Unit 7, chapter 29

Make a family, then differentiate the family

How differentiation under the integral sign evaluates the integral of (x minus 1) over ln x from 0 to 1 as ln 2, using a parameter family, the Leibniz integral rule, an integrable bound and a known value for the constant.

  • Chapter 29
  • The Gaussian integral, parameter differentiation and contour integration

The integrand (x−1)/ln⁡x(x - 1)/\ln x is positive on (0,1)(0, 1), because its numerator and denominator are both negative there, and its values stay below 11. So the integral ∫01x−1ln⁡x dx\int_0^1 \frac{x - 1}{\ln x}\,dx is a number between 00 and 11, yet none of the earlier techniques produces an antiderivative. This page places the integral in a family F(a)F(a) with a parameter aa, differentiates with respect to aa so that the logarithm cancels, justifies that step with an integrable bound and fixes the constant from a known value. The answer is ln⁡2≈0.6931\ln 2 \approx 0.6931.

Key terms

Parameter family of integrals

Definition A parameter family of integrals is a function of a parameter aa defined by an integral in another variable,

F(a)=∫cdf(x,a) dx,F(a) = \int_c^d f(x, a)\,dx,

for the values of aa in a set PP where the integral converges. The limits c<dc < d may be infinite. The integration variable xx is used up by the integral, so FF depends on aa alone. To evaluate a hard integral this way, choose the family so that the target is F(a1)F(a_1) for one value a1a_1, the partial derivative ∂f/∂a\partial f/\partial a is easier to integrate than ff, and F(a0)F(a_0) is known at some a0a_0.

In plain words Instead of one curve there is a whole collection of curves, one for each value of aa, and F(a)F(a) is the area under the curve with that value. The difficult integral is one member of the collection, and the hope is that the way the area changes with aa is simpler than the area itself.

Example For the target integral, replace xx in the numerator by xax^a and keep the −1-1,

F(a)=∫01xa−1ln⁡x dx,F(a) = \int_0^1 \frac{x^a - 1}{\ln x}\,dx,

for a≥0a \ge 0. Then F(1)F(1) is the target integral, and F(0)=0F(0) = 0 because x0−1=0x^0 - 1 = 0 for every xx in (0,1)(0, 1). The power xax^a is the right choice because its derivative in aa is xaln⁡xx^a \ln x, which contains the logarithm of the denominator.

Removable endpoint limits

Definition Let ff be continuous on an open interval (c,d)(c, d) with finite endpoints. If the one-sided limits lim⁡x→c+f(x)\lim_{x \to c^+} f(x) and lim⁡x→d−f(x)\lim_{x \to d^-} f(x) exist as finite numbers, defining ff at cc and at dd by these limits gives a continuous function on [c,d][c, d]. Its integral over [c,d][c, d] is an ordinary definite integral, it equals the improper integral of ff over (c,d)(c, d), and the values at the two endpoints do not affect it.

In plain words A formula that looks undefined at an endpoint, for example of the form 0/00/0, may still have a sensible value there. When both endpoint limits are finite numbers, the gaps can be filled, the graph becomes a continuous curve on a closed interval, and the integral needs no special treatment.

Example For the family integrand (xa−1)/ln⁡x(x^a - 1)/\ln x with a≥0a \ge 0, put t=ln⁡xt = \ln x, which tends to 00 from below as x→1−x \to 1^-. Then

lim⁡x→1−xa−1ln⁡x=lim⁡t→0−eat−1t=a,\begin{aligned} &\lim_{x \to 1^-} \frac{x^a - 1}{\ln x} \\[0.4em] &= \lim_{t \to 0^-} \frac{e^{at} - 1}{t} = a, \end{aligned}

the derivative of eate^{at} at t=0t = 0. As x→0+x \to 0^+ with a>0a > 0, the numerator tends to −1-1 while ln⁡x→−∞\ln x \to -\infty, so the quotient tends to 00, and for a=0a = 0 the integrand is 00 everywhere. For a=1a = 1 the target integrand extends to a continuous function on [0,1][0, 1] that increases from 00 to 11, so its integral lies between 00 and 11.

Differentiation under the integral sign

Definition Let PP be an interval of parameter values and (c,d)(c, d) an interval of xx values, bounded or not. Suppose that

  • the functions f(x,a)f(x, a) and ∂f∂a(x,a)\frac{\partial f}{\partial a}(x, a) are continuous for c<x<dc < x < d and aa in PP,
  • the integral F(a)=∫cdf(x,a) dxF(a) = \int_c^d f(x, a)\,dx converges for every aa in PP,
  • some function gg with ∫cdg(x) dx<∞\int_c^d g(x)\,dx < \infty satisfies ∣∂f∂a(x,a)∣≤g(x)\left\lvert \frac{\partial f}{\partial a}(x, a) \right\rvert \le g(x) for all c<x<dc < x < d and all aa in PP.

Then FF is differentiable on PP, with a one-sided derivative at an endpoint of PP that belongs to PP, and

F′(a)=∫cd∂f∂a(x,a) dx.F'(a) = \int_c^d \frac{\partial f}{\partial a}(x, a)\,dx.

When [c,d][c, d] and PP are closed bounded intervals and ff and ∂f/∂a\partial f/\partial a are continuous on the whole rectangle [c,d]×P[c, d] \times P, the bound is automatic, because a continuous function on a closed bounded rectangle is bounded and a constant is integrable over [c,d][c, d].

In plain words The rate at which an area changes with a parameter is the area under the rate at which the curve changes, provided one fixed function with finite area lies above all of those rates. The method is also known as Feynman’s trick.

Example For the family F(a)F(a) with aa in P=[0,∞)P = [0, \infty), write xa=ealn⁡xx^a = e^{a \ln x} for 0<x<10 < x < 1. Then

∂∂a xa−1ln⁡x=xaln⁡xln⁡x=xa,\frac{\partial}{\partial a}\, \frac{x^a - 1}{\ln x} = \frac{x^a \ln x}{\ln x} = x^a,

and 0<xa≤10 < x^a \le 1, so g(x)=1g(x) = 1 is an integrable bound on (0,1)(0, 1). The theorem gives, for every a≥0a \ge 0 and with a right derivative at a=0a = 0,

F′(a)=∫01xa dx=1a+1,F'(a) = \int_0^1 x^a\,dx = \frac{1}{a + 1},

where the lower limit contributes nothing because a+1>0a + 1 > 0. The functions on [0,∞)[0, \infty) with derivative 1/(a+1)1/(a + 1) are ln⁡(a+1)+C\ln(a + 1) + C, and F(0)=0F(0) = 0 gives C=0C = 0. So F(a)=ln⁡(a+1)F(a) = \ln(a + 1) for a≥0a \ge 0, and the target integral is

∫01x−1ln⁡x dx=F(1)=ln⁡2≈0.6931.\begin{aligned} &\int_0^1 \frac{x - 1}{\ln x}\,dx = F(1) \\[0.4em] &= \ln 2 \approx 0.6931. \end{aligned}

Integrable bound

Definition For a family f(x,a)f(x, a) with xx in (c,d)(c, d) and aa in PP, an integrable bound for the parameter derivative is a function gg that does not depend on aa, with ∫cdg(x) dx<∞\int_c^d g(x)\,dx < \infty and

∣∂f∂a(x,a)∣≤g(x)\left\lvert \frac{\partial f}{\partial a}(x, a) \right\rvert \le g(x)

for all xx in (c,d)(c, d) and all aa in PP. Differentiability is a local property, so it is enough that each value a0a_0 in PP has such a bound for the aa in PP within some fixed distance of a0a_0.

In plain words The bound is one ceiling that lies above the rate of change for every parameter value at once, and the area under the ceiling is finite. Without such a ceiling, the derivative of the integral and the integral of the derivative can differ, or the second one may not exist.

Example In the target family the bound is g(x)=1g(x) = 1 on (0,1)(0, 1), since 0<xa≤10 < x^a \le 1 for all a≥0a \ge 0. Without a bound the step can fail. For every a>0a > 0 the Dirichlet integral

∫0∞sin⁡(ax)x dx=π2\int_0^{\infty} \frac{\sin(ax)}{x}\,dx = \frac{\pi}{2}

does not depend on aa, so its derivative is 00. Differentiating inside the integral instead gives ∫0∞cos⁡(ax) dx\int_0^{\infty} \cos(ax)\,dx, which diverges. No integrable bound exists here, since a bound would have to lie above ∣cos⁡(a0x)∣\lvert \cos(a_0 x) \rvert, whose integral over (0,∞)(0, \infty) is infinite.

Leibniz integral rule

Definition Let ff and ∂f/∂a\partial f/\partial a be continuous on a rectangle [c,d]×[α,β][c, d] \times [\alpha, \beta], and let u(a)u(a) and v(a)v(a) be differentiable on [α,β][\alpha, \beta] with values in [c,d][c, d]. Then for aa in [α,β][\alpha, \beta], with one-sided derivatives at α\alpha and β\beta,

dda∫u(a)v(a)f(x,a) dx=f(v(a),a) v′(a)−f(u(a),a) u′(a)+∫u(a)v(a)∂f∂a(x,a) dx.\begin{aligned} &\frac{d}{da} \int_{u(a)}^{v(a)} f(x, a)\,dx \\[0.4em] &= f(v(a), a)\, v'(a) \\[0.4em] &\quad - f(u(a), a)\, u'(a) \\[0.4em] &\quad + \int_{u(a)}^{v(a)} \frac{\partial f}{\partial a}(x, a)\,dx. \end{aligned}

With fixed limits uu and vv the two boundary terms are zero, which is differentiation under the integral sign on a closed bounded interval.

In plain words When the limits move with the parameter, the area changes for two reasons. The curve changes height, which the integral of ∂f/∂a\partial f/\partial a measures, and the edges move, which adds or removes a thin strip at each end with area about the height there times the distance the edge moves. In the target family the limits 00 and 11 stay fixed, so no edge terms appear.

Example For a>0a > 0 let

G(a)=∫0adxa+x.G(a) = \int_0^a \frac{dx}{a + x}.

Directly, G(a)=ln⁡(2a)−ln⁡a=ln⁡2G(a) = \ln(2a) - \ln a = \ln 2, so G′(a)=0G'(a) = 0. The rule applies on every rectangle [0,β]×[α,β][0, \beta] \times [\alpha, \beta] with 0<α<β0 < \alpha < \beta, where f(x,a)=1/(a+x)f(x, a) = 1/(a + x) and its partial derivative are continuous, and it finds the same derivative as a balance of two terms,

G′(a)=12a−∫0adx(a+x)2=12a−12a=0.\begin{aligned} G'(a) &= \frac{1}{2a} - \int_0^a \frac{dx}{(a + x)^2} \\[0.4em] &= \frac{1}{2a} - \frac{1}{2a} = 0. \end{aligned}

The moving upper edge contributes f(a,a)⋅v′(a)=1/(2a)f(a, a) \cdot v'(a) = 1/(2a), and the falling heights inside take away the same amount.

Constant from a known value

Definition Let FF be continuous on an interval PP with F′(a)=h(a)F'(a) = h(a) at every interior point, and let HH be continuous on PP with H′=hH' = h there as well. Then F(a)=H(a)+CF(a) = H(a) + C for a single constant CC, and one known value F(a0)F(a_0) at a point a0a_0 of PP gives C=F(a0)−H(a0)C = F(a_0) - H(a_0). The derivative alone determines FF only up to this constant, and the known value has to be found independently of the formula being derived.

In plain words Knowing how fast a quantity changes tells you its shape but not its starting height. One value where the integral is easy, often because the integrand vanishes there, pins down the height.

Example In the target family F(a)=ln⁡(a+1)+CF(a) = \ln(a + 1) + C, and at a=0a = 0 the integrand (x0−1)/ln⁡x(x^0 - 1)/\ln x is 00 on (0,1)(0, 1), so F(0)=0F(0) = 0. Since ln⁡1=0\ln 1 = 0, the constant is C=0C = 0. Dropping this step and keeping, say, ln⁡(a+1)+5\ln(a + 1) + 5 would give F(1)=ln⁡2+5F(1) = \ln 2 + 5, which is impossible because the integrand stays below 11.

Check yourself

  1. Evaluate ∫01x2−xln⁡x dx\int_0^1 \frac{x^2 - x}{\ln x}\,dx with the family F(a)=ln⁡(a+1)F(a) = \ln(a + 1) found on this page.
Show the answer

Write the integrand as the difference of two members of the family,

x2−xln⁡x=x2−1ln⁡x−x−1ln⁡x.\frac{x^2 - x}{\ln x} = \frac{x^2 - 1}{\ln x} - \frac{x - 1}{\ln x}.

Both integrals converge, so the integral is F(2)−F(1)=ln⁡3−ln⁡2=ln⁡32≈0.4055F(2) - F(1) = \ln 3 - \ln 2 = \ln \tfrac32 \approx 0.4055.

  1. For b>0b > 0 let
G(b)=∫0∞e−x−e−bxx dx.G(b) = \int_0^{\infty} \frac{e^{-x} - e^{-bx}}{x}\,dx.

Find G(b)G(b) by differentiating with respect to bb. Which integrable bound justifies the step, and which known value fixes the constant?

Show the answer

The integrand has the removable limit b−1b - 1 as x→0+x \to 0^+ and decays exponentially as x→∞x \to \infty, so G(b)G(b) converges. Its derivative with respect to bb is e−bxe^{-bx}. One bound for all b>0b > 0 is impossible, because such a bound would be at least 11 for every xx, which is not integrable over (0,∞)(0, \infty). Instead fix any b0>0b_0 > 0. For b≥b0b \ge b_0,

0<e−bx≤e−b0x,0 < e^{-bx} \le e^{-b_0 x},

and ∫0∞e−b0x dx=1/b0\int_0^{\infty} e^{-b_0 x}\,dx = 1/b_0 is finite. Since b0>0b_0 > 0 was arbitrary, for every b>0b > 0

G′(b)=lim⁡t→∞∫0te−bx dx=1b,\begin{aligned} G'(b) &= \lim_{t \to \infty} \int_0^t e^{-bx}\,dx \\[0.4em] &= \frac{1}{b}, \end{aligned}

and G(b)=ln⁡b+CG(b) = \ln b + C. The known value is G(1)=0G(1) = 0, because the integrand is 00 when b=1b = 1. Hence C=0C = 0 and G(b)=ln⁡bG(b) = \ln b, for example G(2)=ln⁡2≈0.6931G(2) = \ln 2 \approx 0.6931.

  1. Let H(b)=∫0∞e−x2cos⁡(bx) dxH(b) = \int_0^{\infty} e^{-x^2} \cos(bx)\,dx for real bb. Use differentiation under the integral sign, integration by parts and the value H(0)=π/2H(0) = \sqrt{\pi}/2 to find H(b)H(b).
Show the answer

The derivative of the integrand with respect to bb is −xe−x2sin⁡(bx)-x e^{-x^2} \sin(bx), and its absolute value is at most xe−x2x e^{-x^2} for every bb. That bound is integrable, since ∫0∞xe−x2 dx=12\int_0^{\infty} x e^{-x^2}\,dx = \tfrac12, so H′(b)H'(b) is the integral of −xe−x2sin⁡(bx)-x e^{-x^2} \sin(bx) from 00 to ∞\infty. Integrate by parts with u=sin⁡(bx)u = \sin(bx) and v=12e−x2v = \tfrac12 e^{-x^2}, so that dv=−xe−x2 dxdv = -x e^{-x^2}\,dx and du=bcos⁡(bx) dxdu = b\cos(bx)\,dx. The boundary term uv=12e−x2sin⁡(bx)uv = \tfrac12 e^{-x^2} \sin(bx) is 00 at x=0x = 0 and tends to 00 as x→∞x \to \infty. What remains is −∫0∞v du-\int_0^{\infty} v\,du, which is −b2-\tfrac{b}{2} times the integral defining H(b)H(b), so

H′(b)=−b2 H(b).H'(b) = -\frac{b}{2}\, H(b).

By the product rule, the derivative of eb2/4H(b)e^{b^2/4} H(b) is eb2/4(H′(b)+b2H(b))=0e^{b^2/4}\left(H'(b) + \tfrac{b}{2} H(b)\right) = 0, so eb2/4H(b)e^{b^2/4} H(b) is constant. Its value at b=0b = 0 is H(0)=π/2H(0) = \sqrt{\pi}/2, the half-line Gaussian integral. Therefore

∫0∞e−x2cos⁡(bx) dx=π2 e−b2/4,\begin{aligned} &\int_0^{\infty} e^{-x^2} \cos(bx)\,dx \\[0.4em] &= \frac{\sqrt{\pi}}{2}\, e^{-b^2/4}, \end{aligned}

and for example H(2)=π/(2e)≈0.3260H(2) = \sqrt{\pi}/(2e) \approx 0.3260.

How this lesson connects