5
$\begingroup$

As presented in the question's title, I wish to find a function $\lambda(\cdot): [0, \infty) \to [0, \infty)$ which satisfies the integral equation:

\begin{equation} \frac{1}{\lambda(y)} = c_1 \int_0^\infty \lambda(x) \exp(-c_2 y x) \, dx \end{equation}

where $c_1$ and $c_2$ are positive constants.

Unfortunately, I have no clear idea about how to systematically tackle this question. Any help is greatly appreciated!

$\endgroup$
  • 1
    $\begingroup$ You can easily get rid of $c_1$ and $c_2$. Just an idea. $\endgroup$ – iamvegan Oct 13 '16 at 14:30
  • $\begingroup$ I agree about getting rid of $c_1$. As far as $c_2$ is concerned, I don't see a way; for instance any substitution involving $c_2 x$ inevitably places $c_2$ into the argument of the integrated $\lambda$. $\endgroup$ – R.G. Oct 13 '16 at 14:35
3
$\begingroup$

Let me show that no such function exists. We argue this by contradiction. Assume that there is a function $\lambda : [0, \infty) \to [0, \infty)$ satisfying

$$ \frac{1}{\lambda(s)} = c_1 \int_{0}^{\infty} \lambda(x) e^{-c_2 sx} \, dx \quad \forall s \geq 0 \tag{*}$$

with the convention that $1/0 = \infty$. Then

Step 1. In this step, we normalize $\lambda$ and reveal some useful facts on it.

  • Since the right-hand side of $\text{(*)}$ is decreasing, $\lambda$ is increasing.

  • Since the left-hand side of $\text{(*)}$ is always positive (or possibly infinite), $\lambda$ cannot be identically zero.

  • From these two properties, $\alpha$ defined by $$ \alpha := \inf \{ s > 0 : \lambda(s) > 0 \} $$ Is a non-zero real number such that $\lambda(s) = 0$ for all $s \in [0, \alpha)$. Moreover, if $\alpha > 0$ then the modified version $\tilde{\lambda}(s) = e^{-c_2 \alpha s}\lambda(s+\alpha)$ satisfies \begin{align*} \frac{1}{\tilde{\lambda}(s)} = \frac{e^{c_2 \alpha s}}{\lambda(s+\alpha)} &= c_1 e^{c_2 \alpha s} \int_{\alpha}^{\infty} \lambda(x) e^{-c_2 (s+\alpha) x} \, dx \\ &= c_1 e^{c_2 \alpha s} \int_{0}^{\infty} \lambda(x+\alpha) e^{-c_2 (s+\alpha)(x+\alpha)} \, dx \\ &= c_1 e^{-c_2 \alpha^2} \int_{0}^{\infty} \tilde{\lambda}(x) e^{-c_2 sx} \, dx. \end{align*} So by changing the value of $c_1$ if needed, we may assume that $\alpha = 0$. Then by @guestDiego's computation, we may further assume that $c_1 = c_2 = 1$ and we do so.

  • If $\lambda(0) > 0$, then $$ \infty > \frac{1}{\lambda(0)} = \int_{0}^{\infty} \lambda(x) \, dx \geq \int_{0}^{\infty} \lambda(0) \, dx = \infty $$ and we get a contradiction. Thus $\lambda(0) = 0$.

  • From the standard theory of Laplace transform, if $f \geq 0$ is measurable and $$\mathcal{L}\{f\}(s) := \int_{0}^{\infty} f(x)e^{-sx} \, dx$$ Is finite for all $s > 0$, then $\mathcal{L}\{f\}(s)$ converges for $\Re(s) > 0$ and defines an analytic function on the same region. Moreover, differentiation can be computed by using Leibniz's integral rule: $$ \frac{d^n}{ds^n} \mathcal{L}\{f\}(s) = (-1)^n \int_{0}^{\infty} x^n f(x) e^{-sx} \, dx. $$

Step 2. Now we are ready to establish a contradiction. First, we have $\lambda'(s) \geq 0$ because $\lambda$ is increasing. Then by the Tonelli's theorem,

\begin{align*} \frac{s}{\lambda(s)} &= \int_{0}^{\infty} \lambda(t) s e^{-st} \, dt \\ &= \int_{0}^{\infty} \bigg( \int_{0}^{t} \lambda'(x) \, dx \bigg) s e^{-st} \, dt \\ &= \int_{0}^{\infty} \lambda'(x) \bigg( \int_{x}^{\infty} s e^{-st} \, dt \bigg) dx \\ &= \int_{0}^{\infty} \lambda'(x) e^{-sx} \, dx. \end{align*}

Taking log-differentiation to both sides, we get

$$ \frac{\lambda'(s)}{\lambda(s)} = \frac{\int_{0}^{\infty} x \lambda'(x) e^{-sx} \, dx}{\int_{0}^{\infty} \lambda'(x) e^{-sx} \, dx} + \frac{1}{s}. \tag{1}$$

This is our key ingredient toward a contradiction. Using this, we inductively prove that

Claim. For any $n = 1, 2, 3, \cdots$ and $s > 0$, we have $\frac{\lambda'(s)}{\lambda(s)} \geq \frac{n}{s}$.

The base case $n = 1$ is straightforward from $\text{(1)}$ since the ratio between two integra in the RHS of $\text{(1)}$ is non-negative. Next, assuming the claim for $n$, we have

$$ \frac{\lambda'(s)}{\lambda(s)} \geq \frac{\int_{0}^{\infty} n \lambda(x) e^{-sx} \, dx}{\int_{0}^{\infty} \lambda'(x) e^{-sx} \, dx} + \frac{1}{s} = \frac{n/\lambda(s)}{s/\lambda(s)} + \frac{1}{s} = \frac{n+1}{s}. $$

Then the claim follows from mathematical induction.

Now the contradiction is obvious: $\lambda'(s)/\lambda(s)$ is finite for $s > 0$ while $n/s$ can be arbitrary large! Therefore no such $\lambda$ can exist.

$\endgroup$
  • $\begingroup$ Very nice argument! May I ask what the intuition for showing $\frac{s}{\lambda(s)} = \int_0^\infty \lambda'(x) \exp(-s*x) \, dx$ is? $\endgroup$ – R.G. Oct 18 '16 at 12:37
  • $\begingroup$ @R.G., it is essentially integration by parts. But you can avoid introducing some nebulous intermediate term if you use Fubini's theorem. $\endgroup$ – Sangchul Lee Oct 18 '16 at 13:03
  • $\begingroup$ I'm sorry for the being unclear with my question. It is not the calculation I want to know about but rather why you thought that evaluating this will help you later on. $\endgroup$ – R.G. Oct 18 '16 at 13:13
  • $\begingroup$ @R.G. Now I see what you asked. I had no clear intuition, but somehow I believed that higher derivatives of $\lambda$ would reveal erratic behavior which we may employ to derive a contradiction. (To be precise, I believed that $\lambda^{(n)}$ will change sign for some $n$ so that the Laplace transform of it becomes negative, which is impossible in view of $s^n / \lambda(s) = \mathcal{L}\{\lambda^{(n)}\}(s)$.) So it was natural for me to consider Laplace transform of derivatives of $\lambda$. It was totally unclear to me where it will lead me, but very fortunately $\text{(1)}$ was enough! $\endgroup$ – Sangchul Lee Oct 18 '16 at 13:21
  • $\begingroup$ Thanks for this insight! I hope to reach a similar level of mathematical maturity one day. $\endgroup$ – R.G. Oct 19 '16 at 7:51
2
$\begingroup$

I make explicit the hint of iamavegan.

If $\lambda_1(y)=\sqrt{c_1}\lambda(y)$, then $$ \frac{1}{\lambda_1(y)} = \int_0^\infty \lambda_1(x) \exp(-c_2 y x) \, dx= \int_0^\infty \frac{\lambda_1(s/\sqrt{c_2})}{\sqrt{c_2}} \exp(-\sqrt{c_2} y s) \, ds $$ Define $$ \lambda_2(z)=\frac{\lambda_1(z/\sqrt{c_2})}{c_2^{1/4}}. $$ Then, with $z=y\sqrt{c_2} $, $$ \frac{1}{\lambda_2(z)} = \int_0^\infty \lambda_2(s) \exp(-sz) \, ds. $$ and $$ \lambda(y)=c_1^{-1/2}c_2^{1/4}\lambda_2(y c_2^{1/2}) $$

$\endgroup$
  • 1
    $\begingroup$ Thank you very much for this elaboration! $\endgroup$ – R.G. Oct 13 '16 at 15:08
1
$\begingroup$

HINT: iterate the $\lambda$ into the integral. \begin{align*} \frac1{\lambda(y)} &=c_1\int_0^{+\infty}\lambda(x)e^{-c_2xy}\,dx\\ &=c_1\int_0^{+\infty}\frac{e^{-c_2xy}}{c_1\int_0^{+\infty}\lambda(z)e^{-c_2zx}\,dz\\}\,dx\\ &=\int_0^{+\infty}\frac{e^{-c_2xy}}{\int_0^{+\infty}\lambda(z)e^{-c_2zx}\,dz\\}\,dx\\ \end{align*} go on from here

$\endgroup$
  • $\begingroup$ Thank you for this hint. I'll see where this will take me. $\endgroup$ – R.G. Oct 13 '16 at 15:09

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.