2
$\begingroup$

I am trying to solve the equation

$$ \sum_i^n \frac{a_i}{\lambda - \lambda_i} = \lambda $$

for some real constants $a_i, \lambda_i$, with $\lambda_1 > \lambda_2 > \lambda_3 > ... $ I have the hunch that I can approximate one of its solutions by taking only the first term and solving

$$ \frac{a_1}{\lambda - \lambda_1} = \lambda $$

but I'm having trouble formalizing this step. Under what conditions can I do so and how can I justify this formally? For example, if $\lambda_1 \gg \lambda_2 > \lambda_3 > ...$ then, whenever $\lambda \approx \lambda_1$, the original equation is approximately

$$ \frac{a_1}{\lambda - \lambda_1} + \frac{\sum_{i>1} a_i}{\lambda_1} = \lambda $$

Also, can I find other solutions to the original equation by focusing on other terms? Concretely, what can the solution to the following tell me?

$$ \frac{a_2}{\lambda - \lambda_2} = \lambda. $$

EDIT: I'm especially interested in the case where solving $a_1/(\lambda - \lambda_1) $ is a "good enough" approximation, as computing every $\lambda_1$ might be expensive.

$\endgroup$
  • $\begingroup$ Do you know the behavior of $a_i$? If $a_i$ was very small (went to zero) after one or two terms then you can neglect them. But since the $\lambda_i$ are just getting smaller, the component terms of the sum are simply approaching $\dfrac{a_i}{\lambda}$. $\endgroup$ – JacobCheverie Apr 15 at 15:04
  • $\begingroup$ Can we assume $a_i, \lambda_i$ are positive? $\endgroup$ – Tom Chen Apr 15 at 15:06
  • $\begingroup$ @TomChen Yes, although I'm interested in the general case too. $\endgroup$ – Leo Apr 15 at 15:09
  • $\begingroup$ @JacobCheverie this is my intuition too though I'm not sure how large the difference $\lambda_1 - \lambda_2$ must be for me to make that approximation. And in that case, what's the size of the error term. $\endgroup$ – Leo Apr 15 at 15:12
3
$\begingroup$

Something I've found to be helpful. We may re-express \begin{align*} \sum_{i=1}^{n}\left\{\frac{a_i}{\lambda - \lambda_i} - \frac{\lambda}{n}\right\} = 0 \end{align*} Solving each summand individually (for positive solutions of $\lambda$): \begin{align*} f_i(\lambda) \overset{\text{def}}{=}\frac{a_i}{\lambda - \lambda_i} - \frac{\lambda}{n} = 0 \implies \lambda_i^* = \frac{1}{2}\left(\sqrt{4a_in + \lambda_i^2} + \lambda_i\right) \end{align*} Next, approximate $f_i(\lambda)$ with a Taylor series around $\lambda = \lambda_i^*$, \begin{align*} f_i(\lambda) &\approx a_i\left\{\frac{1}{\lambda_i^* - \lambda_i} - \frac{\lambda - \lambda_i^*}{\lambda_i^*-\lambda_i} + \frac{(\lambda - \lambda_i^*)^2}{(\lambda_i^*-\lambda_i)^3}\right\} \\ &=A_i\lambda^2 + B_i\lambda + C_i \end{align*} for constants $A_i, B_i, C_i$ after collecting the terms. Finally, you then solve \begin{align*} 0 = \sum_{i=1}^{n}f_i(\lambda) \approx \sum_{i=1}^{n}\{A_i \lambda^2 + B_i \lambda + C_i\} = A\lambda^2 + B\lambda + C \end{align*} In reality, Newton-Raphson could get a very close approximation within 4-5 iterations, but with this, you can get a closed-form approximation.

$\endgroup$
1
$\begingroup$

Let $f(\lambda) = \sum_i a_i/(\lambda - \lambda_i)$. Suppose $a_1 > 0$. Then $\lim_{\lambda \to \lambda_1+} f(\lambda) = +\infty$, so $f(\lambda) > \lambda$ for $\lambda$ near $\lambda_1$ and slightly greater. On the other hand, $f(\lambda) \to 0$ as $\lambda \to +\infty$, so $f(\lambda) < \lambda$ when $\lambda$ is sufficiently large. By the Intermediate Value Theorem, there exists a solution to $f(\lambda) = \lambda$ somewhere in the interval $(\lambda_1, +\infty)$.

EDIT: You may be able to use some estimates to show that there is a solution near a solution to $a_1/(\lambda-\lambda_1) = \lambda$. Write $f(\lambda) = \frac{a_1}{\lambda - \lambda_1} + g(\lambda)$ where $$ g(\lambda) = \sum_{i=2}^n \frac{a_i}{\lambda - \lambda_i}$$ Suppose $p$ is a solution to $a_1/(\lambda - \lambda_1) = \lambda$, and in some interval $[p-\delta, p+\delta]$ around $p$ we have $f'(\lambda) - 1 < -m < 0$ and $|g(\lambda)| < \epsilon$. The $f(p+\delta) - (p+\delta) < -m \delta + \epsilon$ while $f(p-\delta) - (p-\delta) > m \delta - \epsilon$. If $m \delta - \epsilon > 0$, there must be a solution to $f(\lambda) = \lambda$ in the interval $(p - \delta, p+\delta)$. For example, this will work if $a_2,\ldots,a_n$ are sufficiently small.

$\endgroup$
  • $\begingroup$ Excellent! Any ideas on how to approximate that solution using the solution to $a_1/(\lambda - \lambda_1) = \lambda$? Or any other methods? $\endgroup$ – Leo Apr 15 at 15:22
1
$\begingroup$

Just a few ideas.

For the case where the $a_i$'s are positive and $\lambda_1 \gg \lambda_2 >\cdots > \lambda_n$, the solution would be just above $\lambda_1$ and a possible first approximation would be given by the solution of $$\frac{a_1}{\lambda - \lambda_1}-(\lambda - \lambda_1)+\left(\sum_{i=2}^n \frac{a_i}{\lambda_1 - \lambda_i}-\lambda_1\right)=0\tag 1$$ which is quadratic in $(\lambda-\lambda_1)$ and we select the positive solution of it. Another possiblity would be to expand the equation around $\lambda_1$ using Taylor series to get $$\frac{a_i}{\lambda - \lambda_i}=\frac{a_i}{\lambda_i-\lambda_i }-\frac{a_i }{(\lambda_1 -\lambda_i)^2}(\lambda-\lambda_1)+O\left((\lambda-\lambda_1)^2\right)$$ making the equation to be $$\frac{a_1}{\lambda - \lambda_1}-(\lambda - \lambda_1)\left(1+\sum_{i=2}^n \frac{a_i}{(\lambda_1 - \lambda_i)^2} \right)+\left(\sum_{i=2}^n \frac{a_i}{\lambda_1 - \lambda_i}-\lambda_1\right)=0\tag 2$$

Just to test, for a few values of $n$, I used $a_i=1 \,\, \forall i$, $\lambda_1=2p_{n}$ and $\lambda_i=p_{n-i+1}$. Below are given some results

$$\left( \begin{array}{ccc} n & \text{using } (1)& \text{using } (2) & \text{exact} \\ 2 & 6.1622777 & 6.1622777 & 6.1686686 \\ 3 & 10.100248 & 10.098702 & 10.101653 \\ 4 & 14.071954 & 14.070869 & 14.072521 \\ 5 & 22.045695 & 22.045297 & 22.045833 \\ 6 & 26.038680 & 26.038346 & 26.038779 \\ 7 & 34.029541 & 34.029351 & 34.029583 \\ 8 & 38.026434 & 38.026262 & 38.026467 \\ 9 & 46.021819 & 46.021705 & 46.021836 \end{array} \right)$$ AT least for these specific cases, it seems that $(1)$ is more than sufficient.

$\endgroup$

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.