Chapter VIII · The convergence of infinite series and infinite integrals

Multiplication of series

§202 30 min

Two absolutely convergent series may be multiplied term by term and collected along diagonals, as polynomials are; for conditionally convergent series this can fail.

Why this matters

To multiply two polynomials, multiply every term of one by every term of the other and collect equal powers. Power series are “infinite polynomials”, and we would like to multiply them the same way: that is how Hardy proves the binomial theorem for negative exponents and how, in Chapter IX, the exponential series is shown to satisfy exey=ex+ye^{x}e^{y} = e^{x+y}. §177 justified this for series of positive terms. §202 extends it to all absolutely convergent series, and an example shows that the word “absolutely” cannot be dropped.

The idea

Write all the products umvnu_m v_n in a square array, with mm numbering the columns and nn the rows (as in Hardy’s diagram). The product (∑um)(∑vn)\big(\sum u_m\big)\big(\sum v_n\big) is, in some sense, the sum of the whole array, but an infinite array has no built-in order. Two orders are natural.

  • By squares. The products with m,n≤km, n \le k fill a square, and their total is (u0+⋯+uk)(v0+⋯+vk)=sktk(u_0 + \cdots + u_k)(v_0 + \cdots + v_k) = s_k t_k. This always tends to stst, the product of the two sums.
  • By diagonals. The products with m+n=km + n = k are exactly the terms that contribute to zkz^k when the series are power series. Their total is
wk=u0vk+u1vk−1+⋯+ukv0,w_k = u_0 v_k + u_1 v_{k-1} + \cdots + u_k v_0 ,

and ∑wk\sum w_k is the Cauchy product of the two series.

The question is whether the two orders give the same answer. For absolutely convergent series the array behaves like a series of positive terms, and order is irrelevant.

Explore Adding up the array in two orders

Try this. For each pair of series, switch between “diagonals” and “squares” and step kk from 00 to 77. Then look at the plot: the filled dots are partial sums of the Cauchy product, the hollow dots the products sktks_k t_k. Finish with the last pair, ∑(−1)n/n+1\sum (-1)^n/\sqrt{n+1} squared.

um=2−m,vn=3−nu_m = 2^{-m}, \quad v_n = 3^{-n}m=00m=11m=22m=33m=44m=55m=66m=77n1.000.330.110.040.01···0.500.170.060.020.01···0.250.080.030.01····0.130.040.01·····0.060.020.01·····0.030.01······0.020.01······0.01·······
ksums
  • w₀ + … + wₖ (diagonals: the Cauchy product)
  • sₖ tₖ (squares: product of partial sums)
  • st, the product of the sums
Group the products by
The diagonal m+n=3m + n = 3 sums to w3=0.30093w_{3} = 0.30093, and w0+⋯+w3=2.662w_0 + \cdots + w_{3} = 2.662. After 40 diagonals the Cauchy product has reached 3; the product of the sums is 2×32=3≈32 \times \tfrac32 = 3 \approx 3. Both series are absolutely convergent, so every way of adding up the array gives st.
What to notice

For the first three pairs (absolutely convergent) both kinds of dots close in on the dashed line stst. For the alternating harmonic series squared the Cauchy product still converges, slowly. For ∑(−1)n/n+1\sum(-1)^n/\sqrt{n+1} the cells on each diagonal all have the same sign, the diagonal sums wkw_k never shrink, and the filled dots swing up and down for ever, although the hollow dots still converge.

Precise statements

Theorem Multiplication of absolutely convergent series Hardy: §202

If ∑un\sum u_n and ∑vn\sum v_n are absolutely convergent, with sums ss and tt, then their Cauchy product ∑wn\sum w_n, where wn=u0vn+u1vn−1+⋯+unv0w_n = u_0v_n + u_1v_{n-1} + \cdots + u_nv_0, is absolutely convergent and its sum is stst.

Corollary Multiplying power series Hardy: Examples LXXXI, 1

If ∣z∣|z| is less than the radius of convergence of both ∑anzn\sum a_n z^n and ∑bnzn\sum b_n z^n, then

(∑anzn)(∑bnzn)=∑cnzn,cn=a0bn+a1bn−1+⋯+anb0.\Big(\sum a_n z^n\Big)\Big(\sum b_n z^n\Big) = \sum c_n z^n, \qquad c_n = a_0b_n + a_1b_{n-1} + \cdots + a_nb_0 .

The corollary is the theorem with un=anznu_n = a_n z^n and vn=bnznv_n = b_n z^n, which are absolutely convergent inside both circles of convergence (§200); then wn=cnznw_n = c_n z^n.

Proofs

Proof Multiplication of absolutely convergent series
Strategy

Treat the whole array umvnu_m v_n as one series. It is absolutely convergent, so (§192) its sum does not depend on the order. Taken by squares it gives stst; taken by diagonals it gives the Cauchy product.

The array of moduli has a finite total

Let σ=∑∣un∣\sigma = \sum |u_n| and τ=∑∣vn∣\tau = \sum |v_n|. Any finite collection of products ∣umvn∣|u_m v_n| lies inside some square m,n≤Km, n \le K, whose total is (∣u0∣+⋯+∣uK∣)(∣v0∣+⋯+∣vK∣)≤στ(|u_0| + \cdots + |u_K|)(|v_0| + \cdots + |v_K|) \le \sigma\tau.

So any arrangement of the array is absolutely convergent

Arrange all the products umvnu_m v_n in any order as a single series. By the previous step its series of moduli has bounded partial sums, so it converges: the series is absolutely convergent, and by §192 all arrangements have the same sum SS.

Taken by squares, the sum is stst

Arrange the products square by square (first m,n≤0m, n \le 0, then the new products with max⁡(m,n)=1\max(m, n) = 1, and so on). The partial sums at the end of the kk-th square are sktk→sts_k t_k \to st. A subsequence of the partial sums of a convergent series has the same limit, so S=stS = st.

Taken by diagonals, the sum is the Cauchy product

Arrange the products diagonal by diagonal. The partial sums at the end of the diagonals are w0+⋯+wkw_0 + \cdots + w_k, a subsequence of the partial sums of a series with sum SS. So ∑wk\sum w_k converges to S=stS = st.

Absolutely

∣wk∣≤∣u0∣∣vk∣+⋯+∣uk∣∣v0∣|w_k| \le |u_0||v_k| + \cdots + |u_k||v_0|, and the right-hand sides are the diagonal sums of the array of moduli, whose series converges to στ\sigma\tau. By comparison ∑∣wk∣\sum |w_k| converges.

The proof is Hardy’s for positive series (§177), with Dirichlet’s theorem for absolutely convergent series (§192) in place of the positive one. Its weak point is visible: everything rests on the array having a finite total of moduli. For ∑(−1)n/n+1\sum (-1)^n/\sqrt{n+1} the array of moduli 1/(m+1)(n+1)1/\sqrt{(m+1)(n+1)} has an infinite total, and the diagonals and the squares then part company, as the last exercise shows.

Worked examples and exercises

Exercise Squaring the geometric series ●○○ Hardy: Examples LXXXI, 3–4

Show that (1−z)−2=1+2z+3z2+⋯(1 - z)^{-2} = 1 + 2z + 3z^2 + \cdots and (1−z)−3=1+3z+6z2+⋯(1 - z)^{-3} = 1 + 3z + 6z^2 + \cdots for ∣z∣<1|z| < 1, the coefficient of znz^n in the second being 12(n+1)(n+2)\tfrac12(n+1)(n+2).

Hint
Multiply ∑zn\sum z^n by itself, then multiply the result by ∑zn\sum z^n again.

Solution

∑zn\sum z^n is absolutely convergent for ∣z∣<1|z| < 1 with sum 1/(1−z)1/(1-z). Squaring by the corollary, cn=∑k=0n1⋅1=n+1c_n = \sum_{k=0}^{n} 1\cdot 1 = n + 1, so (1−z)−2=∑(n+1)zn(1-z)^{-2} = \sum (n+1)z^n, again absolutely convergent for ∣z∣<1|z| < 1 (ratio test). Multiplying this by ∑zn\sum z^n: the coefficient of znz^n is ∑k=0n(k+1)=12(n+1)(n+2)\sum_{k=0}^n (k+1) = \tfrac12(n+1)(n+2), so (1−z)−3=1+3z+6z2+10z3+⋯(1 - z)^{-3} = 1 + 3z + 6z^2 + 10z^3 + \cdots.

Exercise Dividing by 1−z1 - z ●○○ Hardy: Examples LXXXI, 2

If f(z)=∑anznf(z) = \sum a_n z^n for ∣z∣<R|z| < R, show that for ∣z∣<min⁡(R,1)|z| < \min(R, 1)

f(z)1−z=∑snzn,sn=a0+a1+⋯+an.\frac{f(z)}{1 - z} = \sum s_n z^n, \qquad s_n = a_0 + a_1 + \cdots + a_n .
Solution

For such zz both ∑anzn\sum a_n z^n and ∑zn\sum z^n are absolutely convergent. Their Cauchy product has cn=a0⋅1+a1⋅1+⋯+an⋅1=snc_n = a_0\cdot 1 + a_1 \cdot 1 + \cdots + a_n \cdot 1 = s_n, and its sum is f(z)⋅11−zf(z)\cdot\frac{1}{1-z}.

Exercise Negative integral exponents ●●○ Hardy: Examples LXXXI, 5

Prove that for ∣z∣<1|z| < 1 and every positive integer mm,

1(1−z)m=∑n≥0(m+n−1n)zn=1+mz+m(m+1)1⋅2z2+⋯ .\frac{1}{(1 - z)^m} = \sum_{n \ge 0} \binom{m + n - 1}{n} z^n = 1 + mz + \frac{m(m+1)}{1\cdot 2}z^2 + \cdots .
Hint
Induction on mm, using the previous exercise.
Hint
You will need ∑k=0n(m+k−1k)=(m+nn)\sum_{k=0}^n \binom{m + k - 1}{k} = \binom{m + n}{n}. Prove it by induction on nn, using Pascal’s rule.
Solution

For m=1m = 1 this is the geometric series. Suppose it holds for mm. By the previous exercise with f(z)=(1−z)−mf(z) = (1-z)^{-m} (radius 11),

1(1−z)m+1=∑nsnzn,sn=∑k=0n(m+k−1k).\frac{1}{(1-z)^{m+1}} = \sum_n s_n z^n, \qquad s_n = \sum_{k=0}^{n}\binom{m+k-1}{k} .

The identity sn=(m+nn)s_n = \binom{m+n}{n} holds for n=0n = 0 (both are 11), and if it holds for n−1n - 1 then sn=(m+n−1n−1)+(m+n−1n)=(m+nn)s_n = \binom{m+n-1}{n-1} + \binom{m+n-1}{n} = \binom{m+n}{n} by Pascal’s rule. Since ((m+1)+n−1n)=(m+nn)\binom{(m+1) + n - 1}{n} = \binom{m+n}{n}, this is the formula for m+1m + 1.

Exercise The addition theorem for the exponential series ●●○ Hardy: Examples LXXXI, 7

Let f(z)=∑n≥0zn/n!f(z) = \sum_{n\ge0} z^n/n!. Show that f(z)f(z′)=f(z+z′)f(z)f(z') = f(z + z') for all complex z,z′z, z'.

Hint
Both series are absolutely convergent everywhere. Compute the diagonal sum wnw_n and recognise the binomial theorem.

Solution

The series converges absolutely for every zz (§198), so the theorem applies:

wn=∑k=0nzkk!⋅z′ n−k(n−k)!=1n!∑k=0n(nk)zkz′ n−k=(z+z′)nn!.w_n = \sum_{k=0}^{n} \frac{z^k}{k!}\cdot\frac{z'^{\,n-k}}{(n-k)!} = \frac{1}{n!}\sum_{k=0}^{n}\binom{n}{k} z^k z'^{\,n-k} = \frac{(z + z')^n}{n!} .

So f(z)f(z′)=∑(z+z′)n/n!=f(z+z′)f(z)f(z') = \sum (z + z')^n/n! = f(z + z'). This is the key property of eze^z, used in Chapter IX; the second pair in the figure (with z=z′=1z = z' = 1) is e⋅e=e2e \cdot e = e^2.

Exercise Failure of the multiplication theorem ●●● Hardy: Examples LXXXI, 9

Let un=vn=(−1)n/n+1u_n = v_n = (-1)^n/\sqrt{n+1}. Both series converge (alternating series test). Show that their Cauchy product does not converge.

Hint
Write wnw_n as (−1)n(-1)^n times a sum of n+1n + 1 positive terms.
Hint
By the AM–GM inequality, (r+1)(n+1−r)≤12(n+2)\sqrt{(r+1)(n+1-r)} \le \tfrac12(n+2).

Solution

urvn−r=(−1)r(−1)n−r/(r+1)(n−r+1)=(−1)n/(r+1)(n+1−r)u_r v_{n-r} = (-1)^r(-1)^{n-r}/\sqrt{(r+1)(n-r+1)} = (-1)^n/\sqrt{(r+1)(n+1-r)}, so

wn=(−1)n∑r=0n1(r+1)(n+1−r).w_n = (-1)^n\sum_{r=0}^{n}\frac{1}{\sqrt{(r+1)(n+1-r)}} .

The two factors under the root add up to n+2n + 2, so their geometric mean is at most 12(n+2)\tfrac12(n+2), and each of the n+1n + 1 terms is at least 2/(n+2)2/(n+2). Hence ∣wn∣≥2(n+1)/(n+2)≥1|w_n| \ge 2(n+1)/(n+2) \ge 1 for all nn, the terms do not tend to 00, and ∑wn\sum w_n cannot converge. (In fact ∣wn∣→π|w_n| \to \pi, as the figure suggests: the sum is a Riemann sum for ∫01dx/x(1−x)=π\int_0^1 dx/\sqrt{x(1-x)} = \pi.)

Check yourself

Check yourself Enter a number

What is the coefficient of z3z^3 in (1+z+z2+⋯ )2(1 + z + z^2 + \cdots)^2?

Why

c3=a0b3+a1b2+a2b1+a3b0=4c_3 = a_0b_3 + a_1b_2 + a_2b_1 + a_3b_0 = 4: one product for each cell on the diagonal m+n=3m + n = 3. In general cn=n+1c_n = n + 1.

Check yourself True or false?

If ∑un\sum u_n and ∑vn\sum v_n both converge, then their Cauchy product converges.

Why

False: ∑(−1)n/n+1\sum (-1)^n/\sqrt{n+1} multiplied by itself gives terms of modulus at least 11. The theorem needs absolute convergence (Mertens later showed that absolute convergence of one of the two series is enough).

Check yourself Choose one

For any two convergent series with sums ss and tt, which of these sequences always tends to stst?

Why

sktks_k t_k is a product of two convergent sequences, so it tends to stst with no further conditions. The diagonal partial sums are a different selection of products, and for conditionally convergent series they may fail to converge at all.

Hardy’s language

Summary

Summary
  • The Cauchy product of ∑un\sum u_n and ∑vn\sum v_n is ∑wk\sum w_k with wk=u0vk+⋯+ukv0w_k = u_0v_k + \cdots + u_kv_0: the products along the diagonal m+n=km + n = k of the array.
  • If both series are absolutely convergent, the Cauchy product is absolutely convergent with sum stst.
  • Proof: the array has a finite total of moduli, so its sum does not depend on the order, and the square order gives sktk→sts_kt_k \to st.
  • Power series may be multiplied like polynomials inside both circles of convergence; for instance (1−z)−m=∑(m+n−1n)zn(1-z)^{-m} = \sum \binom{m+n-1}{n}z^n and ezez′=ez+z′e^z e^{z'} = e^{z+z'}.
  • Without absolute convergence it can fail: (∑(−1)n/n+1)2\big(\sum (-1)^n/\sqrt{n+1}\big)^2 has a divergent Cauchy product.