It is needed just to implement actions described in statement. You had to read data and to calculate number of members of team, which were sure about the solution, for every task. If this number is greater than one, the answer must be increased by one.

231B - Magic, Wizardry and Wonders

Let's see, what will be the last number of array after *i* iterations. After the first iteration it will be *a*_{n - 1}–*a*_{n} (and total number of elements will be decreased by one). After the second iteration the last number will be *a*_{n - 2}–*a*_{n - 1} + *a*_{n}. It is not hard to see, that after *n* - 1 iterations remain *a*_{1}–*a*_{2} + *a*_{3}–*a*_{4} + ... + ( - 1)^{n + 1}·*a*_{n}. In a such way, our task is to put numbers from 1 to *l* in array so, that sum of numbers in odd positions minus sum of numbers in even positions will equal to given *d*. This means sum of numbers in odd positions must be equal . But the minimal sum can be , and the maximal — .

Because of this we should choose *a*_{2·k} so, that *s* fits the boundaries. Constrains allow to do it in a such manner. Firstly, put ones on the even positions. If *s* > *maxv* after that, the answer is - 1. Otherwise, let's increase each *a*_{2·k} by one until *s* = *minv*. If we put *l* in all even positions and *s* < *minv*, than answer is - 1 too. After we put numbers on even positions, let's write 1 in all odd positions, and while sum of this elements is less than *s* increase each one by fitting value.

One of the main observations, needed to solve this problem, is that the second number in answer always coincides with someone *a*_{j}. Let's see why it is true. Suppose, the second number of the answer is *a*_{j} + *d* for someone *j* and *a*_{j} + *d* ≠ *a*_{i} for all *i*. This means, we increased some numbers, which is less than *a*_{j}, so that they became equal to *a*_{j}, and then all this numbers and some numbers, which is equal to *a*_{j}, we increased to *a*_{j} + *d*. But if we didn't increase all this numbers to *a*_{j} + *d* and remain they equal to *a*_{j}, we'd perform less operations and the answer would be better.

Due to this fact we can solve problem in a such manner. Sort array in non-decreasing order. Iterate over *a*_{i} and calculate, what is the maximal number of *a*_{i} we can obtain. For maximizing first number of answer, we must increase some lesser numbers to *a*_{i} and perform not greater than *k* operations. It is obvious that firstly we should increase such *a*_{j} that *a*_{i}–*a*_{j} is minimal. So, if we can solve problem in *O*(*n*^{2}), we would iterate *j* from *i* to 0 and increase *a*_{j} to *a*_{i}, while we could. But the solution must be faster, and we will use binary search. We will brute the number of numbers, which we must do equal to *a*_{i}. Suppose we fix *cnt* this value. Now we have to check if we can do *cnt* numbers equal to *a*_{i} by not greater than *k* operations. For doing this, let’s calculate . If this value not greater than *k*, we can do it. For calculating sum quickly, we can save prefix sums and than *s*_{i - cnt + 1, i} = *s*_{i}–*s*_{i–cnt}. Finally we solved this problem in *O*(*n*·*logn*).

The main subtask of this problem is to check whether we can observe the center of face of parallelepiped from point *p* = (*x*, *y*, *z*). Let’s see the case, when the face belongs to plane *z* = *z*_{1}. For performing all calculations in integer numbers, multiply all coordinates *x*, *y*, *z*, *x*_{1}, *y*_{1}, *z*_{1} by 2. Take the point and normal to plane, containing the fixed face, which is directed out of interior of parallelepiped, that is . Also take vector . If undirected angle between this vectors is less than 90 degrees, we can observe *a* from *p*. For checking this we can use scalar product. If scalar product of and is strictly greater than zero, than that angle is fitting.

In this problem you should find the number of simple paths between some pair of vertices in vertex cactus. If you learn the structure of these graphs, it is not hard to see, that if we’ll squeeze each cycle in one vertex, we get a tree. So let’s squeeze all cycles in source graph and get this tree. Also every vertex of this tree we’ll mark, if it is squeezed cycle (let’s call this vertices *1-vertices*) or single vertex in source graph (this vertices we’ll call *0-vertices*).

Then, we’ll do the following to find the number of paths between vertices *a* and *b* in source graph. Suppose *c* is a vertex, corresponding to *a* in obtained tree (it can be a single vertex or a vertex, corresponding to a squeezed cycle with *a*), and *d* is a vertex, corresponding to *b*. Let’s denote *deg* is the number of 1-vertices in path from *c* to *d* in tree. Than it is easy to understand, that the answer for query is , because every cycle (1-vertex) increase the number of possible ways twice (you can go from one vertex to other by two ways in cycle).

It means that we need to count the number of 1-vertex on the path from one vertex to other in tree quickly to answer a query. We can do it in a following way. Hang our tree for one vertex, which we’ll call a root. Denote for every vertex *cnt*_{v} is the number of 1-vertex on the way to the root (including root and vertex itself). Suppose we want to find the number of 1-vertex on the path from *a* to *b*. Denote *c* is the least common ancestor of vertices *a* and *b*. Than number of 1-vertex on the way from *a* to *b* is equal to *cnt*_{a} + *cnt*_{b}–2·*cnt*_{c}, if *c* is 0-vertex and *cnt*_{a} + *cnt*_{b}–2·*cnt*_{c} + 1, if *c* is 1-vertex. The least common ancestor can be found by standard method — binary method recovery. Finally we have *O*(*m* + *k*·*logn*) solution.

I solved Problem B by dfs :

Lets maintain two arrays ,d[] the array of differences and a[] the array of numbers . Initially d[0]=d . Now we know that a[0] lies in the range of [1...L] and a[0]-d[1]=d[0] => d[1]=a[0]-d[0].So we need to iterate over all possible values of a[0] i.e [1..L] and check if it is going good .Similarly we check for d[1] and keep digging deeper until we reach the base case ( index=n-1) . My code implementing the above idea code ..As you see the code exceeded time on pretest 10 (call it lucky or unlucky).

So other thing to observe is what is the maximum ans minimum values that the difference array can take. With this we can prune the tree into avoid going to some branches which are unncessary . Lets take the trivial values of max and min for the last location . i.e . d[n-1] as max[n-1]=L ans min[n-1]=1. Now max[n-2]=L-min[n-1] and min[n-2]=1-max[n-1] .So like this , we make up these two arrays . And everytime we are calling the dfs function we check if the d value is in the range of max and min .Code

It will be great if you can post the implementation code for these solutions as well.

If you really need them, you could see other participants' solutions in the standing page.

Why there is no English tutorial, I can't understand Russian...

http://codeforces.com/blog/entry/5486

Change the url from http://codeforces.com/blog/entry/5486 to http://codeforces.com/blog/entry/5486

Or simply press the British flag on the top right corner of the page

Warning — I am a complete newb at online coding competitions, so if dealing with people like me causes you lots of frustration, please just ignore my comment and pass on!

I competed in this competition (only submitted on 231A Team) I was wondering if someone might hop into a 'talk' me and help me understand why my solution was incorrect (my code is so ugly I dont want to post it here)? I had it working fine my version of Visual Studio, but when I submitted it didn't pass the first test...think it's probably something to do with how I read the stdin...

Any help gratefully received — thanks. ahab

There are spaces between the numbers in the input data. Your method of inputting data is weird. Why use

`string`

s and`stringstream`

s if you can read numbers directly from`cin`

?Thanks for the reply — you're right. It was weird to use

`string`

(The perils of being newb coder!) and I just realised my error this morning with the spaces between the letters when I started stepping through someone else's succussful submission. Thank you for taking the time to look.Out of interest, when you are coding and debugging your potential submissions, how do you replicate the standard input? Do you set up the examples given in the problem in a file and point

`cin`

to that, or do you type them in by hand as required by your code as you execute a trial run?One popular way to do it is to store test input in a file and then insert

at the beginning of

`main()`

.Then you would use

`cin`

as usual. The`ONLINE_JUDGE`

macro is defined on Codeforces, so this code will redirect`cin`

if you compile it on your computer and will still read from standard input if compiled on Codeforces. You can also do the same for output with`freopen("output.txt", "w", stdout);`

.Then there is a problem of inputting multiple test cases. Many people deal with it like this: they store all tests in one file, and in the main program write

While there is something yet to be read from file, this loop will begin anew, allowing to process many tests at once. But be careful to re-initialize all global variables each time. I won't say this is the best solution, but it is simple and quick enough to write.

Could somebody help me with the sample for problem E: points 3-5, isn't the number of simple paths equal to 3? 3-2-1-4-3-5, 3-4-1-2-3-5, 3-5? what am i missing?

For E — Cactus, can someone explain the algorithm needed to "squeeze" the cycles?

Could anyone explain this part in the editorial for 231C please?

" Suppose we fix cnt this value. Now we have to check if we can do cnt numbers equal to ai by not greater than k operations. For doing this, let’s calculate . If this value not greater than k, we can do it. For calculating sum quickly, we can save prefix sums and than si - cnt + 1, i = si–si–cnt."

P231C is possible in

O(n)using two pointers. First Input n,k,array (a). Sort array a.You need mocc and mnum.

I didn't understand the part:

please explain, thank you.

we are restoring back the total difference, tdiff.

While your

`sort`

is in ..., the time complexity still same as the official solution:(Solution to div2 C using 2 pointers . Works in O(n) apart from sorting

code

Can you explain how you used two pointers in this?I solved using the author's solution

A much easier to solution D is to simply check for each dimension whether the observer's position for that dimension is less than 0, more than x1/y1/z1, or in between.

For example, if observer's x is less than 0, the a5 side will be visible no matter the observer's y and z, if the observer's x is more than x1 (where the cube ends) then the a6 side will be visible no matter the observer's y and z. Do the same for the other 2 dimensions and you get the sum. No angle calculations necessary.

Program: https://codeforces.com/problemset/submission/231/53533694