5

I would like to know how to justify rigorously that $$-\frac{1}{\pi}\int_0^\infty t\sin\left(\frac{t^3}{3}+0\right)dt=\operatorname{Ai}'(0).\tag{1}$$

I've performed the differentiation under the integral sign for the integral representation of the Airy function $\operatorname{Ai}(x)$ and with the help of Wolfram Alpha online calculator:

int -1/pi t*sin(t^3/3)dt, from t=0 to infinity

I did a comparison. I've used the formula from this Wikipedia.

Question. I know that numerically my statement $(1)$ is right, but what details are required for a rigorous proof? I only request a guide with hints about what is required to check, even the convergence of the resulting improper integral. Thank you very much.

Hans
  • 10,484
  • Many thanks @Chappers –  Oct 06 '17 at 22:01
  • 1
    I am putting my input as a comment because I have not finished working out the details. One way to start is 'integration by parts'. Transform the integral to an absolutely convergent integral and try using the Dominated Convergence Theorem. – Sungjin Kim Oct 06 '17 at 22:47
  • Please @i707107 add your answer when you want, and then I and the other users interested in the justification, can to read and study your reasoning. Many thanks. –  Oct 07 '17 at 10:03
  • Many thanks @Hans improving the post. –  Oct 07 '17 at 10:04
  • You are welcome. I have put up a proof of a slightly more general formulation. Check it out if you are still interested. – Hans Oct 08 '17 at 06:44

3 Answers3

1

By addition formula for $\cos$, we have $\cos(t^3/3 + xt) = \cos(t^3/3)\cos xt - \sin(t^3/3) \sin xt$.

If we can prove that $f(x)=\int_0^{\infty} \cos(t^3/3) \cos xt dt$ is differentiable under integral sign, then the same argument will work for $g(x) = \int_0^{\infty} \sin(t^3/3)\sin xt dt$. For $g(x)$, we need to split the integral into $\int_0^1$ and $\int_1^{\infty}$ due to the singularity at $t=0$.

Integration by parts ($u=(\cos xt)/t^2$, $dv=t^2\cos(t^3/3)dt$) --- my previous answer had an error in finding $du$ --- gives $$ f(x) = \frac{\sin(t^3/3)\cos xt}{t^2} \Bigg\vert_0^{\infty} + \int_0^{\infty} \frac{ x t^2\sin xt+2t\cos xt}{t^4} \sin(t^3/3) dt. $$

Let us consider $\lim_{x\rightarrow 0+} \frac{f(x)-f(0)}x$, since $x\rightarrow 0-$ can be treated the same way. So, let $x_n$ be a sequence of positive real numbers converging to $0$. Then $$ \begin{align} \frac{f(x_n)-f(0)}{x_n} &= \int_0^{\infty} \left( \frac{ x_n t^2\sin x_n t+2t(\cos x_n t -1)}{x_nt^4}\right) \sin (t^3/3) dt \end{align} $$ By the Mean Value Theorem, inside of the big paranthesis is bounded by $C/t^2$ uniformly in $n$, for some $C>0$. Thus, we can apply the Dominated Convergence Theorem to obtain $$\begin{align} \lim_{n\rightarrow\infty} \frac{f(x_n)-f(0)}{x_n} &= \int_0^{\infty} \lim_{n\rightarrow\infty} \left( \frac{ x_n t^2\sin x_n t+2t(\cos x_n t -1)}{x_nt^4}\right) \sin (t^3/3) dt < \infty.\end{align} $$ Thus, the results are the same with $x=0$ plugged in to the expression of $f'$ after differentiated under integral.

Let $g(x)=g_1(x)+g_2(x)$ where the integral is performed as $\int_0^1$ for $g_1$ and $\int_1^{\infty}$ for $g_2$. Then $g_1'$ under integral sign is justified by the Dominated Convergence Theorem without resorting to the integration by parts. We treat $g_2$ with the integration parts.

Integration by parts ($u=(\sin xt)/t^2$, $dv=t^2\sin(t^3/3)$) gives $$ g_2(x) = \frac{-\cos(t^3/3)\sin xt}{t^2} \Bigg\vert_1^{\infty} + \int_1^{\infty} \frac{xt^2\cos xt-2t\sin xt}{t^4} \cos(t^3/3) dt. $$

Let us consider $\lim_{x\rightarrow 0+} \frac{g_2(x)-g_2(0)}x$, since $x\rightarrow 0-$ can be treated the same way. So, let $x_n$ be a sequence of positive real numbers converging to $0$. Then $$ \begin{align} \frac{g_2(x_n)-g_2(0)}{x_n} &= \cos(1/3)\frac{\sin x_n}{x_n}+\int_1^{\infty} \left( \frac{x_nt^2\cos x_nt - 2t\sin x_n t}{x_nt^4}\right) \cos (t^3/3) dt \end{align} $$ By the Mean Value Theorem, inside of the big paranthesis is bounded by $C/t^2$ uniformly in $n$, for some $C>0$. Thus, we can apply the Dominated Convergence Theorem to obtain $$\begin{align} \lim_{n\rightarrow\infty} \frac{g_2(x_n)-g_2(0)}{x_n} &= \cos(1/3)\frac{\sin x_n}{x_n}+ \int_1^{\infty} \lim_{n\rightarrow\infty} \left( \frac{x_nt^2\cos x_nt - 2t\sin x_n t}{x_nt^4}\right) \cos (t^3/3) dt\\ &= \cos(1/3) + \int_1^{\infty} \left(\frac1{t^2}-\frac2{t^2}\right)\cos(t^3/3)dt\end{align} $$ Thus, the results are the same with $x=0$ plugged in to the expression of $g_2'$ after differentiated under integral. In fact, integration by parts ($u=1/t $, $dv=t^2\sin(t^3/3)dt$) $$ \int_1^{\infty} t\sin(t^3/3) dt=\cos(1/3) + \int_1^{\infty} \left(-\frac1{t^2} \right)\cos(t^3/3)dt. $$

Hans
  • 10,484
Sungjin Kim
  • 20,850
  • You missed to divide by $x_n$ "inside of the big parenthesis". $\frac{\sin x_nt}{x_nt}$ "inside of the big parenthesis" is bounded not by $C_1/t$ but $C_1$, so "inside of the big parenthesis" is not bounded by $C/t^2$ but $C/t$. So it is not sufficient to use the dominated convergence theorem. You may need another round of integration by parts. – Hans Oct 07 '17 at 05:23
  • @Hans Thank you for this, you are right. I will edit this. – Sungjin Kim Oct 07 '17 at 06:00
  • If $u=\frac{\cos(xt)}{t^2}$, then $u'(t)=-,\frac{xt^2\sin(xt)+2t\cos(xt)}{t^4}\ne\frac{xt^2\sin(xt)-2t\cos(xt)}{t^4} $. – Mark Viola Oct 07 '17 at 17:36
  • @MarkViola Thank you. – Sungjin Kim Oct 07 '17 at 17:38
  • @Hans It turns out it is bounded by $C/t^2$. I made mistakes in IBP previously. It is edited now. – Sungjin Kim Oct 07 '17 at 17:40
  • I am going to study your answer. Many thanks to you, Mark and to AngusTheMan, thanks as users as you this site is great. –  Oct 07 '17 at 18:14
  • Yes, this current version is correct. +1 I am finishing up an answer for a more general formulation. I will ask you to review it when I complete it. – Hans Oct 08 '17 at 02:25
  • I have now completed my proof of a more general formulation. I would appreciate it if you would care to review it. – Hans Oct 08 '17 at 06:42
1

We shall prove a more general result. Let $\varphi\in C^{(1)}[0,\infty) \land \big(|\varphi(t)|<1,\,\forall t\in[0,\infty)\big)$. Let $f(t,x) = \varphi\big(\frac{t^3}3+xt\big)$. We will show the differentiation under the integral works for all $x\ge 0$. $I(x)=\int_0^\infty f(t,x)dt$. We need to show $$\lim_{\delta\to 0}\int_0^\infty \frac{f(t,x+\delta)-f(t,x)}{\delta}dt = \int_0^\infty \lim_{\delta\to 0}\frac{f(t,x+\delta)-f(t,x)}{\delta}dt. \tag1$$

Define \begin{align} I_1(x) &:= \int_0^1\varphi\big(\frac{t^3}3+xt\big)\,dt \\ I_2(x) &:= \int_1^\infty\varphi\big(\frac{t^3}3+xt\big)\,dt \end{align} Consider $I_2$ first.

Let $u:=\frac{t^3}3+xt$. Surely $u>0$. $$u<\frac{(t+x)^3}3\Longleftrightarrow (3u)^\frac13-x<t,\,\forall (t>1\land x\ge0). \tag2$$ Now $(u,x)$ is the independent variable, while $t$ becomes a function of $(u,x)$ and is denoted as $t(u,x)$. $$1 = (t^2+x)\frac{\partial t}{\partial u}$$ $$0=t^2\frac{\partial t}{\partial x}+t+x\frac{\partial t}{\partial x}\Longleftrightarrow \frac{\partial t}{\partial u}=-\frac{t}{t^2+x}$$ $$\frac{\partial}{\partial x}\frac1{t(u,x)^2+x}=-\frac{2t\frac{\partial t}{\partial x}+1}{(t^2+x)^2}=\frac1{(t^2+x)^2}\frac{t^2-x}{t^2+x},$$

$$I_2(x)=\int_{\frac13+x}^\infty \frac{\varphi(u)}{t^2+x}du,$$ and $$\frac{I_2(x+\delta)-I_2(x)}\delta=J_{2,1}-J_{2,2}$$ where \begin{align} J_{2,1}&:=\int_{\frac13+x}^\infty\varphi(u)g(u,x,\delta)du \\ J_{2,2}&:= \frac1\delta\int_{\frac13+x}^{\frac13+x+\delta}\frac{\varphi(u)}{t(u,x+\delta)^2+x+\delta}\,du, \end{align} where \begin{align} g(u,x,\delta) &:= \frac1\delta\Big(\frac1{t(u,x+\delta)^2+x+\delta}-\frac1{t(u,x)+x}\Big) \\ &=\frac1{(t(u,x+\theta\delta)^2+x+\theta\delta)^2}\frac{t(u,x+\theta\delta)^2-(x+\theta\delta)}{t(u,x+\theta\delta)^2+(x+\theta\delta)} \end{align} for some function $\theta:=\theta(u,x,\delta)\in[0,1]$ with the second equation coming from Taylor expansion to the first order.

By inequality (2), for some positive function $C$, $\forall u>C(x)$, $$t(u,x)^2+x>((3u)^\frac13-x)^2+x>u^\frac23,$$ $$0<\frac{t(u,x)^2-x}{t(u,x)^2+x}\le 1$$ $\forall u>C(u+\Delta),\ \forall \delta<\Delta$, for some $\Delta>0$, $$0<g(u,x,\delta)<\frac1{u^\frac43}$$ So $$\int_{C(x+\Delta)}^\infty g(u,x,\delta)\varphi(u)du<\int_{C(x+\Delta)}^\infty\frac1{u^\frac43}du<\infty$$ For $g(u,x,\delta)$ is continuous for $(u,\delta)$ on a compact set $\big[\frac13+x,C(x+\Delta)\big]\times[0,\Delta]$ and thus is bounded for any given $x>0$. Hence we can apply the Lebesgue's dominated convergence theorem to the integral $J_{2,1}$ as $\delta\to0$.

It is easy to see $J_{2,2}\to\frac{\varphi(\frac13+x)}{1+x}$ as $\delta\to0$ since the integrand is continuous on a compact set.

Now we consider $$\frac{I_1(x+\delta)-I_1(x)}{\delta} =\int_0^1\frac1\delta\Big(\varphi\Big(\frac{t^3}3+x+\delta\Big)-\varphi\Big(\frac{t^3}3+x\Big)\Big)\, dt =\int_0^1\varphi'\Big(\frac{t^3}3+x+\theta_1\delta\Big)\, dt $$ for some function $\theta_1:=\theta_1(\frac{t^3}3+x,\delta)\in[0,1]$. As $\varphi\in C^{(1)}$ on a compact set, it is uniformly bounded with respect to $(t,\delta)$ for any given $x$. We can thus apply again Lebesgue's dominated convergence theorem to the above integral.

Summing up equation (1) is proved.

Hans
  • 10,484
  • There are few typos I noticed. After (1), the long display $\partial u/\partial x = -t/(t^2+x)$, $(u,x,\delta)$ the function name is missing, Lastly, $(I_1(x+\delta)-I_1(x))/\delta$. With these fixed, it would be much easier to understand. – Sungjin Kim Oct 08 '17 at 07:46
  • @i707107: Thank you for spotting the typos. I have corrected these as well as others I have seen just now, and the equation numbers. Let me know when you spot any further mistakes. – Hans Oct 08 '17 at 09:04
  • Many thanks for this effort, and also thanks to i707107. I believe that this kind of post are very useful for students. To me seem a miracle of mathematical precision. Many thanks I am going to read your answer. –  Oct 08 '17 at 09:20
  • @Hans +1 good answer! – AngusTheMan Oct 08 '17 at 11:36
0

If I understand your question correctly, you are asking how we derive the solution to Airy function as a complex Riemannian integral?

If so, consider the Airy equation:

$$ (1) \qquad \frac{d^2 y}{dx^2} - xy = 0 $$

If we search for a contour integral solution in some unknown function, over some general contour $C$ in the complex plane.

$$ y(x) = \int _C f(z) e^{xz} dz $$

Inserting this into (1) we find:

$$ \int _C ( z^2 -x) f(z) e^{xz} dz = 0 $$

We then integrate the negative bit by parts:

$$ 0 = - f(z)e^{xz} \bigg|_C + \int _C \bigg( z^2f + \frac{df}{dz}\bigg) e^{xz} dz $$

If we choose $f(z) = e^{- z^3/3}$ then we must insist that at the end points of the contour:

$$ f(z)e^{xz} \bigg|_C = e^{-z^3/3 + zx} \bigg|_C = 0 $$

There are then three regions $j_{1,2,3}$ for a positive cosine where the contour vanishes in the complex plane giving us three functions; one for each region.

$$ f_j = \frac{1}{2\pi} i \int _{C_j} e^{- z^3/3+ xz} dz $$

We define the Airy function as the above solution evaluated in one of these regions. The function is expressed as:

$$ \operatorname{Ai}(x) = \frac 1 \pi \int ^\infty_0 \cos ( z^3/ + xz) dz = \frac {1}{2\pi} \int ^\infty _0 e^{i(z^3/3 _xz) }dz + \frac {1}{2\pi} \int ^\infty _ 0 e^{i(z^3/3 _xz)} dz $$

Using Jordan's lemma rotate the integral anti-clockwise in the first and clockwise in the second integrand. This will result in:

$$ \operatorname{Ai}(x) = \frac 1{2\pi} \int ^\infty _0 e^{ i (z^3/3 e^{3i\theta} + xze^{i\theta})} + \frac 1{2\pi} \int ^\infty _0 e^{- i (z^3/3 e^{-3i\theta} + xze^{-i\theta})} e^{-i\theta}dz $$

We then tidy them exponents to converge:

$$ \operatorname{Ai}(x) = \frac 1 \pi \Re \bigg( \int ^\infty_ 0 e^{\frac{z^3}{3} ( - \sin(3\theta) + i \cos (3\theta)) + xz(-\sin (\theta) + i \cos(\theta)) + i\theta} dz\bigg) $$

Now we use the following integral assured of the convergence:

$$ \int ^\infty _0 z^\lambda \cos (\frac {z^3}{3} + xz)dz = \frac 1 3 \int ^\infty_0 z^{\frac{(\lambda - 2)}{3}} \cos( \frac 1 3 z + x z^{\frac 1 3})dz $$

In specific response to your question, differentiating under the integral sign:

$$ \frac{d}{dx}\int ^b_a f(x,z) dz = \int ^b_a \frac{\partial f(x,z)}{\partial x} dz $$

Cannot be used since we are not certain that the R.H.S exists. That is why we perform the rotation to assure us of an exponentially convergent integrand.

  • I do not understand how this answers OP's question about differentiation under integral sign. – Sungjin Kim Oct 06 '17 at 22:41
  • This doesn't answer the question - at all. – Mark Viola Oct 06 '17 at 22:42
  • @AngusTheMan It was extraordinarily explicit that the question regarded the justification for differentiating under the integral. – Mark Viola Oct 06 '17 at 22:50
  • @ AngusTheMan The OP probably didn't get we can change of variable $\tau = t^3$ to obtain something whose convergence is obvious. @MarkViola I'd say instead the question is a nonsense (trivial question about a non-trivial function) – reuns Oct 06 '17 at 23:15
  • OP is asking for the justification of the integral representation of Ai' or the derivative of Ai, not of Ai itself. You are showing the derivation of the latter. It is not what the OP is asking for. – Hans Oct 06 '17 at 23:47
  • It’s good practice to separate entities that consist of more than one symbol with \,. Also, you can make custom upright operators in MathJax like $\operatorname{Ai}(z)$ by using the markdown \operatorname{Ai}(z). – gen-ℤ ready to perish Oct 06 '17 at 23:49
  • In case you’re interested in learned lots more tricks, consider browsing through the MathJax tutorial. – gen-ℤ ready to perish Oct 06 '17 at 23:50
  • Many thaks for your attention and contribution. I am interested in the answer that provide us a explanation if we can get the identity using differentiation under the integral sign. Any case your contribution seems to me valuable. –  Oct 07 '17 at 10:00
  • @user243301 You are welcome, I hope it helped! – AngusTheMan Oct 07 '17 at 10:17