Integrals of unbounded functions, and a warning about substitution
§§187–18940 min
Integrals whose integrand becomes infinite at a point, defined as limits; substitutions that turn one kind of infinite integral into another or into an ordinary integral; and why a substitution must be monotone.
The definite integral of Chapter VII needs two things: a bounded interval, and a continuous integrand. The last lesson dropped the first condition. §§187–188 drop the second: what does ∫01dx/x mean, when 1/x is unbounded near 0? The answer is again a limit of ordinary integrals, and again the yardstick is a power, now (x−a)−s with the dividing line at s=1, the other way round. These infinite integrals of the second kind appear constantly: the Beta and Gamma integrals, ∫01logxdx, and every substitution that maps an infinite range onto a finite one. §189 ends with a warning about substitutions that is worth learning once and for all.
The idea
Near x=0 the graph of x−s shoots up to infinity. The region under it between 0 and 1 is infinitely tall, just as the region in the last lesson was infinitely long. Cut it off at x=ε: the area from ε to 1 is an ordinary integral. Now let ε→0. If these areas approach a limit, that limit is what we mean by ∫01x−sdx.
ExploreAn infinitely tall region
Try this.With s=0.5, slide ε towards 10−4 and watch the lower plot. Then raise s through 1.
∫ε1x−0.5dx: the integrand is unbounded near 0
xφ(x)
ε (log scale)∫ from ε to 1
the integrand (cut off at height 8); shaded: from ε to 1
∫ε1ϕ as ε shrinks (to the left)
its limit
∫0.11ϕ=1.3675. For s<1 this tends to 1/(1−s)=2 as ε→0: the region is infinitely tall but has finite area, and the integral of the second kind converges.
What to notice
For s=0.5 the area approaches 2, although the shaded region becomes ever taller and thinner near 0. For s≥1 it grows without bound. Compare the last lesson: there ∫1∞x−s converged for s>1; here ∫01x−s converges for s<1. The substitution x=1/y turns one region into the other.
Why does the dividing line flip? Under x=1/y, the interval ε≤x≤1 becomes 1≤y≤1/ε, and ∫ε1x−sdx=∫11/εys−2dy. The new integrand is y−(2−s), which has finite area out to infinity exactly when 2−s>1, that is s<1.
Precise statements
DefinitionInfinite integral of the second kindHardy: §187
Let ϕ be continuous on a<x≤A, and ϕ(x)→∞ (or −∞) as x→a from the right. If
ε→+0lim∫a+εAϕ(x)dx=l
exists, we say that ∫aAϕ(x)dxconverges and has the value l. Similarly at the upper end, with ∫aA−ε. If ϕ becomes infinite at an interior point ξ, we define ∫aAϕ=∫aξϕ+∫ξAϕ when both converge; and in the same way for finitely many such points.
The same definition serves for other discontinuities: for a jump, or for an oscillating function like sin(1/x). If both a finite end and an infinite range are involved, as in ∫0∞xs−1/(1+x)dx, split the range so that each piece has only one difficulty, and require each piece to converge. Most of Hardy’s remarks on infinite integrals of the first kind (last lesson) carry over word for word.
TheoremThe standard integral of the second kindHardy: §187; Examples LXXVI, 2
∫aA(x−a)sdx converges if s<1, with value 1−s(A−a)1−s, and diverges if s≥1.
TheoremComparison near a singular pointHardy: Examples LXXVI, 1 and 3
Let ϕ be continuous on a<x≤A.
If ϕ≥0, then ∫aAϕ converges if and only if ∫a+εAϕ<K for some constant K and every ε>0; otherwise it diverges to ∞.
If 0≤ϕ(x)<K(x−a)−s with s<1, the integral converges.
If ϕ(x)>K(x−a)−s with K>0 and s≥1, the integral diverges.
Substitution can move a difficulty from one place to another, or remove it altogether (§188).
TheoremSubstitution onto a finite rangeHardy: §188
Suppose ∫a∞ϕ(x)dx converges, and x=f(t) maps b≤t<c increasingly onto a≤x<∞, with f′ continuous. Then
∫a∞ϕ(x)dx=∫bcϕ{f(t)}f′(t)dt,
where the right-hand side is an infinite integral of the second kind if ϕ{f(t)}f′(t) is unbounded near c, and an ordinary integral if it extends continuously to t=c.
For example, with x=t/(1−t),
∫0∞(1+x)mdx=∫01(1−t)m−2dt(m>1),
an integral of the second kind when 1<m<2 and an ordinary integral when m≥2. Both sides equal 1/(m−1).
A substitution must run one way
The substitution rule ∫ϕ(x)dx=∫ϕ{f(t)}f′(t)dt is proved for a function x=f(t) defined on the whole range. If instead we start from y=g(x) and solve for x, there may be several solutions, and we must choose the right one on each part of the range. Hardy’s example is J=∫17(x2−6x+13)dx=48 with y=x2−6x+13.
ExploreWhich area did the substitution compute?
Try this.Choose each of the three ways of computing J in the variable y, and look at which area is shaded.
J=∫17(x2−6x+13)dx with the substitution y=x2−6x+13
xy
y=x2−6x+13
the levels y=8 (at x=1) and y=20 (at x=7)
the minimum y=4 at x=3, where dx/dy changes sign
Compute J in the variable y by
As x runs from 1 to 3, y falls from 8 to 4 on the branch x=3−y−4, with dx/dy<0; as x runs from 3 to 7, y rises from 4 to 20 on the other branch. So J=∫84y(−2y−41)dy+∫4202y−4ydy=332+3112=48, the correct value.
What to notice
As x runs from 1 to 7, y does not simply run from 8 to 20: it falls to 4 at x=3 and then rises. Using one branch for the whole range computes the area over a wrong interval (5≤x≤7, or even −1≤x≤1), which is why both signs give ±80/3. Splitting at the turning point, with the right branch on each side, gives 48.
RuleSplit a substitution where it turnsHardy: §189
When transforming ∫abϕ(x)dx by a relation y=g(x), split [a,b] into intervals on which g is monotone. On each, x is a single-valued function of y with dx/dy of one sign, and the substitution rule applies to that piece.
Proofs
ProofThe standard integral of the second kind
Strategy
Compute the ordinary integral from a+ε to A and let ε→0. Everything depends on whether ε1−s tends to 0.
The integral from a+ε
For s=1, an antiderivative of (x−a)−s is (x−a)1−s/(1−s), so
∫a+εA(x−a)sdx=1−s(A−a)1−s−ε1−s.
s<1: convergence
Then 1−s>0 and ε1−s→0, so the integral tends to (A−a)1−s/(1−s).
s>1: divergence
Then ε1−s=1/εs−1→∞, and since 1−s<0 the whole expression tends to +∞.
s=1: divergence
The substitution x−a=1/y turns ∫a+εAdx/(x−a) into ∫1/(A−a)1/εdy/y, which tends to ∞ as ε→0 by the harmonic case of §181.
∎
ProofComparison near a singular point
Strategy
Exactly as for the first kind: for positive ϕ the integral from a+ε increases as ε decreases, so it either has a limit or tends to ∞; then compare with (x−a)−s.
Monotonicity in ε
If ε′<ε then ∫a+ε′Aϕ−∫a+εAϕ=∫a+ε′a+εϕ≥0. So I(ε)=∫a+εAϕ increases as ε→0, and tends to a limit if it is bounded, to ∞ if not. This is part 1.
Comparison
If 0≤ϕ<K(x−a)−s with s<1, then I(ε)≤K∫a+εA(x−a)−sdx≤K(A−a)1−s/(1−s), a bound independent of ε. If ϕ>K(x−a)−s with s≥1, then I(ε)≥K∫a+εA(x−a)−sdx→∞.
∎
ProofSubstitution onto a finite range
Strategy
Apply the ordinary rule on [b,τ] with τ<c, and let τ→c.
Finite pieces
For b≤τ<c, with ξ=f(τ), §166 gives ∫aξϕ(x)dx=∫bτϕ{f(t)}f′(t)dt.
Let τ→c
As τ→c, ξ=f(τ)→∞, so the left side tends to ∫a∞ϕ. Hence the right side tends to the same limit: by definition, the integral of the second kind ∫bc converges and has that value.
When the new integrand is continuous at c
If ϕ{f(t)}f′(t) extends to a continuous function on [b,c], then ∫bτ is a continuous function of τ (§165), so its limit as τ→c is the ordinary integral ∫bc. The substitution has turned an infinite integral into a finite one.
∎
ProofSplitting a substitution at its turning point
Strategy
On an interval where g is strictly monotone, it has a continuous inverse, and the ordinary substitution rule applies with that inverse. Then add the pieces.
The pieces of Hardy’s example
y=x2−6x+13=(x−3)2+4 decreases on [1,3] (from 8 to 4) and increases on [3,7] (from 4 to 20). The inverses are x=3−y−4 and x=3+y−4, with dx/dy=∓21(y−4)−1/2.
Look at each end separately and find the power of (x−a) or (A−x) that the integrand behaves like there.
Solution
First. Near a the integrand is about (A−a)−1/2(x−a)−1/2; near A it is about (A−a)−1/2(A−x)−1/2. Both exponents are 21<1: converges. (Its value is π, Examples LXXVI, 22.)
Second. Near a it behaves like (x−a)−1/3, fine; near A like (A−x)−1: diverges.
Third. The integrand is (A−x)−4/3; the exponent exceeds 1: diverges.
Exercise●●○Hardy: Examples LXXVI, 7
Show that ∫0π(sinx)sdx converges if and only if s<1. (In particular ∫0πdx/sinx converges: Examples LXXVI, 6.)
Hint
sinx/x→1 as x→0, and the integrand is symmetric about 21π.
Solution
For s≤0 the integrand is continuous and there is nothing to prove, so let s>0. The integrand is infinite at both ends, so split at 21π; by symmetry (sin(π−x)=sinx) it is enough to study ∫0π/2. On (0,21π] we have π2x≤sinx≤x. So x−s≤(sinx)−s≤(2π)sx−s, and by the comparison theorem the integral converges exactly when ∫0x−sdx does, that is, when s<1.
Exercise●●○Hardy: Examples LXXVI, 11
Show that ∫0∞1+xxs−1dx converges if and only if 0<s<1.
Hint
Split at x=1. Near 0 the integrand behaves like xs−1; near ∞ like xs−2.
Solution
By definition it is ∫01+∫1∞, and both must converge. On (0,1], 21xs−1≤1+xxs−1≤xs−1, so ∫01 converges iff ∫01x−(1−s)dx does, that is iff 1−s<1, i.e. s>0. On [1,∞), x2−s⋅1+xxs−1=1+xx→1, so ∫1∞ converges iff 2−s>1, i.e. s<1. Both together: 0<s<1. (Its value, π/sinsπ, comes much later.)
Exercise●●○Hardy: Examples LXXVI, 20
Show by the substitution x=t/(1−t) that, for l,m>0,
∫0∞(1+x)l+mxl−1dx=∫01tl−1(1−t)m−1dt.Solution
x=t/(1−t) maps 0≤t<1 increasingly onto 0≤x<∞, with 1+x=1/(1−t) and dx=dt/(1−t)2. Then
Both integrals converge: the left one behaves like xl−1 near 0 and like x−m−1 at infinity; the right one like tl−1 near 0 and (1−t)m−1 near 1. Applying the substitution on [ε,τ] and letting ε→0, τ→1 gives the equality. The right-hand side is Euler’s Beta integral B(l,m).
Exercise●●○Hardy: §189, Example
Verify the rule of §189 on ∫01(4x2−x+161)dx with the substitution y=4x2−x+161.
Hint
y=4(x−81)2. Where is the turning point, and what are the two branches?
Solution
Directly: 34−21+161=4843.
Now y=4(x−81)2, which falls from 161 to 0 on [0,81] and rises from 0 to 1649 on [81,1]. The branches are x=81∓21y, with dx/dy=∓4y1. Hence
With ∫0cydy=32c3/2: 61(641+64343)=384344=4843. A single branch over y from 161 to 1649 would have given 61⋅64342=6457, which is wrong.
Check yourself
Check yourselfChoose one
What is ∫−11x2dx?
Why
The integrand is unbounded at the interior point 0, so the integral means ∫−10+∫01, and ∫ε1x−2dx=1/ε−1→∞. The answer −2 (negative, for a positive integrand!) comes from applying the fundamental theorem across a point where it does not hold. Symmetry would only matter if both halves converged.
Check yourselfEnter a number
What is ∫01xdx?
Why
∫ε1x−1/2dx=2−2ε→2. This is the case s=21 of the standard integral: (A−a)1−s/(1−s)=1/21=2.
Check yourselfTrue or false?
∫01x−sdx and ∫1∞x−sdx converge for the same values of s.
Why
False: the first converges for s<1, the second for s>1, and neither for s=1. Near 0 a smaller power is milder; near ∞ a larger power decays faster. So ∫0∞x−sdx converges for no s at all.
Check yourselfChoose one
Hardy’s substitution y=x2−6x+13 in ∫17ydx, using x=3+y−4 for y from 8 to 20, gives 80/3 instead of 48. What went wrong?
Why
The relation y=g(x) must be inverted separately on each interval where g is monotone. The unboundedness at y=4 is harmless (it is a convergent integral of the second kind) and appears only in the correct, split computation.
Hardy’s language
Some care has occasionally to be exercised in applying the rule for transformation by substitution.
— G. H. Hardy, §189
Summary
Summary
If ϕ is unbounded near a, ∫aAϕ means limε→+0∫a+εAϕ; with several bad points, split the range and require every piece to converge.
∫aA(x−a)−sdx converges iff s<1, the opposite of the condition at infinity; comparison with it decides most integrals.
A substitution can turn an infinite range into a finite one with an unbounded integrand (or a perfectly ordinary one), and back.
A substitution defined by y=g(x) must be split where g turns, with the correct branch on each piece.