Let's suppose we have $n+1$ real values $d_i$ ($i=0,\ldots,n$) with $0 \le m \le d_0 \lt d1 \lt \ldots \lt d_{n-1} \lt d_n \le M$ (we don't know anymore about $d_i$), and $n$ real values $r_i \ge 1$ ($i=1,\ldots,n$) with no specified order. Then compute $c$ as follows: $$ c = d_0 + \sum_{i=1}^n (d_i-d_{i-1})r_i $$

I want to get a function $S(m, M, r_1, \ldots, r_n)$ such that: $$ S(m, M, r_1, \ldots, r_n) \ge c $$

What is the best (i.e. the lowest) $S(m, M, r_1, \ldots, r_n)$ for any $d_i$ satisfying the constraint above?

I suppose the best I can do is: $$ S(m, M, r_1, \ldots, r_n) = (M-m) \cdot \max_{i=1}^n r_i + m$$

But I am not able to demonstrate it (not even sure it is always $\ge c$).

Any hint? Thank you!

@Michael Seifert, following your observation and noting that we can write: $$ (r_i-r_{i+1})d_i = (r_i-r_{i+1} + |r_i-r_{i+1}|)\frac{d_i}{2}+(r_i-r_{i+1} - |r_i-r_{i+1}|)\frac{d_i}{2} $$

where the first term is always $\ge 0$ and the second term is always $\le 0$, and being $m \le d_i \le M$ then: $$ (r_i-r_{i+1})d_i = (r_i-r_{i+1} + |r_i-r_{i+1}|)\frac{d_i}{2}+(r_i-r_{i+1} - |r_i-r_{i+1}|)\frac{d_i}{2} \le (r_i-r_{i+1} + |r_i-r_{i+1}|)\frac{M}{2}+(r_i-r_{i+1} - |r_i-r_{i+1}|)\frac{m}{2} = $$$$ (r_i-r_{i+1})\frac{M+m}{2}+|r_i-r_{i+1}|\frac{M-m}{2} $$ and $$ \sum_{i=0}^n (r_i-r_{i+1})d_i \le \frac{M+m}{2}(1-r_1+r_1-r_2+r_2-\ldots-r_n+r_n)+\frac{M-m}{2}\sum_{i=0}^n |r_i-r_{i+1}|=$$$$\frac{M+m}{2}+\frac{M-m}{2}\sum_{i=0}^n |r_i-r_{i+1}|\qquad (1)$$

And only if there exists $j$ such that $r_j=\max_{i=1}^nr_i$ and $1 \le r_1 \le r_2 \le \ldots \le r_{j-1} \le r_j$ and $r_j \ge r_{j+1} \ge r_{n-1} \ge r_n$ then (1) can be simplified into: $$ (M-m)\max_{i=1}^nr_i+m \qquad (2) $$

However, making some numerical tests, I think the right solution is (2).

  • $\begingroup$ Just an observation: note that if you define $r_0 = 1$ and $r_{n+1} = 0$, then your expression for $c$ can be rewritten as $$c = \sum_{i = 0}^{n} (r_i - r_{i+1}) d_i,$$ which may be more conducive to bound-placing. $\endgroup$ – Michael Seifert Jun 29 '18 at 15:22

Finally it turned out to be easy: $$ c = d_0 + \sum_{i=1}^n (d_i-d_{i-1})r_i \le d_0 + \max_{i=1}^nr_i \cdot \sum_{i=1}^n (d_i-d_{i-1}) = d_0 + \max_{i=1}^nr_i \cdot (d_n-d_0) = d_0 (1 - \max_{i=1}^nr_i) + d_n \max_{i=1}^nr_i $$

and since the first term is always $\le 0$ and the second term is always $\ge 0$ then: $$ d_0 (1 - \max_{i=1}^nr_i) + d_n \cdot \max_{i=1}^nr_i \le m (1 - \max_{i=1}^nr_i) + M \cdot\max_{i=1}^nr_i = (M - m) \cdot\max_{i=1}^nr_i + m $$

And if $r_j=\max_{i=1}^nr_i$ choosing $d_0 = \ldots = d_{j-1} = m$ and $d_j = \ldots = d_n = M$ gives exactly $c = (M - m) \cdot r_j + m = (M - m) \cdot\max_{i=1}^nr_i + m$ so there is no better boundary.


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.