So far ex has been defined indirectly, as the inverse of the area function logx. That definition is good for proving properties but useless for computing: it does not tell you what e0.3 is to six decimals. In §219 Hardy shows that the exponential function is the sum of a power series that converges for every real x, and very fast. This one formula is the working definition of ex in most of mathematics, and in §223 it will become the starting point of a second, independent theory of the exponential function.
The idea
Every derivative of ex is ex again, so every derivative at 0 equals 1. The Taylor polynomial of degree n−1 about 0 is therefore
1+x+2!x2+⋯+(n−1)!xn−1.
The question is whether these polynomials actually converge to ex, and where. For log(1+x) or 1/(1+x2) the answer is “only on an interval”, as you saw in §152. For ex it is “everywhere”, and the reason is that n! eventually beats any power xn.
ExploreTaylor polynomials of ex
Try this.Start at degree 1 and raise the degree one step at a time. Watch where the polynomial leaves the curve, on the right and on the left. Then drag the centre away from 0.
f(x)=ex, centre a=0, degree n=1
xy
f(x)
Taylor polynomial Pn
remainder f − Pn
centre a (drag)
probe x (drag)
P1(x)=1+x At x=2.4: f(x)=11.023, P1(x)=3.4, error ∣f(x)−P1(x)∣=7.62.
What to notice
Each new term pushes the region of good agreement further out, on both sides, and no interval of convergence is ever marked: unlike log(1+x), the series for ex has no barrier. On the left the polynomials alternate above and below the curve, because the terms xn/n! alternate in sign when x<0.
Taylor’s theorem with Lagrange’s remainder (§150) makes this precise. For each n and each x there is a θ between 0 and 1 with
ex=1+x+2!x2+⋯+(n−1)!xn−1+n!xneθx.
We know nothing about θ except that it lies in (0,1). That is enough: eθx is at most e∣x∣ whatever θ is, and xn/n!→0 for every fixed x (Examples XXVII, 12). So the remainder is squeezed to 0.
Precise statements
TheoremThe exponential seriesHardy: §219 (1)–(3)
For every real x,
ex=1+x+2!x2+3!x3+⋯=n=0∑∞n!xn.
In particular e=1+1+2!1+3!1+…, and, since exey=ex+y,
(n=0∑∞n!xn)(n=0∑∞n!yn)=n=0∑∞n!(x+y)n.
Hardy calls the last identity the exponential theorem. Combined with ax=exloga (§214) it gives a series for every power:
CorollarySeries for axHardy: §219 (4)
For a>0 and every real x,
ax=1+xloga+2!(xloga)2+3!(xloga)3+….
Proofs
ProofThe exponential series, by Taylor's theorem
Strategy
Write ex as the Taylor polynomial plus Lagrange’s remainder, bound the remainder by something that does not involve the unknown θ, and show that this bound tends to 0.
Taylor's theorem applies
f(x)=ex has derivatives of every order, all equal to ex (§213), so f(k)(0)=1 for every k. Lagrange’s form of Taylor’s theorem about 0 gives the displayed formula above, with some θ=θ(n,x) in (0,1).
Bound the remainder independently of θ
Since et increases, eθx≤e∣x∣: if x≥0 then θx≤x, and if x<0 then θx<0, so eθx<1. Hence
ex−k=0∑n−1k!xk≤n!∣x∣ne∣x∣.
∣x∣n/n!→0
Fix x and choose an integer N≥2∣x∣. For n>N, each extra factor ∣x∣/n is at most 21, so ∣x∣n/n!≤(∣x∣N/N!)(21)n−N, which tends to 0.
Conclude
The bound in step 2 is a fixed number e∣x∣ times a sequence tending to 0, so the partial sums tend to ex. Taking x=1 gives the series for e, and the exponential theorem is just exey=ex+y rewritten with the series.
∎
Hardy then gives a second proof that avoids Taylor’s theorem altogether. It uses only the limit (1+x/n)n→ex of §215 and the binomial theorem. Write En(x)=1+x+⋯+xn/n! for the partial sum ending with xn/n!.
ExploreThe partial sums caught in a vice
Try this.With x=2, step n from 1 upwards. The filled ochre dots are (1+x/n)n, the hollow ones (1−x/n)−n. Then raise x to 4 or 5 and watch the upper dots appear only once n>x.
nvalue
partial sum En(x)
lower (1+x/n)n (filled), upper (1−x/n)−n (hollow, only for n > x)
ex
At n=8: lower (1+x/n)n=5.960464, sum En(x)=7.387302, upper (1−x/n)−n=9.988721. The target is ex=7.389056; the sum is off by 0.00175, the lower bound by 1.43.
What to notice
The blue partial sums always sit between the two ochre sequences. Both ochre sequences tend to ex (dashed), so the blue ones must too. Notice also how much faster the blue dots arrive: the squeeze proves convergence, but the series itself converges far more quickly than either bound.
ProofThe exponential series, by squeezing (Hardy's second proof)
Strategy
For x>0, compare En(x) term by term with the binomial expansions of (1+x/n)n and (1−x/n)−n. Both bounds tend to ex. Negative x is then handled by the functional equation.
Lower bound: (1+x/n)n≤En(x)
By the binomial theorem,
(1+nx)n=k=0∑nnkn(n−1)⋯(n−k+1)k!xk.
Each fraction n(n−1)⋯(n−k+1)/nk is a product of numbers at most 1, so each term is at most xk/k!, the matching term of En(x). (For n>1 the term with k=2 is strictly smaller, so the inequality is strict.)
Upper bound: En(x)<(1−x/n)−n when n>x
When 0<x/n<1, the binomial series for a negative integral exponent gives
(1−nx)−n=k=0∑∞nkn(n+1)⋯(n+k−1)k!xk.
Now every fraction is at least 1, so every term is at least xk/k!; and the series has infinitely many further positive terms beyond k=n.
Squeeze
By §215, (1+x/n)n→ex and (1−x/n)−n=1/(1−x/n)n→1/e−x=ex. A sequence trapped between two sequences with the same limit has that limit too, so En(x)→ex for every x>0 (and trivially for x=0).
Negative x
Write F(x)=∑xn/n!, which converges absolutely for every x. Multiplying two such series (Examples LXXXI, 7, at §202) gives F(x)F(y)=F(x+y); this is worked out in §223. So for x>0, F(−x)=F(0)/F(x)=1/ex=e−x.
∎
Worked examples and exercises
ExerciseThe largest term●○○Hardy: Examples XCI, 2
Let x>0. Show that the largest term of ∑xn/n! is x[x]/[x]!, and that when x is an integer the term before it, xx−1/(x−1)!, is equally large.
Hint
Compare consecutive terms: the ratio of xn/n! to xn−1/(n−1)! is x/n.
Solution
The ratio un/un−1 of consecutive terms un=xn/n! is x/n. So the terms increase while n<x and decrease once n>x. The last increase happens at n=[x] (when x is not an integer, [x]<x<[x]+1), so u[x] is the largest term. In Hardy’s counting, which starts with u0=1 as the first term, this is the ([x]+1)th term. If x is an integer, ux/ux−1=x/x=1, so the two are equal. You can see the hump in the widget below: for x=10 the terms rise to about 2756 at n=9,10 before the factorial takes over.
ExploreBig x: the terms rise before they fall
Try this.Show the terms with the log scale on, at x=10. Then switch to the error and read off how many terms give six correct decimals.
∑n=0∞n!xn
nuₙ (log scale)
uₙ for n!xn
Plot
After 30 terms s29=22026.46027; the sum is 22026.466, so ∣s−s29∣ = 0.00553. The series converges to 22026.466.
What to notice
The terms climb to a peak at n=9 and 10 and then fall faster than any geometric sequence (the log plot bends downwards). The error stays enormous for the first ten terms; convergence is guaranteed, but it starts late when x is large.
Exercise●○○Hardy: Examples XCI, 3
Show that n!>(n/e)n for every positive integer n.
Hint
nn/n! is one of the terms of the series for en.
Solution
All terms of en=∑knk/k! are positive, so the series is larger than any one of its terms (there are others): en>nn/n!. Rearranged, n!>nn/en=(n/e)n. This crude bound is already the main factor in Stirling’s formula n!∼2πn(n/e)n.
Exercise●○○Hardy: Examples XCI, 5
Use the series to show that ex tends to infinity faster than any power of x: x−kex→∞ as x→∞, for every k>0.
Hint
Pick an integer n>k and keep just one term of the series.
Solution
Let n be an integer with n>k. For x>0 every term is positive, so ex>xn/n!, and x−kex>xn−k/n!. Since n−k>0, the right-hand side tends to ∞ as x→∞. Compare the proof of the same fact in §215: the series gives it in one line.
Exercisee is irrational●●○Hardy: Examples XCI, 6
Show that e is not a rational number.
Hint
Suppose e=p/q with integers p,q≥1. Multiply e=∑1/n! by q! and split the sum at n=q.
Hint
The part with n≤q is an integer. Bound the rest by a geometric series.
Solution
Suppose e=p/q. Then
q!e−q!(1+1+2!1+⋯+q!1)=q+11+(q+1)(q+2)1+….
The left side is an integer: q!e=(q−1)!p, and each q!/k! with k≤q is an integer. The right side is positive and less than q+11+(q+1)21+⋯=q1≤1. A positive integer less than 1 is impossible, so e is irrational.
Exercise●●○Hardy: Examples XCI, 9
Show that ∑n≥1n/n!=e, ∑n2/n!=2e, ∑n3/n!=5e, and that ∑nk/n! is a positive integer multiple of e for every positive integer k.
Hint
n/n!=1/(n−1)!, and more generally n(n−1)⋯(n−j+1)/n!=1/(n−j)! for n≥j.
Hint
Write n2=n(n−1)+n and n3=n(n−1)(n−2)+3n(n−1)+n.
Solution
For each j≥1, ∑n≥0n!n(n−1)⋯(n−j+1)=∑n≥j(n−j)!1=e, since the terms with n<j vanish. Hence ∑n/n!=e; ∑n2/n!=e+e=2e; and ∑n3/n!=e+3e+e=5e (check: n(n−1)(n−2)+3n(n−1)+n=n3−3n2+2n+3n2−3n+n=n3).
In general nk can be written as ∑j=1kck,jn(n−1)⋯(n−j+1) with non-negative integer coefficients ck,j and ck,k=1. (Induction: multiply by n=(n−j)+j and use n(n−1)⋯(n−j+1)⋅(n−j)= the next falling product.) Then ∑nk/n!=e∑jck,j, a positive integer multiple of e. Hardy’s Ex. 7 uses the same idea to sum ∑P(n)xn/n! for any polynomial P.
Check yourself
Check yourselfEnter a number
What is the smallest N for which 1+1+2!1+⋯+N!1 differs from e by less than 10−6?
Why
The error is the tail (N+1)!1+(N+2)!1+…, which is a little more than its first term. For N=8 it is about 3.06×10−6 (the first term alone is 1/9!≈2.76×10−6), too big; for N=9 it is about 3.03×10−7. So N=9: ten terms give e to six decimals.
Check yourselfTrue or false?
For x=50 the terms 50n/n! increase for the first fifty steps, so the series for e50 diverges.
Why
False. Convergence depends only on the tail. The terms increase while n<50, but from then on each is multiplied by 50/n, which falls below 21 once n>100; the tail is dominated by a geometric series. A series can start by growing and still converge; it simply converges late.
Check yourselfChoose one
In the Taylor proof, the remainder is n!xneθx with an unknown θ∈(0,1) that depends on n. Why does that not spoil the argument?
Why
We never need to know θ, only a bound for eθx valid for all possible θ. Nothing is claimed about how θ behaves as n changes (in fact θ(n,x) does tend to 0, but proving that is harder and unnecessary). The remainder is negative for some x<0 and odd n, so the last option is false.
Check yourselfChoose one
Hardy’s squeeze proof works directly only for x>0. What breaks for x<0?
Why
The inequality “each term of (1+x/n)n is at most the matching term of En(x)” turns into a statement about sums only when the terms are positive. For negative x the limit in §215 is still true and the series still converges; the proof is simply finished by the functional equation F(x)F(−x)=F(0)=1.
Hardy’s language
The reader will observe that the exponential series has the property of reproducing itself when every term is differentiated, and that no other series of powers of x would possess this property.
— G. H. Hardy, §219
Summary
Summary
ex=∑n≥0xn/n! for every real x; in particular e=∑1/n!, and ax=∑(xloga)n/n!.
Taylor proof: the Lagrange remainder is at most e∣x∣∣x∣n/n!, and ∣x∣n/n!→0 because the factorial eventually wins.
Squeeze proof: (1+x/n)n<En(x)<(1−x/n)−n for x>0, n>x; both bounds tend to ex.
For large x the terms first rise, peaking at n=[x], so many terms are needed; for small x convergence is extremely fast.
One term of the series already gives useful inequalities: ex>xn/n!, n!>(n/e)n; and the series shows that e is irrational.