HOLDEN

the Sobolev embedding inequality to the function x 7→ f (x)χ(|x| − R), where χ ∈

C ∞ (R), 0 ≤ χ ≤ 1, χ(x) = 0 for x < 0 and χ(x) = 1 for x > 1.
If we apply the Sobolev embedding inequality to the function x 7→ f (x/λ) where
λ > 0 and change variables in the resulting integrals, we obtain
 Z Z 1/p
n/q n p n−p
λ kf kq ≤ C λ |f (x)| dx + λ |∇f (x)|pp dx (1)
Rn Rn
We shall apply the above inequality not to f , but to x 7→ f (x + y) − f (x), where
f ∈ F.
Now let ε > 0 be given. By picking λ sufficiently large we can ensure that
 Z 1/p
C λn−p |∇f (x + y) − ∇f (x))|pp dx ≤ ελn/q (2)
Rn
for all f ∈ F, since the integral in this expression is bounded uniformly for f ∈ F.
Next, we find (using the Jensen and Hölder inequalities, then Fubini’s theorem)
Z Z Z 1 p
|f (x + y) − f (x)|p dx = y · ∇f (x + ty) dt dx
 

Rn Rn 0
Z 1 Z
≤ |y|pp′ |∇f (x + ty)|pp dx dt
0 Rn
Z
= |y|pp′ |∇f (x)|pp dx
Rn
(where p and p′ are conjugate exponents) for any test function f , and hence for
any f ∈ W 1,p . The integrals on the right-hand side of this inequality are uniformly
bounded for f ∈ F, and so we can find some δ > 0 so that |y| < δ implies
 Z 1/p
C λn |f (x + y) − f (x)|p dx ≤ ελn/q . (3)
Rn
For such y and f , (1) applied to x 7→ f (x + y) − f (x) combined with (2) and (3) to
yield
λn/q kf (· + y) − f (·)kq ≤ 21/p ελn/q ,
and so assumption (iii) of Theorem 5 is satisfied. 

6. Helly’s theorem
Helly’s theorem is often referred to as Helly’s selection principle, in order to
avoid confusion with another theorem by Helly, stating that, given a collection of
convex sets in Rn so that any n + 1 of them have a point in common, then any finite
subcollection has nonempty intersection. Helly’s selection principle is essentially a
corollary of the Kolmogorov–Riesz theorem, though historically it was not derived
that way.
Recall that an integrable function f on the line is of bounded variation if it has
finite essential or total variation, that is, if
Xm
TV(f ) = sup |f (xj+1 ) − f (xj )| < ∞,
j=1

where the supremum is taken over all finite partitions xj < xj+1 such that each xj is
a point of approximate continuity of f (that is, δ −1 |{x : |x−xj | < δ, |f (x)−f (xj )| ≥
ε}| → 0 for every ε > 0 as δ → 0. See, e.g., [10, p. 47]). We need a lemma:
Lemma 11. Let u be function of bounded variation on R. Then
Z ∞
|u(x + y) − u(x)| dx ≤ |y| TV(u)
−∞
for all y ∈ R.
 THE KOLMOGOROV–RIESZ COMPACTNESS THEOREM 9

Proof. We may assume y > 0 without loss of generality. The calculation

Z ∞ ∞ Z y
X
|u(x + y) − u(x)| dx = |u(x + jy + y) − u(x + jy)| dx
−∞ j=−∞ 0
Z y ∞
X
= |u(x + (j + 1)y) − u(x + jy)| dx
0 j=−∞
Z y
≤ TV(u) dx = y TV(u),
0
finishes the proof. 

Theorem 12 (Helly). Let (un ) be a sequence of functions of bounded variation on

the bounded real interval [a, b]. If there is a constant M so that TV(un ) ≤ M and
kun k∞ ≤ M for all n, then there is a subsequence of (un ) which converges pointwise
everywhere and in L1 norm in [a, b] to a function of bounded variation.
Proof. Extend each function un to all of R by setting it to zero outside [a, b]. By
Lemma 11, the set of all the functions un satisfy Condition (iii) of Theorem 5
(with p = 1), while (i) holds by assumption and (ii) is trivial. Hence there is a
subsequence of (un ) which converges in L1 ([a, b]). Moreover, integration theory
tells us that we also get pointwise convergence almost everywhere, possibly after
passing to a subsequence once more. However, this is not quite enough.
Write instead un = vn − wn where each vn , wn is an non-decreasing function:
vn (x) is un (a) plus the positive variation of un on the interval [a, x], and wn (x) is
the negative variation on the same interval. Then the sequences (vn ) and (wn ) both
satisfy the conditions of the present theorem, and so, by the result of the previous
paragraph, we may pass to a subsequence so that (vn ) and (wn ) both converge in
L1 ([a, b]), as well as pointwise almost everywhere.
Let v be the limit of the sequence (vn ). Clearly, v is non-decreasing on the set
where pointwise convergence holds, and so we may assume that v is non-decreasing
everywhere, after possibly redefining it on a set of measure zero.
Now it is clear that vn (x) → v(x) for any point of continuity x for v: Given ε > 0,
pick δ > 0 so that |y − x| < δ implies |v(y) − v(x)| < ε, let x − δ < y < x < z < x + δ
with vn (y) → v(y) and vn (z) → v(z), and note that for n large enough we get
v(x) − 2ε < v(y) − ε < vn (y) ≤ vn (x) ≤ vn (z) < v(z) + ε < v(x) + 2ε, so that
|vn (x) − v(x)| < 2ε.
Since v has at most a countable number of discontinuities, a diagonal argument
yields a further subsequence which converges at all the discontinuities of v as well,
and so we have pointwise convergence everywhere.
In the same way we show that wn (x) → w(x) for all x. Thus un → v − w
pointwise, and v − w has bounded variation. 

Remark 13. The above proof is probably not the most natural one, but it does
make clear the connection with the Kolmogorov–Riesz theorem. In a sense L1
convergence is irrelevant: Pointwise convergence is the key, and L1 convergence
follows from the bounded convergence theorem.
It should be noted, however, that Helly’s theorem, without pointwise conver-
gence, is also true in higher dimensions [10, p. 176].
A recent generalization of Helly’s selection principle (in one dimension) can be
found in [29].

References
[1] A. Adams and J. F. Fournier. Sobolev Spaces. Second edition. Academic Press 2003.
10 HANCHE-OLSEN AND HOLDEN

[2] C. Arzelà. Sulle funzione de linee. (In Italian) Mem. Accad. Sci. Bologna 5-5 (1894/95)
225–244.
[3] G. Ascoli. Le curve limiti di una varietá data di curve. (In Italian) Rend. Accad. Lincei 18
(1884) 521–586.
[4] E. DiBenedetto. Real Analysis. Birkhäuser, Boston, 2002.
[5] G. Bruno and R. Grande. A compactness criterion in Bap q
spaces. Rend. Accad. Naz. Sci. XL
Mem. Mat. Appl. 20 (1996), 95–121.
[6] G. Bruno and R. Grande. Compact embedding theorems for Sobolev–Besicovitch spaces of
almost periodic functions. Rend. Accad. Naz. Sci. XL Mem. Mat. Appl. 20 (1996) 157–173.
[7] M. Dörfler, H. G. Feichtinger, K. Gröchenig. Compactness criteria in function spaces. Coll.
Math. 94 (2002) 37–50.
[8] N. Dunford and J. T. Schwartz. Linear Operators. Part I: General Theory. Wiley, New York,
1988.
[9] R. E. Edwards. Functional Analysis. Holt, Rinehart and Winston, Chicago, 1965.
[10] L. C. Evans and R. F. Gariepy. Measure Theory and Fine Properties of Functions. CRC
Press, Boca Raton, 1992.
[11] H. G. Feichtinger, Compactness in translation invariant Banach spaces of distributions and
compact multipliers. J. Math. Anal. Appl. 102 (1984) 289–327.
[12] M. Fréchet. Essai de geometrie analytique. (In French) Nouv. ann. math. 4 (1908) 97–116,
289–317.
[13] M. Fréchet. Sur les ensembles compacts de fonctions de carrés sommables. (In French) Acta
Litt. Sci. Szeged. 8 (1937), 116–126.
[14] J. Glimm. Solutions in the large for nonlinear hyperbolic systems of equations. Comm. Pure
Appl. Math. 18 (1965), 697–715.
[15] E. H. Hanson. A note on compactness. Bull. Amer. Math. Soc. 39 (1933), 397–400.
[16] E. Helly. Über lineare Funktionaloperationen. (In German) Österreich. Akad. Wiss. Math.-
Natur. Kl. S.-B. IIa 121 (1912) 265–297.
[17] S. Izumi. On the compactness of a class of functions. Proc. Imp. Acad. Tokyo 15 (1939)
111–113.
[18] A. N. Kolmogorov. Über Kompaktheit der Funktionenmengen bei der Konvergenz im Mittel,
Nachr. Ges. Wiss. Göttingen 9 (1931), 60–63.
English translation: On the compactness of sets of functions in the case of convergence in
the mean, in V. M. Tikhomirov (ed.), Selected Works of A. N. Kolmogorov, Vol. I, Kluwer,
Dordrecht, 1991, pp. 147–150.
[19] W. Kondrachov. Sur certaines propriétés des fonctions dans l’espace. C. R. Acad. Sci. URSS.,
n. Sér. 48 (1945) 535–538.
[20] E. Maitre. On a nonlinear compactness lemma in Lp (0, T ; B). Int. J. Math. Math. Sci. no. 27
(2003) 1725–1730.
[21] M. Nicolescu. On the criterion of compactness of A. Kolmogorov. (In Romanian) Acad. Repub.
Pop. Române. Bul. Şti. Ser. Math. Fiz. Chim. 2 (1950) 407–415.
[22] R. L. Pego. Compactness in L2 and the Fourier Transform. Proc. Amer. Math. Soc. 95 (1985)
252–254.
[23] R. S. Phillips. On linear transforms. Trans. Amer. Math. Soc. 48 (1940) 516–541.
[24] F. Rellich. Ein Satz über mittlere Konvergenz. Nachrichten von der Gesellschaft der Wis-
senschaften zu Göttingen, Mathematisch-Physikalische Klasse (1930) 30–35.
[25] M. Riesz. Sur les ensembles compacts de fonctions sommables. (In French) Acta Szeged Sect.
Math. 6 (1933), 136–142.
Also in L. Gårding, L. Hörmander (eds.), Marcel Riesz—Collected Papers, Springer, Berlin
(1988).
[26] J. Simon. Compact sets in the space Lp (0, T ; B). Ann. Mat. Pura Appl. 146 (1987), 65–96.
[27] T. Takahashi. On the compactness of the function-set by the convergence in mean of general
type. Studio Math. 5 (1934) 141–150.
[28] J. D. Tamarkin. On the compactness of the space Lp . Bull. Amer. Math. Soc. 32 (1932),
79–84.
[29] Yu. V. Tret’yachenko and V. V. Chistyakov. Selection principle for pointwise bounded se-
quences of functions. Math. Notes 84 (2008) 396–406.
[30] M. Tsuji. On the compactness of space Lp (p > 0) and its application to integral operators.
Kodai Math. J. (1951) 33–36.
[31] A. Tulajkov. Zur Kompaktheit im Raum Lp für p = 1. (In German) Nachr. Ges. Wiss.
Göttingen, Math. Phys. Kl. I 1933, nr. 39, 167–170.
[32] P. Veress. Über Functionenmengen. (In German) Acta scient. math. 3 (1927) 177–192.
[33] A. Weil. L’intégration dans les groupes topologiques et ses applications. Hermann et Cie.,
Paris, 1940.
 THE KOLMOGOROV–RIESZ COMPACTNESS THEOREM 11

[34] K. Yosida. Functional Analysis. Springer, Berlin, 1980.

(Hanche-Olsen)
Department of Mathematical Sciences
Norwegian University of Science and Technology
NO–7491 Trondheim, Norway
E-mail address: [email protected]
URL: http://www.math.ntnu.no/˜hanche/

(Holden)
Department of Mathematical Sciences
Norwegian University of Science and Technology
NO–7491 Trondheim, Norway,
and
Centre of Mathematics for Applications
University of Oslo
P.O. Box 1053, Blindern
NO–0316 Oslo, Norway
E-mail address: [email protected]
URL: http://www.math.ntnu.no/˜holden/