7
$\begingroup$

suppose we have the following two sequences

$$\alpha_k = (k-1)\left(1-\frac {1}{1+(k+1)l}\right) \quad , k \geq 2$$ $$\beta_k = (k-1)\left(1+\frac {1}{1+(k-1)l}\right) \quad , k \geq 2$$

where $l$ is a positive constant and define the sequence $c_k$ recursively by: $$c_2 = - 1/\beta_2 $$ $$c_3 = 0 $$ $$c_{k+1} = \frac{\alpha_{k-1}}{\beta_{k+1}}c_{k-1} \quad , k \geq 3 $$

it is not hard to see that this would give $$c_2 = - 1/\beta_2$$ $$c_{2k} = -\frac{\alpha_2}{\beta_2 }\cdot\frac{\alpha_4}{\beta_4 }\cdot\cdot\cdot \frac{\alpha_{2k-2}}{\beta_{2k-2} }\cdot \frac{1} {\beta_{2k}} \quad , k \geq 2$$

$$c_{2k+1} = 0 \quad , k \geq 1 $$

apparently we must have $d_k = c_{2k} \sim k^{-(1+1/l)}$ but I have no idea how to show this. can anyone shed some light on this?

also a similar question for the recursion $$c_2 = - 1/\beta_2 $$ $$c_3 = \frac{\gamma_2}{\beta_3}c_2 $$ $$c_{k+1} = \frac{\alpha_{k-1}}{\beta_{k+1}}c_{k-1} + \frac{\gamma_k}{\beta_{k+1}}c_k \quad , k \geq 3 $$

where $$\gamma_k = \sigma \frac{l(k^3-k)}{1+kl} \quad , k \geq 2$$ $\sigma$ being also a positive constant

How would the asymptotics look like in this case?

$\endgroup$
6
$\begingroup$

For the first sequence, $$d_k=\frac{\alpha_{2k-2}}{\beta_{2k}}d_{k-1}=\frac{k-3/2}{k+1/2-1/\ell}d_{k-1}$$ so one has $$d_k=C\frac{\Gamma(k-1/2)}{\Gamma(k+3/2-1/\ell)}\ ,$$ the constant $C$ being determined by the initial condition $d_1$, namely $$C=d_1\frac{\Gamma(5/2- 1/\ell)}{\Gamma(1/2)}\ . $$ Recall that $\Gamma(x+a)=x^a\Gamma(x)(1+o(1))$ as $x\to+\infty$, so $$d_k=Ck^{-1-1/\ell}(1+o(1)). $$

For the second sequence, $$c_{k+1}=\frac{(k-1)(k+1)}{k+2/\ell }c_k+\frac{k-2}{k+2/\ell}c_{k-1}$$ which implies $c_{k-1}/c_k=O(1/k)$; if we plug this in the recursion again we have $$c_{k+1}=\frac{(k-1)(k+1)}{k+2/\ell }(1+O(1/k^2))\ , $$ whence $$c_k= A\frac{\Gamma(k+1)\Gamma(k-1)}{\Gamma(k+2/\ell) }(1+o(1))\ ,$$ because the infinite product of $1+O(1/k^2)$ is convergent. By the Stirling formula $$c_k= B k^{k-1/2+2/\ell}e^{-k} (1+o(1))$$ for a certain constant $B$.

$\endgroup$
  • $\begingroup$ is it just me or is the sign in the power of $k$ wrong in your answer for the first question? $\endgroup$ – Koen Van Tuijl Jan 1 '16 at 0:17
  • $\begingroup$ it is… fixed now $\endgroup$ – Pietro Majer Jan 1 '16 at 7:37
3
$\begingroup$

The first factor telescopes completely, and the second becomes easier by combining $\alpha_{2\kappa}$ with $\beta_{2\kappa+2}$, so $$ c_{2k} = \frac{1+\ell}{(2k-1)(2+\ell)}\prod_{\kappa=1}^k \left(1-\frac{2}{2+(2\kappa+1)\ell}\right). $$ Now take logarithms, and apply Euler-MacLaurin summation.

$\endgroup$
  • $\begingroup$ Do you mean by applying the Eucler-Maclaurin summation that I should replace the sum of logarithms by an integral over a logarithm? $\endgroup$ – Koen Van Tuijl Dec 31 '15 at 16:48
  • $\begingroup$ If $f$ is a smooth function, then $\int f(t)$ is the first approximation to $\sum f(n)$. Euler-Maclaurin gives a sequence of correction terms, involving higher and higher derivatives of $f$. When applied to a sum of length $N$, using $k$-th derivatives usually gives an error of size $k!^C N^{-k}$, so taking $k$ too big does not pay off, but most of the time one or two terms suffice. $\endgroup$ – Jan-Christoph Schlage-Puchta Jan 1 '16 at 11:09
  • $\begingroup$ Anyway, Majer's answer is better. The final result is the same, as the derivation of Stirling involves Euler-Maclaurin, but this way you save yourself a lot of work. $\endgroup$ – Jan-Christoph Schlage-Puchta Jan 1 '16 at 11:13

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.