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

exe^x as a limit: (1+x/n)n→ex(1 + x/n)^n \to e^x

§215 35 min

The compound-interest limit (1+x/n)n→ex(1 + x/n)^n \to e^x, proved twice, once from log⁡(1+h)/h→1\log(1+h)/h \to 1 and once by an explicit squeeze; and the identification of ee with the limit of Chapter IV.

Why this matters

In §73 Hardy proved that (1+1/n)n(1 + 1/n)^n tends to a limit, and called it ee “provisionally”. Now we have a second ee, defined by log⁡e=1\log e = 1. §215 shows they are the same number, and more: (1+x/n)n→ex(1 + x/n)^n \to e^x for every real xx. This is the formula behind compound interest, population growth and the probability of “no success in nn tries”, and it is often taken as the definition of exe^x. Hardy gives two proofs; the second gives explicit error bounds.

The idea

Taking logarithms turns the power into a product:

log⁡(1+xn)n=nlog⁡(1+xn)=x⋅log⁡(1+h)h,h=xn.\log\left(1 + \frac{x}{n}\right)^n = n\log\left(1 + \frac{x}{n}\right) = x \cdot \frac{\log(1 + h)}{h}, \qquad h = \frac{x}{n}.

As n→∞n \to \infty, h→0h \to 0, and log⁡(1+h)/h→1\log(1+h)/h \to 1 (the slope of log⁡\log at 11). So the logarithm of (1+x/n)n(1 + x/n)^n tends to xx, and the power itself tends to exe^x.

The second proof is more concrete. The basic inequality log⁡(1+u)<u\log(1 + u) < u of the first lesson, read through the exponential, says 1+u<eu1 + u < e^u. Applied with u=y/nu = y/n and with u=−y/nu = -y/n it traps eye^y between two explicit expressions.

Explore Squeeze eye^y between two powers

Try this. Start with n=4n = 4 and drag the point along eye^y. Then increase nn to 1010, 100100, 10001000 and watch the green band between the two outer curves.

y
  • (1+y/n)n(1 + y/n)^n
  • eye^y
  • (1−y/n)−n(1 - y/n)^{-n}
  • the squeeze
(1+yn)n=2.44141  <  ey=2.71828  <  (1−yn)−n=3.16049(1 + \tfrac{y}{n})^n = 2.44141 \;<\; e^{y} = 2.71828 \;<\; (1 - \tfrac{y}{n})^{-n} = 3.16049. Gap 0.71910.7191, less than the bound y2eyn−y2=0.9061\dfrac{y^2e^y}{n - y^2} = 0.9061.
What to notice

The lower curve (1+y/n)n(1 + y/n)^n always stays below eye^y, the upper curve (1−y/n)−n(1 - y/n)^{-n} always above, as long as ∣y∣<n|y| < n. The band narrows roughly in proportion to 1/n1/n: multiplying nn by 1010 divides the gap at y=1y = 1 by about 1010. Convergence is slowest where yy is large: there y2eyy^2 e^y is large.

Explore The same convergence as a sequence

Try this. With x=1x = 1, find the n0n_0 needed for δ=0.01\delta = 0.01. Then set x=−2x = -2 or x=3x = 3.

ϕ(n)=(1+xn)n\phi(n) = \left(1+\frac xn\right)^n , candidate limit l=2.718l = 2.718
nφ(n)
  • band l ± δ
  • l
  • n₀ (drag)
  • n ≥ n₀, good
  • n ≥ n₀, bad
134 terms with n ≥ n₀ fall outside the band: n = 1, 2, 3, 4, 5, ….
What to notice

For x=1x = 1 the terms (1+1/n)n(1 + 1/n)^n creep up to e=2.718…e = 2.718\ldots from below: you need n0n_0 around 135135 for δ=0.01\delta = 0.01, because the error is about e/(2n)e/(2n). For larger xx the error is larger, roughly x2ex/(2n)x^2e^x/(2n).

Precise statements

Theorem The exponential limit Hardy: §215 (1), (2)

For every real xx,

lim⁡n→∞(1+xn)n=lim⁡n→∞(1−xn)−n=ex,\lim_{n\to\infty}\left(1 + \frac{x}{n}\right)^n = \lim_{n\to\infty}\left(1 - \frac{x}{n}\right)^{-n} = e^{x},

and more generally (1+xξ)ξ→ex\left(1 + \dfrac{x}{\xi}\right)^{\xi} \to e^x as the real variable ξ→∞\xi \to \infty or ξ→−∞\xi \to -\infty. In particular (1+1/n)n→e(1 + 1/n)^n \to e: the number of §73 is the number defined by log⁡e=1\log e = 1.

Theorem The squeeze, with an error bound Hardy: §215 (4)

If y≠0y \ne 0 and n>∣y∣n > |y|, then

(1+yn)n<ey<(1−yn)−n,\left(1 + \frac{y}{n}\right)^{n} < e^{y} < \left(1 - \frac{y}{n}\right)^{-n},

and if also n>y2n > y^2, the difference between the outer terms is less than y2eyn−y2\dfrac{y^2e^{y}}{n - y^2}.

Proofs

Proof First proof: through log⁡(1+h)/h→1\log(1+h)/h \to 1
Strategy

Take logarithms, use the known limit log⁡(1+h)/h→1\log(1+h)/h \to 1, and return through the continuity of the exponential.

The key limit

log⁡(1+h)h→1\dfrac{\log(1 + h)}{h} \to 1 as h→0h \to 0 (first lesson, Examples LXXXIII, 3). For x=0x = 0 the theorem is trivial, so let x≠0x \ne 0.

Put h=x/ξh = x/\xi

For ∣ξ∣>∣x∣|\xi| > |x| we have 1+x/ξ>01 + x/\xi > 0, and

ξlog⁡(1+xξ)=x log⁡(1+h)h.\xi\log\left(1 + \frac{x}{\xi}\right) = x\,\frac{\log(1 + h)}{h}.

As ξ→±∞\xi \to \pm\infty, h→0h \to 0, so this tends to xx.

Exponentiate

(1+x/ξ)ξ=eξlog⁡(1+x/ξ)(1 + x/\xi)^{\xi} = e^{\xi\log(1 + x/\xi)} by the definition of a general power. The exponent tends to xx and eye^y is continuous, so the power tends to exe^x.

Special cases

Letting ξ\xi run through the positive integers gives (1+x/n)n→ex(1 + x/n)^n \to e^x; letting it run through the negative integers ξ=−n\xi = -n gives (1−x/n)−n→ex(1 - x/n)^{-n} \to e^x.

Proof Second proof: the squeeze
Strategy

Use 1+u<eu1 + u < e^u for u≠0u \ne 0, which is the inequality log⁡(1+u)<u\log(1 + u) < u in exponential form, once with u=y/nu = y/n and once with u=−y/nu = -y/n. Then estimate the ratio of the outer terms with Bernoulli’s inequality.

1+u<eu1 + u < e^u for every u≠0u \ne 0

For u>−1u > -1, log⁡(1+u)<u\log(1 + u) < u (first lesson); apply the increasing function exp⁡\exp. For u≤−1u \le -1 the left side is ≤0<eu\le 0 < e^u.

Lower bound

With u=y/nu = y/n: 1+y/n<ey/n1 + y/n < e^{y/n}. Since n>∣y∣n > |y|, the left side is positive, and raising both sides to the nn-th power keeps the order: (1+y/n)n<ey(1 + y/n)^n < e^{y}.

Upper bound

With u=−y/nu = -y/n: 0<1−y/n<e−y/n0 < 1 - y/n < e^{-y/n}, so (1−y/n)n<e−y(1 - y/n)^n < e^{-y}. Taking reciprocals reverses the inequality: ey<(1−y/n)−ne^{y} < (1 - y/n)^{-n}.

The ratio of the outer terms

(1−y/n)−n(1+y/n)n=(1−y2n2)−n.\frac{(1 - y/n)^{-n}}{(1 + y/n)^{n}} = \left(1 - \frac{y^2}{n^2}\right)^{-n}.

Bernoulli’s inequality (1−η)n>1−nη(1 - \eta)^n > 1 - n\eta (§74), with η=y2/n2\eta = y^2/n^2 and nη=y2/n<1n\eta = y^2/n < 1, gives (1−y2n2)n>1−y2n\left(1 - \frac{y^2}{n^2}\right)^{n} > 1 - \frac{y^2}{n}, so the ratio is less than nn−y2\dfrac{n}{n - y^2}.

The gap

(1−yn)−n−(1+yn)n=(1+yn)n{(1−y2n2)−n−1}<ey(nn−y2−1)=y2eyn−y2.\left(1 - \tfrac{y}{n}\right)^{-n} - \left(1 + \tfrac{y}{n}\right)^{n} = \left(1 + \tfrac{y}{n}\right)^{n}\left\{\left(1 - \tfrac{y^2}{n^2}\right)^{-n} - 1\right\} < e^{y}\left(\frac{n}{n - y^2} - 1\right) = \frac{y^2e^{y}}{n - y^2}.

This tends to 00 as n→∞n \to \infty, and eye^y lies between the two outer terms, so both tend to eye^y.

Hardy reaches the squeeze through the integral inequalities n(1−x−1/n)<log⁡x<n(x1/n−1)n(1 - x^{-1/n}) < \log x < n(x^{1/n} - 1), obtained by comparing 1/t1/t with t−1±1/nt^{-1 \pm 1/n}; with x=eyx = e^y they are exactly the two bounds above. That route is the subject of the next lesson. He proves the case y>0y > 0 and leaves y<0y < 0 to the reader; the argument above covers both at once, because it only uses ∣y∣<n|y| < n.

Worked examples and exercises

Exercise ●○○ Hardy: Examples LXXXVII, 1

Use the squeeze with y=1y = 1 and n=6n = 6 to bound ee. How large must nn be for the upper bound to fall below 2.92.9?

Hint
The bounds are (7/6)6(7/6)^6 and (5/6)−6=(6/5)6(5/6)^{-6} = (6/5)^6.

Solution

(7/6)6=117 649/46 656=2.5216…(7/6)^6 = 117\,649/46\,656 = 2.5216\ldots and (6/5)6=1.26=2.9860(6/5)^6 = 1.2^6 = 2.9860, so 2.52<e<2.992.52 < e < 2.99. The book prints the upper bound as 2.92.9; with n=6n = 6 the method gives only 2.9862.986. The upper bound (nn−1)n\left(\frac{n}{n-1}\right)^n falls below 2.92.9 first at n=9n = 9: (9/8)9=2.8865(9/8)^9 = 2.8865, while (8/7)8=2.9103(8/7)^8 = 2.9103. Either way the width of the bracket shrinks only like 1/n1/n: the error bound with y=1y = 1 is e/(n−1)e/(n - 1).

Exercise ●●○ Hardy: Examples LXXXVII, 2

Show that if nξn→ln\xi_n \to l as n→∞n \to \infty, then (1+ξn)n→el(1 + \xi_n)^n \to e^{l}.

Hint
Write nlog⁡(1+ξn)=nξn⋅log⁡(1+ξn)ξnn\log(1 + \xi_n) = n\xi_n \cdot \dfrac{\log(1 + \xi_n)}{\xi_n}.

Solution

Since nξn→ln\xi_n \to l, ξn=(nξn)/n→0\xi_n = (n\xi_n)/n \to 0; in particular 1+ξn>01 + \xi_n > 0 for large nn. If ξn≠0\xi_n \ne 0,

nlog⁡(1+ξn)=nξn⋅log⁡(1+ξn)ξn,n\log(1 + \xi_n) = n\xi_n\cdot\frac{\log(1 + \xi_n)}{\xi_n},

and the second factor tends to 11 because ξn→0\xi_n \to 0. When ξn=0\xi_n = 0 the left side is 0=nξn0 = n\xi_n. In all cases nlog⁡(1+ξn)→ln\log(1 + \xi_n) \to l, and by continuity of the exponential (1+ξn)n=enlog⁡(1+ξn)→el(1 + \xi_n)^n = e^{n\log(1 + \xi_n)} \to e^{l}. The theorem is the case ξn=x/n\xi_n = x/n.

Exercise ●●○ Hardy: Examples LXXXVII, 3

Show that if nξn→∞n\xi_n \to \infty then (1+ξn)n→∞(1 + \xi_n)^n \to \infty; and if 1+ξn>01 + \xi_n > 0 and nξn→−∞n\xi_n \to -\infty then (1+ξn)n→0(1 + \xi_n)^n \to 0. Deduce that (1+1/n)n2→∞(1 + 1/n)^{n^2} \to \infty and (1−1/n)n→0(1 - 1/\sqrt n)^{n} \to 0.

Hint
Use h1+h<log⁡(1+h)<h\dfrac{h}{1+h} < \log(1 + h) < h. For the first part, treat ξn≤1\xi_n \le 1 and ξn>1\xi_n > 1 separately.

Solution

First part. For large nn, ξn>0\xi_n > 0. If 0<ξn≤10 < \xi_n \le 1, then log⁡(1+ξn)>ξn/(1+ξn)≥ξn/2\log(1 + \xi_n) > \xi_n/(1 + \xi_n) \ge \xi_n/2, so nlog⁡(1+ξn)>nξn/2n\log(1 + \xi_n) > n\xi_n/2. If ξn>1\xi_n > 1, then nlog⁡(1+ξn)>nlog⁡2n\log(1 + \xi_n) > n\log 2. In either case nlog⁡(1+ξn)>min⁡(nξn/2, nlog⁡2)→∞n\log(1 + \xi_n) > \min(n\xi_n/2,\ n\log 2) \to \infty, so (1+ξn)n→∞(1 + \xi_n)^n \to \infty.

Second part. log⁡(1+ξn)<ξn\log(1 + \xi_n) < \xi_n (or =0= 0 when ξn=0\xi_n = 0), so nlog⁡(1+ξn)≤nξn→−∞n\log(1 + \xi_n) \le n\xi_n \to -\infty, and (1+ξn)n=enlog⁡(1+ξn)→0(1 + \xi_n)^n = e^{n\log(1 + \xi_n)} \to 0.

Examples. (1+1/n)n2=(1+ξm)m(1 + 1/n)^{n^2} = (1 + \xi_m)^{m} with m=n2m = n^2 and ξm=1/n=1/m\xi_m = 1/n = 1/\sqrt m, so mξm=m→∞m\xi_m = \sqrt m \to \infty. Directly: it equals {(1+1/n)n}n≥2n\{(1 + 1/n)^n\}^n \ge 2^n. And (1−1/n)n(1 - 1/\sqrt n)^n has ξn=−1/n\xi_n = -1/\sqrt n, nξn=−n→−∞n\xi_n = -\sqrt n \to -\infty, so it tends to 00.

Exercise ●○○ Hardy: Examples LXXXVII, 4

Deduce from the inequality (1+y/n)n<ey(1 + y/n)^n < e^y that eye^y tends to infinity faster than any power of yy.

Hint
Fix an integer nn larger than the power α\alpha, and drop the 11.

Solution

Let α\alpha be given and fix an integer n>αn > \alpha. For y>0y > 0, ey>(1+y/n)n>(y/n)ne^y > (1 + y/n)^n > (y/n)^n. So ey/yα>yn−α/nn→∞e^y/y^\alpha > y^{n - \alpha}/n^n \to \infty as y→∞y \to \infty, because n−α>0n - \alpha > 0.

Exercise Compound interest ●○○ Hardy: §215

A sum of money earns interest at 100%100\% per year, compounded nn times a year. Show that, however often the interest is compounded, the sum is multiplied by less than ee in a year, and that monthly compounding already gives more than 2.62.6 times.

Hint
Compounding nn times multiplies the sum by (1+1/n)n(1 + 1/n)^n.

Solution

By the squeeze with y=1y = 1, (1+1/n)n<e(1 + 1/n)^n < e for every nn. Monthly: (13/12)12=2.613…(13/12)^{12} = 2.613\ldots. Daily: (1+1/365)365=2.7146…(1 + 1/365)^{365} = 2.7146\ldots, already within 0.0040.004 of e=2.71828…e = 2.71828\ldots, in line with the error estimate e/(2n)e/(2n).

Check yourself

Check yourself True or false?

Since 1+1/n→11 + 1/n \to 1 and 1n=11^n = 1, we have (1+1/n)n→1(1 + 1/n)^n \to 1.

Why

False. A limit of a power cannot be taken base first when the exponent also changes: the base is never exactly 11, and the small excess 1/n1/n is raised to an ever larger power. The logarithm nlog⁡(1+1/n)→1n\log(1 + 1/n) \to 1, not 00, so the limit is ee. Forms like "1∞1^\infty" are indeterminate.

Check yourself Enter a number

Compute (1+2/100)100(1 + 2/100)^{100} to four decimal places. (For comparison, e2=7.3891e^2 = 7.3891.)

Why

1.02100=7.2446…1.02^{100} = 7.2446\ldots, below e2e^2 as the squeeze predicts. The gap 0.14440.1444 is below the bound y2ey/(n−y2)=4e2/96≈0.31y^2e^y/(n - y^2) = 4e^2/96 \approx 0.31 (that bound is for the whole bracket, so it is generous).

Check yourself Choose one

For a fixed y>0y > 0, how do the sequences (1+y/n)n(1 + y/n)^n and (1−y/n)−n(1 - y/n)^{-n} approach eye^y?

Why

The squeeze puts (1+y/n)n(1 + y/n)^n below eye^y and (1−y/n)−n(1 - y/n)^{-n} above it for every n>yn > y. (In fact the first increases with nn and the second decreases, as for y=1y = 1 in §73.) This holds for negative yy too, as long as n>∣y∣n > |y|.

Check yourself Choose one

Roughly how does the error e−(1+1/n)ne - (1 + 1/n)^n behave for large nn?

Why

nlog⁡(1+1/n)=1−12n+O(1/n2)n\log(1 + 1/n) = 1 - \frac{1}{2n} + O(1/n^2), so (1+1/n)n=e⋅e−1/(2n)+…≈e(1−12n)(1 + 1/n)^n = e\cdot e^{-1/(2n) + \dots} \approx e(1 - \frac{1}{2n}). The error is about e/(2n)e/(2n): halving it costs doubling nn. The squeeze’s bound e/(n−1)e/(n - 1) is about twice the truth.

Hardy’s language

Summary

Summary
  • (1+x/n)n→ex(1 + x/n)^n \to e^x and (1−x/n)−n→ex(1 - x/n)^{-n} \to e^x for every real xx; with x=1x = 1 this identifies the ee of §73 with the ee of log⁡e=1\log e = 1.
  • Proof 1: nlog⁡(1+x/n)=x⋅log⁡(1+h)h→xn\log(1 + x/n) = x\cdot\frac{\log(1+h)}{h} \to x, then continuity of exp⁡\exp.
  • Proof 2: 1+u<eu1 + u < e^u gives (1+y/n)n<ey<(1−y/n)−n(1 + y/n)^n < e^y < (1 - y/n)^{-n} for ∣y∣<n|y| < n, with gap below y2ey/(n−y2)y^2e^y/(n - y^2).
  • More generally, (1+ξn)n→el(1 + \xi_n)^n \to e^l whenever nξn→ln\xi_n \to l.