Chapter IX · The logarithmic, exponential and circular functions of a real variable

log⁡x\log x as a limit: n(xn−1)→log⁡xn(\sqrt[n]{x} - 1) \to \log x

§216 25 min

The logarithm as the limit of n(x1/n−1)n(x^{1/n} - 1), squeezed between two integrals; with the old method of computing logarithms by repeated square roots, and limits of nn-th roots.

Why this matters

In §75 Hardy showed that n(xn−1)n(\sqrt[n]{x} - 1) tends to a limit for every x>0x > 0, without being able to say what the limit was. §216 names it: it is log⁡x\log x. This is the mirror image of the previous lesson (exe^x as a limit of powers, log⁡x\log x as a limit of roots), and it is how logarithms were first computed: take enough square roots of xx that the result is very close to 11, subtract 11, and multiply back up.

The idea

The integrand 1/t=t−11/t = t^{-1} has no power antiderivative, but its neighbours t−1+1/nt^{-1+1/n} and t−1−1/nt^{-1-1/n} do. For t>1t > 1 the first lies a little above 1/t1/t and the second a little below, and both approach 1/t1/t as nn grows. Integrating from 11 to xx:

∫1xt−1−1/n dt=n(1−x−1/n),∫1xt−1+1/n dt=n(x1/n−1).\int_1^x t^{-1-1/n}\,dt = n\left(1 - x^{-1/n}\right), \qquad \int_1^x t^{-1+1/n}\,dt = n\left(x^{1/n} - 1\right).

So the area log⁡x\log x is squeezed between two expressions built from nn-th roots alone.

Explore Squeeze the logarithm between two power integrands

Try this. With x=3x = 3 and n=2n = 2, compare the three areas from 11 to xx in the readout. Then increase nn, and try xx less than 11.

t
  • t−1+1/nt^{-1+1/n}
  • 1/t1/t
  • t−1−1/nt^{-1-1/n}
  • x (drag)
n(1−x−1/n)=0.845299  <  log⁡x=1.09861  <  n(x1/n−1)=1.4641n(1 - x^{-1/n}) = 0.845299 \;<\; \log x = 1.09861 \;<\; n(x^{1/n} - 1) = 1.4641. Width of the bracket: 0.6190.619, which shrinks roughly like (log⁡x)2/n(\log x)^2/n.
What to notice

For small nn the bracket is wide (n=2n = 2: 0.845<log⁡3<1.4640.845 < \log 3 < 1.464), but it narrows roughly like (log⁡x)2/n(\log x)^2/n, and both ends converge on log⁡3=1.0986\log 3 = 1.0986. For x<1x < 1 the curves swap order on [x,1][x, 1] but, the area now counting negatively, the same inequalities hold.

Explore The sequence n(xn−1)n(\sqrt[n]{x} - 1)

Try this. This is the sequence of §75 for x=2x = 2. Find an n0n_0 for δ=0.01\delta = 0.01, then change xx.

ϕ(n)=n(x1/n−1)\phi(n) = n\left(x^{1/n}-1\right) , candidate limit l=0.6931l = 0.6931
nφ(n)
  • band l ± δ
  • l
  • n₀ (drag)
  • n ≥ n₀, good
  • n ≥ n₀, bad
24 terms with n ≥ n₀ fall outside the band: n = 1, 2, 3, 4, 5, ….
What to notice

The terms decrease towards log⁡2=0.6931\log 2 = 0.6931. The error is about (log⁡x)2/(2n)(\log x)^2/(2n), so for δ=0.01\delta = 0.01 and x=2x = 2 you need nn around 2525.

Precise statements

Theorem The integral inequalities Hardy: §215 (3)

For every x>0x > 0 with x≠1x \ne 1 and every positive integer nn,

n(1−x−1/n)<log⁡x<n(x1/n−1).n\left(1 - x^{-1/n}\right) < \log x < n\left(x^{1/n} - 1\right).
Theorem The logarithmic limit Hardy: §216

For every x>0x > 0,

lim⁡n→∞n(1−x−1/n)=lim⁡n→∞n(x1/n−1)=log⁡x.\lim_{n\to\infty} n\left(1 - x^{-1/n}\right) = \lim_{n\to\infty} n\left(x^{1/n} - 1\right) = \log x .

Proofs

Proof The integral inequalities
Strategy

Compare integrands on [1,x][1, x] and integrate. For x<1x < 1, reduce to the case x>1x > 1 by the substitution x=1/zx = 1/z.

Compare integrands for t>1t > 1

t1/n>1t^{1/n} > 1, so t−1−1/n=t−1/t1/n<t−1<t−1 t1/n=t−1+1/nt^{-1-1/n} = t^{-1}/t^{1/n} < t^{-1} < t^{-1}\,t^{1/n} = t^{-1+1/n}.

Integrate over [1,x][1, x], x>1x > 1

The antiderivatives are −n t−1/n-n\,t^{-1/n} and n t1/nn\,t^{1/n}, so

n(1−x−1/n)=∫1xt−1−1/ndt<log⁡x<∫1xt−1+1/ndt=n(x1/n−1).n(1 - x^{-1/n}) = \int_1^x t^{-1-1/n}dt < \log x < \int_1^x t^{-1+1/n}dt = n(x^{1/n} - 1).

The case 0<x<10 < x < 1

Put z=1/x>1z = 1/x > 1. The case just proved gives n(1−z−1/n)<log⁡z<n(z1/n−1)n(1 - z^{-1/n}) < \log z < n(z^{1/n} - 1). Multiply by −1-1, reversing the inequalities, and use log⁡z=−log⁡x\log z = -\log x, z−1/n=x1/nz^{-1/n} = x^{1/n}, z1/n=x−1/nz^{1/n} = x^{-1/n}: we get n(x1/n−1)>log⁡x>n(1−x−1/n)n(x^{1/n} - 1) > \log x > n(1 - x^{-1/n}), the same statement.

With x=eyx = e^y these are exactly the two bounds (1+y/n)n<ey<(1−y/n)−n(1 + y/n)^n < e^y < (1 - y/n)^{-n} of the previous lesson, rewritten: so the two limits of §215 and §216 are two readings of one inequality.

Proof The logarithmic limit
Strategy

Show that the difference of the two outer terms tends to 00. Since log⁡x\log x is trapped between them, both tend to log⁡x\log x.

Factor the difference

n(x1/n−1)−n(1−x−1/n)=n(x1/n−1) (1−x−1/n).n(x^{1/n} - 1) - n(1 - x^{-1/n}) = n(x^{1/n} - 1)\,(1 - x^{-1/n}).

(Multiply out: n(x1/n−1−1+x−1/n)=n(x1/n−1)(1−x−1/n)n(x^{1/n} - 1 - 1 + x^{-1/n}) = n(x^{1/n} - 1)(1 - x^{-1/n}), since (x1/n−1)(1−x−1/n)=x1/n−2+x−1/n(x^{1/n} - 1)(1 - x^{-1/n}) = x^{1/n} - 2 + x^{-1/n}.)

The first factor is bounded

By §75, n(x1/n−1)n(x^{1/n} - 1) tends to a limit, so it is bounded. (Without that result: it equals ehlog⁡x−1h\frac{e^{h\log x} - 1}{h} with h=1/nh = 1/n, which tends to the derivative of ehlog⁡xe^{h\log x} at h=0h = 0, namely log⁡x\log x.)

The second factor tends to 00

x−1/n→1x^{-1/n} \to 1 (Examples XXVII, 10), so 1−x−1/n→01 - x^{-1/n} \to 0.

Squeeze

A bounded sequence times one tending to 00 tends to 00. So the width of the bracket tends to 00, and both ends converge to the number log⁡x\log x that lies between them.

Worked examples and exercises

Exercise Logarithms by square roots ●●○ Hardy: §216

Using n=210=1024n = 2^{10} = 1024 (that is, ten successive square roots of 22), show that 0.6929<log⁡2<0.69340.6929 < \log 2 < 0.6934. How many correct decimals does this give?

Hint
21/1024=1.000677130…2^{1/1024} = 1.000677130\ldots, obtained by ten square roots. Then compute both sides of the integral inequality.

Solution

With 21/1024=1.00067713062^{1/1024} = 1.0006771306 and 2−1/1024=0.99932332752^{-1/1024} = 0.9993233275:

1024 (1−0.9993233275)=0.69291<log⁡2<1024 (1.0006771306−1)=0.69338.1024\,(1 - 0.9993233275) = 0.69291 < \log 2 < 1024\,(1.0006771306 - 1) = 0.69338.

The bracket has width 0.00047≈(log⁡2)2/10240.00047 \approx (\log 2)^2/1024, so it fixes log⁡2=0.693…\log 2 = 0.693\ldots to three decimals; the true value is 0.6931470.693147. The average of the two ends, 0.6931470.693147, is far better, because the errors at the two ends are nearly equal and opposite (each about (log⁡2)2/(2n)(\log 2)^2/(2n)). This is essentially the method by which Henry Briggs computed his tables of logarithms in the early seventeenth century, by hand, with dozens of successive square roots.

Exercise ●●○ Hardy: Examples LXXXVIII, 33

Show that if un>0u_n > 0 and un+1/un→lu_{n+1}/u_n \to l, where l>0l > 0, then unn→l\sqrt[n]{u_n} \to l.

Hint
Take logarithms: log⁡un+1−log⁡un→log⁡l\log u_{n+1} - \log u_n \to \log l.
Hint
If the differences an+1−ana_{n+1} - a_n of a sequence tend to λ\lambda, then an/n→λa_n/n \to \lambda (Ch. IV, Misc. Ex. 17, an averaging argument).

Solution

Let an=log⁡una_n = \log u_n. Since log⁡\log is continuous at l>0l > 0, an+1−an=log⁡(un+1/un)→log⁡l=λa_{n+1} - a_n = \log(u_{n+1}/u_n) \to \log l = \lambda.

Averaging step. Given δ>0\delta > 0 choose NN with ∣ak+1−ak−λ∣<δ|a_{k+1} - a_k - \lambda| < \delta for k≥Nk \ge N. Then for n>Nn > N,

an=aN+∑k=Nn−1(ak+1−ak)=aN+(n−N)λ+En,∣En∣<(n−N)δ,a_n = a_N + \sum_{k=N}^{n-1}(a_{k+1} - a_k) = a_N + (n - N)\lambda + E_n, \qquad |E_n| < (n - N)\delta,

so ∣ann−λ∣≤∣aN∣+N∣λ∣n+δ<2δ\left|\dfrac{a_n}{n} - \lambda\right| \le \dfrac{|a_N| + N|\lambda|}{n} + \delta < 2\delta for large nn. Hence an/n→λa_n/n \to \lambda.

Finally unn=ean/n→eλ=l\sqrt[n]{u_n} = e^{a_n/n} \to e^{\lambda} = l, by continuity of the exponential.

Exercise ●●○ Hardy: Examples LXXXVIII, 34

Show that n!n∼n/e\sqrt[n]{n!} \sim n/e, that is, n!nn→1e\dfrac{\sqrt[n]{n!}}{n} \to \dfrac1e.

Hint
Apply the previous exercise to un=n!/nnu_n = n!/n^n.

Solution

With un=n!/nnu_n = n!/n^n,

un+1un=(n+1)!(n+1)n+1⋅nnn!=nn(n+1)n=(1+1n)−n→1e,\frac{u_{n+1}}{u_n} = \frac{(n+1)!}{(n+1)^{n+1}}\cdot\frac{n^n}{n!} = \frac{n^n}{(n+1)^n} = \left(1 + \frac1n\right)^{-n} \to \frac1e,

by the previous lesson. So unn=n!n/n→1/e\sqrt[n]{u_n} = \sqrt[n]{n!}/n \to 1/e. (For n=100n = 100 the ratio is 0.37990.3799, against 1/e=0.36791/e = 0.3679: the convergence is slow, with error about log⁡(2πn)2n\frac{\log(2\pi n)}{2n} in the logarithm, as Stirling’s formula shows.)

Exercise ●○○ Hardy: Examples LXXXVIII, 35

Show that (2n)!(n!)2n→4\sqrt[n]{\dfrac{(2n)!}{(n!)^2}} \to 4.

Hint
Use the ratio criterion of Examples LXXXVIII, 33.

Solution

With un=(2n)!/(n!)2u_n = (2n)!/(n!)^2,

un+1un=(2n+2)(2n+1)(n+1)2=2(2n+1)n+1→4.\frac{u_{n+1}}{u_n} = \frac{(2n+2)(2n+1)}{(n+1)^2} = \frac{2(2n+1)}{n+1} \to 4 .

So unn→4\sqrt[n]{u_n} \to 4. (Consistent with the binomial theorem: unu_n is the largest of the 2n+12n+1 coefficients of (1+1)2n=4n(1 + 1)^{2n} = 4^n, so 4n/(2n+1)≤un≤4n4^n/(2n+1) \le u_n \le 4^n.)

Check yourself

Check yourself Choose one

For x>1x > 1, on which side of log⁡x\log x does n(x1/n−1)n(x^{1/n} - 1) lie?

Why

It is the integral of t−1+1/nt^{-1+1/n}, which exceeds 1/t1/t on (1,x)(1, x), so it is always above log⁡x\log x, approaching from above as nn grows. For 0<x<10 < x < 1 it is also above (both are negative there, and it is the less negative).

Check yourself Enter a number

Compute the upper bound n(x1/n−1)n(x^{1/n} - 1) for log⁡3\log 3 when n=2n = 2.

Why

2(3−1)=2×0.73205=1.46412(\sqrt3 - 1) = 2 \times 0.73205 = 1.4641. The lower bound is 2(1−1/3)=0.84532(1 - 1/\sqrt3) = 0.8453, and log⁡3=1.0986\log 3 = 1.0986 lies between.

Check yourself True or false?

The inequalities n(1−x−1/n)<log⁡x<n(x1/n−1)n(1 - x^{-1/n}) < \log x < n(x^{1/n} - 1) and (1+y/n)n<ey<(1−y/n)−n(1 + y/n)^n < e^y < (1 - y/n)^{-n} are equivalent (with x=eyx = e^y, ∣y∣<n|y| < n).

Why

True. Putting x=eyx = e^y, the right-hand inequality reads y/n<ey/n−1y/n < e^{y/n} - 1, i.e. 1+y/n<ey/n1 + y/n < e^{y/n}, and raising to the nn-th power (both sides positive) gives (1+y/n)n<ey(1 + y/n)^n < e^y. The left-hand one becomes 1−y/n<e−y/n1 - y/n < e^{-y/n}, which gives the other bound. Both are forms of 1+u<eu1 + u < e^u.

Hardy’s language

Summary

Summary
  • Comparing 1/t1/t with t−1∓1/nt^{-1 \mp 1/n} gives n(1−x−1/n)<log⁡x<n(x1/n−1)n(1 - x^{-1/n}) < \log x < n(x^{1/n} - 1) for every x>0x > 0, x≠1x \ne 1.
  • The bracket has width n(x1/n−1)(1−x−1/n)→0n(x^{1/n} - 1)(1 - x^{-1/n}) \to 0, so both ends tend to log⁡x\log x: the limit of §75 is the logarithm.
  • With x=eyx = e^y this is the same inequality as (1+y/n)n<ey<(1−y/n)−n(1 + y/n)^n < e^y < (1 - y/n)^{-n}.
  • Taking logarithms turns ratio limits into nn-th-root limits: un+1/un→lu_{n+1}/u_n \to l implies unn→l\sqrt[n]{u_n} \to l; hence n!n∼n/e\sqrt[n]{n!} \sim n/e.