MATH 131BH notes 140

25. U NIFORM AND C AUCHY CONTINUITY

We have seen above that continuous functions are exactly those that convert conver-
gent sequences to convergent sequences. A natural variation on this is: What functions
convert Cauchy sequences to Cauchy sequences. We will give these a name:
Definition 25.1 (Cauchy continuity) Let f : X Ñ Y be a function between metric spaces X
and Y. We say that f is Cauchy continuous if

@txn unPN P Domp f qN : txn unPN Cauchy (in X) ñ t f pxn qunPN Cauchy (in Y) (25.1)

i.e., if f turns Cauchy sequences to Cauchy sequences.

We immediately note:
Lemma 25.2 (AC) A Cauchy continuous function is continuous.

Proof. Let f : X Ñ Y be Cauchy continuous and let x P Domp f q. Consider a se-

quence txn unPN P Domp f qN such that xn Ñ x and define tyn unPN by y2n :“ xn and
y2n`1 :“ x. Then also yn Ñ x. But then t f pyn qunPN is Cauchy which means
` ˘ ` ˘
rY f pxn q, f pxq “ rY f py2n q, f py2n`1 q Ñ 0 (25.2)

Hence f pxn q Ñ f pxq. By Theorem 24.7 (which is where we need to call upon the Axiom
of Choice), f is continuous at x. ⇤
However, the converse does not hold. Indeed consider a function f : R r t0u Ñ R
defined by
#
1, if x ° 0,
f pxq :“ (25.3)
0, if x † 0,
Then f is continuous yet not Cauchy continuous because the sequence xn :“ p´2q´n is
Cauchy yet t f pxn qunPN is NOT.
The concept of Cauchy continuity is somewhat special and in practice we typically
use different concepts that address directly continuity in somewhat quantitative way.
The simplest of these is:
Definition 25.3 (Uniform continuity) A function f : X Ñ Y between metric spaces pX, r X q
and pY, rY q is said to be uniformly continuous if
` ˘
@e ° 0 Dd ° 0 @x, y P X : r X px, yq † d ñ rY f pxq, f pyq † e (25.4)
As it turns out, this differs from the definition of continuity in a rather inconspicuous
yet very important way. Indeed, a function f : X Ñ Y is continuous if
` ˘
@e ° 0 @x P X Dd ° 0 @y P X : r X px, yq † d ñ rY f pxq, f pyq † e (25.5)

which differs from (25.4) by a swap of Dd ° 0 and @x P X quantifiers. This swap amounts
to the fact that d in (25.5) may depend on x yet in (25.4) one d must work for all x (and y)
simultaneously. This immediately gives:
Lemma 25.4 A uniformly continuous function is continuous.

Preliminary version (subject to change anytime!) Typeset: April 13, 2023

141 MATH 131BH notes

Again, the converse to this fails as witnessed by the example

f pxq :“ x2 (25.6)
with Domp f q “ R. Indeed, px ` dq2 ´ x2 “ 2xd ` d2 which cannot be made smaller
than e without restricting the size of x. We in fact have a stronger conclusion:
Lemma 25.5 A uniformly continuous function is Cauchy continuous.

Proof. Let f : X Ñ Y be uniformly continuous. Given e ° 0 let d ° 0 be such that

r X px, yq † d implies rY p f pxq, f pyqq † e. Given a Cauchy sequence txn unPN , there ex-
ists n0 • 0 such that r X pxn , xm q † d for all n, m • n0 . But then r X p f pxn q, f pxm qq † e for
all n, m • n0 . As this holds for all e ° 0, t f pxn qunPN is Cauchy as desired. ⇤
The example (25.6) shows that the converse to this does not hold; indeed, the function
there is Cauchy continuous yet not uniformly continuous. However, this would not be
the case if we restrict that f to any bounded subset of R. Indeed, we have:
Lemma 25.6 Let X and Y be metric space with X compact. The for all f : X Ñ Y:
f continuous ñ f uniformly continuous (25.7)
(Note that, thanks to Lemma 25.4, we can even write ô in (25.7).)

Proof. Assume X to be compact and let f : X Ñ Y be continuous. Given e ° 0, for

each x P X let
(
D x :“ d P p0, 8q : diam f pBX px, 2dqq † e (25.8)
The continuity of f at x ensures that D x ‰ H for all x P X. This, along with the fact that
x P BX px, dq once d ° 0 shows that tBX px, dq : x P X ^ d P D x u is an open cover X. The
assumed compactness of X ensures existence of n P N, z0 , . . . , zn P X and d0 , . . . , dn P
p0, 8q such that
§n
X“ BX pzi , di q ^ @i “ 0, . . . , n : di P Dzi (25.9)
i“0
Define d :“ mini“0,...,n di and note that d ° 0. Then pick any x, y P X with r X px, yq † d
and observe that for
(
i :“ min j “ 0, . . . , n : x P Bpz j , dj q (25.10)
we have
rpy, zi q § rpx, yq ` rpx, zi q † d ` di § 2di (25.11)
and, since r X px, zi q † d § di , we have x, y P BX pxi , 2di q. The fact that di P Dzi then gives
` ˘ ` ˘
rY f pxq, f pyq § diam f BX pzi , 2di q † e. (25.12)
As x, y P X were arbitrary points with r X px, yq † d, this proves uniform continuity of f
on X. ⇤
The assumption of compactness of X can be relaxed and the conclusion strengthened
to the following form:
Lemma 25.7 Let X and Y be metric space with X totally bounded. The for all f : X Ñ Y:
f Cauchy continuous ñ f uniformly continuous (25.13)

Preliminary version (subject to change anytime!) Typeset: April 13, 2023

MATH 131BH notes 142

We leave the proof of this lemma to homework. A key reason for dealing with all of
these concepts comes in:
Theorem 25.8 (AC)(Existence/uniqueness of continuous extension) Let X and Y be met-
ric spaces, A Ñ X a non-empty set and f : A Ñ Y a function with Domp f q “ A. Assume:
(1) A is dense in X; i.e., A “ X,
(2) Y is complete
(3) f is uniformly continuous (or even just Cauchy continuous).
Then there exists a continuous function f¯ : X Ñ Y such that
@x P X : f¯pxq “ f pxq (25.14)
Moreover, f¯ is unique in the sense that if g : X Ñ Y is a continuous function with Dompgq “ X
and g “ f on A, then g “ f¯ on X.

Proof. Let pX, r X q and pY, rY q be metric spaces as is the setting of the theorem and assume
that A and Y are such that the conditions (1-2) above apply. Let f : A Ñ Y be a function
with Domp f q “ A and assume that f is Cauchy continuous (which is implied by uniform
continuity). We start with the construction of f¯.
Recall that f : X Ñ Y with Domp f q “ A is technically a relation G Ñ X ˆ Y of the
specific form (
G :“ px, f pxqq P X ˆ Y : x P A (25.15)
that we typically refer to as the graph of f . Next note that X ˆ Y is a metric space relative
to the metric ` ˘
r px, yq, px̃, ỹq “ r X px, x̃q ` rY py, ỹq (25.16)
This permits us to consider the closure G of G in X ˆ Y. We now claim:
Claim 1: G is the graph of a function
Indeed, suppose px, yq, px, ỹq P G. Then the fact that these are adherent points of G,
which is the graph of f , imply (by the AC) existence of sequences txn unPN , tx̃n unPN P AN
such that ` ˘ ` ˘
xn , f pxn q ›Ñ px, yq ^ x̃n , f px̃n q ›Ñ px, ỹq (25.17)
with the convergences in the metric space pX ˆ Y, rq. But this means that xn Ñ x
and x̃n Ñ x and thus also that the sequence tzn unPN defined by
z2n :“ xn ^ z2n`1 “ x̃n (25.18)
obeys zn Ñ x. But the fact that f is Cauchy then implies that t f pzn qunPN is Cauchy and,
since also f pxn q Ñ y and f px̃n q Ñ ỹ, shows
` ˘ ` ˘
rY py, ỹq “ lim rY f pxn q, f px̃n q “ lim rY f pz2n q, f pz2n`1 q “ 0 (25.19)
nÑ8 nÑ8
This proves that
@px, yq, px, ỹq P G : y “ ỹ (25.20)
¯
and so G is the graph of a function. Let us denote this function by f . Next we note:
Claim 2: Domp f¯q “ X
To prove this, let x P X. Then by (1) above x is adherent to A and so (by the AC) there ex-
ists txn unPN P AN such that xn Ñ x. The Cauchy continuity of f implies that t f pxn qunPN

Preliminary version (subject to change anytime!) Typeset: April 13, 2023

143 MATH 131BH notes

is a Cauchy sequence and so, by the completeness of Y assumed in (2) above, there
exists y P Y such that f pxn q Ñ Y. But then pxn , f pxn qq Ñ px, yq in X ˆ Y proving
that px, yq P G and thus x P Domp f¯q.
Claim 3: f¯ is continuous Since G Ñ G, the function f¯ is an extension of f to all points
of X. To prove continuity let x P X assume for contradiction that there is txn unPN P X N
such that
xn Ñ x ^ @n P N : rY p f¯pxn q, f¯pxqq • e (25.21)
By the construction of G, there exists tx̃n unPN P AN such that
` ˘
@n P N : r X pxn , x̃n q † 2´n ^ rY f¯pxn q, f px̃n q † e{2 (25.22)
But then x̃n Ñ x and, by Cauchy continuity of f and completeness of Y, there exists y P Y
such that f px̃n q Ñ y. It follows that f¯pxq “ y yet the second halves of (25.21–25.22),
rY py, f¯pxqq • e{2, and thus y ‰ f¯pxq. As px, f¯pxqq, px, yq P G, this contradicts that G is
the graph of a function. It follows that f¯ is continuous.
For the uniqueness it suffices to note that two continuous functions defined on a clo-
sure of the set A agree once they agree on A. We leave this detail to the reader. ⇤
We remark that uniform continuity allows us to avoid the use of the Axiom of Choice
in the proof of Claim 1 but does not seem to do that for the rest of the claim, due to the
fact that the completeness of Y requires working with Cauchy sequences to begin with.
To demonstrate the power of this result, let us prove one more time the existence of
exponential function:
Lemma 25.9 Let a ° 0 be real and recall that f pxq :“ a x is well defined for all x P Q. Then f
is Cauchy continuous and thus extends continuously to a unique continuous function on R (still
written as x ބ a x ).

Proof. Assume without (much) loss of generality that a • 1. We will rely on the fact that
a x`y “ a x ay which is checked algebraically for x, y P Q. This shows
ay ´ a x “ a x pay´x ´ 1q (25.23)
1
Next note that for each d P p0, 1q there is Npdq P N such that 1 ´ d † a † 1 ` d.
Npdq`1

Indeed, if note than either a ° p1 ` dq 1`n for all n • 0 or a † p1 ´ dq n`1 for all n • 0
which is impossible due to the fact that a is positive and finite.
Now let txn unPN P QN be a Cauchy sequence. This sequence is bounded by, say,
M P N. Pick e ° 0 and set d :“ ea´M . The fact that txn unPN is Cauchy implies
1
Dn0 P N @n, m • n0 : |xn ´ xm | † (25.24)
Npdq ` 1
But then for m, n • n0 the above shows
ˇ ˇ
ˇ f pxm q ´ f pxn qˇ “ a xm |a xn ´xm ´ 1| † a M d “ e (25.25)
thus proving that t f pxn qunPN is Cauchy. Theorem 25.7 now yields the desired unique
continuous extension of f to all of R. ⇤
To finish the discussion of uniform continuity, we introduce a couple of standard no-
tions that give a quantitative form of the dependence of d on e in (25.4).

Preliminary version (subject to change anytime!) Typeset: April 13, 2023

MATH 131BH notes 144

Definition 25.10 (Lipschitz/Hölder functions) Let f : X Ñ Y be a function between

metric spaces pX, r X q and pY, rY q. We say that f is Lipschitz continuous if
` ˘
Dl ° 0@x, y P X : rY f pxq, f pyq § lr X px, yq (25.26)
and, given a ° 0, is a-Hölder continuous if
` ˘
Dl ° 0@x, y P X : rY f pxq, f pyq § lr X px, yqa (25.27)
Note that Lipschitz continuity is a special case of Hölder continuity (corresponding
to a :“ 1) but the vernacular is used in this form throughout mathematics. Both concepts
give a quantitative form of dependence of d on e in (25.4); namely, e “ ld for (25.26) and
e “ lda in (25.27). A Lipschitz/Hölder continuous function is uniformly continuous
and thus also continuous.
The smallest constant l that one can put into these expression is sometimes called
the Lipschitz/Hölder norm. The cases with a P p0, 1s are most natural because x, y fiÑ
r X px, yqa is a metric. As we will see, this is all there is for functions R Ñ R; indeed, an
a-Hölder function R Ñ R (relative to Euclidean metric) for a ° 1 is necessarily constant.
Both properties above try to estimate the distance of function values by a function of
the distance of the arguments. This naturally leads to the following generalization:
Definition 25.11 (Modulus of continuity) Given a function f : X Ñ Y between metric
spaces pX, r X q and pY, rY q and a continuous non-decreasing map w : r0, 8q Ñ r0, 8q
with wp0q “ 0 and wptq ° 0 for t ° 0, we say that w is a modulus of continuity of f if
` ˘ ` ˘
@x, y P X : rY f pxq, f pyq § w r X px, yq (25.28)
So wptq “ lt in (25.26) and wptq “ lta in (25.27). The objective above is of course
to find a “best” function w that works for the given f ; if for each l P p0, 1q there is a
pair x, y P X with r X px, yq ° 0 and rY p f pxq, f pyqq • lwpr X px, yqq, then we sometimes
talk about the modulus of continuity of f . (Other interpretations of this term do exist,
though.) Another way to study the behavior of functions is using:
Definition 25.12 (Oscillation) Let f : X Ñ Y be a function between metric spaces pX, r X q
and pY, rY q and let A Ñ X. The oscillation of f on A is the function r fiÑ osc f pA, rq where
! ` ˘ )
osc f pA, rq :“ sup rY f pxq, f pyq : x, y P A ^ r X px, yq † r (25.29)
Here r is restricted to positive reals.
We note that all of these concepts are just ways to express properties of a function in
somewhat more condensed way.

Preliminary version (subject to change anytime!) Typeset: April 13, 2023