Do anyone have any idea how to solve this recurrence?$$T(k) = \sum_{i=1}^{k/2} {k \choose i}T(i)T(k-i)$$

I am working with this kind of recurrence for the first time and don't have much idea of how to proceed. But, for beginning I was trying to guess the solution and use substitution method. But, in the guessing part itself I am stuck. It seems like the solution should be exponential to $k$.


Everything becomes much simpler if when $n$ is odd we stop the sum at $\lfloor n/2 \rfloor$, and when $k$ is even we discount the term corresponding to $k/2$ by a half. In that case you can write $$ 2T(k) = \sum_{i=1}^{k-1} \binom{k}{i} T(i) T(k-i) $$ and so, putting $T(0) = 0$, we get that for $k \geq 2$, $$ 2T(k) = \sum_{i=0}^k \binom{k}{i} T(i) T(k-i). \tag{1}$$ (When $k = 1$ this reads $2T(1) = 0$.)

Now define the exponential generating function $$ P = \sum_{n \geq 0} T(n) \frac{x^n}{n!}. $$ Put $T(1) = c$. The identity (1) implies the equation $$ 2(P-cx) = P^2. $$ Solving the quadratic, we obtain $$ P = 1 - \sqrt{1 - 2cx} = 1 - \sum_{n \geq 0} \frac{(2n)!}{(1-2n)(n!)^24^n} (2cx)^n = \sum_{n \geq 1} \frac{(2n)!(c/2)^n}{(2n-1)n!} \frac{x^n}{n!}. $$ We conclude that for $n \geq 1$, $$ T(n) = \frac{(2n)!(c/2)^n}{(2n-1)n!} \sim \frac{1}{\sqrt{2} n} \left( \frac{2c}{e} n \right)^n. $$ The first few terms (starting with $T(1)$) are $$ c,c^2,3c^3,15c^4,105c^5,\ldots $$

  • $\begingroup$ Can you explain how did you get $2(P-cx)=P^2$ from identity 1. And, what is the intuition behind choosing this generating function? $\endgroup$ – Sayan Bandyapadhyay May 10 '14 at 17:08
  • $\begingroup$ For my first question, in particular I can't see why you are subtracting $cx$. As $T(0)=0$, I am getting $2P=P^2$. $\endgroup$ – Sayan Bandyapadhyay May 10 '14 at 17:51
  • $\begingroup$ I am subtracting $cx$ since identity (1) fails for $k = 1$, which in fact produces $0$ instead of $T(1)$. For further information, I suggest checking out the book generatingfunctionology. I'm using an exponential generating function since then $P^2$ corresponds to the convolution in (1) with the weights $\binom{k}{i}$. If identity (1) didn't have these weights, I would use ordinary generating functions. $\endgroup$ – Yuval Filmus May 11 '14 at 0:10

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.