1
$\begingroup$

I have the following equation that I would like to solve for $y$: $$\alpha=\sum_{i=1}^{n}\exp(-y\beta_i)$$ $\alpha$, $y$ and $\beta_i$ are all positive reals (the $\beta_i$'s are independent random numbers), and $n$ is a large positive integer. I'm unable to find a way to isolate $y$. Is this equation solvable?

Probability distribution of $\beta_i$:

$\beta_i$ is defined as $\beta_i=\sqrt{2\omega_i}$, where $\omega_i$ assumes a uniform probability distribution in some interval $[\omega_a,\omega_b]$, where $0\leq\omega_a<\omega_b$.

$\endgroup$
  • 1
    $\begingroup$ No, in general this equation is not solvable for y. Espacially when $n$ is a large number. $\endgroup$ – Cornman Jun 17 at 13:07
  • $\begingroup$ As @Cornman said, you won't be able to get rid of that sum of exponentials $\endgroup$ – David Jun 17 at 13:18
  • 1
    $\begingroup$ If $\beta_i$ were all rational numbers, then this will become equivalent to solving a polynomial of a potentially enormous degree (probably much larger than $n$). Naturally, there is no formula for solving polynomials of degree $5$ or more, so you might be out of luck. $\endgroup$ – Theo Bendit Jun 17 at 13:20
  • $\begingroup$ hmm.. ok, thank you $\endgroup$ – Tofi Jun 17 at 13:25
  • 3
    $\begingroup$ It is worth noting that there is a unique real solution $y$ for any values of $\alpha$ and the $\beta_i$. Given $\alpha$ and the $\beta_i$ you can approximate $y$ with the help of a computer. An explicit expression for $y$ in terms of $\alpha$ and the $\beta_i$ is too much to ask, as the other comments explain. $\endgroup$ – Inactive - avoiding CoC Jun 17 at 13:29
1
$\begingroup$

Updated post

After publishing my original post the author of the OP has specified the PDF for the $\beta_i$, also hereby he has assumed that all $\beta_i$ have the same distribution (which was my main simplifying assumption).

Here I take up this definition and update my solution.

EDIT (18.06.19): the following through derivation yields the corrected result which is even simpler than the previous one.

The distribution function for $b$ is from the definition $w = \frac{1}{2}b^2$ and $f(w)$ flat between $w_a$ and $w_b$ given from $dw = b\; db$ after normalization

$$f(b) = \frac{2 b}{b_b^2-b_a^2} = \frac{b}{w_b-w_a}$$

Now we can do the integral $(2)$ which gives

$$\alpha = n \int_{b_a}^{b_b} f(b) e^{- y b} \, db \\ = \frac{2 n \left(e^{- y b_a } (y b_a+1)-e^{-y b_b} (y b_b +1)\right)}{y^2 \left(b_b^2-b_a^2\right)}\tag{4a} $$

We have obtained a transcendental equation for $y(\alpha)$. This is best solved in the inverted form $(4a)$ for $\alpha (y)$.

In the limit of a very narrow distribution ($w_a \to w_b (=w)$) we get from $(4a)$

$$\alpha = n e^{-b_b y}\tag{4b} $$

which can be inverted to give the explicit solution

$$y = \frac{1}{b_b}\log(\frac{\alpha}{n})\tag{4c}$$

My original post

Making the simplifying assumptions that the $\beta_i$ have the same PDF $f(\beta)$ we can take the expectation value of the equation over all $\beta_i$ and get

$$E(\alpha) = \alpha = \sum_{i=1}^n E(\exp(- y \beta_i))= n E(\exp(- y \beta))\tag{1}$$

More explicitly we have for a continuous variable $\beta$

$$E(\exp(- y \beta))=\int_{0}^\infty f(\beta) \exp(- y \beta))^\,d\beta\tag{2}$$

Assuming for simplicity that $\beta$ is exponentially distributed, i.e.

$$f(\beta) = a \exp(- a \beta )\tag{3}$$

the integral can be done, and we obtain

$$\alpha = n \frac{a}{a + y}\tag{4}$$

which can be solved explicitly for $y$ to give

$$y = a (\frac{n}{\alpha}-1)\tag{5}$$

$\endgroup$
  • $\begingroup$ Thank you, it looks like this equation is just too much! $\endgroup$ – Tofi Jun 17 at 23:42
  • $\begingroup$ @ Tofi Why "too much"? The specific problem is solved by equation $(4a)$ leaving just the standard problem of solving numerically a transcendental equation. $\endgroup$ – Dr. Wolfgang Hintze Jun 18 at 8:06
  • $\begingroup$ Yeah, but what I wanted was a nice analytical expression. By the way, the value I found for the PDF of $\beta$ is $\dfrac{\beta}{\omega_b-\omega_a}$, so is this the same as yours? $\endgroup$ – Tofi Jun 18 at 11:36

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.