In §168 the binomial theorem (1+x)m=∑(nm)xn was proved for −1<x<1 and rationalm, because that was the only kind of power then available. Chapter IX has since defined xm for every real m (§214), so two questions remain, and §222 answers both. Does the theorem hold for irrational m? (Yes, with no new work.) And what happens on the boundary x=±1 of the interval, where the ratio test is silent? There the answer depends on m in an interesting way, and it is a good example of a power series behaving differently at the two ends of its interval of convergence.
Throughout, (nm)=n!m(m−1)⋯(m−n+1) for any real m, and m is not a non-negative integer (otherwise the series stops and there is nothing to discuss).
The idea
Inside the interval the terms (nm)xn shrink geometrically, whatever m is. On the boundary the factor xn no longer helps, and everything depends on how the coefficients (nm) themselves behave. Their size is governed by the ratio
(nm)(n+1m)=n+1∣m−n∣=1−n+1m+1(n>m),
which is just below 1 when m>−1 and at least 1 when m≤−1. In fact ∣(nm)∣ behaves like a constant times n−(m+1).
ExploreOn the edge of convergence
Try this.Start at x=1, m=−0.5. Slide m down through −1 and watch the partial sums. Then set x=−1 and slide m through 0. Finally pick x=0.95 and a negative m such as −2.
∑n≥0(n−0.5)(1)n
nsₙ
partial sums sₙ (red when the series diverges)
(1+x)m
endpoint: converges
endpoint: diverges
your m
The series converges at x = 1, because m > −1; its sum is (1+x)m=0.7071068. After 40 terms s39=0.6623667, and the last term has size 0.0901.
What to notice
At x=1 the partial sums zig-zag towards 2m for every m>−1, more and more slowly as m approaches −1; at m=−1 they flip between 1 and 0 for ever, and below −1 the zig-zag widens. At x=−1 the sums tend to 0 when m>0, but climb steadily to +∞ when m<0. Inside the interval everything converges, though slowly near the edge.
Precise statements
LemmaSize of the binomial coefficientsHardy: cf. Examples LXXXIV, 3
Let m be real and not a non-negative integer. Then (nm)→0 as n→∞ if m>−1, while (nm) never decreases from some n on (so does not tend to 0) if m≤−1. For n>m+1 the coefficients alternate in sign.
TheoremThe binomial series for any real exponentHardy: §222
Let m be real and not a non-negative integer.
If −1<x<1 then (1+x)m=∑n=0∞(nm)xn.
At x=1 the series converges if and only if m>−1, and its sum is then 2m.
At x=−1 the series converges if and only if m>0, and its sum is then 0. For m<0 its partial sums tend to +∞.
Including the trivial case of a non-negative integer m, the series converges at x=−1 exactly when m≥0 (with sum 1 for m=0), as Hardy states it.
Proofs
ProofThe lemma
Strategy
Look at the ratio of consecutive coefficients. When it is at least 1 the coefficients cannot shrink; when it is 1−(m+1)/(n+1) with m+1>0, the product of these ratios tends to 0 because ∑1/n diverges.
The ratio
(n+1m)=(nm)⋅n+1m−n. For n>m the factor n+1m−n is negative (so the signs alternate from then on) and its size is n+1n−m=1−n+1m+1. The coefficients are never 0, because m is not a non-negative integer.
m≤−1
Then 1−n+1m+1≥1, so (nm) never decreases once n>m, and stays at least as large as some fixed non-zero coefficient.
m>−1
Put c=m+1>0 and fix n1>m. For n>n1, using 1−u≤e−u,
and the sum tends to ∞ with n (the harmonic series diverges), so the right side tends to 0.
∎
ProofThe theorem
Strategy
Part 1 is the old proof once the derivative of (1+x)m is known for irrational m. Part 2 uses the integral form of the remainder at the single point x=1. Part 3 uses an exact formula for the partial sums.
Part 1: irrational m
Now (1+x)m=emlog(1+x) (§214), so by the chain rule its derivative is emlog(1+x)⋅1+xm=m(1+x)m−1, the same rule as for rational m. The proof of §168 used nothing else about the power, so it applies unchanged.
Part 2, m≤−1: divergence at x=1
The terms (nm) do not tend to 0 (lemma), so the series cannot converge.
Part 2, m>−1: the sum is 2m
Taylor’s theorem with the integral remainder (§167), applied to f(t)=(1+t)m on [0,1], gives 2m=∑k<n(km)+Rn with
Rn=(n−1)!m(m−1)⋯(m−n+1)∫01(1−t)n−1(1+t)m−ndt.
When n>m we have (1+t)m−n≤1 on [0,1], so the integral is at most ∫01(1−t)n−1dt=1/n. Hence ∣Rn∣≤(nm), which tends to 0 by the lemma.
Part 3: a closed form for the partial sums at x=−1
By Pascal’s rule (km)=(km−1)+(k−1m−1), the sum ∑k=0n(−1)k(km) telescopes:
If m>0 then m−1>−1, so sn=±(nm−1)→0 by the lemma, and 0=(1−1)m as it should be. If m<0, every factor 1+∣m∣/j exceeds 1, so sn increases, and sn≥1+∣m∣∑j≤n1/j→∞.
∎
The picture at x=−1 matches the function: (1+x)m→0 as x→−1 when m>0, but →+∞ when m<0. At x=1 the function is perfectly finite (2m) for every m, and yet the series diverges for m≤−1: a power series can fail to represent a function at a point where the function is harmless.
Worked examples and exercises
ExerciseTwo endpoint sums●○○Hardy: §222
Find the sums of 1−21+2⋅41⋅3−2⋅4⋅61⋅3⋅5+… and of ∑n≥0(−1)n(n1/2).
Hint
Identify each as a binomial series at x=±1: what are (n−1/2) and m?
Solution
(n−1/2)=(−1)n2⋅4⋯(2n)1⋅3⋯(2n−1), so the first series is the binomial series with m=−21 at x=1. Since m>−1 it converges, to 2−1/2=1/2≈0.7071. The second is the series with m=21 at x=−1; since m>0 it converges, to 0. (By contrast m=−21 at x=−1 gives ∑2⋅4⋯(2n)1⋅3⋯(2n−1), which diverges.)
With m=−21, (n−1/2)=(−1)n2⋅4⋯(2n)1⋅3⋯(2n−1). Substituting x2 (which lies in [0,1)) gives the first series and −x2 gives the second. At x=±1 the first becomes the series of the previous exercise at x=1, convergent to 1/2 since m>−1. The second becomes the series at x=−1 with m<0, which diverges to +∞, matching 1/1−x2→∞.
ExerciseSquare roots by hand●●○Hardy: Examples XCIII, 2
Write 67=81+3/64 and expand. Show that the first two terms give 8163=8.1875, that this is too large, and that the error is less than (3/64)2<0.003.
Hint
With u=3/64, 1+u=1+21u−81u2+161u3−…. Is this an alternating series with decreasing terms?
Solution
81+u=8(1+21u−81u2+161u3−…) with u=643. After the first term the signs alternate (the coefficients of (1+u)1/2 alternate from n=1 on), and the terms decrease in size, since the ratio of successive terms is n+1n−1/2u<u<1. So the sum lies between consecutive partial sums. Stopping after 8(1+21u)=8+163=8.1875, the next term −8⋅81u2=−u2 is negative, so the approximation is too large, by less than u2=40969≈0.0022. Indeed 67=8.18535…, an error of 0.00215.
Exercise●●○Hardy: §222 (2); Examples LXXXI, 5
Prove the formula ∑k=0n(−1)k(km)=(−1)n(nm−1) used in part 3, and check it for m=−1.
Hint
Induction on n, using (nm−1)+(n+1m−1)=(n+1m).
Solution
For n=0 both sides are 1. If the formula holds for n, then
by Pascal’s rule (which holds for real m: (nm−1)+(n+1m−1)=(nm−1)n+1(n+1)+(m−1−n)=(n+1m)). For m=−1: (k−1)=(−1)k, so the left side is n+1; and (−1)n(n−2)=(−1)n(−1)n(n+1)=n+1.
Check yourself
Check yourselfChoose one
For which m (not a non-negative integer) does 1+m+2!m(m−1)+3!m(m−1)(m−2)+… converge?
Why
This is the series at x=1. It converges exactly when the terms (nm) tend to 0, which happens exactly when m>−1. At m=−1 the terms are ±1 and the partial sums alternate between 1 and 0. That 2m exists for all m is irrelevant: the series need not represent the function at the edge.
Check yourselfEnter a number
What is the sum of 1−21+83−165+12835−…, the binomial series with m=−21 at x=1? (Decimals are fine.)
Why
m=−21>−1, so the series converges to 2−1/2=1/2≈0.7071. It converges very slowly: the terms decrease like n−1/2, the case m+1=21 of the rule ∣(nm)∣≈Cn−(m+1).
Check yourselfTrue or false?
The binomial series for (1+x)−1/2 converges at x=−1.
Why
False. At x=−1 the series converges only for m≥0. For m=−21 all its terms are positive and its partial sums ∏j≤n(1+2j1) grow without bound, as does (1+x)−1/2 when x→−1.
Check yourselfChoose one
What new fact was needed to extend the binomial theorem of §168 to irrational m?
Why
The old proof used only the derivatives of (1+x)m. Once the power is defined for irrational m via the exponential and logarithm, the chain rule gives the same derivatives, and the proof goes through unchanged. A limit argument in m would also be possible but is not needed.
Hardy’s language
If m+1>0 then un ultimately alternates in sign and decreases steadily to zero, so that the series is convergent.
— G. H. Hardy, §222
Summary
Summary
For every real m and −1<x<1, (1+x)m=∑(nm)xn: the proof of §168 needs only dxd(1+x)m=m(1+x)m−1.
(nm)→0 exactly when m>−1 (for m not a non-negative integer); roughly ∣(nm)∣≈Cn−(m+1).
At x=1: convergent exactly when m>−1, with sum 2m; the integral remainder is at most ∣(nm)∣.
At x=−1: the partial sums are ∏j≤n(1−m/j); convergent (to 0) for m>0, divergent to +∞ for m<0.