Closed form for ∫∞0lnJμ(x)2+Yμ(x)2Jν(x)2+Yν(x)2dx\int_0^\infty\ln\frac{J_\mu(x)^2+Y_\mu(x)^2}{J_\nu(x)^2+Y_\nu(x)^2}\mathrm dx

Consider the following integral:
where Jμ(x) is the Bessel function of the first kind:
and Yμ(x) is the Bessel function of the second kind:
I was not able to rigorously establish a closed form for I(μ,ν), but based on numerical integration I made a conjecture:
Could you please help me to find out if this conjecture is true?

If the conjecture is true, then taking the derivative with respect to μ at μ=1 we get the following corollary:

As pointed by O.L., the conjecture is equivalent to
where H(1)ν(x)=Jν(x)+iYν(x) and H(2)ν(x)=Jν(x)iYν(x) are the Hankel functions of the first and second kind.


I will assume that ν is real, as in the formulation
of the question. Something similar may be true for
complex ν by a similar argument, but there would be an extra complication (perhaps
only notational) of working with ν and
its conjugate ¯ν instead of just ν.
The Hankel functions H(i)ν(z) are entire
except for a branch cut along the negative real axis.
We introduce the following notation:
for a function F(z) with a branch cut along the negative axis, we let F(x+)
and F(x) denote the limit of F(z) as zx from the region
Im(z)>0 and Im(z)<0 respectively.
There are identities as follows:
Eq. 1:A(1)ν(x+)=¯A(2)ν(x)=A(2)ν(x)=¯A(1)ν(x),
and the same equations hold for B. Note that
it's not true that
A(2)ν(x+)=A(2)ν(x), the lack of symmetry
here is related to the branch cut. This is an important point. The behavior of A(1)ν(z) is bad in the region near x, and correspondingly
A(2)ν(z) is bad near x+.
The Hankel function has nice asymptotic expansions for large z.
The ratio of such functions at arguments
ν differing by integers is particularly nice, because the complex phase cancels.
In particular, one has the following:
This holds outside the bad regions mentioned above. In particular,
it holds for A(1)ν(z) and B(1)ν(z) for z with argument
in [π+ϵ,π], and for A(2)ν(z) and B(2)ν(z) with argument
in [π,πϵ].
Let CR be the semi-circle
with centre 0 and radius R in the upper half plane, oriented anti-clockwise, and thought of as lying
above the branch cut in (,0]). Note that this is contained within the range where the asymptotic holds for A(1)ν(z),
and hence
The main term comes from the residue theorem
(applied to a half circle, hence the \pi i rather than 2 \pi i factor), and
the error term comes from the fact that the integral of O(z^{-2}) over a
half-circle of radius
R and circumference \pi R is O(R^{-1}).

We now use the following fact: H^{(1)}_{\nu}(z) has no zeros in the upper half plane.
I say that it is a fact, but I couldn't find a reference (Edit: proof of this fact included at the end of this answer). I proved it rigorously by an explicit contour integral
computation for various ranges of values of \nu, however. (Certainly, by the asymptotic expansion which is valid in the entire upper half plane, it follows that any such zeros, if they exist, must be within some small radius, which one can eliminate by computing (2 \pi i)^{-1} \oint d \log(f).)
By the residue theorem (taking C above to be the circle in the upper half plane), we get,
for any holomorphic integrand,
0 = \oint = \int_{C} + \int^{R}_{-R},
and hence
\lim_{R \rightarrow \infty} \int^{R}_{-R} z \left(\frac{H^{(1)}_{\nu-1}(z)}{ H^{(1)}_{\nu}(z)} - i \right)
- \frac{(2 \nu - 1)}{2} dz
= - \frac{\pi (4 \nu^2 - 1)}{8} .

Note that the integrand has order O(1/z), so one really has to take the integrand
from -R to R and then take the limit for this to make sense.
One may apply the same analysis to H^{(2)}_{\nu}, except now the zero free region
of H^{(2)}_{\nu}
is in the lower half plane --- this follows by symmetry from the
identity H^{(1)}_{\nu}(\overline{z}) = \overline{H^{(2)}_{\nu}(z)}, noting
that we are once again in the correct region as far as asymptotics goes.
Hence we deduce that
\lim_{R \rightarrow \infty} \int^{-R}_{R} z \left(\frac{H^{(2)}_{\nu+1}(z)}{ H^{(2)}_{\nu}(z)} - i \right)
- \frac{(2 \nu + 1)}{2} dz = - \frac{\pi (4 \nu^2 - 1)}{8} .

Note that the direction of the integral has changed, for orientation reasons.

Warning! There's also another difference between this and the previous integral. The first integral was above the branch cut and this integral is below
the branch cut. However, in the first case,
we were integrating values of the form A^{(1)}_{\nu}(x^+), which was the good value (in the
sense that it was related to three other values by symmetry in equation 1), and
here we are integrating B^{(2)}_{\nu}(x^{-}), which also is related to three
other values by the same equations.

the order of the second integrand and then subtracting the results, we get
\lim_{R \rightarrow \infty} \int^{R}_{-R} 1 + z
\left(\frac{H^{(1)}_{\nu-1}(z)}{ H^{(1)}_{\nu}(z)} - \frac{H^{(2)}_{\nu+1}(z)}{ H^{(2)}_{\nu}(z)} \right)
dz = - 2 \cdot \frac{\pi (4 \nu^2 - 1)}{8}.

We now make two observations: the integrand is now O(z^{-2}) for large z, and hence
it exists as a definite integral. Moreover, the integrand is even. In light of the warning,
we should really specify that the
integrand for values x \in (-\infty,0] is precisely:
1 + x \left(A^{(1)}_{\nu}(x^{+}) - B^{(2)}_{\nu}(x^{-})\right).
(One should check this is the correct function to make the integrand
even.) We deduce that

- \int^{\infty}_{0} 1 + z \left(\frac{H^{(1)}_{\nu-1}(z)}{ H^{(1)}_{\nu}(z)} -
\frac{H^{(2)}_{\nu+1}(z)}{ H^{(2)}_{\nu}(z)} \right) dz = \frac{\pi (4 \nu^2 - 1)}{8}.

I(\nu) = \int^{\infty}_{0} \log \frac{\pi x H^{(1)}_{\nu}(x) H^{(2)}_{\nu}(x)}{2} \cdot dx.
Integrating by parts, and being a little bit careful about what happens at 0, and
expressing the derivatives on Hankel functions in terms of Hankel functions of other arguments, we find that
I(\nu) = - \int^{\infty}_{0} 1 + x \left(\frac{H^{(1)}_{\nu-1}(x)}{ H^{(1)}_{\nu}(x)} -
\frac{H^{(2)}_{\nu+1}(x)}{ H^{(2)}_{\nu}(x)} \right) dx = \frac{\pi (4 \nu^2 - 1)}{8}.

As noted in the comments, this was the identity to be proved.
Alternatively, integration by parts also shows that
\frac{\pi (\mu^2 - \nu^2)}{2} = I(\mu) - I(\nu)
= \int^{\infty}_{0} \log \frac{H^{(1)}_{\mu}(x) H^{(2)}_{\mu}(x)}{H^{(1)}_{\nu}(x) H^{(2)}_{\nu}(x)} \cdot dx,

and hence
\int^{\infty}_{0} \log \frac{J_{\mu}(x)^2 + Y_{\mu}(x)^2}{J_{\nu}(x)^2 + Y_{\nu}(x)^2} \cdot
dx = \frac{\pi (\mu^2 - \nu^2)}{2}.

Edit: Proof that H^{(1)}_{\nu}(z) has no zeros in the upper half plane for real \nu > 0.

I realized that the proof can be completed in a similar manner. Let
G^{(1)}_{\nu}(z) = d \log H^{(1)}_{\nu}(z)
=\frac{1}{2} \left(A^{(1)}_{\nu}(z) - B^{(1)}_{\nu}(z)\right).

Then we have an asymptotic formula, as before:
A^{(1)}_{\nu}(z) \sim
i \left(1 - \frac{(4 \nu^2 - 1)}{8} \cdot \frac{1}{z^2} + \ldots \right)
+ \frac{(2 \nu - 1)}{2} \cdot \frac{1}{z} + O(z^{-3}),

B^{(1)}_{\nu}(z) \sim
- i \left(1 - \frac{(4 \nu^2 - 1)}{8} \cdot \frac{1}{z^2} + \ldots \right)
+ \frac{(2 \nu + 1)}{2} \cdot \frac{1}{z} + O(z^{-3}),

and thus
G^{(1)}_{\nu}(z) \sim i \left(1 - \frac{(4 \nu^2 - 1)}{8} \cdot \frac{1}{z^2} + \ldots \right) - \frac{1}{2z} + O(z^{-3}),
This is valid for z with argument in [-\pi + \epsilon,\pi], so in particular is valid in the upper half plane.
If C_R denotes the circle in the upper half plane, we find that:
\lim_{R \rightarrow \infty} \int_{C_R} (G^{(1)}_{\nu}(z) - i) dz = - \frac{\pi i}{2},
because it is O(R^{-1}) plus the contribution from the 1/(2z) term.
Let \Omega_R denote the boundary of the region bounded by by the interval [-R,R] and C_R.
The function H^{(1)}_{\nu}(z) and thus H^{(1)}_{\nu}(z) e^{-iz} is holomorphic in \Omega_R
away from z = 0. At zero (and \nu > 0) we have an asymptotic
H^{(1)}_{\nu}(z) \sim - i \cdot \frac{\Gamma(\nu)}{\pi} \left(\frac{2}{z}\right)^{\nu}.

It follows that the number of zeros
in the upper half plane is given, accounting for the singularity at 0 (modified by a factor
of two since this integral only accounts for half of the singularity) by
\frac{\nu}{2} + \lim_{R \rightarrow \infty} \frac{1}{2 \pi i} \oint_{\Omega_R} d \log (H^{(1)}_{\nu}(z) e^{-iz}) dz
= \frac{\nu}{2} + \frac{1}{2 \pi i} \left( \lim_{R \rightarrow \infty} \int_{C_R} (G^{(1)}_{\nu}(z) - i) dz +
\lim_{R \rightarrow \infty} \int_{-R}^{R} (G^{(1)}_{\nu}(z) - i) dz\right).

We computed the first integral above, hence it suffices to compute:
\frac{\nu}{2} -\frac{1}{4} + \frac{1}{2 \pi i} \lim_{R \rightarrow \infty} \int_{-R}^{R} (G^{(1)}_{\nu}(z) - i) dz.
We may write this as the integral
\frac{\nu}{2} -\frac{1}{4} + \frac{1}{2 \pi i} \int_{0}^{\infty} \left( G^{(1)}_{\nu}(z) + G^{(1)}_{\nu}(-z) - 2 i \right) dz.
This expression evaluates to an integer, which is the number of zeros of H^{(1)}_{\nu}(z)
in the upper half plane.
In particular, evaluating this integral numerically for a random value (say \nu = \pi) shows
that, for this value, it is equal to zero. Now suppose that we vary \nu. Since this integral
evaluates to an integer, to complete the proof, it suffices to show that it varies continuously.
For the integral over [R,\infty) this is clear for large enough R from the asymptotic formula. For small values, it suffices to show that H^{(1)}_{\nu}(z) doesn't have any zeros on
the real line [0,R], since otherwise the integrand is continuous in \nu, and the continuity
of the integral is clear. Now on the real axis, we have, by definition,
H^{(1)}_{\nu}(z) = J_{\nu}(z) + i \cdot Y_{\nu}(z).
Since \nu is real, it suffices to show that J_{\nu}(z) and Y_{\nu}(z)
do not have any simultaneous zeros. However, the zeros of these Bessel functions
are well known to interlace (see Watson, A treatise on the theory of Bessel functions),
and the result is established. Edit: Actually, there's a much easier proof that J_{\nu}(z) and Y_{\nu}(z)
do not have any common zeros --- they are linearly independent solutions to a second order ODE!

Source : Link , Question Author : Vladimir Reshetnikov , Answer Author :
3 revs

Leave a Comment