In §98 Hardy measured how fast a function tends to infinity by comparing it with powers: f is “of order α” if f(x)/xα tends to a non-zero limit. The powers xα, with α as small or as large as we please, look like a complete measuring stick. The logarithm shows that they are not: logx tends to infinity, yet it is beaten by x0.0000001. This single fact, proved in §208 in three lines, is used in nearly every estimate in analysis.
The idea
Compare logx with a small power such as x0.1. For moderate x the logarithm is far ahead: at x=1010, logx≈23 while x0.1=10. It looks as if the logarithm wins. But plot both on log–log axes (horizontally log10x, vertically log10 of the value) and the picture changes. A power xα becomes a straight line of slope α, because log10xα=αlog10x. The logarithm becomes log10(logx), a curve that keeps bending over. A straight line rising at any positive slope must overtake a curve that flattens out.
ExploreRace a power of logx against a small power of x
Try this.Start with k=3 and α=0.25 and find the last crossing. Then make α smaller (try 0.05) or k larger, and widen the view until the crossing reappears.
(logx)3againstx0.25
log₁₀ xlog₁₀ (value)
(logx)3
x0.25
last crossing
read-out point
At x=1010: (logx)3/x0.25≈101.587 (the slower function is still ahead). The last crossing is at x≈1019.94; beyond it the ratio only falls, towards 0. The ratio (slower ÷ faster) is largest at x=105.21, i.e. logx=k/α.
What to notice
However you handicap the power, it wins in the end: the crossing moves out to astronomical x (for logx against x0.1 it is at about x=3.4×1015), but it is always there, and after it the ratio falls without limit. The ratio peaks where logx=k/α, then declines for ever.
So the picture suggests two things: the race is always won by the power, and the win may come very late. The theorem below captures the first; the second is a warning about trusting tables of values.
Precise statements
TheoremThe logarithm is slower than every powerHardy: §207
logx→∞ as x→∞, but for every positive α
xαlogx→0(x→∞).
Hardy states this for rational α, since xα has not yet been defined for irrational α; once it has (§214), the same proof works for every real α>0.
CorollaryAny power of the logarithm
For every k>0 and α>0, xα(logx)k→0 as x→∞.
TheoremBehaviour as x→+0Hardy: §209
For every positive α, xαlogx→0 as x→+0. Thus log(1/x)→∞ as x→+0, but more slowly than any positive power of 1/x.
Proofs
The whole difficulty is that log is defined by an integral we cannot evaluate. Hardy’s trick is to compare the integrand 1/t with an integrand we can integrate, tβ−1, which is only a little larger.
Proofx−αlogx→0
Strategy
Bound logx above by a slightly smaller power xβ/β with 0<β<α. Then x−αlogx is at most a constant times xβ−α, and a negative power tends to 0.
For t>1 and β>0: 1/t<tβ−1
Because tβ−1=tβ/t and tβ>1 when t>1.
Integrate from 1 to x (for x>1)
logx=∫1xtdt<∫1xtβ−1dt=βxβ−1<βxβ.
Choose β between 0 and α
For example β=α/2. Then for x>1,
0<x−αlogx<βxβ−α=α2x−α/2.
Squeeze
Since α/2>0, x−α/2→0 as x→∞, so the right-hand side tends to 0, and x−αlogx, trapped between 0 and it, tends to 0 as well.
∎
The bound logx<xβ/β is worth remembering by itself. With β=21 it says logx<2x for all x>1; with β=1, logx<x.
ProofThe corollary: (logx)k/xα→0
Strategy
Take the k-th root to reduce to the theorem.
Rewrite
For x>1, xα(logx)k=(xα/klogx)k.
Apply the theorem with exponent α/k
The bracket is positive and tends to 0, so its k-th power tends to 0 (for rational k by the rules for limits; for any k>0 because u↦uk is continuous at 0 with value 0).
∎
Proofxαlogx→0 as x→+0
Strategy
Turn x→+0 into y→∞ by putting x=1/y, and use log(1/y)=−logy.
Substitute x=1/y
As x→+0, y→+∞. And xαlogx=y−αlog(1/y)=−y−αlogy.
Apply the theorem
y−αlogy→0 as y→∞, so xαlogx→0. Equivalently log(1/x)/(1/x)α→0: the logarithm of 1/x is beaten by every power of 1/x.
∎
Explorexlogx near 0
Try this.Move the trace point towards x=0 from the right and read off xlogx.
y=xlogxHardy: Examples LXXXVIII, 16
xy
x log x
trace point (drag)
At x=0.3: y=−0.36119
What to notice
Although logx→−∞, the factor x→0 wins: the product creeps back up to 0. Its lowest point is at x=1/e≈0.37, where the value is −1/e. The same happens with x0.1logx, only the dip is much deeper and the recovery much later.
Worked examples and exercises
ExerciseAn explicit bound●○○Hardy: §208
Show that logx<2x for every x>0, and that logx<10x0.1 for every x>0.
Hint
Use logx<(xβ−1)/β for x>1. What about 0<x≤1?
Solution
For x>1 the proof above gives logx<(xβ−1)/β<xβ/β. With β=21 this is logx<2x; with β=0.1 it is logx<10x0.1. For 0<x≤1 we have logx≤0, while 2x and 10x0.1 are positive. So both inequalities hold for all x>0.
ExerciseHow late is “eventually”?●●○Hardy: §207
Show that logx>x0.01 at x=10100, although logx/x0.01→0. Estimate where logx finally drops below x0.01 for good.
Hint
Write u=logx. The inequality logx<x0.01 becomes u<e0.01u, or logu<0.01u, that is, u>100logu.
Hint
log10≈2.3026, so log10100≈230.
Solution
At x=10100: logx=100log10≈230.3, while x0.01=101=10. So logx is far ahead.
With u=logx, we have x0.01=e0.01u, and logx<x0.01 means logu<0.01u, i.e. u>100logu. The function u−100logu decreases until u=100 and then increases, so there is a last crossing u1>100 with u1=100logu1. Trying values: at u=600, 100log600≈639.7>600; at u=650, 100log650≈647.7<650. Bisection gives u1≈647.3. So logx<x0.01 for good only once x>e647.3≈10281. The theorem is about the limit, and says nothing about where the limit “starts to show”.
Exercise●○○Hardy: §209
Show that xα{log(1/x)}k→0 as x→+0, for every α>0 and k>0.
Hint
Put x=1/y and use the corollary.
Solution
With x=1/y, y→∞ and xα{log(1/x)}k=(logy)k/yα→0 by the corollary.
ExerciseSequences●○○Hardy: §208
Show that 100n(logn)10→0 as n→∞, and that ∑n≥2n2(logn)10 converges.
Hint
A limit of a function as x→∞ gives the same limit along x=n. For the series, compare with ∑n−3/2.
Solution
The corollary with k=10, α=0.01 gives (logx)10/x0.01→0 as x→∞, hence also along the integers.
For the series: by the corollary with α=21, (logn)10/n1/2→0, so it is less than 1 for n≥n0. Then (logn)10/n2<n1/2/n2=n−3/2 for n≥n0, and ∑n−3/2 converges (§181). By comparison the series converges. Logarithmic factors never change the convergence of ∑n−s when s=1; the case s=1 is the subject of §218.
Check yourself
Check yourselfChoose one
Which of these tends to 0 as x→∞?
Why
Any power of logx over any positive power of x tends to 0, so the second tends to 0 and the first (its reciprocal) to ∞. The third is the constant logx/(0.01logx)=100. The large exponent 100 only delays the moment when x0.01 takes over, to around x=e116700≈1050670 (the ratio peaks where logx=100/0.01=104).
Check yourselfTrue or false?
At x=1010 we have logx≈23>10=x0.1. This shows that logx/x0.1 does not tend to 0.
Why
False. A limit as x→∞ is about all sufficiently large x, and no single value decides it. The ratio rises until logx=10 (that is x=e10≈22000) and then falls for ever; it drops below 1 near x≈3.4×1015 and tends to 0.
Check yourselfEnter a number
The inequality logx<xβ/β with β=21 gives an upper bound for log106. What is it?
Why
x1/2/(1/2)=2106=2000. The true value is log106≈13.8, so the bound is very crude, but crude bounds are all a limit proof needs.
Check yourselfChoose one
As x→+0, xlogx
Why
Putting x=1/y gives −logy/y→0. Any positive power of x beats the logarithm at 0, just as at infinity.
Hardy’s language
Summary
Summary
logx→∞, but logx/xα→0 for every α>0; also (logx)k/xα→0 for every k.
Proof: 1/t<tβ−1 for t>1, so logx<xβ/β; choose 0<β<α.
At 0: xαlogx→0 as x→+0 (put x=1/y).
“Eventually” can be very late: logx stays ahead of x0.01 until about x=10281.