Maximize multivariable integer function

229 Views Asked by At

Let $L\ge 2$ be an integer, $a_0=10$, $a_L=2$, and $a_1,...,a_{L-1}\ge 2$ be integers summing to $36$. Maximize $\sum_{i=0}^{L-1}a_ia_{i+1}$ over $L$ and $a_1,...,a_{L-1}$.

I think the maximum is $556$ achieved at $L=3$ and $(a_1,a_2)=(22,14)$. It’s easy to show if $L\le 3$. For $L>3$, I tried to increase the sum by using $a_1,...,a_{L-3},a_{L-2}+a_{L-1}$ instead of $a_1,...,a_{L-1}$, but it doesn’t work for arbitrary $a_1,...,a_{L-1}$, so I tried to first increase the sum by swapping $a_i$ with $a_j$ when they were out of order in some sense, like the proof of the rearrangement inequality, but I couldn’t figure out how. Maybe there is a way using calculus.

2

There are 2 best solutions below

0
On

Here's a solution with calculus. This confirms that indeed the maximum is achieved, as stated by the author, for $L=3$ and $(a_1,a_2)=(22,14)$. However, I'd be grateful for explanations from MSE readers (see text below and case 3) where I do not understand some results.

First note that, as $a_1,...,a_{L-1}\ge 2$ and $\sum_{i=1}^{L-1}a_i = 36$, we need to consider only $L\le 19$.

First, relax the condition $a_i \ge2$. Then, enforce the sum condition with Lagrange parameter $\lambda$. The function which needs to be maximized is then

$$ S = \sum_{i=0}^{L-1}a_ia_{i+1} - \lambda(-36 + \sum_{i=1}^{L-1}a_i) \\ = 10 a_1 + 2 a_{L-1} + \sum_{i=1}^{L-2}a_ia_{i+1} - \lambda(-36 + \sum_{i=1}^{L-1}a_i) \\ $$ Maximization goes by taking partial derivatives which gives $$ \frac{\partial S}{\partial a_j} = a_{j-1} + a_{j+1} - \lambda = 0 \quad \forall j=2\cdots L-2\\ \frac{\partial S}{\partial a_1} = 10 + a_2 - \lambda = 0\\ \frac{\partial S}{\partial a_{L-1}} = 2 + a_{L-2} - \lambda = 0 $$

The third equation can only be enforced for $L\ne4$ because otherwise it contradicts the second equation.

For $L \ne 4$, the sum of all these equations gives $2 \sum_{i=1}^{L-1}a_i - (a_1 + a_{L-1}) + 12= (L-1)\lambda$ or, plugging in the sum condition, $84 - (a_1 + a_{L-1})= (L-1)\lambda$. Now let us inspect the partial derivatives above. Start with unknown $a_1$ and $\lambda$. Then we get the following equations, which can be used up to $a_j$ with $j = L-1$: $$ a_2 = \lambda -10\\ a_3 = \lambda -a_1\\ a_4 = \lambda -a_2 = 10\\ a_5 = \lambda -a_3 = a_1\\ a_6 = \lambda -a_4 = \lambda -10\\ a_7 = \lambda -a_5 = \lambda -a_1 $$

It is indeed very strange to see that $a_4=10$ which will come in action for $L \ge 5$. As we shall see in case 3 (below), this produces errors which I cannot explain.

Since the recursion length is two, and since the last two values repeat the first two, we see that in general, $a_k = a_{k \, \text{mod}\, 4}$, so we need only consider 4 cases.

Case 1: $L-1 =2 + 4 n$. Then $a_{L-1} = a_2 = \lambda-10$ and we have $84 - (a_1 + a_{L-1}) = 84 - (a_1 +\lambda-10) = (L-1)\lambda$ or $ 94 - a_1 = L \lambda = (3 + 4n) \lambda$ . Further, we have $36 = \sum_{i=1}^{L-1}a_i = a_1 + a_2 + n\cdot(2 \lambda) = -10 + a_1 + (2n +1)\lambda$ or $ 46 - a_1 = (1 + 2n) \lambda$ . Combining the two equations gives $$ \lambda = \frac{24}{1+n}\\ a_1 = -2 +\frac{24}{1+n} $$

For $n=0 \; (L=3)$ we get $a_1 = 22$ and $a_2 = \lambda -10 = 14$ which is the result that the question text already stated.

For $n=1 \; (L=7)$ we get $\lambda = 12$ and $a_1 = 10$ which gives the $a$-sequence (starting with $a_0$) $10, 10, 2,2,10, 10,2, 2$ and hence $\sum_{i=0}^{6}a_ia_{i+1} = 100 + 20+4+20+100+20+4 = 268$.

For $n=2 \; (L=11)$ we get $\lambda = 8$ and $a_1 = 6$, $a_2 = -2$ so this does not allow unrestricted optimization any more.

Case 2: $L-1 =3 + 4 n$. Then $a_{L-1} = a_3 = \lambda-a_1$ and we have $36 = \sum_{i=1}^{L-1}a_i = a_1 + a_2 + a_3 + n\cdot(2 \lambda) = -10 + (2n +2)\lambda$ or $ 23 = (n +1)\lambda$. The parameter $a_1$ remains arbitrary.

For $n=0$ $(L =4)$, we have $\lambda = 23$ and hence $a_2 = 13, a_3 = 23 - a_1$. So $S = a_0 a_1 +a_1 a_2 +a_2 a_3 +a_3 a_4 = 23 a_1 + 13 (23 - a_1) + 2 (23 - a_1) = 8 a_1 + 345$. So $S$ rises with $a_1$ and the limiting factor is $a_3 = 23 - a_1 \ge 2$ which allows $a_1 = 21$, giving $S=513$ .

For $n=1$ $(L =8)$, we have $\lambda = 11.5$ and hence $a_2 = 1.5$ which doesn't fit. Let $\lambda = 12$ and we have $a_2 = 2, a_3 = 12 - a_1$. So $S = a_0 a_1 +a_1 a_2 +a_2 a_3 +a_3 a_4 = 12 a_1 + 2 (12 - a_1) + 2 (12 - a_1) = 8 a_1 + 48$. So $S$ rises with $a_1$ and the limiting factor is $a_3 = 12 - a_1 \ge 2$ which allows $a_1 = 10$, giving $S=128$ .

Case 3: $L-1 =4 + 4 n$. Then $a_{L-1} = a_4 = 10$ and we have $84 - (a_1 + a_{L-1}) = 84 - (a_1 +10) = (L-1)\lambda$ or $ 74 - a_1 = (L-1) \lambda = (4+ 4n) \lambda$ . Further, we have $36 = \sum_{i=1}^{L-1}a_i = a_1 + a_2 + a_3 + a_4 + n\cdot(2 \lambda) = (2n +2)\lambda$ or $18 = (1 + n) \lambda$ . Combining the two equations gives $$ \lambda = \frac{18}{1+n}\\ a_1 = 2 $$

For $n=0 \; (L=5)$ we get $a_1 = 2$ and $\lambda = 18$ which gives the $a$-sequence (starting with $a_0$) $10, 2, 8, 16, 10,2$ and hence $\sum_{i=0}^{4}a_ia_{i+1} = 20 + 16 + 128 + 160 + 20 = 344$. However, by numerical evaluation, the sequence $a_1=20,a_2=12,a_3=2,a_4=2$ gives $S=472$ and I do not see how this comes about so I'll stop here. I'd be grateful for some explanation.

As already noted by the author, it is unlikely that increasing $L$ will lead to higher values of $S$, given that with $a_i \ge 2$ an increasing portion of the sum (36) must be held back for ever more values $a_i$, which will not allow products of high numbers.

0
On

For a particular numerical value of $L$, this can be formulated and solved as a Mixed Integer non-convex Quadratic Programming problem, a.k.a.non-convex MIQP, using readily available non-convex MIQP solvers.

CPLEX, using optimalitytarget = 3 can be used to (try to) solve this to global optimality., or using optimalitytarget = 2 to local optimality. BARON could also be used to solve for the global optimum. Whether or not the global optimum is found and claimed to be found depends on available time and memory. It could be fast or take a very long time.