Chapter VII · Additional theorems in the differential and integral calculus

The binomial series by Cauchy's remainder

§168 30 min

Cauchy's form of the remainder finishes the proof that (1+x)m=∑(mk)xk(1+x)^m = \sum \binom{m}{k}x^k for every xx between −1-1 and 11, exactly where Lagrange's form gave up.

Why this matters

In §152 the binomial series

(1+x)m=1+mx+m(m−1)2!x2+m(m−1)(m−2)3!x3+…(1+x)^m = 1 + mx + \frac{m(m-1)}{2!}x^2 + \frac{m(m-1)(m-2)}{3!}x^3 + \dots

was proved for 0≤x<10 \le x < 1, but the proof broke down for negative xx. Lagrange’s remainder contains (1+θx)m−n(1+\theta x)^{m-n}, and when x<0x < 0 and nn is large this factor could be enormous if θ\theta happened to be close to 11; knowing only 0<θ<10 < \theta < 1, we could not rule that out. §168 closes the gap with the Cauchy form of the remainder from the previous lesson. It is a short section, but it is the payoff of the whole integral-remainder machinery, and the binomial series is used constantly later (Chapters VIII and IX).

The idea

Cauchy’s form of the remainder for (1+x)m(1+x)^m contains the factor

(1−θ1+θx)n−1.\left(\frac{1-\theta}{1+\theta x}\right)^{n-1}.

This is the factor that rescues us. Whether xx is positive or negative, as long as x>−1x > -1 the fraction is less than 11, so its power is less than 11 whatever θ\theta is. All the dangerous dependence on θ\theta is in a quantity we can bound by 11; what is left is an explicit number ρn\rho_n, which tends to 00 by a ratio argument. Lagrange’s form has no such factor: its (1+θx)−n(1+\theta x)^{-n} can only be bounded by (1−∣x∣)−n(1-|x|)^{-n}, which grows faster than ∣x∣n|x|^n shrinks when ∣x∣>12|x| > \tfrac12.

Explore Three sizes of the same remainder

Try this. Start at x=−0.8x = -0.8, m=12m = \tfrac12. Compare the three families of dots: the true remainder, the bound Lagrange’s form guarantees, and Cauchy’s bound. Move xx slowly from −0.98-0.98 to 0.980.98, watching the hollow orange dots; try other values of mm.

(1+x)0.5(1+x)^{0.5} at x=−0.8x = -0.8: remainder after nn terms (log scale)
nlog₁₀
  • true |Rₙ|
  • Lagrange: the best bound when only 0 < θ < 1 is known
  • Cauchy’s bound ρₙ
  • size 1 (log₁₀ = 0)
At n=40n = 40: true ∣Rn∣=6.59×10−7|R_n| = 6.59 \times 10^{-7}, Cauchy’s bound ρn=1.34×10−5\rho_n = 1.34 \times 10^{-5}, Lagrange’s bound 6.09×10206.09 \times 10^{20}. Lagrange’s bound does not tend to 00 (here x≤−12x \le -\tfrac12): it cannot prove Rn→0R_n \to 0, although the true remainder (and Cauchy’s bound) tend to 00.
What to notice

The true remainder always tends to 00 for ∣x∣<1|x| < 1, and Cauchy’s bound ρn\rho_n follows it down, always a little above it. Lagrange’s bound tends to 00 for x>−12x > -\tfrac12 but grows without limit for x<−12x < -\tfrac12 (watch it turn upwards as xx crosses −12-\tfrac12). A bound that grows proves nothing, even though the quantity it bounds is small.

Precise statements

Here (mk)=m(m−1)⋯(m−k+1)k!\binom{m}{k} = \dfrac{m(m-1)\cdots(m-k+1)}{k!} for any real mm, and (m0)=1\binom{m}{0} = 1.

Theorem The binomial theorem for any rational exponent Hardy: §168; §152 (2)

Let mm be rational and not a non-negative integer. Then for −1<x<1-1 < x < 1,

(1+x)m=∑k=0∞(mk)xk,(1+x)^m = \sum_{k=0}^{\infty}\binom{m}{k}x^k ,

that is, the remainder RnR_n after nn terms tends to 00 as n→∞n \to \infty. More precisely, ∣Rn∣≤ρn|R_n| \le \rho_n where

ρn=∣m∣ (1±∣x∣)m−1∣(m−1n−1)∣∣x∣n(+ if m>1, − if m<1),\rho_n = |m|\,(1 \pm |x|)^{m-1}\left|\binom{m-1}{n-1}\right||x|^n \qquad (+ \text{ if } m > 1,\ - \text{ if } m < 1),

and ρn→0\rho_n \to 0.

When mm is a non-negative integer the series stops after the term in xmx^m and the formula holds for every xx. Once xmx^m is defined for irrational mm (Chapter IX), the same proof works word for word. What happens at x=±1x = \pm 1 depends on mm; that is a question about convergence of series (§222).

Proofs

Proof The binomial series for −1<x<1-1 < x < 1
Strategy

Write Cauchy’s remainder for f(x)=(1+x)mf(x) = (1+x)^m and regroup it so that every appearance of the unknown θ\theta is in a factor we can bound independently of θ\theta. What remains is ρn\rho_n, a fixed number for each nn, and a ratio argument shows ρn→0\rho_n \to 0.

Cauchy’s remainder for (1+x)m(1+x)^m

f(n)(y)=m(m−1)⋯(m−n+1)(1+y)m−nf^{(n)}(y) = m(m-1)\cdots(m-n+1)(1+y)^{m-n}. With a=0a = 0 and h=xh = x, Cauchy’s form (§167) gives, for some θ\theta in [0,1][0, 1],

Rn=m(m−1)⋯(m−n+1)(n−1)! (1−θ)n−1 xn(1+θx)n−m.R_n = \frac{m(m-1)\cdots(m-n+1)}{(n-1)!}\,(1-\theta)^{n-1}\,\frac{x^n}{(1+\theta x)^{n-m}} .

Regroup

Note that m(m−1)⋯(m−n+1)(n−1)!=m(m−1n−1)\dfrac{m(m-1)\cdots(m-n+1)}{(n-1)!} = m\binom{m-1}{n-1}, and (1+θx)m−n=(1+θx)−(n−1)(1+θx)m−1(1+\theta x)^{m-n} = (1+\theta x)^{-(n-1)}(1+\theta x)^{m-1}. So

Rn=m(m−1n−1)xn(1−θ1+θx)n−1(1+θx)m−1.R_n = m\binom{m-1}{n-1}x^n\left(\frac{1-\theta}{1+\theta x}\right)^{n-1}(1+\theta x)^{m-1}.

The key factor is at most 11

1−θ≤1+θx1 - \theta \le 1 + \theta x is the same as −θ≤θx-\theta \le \theta x, which is true because x>−1x > -1 and θ≥0\theta \ge 0. Both sides are positive, so 0≤1−θ1+θx≤10 \le \dfrac{1-\theta}{1+\theta x} \le 1, for positive and negative xx alike, and its (n−1)(n-1)-th power is at most 11.

Bound the last factor

1+θx1 + \theta x lies between 1−∣x∣1 - |x| and 1+∣x∣1 + |x|, both positive. If m>1m > 1 the power (1+θx)m−1(1+\theta x)^{m-1} is at most (1+∣x∣)m−1(1+|x|)^{m-1}; if m<1m < 1 it is at most (1−∣x∣)m−1(1-|x|)^{m-1}; if m=1m = 1 it is 11. Hence ∣Rn∣≤ρn|R_n| \le \rho_n, and θ\theta has disappeared.

ρn→0\rho_n \to 0

The ratio of consecutive values is ρn+1ρn=∣m−n∣n ∣x∣→∣x∣<1\dfrac{\rho_{n+1}}{\rho_n} = \dfrac{|m - n|}{n}\,|x| \to |x| < 1. Choose qq with ∣x∣<q<1|x| < q < 1; then ρn+1<qρn\rho_{n+1} < q\rho_n for all n≥n \ge some NN, so ρn<ρNqn−N→0\rho_n < \rho_N q^{n-N} \to 0 (this is Examples XXVII, 13 in Hardy). Therefore Rn→0R_n \to 0.

The proof never uses the sign of xx except in the third step, and there only x>−1x > -1. That is the whole difference from §152.

Worked examples and exercises

Exercise Why Lagrange fails for x<−12x < -\tfrac12 ●●○ Hardy: §152, §168

Let −1<x<0-1 < x < 0 and n>mn > m. Show that the best bound Lagrange’s form gives, knowing only 0<θ<10 < \theta < 1, is ∣(mn)∣∣x∣n(1−∣x∣)m−n\left|\binom mn\right||x|^n(1-|x|)^{m-n}, and that it tends to infinity when ∣x∣>12|x| > \tfrac12.

Hint
Lagrange: Rn=(mn)xn(1+θx)m−nR_n = \binom mn x^n(1+\theta x)^{m-n}. For n>mn > m the exponent is negative, and 1+θx1 + \theta x can be as small as (nearly) 1−∣x∣1 - |x|.
Hint
Write the bound as (1−∣x∣)m∣(mn)∣(∣x∣1−∣x∣)n(1-|x|)^m\left|\binom mn\right|\left(\dfrac{|x|}{1-|x|}\right)^n and look at the ratio of consecutive terms.

Solution

Since m−n<0m - n < 0 and 1+θx1 + \theta x decreases towards 1−∣x∣1 - |x| as θ→1\theta \to 1, the factor (1+θx)m−n(1+\theta x)^{m-n} can be as large as (nearly) (1−∣x∣)m−n(1-|x|)^{m-n}, and nothing in the information 0<θ<10 < \theta < 1 excludes this. So the best bound is Bn=∣(mn)∣∣x∣n(1−∣x∣)m−n=(1−∣x∣)m∣(mn)∣qnB_n = \left|\binom mn\right||x|^n(1-|x|)^{m-n} = (1-|x|)^m\left|\binom mn\right|q^n with q=∣x∣/(1−∣x∣)q = |x|/(1-|x|).

The ratio Bn+1/Bn=∣m−n∣n+1 q→qB_{n+1}/B_n = \frac{|m-n|}{n+1}\,q \to q. If ∣x∣>12|x| > \tfrac12 then q>1q > 1, so BnB_n eventually grows geometrically and tends to infinity. (If ∣x∣<12|x| < \tfrac12, q<1q < 1 and Lagrange’s bound works; this is why §152 could have handled −12<x<0-\tfrac12 < x < 0 too.)

Exercise How many terms for 1/2\sqrt{1/2}? ●●○ Hardy: §168

Use the series for (1−12)1/2(1 - \tfrac12)^{1/2} to compute 1/2\sqrt{1/2}. How many terms does Cauchy’s bound guarantee for an error below 0.0010.001?

Hint
Here m=12<1m = \tfrac12 < 1, x=−12x = -\tfrac12, so ρn=12⋅(12)−1/2∣(−1/2n−1)∣2−n\rho_n = \tfrac12\cdot(\tfrac12)^{-1/2}\left|\binom{-1/2}{n-1}\right|2^{-n}.
Hint
∣(−1/2k)∣=1⋅3⋅5⋯(2k−1)2⋅4⋅6⋯(2k)\left|\binom{-1/2}{k}\right| = \dfrac{1\cdot3\cdot5\cdots(2k-1)}{2\cdot4\cdot6\cdots(2k)}.

Solution

ρn=22∣(−1/2n−1)∣2−n\rho_n = \tfrac{\sqrt2}{2}\left|\binom{-1/2}{n-1}\right|2^{-n}. With ∣(−1/26)∣=2311024≈0.2256\left|\binom{-1/2}{6}\right| = \tfrac{231}{1024} \approx 0.2256 and ∣(−1/27)∣=4292048≈0.2095\left|\binom{-1/2}{7}\right| = \tfrac{429}{2048} \approx 0.2095:

ρ7=0.7071×0.2256/128≈0.00125,ρ8=0.7071×0.2095/256≈0.00058.\rho_7 = 0.7071 \times 0.2256 / 128 \approx 0.00125, \qquad \rho_8 = 0.7071\times 0.2095/256 \approx 0.00058 .

So 88 terms (1,−14,−132,…1, -\tfrac14, -\tfrac1{32}, \dots, up to the term in x7x^7) are guaranteed to give 1/2=0.70711…\sqrt{1/2} = 0.70711\ldots to within 0.0010.001. The true error after 88 terms is about 0.000090.00009, and 66 terms would already suffice: a bound is a guarantee, not an estimate.

Exercise Coefficients of (1+x)−1/2(1+x)^{-1/2} ●○○ Hardy: §168

Show that (−1/2k)=(−1)k1⋅3⋅5⋯(2k−1)2⋅4⋅6⋯(2k)\binom{-1/2}{k} = (-1)^k\dfrac{1\cdot3\cdot5\cdots(2k-1)}{2\cdot4\cdot6\cdots(2k)}, and write down the expansion of (1−x)−1/2(1-x)^{-1/2} for ∣x∣<1|x| < 1. Why are all its coefficients positive?

Hint
(−1/2k)=(−12)(−32)⋯(−2k−12)k!\binom{-1/2}{k} = \dfrac{(-\frac12)(-\frac32)\cdots(-\frac{2k-1}{2})}{k!}.

Solution

The numerator has kk negative factors −2j−12-\tfrac{2j-1}{2}, j=1,…,kj = 1, \dots, k, so it is (−1)k1⋅3⋯(2k−1)2k(-1)^k\dfrac{1\cdot3\cdots(2k-1)}{2^k}; and 2kk!=2⋅4⋯(2k)2^k k! = 2\cdot4\cdots(2k). Hence the formula. Replacing xx by −x-x multiplies the kk-th term by (−1)k(-1)^k, which cancels the sign:

(1−x)−1/2=1+12x+1⋅32⋅4x2+1⋅3⋅52⋅4⋅6x3+…(∣x∣<1).(1-x)^{-1/2} = 1 + \frac12x + \frac{1\cdot3}{2\cdot4}x^2 + \frac{1\cdot3\cdot5}{2\cdot4\cdot6}x^3 + \dots \qquad (|x| < 1).

All coefficients are positive because (1−x)−1/2(1-x)^{-1/2} and all its derivatives are positive at 00; each derivative of (1−x)−1/2(1-x)^{-1/2} is a positive multiple of a negative power of 1−x1 - x. Here the theorem of this lesson (rather than Lagrange’s form) is needed precisely for 12<x<1\tfrac12 < x < 1, which corresponds to −1<−x<−12-1 < -x < -\tfrac12 in the series for (1+y)−1/2(1 + y)^{-1/2}; there all the terms are positive.

Exercise ρn→0\rho_n \to 0 in detail ●○○ Hardy: Examples XXVII, 13

Show that (mn)xn→0\binom{m}{n}x^n \to 0 as n→∞n \to \infty when ∣x∣<1|x| < 1, for any real mm.

Hint
Look at the ratio of the (n+1)(n+1)-th term to the nn-th.

Solution

Let un=∣(mn)xn∣u_n = \left|\binom mn x^n\right|. If mm is a non-negative integer, un=0u_n = 0 for n>mn > m. Otherwise no unu_n vanishes, and un+1/un=∣m−n∣n+1∣x∣→∣x∣u_{n+1}/u_n = \frac{|m-n|}{n+1}|x| \to |x|. Pick qq with ∣x∣<q<1|x| < q < 1. There is an NN with un+1<qunu_{n+1} < qu_n for n≥Nn \ge N, hence 0<un<uNqn−N0 < u_n < u_N q^{n-N} for n>Nn > N, and qn−N→0q^{n-N} \to 0 (§72). So un→0u_n \to 0. The same argument applied to (m−1n−1)xn\binom{m-1}{n-1}x^n shows ρn→0\rho_n \to 0.

Check yourself

Check yourself True or false?

Lagrange’s form of the remainder, with only the information 0<θ<10 < \theta < 1, proves the binomial series for every xx in (−1,1)(-1, 1).

Why

False. For x<−12x < -\tfrac12 the best bound it gives grows without limit (first exercise). The series is still valid there; proving it needs Cauchy’s form, whose factor {(1−θ)/(1+θx)}n−1≤1\{(1-\theta)/(1+\theta x)\}^{n-1} \le 1 absorbs the dangerous dependence on θ\theta.

Check yourself Choose one

The inequality 1−θ1+θx≤1\dfrac{1-\theta}{1+\theta x} \le 1 (for 0≤θ≤10 \le \theta \le 1) holds exactly when:

Why

Given 1+θx>01 + \theta x > 0, the inequality is 1−θ≤1+θx1 - \theta \le 1 + \theta x, that is, −θ≤θx-\theta \le \theta x, true for x≥−1x \ge -1. It holds for negative xx too: that is the whole point of using Cauchy’s form.

Check yourself Enter a number

What is the coefficient of x2x^2 in the expansion of (1+x)1/2(1+x)^{1/2}?

Why

(1/22)=12⋅(−12)2=−18\binom{1/2}{2} = \dfrac{\frac12\cdot(-\frac12)}{2} = -\tfrac18. So 1+x=1+12x−18x2+…\sqrt{1+x} = 1 + \tfrac12x - \tfrac18x^2 + \dots.

Check yourself Choose one

In the widget, Cauchy’s bound ρn\rho_n always lies above the true remainder ∣Rn∣|R_n|. Why?

Why

∣Rn∣=∣m∣∣(m−1n−1)∣∣x∣n⋅(1−θ1+θx)n−1⋅(1+θx)m−1|R_n| = |m|\left|\binom{m-1}{n-1}\right||x|^n\cdot\left(\frac{1-\theta}{1+\theta x}\right)^{n-1}\cdot(1+\theta x)^{m-1} exactly, for Cauchy’s θ\theta. Replacing the second factor by 11 and the third by its largest possible value can only increase it. That is what “bound” means; the theorem says nothing about how close the bound is.

Hardy’s language

Summary

Summary
  • For (1+x)m(1+x)^m, Cauchy’s remainder is m(m−1n−1)xn(1−θ1+θx)n−1(1+θx)m−1m\binom{m-1}{n-1}x^n\left(\frac{1-\theta}{1+\theta x}\right)^{n-1}(1+\theta x)^{m-1}.
  • For −1<x<1-1 < x < 1, the factor 1−θ1+θx\frac{1-\theta}{1+\theta x} is at most 11, whatever θ\theta and whatever the sign of xx; this removes θ\theta from the bound.
  • The resulting ρn\rho_n tends to 00 by a ratio argument, so (1+x)m=∑(mk)xk(1+x)^m = \sum\binom mk x^k on the whole interval (−1,1)(-1, 1).
  • Lagrange’s form fails for x<−12x < -\tfrac12 not because the remainder is large but because its bound is too crude: the choice of remainder form matters.