3
$\begingroup$

Prove that$$\lim_{n\to\infty}\frac1n\sum_{i_1,i_2,...i_k=1}^n\lambda_1^{|i_1-i_2-s_1|}\lambda_2^{|i_2-i_3-s_2|}...\lambda_k^{|i_k-i_1-s_k|}$$is equal to$$\sum_{j=1}^k\lambda_j^{S+k-1}\prod_{l=1,l\ne j}^k\frac{1-\lambda_l^2}{(\lambda_j-\lambda_l)(1-\lambda_j\lambda_l)}$$where all $\lambda$s are, in absolute value, smaller than 1, distinct from one another, and $S=|s_1+s_2+...+s_k|$. This can be done by 'brute force' for $k=2$, $k=3$ and $k=4$ but I have no idea how to continue; yet I am reasonably confident that the formula is correct for any $k$.

$\endgroup$
1

1 Answer 1

5
$\begingroup$

Denote $$A=\sum_{p_1+\ldots+p_k=S} \lambda_1^{|p_1|}\lambda_2^{|p_2|}\ldots \lambda_k^{|p_k|}.\quad (1)$$ We prove two things:

1) Your limit equals $A$.

2) $A$ equals to what you write.

Start with 1). Note that any summand in your expression is of the form $\lambda_1^{|p_1|}\lambda_2^{|p_2|}\ldots \lambda_k^{|p_k|}$ for $\sum p_i=S$ (if $S$ is positive, define $p_m=i_{m+1}+s_m-i_m$, otherwise $p_m=i_m-i_{m+1}-s_m$ for $m=1,\ldots,k$, indices are cyclic), and each term appears at most $n$ times, since for fixed $i_1$ it appears at most once. Thus $$ \frac1n\sum_{i_1,i_2,...i_k=1}^n\lambda_1^{|i_1-i_2-s_1|}\lambda_2^{|i_2-i_3-s_2|}...\lambda_k^{|i_k-i_1-s_k|}\leqslant A, $$ and the upper limit does not exceed $A$. On the other hand, for any $B<A$ we may find finitely many terms in the sum (1) whose sum exceeds $B$. Each of this terms appears in your sum $n-O(1)$ times, thus we get that the lower limit is at least $B$. Since $B<A$ was arbitrary, it implies that the limit exists and equals to $A$.

2) $A$ is the coefficient of $t^S$ in the Laurent series $F:=\prod_i \sum_{p\in \mathbb{Z}} t^p \lambda_i^{|p|}$ (which is defined for $|t|$ close enough to 1), $i$-th multiple equals $1/(1-t\lambda_i)+\lambda_i/(t-\lambda_i)=t(1-\lambda_i^2)/((1-t\lambda_i)(t-\lambda_i))$. The product of these guys is a rational function in $t$ which is a linear combinations of the functions $1/(1-\lambda_jt)$ and $1/(\lambda_j-t)$: $$ F=\sum_j C_j/(1-\lambda_jt)+D_j/(\lambda_j-t).\quad (2) $$

$D_j/(\lambda_j-t)$ does not provide a coefficient of $t^S$ (it is expressed via negative powers of $t$), the coefficient of $t^S$ may come from $1/(1-\lambda_jt)=\sum_{n=0}^\infty \lambda_j^n t^n$. It remains to calculate the coefficient $C_j$ in the decomposition (2). We have $C_j=(F(t)(1-\lambda_jt))_{t=1/\lambda_j}$ that gives your formula.

$\endgroup$
2
  • $\begingroup$ Thanks Fedor, a brilliant solution. Is it fair to say that the answer can be built as a sum of $\lambda_j^S$ multiplied by the coefficient of $(1-t*\lambda_j)^{-1}$ in the usual partial-fraction expansion of your F (thus bypassing Laurent - but the same thing I guess)? $\endgroup$
    – Honza
    Jul 10, 2019 at 14:14
  • $\begingroup$ I do not understand how to bypass Laurent. $\endgroup$ Jul 10, 2019 at 15:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.