Bivariate generating function for squared binomial coefficients

binomial-coefficientsgenerating-functionslegendre polynomials

I want to find a closed form to the bivariate generating function
$$
G(x, y) = \sum\limits_{i, j} \binom{i+j}{i}^2 x^i y^j.
$$

Ideally, I would prefer a direct approach that is based on the definition above.

I know that there is a closed form here, as one can reduce the summation to Legendre polynomials:
$$
G(x, y) = \sum\limits_n y^n \sum\limits_k \binom{n}{k}^2 \left(\frac{x}{y}\right)^k = \sum\limits_n (y-x)^n P_n \left(\frac{y+x}{y-x}\right),
$$

where $P_n(x)$ is the $n$-th Legendre polynomial. From this, using the generating function formula
$$
\sum\limits_n P_n(x) t^n = \frac{1}{\sqrt{1-2xt+t^2}},
$$

we get the closed-form expression for $G(x, y)$ as
$$
\boxed{G(x, y) = \frac{1}{\sqrt{1-2(y+x)+(y-x)^2}}}
$$

But I totally fail to see any meaningful way to derive it in a more direct and self-contained way. Any hints? And while we're at it, are there similar closed-form expressions for higher powers of $\binom{n}{k}$?

Best Answer

After some thoughts, I couldn't find any reasonable way to do it directly, but at least now I understand how to fill the gap in the derivation of the genfunc of Legendre polynomials. First, let's expand the genfunc.

$$ G(t, x) = \frac{1}{\sqrt{1-2xt+t^2}} = \sum\limits_k \frac{t^k(2x-t)^k}{4^k}\binom{2k}{k}. $$ If we want to extract the coefficient near $t^n$, we get $$ [t^n] G(t, x) = \frac{1}{2^n}\sum\limits_k \binom{2k}{k} \binom{k}{n-k} (-1)^{n-k} x^{2k-n}, $$ where $k$ goes from $\lceil n/2 \rceil$ to $n$. Changing $k \to n-k$, we get $$ [t^n] G(t, x) = \frac{1}{2^n} \sum\limits_k (-1)^k \binom{2n-2k}{n-k} \binom{n-k}{k} x^{n-2k}. $$ On the other hand, $$ \binom{2n-2k}{n-k}\binom{n-k}{k}=\binom{2n-2k}{n-k,k,n-2k}=\binom{2n-2k}{n}\binom{n}{k}, $$ which gives one of the standard expressions $$ [t^n] G(t, x) =\boxed{ \frac{1}{2^n} \sum\limits_k (-1)^k \binom{n}{k}\binom{2n-2k}{n} x^{n-2k}} $$

Now, we need to connect the dots. To make it relevant to $\binom{n}{k}^2$, we define Legendre polynomials as $$ P_n(x) = \frac{1}{2^n} \sum\limits_{k=0}^n \binom{n}{k}^2(x-1)^{n-k} (x+1)^k. $$ It's easy to check that it compacts into $$ P_n(x) = [t^n] \frac{(t+x-1)^n(t+x+1)^n}{2^n} = [t^n] \frac{((t+x)^2-1)^n}{2^n}. $$

We can now expand it back to get the same expression: $$ [t^n]\frac{((x+t)^2-1)^n}{2^n} = [t^n]\frac{1}{2^n}\sum\limits_k (-1)^k \binom{n}{k}(x+t)^{2n-2k} =\boxed{ \frac{1}{2^n} \sum\limits_k (-1)^k \binom{n}{k}\binom{2n-2k}{n} x^{n-2k}} $$ I still don't like this approach, as I needed to know $G(t, x)$ in advance to justify it, and it's quite far from being direct in terms of the genfunc for $\binom{n}{k}^2$, but I suppose it's still nice to have proofs that are directly based on binomial identities, rather than physics...