Repeated Differentiation Requires Bounds at Every Order
For an integral depending on a real parameter, differentiating once often introduces a new factor into the integrand. Repeating the operation introduces further derivatives, and each step needs its own justification. A bound that works for the original integrand need not control all its parameter derivatives. The useful condition is local in the parameter: on each compact parameter interval, every derivative under consideration has an integrable majorant independent of the parameter.
We work on an arbitrary measure space \((X,\mathcal{M},\mu)\); no finiteness or sigma-finiteness assumption on \(\mu\) is needed. All functions in the theorem are real-valued, and all pointwise-in-\(x\) statements may exclude one fixed \(\mu\)-null set. We use the Dominated Convergence Theorem in the following form: if measurable functions converge \(\mu\)-a.e. and their absolute values are bounded by one integrable function, their integrals converge to the integral of their limit.
\(F^{(j)}(t)=\int_X f_j(t,x)\,d\mu(x)\) for \(j=0,\ldots,k\).
Moreover, \(F^{(j)}\) is continuous for \(0\le j<k\). If, in addition, \(t\mapsto f_k(t,x)\) is continuous for \(\mu\)-a.e. \(x\), then \(F^{(k)}\) is continuous as well, so \(F\in C^k(I)\).
Proof. Fix a compact interval \(K\subset I\), and first consider \(t\) in its interior. For \(0\le j<k\), the pointwise function \(t\mapsto f_j(t,x)\) is continuous outside \(N\), since it has derivative \(f_{j+1}(t,x)\). If \(t_n\to t\) with \(t_n,t\in K\), then \(f_j(t_n,x)\to f_j(t,x)\) for \(x\notin N\), and \(|f_j(t_n,x)|\le g_{K,j}(x)\). The Dominated Convergence Theorem gives \(\int f_j(t_n,x)\,d\mu(x)\to\int f_j(t,x)\,d\mu(x)\). Thus each such integral is continuous in \(t\).
We now establish the derivative identity successively. Suppose \(0\le j<k\), and define \(G_j(t)=\int_X f_j(t,x)\,d\mu(x)\). For nonzero \(h\) small enough that the segment from \(t\) to \(t+h\) lies in a compact interval \(K\subset I\), the scalar mean value theorem, applied for each \(x\notin N\), gives \[ \left|\frac{f_j(t+h,x)-f_j(t,x)}{h}\right|\le g_{K,j+1}(x). \] The difference quotient converges pointwise outside \(N\) to \(f_{j+1}(t,x)\). Its terms are measurable, and the displayed majorant is integrable. Dominated Convergence therefore gives \[ \lim_{h\to0}\int_X\frac{f_j(t+h,x)-f_j(t,x)}{h}\,d\mu(x) =\int_X f_{j+1}(t,x)\,d\mu(x). \] Both integrals in the quotient are finite by the assumed bounds, so linearity of the integral identifies the left side with \(G_j'(t)\). Starting with \(G_0=F\) and repeating this argument proves all the derivative identities. The continuity argument in the first paragraph applies to \(G_j\) for \(j<k\). If \(f_k(t_n,x)\to f_k(t,x)\) a.e. whenever \(t_n\to t\), the same argument with \(g_{K,k}\) proves continuity of \(G_k\). This proves the assertions. \(\square\)
Finite-Interval Exponential Kernels
Worked Example: Differentiating an Exponential Integral at Zero
For \(t\in\mathbb{R}\), let \(F(t)=\int_0^1 e^{tx}\,dx\), with Lebesgue measure \(m\). Compute \(F'(0)\) and \(F''(0)\), justifying the differentiations.
For each integer \(j\ge0\), the \(j\)th parameter derivative of the integrand is \(f_j(t,x)=x^j e^{tx}\). On any compact \(K\subset\mathbb{R}\), let \(b=\max_{t\in K}|t|\). Since \(0\le x\le1\), \(|x^j e^{tx}|\le e^b\). The constant function \(e^b\) is integrable on \([0,1]\), so the theorem applies at every finite order. In particular, \[ F'(0)=\int_0^1 x\,dx=\left[\frac{x^2}{2}\right]_0^1=\frac12, \qquad F''(0)=\int_0^1 x^2\,dx=\left[\frac{x^3}{3}\right]_0^1=\frac13. \] For a check, when \(t\ne0\), direct integration gives \(F(t)=(e^t-1)/t\). The expansion \(e^t=1+t+t^2/2+t^3/6+\cdots\) gives \(F(t)=1+t/2+t^2/6+\cdots\) near zero, which yields \(F'(0)=1/2\) and \(F''(0)=1/3\), agreeing with the integrals.
Repeated Derivatives of Laplace Integrals
For exponential kernels on an unbounded domain, the integrable majorants must control both the polynomial factors produced by differentiation and the exponential tail. A positive lower bound on the parameter provides that control locally.
Worked Example: The Laplace Transform of a Quadratic
For \(t>0\), set \(F(t)=\int_0^\infty x^2e^{-tx}\,dm(x)\). Find \(F''(t)\) by differentiating under the integral, and evaluate it at \(t=2\).
The \(j\)th derivative of the integrand is \(f_j(t,x)=(-x)^j x^2e^{-tx}\). If \(K\subset(0,\infty)\) is compact and \(a=\min K>0\), then \(|f_j(t,x)|\le x^{j+2}e^{-ax}\) for \(t\in K\). This majorant is integrable: repeated integration by parts gives \(\int_0^\infty x^n e^{-ax}\,dx=n!/a^{n+1}\) for every integer \(n\ge0\). Thus the theorem applies at every order, and \[ F''(t)=\int_0^\infty x^4e^{-tx}\,dm(x). \] To evaluate the integral, integration by parts, with the boundary terms vanishing at infinity because \(x^n e^{-tx}\to0\), gives the recurrence \(I_n(t)=\frac{n}{t}I_{n-1}(t)\), where \(I_n(t)=\int_0^\infty x^n e^{-tx}\,dm(x)\) and \(I_0(t)=1/t\). Hence \(F''(t)=I_4(t)=24/t^5\). At \(t=2\), this is \(24/2^5=24/32=3/4\). The integral formula gives the same check directly: \(\int_0^\infty x^4e^{-2x}\,dm(x)=4!/2^5=24/32=3/4\).
Worked Example: A Sum of Two Exponential Moments
For \(t>0\), let \(H(t)=\int_0^\infty (1+x)e^{-tx}\,dm(x)\). Compute \(H''(1)\) and verify it from the differentiated integrand.
For \(j\ge0\), the derivative of order \(j\) of the integrand is \((-x)^j(1+x)e^{-tx}\). On a compact parameter set \(K\subset(0,\infty)\), with \(a=\min K\), its absolute value is at most \(x^j(1+x)e^{-ax}\), an integrable function by the integration-by-parts formula above. Therefore repeated differentiation is justified. First, \[ H(t)=\int_0^\infty e^{-tx}\,dm(x)+\int_0^\infty xe^{-tx}\,dm(x) =\frac1t+\frac1{t^2}. \] Consequently \(H''(t)=2/t^3+6/t^4\), so \(H''(1)=2+6=8\). On the integral side, two differentiations give \(H''(1)=\int_0^\infty x^2(1+x)e^{-x}\,dm(x)\). Evaluating term by term gives \(2!+3!=2+6=8\), the same value.
What the Bounds Do—and Do Not—Guarantee
The bound for \(f_j\) is used to establish continuity of the integral of \(f_j\); the bound for \(f_{j+1}\) controls the difference quotients used to differentiate that integral. Thus bounds are required at every order up to \(k\), not just for the original integrand. The theorem does not require the measure space to be finite: what matters is that the chosen majorants are integrable.
A common mistake is to infer continuity of the highest derivative without assuming pointwise continuity of the highest derivative. For a concrete example, take \(X\) to be a singleton with measure one, \(I=(-1,1)\), and \[ f_0(t)= \begin{cases} t^2\sin(1/t),&t\ne0,\\ 0,&t=0. \end{cases} \] Its derivative is \(f_1(t)=2t\sin(1/t)-\cos(1/t)\) for \(t\ne0\), and \(f_1(0)=0\). Both functions are bounded, so the local integrable bounds for \(k=1\) hold on this measure space. But for \(t_n=1/(2\pi n)\), \(f_1(t_n)=-1\), whereas \(f_1(0)=0\). Thus \(f_1\) is not continuous at zero. The theorem correctly gives \(F'=f_1\), but it does not give \(F\in C^1\) without the extra continuity hypothesis on \(f_1\).
A second error is to differentiate an integral merely because the integrand has pointwise derivatives. Pointwise differentiability alone does not provide an integrable majorant for the difference quotients. In the proof, the mean value theorem and the bound for the next derivative supply exactly that majorant; the Dominated Convergence Theorem then justifies the limit of the integrals. Omitting either step leaves the interchange unjustified.
For exponential kernels, the practical check is systematic: restrict the parameter to a compact subset of \((0,\infty)\), take its positive minimum \(a\), and bound each differentiated integrand by a polynomial times \(e^{-ax}\). On a bounded \(x\)-interval, boundedness of the parameter gives an integrable constant bound instead. These checks explain why repeated differentiation is available in both finite-interval integrals and Laplace integrals, while making explicit which conditions each setting supplies.
Check Your Understanding
For each question, identify the measure space, the derivative integrand, and the bound that permits the relevant limit.
- For \(F(t)=\int_0^1 e^{tx}\,dm(x)\), compute \(F^{(3)}(0)\) by differentiating under the integral.
- For \(G(t)=\int_0^\infty e^{-tx}\,dm(x)\), \(t>0\), find \(G^{(3)}(t)\) and state an integrable local majorant for the third derivative.
- Why does the higher differentiation theorem guarantee continuity of \(F^{(j)}\) for \(j<k\) under its assumptions, but require an extra assumption for continuity of \(F^{(k)}\)?
- In the singleton example above, calculate \(f_1(1/(2\pi n))\) and explain why this contradicts continuity at zero.
- Which derivative-order bound controls the difference quotients used to prove \(F^{(j+1)}(t)=\int_X f_{j+1}(t,x)\,d\mu(x)\)?