Every power series has a radius of convergence R; it converges absolutely inside the circle |z| = R and not at all outside, with anything possible on the circle; and a function has at most one power series.
The previous lesson showed that convergence of a power series at one point spreads inwards. §200 pushes this to its conclusion: the region of convergence is always a disc, the inside of the circle of convergence, possibly together with some points of its boundary. This one number, the radius R, is the first thing to find about any power series, and the ratio and root tests usually find it at once. §201 then proves that the coefficients of a power series are determined by its sum, which is what makes “the power series of a function” a meaningful phrase.
The idea
Fix a series and walk out from 0 along the positive real axis. By Theorem B, once the series fails to converge at some point r, it fails at every point beyond it (otherwise convergence further out would spread back to r). So the points of the positive axis split into an initial stretch where the series converges and a final stretch where it does not, separated by a single number R. Theorem B again, applied in every direction, turns the stretch [0,R) into the open disc ∣z∣<R.
On the circle ∣z∣=R itself, the argument is silent, and the examples show that it has to be.
ExploreInside, outside, and on the circle
Try this.Drag z around inside the green circle, then outside it, for ∑zn. Then drag it onto the circle (it snaps there) and compare four series: ∑zn, the logarithmic series (try z near −1), ∑zn/n2, and the binomial series. Use N=1000 on the circle.
1+z+z2+⋯
z-plane: drag z
Re zIm z
w-plane: partial sums s₀, s₁, …, s_N
ReIm
circle of convergence, R = 1
z, and the partial sum s_N
path of partial sums
the sum f(z)
z = 0.6+0.45i. |z| = 0.75 < R = 1: the series converges absolutely, and the path homes in on the sum 1.1034+1.2414i.
What to notice
Inside, the path of partial sums closes in on the sum f(z); outside, the terms grow and the path runs off. On the circle the four series behave in four ways: ∑zn never converges; the logarithmic series converges everywhere except at z=−1, though the path is infinitely long; ∑zn/n2 and the binomial series converge absolutely all round the circle.
Precise statements
TheoremThe circle of convergenceHardy: §200
A power series ∑anzn either
converges for z=0 and for no other z; or
converges absolutely for every z; or
converges absolutely for every z with ∣z∣<R, and does not converge for any z with ∣z∣>R, for a certain number R>0.
In case 3 the circle ∣z∣=R is the circle of convergence and R the radius of convergence. It is convenient to say R=0 in case 1 and R=∞ in case 2. The theorem gives no information about points on the circle.
TheoremFinding the radiusHardy: Examples LXXX, 3
If ∣an+1∣/∣an∣→λ, or if ∣an∣1/n→λ, then the radius of convergence is R=1/λ (meaning R=∞ if λ=0, and R=0 if λ=∞).
TheoremBehaviour near the centreHardy: §201
If ∑anzn has radius of convergence R>0 and sum f(z) for ∣z∣<R, then for every m≥0
f(z)=a0+a1z+⋯+amzm+O(∣z∣m+1)(z→0).
CorollaryUniqueness of power seriesHardy: §201
If ∑anzn=∑bnzn for all z with ∣z∣<μ (some μ>0), then an=bn for every n. The same function cannot be represented by two different power series. It is even enough that the two sums agree for real positive z near 0.
Proofs
Hardy splits the positive real axis into two classes, as in a Dedekind section (§17). The same argument can be put in terms of a supremum.
ProofThe circle of convergence
Strategy
Define R as the supremum of the radii r≥0 at which the series converges, and use Theorem B to show that the series converges absolutely inside ∣z∣=R and not at all outside.
Define R
Let S be the set of real r≥0 at which ∑anrn converges; 0∈S. Put R=supS, with R=∞ if S is unbounded. Case 1 is R=0, case 2 is R=∞.
Absolute convergence inside
Let ∣z∣<R. Since ∣z∣ is less than the supremum, some r∈S has r>∣z∣. The series converges at r, so by Theorem B it converges absolutely at z.
No convergence outside
Let ∣z∣>R and suppose the series converged at z. Choose a real r with R<r<∣z∣. By Theorem B the series would converge (absolutely) at r, so r∈S and r≤R, a contradiction.
The three cases
If R=0, the second step is empty and the third says the series converges only at 0. If R=∞, the third step is empty and the second gives absolute convergence everywhere. Otherwise we are in case 3.
∎
ProofFinding the radius by the ratio test
Strategy
Apply d’Alembert’s test to the moduli ∣anzn∣.
The ratio of consecutive terms
For z=0: ∣anzn∣∣an+1zn+1∣=∣an∣∣an+1∣∣z∣→λ∣z∣.
Inside
If λ∣z∣<1, the ratio test (§175) makes ∑∣anzn∣ converge.
Outside
If λ∣z∣>1, the ratio exceeds 1 from some n on, so ∣anzn∣ increases from then on and cannot tend to 0; the series cannot converge.
Conclude
The series converges absolutely for ∣z∣<1/λ and does not converge for ∣z∣>1/λ, so R=1/λ. The root test (§174) gives the same with ∣anzn∣1/n→λ∣z∣.
∎
ProofBehaviour near the centre, and uniqueness
Strategy
The tail ∑n>manzn is dominated, near 0, by a geometric series whose first term is of size ∣z∣m+1.
A bound on the coefficients
Fix μ with 0<μ<R. The series converges at μ, so its terms are bounded: ∣an∣μn≤K for all n.
For ∣z∣≤μ/2 the right-hand side is at most μm+12K∣z∣m+1, which is O(∣z∣m+1).
Uniqueness
Let cn=an−bn, so that ∑cnzn=0 for ∣z∣<μ. If not all cn were 0, let cm be the first that is not. By the previous step applied to ∑cnzn (whose sum is 0), 0=cmzm+O(∣z∣m+1). Dividing by zm gives cm=O(∣z∣), and letting z→0 (along the positive reals, if we like) gives cm=0: a contradiction.
∎
The figure below is the logarithmic series of Examples LXXX, 4. Its radius is 1, and it shows the most delicate behaviour possible on the circle.
ExploreThe logarithmic series on its circle
Try this.With z on the circle, move it slowly towards −1 from above, with N=1000. Then go back inside the circle.
z−2z2+3z3−⋯
z-plane: drag z
Re zIm z
w-plane: partial sums s₀, s₁, …, s_N
ReIm
circle of convergence, R = 1
z, and the partial sum s_N
path of partial sums
the sum f(z)
z = −0.8+0.6i. z is on the circle of convergence: the series converges here (conditionally, by Dirichlet’s test), although the path has infinite length.
What to notice
Everywhere on the circle except −1 the path spirals in to log(1+z), the hollow point; the closer z is to −1, the further away that point is and the slower the spiral. At z=−1 itself the series is −(1+21+31+⋯) and diverges.
Let a>0. Show that 1+az+a2z2+⋯ has radius of convergence 1/a, and that it does not converge anywhere on its circle of convergence: it diverges at z=1/a and oscillates finitely at the other points of the circle.
Hint
Put w=az. On the circle, ∣w∣=1.
Solution
With w=az the series is ∑wn, which converges exactly when ∣w∣<1, that is ∣z∣<1/a. So R=1/a. On the circle ∣w∣=1 and the terms have modulus 1, so the series cannot converge. At w=1 (that is, z=1/a) it is 1+1+⋯, which diverges. At any other w on the unit circle, ∣1+w+⋯+wn∣=∣1−wn+1∣/∣1−w∣≤2/∣1−w∣: bounded partial sums without a limit, finite oscillation.
ExerciseRadii by the ratio test●○○Hardy: Examples LXXX, 2–3
Find the radius of convergence of (i) ∑zn/n2, (ii) ∑zn/(2n+3n), (iii) ∑nnzn. For (i), describe what happens on the circle.
Hint
(ii): divide numerator and denominator of the ratio by 3n. (iii): use the root test.
Solution
(i) ∣an+1/an∣=n2/(n+1)2→1, so R=1. On ∣z∣=1, ∣zn/n2∣=1/n2, so the series converges absolutely at every point of the circle (Examples LXXX, 2).
(ii) anan+1=2n+1+3n+12n+3n=2(2/3)n+3(2/3)n+1→31, so R=3.
(iii) ∣an∣1/n=n→∞, so R=0: the series converges only at z=0.
Show that z−21z2+31z3−⋯ has radius of convergence 1, converges at every point of the circle ∣z∣=1 except z=−1, and is not absolutely convergent anywhere on the circle.
Hint
On the circle write z=cosθ+isinθ, and look at the real and imaginary parts as series of the kind treated in the previous lesson.
Hint
(−1)n−1cosnθ=−cosn(θ+π).
Solution
∣an+1/an∣=n/(n+1)→1, so R=1. On the circle ∣anzn∣=1/n, so there is no absolute convergence there. With z=cosθ+isinθ the real and imaginary parts are
By Dirichlet’s test (§196, Examples LXXIX, 3 with s=1) both converge unless θ+π is a multiple of 2π (for the cosine series), that is, unless z=−1. At z=−1 the series is −1−21−31−⋯, which diverges to −∞.
ExerciseEven functions●●○Hardy: §201
Suppose f(z)=∑anzn for ∣z∣<R and f(−z)=f(z) there. Show that an=0 for every odd n.
Hint
Write down the power series of f(−z) and use uniqueness.
Solution
For ∣z∣<R, f(−z)=∑an(−z)n=∑(−1)nanzn. Since f(−z)=f(z), the two power series ∑(−1)nanzn and ∑anzn have the same sum on the disc ∣z∣<R, so by uniqueness (−1)nan=an for every n. For odd n this gives an=−an, so an=0. (Example: the series for cosz, which you will meet in Chapter IX, contains only even powers.)
Check yourself
Check yourselfEnter a number
What is the radius of convergence of ∑n≥03nn2zn?
Why
∣an∣∣an+1∣=3n2(n+1)2→31, so R=3. Polynomial factors such as n2 never change the radius.
Check yourselfChoose one
A power series has radius of convergence 1. What can you conclude about its behaviour at z=1?
Why
On the circle everything is possible: ∑zn/n2 converges at 1, ∑zn diverges there, and ∑zn/n diverges at 1 but converges at −1, which also rules out the last option.
Check yourselfTrue or false?
The set of points where a power series converges is always an open disc (or the whole plane, or the single point 0).
Why
False: the set always contains the open disc and lies inside the closed disc, but it may include some, all or none of the boundary circle. For the logarithmic series it is the closed disc minus the single point −1.
Check yourselfTrue or false?
If two power series have the same sum for every real x with 0<x<0.001, their coefficients are equal.
Why
True. Hardy notes that the estimate f(z)=∑n≤manzn+O(∣z∣m+1) holds in particular for real positive z, and the uniqueness proof only lets z→0, so agreement on any interval (0,ϵ) is enough.
Hardy’s language
Summary
Summary
Every power series has a radius of convergence R∈[0,∞]: absolute convergence for ∣z∣<R, no convergence for ∣z∣>R.
On the circle ∣z∣=R anything can happen: ∑zn (nowhere), the logarithmic series (everywhere except −1), ∑zn/n2 (everywhere, absolutely).
If ∣an+1/an∣→λ or ∣an∣1/n→λ, then R=1/λ.
Near 0, f(z)=a0+⋯+amzm+O(∣z∣m+1); hence a function has at most one power series about 0.