Problem 0: Richard has been infected with bovine spongiform encephalopathy. Help Kevin understand what he's saying!

## Div 2 A

**Hint:** Just do it! But if you're having trouble, try doing your computations using only integers.

This problem is straightforward implementation---just code what's described in the problem statement. However, floating point error is one place where you can trip up. Avoid it by rounding (adding 0.5 before casting to int), or by doing all calculations with integers. The latter is possible since 250 always divides the maximum point value of a problem. Thus when we rewrite our formula for score as , it is easy to check that we only have integers as intermediate values.

**Code:** Coming soon!

## Div 2 B

**Hint:** Try thinking about a sorted list of cowbells. What do we do with the largest ones?

Intuitively, we want to use as many boxes as we can and put the largest cowbells by themselves. Then, we want to pair the leftover cowbells so that the largest sum of a pair is minimized.This leads to the following greedy algorithm:

First, if *k* ≥ *n*, then each cowbell can go into its own box, so our answer is *max*(*s*_{1}, *s*_{2}, ..., *s*_{n}). Otherwise, we can have at most 2*k* - *n* boxes that contain one cowbell. So we sort the cowbells by size and put the 2*k* - *n* largest into their own boxes. For the remaining *n* - (2*k* - *n*) = 2(*n* - *k*) cowbells, we pair the *i* th largest cowbell with the (2(*n* - *k*) - *i* + 1) th largest. In other words, we match the smallest remaining cowbell with the largest, the second smallest with the second largest, and so on. Given these pairings, we can loop through them to find the largest box we'll need. The complexity this algorithm is due to sorting.

To prove that this greedy works, think about the cowbell the the largest one gets paired with. If it's not the smallest, we can perform a swap so that the largest cowbell is paired with the smallest and not make our answer worse. After we've paired the largest cowbell, we can apply the same logic to the second largest, third largest, etc. until we're done.

**Code:** Coming soon!

## Div 2 C/Div 1 A

**Hint:** Is there any easy way to describe the longest alternating subsequence of a string? What happens at the endpoints of the substring that we flip?

Imagine compressing each contiguous block of the same character into a single character. For example, the first sample case 10000011 gets mapped to 101. Then the longest alternating subsequence of our string is equal to the length of our compressed string. So what does flipping a substring do to our compressed string? To answer this, we can think about flipping a substring as flipping two (possibly empty) prefixes. As an example, consider the string 10000011. Flipping the bolded substring 100 **00** 011 is equivalent to flipping the two bolded prefixes **100**00011 and **10000**.

For the most part, flipping the prefix of a string also flips the corresponding portion of the compressed string. The interesting case occurs at the endpoint of the prefix. Here, we have two possibilities: the two characters on either side of the endpoint are the same or different. If they are the same (00 or 11), then flipping this prefix adds an extra character into our compressed string. If they are different (01 or 10), we merge two characters in our compressed string. These increase and decrease, respectively, the length of the longest alternating subsequence by one. There is actually one more case that we left out: when the endpoint of our prefix is also an endpoint of the string. Then it is easy to check that the length of the longest alternating subsequence doesn't change.

With these observations, we see that we want to flip prefixes that end between 00 or 11 substrings. Each such substring allows us to increase our result by one, up to a maximum of two, since we only have two flips. If there exist no such substrings that we can flip, we can always flip the entire string and have our result stay the same. Thus our answer is the length of the initial longest alternating subsequence plus .

A very easy way to even simplify the above is to notice that if the initial longest alternating subsequence has length *len* - 2, then there will definitely be two 00 or 11 substrings. If it has length *n* - 1, then it has exactly one 00 or 11 substring. So our answer can be seen as the even easier .

**Code:** Coming soon!

## Div 2 D/Div 1 B

**Hint:** First there are special cases *k* = 0 and *k* = 1. After clearing these out, think about the following: given the value of *f*(*n*) for some *n*, how many other values of *f* can we find?

We first have the degenerate cases where *k* = 0 and *k* = 1. If *k* = 0, then the functional equaton is equivalent to *f*(0) = 0. Therefore, *p*^{p - 1} functions satisfy this, because the values *f*(1), *f*(2), ..., *f*(*p* - 1) can be anything in {0, 1, 2, ..., *p* - 1}.

If *k* = 1, then the equation is just *f*(*x*) = *f*(*x*). Therefore *p*^{p} functions satisfy this, because the values *f*(0), *f*(1), *f*(2), ..., *f*(*p* - 1) can be anything in {0, 1, 2, ..., *p* - 1}.

Now assume that *k* ≥ 2, and let *m* be the least positive integer such that This is called the \emph{order} of First, plug in *x* = 0 to find that as *p* is prime, and . Now for some integer , choose a value for *f*(*n*). Given this value, we can easily show that just by plugging in *x* = *k*^{i - 1}*n* into the functional equation and using induction. Note that the numbers *n*, *kn*, *k*^{2}*n*, ..., *k*^{m - 1}*n* are distinct , since *m* is the smallest number such that . Therefore, if we choose the value of *f*(*n*), we get the value of *m* numbers (). Therefore, if we choose *f*(*n*) of integers *n*, we can get the value of all *p* - 1 nonzero integers. Since *f*(*n*) can be chosen in *p* ways for each of the integers, the answer is .

Another way to think about this idea is to view each integer from 0 to *p* - 1 as a vertex in a graph, where *n* is connected to for every integer *i*. If we fix the value of *f*(*n*) for some *n*, then *f* also becomes fixed for all other vertices in its connected component. Thus our answer is *p* raised to the the number of connected components in the graph.

**Code:** Coming soon!

**Challenge**: How quickly can we find *m*? For this problem we let *O*(*p*) pass, but faster methods certainly exist. (This is a very open-ended question.)

## Div 2 E/Div 1 C

**Hint:** Is there a way to determine the winner of a game with many piles but looking at only one pile at a time?

We'll use the concepts of Grundy numbers and Sprague-Grundy's Theorem in this solution. The idea is that every game state can be assigned an integer number, and if there are many piles of a game, then the value assigned to that total game state is the xor of the values of each pile individually. The Grundy number of a state is the minimum number that is not achieved among any state that the state can move to.

Given this brief intro (which you can read more about many places), we have to separate the problem into 2 cases, *k* even and odd. Let *f*(*n*) denote the Grundy number of a pile of size *n*. By definition *f*(0) = 0.

If *k* is even, then when you split the pile of size 2*n* into *k* piles of size *n*, the resulting Grundy number of that state is as *k* is even. Given this, it is easy to compute that *f*(0) = 0, *f*(1) = 1, *f*(2) = 2, *f*(3) = 0, *f*(4) = 1. Now I will show by induction that for *n* ≥ 2, *f*(2*n* - 1) = 0, *f*(2*n*) = 1. The base cases are clear. For *f*(2*n* - 1), the only state that can be moved to from here is that with 2*n* - 2 cows. By induction, *f*(2*n* - 2) = 1 > 0, so *f*(2*n* - 1) = 0. On the other hand, for 2*n*, removing one stone gets to a state with 2*n* - 1 stones, with Grundy number *f*(2*n* - 1) = 0 by induction. Using the second operation gives a Grundy number of 0 as explained above, so the smallest positive integer not achieveable is 1, so *f*(2*n*) = 1.

The case where *k* is odd is similar but requires more work. Let's look at the splitting operation first. This time, from a pile of size 2*n* we can move to *k* piles of size *n*, with Grundy number as *k* is odd. So from 2*n* we can achieve the Grundy numbers *f*(2*n* - 1) and *f*(*n*). Using this discussion, we can easily compute the first few Grundy numbers. *f*(0) = 0, *f*(1) = 1, *f*(2) = 0, *f*(3) = 1, *f*(4) = 2, *f*(5) = 0. I'll prove that for *n* ≥ 2, *f*(2*n*) > 0, *f*(2*n* + 1) = 0 by induction. The base cases are clear. Now, for *n* ≥ 3, since a pile of size 2*n* + 1 can only move to a pile of size 2*n*, which by induction has Grundy number *f*(2*n*) > 0, *f*(2*n* + 1) = 0. Similarly, because from a pile of size 2*n*, you can move to a pile of size 2*n* - 1, which has Grundy number *f*(2*n* - 1) = 0, *f*(2*n*) > 0. Now computing the general Grundy number *f*(*n*) for any *n* is easy. If *n* ≤ 4, we have them precomputed. If *n* is odd and *n* > 4, *f*(*n*) = 0. In *n* is even and *n* ≥ 6, then *f*(*n*) is the minimal excludent of *f*(*n* - 1) = 0 and *f*(*n* / 2) (because *n* - 1 is odd and ≥ 5, so *f*(*n* - 1) = 0.) We can do this recursively,

The complexity is *O*(*n*) in the *k* even case and in the *k* odd case.

**Code:** Coming soon!

## Div 1 D

**Hint:** It seems like this would be *O*(*n*^{3}) because of triples of lines. Can you reduce that with some geometric observations? Think of results from Euclidean geometry relating to 4 points lying on the same circle.

First, we will prove a lemma, known as the Simson Line:

**Lemma:**

Given points *A*, *B*, *C*, *P* in the plane with *D*, *E*, *F* on lines *BC*, *CA*, and *AB*, respectively such that , then *P* lies on the circumcircle of if and only if *D*, *E*, and *F* are collinear.

**Proof:**

Assume that the points are configured as shown, and other other configurations follow similarly. Recall that a quadrilateral *ABCP* is cyclic if and only if . Note that this implies that a quadrilateral with two opposite right angles is cyclic, so in particular quadrilaterals *AEPF*, *BFPD*, *CDPE* are cyclic. Because

we get that *ABPC* is cyclic if and only if , if and only if .

Now note that (again since *BFPD* is cyclic) and , so if and only if , if and only if *ABPC* is cyclic. Thus the lemma is proven.

This lemma provides us with an efficient way to test if the circumcircle of the triangle formed by three lines in the plane passes through the origin. Specifically, for a line , let *X*_{i} be the projection of the origin onto . Then form an original triangle if and only if *X*_{i}, *X*_{j}, and *X*_{k} are collinear. Thus the problem reduces to finding the number of triples *i* < *j* < *k* with *X*_{i}, *X*_{j}, *X*_{k} collinear. The points *X*_{i} are all distinct, except that possibly two lines may pass through the origin, so we can have up to two points *X*_{i} at the origin.

Let us first solve the problem in the case that all points *X*_{i} are distinct. In this case, consider each *i*, and store for each slope how many points *X*_{j} with *i* < *j* the line *X*_{i} *X*_{j} has this slope. This storage can be done in *O*(1) or , depending on how hashing is done. Note also that we must consider a vertical line to have the valid slope . If are the number of points corresponding to the distinct slopes through *X*_{i} (for points *X*_{j} with *i* < *j*), then for *X*_{i} we add to the total count

If the *X*_{i} are not distinct, we only encounter an issue in the above solutions when we consider the slope through points *X*_{i} and *X*_{j} where *X*_{i} = *X*_{j}. In this case, for any third *k*, *X*_{i}, *X*_{j}, and *X*_{k} are collinear. So when considering slopes from *X*_{i} in the original algorithm, we simply run the algorithm on all slopes other than the one through *X*_{j}, and simply add *n* - 2 to the count afterwards to account for the *n* - 2 points which are collinear with *X*_{i} and *X*_{j}.

Running the above, we get an algorithm that runs in *O*(*n*^{2}) or .

**Code:** Coming soon!

## Div 1 E

**Hint:** What is a necessary and sufficient condition for Kevin to be able to pave paths so that each edge is incident to an odd number of them? Does this problem remind you of constructing a minimum spanning tree?

We represent this problem on a graph with pastures as vertices and paths as edges. Call a paving where each vertex is incident to an odd number of paved edges an \emph{odd paving}. We start with a lemma about such pavings:

A connected graph has an odd paving if and only if it has an even number of vertices.

For connected graphs with even numbers of vertices, we can prove this observation by considering a spanning tree of the graph. To construct an odd paving, start from the leaves of the tree and greedily pave edges so that each vertex but the root is incident to an odd number of paved edges. Now consider the graph consisting only of paved edges. Since the sum of all vertex degrees in this graph must be even, it follows that the root is also incident to an odd number of paved edges, so the paving is odd.

Now we prove that no odd paving exists in the case of an odd number of vertices. Suppose for the sake of contradiction that one existed. Then the sum of the vertex degrees in the graph consisting only of paved edges would be odd, which is impossible. Thus no odd paving exists for graphs with odd numbers of vertices.

Note that this observation turns the degree condition into a condition on the parity of connected component sizes. We finish the problem using this equivalent condition.

Suppose we only want to solve this problem once, after all *m* edges are added. Then we can use Kruskal's algorithm to build a minimum spanning forest by adding edges in order of increasing length. We stop once each tree in the forest contains an even number of vertices, since the graph now satisfies the conditions of the lemma. If there are still odd-sized components by the time we add all the edges, then no odd paving exists. This algorithm, however, runs in per query, which is too slow if we want to answer after adding each edge.

To speed things up, we maintain the ending position of our version of Kruskal's as we add edges online. We do this using a data structure called a link-cut tree. This data structure allows us to add and delete edges from a forest while handling path and connectivity queries. All of these operations take only time per operation. (A path query asks for something like maximum-weight edge on the path between *u* and *v*; a connectivity query asks if *u* and *v* are connected.)

First, let's look at how we can solve the online minimum spanning tree problem with a link-cut tree. We augment our data structure to support finding the maximum-weight edge on the path between vertices *u* and *v* in . Adding an edge then works as follows: If *u* and *v* are not connected, connect *u* and *v*; otherwise, if the new edge is cheaper, delete the maximum-weight edge on the path between *u* and *v* and add the new edge. To make implementation easier, we can represent edges as vertices in the link-cut tree. For example, if *u* and *v* are connected, in the link-cut tree they would be connected as *u*--*e*--*v*, where *e* is a vertex representing edge *u*--*v*.

We solve our original problem with a similar idea. Note that the end state of our variation on Kruskal's is a minimum spanning forest after adding *k* edges. (We no longer care about the edges that are longer than the longest of these *k* edges, since the answer is monotonically decreasing---more edges never hurt.) So when we add another edge to the forest, we can use the online minimum spanning tree idea to get the minimum spanning forest that uses the old cheapest *k* edges and our new edge. Note that inserting the edge never increases the number of odd components: even linked to even is even, even linked to odd is odd, odd linked to odd is even.

Now, pretend that we got this arrangement by running Kruskal's algorithm, adding the edges one-by-one. We can "roll back" the steps of the algorithm by deleting the longest edge until deleting another edge would give us an odd-sized component. (If we started with an odd-sized component, we don't delete anything.) This gives us an ending position for our Kruskal-like algorithm that uses a minimal number of edges so that all components have even size---we're ready to add another edge. ("But wait a minute!" you may say. "What if edges have the same weight?" In this case, if we can't remove one of possibly many longest edges, then we can't lower our answer anyway, so we stop.)

Note that all of this takes amortized time per edge. The path queries and the insertion of the new edge involve a constant number of link-cut tree operations. To know which edge to delete, the set of edges currently in the forest can be stored easily in an STL set sorted by length. When adding an edge, we also pay for the cost of deleting that edge, so the "rolling back" phase gets accounted for. Therefore, this algorithm runs in .

You may have noticed that executing this algorithm involves checking the size of a connected component in the link-cut tree. This is a detail that needs to be resolved carefully, since link-cut trees usually only handle path operations, not operations that involve subtrees. Here, we stop treating link-cut trees as a black box. (If you aren't familiar with the data structure, you should read about it at https://courses.csail.mit.edu/6.851/spring12/scribe/L19.pdf.) At each vertex, we track the size of its virtual subtree, as well as the sum of the real subtree sizes of its non-preferred children. We can update these values while linking and exposing (a.k.a. accessing), allowing us to perform root-change operations while keeping real subtree sizes. To get the size of a component, we just query for the real subtree size of the root.

Since the implementation of this algorithm can be rather challenging, here is a link to a documented version of my code:

**Code:** Coming soon!