Remainder in Taylor Formula
Remainder in Taylor Formula
Remainder in Taylor Formula
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
Mathematical Association of America is collaborating with JSTOR to digitize, preserve and extend access
to The American Mathematical Monthly
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
THEOREM. (Taylor-Lagrange formula). Iff is continuous in [0, x], f (n- 1)(O) exists
and f (n)(t) exists in (0, x), then there exists a t in (0, x) such that
THEOREM 1. With notation as in the previous theorem, if f (n+ 1)(t) exists in [0, x
is continuous from the right at t = 0 and if f (n +1)(O) # 0, then
li1
x o+ x n +1
Remark. The assumptions in this theorem imply that the number t is uniquely
determined for x small enough.
*This work was supported in part by the Byron K. Trippet Research Stipend at Wabash College.
205
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
206
ESTEBAN
Proof
of
I.
POFFALD
the
[March
theorem.
In
(1),
Xn~~~n
n!
n!
f (n (T) f (na ) n + 1
Since f(n?+1) is right-continuous and nonzero at t = 0, the conclusion follows.
With notation as above, f(n)(4) n can be viewed as the error made in approximating f(x) by p-1(x). In view of Theorem 1, one would expect that replacing t
by x/(n + 1) and approximating f(x) by
X xn
f(x)
xn+2
A0(x)
2(n
1)
n + 1 n! ?2(n + ) (n + 2)!
xn
x n+ ( F2 x 1
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
but
1
?!
f-
.f n+lf(n+2)(t)[(+t) - 1![n]]d]
n+ (n + 2)! lX(t) t dt-)nl
( (x- )~ X t -n1 x
f (X)-Ao(x) = ! x f (n 2(1+lg(t)
2g(t) ndtn ? 1 -( + n 1 - t J
Then g(O) = 0 and g'(t) > 0 in (0, x], so that g(t) > 0 in that interval.
? 1)!for
j integrals to deduce that there are
we can apply the Mean-Value(n
theorem
n+1 ~ n?
numbers 41 and 42 in (0, x) such that
xn?2
where for a given n, the coefficients M21 and c2t depend only on i. Clearly, w
M= 1 and c0 = ?1
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
208
ESTEBAN
collecting
recursively from
I.
POFFALD
terms,
[March
we
see
that
cok?1
M2
c2k1
M2k
C2k
(3)
Then, one verifies readily that if f(n+2k+2) exists and is continuous at 0,
THEOREM 3. If f (n+2k)(t) exists and is continuous in [0, x], then there exists a
number t in (0, x) such that
x n+2k
the result holds for k > 1. Assume also that 0 < M2j < 1 for j = 1,..., k and
0 < c2j < 1 for j = 0,..., k - 1. Finally, as part of the induction hypothesis,
Therefore, 0 s< C2k < 1. This, together with the induction hypotheses, implies that
gk(O) = gk(l) = 0. Notice also that g"(X) exists for X + C2k and g9k(X) = gk-l(X)*
This implies that gk is convex in [0, C2k] and in [C2k, 11.
If C2k + 0,1, then gk'(X) > 0 for X + 0, C2k, 1 implies that g' is increasing in
[0, C2k] and in [C2k, 1]. We also have that g'(0) = g'(1) = 0, which implies that gk is
convex and increasing in [0, C2k] and it is convex and decreasing in [c2k, 1]. The
conclusion is that gk(X) > 0 in [0,11 with equality only for A = 0 and 1.
If C2k were 0 or 1, then we could carry out the same analysis as above, using
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
n+2j
Now, expand each term in the right-hand side into its Taylor polynomial of degree
In each of the integrals above, substitute t = Xx. After rearranging, this gives
formula for the sequences {c2j} and {M2j} given by (2) and (3). Numerical
computations suggest that for fixed n the sequence { c2j } is monotonically increasing while { M2 j } decreases to 0.
2. If f is analytic in a disk with center at the origin and radius R > 0, then one
can use the Cauchy-Hadamard estimates to deduce that the expansion
0O
1(x)
is
valid
in
n+2j
Pn-l(X)
the
disk
jxl
this expansion appears to be at least as large as that of the power series expansion of
f. (Cf. Remark 4 below.) The validity of this statement may depend on the behavior
3. For n = 1, both the statement and proof of our main result are simpler:
elementary combinatorial identities reveal that in this case C2j = 1/2 and M
1/4j satisfy (2) and (3). This suggests the following proof for n = 1: In the
Taylor-Lagrange formula
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
210
ESTEBAN
I.
POFFALD
[March
that
2k 2 X X 2j+1
f =0 (2j + 1)(! 2
2
(X]
2k1
+ f (2k+1) ( )
+(2k ? 1)!2
It follows that if
x)[X]2j1?
P be a point in the complex plane where f has a singularity and let 0 denote the
origin. Draw the line A/p perpendicular to OP at P. Then, a point x is in the Borel
polygon of summability if and only if x is in the half plane determined by A p and 0,
for all the singular points P of f. It then follows that x is in the interior of the Borel
polygon if and only if f is analytic at every point of the closed disk with diameter Ox.
Therefore, if x belongs to the Borel polygon of f, then
holds for all z such that Iz - (x/2)1 < Ix/21. As above, putting z = x and then
z = 0 and subtracting gives the result.
Example. If f(x) = ln(l + x), then
f (2k+1)(x) (2k)!
Applying (5), we obtain
00 1 x 2n+1
ln(l + x) = 2 O 2n + 1 2 +
The Borel polygon of f is the half plane Re(x) > -1. Therefore, this expansion is
valid for all x with Re(x) > -1. This should be contrasted with the circle of
convergence of the Taylor series expansion for f(x) about x = 0. Even in the
common region of convergence, this expansion converges faster. Since
1
we also obtain
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
which is valid for -1 < lm(x) < 1. In particular, for all x real, arctan x may be
computed using
oo1
arctanx= A 2 + 1An(x)
n=02n+
where
4x
Ao()
X2'
and
arctan2= k(l[kl + 8
and
16 00 1 f2n+1 (4)j
5 n=0 =O 2n + 1 2j +1 52n
the local discretization error). We will omit the derivation of the error term. Not
however, that in each case the error term can be derived using the main ideas in
(i). The Modified Euler's Method: for n = 1 and k= 1, we apply (4) (in the
interval [xi1, xi+]) to obtain the approximation Yj1? = Yj-i + 2hfj with E =
ing, we get Yj+1 = Yj-2 + (3h/4)[fj-2 + 3fj1 with the error term E =
This approximation does not appear to be one of the standard multistep methods
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
212
ESTEBAN
I.
POFFALD
[March
we obtain
~~8
which
is
the
31k3)
well-known
Sim
(iv). Now, we use (4) with n = 1, k = 2 and the interval [xj_3, xj+11 to get
(4h)3
(v). Again, we take n = 1 and k = 2, but we use the interval [xj>1, xj+11. Thi
using this formula in conjunction with the one derived in (iv), we obtain a
well-known predictor-corrector method (Milne's method):
4h
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms
fbf( dt (b-a)
Now, the error of this approximation is -f"(()(b - a)3/12, for some ( in (a, b). It
can then be shown that if f(3) is continuous and nonzero at a, then, as b > a+, c
approaches the midpoint of [a, b], that is,
(c-a) 1
b-a+ (b-a) 2
1. R. Ellis and D. Gulick, Calculus with Analytic Geometry, 3rd ed., H. B. J. Publi
2. J. M. H. Olmsted, Real Variables, Appleton-Century-Crofts, Inc., 1959.
This content downloaded from 190.114.221.112 on Tue, 25 Oct 2016 21:34:12 UTC
All use subject to http://about.jstor.org/terms