MEAN VALUE THEOREMS FOR

VECTOR VALUED FUNCTIONS

by ROBERT M. McLEOD
(Received 28th April 1964)

1. Introduction
The object of this paper is to give a generalisation to vector valued functions
of the classical mean value theorem of differential calculus. In that theorem
we have

for some c in the open interval ]a, 6[ w h e n / is a real valued function which is
continuous on the closed interval [a, b~\ and differentiable on the open interval.
The counterpart to (1) when / has values in an n-dimensional vector space
turns out to be
f(b)-f(a) = (b-a) £ XJ'{ck) (2)
k = 1
n
where ck e ]a, b[_, 0^Xk, and £ ^* = *•
The key to the algebraic conclusion (2) lies in geometric ideas. Convex
sets play a basic role in the statements and proofs of the mean value theorems
developed in § 3. Let us recall a few definitions here.
Definitions, (a) Let x and y be elements of a vector space F. The segment
joining x and y is the set of elements Xx + {\—X)y for all X such that O g l ^ 1.
(b) A subset A of F is convex provided the segment joining each pair of points
in A is contained in A. (c) If A is any subset of F, the convex cover HA of A
is the minimal convex set containing A, that is, the intersection of all convex
sets containing A. (d) When F is a topological vector space the closed convex
cover KA of A is the minimal closed convex set containing A.
The equation (2) is the end of a three step development. The first step
is to prove that (f(b)—f(a))/(b—a) is in the closed convex cover of the set of
values of the derivative. At this stage / may have values in any locally convex
topological linear space. Then under additional hypotheses, for example
that the space is finite dimensional, it is also true that (f(b)—f(a))l(b—a) is
in the convex cover of the set of values of the derivative. The algebraic form
of this conclusion is
Kb) -f(a) = {b-a) "Z XJ'(ck) (3)
k = 1

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 198 R. M. McLEOD
when / has values in n-dimensional space. The final step is to show that the
number of terms in the sum can be reduced from n +1 to n.
In the first two steps it is not necessary to require the existence of a derivative
throughout ]a, b\_. A right-hand • or left-hand derivative can take the place
of the ordinary derivative and there may be a countable subset of ]a, b[ in
which no derivative of any sort need exist. The transition from (3) to (2) is
possible under two additional assumptions: (a) The derivative / '(0 exists for
all t in ]a, 2>[—M where M is a set having at most n— 1 points, (b) When
n^.2 the derivative/' is continuous from the right at every point of ]a, b\_ — M
or continuous from the left at every point of ]a, b\_—M.
The condition (a) is essential for (2). In the case n = 1, let / b e the real
valued function given by f(t) = \t\ for — l^t^l. Now (3) does hold.
Indeed, / ( l ) - / ( - l ) = 0 and 0 = 2(A,/'(c 1 ) + A 2 /'(c 2 )) with At = X2 = i,
— l < C ! < 0 , and 0 < c 2 < l . But (2) is impossible for all c in ] - l , l[-{0}.
Similar examples can be given of functions having values in En for any n.
Mean value theorems of other kinds have already been given for vector
valued functions. Bourbaki (3) and Dieudonne (4) give one in which the
conclusion
\\f(b)-f(a)\\^g(b)-g(a)
is drawn from the hypothesis ||/'(0||^ff'(0 f° r au" l m ]a> b[ — M where M
is a countable set. See Theorem B in the next section for a full statement of
a theorem of this type. It will be convenient to refer to theorems of this kind
as increment theorems.
Aziz and Diaz have given a third kind in (2). It is possible to generalise
their result by using the version (3) of the mean value theorem for real valued
functions. Details are given in Theorem 5.
§ 2 is preparatory. The main results are in § 3 and Theorem 1 is chief
among them. Thefinitedimensional case comes in Theorem 3 and Theorem 4.

2. Increment theorems
It is usual to prove that a function with a positive derivative is increasing
by using the mean value theorem. Here we shall reverse the procedure. From
a sufficient condition for a function to be increasing we can obtain mean
value theorems for vector valued functions. From the same starting point
and by a very similar method it is also possible to prove increment theorems
for vector valued functions. In fact, for economy of exposition it is advantage-
ous to derive the mean value theorems from an increment theorem for vector
valued functions.
The increasing function theorem that is suited to the present purpose,
Theorem A below, is a special case of theorems given by Aumann (1, p. 222)
and Gal (6, p. 310). The principal simplification is in replacing a semi-
continuity condition on / by the assumption of continuity.
Definitions, (a) Let / be a continuous function from an interval [a, ft]
into a topological linear space F over the real field R. Let / e [a, ft[. An

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 THEOREMS FOR VECTOR VALUED FUNCTIONS 199

element y of F is a right-hand derivative value o f / a t t if there exists a sequence

t\, t2, t3, ... decreasing to t such that
y = um / i
n-co *„ —
R
(b) For a real valued function/let £) /be the upper right derivative, i.e.

S —f

(£>"/ may take on the values — oo and + oo but it will be assumed throughout
that right-hand derivative values are finite when / is real valued.)
Theorem A. Let f be a continuous function from [a, ft] into R. Let M be
a countable subset of ]a, ft[ and RDf a real valued function such that RDf(f)
is a right-hand derivative value of f for all t in ]a, b\_ — M. If there is a set N
of Lebesgue measure zero such that 0^DRf(t) for all t in ]a, b\_— N, then f is
an increasing function on [a, b~\. Moreover, if 0 <DRf(t) for at least one t in
-\a,blthenf(a)<fQ>).
Definition. A sublinear funtional on a linear space F is a function p from
F into R such that
p(Xx) = Xp(x) and p(x+y)^p(x)+p(y)
for all A^O and all x and y in F. (p is a linear functional when the first relation
holds for all A in R and equality holds for all x and y in the second relation.)
The following theorem differs in detail from those stated by Bourbaki
and Dieudonne. The proof is the one suggested in (4), p. 156, problem 6.
Theorem B. Let F be a topological linear space. Let p be a continuous
sublinear functional on F. Let f and g be continuous functions from [a, b~\ into
F and R respectively. Suppose there is a countable subset M of~\a, b\_ such that,
for each t in ]a, ft[ —M, / and g have right-hand derivative values RDf(t) and
RDg(f) associated with the same sequence decreasing to t. Suppose also there
is a set N of Lebesgue measure zero such that M^N^]a, b[ and
p{RDf(t))^RDg(f) forallt£~\a,b\_-N. (4)
Then
pU{d)-f(c))^g{d)-g(c)
for all [c, </] contained in \a, b~\. Moreover, if inequality holds in (4) for at least
one value oft, then p{f(b)-f(a))<g(b)-g(a).
Proof, (a) Given x0 e F, let Fo = {Ax0 : A e R}. On the subspace Fo
define u0 by MO0**O) = *P(xo)- Then uo(x)£p(x) for all x in Fo. The Hahn-
Banach theorem (7, p. 42) provides an extension u of w0 to all of F such that
u is a linear functional and u(x)^p(x) for all x in F. Moreover u is continuous
at 0 since p is continuous and
| u(x)\ g max {p(x\ p(- x)}
for all x in F. Therefore u is continuous on F.

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 200 R. M. McLEOD
(b) Take x0 = f(b)-f(a). Set h(i) = g(t)-u(f(t)). Since the right-hand
derivative values of / and g are determined by the same sequence decreasing
to t, RDg{i)-u(RDf(t)) is a right-hand derivative value of h. Call it RDh(t).
Then
0^RDg{t)-p(RDf(t))^RDhfj) for all t e]a, ft[-N.
R
Since RDh{i)f^D h(t) Theorem A is applicable. Then h{a)-^h{b) or
ui:f(b))-uU(d))^g(b)-g(a).
But
<Ab))-ttM) = «(M-M) = P(f(b)-f(a))
and the first conclusion is established for c = a and d = b. The restriction
of/to [c,rf]c [ a > ft] satisfies the same hypotheses. The first conclusion follows.
If inequality holds in (4) then 0<RDh(t)^DRh(t) and the final conclusion
follows from the last assertion of Theorem A.
Note that existence of a right-hand derivative for one of/andgr and existence
of a right-hand derivative value for the other suffices for the existence of right-
hand derivative values on a common sequence.
One important case of Theorem B occurs when p is a norm on F. But
for application to the proofs of mean value theorems it is important that p
can be a linear functional also.

3. Mean value theorems

A few more basic ideas now come into use. (a) Subsets S and U of a topo-
logical linear space/*1 are a closed half-space and a closed hyperplane, respectively,
when
S ={xsF: u(x)^a} and U = {x e F : u(x) = a}
for some non-zero continuous linear functional u and some real a. (b) F
is a locally convex topological linear space when each neighbourhood of 0
contains a convex neighbourhood of 0.
Since norms are sublinear, the balls {x s F : || x || <a} are convex for all
a>0. Thus every normed space is locally convex. In particular the Euclidean
space En of H-tuples {xu x2, •••, xn) of real numbers with

is locally convex.
Theorem 1. Let F be a locally convex topological linear space. Let f be
a continuous function from [a, ft] into F. Let M be a countable subset of~\a, ft[
and RDf a right-hand derivative value function for f on ]a, b[_ — M. Let N be
a set of Lebesgue measure zero such that M^N^~]a, ft[. Set
D= {RDf(t):te-\a,bl-N}
and
Q = {(f(d)-f(c))/(d-c) : aSc<d^b}.
Let KD and KQ be the closed convex covers of D and Q. Then KD = KQ.

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 THEOREMS FOR VECTOR VALUED FUNCTIONS 201
Proof. KD^KQ is elementary. RDf(i) is a limit point of Q for all t in
]a, b[-N. Thus £>£closure of Q^KQ. Now KD^KQ follows from the
definition of the closed convex cover.
If KD = F then trivially KQ^KD. Suppose KD # F. In a locally convex
space each closed convex set is the intersection of all the closed half-spaces
containing it. (See (7) p. 151, problem 5.) Let KD^S where
S = {xeF : w(x)ga}.
Set g(i) = at. Then Theorem B is applicable with u in place of p. The con-
clusion is u{f(d)—f(c))^ad—<x.c for all [c, d~\ contained in [a, b~\. Thus u(q)^a
where q = (f(d) —f(c))l(d— c). This is equivalent to Q £ S. Since S is an arbitrary
half-space containing KD the inclusion Q^KD follows. Therefore KQ^KD
and the proof is complete.
Since there may be more than one right-hand derivative value for each t
the function RDf is not unique. Also many choices may be made for N. Never-
theless the equality KD = KQ guarantees the uniqueness of KD since Q depends
neither on RDf nor on N.
In some circumstances it is possible to conclude further that
(f(b)-f(a))l(b-a)eHD.
The proof turns on HD having an interior point. A convex subset of F may
fail to have interior points merely because it is contained in a hyperplane
and should be seen in a different context. The next theorem is formulated so
as to take care of this difficulty.
Definition. A linear variety in F is a translation x0 + F0 of a subspace Fo
of F by an element x0 of F.
xo + Fo is closed if and only if Fo is closed. The intersection of any collec-
tion of closed linear varieties with a common point is again a closed linear
variety.
Theorem 2. Let F and f be as in Theorem 1. Let HD be the convex cover
of D. Let V be the minimal closed linear variety containing HD, i.e. the inter-
section of all closed linear varieties containing HD. If HD has an interior point
in the relative topology of V, then (f(b)—f(a))l(b — a) is an interior point of HD
relative to V.
Proof. There is no loss in supposing O e D since / can be replaced by the
function t-*f(t) — tRDf(t0) for some t0 in ]a, b[ — M if necessary. With this
agreement V becomes a closed subspace of F. Any continuous linear functional
on V can be extended to a continuous linear functional on all of F. (See
(7) p. 148.) Therefore the minimality of V implies that HB is not contained
in any closed hyperplane in V.
In any topological linear space a convex set having an interior point has
the same interior as its closure. (See (7) p. 133, problem 2.) Since V is closed
the closure of HD in V is KD. By Theorem 1, (f(b)-f(a))l(b-a)eKB. It is
enough to eliminate the possibility that (f(b)-f(a))/(b-a) is a boundary
point of KD relative to V to complete the proof.

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 202 R. M. McLEOD
A closed convex set with interior points has a supporting hyperplane at
each of its boundary points. (See (7) p. 142, problem 2.) That is, if KD is
the set and y is a boundary point of KD, there is a closed half-space S bounded
by a hyperplane U such that KD^S and yeU. Let S = {x e V : u(x)^<x}
and U = {x e V : u(x) = <x) where u is a continuous linear functional on V.
Since HD is not contained in any closed hyperplane in V there is at least one t
in ]a, b[-N such that u(RDf(t)) <a. Apply Theorem B with u in place of p.
From the final assertion of Theorem B,
u(f(b)-f(a))<ab-<xa.
Therefore (f(b)—f(a))/(b — a) is distinct from the boundary point y of KD
lying in U. This completes the proof since y may be any boundary point of KD.
Corollary. Under the hypotheses of Theorem 2 there is an integer m and
numbers ck and Xk,k = 1,2, ..., m, such that

k = 1
m
ck e]a, b[—N, 0^Xk, and Y At = 1.
k = 1
Proof. The convex cover of any set A consists of all sums of the form
YJ Kxk where xk e A, 0^Xk, £ Xk = 1, and m ranges over all the positive
k = 1 * = 1
integers. (See (7) p. 131.) The conclusion follows. (For convenience call
such a sum a convex combination.)
More can be said about m when Fis finite dimensional. Also the hypothesis
that HD has an interior point relative to V becomes superfluous.
If A is a subset of En, each element of HA can be expressed as a convex
combination of elements of A having n + 1 terms. Moreover, if A has at
most n connected components, the convex combinations having n terms yield
all of HA. (See Eggleston (5) p. 35.) Simple examples show no further
reduction of m is possible.
Theorem 3. Replace F by En in the conditions of Theorem 1. Then Q c HD
and i.f(b)—f{dj)l{b — a) is an interior point of HD relative to the minimal linear
variety V containing HD. Moreover,
f(b) -f(a) = (b-a) " l XkRDf(ck) (5)
k= 1
n+1
where cke~]a, b[ — N, 0^1k, and £ Xk = 1. Finally, if D has at most n
k = 1
connected components the sum in (5) may be replaced by a sum having n terms.
Proof. Every convex set in En has an interior point relative to the minimal
variety containing the set (5, p. 16). Moreover, every variety in En is a closed
variety. Thus Theorem 2 can be applied to the restriction o f / t o each sub-
interval [c, d]. Then, with obvious notation,
(f{d) -f(c))/(d- c) e HD{c, d] £ HD

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 THEOREMS FOR VECTOR VALUED FUNCTIONS 203
and thus Q^HD. The remaining conclusions follow directly from Theorem 2,
its corollary, and the facts about convex sets in En set down in the paragraph
preceding Theorem 3.
It does not follow that Q is contained in the interior of HD relative to V.
This is because the minimal variety V\c, d~\ associated with the restriction of
/ t o [c, d~\ need not be V. Let/be the function from [—1, 2] into R given by
f(t)=\t\, - l ^ / ^ 2 . In this example HD = {-\, 1] and V = R, but
HD[0, 2] = {1} and V[0, 2] = {1}. Now (/(2)-/(0))/2 = 1 and 1 is not an
interior point of [—1, 1] relative to R. On the other hand 1 is an interior
point of {1} relative to {1}.
Theorem 4. Let f be a continuous function from [a, b~\ into En. Suppose
there is a finite set M having at most n—\ points such that f '(t) exists for all
t in ~\a, b\_ — M. When n^2 suppose also that f is continuous from the right
on "]a, Z>[ — M or continuous from the left on ]a, £>[ — M. Then (f(b) —f(a))/(b — a)
is equal to a convex combination of n values off.
Proof. In the case n = 1 the derivative/' exists throughout ~\a, b\_. The
theorem is just the classical mean value theorem in this case.
Suppose now that n^2. There is no loss of generality in supposing that
/ ' is right-hand continuous on ]a, &[ — M and in supposing/(a) = f(b). From
Theorem 3,0 = (/(&)-f{a))j(b-a) = £ Xkf\ck) with ck e]«, b{-M, OgAfc,
*= I
n+l
and Y, ^-k = 1- If any proper subset of {/'( c i)> •••> f'(cn+i)} nas a
convex
k = 1
combination equal to 0 the proof is complete. Suppose not. Then, in particular,
hf'(ck) # 0 for A: = 1, 2, ..., n + l . There are at most n open intervals
forming ]ct, 6[—M. Therefore some interval contains two or more of the
numbers ck. Suppose the notation chosen so that cl and c2 are in the same
interval with c1<c2.
Let xk = -f'(ck), k = 1, 2, ..., n + l , and A = {x2, ..., xn + l}. Let C
be the cone subtended at the origin by the convex cover of A. Then C consists
n+l n+- 1
of all points £ nkxk with/z k ^0for2^fc:gn+l. Since YL 4tX* = 0and Xlxl # 0
k = 2 k= 1
we see that — xx e C. We have assumed that 0 does not belong to the convex
cover of any proper subset of {*,, x2, -.., * n +i}- Consequently — x2 $ C.
n+ 1
v x
A is a linearly independent set. Suppose Y k k = 0 with at least one
k = 2
vk not zero. Then

for all real /. Let E be the set of all real t such that l.k + tvk^0 for k = 2,
..., n + l. Then 0 e £ since every kk is positive. There is no loss in generality
in supposing some vk is negative. Then E is bounded above. Let t0 = sup E.

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 204 R. M. McLEOD
By continuity t0 e E and kk + tovk = 0 for at least one integer k. Let

t £ ( * o
k= 2
n+l
Then v^A t >0 and v " 1 ^ * ^ £ V^-Mo* 1 *)** is a convex combination
fc = 2
of a proper subset of {xl5 ..., x n+1 } which is equal to zero. This contradicts
our assumption.
The linearly independent set A contains n elements. Thus any element y
n+l
in En has one and only one representation £ /****• Furthermore y is a
boundary point of C if and only if f-ik^0 for 2^Ar^«+ 1 and nk = 0 for at
least one k.
Let J = sup{f : / ' ( 0 i , t])^C}. Since /'(<a) e C and f'(c2)$C it follows
that ct^d^c2. Clearly/'([f1!, rf[)^C C is a closed convex set. From
Theorem 1, (f(t)-f(d))/(t-d)e C for all f in [c 1; </[. T h u s / V ) e C and
d<c2- On the other hand there exists a sequence fl5 ?2> '3. ••• decreasing to rf
such that / '(tk) $ C. Since / ' is right-hand continuous at d it is true that
n+l
f'(d) belongs to the closure of the complement of C. Therefore/ '{d) = X ^*x*
k = 2
with (ik^0 for all k and nk = 0 for at least one fc. This equation can be
rewritten as

* 2
/ n+l \
Multiplication by 1/1 1 + Y nk) produces the desired convex combination
V *=2 /
of n values of/' equal to 0.
The condition for reduction of the number of terms in the convex com-
bination given in Theorem 4 is not a consequence of the one given in Theorem 3
when n > 1. To establish this it suffices to give an example of a function /
satisfying the hypotheses of Theorem 4 for which f'Qa, b[_-M) has more
than n components. The function / whose construction follows is continuous
on [0, 1/48], differentiable on ]0, l/48[, and its derivative is right-hand con-
tinuous on ]0, l/48[. It has its values in E2. Since E2 is a linear subspace of
En for n>1 we can also consider/as a function having values in £„ for all
«^2. The components of/'(]0, l/48[) are the same whether this set is con-
sidered as a subset of E2 or as a subset of En for n > 2. The collection of
components is countably infinite. Thus this one example serves for every n
greater than 1.
Let a be a real constant. Set g(0, a) = (0, 0) and g(t, a) = r2(cos alt, sin ajt)
for t i=- 0. The map t-+g'{t, a) carries each interval \r, 0] onto {(0, 0)}uS
where 5 is a spiral which approaches the circle centred at (0, 0) with radius a
from the outside. Set
h{t) = g(t, a)-g(r, a) + (t-r)[y-g'(r, d

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 THEOREMS FOR VECTOR VALUED FUNCTIONS 205
where x and y are elements of E2. Then h'(t) = g'(t, a)+y—g'(r, a) and the
image of [r, 0] under W is {y—g'(r, a)}uT where T is the translate of S which
begins at y. Since || g'(t, a)\2 = 4t2 + a2 the diameter of S, and of T, is less
than 2(4r2+a2)i. We want to choose a sequence of functions hk so that the
spiral Tk+i starts at the centre of the limiting circle of Tk and stays inside that
circle. This will be true if yk+l = yk—g'{rk, ak) and 2(4r^+1 +al+l)i<ak.
We may choose ak = 4~k, rk = — 4~k~2, and j t = (0, 0). To complete the
determination of hk we require that hk + 1(rk+1) = hk(0). Equivalently

The choice xl = (0, 0) completes the definition of all the functions.

Now let tk — tk^l = —rk and t0 = 0. Then

lim tk = f 4-k~2 = 1/48.

*-oo k = 1

Define/on [0, 1/48] as follows:

f(t) = hk(t-tk) for tk.i^t<tk,
f(l/4Sy= lim x,.
Jfc-»CD

Then/ satisfies the hypotheses of Theorem 4. B u t / ' maps ]0, l/48[ onto
an infinite union of spirals and each spiral is a connected component of the
union.
In the case n = 1 Theorem 4 is a consequence of Theorem 3. Indeed,
/ '(0 exists throughout ]a, b[ hence, by the Theorem of Darboux, / ' has the
intermediate value property. Thus f'(\a, Z>[) is an interval and the number
of components is 1.
We used right-hand derivative values in Theorems 1, 2 and 3. Left-hand
derivative values can be used equally well. In fact the corresponding theorems
are corollaries of the ones given. Indeed, if/ has left-hand derivative values
LDf(t) for all t in ]a, b\_-M and
9(0 = -/(-'), -b£t£-a,
then g has right-hand derivative values such that RDg(t) = LDf(—t). More-
over (g(d)-g(c))l(d-c) = (f(-d)-f(-c))l(-d+c). Thus application of
Theorems 1, 2 and 3 to g yields corresponding theorems for / in which D,
HD and KD are formed from the values of LDf. Furthermore, if/ has both a
left-hand and a right-hand derivative value function, the two sets KD formed
from LDf and RDf are equal since both are equal to KQ.

4. Applications
Many of the theorems of elementary calculus which are usually derived
from the classical mean value theorem can just as well be got from an increment
theorem like Theorem B. Dieudonne has given an excellent account of their
extensions to Banach spaces in (4) pp. 156 ff. We will avoid duplicating them

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 206 R. M. McLEOD
here and concentrate on some applications which come more readily from mean
value theorems.
The first is a generalisation of the theorem of Aziz and Diaz (2, p. 261).
Theorem 5. Let Fbe a topological linear space. Let f be a continuous function
from \a, b~] into Fhaving a right-hand derivative value function RDf on\a, b[ — M,
where M is countable. Let p be a continuous sublinear functional on F. If N
is any set of Lebesgue measure zero such that Af£JV£]a, b\_, there exists
c e la, 6[ — N such that
p(f(b)-f(a))^p(RDf(cy).
Proof. From part (a) of the proof of Theorem B there is a continuous
linear functional u on .Fsuch that p(f(b) —f(a)) = u{f(b)—f(dj) and u(x)^p(x)
for all x in F. Consider the real valued function h given by h(t) = u(f(t))
for all t in [a, b~\. Since u(RDf(t)) is a right-hand derivative value for h, Theorem
3 is applicable. Thus
uU(b))-u{f(a)) = X1u(RDf(c1)) + X2u(RDJ(c^)
an
with 0^A 1; 0^A2> d At +A2 = 1. Let c be that one of ct and c2 giving the
larger value of u(RDf(t)) for t = ct and t — c2. The left-hand side of the
equation is p(f(b)—f(d)) and the right-hand side is dominated by u(RDf(cJ)
and hence by p(RDf(c)). This completes the proof.
Theorem 6. Let f and F be as in Theorem 1. Let Fo be a closed subspace
of F. The set of values of f is a subset of the variety x0 + Fo if and only if D^F0
andf(t) £ x0 + F0for some t in [a, b].
Proof. Suppose the values of/all lie in xo+Fo. Clearly Q^F0. Since
Fo is closed, .enclosure of Q^F0. Conversely, suppose D^F0 and
f(t0) exo+Fo. Since Fo is closed and convex, KD^F0. By Theorem 1,
Q<=KD. Thus
fit) = fih)+{t- to)y, y e f0,
for each / in \a, b~\. But/(?0) = xo+yo with y0 e Fo. Since Fo is a subspace,
y0 + (f - to)y e Fo. Thus f(t) exo + Fo for all t in [a, b~\.
When F = E3 and Fo is a plane through the origin the geometric content
of Theorem 6 is that / is represented by a plane curve in a plane parallel to
Fo if and only if the values of the function RDf lie in the plane Fo. Note that
it is not merely a matter of the values of RDf lying in a plane. For example, let
f(t) = (cos t, sin t, i), 0^t^2n.
The curve which represents the derivative / ' is a plane curve, a circle in the
plane x3 = 1, and/is represented by a spiral which lies in no plane.
We continue with a geometrical point of view and consider closed curves
in E2. It seems intuitively clear that a continuously turning tangent vector
on a closed curve must turn through at least n radians in one circuit of the
curve. It is possible to make Theorem 4 yield precise information to this
effect.

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 THEOREMS FOR VECTOR VALUED FUNCTIONS 207
Theorem 7. Suppose f is a continuous function from \a, b~] into E2 having
a derivative f' {t) for all t in ]a, b[_ — {c}for some c in ]a, Z>[. Suppose f is right-
hand continous on ]a, &[ — {c} or left-hand continuous on ]a, b\_ — {c}. Suppose
also /'(/) # 0 for all t in ]a, 6[ — {c}. If f(a) = f(b) there are numbers ct and
c2 in ]a, b[ — {c} and a positive number n such that /'(ci) = — nf'(c2). {In
geometrical language, when the curve is closed there is at least one pair of oppositely
directed forward tangent vectors.)
Proof. By Theorem 4,
M-M = 0-<O(Ai/'(ci)+A 2 /'(c 2 ))
with 0^A 1; 0^A2 and Al + 12 = 1. Since f(b) = f(a) the convex combination
is 0. Since neither of the values of the derivative is zero neither kx nor X2 is
zero. Thus 0<At and 0<A2. The conclusion follows with \i = A2/At.
If/maps [a, a^\, \au a 2 ] and [a2, b~\ linearly onto the three successive sides
of a triangle it fails to satisfy the hypotheses of Theorem 7 only in having
no derivative at two points in ]a, Z>[ instead of at one point. The conclusion
fails also.
We turn now to some applications of the mean value theorems to functions
from one norrried space to another. The definition of the derivative, or total
differential, is needed in order to state them.
Definition. Let / be a function from an open set G in a normed space E
into a normed space F. f is said to be differentiate at y in G if there is a
continuous linear mapping u from E into F such that
lim \\f(x)-f(y)-u(x-y)\\l\\x-y\\ =0.

The linear mapping u is called the derivative of/ at y and is denoted by f'(y).
Its value at any x in E is denoted by f'(y) . x.
If g is a function from [a, b~\ into G having a derivative value RDg(t) the
chain rule for the composite function fg reads
RDfg(t)=f'(g(t)).RDg(t).
Definitions, (a) Let x and y belong to the connected open set G. Let
P(x, y) denote the collection of all functions g from [0, 1] into G such that
g(0) = x, g{\) = y, g is continuous on [0, 1], and there exists a countable
set M such that g has a right-hand derivative value function RDg on ]0, 1 [ - M.
(b) Suppose/is a differentiable function from G into F. Let K(x, y; g) denote
the closed convex cover of D(x, y; g) where, for fixed g in P(x, y),
D(x, y; g) = {f'(g{t)). RDg(t) : t e]0, 1[-M}.
Also let
K(x, y)= f) x<>> y'' 9)-
geP(x.y)
Theorem 8. Let f be a differentiable function from the connected open set G
in the normed space E into the normed space F. Then
f(y)-f(x)eK(x,y)

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 208 R. M. McLEOD
for all x and y in G. Moreover, if F = En, f(y)—f(x) is an interior point of
K(x, y; g) relative to the minimal variety containing it for each g in P(x, y).
Also f(y)—f(x) is a convex combination of n+1 elements of D(x, y; g).
Proof. This is a direct application of Theorems 1 and 3 using the chain
rule noted above.
A global sufficient condition for / to be one-to-one on G comes out of
Theorem 8 at once.
Theorem 9. Under the conditions of Theorem 8, / is one-to-one on G if
0 $ K(x, y) for all x and y in G. When F = En it is also sufficient that for each
x and y in G there exists g in P(x, y) such that 0 is not a relative interior point
ofK(x,y;g).
(A local sufficient condition for / to be one-to-one can be obtained in a
useful form from Theorem B. See (4) p. 156 and p. 268.)
Now let/be a holomorphic complex valued function defined on a connected
open set G in the complex plane. There are two concepts of derivative for /,
both denoted by / '(y), which are related so that / '(y) . x can be read correctly
as the value of the linear mapping / '(y) at x or as the product of the complex
number / '(y) and the complex number x. Consequently, Theorems 8 and 9
are applicable to holomorphic functions. Since / ' is continuous we can use
Theorem 4 to get the following fact also.
Theorem 10. Suppose f is a holomorphic function defined on a connected
open set G in the complex plane. Ifzt andz2 are points in G such that the segment
joining them is also in G then
f(z2)-f(zi) = fe-^XWOiHWM)
for some wt and w2 on the segment joining zy and z 2 and some Xt and X2 such
thatO-^ly, 0gA 2 andXy+kz = 1.
The convex combination in this mean value equation cannot be replaced
by a value / '(w) in general. The example f(z) = ez with z2 = zt + 2ni shows
this since/(z 2 )-/(z!) = 0 but (z2-z{)j'(w) = 2niew # 0 for all w.
These facts throw light on why the Jacobian condition |/'(0l 5^0 is a
global sufficient condition for a real valued function of a real variable to be
one-to-one on an interval while the counterpart |/'( z )l T6 0 is only a local
sufficient condition in the case of holomorphic functions.

5. Examples
There are standard examples which show that the countable set M cannot
be replaced by a set of measure zero in Theorem A and its consequences.
Let Tbe the Cantor ternary set in [0, 1]. Let k be the continuous function
on [0, 1] such that A;(0) = 1, ifc(l) = 0, k(t) = \ for \<t<\, k{t) = J for
£ < * < ! , k(t) = | for i < t <f, etc. Then k'(i) = 0 for all t in [0, 1] - T. Since
T is a set of Lebesgue measure zero this example shows that it is not possible

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press

 THEOREMS FOR VECTOR VALUED FUNCTIONS 209
to replace the countable set M by a set of Lebesgue measure zero in Theorem A
and Theorem 1.
Let k2(t) = (/, k(t)) for all / in [0, 1] and k3(t) = (t, k(t), 1) for 0 ^ / ^ 1,
£3(0 = (U 0. Kt— 1)) f° r 1 <t^2. In both of these examples D is contained
in a one dimensional subspace if we use only values of t outside the Cantor
sets. On the other hand the values of k2 span E2 and the values of k3 span E3.
These examples contrast strongly with Theorem 6.

REFERENCES
(1) GEORG AUMANN, Reelle Funktionen (Berlin, 1954).
(2) A. K. Aziz and J. B. DIAZ, On a mean value theorem of the differential
calculus of vector-valued functions, and uniqueness theorems for ordinary differential
equations in a linear-normed space, Contributions to Differential Equations, 1 (1963),
251-269.
(3) N. BOURBAKI, Elements de Mathimatique, Livre IV, Fonctions d'une variable
reelle, Chaps. 1, 2, 3 (2nd ed., Paris, 1958).
(4) J. DIEUDONNE, Foundations of Modern Analysis (New York, 1960).
(5) H. G. EGGLESTON, Convexity (Cambridge, 1958).
(6) ISTVAN S. GAL, On the fundamental theorems of the calculus, Trans. Amer.
Math. Soc. 86 (1957), 309-320.
(7) ANGUS E. TAYLOR, Introduction to Functional Analysis (New York. 1958).

AMERICAN UNIVERSITY OF BEIRUT

BEIRUT, LEBANON

E.M.S.—P

https://doi.org/10.1017/S0013091500008786 Published online by Cambridge University Press