CF1540 Solution

CF1540 Solution

A

If we sort $d_i$ in the increase order, there is a clearly construction:

1. link $1$ to $n$, weight is $d_n$.
2. Expect the situation above,for $i<j$, link no edges between $i$ and $j$; instead, link $j$ to $i$, weight is $d_i-d_j$.
3. the ans is
   $$d_n+\sum _{j=1}^n\sum _{i=1}^{j-1}(a_i-a_j)=d_n+\sum _{i=1}^n(n+1-2i)a_i$$

We can prove that this method is one of the best ones.

B

The problem statement can be transferred to : given a tree, you should choose a root and mark it. Afterwards, extend the subtree of the marked nodes.

At first, we need to choose a root $r$. And the answer becomes $ans=\frac{1}{n}{\sum }_{r=1}^{n}E(r)$, $E(r)$ in it refers to the expected number of inversions according to the process if we fix the root of the tree, that is, $r$.

Now we need to calculate $E(r)$. Let us consider each pair $(u,v),u>v$'s contribution, that is, the expected number that $u$ marks before $v$.

we assume that the latest common ancestor of $u$ and $v$ is $l$. If we haven’t reached $l$, the answer is the same as the situation when we reached $l$. So we consider the latter. We have two stacks, the sizes of which are $dist(u,l)$ and $dist(v,l)$ respectively. Define $f(x,y)$ as the expected number that the stack of size $x$ decrease to 0 faster that size $y$. We suppose the possibility of choosing an unmarked node is $p$, so there are :

$$f(x,y)=p\cdot f(x-1,y)+\cdot f(x,y-1)+(1-2p)\cdot f(x,y)$$

So

$$f(x,y)=\frac{f(x-1,y)+f(x,y-1)}{2}$$

The ans is $f(dist(u,l),dist(v,l))$.

The time complexity is $O(n^3\log n)$ (use $O(\log n)$ LCA) or $O(n^3)$ (use $O(1)$ LCA).

C

What we really do when we take an operation on $a_i$ and $a_{i+1}$?

• if $a_i\le \frac{a_i+a_{i+1}-b_i}{2}$, there is $a_{i+1}\ge \frac{a_i+a_{i+1}+b_i}{2}$, and the contrary are correct, too. So after the operation, $(a_i,a_{i+1})$ becomes $(a_i,a_{i+1})$ or $\left(\frac{a_i+a_{i+1}-b_i}{2},\frac{a_i+a_{i+1}+b_i}{2}\right)$.

We find that $a_i+a_{i+1}$ keeps constant, and we make $(a_{i+1}-a_i)\leftarrow \max (a_{i+1}-a_i,b_i)$ in the operation $i$.

Let us consider the final seqence (suppose that is $f_i$). We know that $f_{i+1}-f_i\ge b_i$.

There is an important conclusion : if $f_{i+1}-f_i>b_i$, then operation $i$ doesn’t matter throughout the process.

prove:

if $a_{i+1}-a_i\le b_i$ stands through the process, we cannot make it up to $f_{i+1}-f_i$, for operation $i-1$ only makes $a_i$ increase and operation $i+1$ only makes $a_{i-1}$ decrease.

For operation $i$ doesn’t matter, $[1,i]$ and $[i+1,n]$ becomes independent parts, for their sums cannot exchange.

If $k$ is the first number which satisfies the condition above, then we get:

$$\sum _{i=1}^k{f}_i=\sum _{i=1}^k{a}_i$$

also we know $f_{i+1}-f_i=b_i,(1\le i<k)$

so

$$f_1+(f_1+b_1)+(f_1+b_1+b_2)+\cdots =\sum _{i=1}^k{a}_i$$

We define ${\sum }_{i=1}^n{\sum }_{j=1}^i{b}_j$ as $B_n$,${\sum }_{i=1}^n{a}_i$ as $A_n$, then

$$f_1=\frac{A_k-B_{k-1}}{k}$$

But we do not know which $k$ satisfies $f_{k+1}-f_k>b_k$ in the end. There is a conclsion : $f_1={\min }_{k=1}^n(\frac{A_k-B_{k-1}}{k})$, why ?

prove:

for all $k\in [1,n]$, ${\sum }_{i=1}^k{f}_i\le {\sum }_{i=1}^k{a}_i$, because operation $k$ only decrease the sum of $[1,k]$ and increase the sum of $[k+1,n]$($k=n$ is a special situation). So $k{f}_1+B_{k-1}\le {\sum }_{i=1}^k{f}_i\le A_k$, $f_1\le \frac{A_k-B_{k-1}}{k}$, and there exists $k_0$ that satisfies $f_1=\frac{A_{k_0}-B_{k_0-1}}{k_0}$, Q.E.D

So the inequation $f_1\ge x$ can be described as $\forall i\in [1,n],\frac{A_i-B_{i-1}}{i}\ge x$, that is, $A_i\ge B_{i-1}+i\cdot x$. We can do a dynamic programming to count the qualified $A_i$s. Call $f(i,s)$ the number of sequences which satisfy $A_i=s$. And we need use prefix summation to optimize the transition. If set $m=\max c_i$, then the time complexity is $O(qm{n}^2)$.

How to optimize the algorithm? There is an important conclusion : we need and only need to calculate $O(m)$ kinds of $x$. Why?

If $nm<B_{n-1}+nx$, the answer is 0; if $0>B_{n-1}+nx$, for all the $x$ which satisfies this condition the ans is the same.so we only require to solve the situation when $B_{n-1}+nx\in [0,nm]$, that is, $x\in [\lfloor -B_{n-1}/n\rfloor ,\lceil m-B_{n-1}/n\rceil ]$. The time complexity is $O(m^2{n}^2+q)$.

D

To be continued…

E

To be continued…