Power Series
Every time your phone works out a sine, a logarithm or a square root, there is no tiny
protractor hidden in the chip — it quietly adds up a never-ending polynomial to reach the
answer. That never-ending polynomial is a power series, and it is how
machines compute almost every function that science and engineering run on.
What if a polynomial never had to stop? Polynomials are the friendliest functions in
mathematics — you can evaluate them with nothing but multiplication and addition,
differentiate them in your head, integrate them by shifting an exponent. Their only flaw is
that they run out: a degree-7 polynomial has nothing to say after its x^7
term. A power series removes the stopping point. Pick a centre
a and an endless supply of coefficients
c_0, c_1, c_2, \dots, and form
f(x) \ = \ \sum_{n=0}^{\infty} c_n (x - a)^n \ = \ c_0 + c_1(x-a) + c_2(x-a)^2 + \cdots.
This is a polynomial of infinite degree. For each fixed x
it is just a numerical series — it either converges or it doesn't — and the set of
x where it converges is the domain of the function
f. The astonishing payoff, which this page unpacks, is that
wherever it converges it behaves exactly like a polynomial: you may differentiate
it, integrate it, and add it term by term. Suddenly the traffic flows both ways —
functions become series (we can compute \ln and
\arctan by adding up powers of x) and
series become functions (an infinite sum acquires a name, a graph, a
derivative). Most of the functions your calculator knows are, under the hood, exactly this.
The mother of all power series
Take a = 0 and every coefficient c_n = 1.
The power series is the
geometric series
in x:
\sum_{n=0}^{\infty} x^n = 1 + x + x^2 + x^3 + \cdots.
Step 1 — the partial sum has a closed form. Multiply
S_N = 1 + x + \cdots + x^N by (1-x);
the middle terms telescope:
(1-x)S_N = 1 - x^{N+1} \quad\Longrightarrow\quad S_N = \frac{1 - x^{N+1}}{1 - x} \quad (x \ne 1).
Step 2 — take the limit. If |x| < 1 then
x^{N+1} \to 0, so S_N \to 1/(1-x). If
|x| \ge 1 the terms x^n do not even tend
to zero and the series diverges. Hence
\sum_{n=0}^{\infty} x^n = \frac{1}{1 - x}, \qquad \text{valid exactly on } (-1, 1).
A simple-looking sum on the left equals a genuine, named function on the right — but only
inside (-1,1). Where a power series converges is not optional; it
is part of what the function is. We call this the mother series because, as
you are about to see, an entire nursery of famous series can be bred from it without ever
doing another limit computation.
Calculus, term by term
The real magic is that, inside its interval of convergence, a power series may be
differentiated and integrated one term at a time, just like a polynomial — and the new series
has the same radius. Differentiate the geometric series:
\frac{d}{dx}\sum_{n=0}^{\infty} x^n = \sum_{n=1}^{\infty} n x^{n-1} = \frac{d}{dx}\,\frac{1}{1-x} = \frac{1}{(1-x)^2}.
We never touched the limit machinery — we differentiated the closed form on the right and the
series on the left, and they agree. Likewise integrating term by term gives a new power series
for -\ln(1-x):
\int_0^x \frac{dt}{1-t} = \sum_{n=0}^{\infty} \frac{x^{n+1}}{n+1} = x + \frac{x^2}{2} + \frac{x^3}{3} + \cdots = -\ln(1-x).
One geometric series, differentiated and integrated, has just handed us power series for
1/(1-x)^2 and \ln. That is the engine of
the whole subject.
A power series f(x) = \sum_{n=0}^{\infty} c_n (x-a)^n defines a
function on the set of x where it converges. On the open interval
where it converges:
-
f is continuous, and indeed
infinitely differentiable.
-
Term-by-term differentiation is valid:
f'(x) = \sum_{n=1}^{\infty} n\, c_n (x-a)^{n-1}, with the same
radius of convergence.
-
Term-by-term integration is valid:
\int f = C + \sum_{n=0}^{\infty} \dfrac{c_n}{n+1}(x-a)^{n+1},
again with the same radius.
-
The geometric series is the prototype:
\sum_{n=0}^{\infty} x^n = \dfrac{1}{1-x} on
(-1,1).
A single convergent power series is automatically C^\infty on the
interior of its interval — you can differentiate it forever and every derivative is again a
power series. This is a small miracle. An arbitrary function built by other means might be
differentiable twice and then break; a power series never does. Each differentiation lowers
every exponent by one and multiplies by it, so
f^{(k)}(x) = \sum_{n=k}^{\infty} n(n-1)\cdots(n-k+1)\, c_n (x-a)^{n-k},
which is still a power series with the same radius. The catch — and it is a sharp one —
is the phrase "inside the interval". The convergence is what licenses every interchange of
limit, derivative and integral above. Step outside the interval and the series is not a
function at all, merely a divergent symbol. The next page,
radius of convergence,
is devoted to measuring exactly how wide that safe interval is.
Watch the series become the function
The faint curve is the true function 1/(1-x). The bold curve is the
truncated power series \sum_{n=0}^{N} x^n — an honest polynomial of
degree N. Drag N upward and watch two
things at once. Inside (-1,1) the polynomial snaps onto the
function — near x = 0 the fit is superb even for small
N, because there the discarded terms x^{N+1} + x^{N+2} + \cdots
are microscopic. But near the boundary x = \pm 1 the polynomial
flails: at x = -1 it flips between 0 and
1 forever, and left of -1 it shoots off
in whichever direction its last term points. No finite N ever tames
the region |x| \ge 1 — that is the boundary of convergence making
itself felt, and no amount of patience crosses it.
A factory for new series
Here is the trade secret of every mathematician who ever wrote down a famous series: almost
none of them were derived from scratch. They were manufactured from the mother series
by two moves — substitution and term-by-term calculus.
Move 1: substitute. The identity
\sum u^n = 1/(1-u) holds for any expression u
with |u| < 1. Feed it u = -x^2:
\frac{1}{1+x^2} = \frac{1}{1-(-x^2)} = \sum_{n=0}^{\infty} (-x^2)^n = 1 - x^2 + x^4 - x^6 + \cdots,
valid when |-x^2| < 1, i.e. on (-1,1)
again. One line of algebra, no limits, and we own a power series for a function that looks
nothing like a geometric sum. (Try u = -x and
u = 2x yourself: you should get
1/(1+x) on (-1,1) and
1/(1-2x) on |x| < \tfrac12 — note how the
substitution rescaled the interval.)
Move 2: integrate term by term. Now let us derive the series for
\ln(1+x) honestly, every step in view. Start from the substitution
u = -t:
\frac{1}{1+t} = \sum_{n=0}^{\infty} (-1)^n t^n = 1 - t + t^2 - t^3 + \cdots, \qquad |t| < 1.
Integrate both sides from 0 to x (with
|x| < 1, so the whole path of integration stays inside the safe
interval). The left side is a calculus classic; the right side goes term by term, each
t^n becoming x^{n+1}/(n+1):
\ln(1+x) \;=\; \int_0^x \frac{dt}{1+t} \;=\; \sum_{n=0}^{\infty} (-1)^n \frac{x^{n+1}}{n+1} \;=\; x - \frac{x^2}{2} + \frac{x^3}{3} - \frac{x^4}{4} + \cdots.
Both constants of integration vanish because both sides are 0 at
x = 0. The same recipe applied to our
1/(1+t^2) series delivers, with no extra effort, one of the most
celebrated formulas in mathematics:
\arctan x \;=\; \int_0^x \frac{dt}{1+t^2} \;=\; x - \frac{x^3}{3} + \frac{x^5}{5} - \frac{x^7}{7} + \cdots, \qquad |x| < 1.
Pause on what just happened. The logarithm and the inverse tangent — transcendental functions,
unreachable by any finite amount of school algebra — have been written as limits of
polynomials, by doing nothing more than substituting into a geometric series and integrating
powers of t. Functions became series.
Before calculus was even public, the young Isaac Newton had discovered that
(1+x)^p has a power series for any exponent
p — not just whole numbers. His generalised binomial series,
(1+x)^{p} = 1 + px + \frac{p(p-1)}{2!}x^2 + \frac{p(p-1)(p-2)}{3!}x^3 + \cdots \qquad (|x|<1),
turns square roots into sums: with p = \tfrac12 it gives
\sqrt{1+x} = 1 + \tfrac{x}{2} - \tfrac{x^2}{8} + \tfrac{x^3}{16} - \cdots.
Newton treated it as a private superpower — he used such series to compute areas, roots and
the digits of \pi to extraordinary accuracy, and confessed he was
"ashamed" to admit how many digits he once computed, "having no other business at the time".
When rivals pressed him for his methods, he famously encoded one claim about series as an
anagram. To Newton, every function worth having was a power series in disguise —
and the pages ahead, on
Taylor series,
largely prove him right.
The factory's product, on screen
Here is the arctangent series earning its keep. The faint curve is the true
\arctan x; the bold one is the polynomial
x - \frac{x^3}{3} + \cdots \pm \frac{x^{2M+1}}{2M+1}. Inside
(-1,1) a handful of terms already hugs the curve. But look just past
x = \pm 1: the polynomial peels away violently, no matter how large
you push M. The series inherited its interval from the geometric
series it was built from, and integrating term by term did not widen it.
Cashing in: evaluating a series at a point
A power series is not just theory — it is a computational device. Any
x inside the interval may be plugged in, and the partial sums then
converge to a genuine number. Let us compute \ln(1.1) by setting
x = 0.1 (comfortably inside (-1,1)) in
our freshly built series:
\ln(1.1) = 0.1 - \frac{(0.1)^2}{2} + \frac{(0.1)^3}{3} - \frac{(0.1)^4}{4} + \cdots = 0.1 - 0.005 + 0.000333\ldots - 0.000025 + \cdots
Three terms give 0.095333\ldots; four give
0.0953083\ldots; the true value is
\ln(1.1) = 0.0953102\ldots Each extra term buys roughly another
digit, because the powers of 0.1 collapse so fast. This is the
general pattern: the deeper inside the interval you are, the faster the series
converges. At x = 0.9 the same series still converges —
but it crawls, and near the edge of the interval it crawls painfully. A power series is a
machine whose efficiency depends on where you stand.
Changing the centre
So far every series has been centred at 0, which is useless if the
function you care about misbehaves there — 1/x, say, which blows up
at the origin. The fix is to expand in powers of (x-a) about a
friendlier centre. Take a = 1 and massage
1/x until a geometric series appears:
\frac{1}{x} = \frac{1}{1 + (x-1)} = \sum_{n=0}^{\infty} (-1)^n (x-1)^n = 1 - (x-1) + (x-1)^2 - (x-1)^3 + \cdots,
valid when |x - 1| < 1, i.e. on the interval
(0, 2). Notice two things. First, the interval is
symmetric about the centre a = 1, not about
0 — a power series always converges on a symmetric interval
(a - R,\, a + R) around its centre (plus possibly an endpoint or
two), where R is its
radius of convergence.
Second, look where the interval stops: exactly at x = 0,
the point where 1/x explodes. The series seems to know
about the singularity and refuses to reach it — the radius here is precisely the distance from
the centre to the nearest catastrophe. Measuring R in general (it
can be 0, a finite number, or \infty) is
the business of the next page.
Every manipulation on this page is licensed only inside the interval of
convergence. Forget that, and algebra becomes a nonsense generator. Plug
x = 2 into \sum x^n = 1/(1-x) and you
"prove"
1 + 2 + 4 + 8 + 16 + \cdots = \frac{1}{1-2} = -1,
an ever-growing sum of positive numbers equalling -1. Nothing in
the algebra failed — the closed form S_N = (1-x^{N+1})/(1-x) is
perfectly true at x=2. What failed is the limit: for
|x| \ge 1 the term x^{N+1} never dies,
so the series diverges and the equation with 1/(1-x) simply does
not apply. Divergent series obey no arithmetic.
A subtler trap: term-by-term calculus preserves the radius but can change
what happens at the endpoints. The geometric series diverges at both
x = 1 and x = -1. Integrate it and you
get x + \frac{x^2}{2} + \frac{x^3}{3} + \cdots = -\ln(1-x): still
radius 1, still divergent at x = 1
(the harmonic series), but at x = -1 it now converges —
it is the alternating harmonic series, summing to -\ln 2.
Integration soothed an endpoint into convergence; differentiation can do the reverse and
destroy one. Moral: the open interval travels with the series for free, but every endpoint
must be re-checked by hand.
Press sin on a calculator and something has to happen inside a chip that can only
add, subtract, multiply and divide. There is no tiny protractor in there. What the chip
evaluates is — a truncated power series (or a polynomial fine-tuned from one). Functions like
\sin, \cos and
e^x all have power series converging for every
x, and near 0 a handful of terms is
spectacular: for |x| \le 0.5,
x - \frac{x^3}{6} + \frac{x^5}{120} already matches
\sin x to about eight decimal places. So the software first uses
identities to fold your input into a small interval around 0
(where convergence is fastest — exactly the "deeper inside, faster" principle above), then
fires off a short polynomial. Your slider experiments on this page are a faithful cartoon of
what happens on every keypress.
A lovely historical aside: many pocket calculators and early flight computers used a
different trick called CORDIC, which computes \sin
and \cos purely by shifting binary digits and adding — no
multiplier circuit needed. But modern processors, with fast multipliers on board, have gone
back to polynomials. Three and a half centuries after Newton, the state of the art is still:
truncate a power series.
See it explained