12
$\begingroup$

For my research, I have to solve many integrals of the Owen's T function. As such, I am having struggles in calculating the integral $$ \int_{0^{-}}^{t}d{s} \ e^{(t+s)} \ \operatorname{T}\left(\frac{x}{\sqrt{2t}}, \sqrt{\frac{s}{2t-s}}\right). $$ I asked a similar question in [1] for the integral $$ \int_{0^{-}}^{t}d{s} \ \operatorname{T}\left(\frac{x}{\sqrt{2t}}, \sqrt{\frac{s}{2t-s}}\right), $$ which was beautifully answered. Nevertheless, if one follows the same procedure the exponential complicates things. What I have done until know is following [1] which yields \begin{align} f(x,t)=\int_{0^{-}}^{t}d{s} \ e^{(t+s)} \ T\left(\frac{x}{\sqrt{2t}}, \sqrt{\frac{s}{2t-s}}\right)=t\int_{0}^{1} d{u} \ e^{t(1+u)} \ T\left(\sqrt{2}y, \sqrt{\frac{u}{2-u}}\right) \end{align} where we performed the change of variable $s=tu$ and took $y=\frac{x}{2\sqrt{t}}$. Substituting the definition for Owen's T function and using Tonelli theorem gives \begin{align} f(x,t)&= \frac{1}{2\pi}t \int_{0}^{1}d{u}\int_{0}^{\sqrt{\frac{u}{2-u}}}d{v} \ e^{t(1+u)} \ \frac{e^{-y^2(1+v^2)}}{1+v^2}\\ &=\frac{1}{2\pi}t \int_{0}^{1}d{v} \ \int_{\frac{2v^2}{1+v^2}}^{1}d{u} \ e^{t(1+u)} \ \frac{e^{-y^2(1+v^2)}}{1+v^2}\\ &= \frac{1}{2\pi}t \int_{0}^{1} d{v} \ \frac{e^{-y^2(1+v^2)}}{1+v^2}\left(\frac{e^{2 t}}{t}-\frac{e^{\frac{2 t v^2}{v^2+1}+t}}{t}\right)\\ &= \frac{1}{2\pi} e^{2 t}\int_{0}^{1} d{v} \ \frac{e^{-y^2(1+v^2)}}{1+v^2}-\frac{1}{2\pi}e^{t} \int_{0}^{1} d{v} \ \frac{e^{-y^2(1+v^2)}}{1+v^2}e^{\frac{2 t v^2}{v^2+1}}\\ &= T\left(\tfrac{x}{\sqrt{2t}}, 1\right)e^{2 t}+\frac{1}{2\pi} e^{t} \color{blue}{\int_{0}^{1} d{v} \frac{e^{-y^2(1+v^2)}}{1+v^2}e^{\frac{2 t v^2}{v^2+1}}}\\ \end{align} The Owen T term, i.e. the first term, can be written in terms of Error function. However, I am struggling in solving the integral in the second term. I have tried integrate by part, but this just yield a more complicated expression. Someone who knows a way forward?

Also I checked above calculation numerically and it indeed holds up!

In[270]:= ClearAll["Global`*"];
int1[x_, t_]:=OwenT[x/Sqrt[2t],1]*Exp[2t]+1/(2\[Pi])*Exp[t]*NIntegrate[ Exp[-(x/(2Sqrt[t]))^2*(1+v^2)]/(1+v^2)*Exp[(2tv^2)/(1+v^2)], {v,0,1},WorkingPrecision->7] ;
int2[x_, t_]:=NIntegrate[Exp[(t+s)]*OwenT[x/Sqrt[2t],Sqrt[s/(2t-s)]], {s,0,t},WorkingPrecision->7] ;
{x,t}=RandomReal[{0,1},2,WorkingPrecision->50]; 
int1[x, t]
int2[x, t]

Out[274]= 0.01602865
Out[275]= 0.01603027
$\endgroup$
3
  • 1
    $\begingroup$ Does pinging when someone else has not talked before work? The only this I could think of is combining the $e^y$ powers and using the maclaurin series for $e^y$. Then maybe binomial, or related, series like here. $\endgroup$ Aug 5 at 18:11
  • $\begingroup$ Thanks you for the suggestion. I will explore it and let you know. $\endgroup$
    – drandran12
    Aug 6 at 6:03
  • 3
    $\begingroup$ I'm afraid I am not too optimistic about this problem. For small $t$ you could try expanding $\mathrm{e}^{-t(1+u)}$. The $t^0$ term then corresponds to the old question, but the integral for the first order term is already quite ugly. $\endgroup$ Aug 6 at 14:50
4
+100
$\begingroup$

The key difference between this integral and the one you linked is the presence of the reciprocal term $(v^2+1)^{-1}$ inside the exponential. Without it, we can easily rely on error-type functions such as Owen's $T$, but as we will see later its presence leads to incomplete gamma-/Bessel-type functions at the very least.

The intuitive substitution $x=v^2+1$ leads to $$\int_0^1\frac{e^{-y^2(v^2+1)+2tv^2/(v^2+1)}}{v^2+1}\,dv=\frac{e^{2t}}2\int_1^2\frac{e^{-y^2x-2t/x}}{x\sqrt{x-1}}\,dx.$$ The issue with this integrand is the $\sqrt{x-1}$ term; without it, we would have $$\int_1^2\frac{e^{-y^2x-2t/x}}x\,dx=\frac{e^{2t}}2\left(\Gamma(0,y^2;2ty^2)-\Gamma(0,2y^2;2ty^2)\right)$$ where $\Gamma(\alpha,x;b):=\int_x^\infty t^{\alpha-1}e^{-t-b/t}\,dt$ is quite a common function in statistical physics (Harris, 2008). In our case, we could try to exploit its inverse Laplace transform $${\cal L}^{-1}\left[\frac1{\sqrt{s-1}}\right]=\frac{e^t}{\sqrt{\pi t}}$$ to write the integral as \begin{align}\int_1^2\frac{e^{-y^2x-2t/x}}{x\sqrt{x-1}}\,dx&=\int_1^2\int_0^\infty\frac{e^{-y^2x-2t/x}}x\cdot\frac{e^{-xu}e^u}{\sqrt{\pi u}}\,du\,dx\\&=\int_0^\infty\frac{e^u}{\sqrt{\pi u}}\int_1^2\frac{e^{-(y^2+u)x-2t/x}}x\,dx\,du\\&=\int_0^\infty\frac{e^u(\Lambda_2(u)-\Lambda_1(u))}{\sqrt{\pi u}}\,du\end{align} where $\Lambda_k(u):=\Gamma(0,k(y^2+u);2t(y^2+u))$. This puts the integral in a $(0,\infty)$-form where there are many more integral tricks, but the nature of the integrand makes the existence of a closed form rather hard to believe.

$\endgroup$
1
  • 1
    $\begingroup$ Wow, this is a great solution! I think OP will have to settle for this if he wants a good answer. $\endgroup$ Aug 31 at 16:02
2
$\begingroup$

Let $I$ be the following:

$$I=\int_{0}^{1} \frac{e^{-y^2(1+v^2)}}{1+v^2}e^{\frac{2 t v^2}{1+v^2}} \textrm{d}v$$

Now use the substitution $v=\tan{u}$:

$$I=\int_{0}^{\frac{\pi}{4}} \frac{e^{-y^2(1+\tan^2u)}}{1+\tan^2u}e^{\frac{2 t \tan^2u}{1+\tan^2u}} \sec^2(u) \textrm{d}u$$

If you are familiar with simple trig identities, you may also notice the following:

$$I=\int_{0}^{\frac{\pi}{4}} \frac{e^{-y^2\sec^2u}}{\sec^2u}e^{\frac{2 t \tan^2u}{\sec^2u}} \sec^2(u) \textrm{d}u$$

$$I=\int_{0}^{\frac{\pi}{4}} e^{-y^2\sec^2u}e^{\frac{2 t \tan^2u}{\sec^2u}} \textrm{d}u$$

$$I=\int_{0}^{\frac{\pi}{4}} e^{-y^2\sec^2u}e^{2 t \tan^2(u)\cos^2u} \textrm{d}u$$

$$I=\int_{0}^{\frac{\pi}{4}} e^{-y^2\sec^2u}e^{2 t \sin^2(u)} \textrm{d}u$$

$$I=\int_{0}^{\frac{\pi}{4}} e^{2 t \sin^2(u)-y^2\sec^2u} \textrm{d}u$$

$$I=\int_{0}^{\frac{\pi}{4}} e^{a\sin^2(v)-b\sec^2v} \textrm{d}v$$

Unfortunately, I don't know if I could do any better than that. I'll leave an update if I do find something, though.

Edit 1: One thing I have noticed about the function $f(x)=e^{a\sin^2(x)-b\sec^2x}$ is that it is a periodic function and that $f(x)=f(\pi-x)$.

Edit 2: I think I'm on to something:

$$I=\int_{0}^{1} \frac{e^{-y^2(1+v^2)}}{1+v^2}e^{\frac{2 t v^2}{1+v^2}} \textrm{d}v = 2\pi\int_{0}^{1} \dfrac{\textrm{d}}{\textrm{d} v}\left(T(y\sqrt{2},v)\right) e^{\frac{2 t v^2}{1+v^2}} \textrm{d}v$$

If we shorten $\dfrac{\textrm{d}}{\textrm{d} v}T(y\sqrt{2},v)$ to just $\dfrac{\textrm{d}T}{\textrm{d}v}$, we can convert the integral to the following:

$$I = 2\pi\int_{0}^{1} \dfrac{\textrm{d}T}{\textrm{d}v}e^{\frac{2 t v^2}{1+v^2}} \textrm{d}v$$

$$I = 2\pi\int_{0}^{1} e^{\frac{2 t v^2}{1+v^2}} \textrm{d}T$$

Then use IBP:

$$ I = 2\pi T(y\sqrt{2},1)e^{t} - 2\pi\int_{0}^{1} T(y\sqrt{2},v)\dfrac{4tv}{\left(1+v^2\right)^2}e^{\frac{2 t v^2}{1+v^2}} \textrm{d}v $$

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.