Power Series and Radius of Convergence

Power series: an infinite series

\sum_{n = 0}^{\infty} a_{n} (x - c)^{n}

that behaves like an "infinite polynomial." Every such series has a radius of convergence

R \in [0, \infty]

such that it converges absolutely for

∣ x - c ∣ < R

and diverges for

∣ x - c ∣ > R

. The number

R

is computed by the Cauchy–Hadamard formula. Power series are the foundation of analytic function theory, Taylor expansions, generating functions, and the solution of differential equations. Central results due to Cauchy, Abel, Hadamard, and Weierstrass (19th century).

A power series centered at $c$ is an expression $\sum_{n = 0}^{\infty} a_{n} (x - c)^{n}$ with fixed coefficients $a_{n}$ and variable $x$ . Its convergence set is always an interval (over $R$ ) or a disk (over $C$ ) centered at $c$ , whose radius $R$ is determined entirely by how fast $∣ a_{n} ∣$ grows. Power series matter because the functions they represent — the analytic functions — are precisely those that are "infinitely differentiable in the strongest possible sense": each is locally equal to its own Taylor series, and may be differentiated and integrated term by term inside the radius of convergence. A surprising fact: a real power series can fail to extend past a point on its interval of convergence purely because of a complex singularity it never "sees" on the real line — the radius of convergence of $\sum x^{2 n} (- 1)^{n} = 1/ (1 + x^{2})$ is $1$ even though $1/ (1 + x^{2})$ is perfectly smooth for all real $x$ .

The Core Insight

The convergence of a power series is governed by a single threshold $R$ . Inside the disk $∣ x - c ∣ < R$ the series converges absolutely and the geometric "tail control" $∣ a_{n} (x - c)^{n} ∣ \sim (ρ / R)^{n}$ for $ρ < R$ makes everything well-behaved (continuity, differentiability, integrability term-by-term). Outside the terms fail to even approach zero, so divergence is automatic. The boundary $∣ x - c ∣ = R$ is the genuinely subtle frontier where anything can happen. The magic is that $R$ depends only on the growth rate of $∣ a_{n} ∣^{1/ n}$ , not on $x$ .

Formal Statement

Power Series

A power series centered at $c \in R$ (or $C$ ) with coefficients $a_{n}$ is the formal sum

f (x) = n = 0 \sum \infty a_{n} (x - c)^{n} = a_{0} + a_{1} (x - c) + a_{2} (x - c)^{2} + \dots

⟳By definition; the symbols are

a_{n}

(the

n

-th coefficient, a fixed scalar),

c

(the center),

x

(the variable),

n

(the summation index running over non-negative integers).

The convergence set is $S = {x : the series [1] converges}$ . We always have $c \in S$ since at $x = c$ the series reduces to $a_{0}$ .

Radius and Interval of Convergence

The radius of convergence is

R = sup {r \geq 0 : n = 0 \sum \infty a_{n} (x - c)^{n} converges for some ∣ x - c ∣ = r}

with the conventions

R = 0

(converges only at

c

) and

R = \infty

(converges everywhere). The interval of convergence is the set

S

itself, which is one of

(c - R, c + R)

[c - R, c + R)

(c - R, c + R]

, or

[c - R, c + R]

over

R

— equal to the open interval together with whichever endpoints converge.

⟳

R

is defined as a supremum over radii of convergence; the four-fold interval cases arise because endpoint behavior is independent on each side.

Cauchy–Hadamard Formula

For any power series $[1]$ , the radius of convergence is given by

\frac{1}{R} = n \to \infty lim sup ∣ a_{n} ∣^{1/ n}

with the conventions

1/0 = \infty

and

1/\infty = 0

. Moreover the series converges absolutely for

∣ x - c ∣ < R

and diverges for

∣ x - c ∣ > R

⟳Apply the root test to

\sum ∣ a_{n} (x - c)^{n} ∣

; the

n

-th root of the general term is

∣ a_{n} ∣^{1/ n} ∣ x - c ∣

, whose limsup is

∣ x - c ∣/ R

Ratio Form

If the limit exists, the radius of convergence may also be computed as

R = n \to \infty lim \frac{a _{n}}{a _{n + 1}}

provided the

a_{n}

are eventually nonzero and the limit (in

[0, \infty]

) exists. This always agrees with

[3]

when the ratio limit exists, but the root formula

[3]

is universally valid.

⟳Apply the ratio test to consecutive terms

a_{n + 1} (x - c)^{n + 1} / a_{n} (x - c)^{n} = (a_{n + 1} / a_{n}) (x - c)

; convergence requires

∣ x - c ∣ < lim ∣ a_{n} / a_{n + 1} ∣

Why It Holds

Cauchy–Hadamard via the Root Test

We prove $[3]$ . Let $L = lim sup_{n \to \infty} ∣ a_{n} ∣^{1/ n}$ and set $R = 1/ L$ .

Step 1 — Set up the root test. Apply the root test to the series of absolute values

\sum_{n = 0}^{\infty} ∣ a_{n} (x - c)^{n} ∣ = \sum_{n} ∣ a_{n} ∣ ∣ x - c ∣^{n}

. The relevant quantity is

α := n \to \infty lim sup (∣ a_{n} ∣ ∣ x - c ∣^{n})^{1/ n} = ∣ x - c ∣ \cdot n \to \infty lim sup ∣ a_{n} ∣^{1/ n} = ∣ x - c ∣ L = \frac{∣ x - c ∣}{R} .

⟳

∣ x - c ∣

is a constant factor that pulls out of the limsup of an

n

-th root; the remaining factor is the definition of

L

Step 2 — Convergence when $∣ x - c ∣ < R$ . Then

α = ∣ x - c ∣/ R < 1

. The root test states that if

lim sup (term)^{1/ n} < 1

the series converges absolutely. Concretely, pick

β

with

α < β < 1

; for all large

n

we have

∣ a_{n} ∣∣ x - c ∣^{n} \leq β^{n}

, and

\sum β^{n}

converges as a geometric series. By comparison,

[1]

converges absolutely. Step 3 — Divergence when $∣ x - c ∣ > R$ . Then

α > 1

. By definition of

lim sup

, infinitely many terms satisfy

(∣ a_{n} ∣∣ x - c ∣^{n})^{1/ n} > 1

, i.e.

∣ a_{n} (x - c)^{n} ∣ > 1

. The general term does not approach

0

, so the series diverges (the $n$ -th term test). Step 4 — Boundary undetermined. When

∣ x - c ∣ = R

α = 1

and the root test is inconclusive; both convergence and divergence are possible (see the examples). This establishes

[2]

and

[3]

■

Term-by-Term Differentiation Inside R

A crucial consequence: inside the disk of convergence, $f$ is differentiable and

f^{'} (x) = n = 1 \sum \infty n a_{n} (x - c)^{n - 1}, ∣ x - c ∣ < R,

and the differentiated series has the same radius $R$ .

⟳

lim sup ∣ n a_{n} ∣^{1/ n} = lim sup n^{1/ n} ∣ a_{n} ∣^{1/ n} = 1 \cdot L = L

since

n^{1/ n} \to 1

; so

R

is unchanged. Equality of

f^{'}

with the term-by-term derivative follows from uniform convergence of the differentiated series on each compact subdisk

∣ x - c ∣ \leq ρ < R

(Weierstrass M-test), justifying interchange of limit and derivative.

Jacques Hadamard | 1865–1963

Though primarily a pure mathematician, Hadamard's 1888 thesis crystallized the formula now bearing his name (with Cauchy), giving the exact radius of convergence via the $lim sup$ of $∣ a_{n} ∣^{1/ n}$ — the first fully general statement, valid even when the ratio $∣ a_{n} / a_{n + 1} ∣$ has no limit. Hadamard also proved the Prime Number Theorem (1896, independently of de la Vallée Poussin), worked extensively on partial differential equations and the calculus of variations, and articulated the influential notion of a well-posed problem. His insight that the growth rate of coefficients alone dictates convergence transformed power series from a computational tool into a structural theory linking real analysis to complex function theory.

Building Intuition

Think of $∣ a_{n} (x - c)^{n} ∣$ as a race between two factors. The coefficients $∣ a_{n} ∣$ try to blow the terms up (if they grow), while the powers $∣ x - c ∣^{n}$ either help (if $∣ x - c ∣ < 1$ ) or hurt (if $∣ x - c ∣ > 1$ ). The radius $R$ is the precise value of $∣ x - c ∣$ at which these forces exactly balance:

If $∣ x - c ∣ < R$ , the powers shrink fast enough to crush whatever growth $∣ a_{n} ∣$ has — the terms decay geometrically, like a convergent geometric series with ratio $∣ x - c ∣/ R < 1$ .
If $∣ x - c ∣ > R$ , the powers grow faster than the coefficients can be tamed, the terms blow up, and the series has no chance.

The number $∣ a_{n} ∣^{1/ n}$ measures the "average geometric growth rate per term" of the coefficients; its reciprocal is exactly the distance at which the powers can offset that growth. This is why $R$ depends only on the coefficients, never on $x$ .

Convergence behavior on the real line: absolute convergence inside, divergence outside, indeterminate at the endpoints.

<svg viewBox="0 0 640 260" xmlns="http://www.w3.org/2000/svg">
<rect width="640" height="260" fill="#0a0e17"/>
<!-- number line -->
<line x1="40" y1="150" x2="600" y2="150" stroke="#94a3b8" stroke-width="2"/>
<polygon points="600,150 590,145 590,155" fill="#94a3b8"/>
<!-- center c -->
<line x1="320" y1="140" x2="320" y2="160" stroke="#e2e8f0" stroke-width="2"/>
<circle cx="320" cy="150" r="3" fill="#e2e8f0"/>
<text x="320" y="180" fill="#e2e8f0" font-size="14" text-anchor="middle">c</text>
<!-- endpoints c-R, c+R -->
<line x1="140" y1="135" x2="140" y2="165" stroke="#d4a853" stroke-width="2"/>
<line x1="500" y1="135" x2="500" y2="165" stroke="#d4a853" stroke-width="2"/>
<text x="140" y="185" fill="#d4a853" font-size="13" text-anchor="middle">c−R</text>
<text x="500" y="185" fill="#d4a853" font-size="13" text-anchor="middle">c+R</text>
<!-- convergence region shading -->
<rect x="140" y="142" width="360" height="16" fill="#d4a853" opacity="0.18"/>
<text x="320" y="120" fill="#d4a853" font-size="14" text-anchor="middle">absolute convergence  |x−c| &lt; R</text>
<!-- divergence regions -->
<text x="85" y="120" fill="#94a3b8" font-size="13" text-anchor="middle">diverges</text>
<text x="555" y="120" fill="#94a3b8" font-size="13" text-anchor="middle">diverges</text>
<!-- endpoint question marks -->
<text x="140" y="220" fill="#e2e8f0" font-size="13" text-anchor="middle">? convergence</text>
<text x="500" y="220" fill="#e2e8f0" font-size="13" text-anchor="middle">? convergence</text>
<text x="320" y="245" fill="#e2e8f0" font-size="12" text-anchor="middle">endpoints must be tested individually</text>
</svg>

Worked Examples

The Exponential Series — Infinite Radius

Given:

f (x) = n = 0 \sum \infty \frac{x ^{n}}{n !}

, so

a_{n} = 1/ n!

c = 0

. Step 1 — Ratio form $[4]$ .

\frac{a _{n}}{a _{n + 1}} = \frac{1/ n !}{1/ ( n + 1 )!} = \frac{( n + 1 )!}{n !} = n + 1.

Step 2 — Take the limit.

R = n \to \infty lim (n + 1) = \infty.

Step 3 — Interpret. The series converges absolutely for every real (and complex)

x

. There are no endpoints to test since

R = \infty

R = \infty, interval of convergence = (- \infty, \infty)

This is consistent with $f (x) = e^{x}$ being defined and analytic everywhere.

Endpoints Differ — A Logarithmic Series

Given:

f (x) = n = 1 \sum \infty \frac{x ^{n}}{n}

, so

a_{n} = 1/ n

c = 0

. Step 1 — Radius via ratio.

R = n \to \infty lim \frac{a _{n}}{a _{n + 1}} = n \to \infty lim \frac{1/ n}{1/ ( n + 1 )} = n \to \infty lim \frac{n + 1}{n} = 1.

So the open interval is

(- 1, 1)

. Step 2 — Test the right endpoint $x = 1$ . The series becomes

\sum 1/ n

, the harmonic series, which diverges. Step 3 — Test the left endpoint $x = - 1$ . The series becomes

\sum (- 1)^{n} / n

, the alternating harmonic series, which converges (conditionally, by the alternating series test).

R = 1, interval of convergence = [- 1, 1)

This illustrates that the two endpoints can — and often do — behave completely differently. (Indeed $f (x) = - ln (1 - x)$ , whose singularity at $x = 1$ explains the divergence there.)

Common Misconceptions

The ratio test is not always applicable. The formula

R = lim ∣ a_{n} / a_{n + 1} ∣

requires the limit to exist. For series with gaps (e.g. only even powers) or oscillating coefficient ratios, the limit may fail. The Cauchy–Hadamard formula

[3]

with the limsup is the universally correct tool. Also note: convergence at

∣ x - c ∣ = R

tells you nothing about the other endpoint or about absolute vs. conditional convergence — each boundary point must be checked by hand.

Ratio Test Failure on a Lacunary Series

Consider

f (x) = n = 0 \sum \infty x^{2^{n}} = x + x^{2} + x^{4} + x^{8} + x^{16} + \dots

Here the coefficient of

x^{k}

a_{k} = 1

k

is a power of

2

and

a_{k} = 0

otherwise. The ratio

∣ a_{k} / a_{k + 1} ∣

is undefined infinitely often (division by zero), so

[4]

cannot be used. But Cauchy–Hadamard

[3]

still works:

lim sup_{k} ∣ a_{k} ∣^{1/ k} = 1

(since

∣ a_{k} ∣^{1/ k}

is either

1

0

, and equals

1

infinitely often), giving

R = 1

. Moreover, the unit circle

∣ x ∣ = 1

is a natural boundary:

f

cannot be analytically continued past it at any boundary point — a phenomenon invisible to the radius alone.

Connections

Power series sit at the crossroads of several theories:

A function equal to its Taylor Series on a neighborhood is precisely an analytic function; power series are the local building blocks of Analytic Functions.
Over $C$ , the radius of convergence equals the distance from the center to the nearest singularity, tying $R$ to Complex Analysis and the residue theory of poles and branch points.
Generating Functions in combinatorics are power series whose coefficients encode counting sequences; $R$ controls the asymptotic growth of those sequences via $a_{n} \sim R^{- n}$ .
The Frobenius method solves linear Ordinary Differential Equations by assuming power-series solutions and recovering recurrences for the $a_{n}$ .
Uniform Convergence on compact subdisks (via the Weierstrass M-test) is the engine that legitimizes term-by-term differentiation $[6]$ and integration.

Real-World Applications

Special-function computation. Libraries evaluate $e^{x}$ , $sin x$ , $cos x$ , $ln (1 + x)$ , and Bessel functions $J_{ν}$ from their power series, using $R$ to bound truncation error and choose argument-reduction strategies.
Signal processing — the $z$ -transform. Discrete signals are encoded as power series $\sum x [n] z^{- n}$ ; the region of convergence (an annulus governed by radii like $[3]$ ) determines causality and stability of filters.
Control theory. Transfer functions and their Taylor/Laurent expansions analyze system response; pole locations set the radius and hence stability margins.
Quantum mechanics & perturbation theory. Energy levels are expanded in powers of a coupling constant $λ$ ; the radius of convergence in $λ$ reveals when perturbation theory breaks down (often $R = 0$ , an asymptotic series).
Numerical ODE solvers. Taylor-series and power-series integration methods propagate solutions step by step, with $R$ limiting the safe step size.
Combinatorics & analysis of algorithms. Generating-function singularities (the radius $R$ ) yield exponential growth rates $\sim R^{- n}$ for counts of trees, paths, and data-structure operations (singularity analysis, Flajolet–Sedgewick).
Financial mathematics. Option-pricing and interest-rate models use series expansions of payoff and value functions; convergence radii bound the validity of analytic approximations.

Historical Development

1668–1671 — Nicolaus Mercator and Isaac Newton develop series for $ln (1 + x)$ and the generalized binomial series, pioneering power series as a computational engine.
1715 — Brook Taylor publishes the Taylor expansion (anticipated by Gregory and Newton), expressing functions as power series in a single point.
1748 — Euler's Introductio in analysin infinitorum treats power series formally and freely, deriving countless identities (often without convergence concerns).
1821 — Augustin-Louis Cauchy's Cours d'analyse rigorously defines convergence and introduces the root test, laying groundwork for what becomes the Cauchy–Hadamard formula.
1826 — Niels Henrik Abel proves Abel's theorem on boundary continuity ( $\sum a_{n} x^{n} \to \sum a_{n}$ as $x \to R^{-}$ when the endpoint series converges) and corrects loose 18th-century practice.
1851–1860s — Weierstrass builds analytic function theory on power series as the primitive objects, proving uniform convergence properties and the M-test that justify term-by-term operations.
1888 — Jacques Hadamard, in his doctoral thesis, states the general formula $1/ R = lim sup ∣ a_{n} ∣^{1/ n}$ , valid without assuming the ratio limit exists.
1892 — Hadamard's gap theorem on natural boundaries shows lacunary series cannot be continued past $∣ x ∣ = R$ .
20th c. — The theory generalizes to power series over $C$ , Banach spaces (formal/analytic functional calculus), and $p$ -adic fields, with the radius concept persisting throughout.

Subtleties & Limits

Several frontiers remain genuinely deep even beyond undergraduate scope:

Boundary behavior is largely undetermined in general: while Abel's theorem handles convergent endpoints and Tauberian theorems give partial converses, classifying exactly which boundary points of a given series converge can be intractable. Carleson's theorem (1966) on almost-everywhere convergence of Fourier series is a celebrated hard result in this circle of ideas.
Divergent yet useful series. Many physical expansions have $R = 0$ (asymptotic series). Borel summation and resurgence theory (Écalle) extract meaning from such series; understanding which divergent series are "resummable" and how is active research.
Natural boundaries. Determining whether a power series with a given coefficient pattern has the unit circle as a natural boundary — and the fine structure of singularities on it — connects to the Fabry gap theorem and remains delicate.
Several complex variables. The single radius $R$ becomes a domain of convergence with intricate (logarithmically convex) geometry; the theory of domains of holomorphy and pseudoconvexity is a rich modern subject.