Math 133 Taylor Series Stewart §11.

10

Series representation of a function. A series writes a given complicated quantity

as an infinite sum of simple terms. To approximate the quantity, we take only the
first few terms of the series, dropping the later terms which give smaller and smaller
corrections. In this section, we finally develop the tool that lets us do this in most
cases: a way to write any reasonable function f (x) as an explicit power series, a
kind of infinte polynomial. This will allow us to compute outputs of the function by
plugging values of x into the series.
Our functions must behave decently near the center point of thePdesired power
series. We say f (x) is analytic at x = a if it is possible to write f (x) = ∞ n=0 cn (x−a)
n

for some coefficients cn , with positive radius of convergence. In practice, any formula
involving standard functions and operations defines an analytic function, provided
the formula gives real number values in a small interval around√x = a. For example
1
x−a is not analytic at x = a, because it gives ±∞ at x = a; and x−a is not analytic
at x = a because for x slightly smaller than a, it gives the square root of a negative
number.∗
Taylor Series Theorem: Let f (x) be a function which is analytic at x = a.
Then we can write f (x) as the following power series, called the Taylor
series of f (x) at x = a:
f 00 (a) f 000 (a) f 0000 (a)
f (x) = f (a)+f 0 (a) (x−a)+ (x−a)2 + (x−a)3 + (x−a)4 +· · · ,
2! 3! 4!
valid for x within a radius of convergence |x−a| < R with R > 0, or
convergent for all x.
If we write the nth derivative of f (x) as f (n) (x), this becomes:
∞
X f (n) (a)
f (x) = cn (x−a)n with coefficients cn = .
n!
n=0

warning: The coefficients are constants with no x, so c1 = f 0 (a), not f 0 (x).

Proof. By hypothesis f (x) is analytic, so f (x) = ∞ n
P
n=0 cn (x−a) Pfor some cn ; we
will derive the desired formula for these coefficients. Since f (a) = ∞ n
n=0 cn (a−a) =
c0 + c1 (0) 2
c2 (0 ) + · · · , we get c0 = f (a). Next, by the Theorem in §11.9, we have
0
P+∞ n−1 , so f 0 (a) = c + 2c (0) + 3c (02 ) + · · · , and c = f 0 (a).
f (x) = n=0 ncP n (x−a) 1 2 3 1
00 ∞ n−2 , so f 00 (a) = (2)(1)c and c = 1 f 00 (a).
Next, f (x) = n=0 n(n−1)c n (x−a) 2 2 2
Continuing, we get:
∞
X
f (N ) (x) = n(n−1) · · · (n−N +1) cn (x−a)n−N .
n=1

Notes by Peter√Magyar magyar@math.msu.edu

∗
The function 3 x−a is also not analytic near x = a, ever though it gives real number values.
2
The problem is it has a vertical tangent at x = a, so it is not differentiable. Also see e−1/x in §11.11.
The terms for n = 0, 1, . . . , N −1 are all zero because of the factors n(n−1) · · · (n−N +
1), so the first non-zero term is for n = N . Plugging in x = a gives: f (N ) (a) =
N (N −1) · · · (1)cN , and cN = N1 ! f (N ) (a) as desired. Q.E.D.
(n)
Once we have a power series for f (x) with known coefficients cn = f n!(a) , we can
approximate f (x) by taking a finite partial sum of the series up to some cutoff term
N . This partial sum is called a Taylor polynomial, denoted TN (x):
N
X f (N ) (a)
f (x) ≈ TN (x) = cn (x−a)n = f (a) + f 0 (a)(x−a) + · · · + (x−a)N .
N!
n=0

Note that T1 (x) = f (a) + f 0 (a)(x − a) is just the linear approximation near x = a,
whose graph is the tangent line (Calculus I §2.9). We can improve this approximation
of f (x) in two ways:

• Take more terms, increasing N .

• Take the center a close to x, giving small (x−a) and tiny (x−a)n .

A Taylor series centered at a = 0 is specially named a Maclaurin series.

Example: sine function. To find Taylor series for a function f (x), we must de-
termine f (n) (a). This is easiest for a function which satisfies a simple differential
equation relating the derivatives to the original function. For example, f (x) = sin(x)
satisfies f 00 (x) = −f (x), so coefficients of the Maclaurin series (center a = 0) are:

n 0 1 2 3 4 5 6 7
f (n) (x) sin(x) cos(x) − sin(x) − cos(x) sin(x) cos(x) − sin(x) − cos(x)
f (n) (0) 0 1 0 −1 0 1 0 −1
f (n) (0)
cn = n! 0 1 0 − 3!1 0 1
5! 0 − 7!1

That is:
∞
x3 x5 x7 X x2n+1
sin(x) = x − + − + ··· = (−1)n .
3! 5! 7! (2n+1)!
n=0

To find the domain of convergence, we apply the Ratio Test (11.6/I):

 
 
 an+1   x2(n+1)+1  x2n+1 
L = lim   = lim  
an  n→∞  (2(n+1)+1)! (2n+1)! 

n→∞

|x|2n+3 (2n+1)! |x|2

= lim · = lim = 0
n→∞ |x|2n+1 (2n+3)! n→∞ (2n+2)(2n+3)

for any fixed x 6= 0. Since L = 0 < 1 regardless of x, the series converges for all x.
This formula for sin(x) astonishes because the right side is a simple algebraic series
having no apparent relation to trigonometry. We can try to understand and check the
series by graphically comparing sin(x) with its Taylor polynomial approximations:

• The Taylor polynomial T1 (x) = x (in red) is just the linear approximation or
tangent line of y = sin(x) at the center point x = 0. The curve and line are
close (to within a couple of decimal places) near the point of tangency and up
to about |x| ≤ 0.5. Once they veer apart, the approximation is useless.
3
• The next Taylor polynomial T3 (x) = x − x3! = x − 16 x3 (in green) matches
y = sin(x) in its first three derivatives at x = 0, and stays close to the original
curve up to about |x| ≤ 1.5 .
3 5
• The next T5 (x) = x − x3! + x5! = x − 16 x3 + 120
1 5
x is even closer to f (x) for even
larger x. Taking enough terms in the Taylor series will give a good approxima-
tion for any x, since the series converges everywhere.

problem: Compute sin(10◦ ). A geometric method would be to construct a right

triangle with a 10◦ angle, and measure the opposite side divided by the hypotenuse;
but this would only produce a couple of decimal places of accuracy. Of course, a
calculator can produce many decimal places, but how does it know? Taylor series!
As always when doing calculus on trig functions, we must first convert to radians
(see end of §2.5): 10◦ = 360
2π π
(10) = 18 π
. Here |x| = 18 ≈ 16 is small, so the Maclaurin
series centered at 0 should converge quickly, giving very accurate approximations:
π π π π 3
sin( 18 ) ≈ T3 ( 18 ) = 18 − 61 ( 18 ) ≈ 0.1736468 .

It turns out this is correct to 5 decimal places (underlined), using only two non-zero
terms of the Taylor series and a good estimate for π. We could verify this by taking
more terms and seeing that these 5 digits do not change, or by applying the Lagrange
Remainder estimates in §11.11.
 √
Example:
√ square roots. Compute 2 to 5 decimal places.† First, we must consider
√
2 to be an output of the function f (x) = x at x = 2. Next, we must choose the
center a for its Taylor series.
√
• a = 0 does not give a series because x is not analytic at x = 0. Indeed, if
√ 2
√ Taylor series x = c0 + c1 x + c22x + · · · , we could plug
there were a convergent
in x = −0.1 to get: −0.1 = c0 + c1 (−0.1) + c2 (−0.1) + · · · , a real value for
the square root of a negative number!

• a = 1 is too far from x = 2: it turns out |x−a| = |2−1| = 1 is beyond the radius
of convergence of the Taylor series.

since writing the Taylor series requires us to know f (n) (2),

• a = 2 is useless, √
including f (2) = 2, the same number we are trying to compute.

• A useful choice of a requires: a > 0 so that the Taylor series exists; a is close
√
to x = 2, making |x−a| small so the series converges quickly; and f (a) = a
is easy to compute so we can find the coefficients. A value satisfying all three
conditions is: a = 94 .
Now we have:

n 0 1 2 3 4
f (n) (x) x1/2 1
2 x−1/2 − 1·1
2·2 x
−3/2 1·1·3
2·2·2 x−5/2 − 2·2·2·2
1·1·3·5 −7/2
x
f (n) ( 94 ) 3
2
1
3
2
− 27 4
81
40
− 729
f (n) ( 94 ) 3 1 1 2 5
cn = n! 2 3 − 27 243 − 2187

Hence:
√ 3 9 2 9 3 9 4
x = 2 + 13 (x− 49 ) − 1
27 (x− 4 ) + 2
243 (x− 4 ) − 5
2187 (x− 4 ) + ···
∞
2n−1
(−1)n−1 (2n−3)!!
X
= 3
2 + 13 (x− 94 ) + n! 32n−1
(x− 49 )n ,
n=2
where we use the odd-factorial notation (2n−3)!! = (1)(3)(5) · · · (2n−3). For x = 2,
we have x− 49 = − 14 , so:
√
2 = 23 + 13 (− 14 ) − 27
1
(− 14 )2 + 243
2
(− 41 )3 − 2187
5
(− 14 )4 + · · ·
3 1 1 1 1 2 1 5 1
≈ 2 − 3 4 − 27 42 − 243 43 − 2187 44 ≈ 1.4142143 ,

which is correct to 5 decimal places (underlined).

†
We saw a faster algorithm for this (but not for functions like sin) in Calculus I §3.8: Newton’s
Method finds approximate solutions to equations like g(x) = x2 −2 = 0 by repeatedly solving a linear
0
approximation of g(x) = 0; this improves approximate solution
2
√ xn to xn+1 = xn − g(xn )/g (xn ) =
xn − (xn −2)/(2xn ). Starting with x0 = 3/2 = 1.5 gives x = 2 accurate to 5 decimal places after
just n = 2 iterations, while the Taylor series requires n = 4. Newton’s Method doubles the number
of accurate places each time, while each term of the series adds a constant number of places (§11.11).
Common Taylor series
∞
1 X
• = xn for |x| < 1 (Geometric Series).
1−x
n=0
∞
X xn
• ln(1+x) = (−1)n−1 for |x| < 1.
n
n=1
∞
X p(p−1) · · · (p−n+1)
• (1+x)p = xn for |x| < 1 (Binomial Series).
n!
n=0
∞
X xn
• exp(x) = for all x.
n!
n=0
∞
X x2n+1
• sin(x) = (−1)n for all x.
(2n+1)!
n=0
∞
X x2n
• cos(x) = (−1)n for all x.
(2n)!
n=0
Extra Topic: Irrationality of e. In §11.2, we saw that repeating infinite decimals
represent rational numbers (fractions), and every fraction can be written by long
division as a repeating decimal. Any infinite decimal which does not repeat cannot
be written as a fraction: it is irrational.
√ However, it is diffficult to prove that any
given interesting number such as π or 2 is irrational.
We can use series to prove the irrationality of the constant e = 2.7182818284590 · · · .
To prove the negative proposition that e is not equal to any possible fraction a/b,
we use the method of contradiction: that is, we assume that there were some fraction
with e = a/b, and use this to deduce an impossible conclusion, which will show that
the original assumption e = a/b was also impossible.
Thus, using the Taylor series definition for e, we assume the possibility, for some
whole numbers a, b, of the equation:
1 1 1 def a
1+ + + + ··· = e = .
1! 2! 3! b
It is easy to show 2 < e < 3, so e is not a whole number, and would have denominator
b > 1. Consider the bth order Taylor approximation, with error or remainder Rb :
∞
1 1 1 1 X 1
1 + + + + · · · + + Rb = e, Rb = .
1! 2! 3! b! n!
n=b+1

1
We multiply by b! to clear denominators up to the b! term:

b! b! b! a
b! + + + · · · + + b!Rb = b! e = b! = (b−1)! a .
1! 2! b! b
b! b!
The terms b! , 1! , 2! , . . . , b!
b! on the left are whole numbers, and (b−1)! a on the right is
a whole number, so the remainder b!Rb must also be a whole number. But it must
also be very small, as we can see from a simple geometric series comparison:‡
∞
P b! 1 1 1
b!Rb = n! = b+1 + (b+1)(b+2) + (b+1)(b+2)(b+3) + ···
n=b+1
1 1 1
< b+1 + (b+1)2
+ (b+1)3
+ ···
1 1 1 b+1 1
= b+1 1−( 1 ) = b+1 b+1−1 = b < 1.
b+1

Thus, we have constructed a positive number b!Rb which is both a whole number and
less than 1, which is impossible. Therefore the original assumption e = a/b must also
be impossible.

‡
Here we do not need the powerful Lagrange remainder formula from §11.11.