2

In trying to solve Exercise #12.13.9 (b), p.247, from the textbook "Probability, An Introduction" by Grimmett and Walsh, an exercise about Random Walks on the edges of 3-dimensional cube, I came up with the following identities involving sums of multinomial coefficients: for each $n\in\mathbb{N}$,

$$ \sum_{i_1+i_2+i_3=2n-1 \\ i_1,i_2 \text{ and } i_3 \text{ odd}}\dfrac{(2n-1)!}{i_1! i_2! i_3!} = \dfrac{3^{2n-1}-3}4 \quad \text{and} \sum_{i_1+i_2+i_3=2n-1\\ \text{ exactly one of } \\ i_1, i_2 \text{ or } i_3 \text{ is odd}}\dfrac{(2n-1)!}{i_1! i_2! i_3!} = \dfrac{3^{2n}+3}4 $$

These two sums give a partition of the sum $ \displaystyle\sum_{i_1,i_2,i_3}\dfrac{(2n-1)!}{i_1! i_2! i_3!} = (1+1+1)^{2n-1}=3^{2n-1}$ into two partial sums. Likewise,

$$ \sum_{i_1+i_2+i_3=2n\\ \text{ exactly one of } \\ i_1,i_2 \text{ or }i_3 \text{ is even}}\dfrac{(2n)!}{i_1! i_2! i_3!} = \dfrac{3^{2n+1}-3}4 \quad \text{and} \sum_{i_1+i_2+i_3=2n \\i_1,i_2 \text{ and }i_3 \text{ even}}\dfrac{(2n)!}{i_1! i_2! i_3!} = \dfrac{3^{2n}+3}4 $$

These two sums give a partition of the sum $\displaystyle\sum_{i_1,i_2,i_3}\dfrac{(2n)!}{i_1! i_2! i_3!} = (1+1+1)^{2n}=3^{2n}$ into two partial sums.

The identities above can be proved by induction on $n$, but I wonder if anybody has a better proof, such as a bijective proof or a proof by generating functions. I am interested in generalizations of the identities as well, for example to cubes of arbitrary dimension, from the case of the 3-dimensional cube as given here.

Any comments will be appreciated.

kiuri
  • 21

2 Answers2

2

Here is a generating function proof. The multinomial theorem implies $$ (1+x+y)^n=\sum_{\substack{i_1,i_2,i_3\ge 0\\ i_1+i_2+i_2=n}} \frac{n!}{i_1!i_2!i_3!} x^{i_1}y^{i_2} \tag1 $$ By applying this twice, once with $x$ and $y$, once with $-x$ and $y$, and averaging the results, we get $$ {(1+x+y)^n+(1-x+y)^n\over 2}=\sum_{\substack{i_1,i_2,i_3\ge 0\\ i_1+i_2+i_2=n}} \frac{n!}{i_1!i_2!i_3!} y^{i_2}\cdot \frac{x^{i_1}+(-x)^{i_1}}{2} \tag2 $$ How does the RHS of $(2)$ compare to the RHS of $(1)$? Well, note that if $i_1$ is even, then $(-x)^{i_1}=x^{i_1}$, so $\frac{x^{i_1}+(-x)^{i_1}}{2}=x^{i_1}$. However, if $i_1$ is odd, then $ \frac{x^{i_1}+(-x)^{i_1}}{2}=0$. The result is that only the terms where $i_1$ is even survive. I will denote this by saying $$ {(1+x+y)^n+(1-x+y)^n\over 2}=\sum_{\text{$i_1$ is even}}\frac{n!}{i_1!i_2!i_3!} x^{i_1}y^{i_2} $$ This is progress. To get further progress, we apply the same trick to the LHS of $(2)$, but this time average between $y$ and $-y$. The result is $$ \frac{(1+x+y)^n+(1-x+y)^n+(1+x-y)^n+(1-x-y)^n}{4} =\sum_{\substack{\text{$i_1$ is even}\\ \text{$i_2$ is even}}} \frac{n!}{i_1!i_2!i_3!} x^{i_1}y^{i_2} \tag3 $$ Finally, substitute $x\gets 1$ and $y\gets 1$ into $(3)$, and we see that $$ \frac{3^n+1^n+1^n+(-1)^n}{4}=\sum_{\substack{\text{$i_1$ is even,}\\ \text{$i_2$ is even}}} \frac{n!}{i_1!i_2!i_3!} \tag4 $$ Equation $(4)$ explains both of your observations for (even, even, even) and (even, even, odd), depending on the parity of $n$.

For the other two cases, you can instead use the opposite trick; if you instead subtract the $x$ and $-x$ versions before dividing by $2$, the only the odd powers survive. Doing this subtraction trick with both $x$ and $y$, you get $$ {(1+x+y)^n-(1-x+y)^n-(1+x-y)^n+(1-x-y)^n\over 4}=\sum_{\substack{\text{$i_1$ is odd,}\\ \text{$i_2$ is odd}}}\frac{n!}{i_1!i_2!i_3!} $$ Finally, substitute $x=1$ and $y=1$ again to derive the other two formulae, depending on the parity of $n$.


All of this generalizes very nicely. For example, if you take the sum of $$\frac{n!}{i_1!\cdot i_2!\cdots i_m!}$$ over all tuples $(i_1,\dots,i_m)$ of nonnegative integers summing to $n$ which are all even, then the result can be proven to be $$ \frac1{2^m}\sum_{k=0}^m\binom{m}k(2k-m)^n $$ using similar methods. However, there is actually a much nicer proof using exponential generating functions; see this answer for that.

Mike Earnest
  • 75,930
  • Great. Thank you very much! I understood your proof of the trinomial case. I will now check out the "exponential generating functions" proof of the general case. – kiuri Aug 10 '21 at 20:57
  • I checked out the answer you cited. Very nice. I see that the method of exponential generating functions (I need to learn more about it) is applicable not only to all even cases but also to the cases of any even-odd combinations with an easy modification. Thank you again! – kiuri Aug 10 '21 at 22:05
  • 1
    @kiuri Indeed! EGF's are very cool, though I found them much less intuitive then ordinary generating functions. Chapter 3 of generatingfunctionology is a good reference. – Mike Earnest Aug 10 '21 at 22:19
  • Will take a look. Thanks for the pointer. – kiuri Aug 11 '21 at 01:19
2

#1

The first sum is the coefficient of $x^{2n-1}$ in the expansion of $$(2n-1)!\left(x+\frac{x^3}{3!}+\frac{x^5}{5!}+\ldots\right)^3$$ Using the fact that $e^x=1+x+\frac{x^2}{2!}+\frac{x^3}{3!}+\ldots$, we can use roots of unity filter to see that $$\frac{e^x-e^{-x}}{2}=x+\frac{x^3}{3!}+\frac{x^5}{5!}+\ldots$$ (If you don't know roots of unity filter, it's not too hard to just expand the taylor series for each of the terms and observe that the above statement is indeed true)

So our generating function is equivalent to $$\frac{(2n-1)!}{8}\left(e^x-e^{-x}\right)^3$$ $$\frac{(2n-1)!}{8}\left(e^{3x}-3e^x+3e^{-x}-e^{-3x}\right)$$ Since the coefficient of $x^{2n-1}$ in $e^{kx}$ is $\frac{k^{2n-1}}{(2n-1)!}$, the coefficient of $x^{2n+1}$ in our generating function is $$\frac{1}{8}(3^{2n-1}-3(1)^{2n-1}+3(-1)^{2n-1}-(-3)^{2n-1})$$ $$\frac{3^{2n-1}-3-3+3^{2n-1}}{8}$$ $$\frac{3^{2n-1}-3}{4}$$

#2 P.S. you might want to include that $i_1,i_2,i_3$ are nonnegative integers.

The second sum can be solved in a similar manner. Just WLOG assume that $i_1$ is odd and $i_2,i_3$ are even. Then multiply by $3$ at the end.

However, there is a faster way to find the result using the expression we found for the first sum.

Consider all the ways to express $2n-1$ as the sum of $3$ nonnegative integers. Using multinomial theorem, $$\sum_{a+b+c=2n-1} \binom{2n-1}{a,b,c}=(1+1+1)^{2n-1}=3^{2n-1}$$

A simple parity argument shows that either exactly $1$ of the summands must be odd, or all $3$ summands must be odd. We know that the sum if all $3$ summands are odd is $$\frac{3^{2n-1}-3}{4}$$

Hence, the sum if exactly $1$ of the summands must be odd is $$3^{2n-1}-\frac{3^{2n-1}-3}{4}$$ $$\frac{4\cdot 3^{2n-1}-3^{2n-1}+3}{4}$$ $$\frac{3\cdot 3^{2n-1}+3}{4}$$ $$\frac{3^{2n}+3}{4}$$

Alan Abraham
  • 5,142
  • 6
  • 20
  • Thanks for the response. I hadn't heard the term roots of unity filter, but I think I know what it means. It's exploiting the property of the character of the regular representation of cyclic groups. I'm sure it will be useful. I will keep it in mind! – kiuri Aug 11 '21 at 02:14
  • 1
    Roots of unity filter is a technique that states that given a generating function $f(x)=\sum_{n=0}^\infty a_nx^n$, then $\sum_{n=0}^\infty a_{kn}x^{kn}=\frac{\sum_{i=0}^k f(\omega_k x)}{k}$ where $\omega_k$ is a primitive $k^\text{th}$ root of unity. In the problem I plug in $x=1$ while using roots of unity filter with $f(x)=xe^x$ and $k=2$. – Alan Abraham Aug 11 '21 at 02:21
  • Yes. I am not that familiar with exponential generating functions. I need to learn more. about them. Thank you again for the detailed solution. – kiuri Aug 11 '21 at 15:27