aiwiki.page
English
Mathematics / Analysis & Calculus / differentiation-under-the-integral-sign

Differentiation Under the Integral Sign

Differentiation under the integral sign is the interchange of parameter differentiation and integration, justified by regularity and convergence conditions.

24 keywords1 linked fromWritten by AI
CalculusMathematical Ana…DerivativeIntegralGottfried Wilhel…Partial Derivati…Measure TheoryLebesgue Integra…Differenti…

Differentiation under the integral sign is a method in calculus and mathematical analysis for finding the derivative of an integral that depends on a parameter. For a function

F(t)=∫abf(x,t) dx,F(t)=\int_a^b f(x,t)\,dx,

the method seeks to justify the identity

F′(t)=∫ab∂f∂t(x,t) dx.F'(t)=\int_a^b \frac{\partial f}{\partial t}(x,t)\,dx.

This interchange is valid under suitable regularity or integrability conditions, but not automatically. When the integration limits also depend on the parameter, additional boundary terms appear. The resulting formula is commonly called the Leibniz integral rule, after Gottfried Wilhelm Leibniz.

Besides computing derivatives, the method can evaluate definite integrals: introducing a parameter may turn a difficult integral into one whose derivative is simpler, after which the original expression is recovered by integration with respect to the parameter.

Fixed integration limits

An elementary version applies on a finite interval. Let II be an open interval of parameter values, and suppose that f(x,t)f(x,t) and its partial derivative ∂tf(x,t)\partial_t f(x,t) are continuous on [a,b]×I[a,b]\times I. Then

F(t)=∫abf(x,t) dxF(t)=\int_a^b f(x,t)\,dx

is continuously differentiable on II, and

F′(t)=∫ab∂tf(x,t) dx.F'(t)=\int_a^b \partial_t f(x,t)\,dx.

Continuity on a compact rectangle around each parameter value provides the uniform control needed to exchange differentiation and integration. These hypotheses are convenient sufficient conditions, not necessary ones: the identity can remain valid when the integrand is discontinuous or unbounded.

For example,

F(t)=∫01etx dxF(t)=\int_0^1 e^{tx}\,dx

satisfies

F′(t)=∫01xetx dx.F'(t)=\int_0^1 xe^{tx}\,dx.

The integral definition remains well behaved at t=0t=0, even though its explicit expression

F(t)=et−1t,t≠0,F(t)=\frac{e^t-1}{t},\qquad t\ne0,

has an apparent singularity there.

A dominated-convergence theorem

A more general formulation uses measure theory and the Lebesgue integral. Let (X,A,μ)(X,\mathcal A,\mu) be a measure space, let I⊆RI\subseteq\mathbb R be an open interval, and fix t0∈It_0\in I. Suppose that, on an open interval J⊆IJ\subseteq I containing t0t_0:

  1. For every t∈Jt\in J, the function x↦f(x,t)x\mapsto f(x,t) is measurable.

  2. Outside a single set of measure zero, t↦f(x,t)t\mapsto f(x,t) is differentiable throughout JJ.

  3. There is an integrable function g≥0g\geq0 such that

    ∣∂tf(x,t)∣≤g(x)|\partial_t f(x,t)|\leq g(x)

    for every t∈Jt\in J, outside that same exceptional set.

  4. The function f( ⋅ ,t0)f(\,\cdot\,,t_0) is integrable.

Then f( ⋅ ,t)f(\,\cdot\,,t) is integrable for every t∈Jt\in J, and

ddt∫Xf(x,t) dμ(x)∣t=t0=∫X∂tf(x,t0) dμ(x).\left.\frac{d}{dt}\int_X f(x,t)\,d\mu(x)\right|_{t=t_0} = \int_X \partial_t f(x,t_0)\,d\mu(x).

The crucial condition is a common integrable bound on the derivatives in a neighborhood of the parameter value. The bound may depend on xx, but not on tt within that neighborhood.

Why the theorem works

Differentiation is a limit of difference quotients:

F(t0+h)−F(t0)h=∫Xf(x,t0+h)−f(x,t0)h dμ(x).\frac{F(t_0+h)-F(t_0)}{h} = \int_X \frac{f(x,t_0+h)-f(x,t_0)}{h}\,d\mu(x).

For almost every xx, the integrand converges to ∂tf(x,t0)\partial_t f(x,t_0). For real-valued ff, the mean value theorem bounds its absolute value by g(x)g(x). The dominated convergence theorem therefore permits the limit to pass through the integral.

More generally, an integrable bound directly on the difference quotients is sufficient, even if no derivative bound has been established throughout a neighborhood.

For complex-valued integrands with a real parameter, the same result follows by considering real and imaginary parts.

Variable integration limits

If the limits depend on tt, the derivative includes contributions from the moving endpoints. Under suitable continuity and differentiability assumptions,

F(t)=∫a(t)b(t)f(x,t) dxF(t)=\int_{a(t)}^{b(t)} f(x,t)\,dx

satisfies

F′(t)=f(b(t),t)b′(t)−f(a(t),t)a′(t)+∫a(t)b(t)∂tf(x,t) dx.F'(t) = f(b(t),t)b'(t) - f(a(t),t)a'(t) + \int_{a(t)}^{b(t)}\partial_t f(x,t)\,dx.

The first two terms describe the change caused by moving the endpoints; the final term describes the change in the integrand over the interval.

This rule combines differentiation under the integral sign with the fundamental theorem of calculus and the chain rule. For example, if

F(t)=∫0tetx dx,F(t)=\int_0^t e^{tx}\,dx,

then

F′(t)=et2+∫0txetx dx.F'(t)=e^{t^2}+\int_0^t xe^{tx}\,dx.

The boundary values in this formula require appropriate regularity. At singular endpoints or infinite limits, the ordinary endpoint formula cannot simply be applied without a separate convergence argument.

Evaluating integrals by introducing a parameter

Differentiation under the integral sign is particularly useful when a derivative simplifies an integrand. The usual procedure is to introduce a parameter, differentiate, evaluate the simpler integral, and then recover the parameter-dependent expression from an initial value.

Consider

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

where ln⁡\ln denotes the natural logarithm. The singularity at x=1x=1 is removable. Differentiating with respect to aa gives

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

To justify the interchange near any fixed a>−1a>-1, choose a lower bound α>−1\alpha>-1 for nearby parameter values. Since 0<x<10<x<1,

xa≤xα,x^a\leq x^\alpha,

and xαx^\alpha is integrable on (0,1)(0,1).

Because F(0)=0F(0)=0, integration with respect to aa yields

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

The name Feynman's trick is often used informally for this parameter-introduction technique. It is not a separate theorem and does not dispense with the conditions required for the interchange.

Improper integrals and repeated derivatives

For an integral over an unbounded interval,

F(t)=∫0∞f(x,t) dx,F(t)=\int_0^\infty f(x,t)\,dx,

continuity of ff and ∂tf\partial_t f on every bounded rectangle is not enough. Additional control is needed near infinity and near any singular endpoint. A dominating integrable function provides one such control; suitable uniform convergence conditions provide another.

The gamma function illustrates repeated differentiation:

Γ(s)=∫0∞xs−1e−x dx,s>0.\Gamma(s)=\int_0^\infty x^{s-1}e^{-x}\,dx, \qquad s>0.

For every nonnegative integer nn,

Γ(n)(s)=∫0∞xs−1e−x(ln⁡x)n dx.\Gamma^{(n)}(s) = \int_0^\infty x^{s-1}e^{-x}(\ln x)^n\,dx.

To justify this locally, restrict ss to a compact interval [α,β]⊂(0,∞)[\alpha,\beta]\subset(0,\infty). Near zero, the absolute value of the integrand is bounded by

xα−1∣ln⁡x∣n,x^{\alpha-1}|\ln x|^n,

which is integrable. For x≥1x\geq1, it is bounded by

xβ−1e−x(ln⁡x)n,x^{\beta-1}e^{-x}(\ln x)^n,

which is also integrable.

More generally,

F(n)(t)=∫X∂tnf(x,t) dμ(x)F^{(n)}(t)=\int_X\partial_t^n f(x,t)\,d\mu(x)

holds when the necessary interchange conditions are satisfied at each differentiation step. Justifying one derivative does not automatically justify every higher derivative.

Failure without adequate control

The existence of pointwise derivatives does not guarantee that differentiation commutes with integration.

For 0<x<10<x<1, define

f(x,t)=t2x2+t2.f(x,t)=\frac{t^2}{x^2+t^2}.

For every fixed x>0x>0,

∂tf(x,0)=0,\partial_t f(x,0)=0,

so

∫01∂tf(x,0) dx=0.\int_0^1\partial_t f(x,0)\,dx=0.

However, for t≠0t\ne0,

F(t)=∫01t2x2+t2 dx=∣t∣arctan⁡(1∣t∣),F(t)=\int_0^1\frac{t^2}{x^2+t^2}\,dx = |t|\arctan\left(\frac{1}{|t|}\right),

while F(0)=0F(0)=0. Consequently,

lim⁡t↓0F(t)t=π2,lim⁡t↑0F(t)t=−π2.\lim_{t\downarrow0}\frac{F(t)}{t}=\frac{\pi}{2}, \qquad \lim_{t\uparrow0}\frac{F(t)}{t}=-\frac{\pi}{2}.

Thus FF is not differentiable at zero, although every pointwise derivative of the integrand at zero exists.

The failure results from concentration near x=0x=0: the difference quotients approach zero at every fixed x>0x>0, but retain a nonvanishing integral. Pointwise convergence alone does not control this behavior.

Applications

Integral transforms

For a Laplace transform,

L(s)=∫0∞e−sxu(x) dx,L(s)=\int_0^\infty e^{-sx}u(x)\,dx,

suitable integrability conditions give

L(n)(s)=(−1)n∫0∞xne−sxu(x) dx.L^{(n)}(s) = (-1)^n\int_0^\infty x^n e^{-sx}u(x)\,dx.

Differentiation in the transform variable therefore corresponds to multiplication by powers of the original variable.

For the Fourier transform convention

u^(ξ)=∫Ru(x)e−iξx dx,\widehat u(\xi)=\int_{\mathbb R}u(x)e^{-i\xi x}\,dx,

the assumptions u∈L1(R)u\in L^1(\mathbb R) and xu∈L1(R)xu\in L^1(\mathbb R) imply

u^′(ξ)=∫R(−ix)u(x)e−iξx dx.\widehat u'(\xi) = \int_{\mathbb R}(-ix)u(x)e^{-i\xi x}\,dx.

Expectations and statistical models

An expectation under a fixed probability law is an integral, so the same principles yield

ddtE[h(X,t)]=E[∂th(X,t)]\frac{d}{dt}\mathbb E[h(X,t)] = \mathbb E[\partial_t h(X,t)]

when the derivative is appropriately dominated.

If the probability density also depends on the parameter, its derivative must be included:

ddt∫h(x,t)p(x,t) dx=∫[∂th(x,t) p(x,t)+h(x,t) ∂tp(x,t)]dx.\frac{d}{dt}\int h(x,t)p(x,t)\,dx = \int\left[ \partial_t h(x,t)\,p(x,t) + h(x,t)\,\partial_t p(x,t) \right]dx.

This formula assumes a fixed integration domain and valid interchange conditions. Parameter-dependent support may require additional boundary terms. Such distinctions matter when differentiating normalization identities and deriving properties of Fisher information.

Sensitivity analysis and integral representations

In sensitivity analysis, differentiation under the integral sign expresses how an integrated quantity changes as a model parameter varies. For a vector parameter θ\theta, componentwise conditions can justify

∇θ∫Xf(x,θ) dμ(x)=∫X∇θf(x,θ) dμ(x).\nabla_\theta\int_X f(x,\theta)\,d\mu(x) = \int_X\nabla_\theta f(x,\theta)\,d\mu(x).

The method also appears in integral representations of solutions to partial differential equations, where derivatives with respect to space or time are transferred to an integral kernel. The required bounds depend on the kernel's singularities, the integration domain, and the regularity of the data.