Top Coder All Tutorials
Top Coder All Tutorials
By lbackstrom — topcoder member
Discuss this article in the forums
Introduction
The first step towards an understanding of why the study and knowledge of algorithms are so important is to define exactly what we
mean by an algorithm. According to the popular algorithms textbook Introduction to Algorithms (Second Edition by Thomas H. Cormen,
Charles E. Leiserson, Ronald L. Rivest, Clifford Stein), "an algorithm is any welldefined computational procedure that takes some value,
or set of values, as input and produces some value, or set of values as output." In other words, algorithms are like road maps for
accomplishing a given, welldefined task. So, a chunk of code that calculates the terms of the Fibonacci sequence is an implementation
of a particular algorithm. Even a simple function for adding two numbers is an algorithm in a sense, albeit a simple one.
Some algorithms, like those that compute the Fibonacci sequences, are intuitive and may be innately embedded into our logical
thinking and problem solving skills. However, for most of us, complex algorithms are best studied so we can use them as building
blocks for more efficient logical problem solving in the future. In fact, you may be surprised to learn just how many complex algorithms
people use every day when they check their email or listen to music on their computers. This article will introduce some basic ideas
related to the analysis of algorithms, and then put these into practice with a few examples illustrating why it is important to know about
algorithms.
Runtime Analysis
One of the most important aspects of an algorithm is how fast it is. It is often easy to come up with an algorithm to solve a problem,
but if the algorithm is too slow, it’s back to the drawing board. Since the exact speed of an algorithm depends on where the algorithm is
run, as well as the exact details of its implementation, computer scientists typically talk about the runtime relative to the size of the
input. For example, if the input consists of N integers, an algorithm might have a runtime proportional to N2, represented as O(N2).
This means that if you were to run an implementation of the algorithm on your computer with an input of size N, it would take C*N2
seconds, where C is some constant that doesn’t change with the size of the input.
However, the execution time of many complex algorithms can vary due to factors other than the size of the input. For example, a
sorting algorithm may run much faster when given a set of integers that are already sorted than it would when given the same set of
integers in a random order. As a result, you often hear people talk about the worstcase runtime, or the averagecase runtime. The
worstcase runtime is how long it would take for the algorithm to run if it were given the most insidious of all possible inputs. The
averagecase runtime is the average of how long it would take the algorithm to run if it were given all possible inputs. Of the two, the
worstcase is often easier to reason about, and therefore is more frequently used as a benchmark for a given algorithm. The process of
determining the worstcase and averagecase runtimes for a given algorithm can be tricky, since it is usually impossible to run an
algorithm on all possible inputs. There are many good online resources that can help you in estimating these values.
Approximate completion time for algorithms, N = 100
O(Log(N)) 107 seconds
O(N) 106 seconds
O(N*Log(N)) 105 seconds
O(N2) 104 seconds
O(N6) 3 minutes
O(2N) 1014 years.
O(N!) 10142 years.
Sorting
Sorting provides a good example of an algorithm that is very frequently used by computer scientists. The simplest way to sort a group
of items is to start by removing the smallest item from the group, and put it first. Then remove the next smallest, and put it next and
so on. Unfortunately, this algorithm is O(N2), meaning that the amount of time it takes is proportional to the number of items squared.
If you had to sort a billion things, this algorithm would take around 1018 operations. To put this in perspective, a desktop PC can do a
little bit over 109 operations per second, and would take years to finish sorting a billion things this way.
Luckily, there are a number of better algorithms (quicksort, heapsort and mergesort, for example) that have been devised over the
years, many of which have a runtime of O(N * Log(N)). This brings the number of operations required to sort a billion items down to a
reasonable number that even a cheap desktop could perform. Instead of a billion squared operations (1018) these algorithms require
only about 10 billion operations (1010), a factor of 100 million faster.
Shortest Path
Algorithms for finding the shortest path from one point to another have been researched for years. Applications abound, but lets keep
things simple by saying we want to find the shortest path from point A to point B in a city with just a few streets and intersections.
There are quite a few different algorithms that have been developed to solve such problems, all with different benefits and drawbacks.
Before we delve into them though, lets consider how long a naive algorithm – one that tries every conceivable option – would take to
run. If the algorithm considered every possible path from A to B (that didn’t go in circles), it would not finish in our lifetimes, even if A
and B were both in a small town. The runtime of this algorithm is exponential in the size of the input, meaning that it is O(CN) for some
C. Even for small values of C, CN becomes astronomical when N gets even moderately large.
One of the fastest algorithms for solving this problem has a runtime of O(E*V*Log(V)), where E is the number of road segments, and V
is the number of intersections. To put this in perspective, the algorithm would take about 2 seconds to find the shortest path in a city
with 10,000 intersections, and 20,000 road segments (there are usually about 2 road segments per intersection). The algorithm, known
as Djikstra’s Algorithm, is fairly complex, and requires the use of a data structure known as a priority queue. In some applications,
however, even this runtime is too slow (consider finding the shortest path from New York City to San Francisco – there are millions of
intersections in the US), and programmers try to do better by using what are known as heuristics. A heuristic is an approximation of
something that is relevant to the problem, and is often computed by an algorithm of its own. In the shortest path problem, for
example, it is useful to know approximately how far a point is from the destination. Knowing this allows for the development of faster
algorithms (such as A*, an algorithm that can sometimes run significantly faster than Djikstra’s algorithm) and so programmers come
up with heuristics to approximate this value. Doing so does not always improve the runtime of the algorithm in the worst case, but it
does make the algorithm faster in most realworld applications.
Approximate algorithms
Sometimes, however, even the most advanced algorithm, with the most advanced heuristics, on the fastest computers is too slow. In
this case, sacrifices must be made that relate to the correctness of the result. Rather than trying to get the shortest path, a
programmer might be satisfied to find a path that is at most 10% longer than the shortest path.
In fact, there are quite a few important problems for which the bestknown algorithm that produces an optimal answer is insufficiently
slow for most purposes. The most famous group of these problems is called NP, which stands for nondeterministic polynomial (don’t
worry about what that means). When a problem is said to be NPcomplete or NPhard, it mean no one knows a good way to solve them
optimally. Furthermore, if someone did figure out an efficient algorithm for one NPcomplete problem, that algorithm would be
applicable to all NPcomplete problems.
A good example of an NPhard problem is the famous traveling salesman problem. A salesman wants to visit N cities, and he knows
how long it takes to get from each city to each other city. The question is "how fast can he visit all of the cities?" Since the fastest
known algorithm for solving this problem is too slow – and many believe this will always be true – programmers look for sufficiently fast
algorithms that give good, but not optimal solutions.
Random Algorithms
Yet another approach to some problems is to randomize an algorithm is some way. While doing so does not improve the algorithm in
the worst case, it often makes very good algorithms in the average case. Quicksort is a good example of an algorithm where
randomization is often used. In the worst case, quicksort sorts a group of items in O(N2), where N is the number of items. If
randomization is incorporated into the algorithm, however, the chances of the worst case actually occurring become diminishingly small,
and on average, quicksort has a runtime of O(N*Log(N)). Other algorithms guarantee a runtime of O(N*Log(N)), even in the worst
case, but they are slower in the average case. Even though both algorithms have a runtime proportional to N*Log(N), quicksort has a
smaller constant factor – that is it requires C*N*Log(N) operations, while other algorithms require more like 2*C*N*Log(N) operations.
Another algorithm that uses random numbers finds the median of a group of numbers with an average runtime of O(N). This is a
significant improvement over sorting the numbers and taking the middle one, which takes O(N*Log(N)). Furthermore, while
deterministic (nonrandom) algorithms exist for finding the median with a runtime of O(N), the random algorithm is attractively simple,
and often faster than the deterministic algorithms.
The basic idea of the median algorithm is to pick one of the numbers in the group at random, and count how many of the numbers in
the group are less than it. Lets say there are N numbers, and K of them are less than or equal to the number we picked at random. If K
is less than half of N, then we know that the median is the (N/2K) th number that is greater than the random number we picked, so
we discard the K numbers less than or equal to the random number. Now, we want to find the (N/2K) th smallest number, instead of
the median. The algorithm is the same though, and we simply pick another number at random, and repeat the above steps.
Compression
Another class of algorithm deals with situations such as data compression. This type of algorithm does not have an expected output
(like a sorting algorithm), but instead tries to optimize some other criteria. In the case of data compression, the algorithm (LZW, for
instance) tries to make the data use as few bytes as possible, in such a way that it can be decompressed to its original form. In some
cases, this type of algorithm will use the same techniques as other algorithms, resulting in output that is good, but potentially sub
optimal. JPG and MP3 compression, for example, both compress data in a way that makes the final result somewhat lower quality than
the original, but they create much smaller files. MP3 compression does not retain every feature of the original song file, but it attempts
to maintain enough of the details to capture most of the quality, while at the same time ensuring the significantly reduced file size that
we all know and love. The JPG image file format follows the same principle, but the details are significantly different since the goal is
image rather than audio compression.
The Importance of Knowing Algorithms
As a computer scientist, it is important to understand all of these types of algorithms so that one can use them properly. If you are
working on an important piece of software, you will likely need to be able to estimate how fast it is going to run. Such an estimate will
be less accurate without an understanding of runtime analysis. Furthermore, you need to understand the details of the algorithms
involved so that you’ll be able to predict if there are special cases in which the software won’t work quickly, or if it will produce
unacceptable results.
Of course, there are often times when you’ll run across a problem that has not been previously studied. In these cases, you have to
come up with a new algorithm, or apply an old algorithm in a new way. The more you know about algorithms in this case, the better
your chances are of finding a good way to solve the problem. In many cases, a new problem can be reduced to an old problem without
too much effort, but you will need to have a fundamental understanding of the old problem in order to do this.
As an example of this, lets consider what a switch does on the Internet. A switch has N cables plugged into it, and receives packets of
data coming in from the cables. The switch has to first analyze the packets, and then send them back out on the correct cables. A
switch, like a computer, is run by a clock with discrete steps – the packets are send out at discrete intervals, rather than continuously.
In a fast switch, we want to send out as many packets as possible during each interval so they don’t stack up and get dropped. The
goal of the algorithm we want to develop is to send out as many packets as possible during each interval, and also to send them out so
that the ones that arrived earlier get sent out earlier. In this case it turns out that an algorithm for a problem that is known as "stable
matching" is directly applicable to our problem, though at first glance this relationship seems unlikely. Only through preexisting
algorithmic knowledge and understanding can such a relationship be discovered.
More Realworld Examples
Other examples of realworld problems with solutions requiring advanced algorithms abound. Almost everything that you do with a
computer relies in some way on an algorithm that someone has worked very hard to figure out. Even the simplest application on a
modern computer would not be possible without algorithms being utilized behind the scenes to manage memory and load data from the
hard drive.
There are dozens of applications of complicated algorithms, but I’m going to discuss two problems that require the same skills as some
past TopCoder problems. The first is known as the maximum flow problem, and the second is related to dynamic programming, a
technique that often solves seemingly impossible problems in blazing speed.
Maximum Flow
The maximum flow problem has to do with determining the best way to get some sort of stuff from one place to another, through a
network of some sort. In more concrete terms, the problem first arose in relation to the rail networks of the Soviet Union, during the
1950′s. The US wanted to know how quickly the Soviet Union could get supplies through its rail network to its satellite states in Eastern
Europe.
In addition, the US wanted to know which rails it could destroy most easily to cut off the satellite states from the rest of the Soviet
Union. It turned out that these two problems were closely related, and that solving the max flow problem also solves the min cut
problem of figuring out the cheapest way to cut off the Soviet Union from its satellites.
The first efficient algorithm for finding the maximum flow was conceived by two Computer Scientists, named Ford and Fulkerson. The
algorithm was subsequently named the FordFulkerson algorithm, and is one of the more famous algorithms in computer science. In the
last 50 years, a number of improvements have been made to the FordFulkerson algorithm to make it faster, some of which are
dauntingly complex.
Since the problem was first posed, many additional applications have been discovered. The algorithm has obvious relevance to the
Internet, where getting as much data as possible from one point to another is important. It also comes up in many business settings,
and is an important part of operations research. For example, if you have N employees and N jobs that need to be done, but not every
employee can do every job, the max flow algorithm will tell you how to assign your N employees to jobs in such a way that every job
gets done, provided that’s possible. Graduation, from SRM 200, is a good example of a TopCoder problem that lends itself to a solution
using max flow.
Sequence comparison
Many coders go their entire careers without ever having to implement an algorithm that uses dynamic programming. However, dynamic
programming pops up in a number of important algorithms. One algorithm that most programmers have probably used, even though
they may not have known it, finds differences between two sequences. More specifically, it calculates the minimum number of
insertions, deletions, and edits required to transform sequence A into sequence B.
For example, lets consider two sequences of letters, "AABAA" and "AAAB". To transform the first sequence into the second, the simplest
thing to do is delete the B in the middle, and change the final A into a B. This algorithm has many applications, including some DNA
problems and plagiarism detection. However, the form in which many programmers use it is when comparing two versions of the same
source code file. If the elements of the sequence are lines in the file, then this algorithm can tell a programmer which lines of code were
removed, which ones were inserted, and which ones were modified to get from one version to the next.
Without dynamic programming, we would have to consider a – you guessed it – exponential number of transformations to get from one
sequence to the other. As it is, however, dynamic programming makes for an algorithm with a runtime of only O(N*M), where N and M
are the numbers of elements in the two sequences.
Conclusion
The different algorithms that people study are as varied as the problems that they solve. However, chances are good that the problem
you are trying to solve is similar to another problem in some respects. By developing a good understanding of a large range of
algorithms, you will be able to choose the right one for a problem and apply it properly. Furthermore, solving problems like those found
in TopCoder’s competitions will help you to hone your skills in this respect. Many of the problems, though they may not seem realistic,
require the same set of algorithmic knowledge that comes up every day in the real world.
How to Dissect a Topcoder Problem Statement
By antimatter — topcoder member
Discuss this article in the forums
How many times has this happened to you: you register for the SRM, go into your assigned room when the system tells you to, and
when the match starts, you open the 250… and find it incomprehensible.
Maybe it’s never happened to you. You may be lucky, or you may already be extremely skilled. However, many experienced competitors
(yes, reds too) might end up just staring at a problem for a long time. This is a pretty serious issue. How can you solve the problem if
you have no idea what it’s asking you to do?
Fortunately, topcoder problem statements are formatted in a very specific way.
Knowing your way around the various pieces will go a long way towards helping you understand what the problem is saying.
The Parts of a Problem Statement
Let’s look at the composition of a typical topcoder problem statement. First off is the introduction. Usually, a problem will be led off with
a highlevel description of a situation. This description may tie into reallife ideas and topics or it may just be a completely fictional
story, serving only as some sort of context. For many problems the backstory itself is not particularly important in understanding the
actual problem at hand.
Next comes the definition section. It gives you the skeleton of the solution you need to write: class name, method name, arguments,
and return type, followed by the complete method signature. At minimum, you will need to declare a class with the given name,
containing a method which conforms to the given method signature. The syntax given will always be correct for your chosen language.
Sometimes notes follow the method definition. They tend to be important reminders of things that you should pay attention to but
might have missed, or they can also be things that are helpful background knowledge that you might not know beforehand. If the notes
section appears, you should make sure to read it – usually the information contained within is extremely important.
The constraints section is arguably the most important. It lists specific constraints on the input variables. This lets you know crucial
details such as how much memory to allocate or how efficient your algorithm will have to be.
Finally, a set of examples is provided. These give sample inputs against which you can test your program. The given parameters will be
in the correct order, followed by the expected return value and, optionally, an explanation of the test case.
Introduction
The problem statement usually begins by motivating the problem. It gives a situation or context for the problem, before diving into
the gory details. This is usually irrelevant to solving the problem, so ignore it if necessary. In some cases, the motivation can
cause serious ambiguities if it is treated as binding – see MatchMaking (SRM 203 Div I Easy / Div II Medium). Also note that for
some simple problems, the initial context may be left out.
The ordering of the rest of this section varies greatly from problem to problem, based on the writing style of the problem author.
There will be a description of what you need to do, in highlevel terms. Take, for example, UserName (SRM 203, Div 2 easy). What
the problem is asking for you to do is to find the first variant of a given username that is not already taken. Note that the problem
has not yet said anything about variable names or types, or input formats.
There will also be a lowlevel description of the input. At the bare minimum, the types and variable names of the inputs will be
given to you, as well as what they correspond to and what they mean. Sometimes much more information about input formats will
be given; this typically occurs in more complicated problems.
Sometimes, even more detailed background information needs to be provided. That is also typically given here, or sometimes in
the Notes section.
The Definition
This is a very barebones description of what topcoder wants you to submit. It gives the class name, the method name to create
inside that class, the parameters it should take, the return value, and a method signature. As mentioned before, the basic form of
a submitted solution is to create a class containing a method with the required signature. Make sure that the class is declared
public if not using C++, and make sure to declare the method public also.
Notes and Constraints
Notes don’t always appear. If they do, READ THEM! Typically they will highlight issues that may have come up during testing, or
they may provide background information that you may not have known beforehand. The constraints section gives a list of
constraints on the input variables. These include constraints on sizes of strings and arrays, or allowed characters, or values of
numbers. These will be checked automatically, so there is no need to worry about writing code to check for these cases.
Be careful of the constraints. Sometimes they may rule out certain algorithms, or make it possible for simpler but less efficient
algorithms to run in time. There can be a very big difference between an input of 50 numbers and an input of 5, both in terms of
solutions that will end up passing, and in terms of ease of coding.
Examples
These are a list of sample test cases to test your program against. It gives the inputs (in the correct order) and then the expected
return value, and sometimes an annotation below, to explain the case further if necessary.
It goes without saying that you should test your code against all of the examples, at the very least. There may be tricky cases,
large cases, or corner cases that you have not considered when writing the solution; fixing issues before you submit is infinitely
preferable to having your solution challenged or having it fail during system testing.
The examples are not always comprehensive! Be aware of this. For some problems, passing the examples is almost the same as
passing every test case.
For others, however, they may intentionally (or not) leave out some test case that you should be aware of. If you are not
completely sure that your code is correct, test extensively, and try to come up with your own test cases as well. You may even be
able to use them in the challenge phase.
Solving a problem
Now we’ll walk through a simple problem and dissect it, bit by bit.
Have a look at BettingMoney, the SRM 191 Division 2 Easy. First we identify the parts of this problem. In the statement itself, we first
have the situation behind the problem – gambling. Then we have a little bit of background information about the betting itself. Then, we
have a description of the input – data types, variable names, and what they represent. After this we have the task: to determine what
the net gain is for the day and return the amount in cents.
Also note the two explanatory paragraphs at the end; the first provides an example of the input format and types, and the second gives
a completely worked example, which should be extremely helpful to your understanding.
The definition section is uninteresting, but it is there for completeness’ sake.
The notes for this problem are fairly comprehensive. In terms of background information, you might not know that there are 100 cents
in a dollar. And in terms of clarification, there is explicit confirmation that the return value may in fact be negative, and that the margin
of victory (the variable finalResult) is all that matters when deciding which payoff to make.
The constraints are fairly straightforward. The input arrays will contain the same number of elements, between 1 and 50, inclusive. (50
is a longstanding topcoder tradition for input sizes). finalResult will be between 0 and that same size minus one (which means, if you
give it a little thought, that someone will win their bet). Each element of each array will be between 0 and 5000, inclusive. This is most
likely to make sure that integer arithmetic will do the job just fine.
Finally, there’s the examples section. Often, the problem statement section will contain an annotated example case, which will become
example case 0. Then there are a couple of other example cases, some with explanation and some without. Also note that one of the
examples tests for negative return values, to supplement the notes.
A More Complicated Example
Now have a look at Poetry, the SRM 170 Div 2 Hard. In this case, you may not be able to actually solve this in the time allotted. That’s
ok – the emphasis should first be on understanding what the problem says, even if you can’t code it in time.
The first section tells you immediately what you want to do – you’ll be given a poem, and you will have to determine what its rhyme
scheme is. The rest of the section clarifies what this actually means, in bottomup fashion (from simpler concepts to more complicated
ones). It defines what a legal word is and how to extract words from a poem, and then it defines what it means when two words rhyme
– that their ending patterns are equal. The concept of ending pattern is then defined. After all this, we find out what it means to have
two lines of the poem rhyme: their last words have to rhyme. Finally, (whew!) we are told how to actually construct the rhyme scheme
and in what format to return it.
This is a problem where a lot of terms need to be defined to get to the heart of things, and so all the definitions deserve at least a
couple of readthroughs, especially if you’re not sure how they all fit together.
The next section is the problem definition section, just for reference. Then there is a single note that clarifies a point that may have
been overlooked when it was stated in the problem statement itself: that blank lines will be labeled with a corresponding space in the
rhyme scheme.
The constraints are fairly standard for topcoder problems: there will be between 1 and 50 lines in the poem, and each line will contain
between 0 and 50 characters. The only allowable characters in the poem will be spaces and letters, and there will be only legal words in
poem.
Finally, there are a number of examples. Usually, problems which are trickier or which have more complex problem statements will have
more examples, to clarify at least some of the finer points of the problem statement. Again, this doesn’t mean that passing the example
cases given is equivalent to having a completely correct solution, but there is a higher chance that you can catch any bugs or trivial
mistakes if there are more examples that you know the answers to.
Try it Yourself
Listed below are a number of additional problems, grouped roughly by difficulty of comprehension. Try them for yourself in the topcoder
Arena Practice Rooms. Even if you can’t solve them, at least work on figuring out what the problem wants by breaking it down in this
manner.
Mentioned in this writeup:
SRM 203 Div 2 Easy – UserName
SRM 191 Div 2 Easy – BettingMoney
SRM 203 Div 1 Easy – MatchMaking
SRM 170 Div 2 Hard – Poetry
Similar tasks:
SRM 146 Div 2 Easy – Yahtzee
SRM 200 Div 2 Easy – NoOrderOfOperations
SRM 185 Div 2 Easy – PassingGrade
SRM 155 Div 2 Easy – Quipu
SRM 147 Div 2 Easy – CCipher
SRM 208 Div 1 Easy – TallPeople
SRM 173 Div 1 Easy – WordForm
SRM 162 Div 1 Easy – PaperFold
More challenging tasks:
SRM 197 Div 2 Hard – QuickSums
SRM 158 Div 1 Hard – Jumper
SRM 170 Div 1 Easy – RecurrenceRelation
SRM 177 Div 1 Easy – TickTick
SRM 169 Div 2 Hard – Twain
SRM 155 Div 1 Med – QuipuReader
How to Find a Solution
By Dumitru — topcoder member
Discuss this article in the forums
Introduction
Straightforward problems that don’t require a special technique
Breadth First Search (BFS)
Flood Fill
Brute Force and Backtracking
Brute Force
Backtracking
Dynamic Programming
Hard Drills
Maximum Flow
Optimal Pair Matching
Linear Programming (Simplex)
Conclusion
Introduction
With many topcoder problems, the solutions may be found instantly just by reading their descriptions. This is possible thanks to a
collection of common traits that problems with similar solutions often have. These traits serve as excellent hints for experienced
problem solvers that are able to observe them. The main focus of this article is to teach the reader to be able to observe them too.
Straightforward problems that don’t require any special technique (e.g. simulation, searching, sorting etc.)
In most cases, these problems will ask you to perform some step by step, straightforward tasks. Their constraints are not high, and
not too low. In most cases the first problems (the easiest ones) in topcoder Single Rounds Matches are of this kind. They test mostly
how fast and properly you code, and not necessarily your algorithmic skills.
Most simple problems of this type are those that ask you just to execute all steps described in the statement.
BusinessTasks – SRM 236 Div 1:
N tasks are written down in the form of a circular list, so the first task is adjacent to the last one. A number n is also given. Starting
with the first task, move clockwise (from element 1 in the list to element 2 in the list and so on), counting from 1 to n. When your
count reaches n, remove that task from the list and start counting from the next available task. Repeat this procedure until one task
remains. Return it.
For N<=1000 this problem is just a matter of coding, no special algorithm is needed – do this operation step by step until one item is
left. Usually these types of problems have a much smaller N, and so we’ll not consider cases where N is very big and for which
complicated solution may be needed. Remember that in topcoder competitions even around 100 millions sets of simple operations (i.e.
some multiplications, attributions or if statements) will run in allowed time.
This category of problems also includes those that need some simple searches.
TallPeople – SRM 208 Div 1:
A group of people stands before you arranged in rows and columns. Looking from above, they form an R by C rectangle of people. Your
job is to return 2 specific heights – the first is computed by finding the shortest person in each row, and then finding the tallest person
among them (the "tallestoftheshortest"); and the second is computed by finding the tallest person in each column, and then finding
the shortest person among them (the "shortestofthetallest").
As you see this is a really simple search problem. What you have to do is just to follow the steps described in the statement and find
those 2 needed heights. Other TC problems may ask you to sort a collection of items by respecting certain given rules. These problems
may be also included in this category, because they too are straightforward – just sort the items respecting the rules! You can do that
with a simple O(N^2) sorting algorithm, or use standard sorting algorithm that exist in your coding language. It’s just a matter of
coding.
Other example(s):
MedalTable – SRM 209 Div 1.
Breadth First Search (BFS)
Problems that use BFS usually ask to find the fewest number of steps (or the shortest path) needed to reach a certain end point (state)
from the starting one. Besides this, certain ways of passing from one point to another are offered, all of them having the same cost of 1
(sometimes it may be equal to another number). Often there is given a N x M table (formed of N lines and M columns) where certain
cells are passable and others are impassable, and the target of the problem is to find the shortest time/path needed to reach the end
point from the start one. Such tables may represent mazes, maps, cities, and other similar things. These may be considered as classical
BFS problems. Because BFS complexity is in most cases linear (sometimes quadratic, or N logN), constraints of N (or M) could be high –
even up to 1 million.
SmartWordToy – SRM 233 Div 1:
A word composed of four Latin lowercase letters is given. With a single button click you can change any letter to the previous or next
letter in alphabetical order (for example ‘c’ can be changed to ‘b’ or ‘d’). The alphabet is circular, thus ‘a’ can become ‘z’, and ‘z’ can
become ‘a’ with one click.
A collection of constraints is also given, each defining a set of forbidden words. A constraint is composed of 4 strings of letters. A word
is forbidden if each of its characters is contained in corresponding string of a single constraint, i.e. first letter is contained in the first
string, the second letter – in the second string, and so on. For example, the constraint "lf a tc e" defines the words "late", "fate", "lace"
and "face".
You should find the minimum number of button presses required to reach the word finish from the word start without passing through
forbidden words, or return 1 if this is not possible.
Problem hints:
Words can be considered as states. There are at most 26^4 different words composed of 4 letters (thus a linear complexity will
run in allowed time).
There are some ways to pass from one state to another.
The cost of passing from a state to another is always 1 (i.e. a single button click).
You need to find the minimum number of steps required to reach the end state from start state.
Everything indicates us that it’s a problem solved by the help of a BFS. Similar things can be found in any other BFS problems. Now
let’s see an interesting case of BFS problems.
CaptureThemAll – SRM 207 Div 2 (3rd problem):
Harry is playing a chess game. He has one knight, and his opponent Joe has a queen and a rook. Find the minimum number of steps
that Harry’s knight has to jump so that it captures both the queen and the rook.
Problem hints: At first sight this may seem like dynamic programming or backtracking. But as always, take a look into the text of the
statement. After a while you should observe the following things:
A table is given.
The knight can jump from one cell to some of its neighbors.
The cost of passing from a cell to another is always 1 (just one jump).
You need to find the minimum number of steps (jumps).
Given this information we can see that the problem can be easily solved by the help of BFS. Don’t get confused by the fact that
connected points are represented by unconnected cells. Think of cells as points in a graph, or states (whatever you want) – and in order
to pass from one point to another, the knight should be able to jump from the first to the second point.
Notice again that the most revealing hint about the BFS solution is the cost of 1 for any jump.
Train yourself in finding the hints of a BFS problem in following examples:
Other example(s):
RevolvingDoors – SRM 223 Div 1
WalkingHome – SRM 222 Div 1
TurntableService – SRM 219 Div 1
Flood Fill
Sometimes you may encounter problems that are solved by the help of Flood Fill, a technique that uses BFS to find all reachable points.
The thing that makes them different from BFS problems described above is that a minimum path/cost is not needed.
For example, imagine a maze where 1 represents impassable cells and 0 passable cells. You need to find all cells that are reachable
from the upperleft corner. The solution is very simple – take onebyone a visited vertex, add its unvisited neighbors to the queue of
visited vertices and proceed with the next one while the queue is still populated. Note that in most cases a DFS (Depth First Search) will
not work for such problems due to stack overflows. Better use a BFS. For inexperienced users it may seem harder to implement, but
after a little training it becomes a "piece of cake". A good example of such problem would be:
grafixMask – SRM 211 Div 1:
A 400 x 600 bitmap is given. A set of rectangles covers certain parts of this bitmap (the corners of rectangles have integer
coordinates). You need to find all contiguous uncovered areas, including their sizes.
Problem hints: What do we have here?
A map (table)
Certain points are impassable (those covered by given rectangles)
Contiguous areas need to be found
It is easy to understand that a problem with such a statement needs a Flood Fill. Usually problems using it are very easy to detect.
Brute Force and Backtracking
I have placed these 2 techniques in the same category because they are very similar. Both do the same thing – try all possible cases
(situations) and choose the best one, or count only those that are needed (depending on the problem). Practically, Backtracking is just
more advanced and optimized than Brute Force. It usually uses recursion and is applied to problems having low constraints (for
example N<=20).
Brute Force
There are many problems that can be solved by the help of a simple brute force. Note that the limits must not be high. How does a
brute force algorithm work? Actually, it tries all possible situations and selects the best one. It’s simple to construct and usually simple
to implement. If there is a problem that asks to enumerate or find all possible ways (situations) of doing a certain thing, and that
doesn’t have high limits – then it’s most probably a brute force problem.
GeneralChess – SRM 197 Div 1:
You are given some knights (at most 8), with their positions on the table (10000<=x, y<=10000). You need to find all positions to
place another one, so that it threatens all given pieces.
Problem hints: Well, this is one of the easiest examples. So which are the hints of this statement?
You need to find all possible situations (positions) that satisfy a certain rule (threatens all given pieces).
The limits are very low – only 8 knights are at most given.
It’s a common Brute Force problem’s statement. Note that x and y limits are not relevant, because you need only try all positions that
threaten one of the knights. For each of these positions see if the knight placed at that position threatens all others too.
Another interesting problem would be:
LargestCircle – SRM 212 Div 2 (3rd problem):
Given a regular square grid, with some number of squares marked, find the largest circle you can draw on the grid that does not pass
through any of the marked squares. The circle must be centered on a grid point (the corner of a square) and the radius must be an
integer. Return the radius of the circle.
The size of the grid is at most 50.
Problem hints: And again one of the most important hints is the low limit of the size of the grid – only 50. This problem is possible to
be solved with the help of the Brute Force because for each cell you can try to find the circle whose center is situated in that cell and
that respects the rules. Among all of these circles found, select the one that has the greatest radius.
Complexity analysis: there are at most 50×50 cells, a circle’s radius is an integer and can be at most 25 units, and you need a linear
time (depending on your implementation) for searching the cells situated on the border of the circle. Total complexity is low and thus
you can apply a simple Brute Force here.
Other example(s):
Cafeteria – SRM 229 Div 1
WordFind – SRM 232 Div 1
Backtracking
This technique may be used in many types of problems. Just take a look at the limits (N, M and other main parameters). They serve as
the main hint of a backtrack problem. If these are very small and you haven’t found a solution that’s easier to implement – then just
don’t waste your time on searching it and implement a straightforward backtracking solution.
Usually problems of this kind ask you to find (similarly to Brute Force):
1. Every possible configuration (subset) of items. These configurations should respect some given rules.
2. The "best" configuration (subset) that respects some given rules.
BridgeCrossing – SRM 146 Div 2 (3rd problem):
A group of people is crossing an old bridge. The bridge cannot hold more than two people at once. It is dark, so they can’t walk without
a flashlight, and they only have one flashlight! Furthermore, the time needed to cross the bridge varies among the people in the group.
When people walk together, they always walk at the speed of the slowest person. It is impossible to toss the flashlight across the
bridge, so one person always has to go back with the flashlight to the others. What is the minimum amount of time needed to get all
the people across the bridge?
There are at most 6 people.
Problem hints:
First look at the constraints – there are at most ONLY 6 people! It’s enough for generating all possible permutations, sets etc.
There are different possible ways to pass the people from one side to another and you need to find the best one.
This is of course a problem solved with a backtracking: at the beginning choose any 2 people to pass the bridge first, and after that at
each step try to pass any of those that have been left on the start side. From all these passages select the one that needs the smallest
amount of time. Note that among persons that have passed over the bridge, the one having the greatest speed should return (it’s
better than returning one having a lower speed). This fact makes the code much easier to implement. After having realized these things
– just code the solution. There may be others – but you will lose more time to find another than to code this one.
MNS – SRM 148 Div 1:
9 numbers need to be arranged in a magic number square. A magic number square is a square of numbers that is arranged such that
every row and column has the same sum. You are given 9 numbers that range from 0 to 9 inclusive. Return the number of distinct ways
that they can be arranged in a magic number square. Two magic number squares are distinct if they differ in value at one or more
positions.
Problem hints: Only 9 numbers are given at most; and every distinct way (configuration) to arrange the numbers so that they form a
magic number square should be found. These are the main properties of a Backtracking problem. If you have observed them – think
about the code. You can generate all permutations of numbers and for each of them check if it forms a magic square. If so – add it to
the answer. Note that it must be unique. A possible way to do that – is to have a list of earlier found configurations, thus for each new
magic square check if it exists in that list and if it doesn’t – add it to the answer. There will not be many distinct magic squares, thus no
additional problems will appear when applying this method.
Other example(s):
WeirdRooks – SRM 234 Div 1
Dynamic Programming
Quite a few problems are solved with the help of this technique. Knowing how to detect this type of problem can be very valuable.
However in order to do so, one has to have some experience in dynamic programming. Usually a DP problem has some main integer
variables (e.g. N) which are neither too small, nor too big – so that a usual DP complexity of N^2, N^3 etc. fits in time. Note that in the
event that N is very small (for TC problems usually less than 30) – then it is likely the problem is not a DP one. Besides that there
should exist states and one or more ways (rules) to reach one greater state from another lower one. In addition, greater states should
depend only upon lower states. What is a socalled state? It’s just a certain configuration or situation. To better understand dynamic
programming, you may want to read this article.
Let’s analyze a simple classic DP problem:
Given a list of N coins with their values (V1, V2, … ,VN), and the total sum S. Find the minimum number of coins the sum of which is S
(you can use as many coins of one type as you want), or report that it’s not possible to select coins in such a way that they sum up to
S.
Let N <= 1,000 and S <= 1,000.
Problem hints:
Two main integer variables are given (N and S). These are neither too small, nor are they too big (i.e. a complexity of N*S fits in
time).
A state can be defined as the minimum number of coins needed to reach a certain sum.
A sum (state) i depends only on lower sums (states) j (j<i).
By adding a coin to a certain sum – another greater sum is reached. This is the way to pass from one state to another.
Thus all properties of a DP problem are uncovered in this statement. Let’s see another (slightly harder) DP problem
ZigZag – 2003 TCCC Semifinals 3:
A sequence of numbers is called a zigzag sequence if the differences between successive numbers strictly alternate between positive
and negative. The first difference (if one exists) may be either positive or negative. A sequence with fewer than two elements is trivially
a zigzag sequence. Given a sequence of integers, return the length of the longest subsequence that is a zigzag sequence. A
subsequence is obtained by deleting some number of elements (possibly zero) from the original sequence, leaving the remaining
elements in their original order. Assume the sequence contains between 1 and 50 elements, inclusive.
Problem hints:
There are N numbers given (1<=N<=50), thus N isn’t too small, nor too big.
A state (i,d) can be defined as the length of the longest zigzag subsequence ending with the ith number, for which the number
before the last one is smaller than it for d=0, and bigger for d=1.
A state i (i.e. a subsequence ending with the ith number) depends only on lower states j (j<i).
By adding a number to the end of a subsequence – another bigger (greater) subsequence is created. This is the way to pass from
one state to another.
As you can see – this statement has almost the same traits (pattern) as in the previous problem. The most difficult part in identifying a
DP problem statement is observing/seeing the states with the properties described above. Once you can do that, the next step is to
construct the algorithm, which is out of the scope of this article.
Other example(s):
ChessMetric – 2003 TCCC Round 4
AvoidRoads – 2003 TCO Semifinals 4
FlowerGarden – 2004 TCCC Round 1
BadNeighbors – 2004 TCCC Round 4
Hard Drills:
Maximum Flow
In many cases it’s hard to detect a problem whose solution uses maximum flow. Often you have to create/define graphs with capacities
based on the problem statement.
Here are some signs of a Maximum Flow problem:
Take a look at the constraints, they have to be appropriate for a O(N^3) or O(N^4) solution, i.e. N shouldn’t be greater than 500
in extreme cases (usually it’s less than 100).
There should be a graph with edges having capacities given, or you should be able to define/create it from data given in the
statement.
You should be finding a maximum value of something.
Sample problem:
You are given a list of water pipes, each having a certain maximum water flow capacity. There are water pipes connected together at
their extremities.
You have to find the maximum amount of water that can flow from start junction to end junction in a unit of time.
Let N<=100.
As you can see – it’s a straightforward maximum flow problem: water pipes represent edges of the graph, their junctions – vertices;
and you have to find the maximum value of amount of water that can flow from start to end vertex in a unit of time.
Optimal Pair Matching:
These problems usually have a list of items (from a set A) for which other items (from a set B) should be assigned under some rules, so
that all (or a maximum possible number of) items from set A have to each be assigned to a certain item from set B.
Mixed:
Some problems need other techniques in addition to network flows.
Parking – SRM 236 Div 1:
N cars and M parking lots are given. They are situated on a rectangular surface (represented by a table), where certain cells are
impassable. You should find a way to assign each car to a parking lot, so that the greatest of the shortest distances from each car to its
assigned parking lot is as small as possible. Each parking lot can have at most one car assigned to it.
Problem hints: By reading this problem one can simply understand the main idea of the solution – it should be something similar to
optimal pair matching, because each car (point from a set A) should be assigned to a parking lot (point from a set B) so that all are
assigned and that there is at most one car assigned to a parking lot. Additionally, there can be cars that can’t reach certain parking lots,
thus some pairs of points (one point from A and the other from B) are not connected. However a graph should be created for optimal
pair matching. The way to make it is clear – an edge exists between a car and a parking lot if only there is a path between them, and
its cost is equal to the shortest distance needed for the car to reach the parking lot. The next step of the solution is a binary search on
the longest edge. Although it may be out of the scope of this article, I will provide a short explanation: At each step delete those edges
of the initial graph that have costs greater than a certain value C (Note that you’ll have to save the initial graph’s state in order to
repeat this step again for other C values). If it’s possible in this case to assign all the cars to parking lots – then take a smaller C, and
repeat the same operation. If not – take a greater C. After a complete binary search, the smallest C for which a complete assignment is
possible will be found. This will be the answer.
Linear Programming (Simplex)
Most of the common traits of problems solved with the help of the linear programming technique are:
You are given collection of items having different costs/weights. There is a certain quantity of each item that must be achieved.
A list of sets is given. These sets are composed of some of the available items, having certain quantities of each of them. Each set
has a certain cost.
The goal of the problem is to find an optimal combination (the cheapest one) of these sets so that the sum of quantities of each of
the items they have is exactly the one needed to achieve.
At first it may seem confusing, but let’s see an example:
Mixture – SRM 231 Div 1):
A precise mixture of a number of different chemicals, each having a certain amount, is needed. Some mixtures of chemicals may be
purchased at a certain price (the chemical components for the mixture might not be available in pure form). Each of them contains
certain amounts of some of the chemicals. You need not purchase the available mixtures in integral amounts. Hence if you purchase a
1.5 of a mixture having a price of 3 and amounts of "2 0 1", you’ll pay 4.5 and get "3 0 1.5" amounts of chemicals. Your task is to
determine the lowest price that the desired mixture can be achieved.
Problem hints:
A collection of items (chemicals).
A list of sets (available mixtures), each containing certain amounts of each of the items, and having a certain cost.
You need to find the lowest price of the desired collection of items achieved by the combination of the available sets. More than
that – you can take also nonintegral amounts of mixtures.
These are exactly the traits described above.
Conclusion
If you have found this article interesting and you have learned new things from it – train yourself on any of the problems in the
topcoder Algorithm Arena. Try hard to see the hints and determine the type of the solution by carefully reading through the problem
statement. Remember, there are still many problems that may not be included properly in any of the categories described above and
may need a different approach.
Planning an Approach to a Topcoder Problem: Part 1
By leadhyena_inran — topcoder member
Discuss this article in the forums
Planning an approach is a finicky art; it can stump the most seasoned coders as much as it stumps the newer ones, and it can be
extremely hard to put into words. It can involve many calculations and backtracks, as well as foresight, intuition, creativity, and even
dumb luck, and when these factors don’t work in concert it can inject a feeling of helplessness in any coder. Sometimes it’s this feeling
of helplessness that discourages coders from even attempting the Div I Hard. There are even coders that stop competing because they
abhor that mental enfeeblement that comes with some problems. However, if one stays diligent, the solution is never really out of the
mind’s reach. This tutorial will attempt to flesh out the concepts that will enable you to pick an approach to attack the problems with a
solid plan.
Pattern Mining and the Wrong Mindset
It is easy to fall into the trap of looking at the algorithm competition as a collection of diverse yet classifiable story problems. For those
that have done a lot of story problems, you know that there are a limited number of forms of problems (especially in classes where the
professor tends to be repetitious), and when you read a problem in a certain form, your mind says, "Oh, this is an X problem, so I find
the numbers that fit the problem and plug and chug." There are many times when this kind of pattern mining pays off; after a number
of topcoder Single Round Matches, most coders will recognize a set of common themes and practice against them, and this method of
problem attack can be successful for many matches.
However, this approach is perilous. There are times when you skim the problem statement and assume it’s of type Q, then start coding
and discover that your code passes none of the examples. That’s when you reread the problem and find out that this problem is unique
to your experience. At that point, you are paralyzed by your practice; being unable to fit any of your problem types to the problem you
are unable to proceed. You’ll see this often when there’s a really original problem that comes down the pipe, and a lot of seasoned
coders fail the problem because they are blinded by their experience.
Pattern mining encourages this kind of mindset that all of the problem concepts have been exhausted, when in reality this is impossible.
Only by unlearning what you have learned (to quote a certain wise old green midget) and by relearning the techniques of critical
thought needed to plan an approach can your rating sustainably rise.
Coding Kata
Here’s your first exercise: take any problem in the Practice Rooms that you haven’t done. Fight through it, no matter how long it takes,
and figure it out (use the editorial from the competition as a last resort). Get it to pass system tests, and then note how long you took
to solve it. Next, clear your solution out, and try to type it in again (obviously cutting and pasting will ruin the effect). Again, get it to
pass system tests. Note how long it took you to finish the second time. Then, clear it out and do the problem a third time, and again
get it to pass system tests. Record this final time.
The time it takes for your first pass is how long it takes you when you have no expectations of the problem and no approach readily in
mind. Your time on the second pass is usually the first time minus the amount of time it took you to understand the problem statement.
(Don’t be surprised at the number of bugs you’ll repeat in the second pass.) That final recorded time is your potential, for you can solve
it this fast in competition if you see the correct approach immediately after reading it. Let that number encourage you; it really is
possible to solve some of these problems this quickly, even without super fast typing ability. But what you should also learn from the
third pass is the feeling that you knew a working strategy, how the code would look, where you would tend to make the mistakes, and
so on. That’s what it feels like to have the right approach, and that feeling is your goal for future problems in competition.
In most martial arts, there’s a practice called kata where the martial artist performs a scripted series of maneuvers in order, usually
pretending to defend (or sometimes actually defending) against an onslaught of fighters, also scripted to come at the artist predictably.
At first this type of practice didn’t make any sense, because it didn’t seem realistic to the chaotic nature of battle. Furthermore it seems
to encourage the type of pattern mining mentioned in the previous section. Only after triplecoding many problems for a while can one
comprehend the true benefit of this coding kata. The kata demonstrates to its practitioners the mental experience of having a plan,
encouraging the type of discipline it takes to sit and think the problem through. This plan of attack is your approach, and it carries you
through your coding, debugging, and submission.
Approach Tactics
Now that you know what an approach feels like and what its contents are, you’ll realize that you know a lot of different types of these
approaches. Do you give them names? "Oh, I used DP (dynamic programming) on that problem." "Really, I could have done that one
greedy?" "Don’t tell me that the bruteforce solution would have passed in time." Really, the name you give an approach to a problem is
a misnomer, because you can’t classify every problem as a type like just greedy or just bruteforce. There are an infinite number of
problem types, even more solution types, and even within each solution type there are an infinite number of different variations. This
name is only a very high level summary of the actual steps it takes to get to the solution.
In some of the better match editorials there is a detailed description of one approach to solving the code. The next time you look at a
match summary, and there is a good writeup of a problem, look for the actual steps and formation of the approach. You start to notice
that there is a granularity in the steps, which suggests a method of cogitation. These grains of insight are approach tactics, or ways to
formulate your approach, transform it, redirect it, and solidify it into code that get you closer to the solution or at least point you away
from the wrong solution. When planning your approach, the idea is that you will use whatever approach tactics are at your disposal to
decide on your approach, the idea being that you are almost prewriting the code in your head before you proceed. It’s almost as if you
are convincing yourself that the code you are about to write will work.
Coders with a math background may recognize this method of thinking, because many of these approach tactics are similar to proof
writing techniques. Chess players may identify it with the use of tactics to look many moves ahead of the current one. Application
designers may already be acquainted with this method when working with design patterns. In many other problem solving domains
there is a similar parallel to this kind of taxonomy.
To practice this type of critical thinking and to decide your preferences among approach tactics, it is very useful to record the solutions
to your problems, and to write up a postSRM analysis of your own performance. Detail in words how each of your solutions work so
that others could understand and reproduce the approach if they wanted to just from your explanations. Not only will writing up your
approaches help you to understand your own thoughts while coding, but this kind of practice also allows you to critique your own
pitfalls and work on them in a constructive manner. Remember, it is difficult to improve that which you don’t understand.
Breaking Down a Problem
Let’s talk about one of the most common approach tactics: breaking down a problem. This is sometimes called topdown programming:
the idea is that your code must execute a series of steps in order, and from simple decisions decide if other steps are necessary, so
start by planning out what your main function needs before you think about how you’ll do the subfunctions. This allows you to
prototype the right functions on the fly (because you only code for what you need and no further), and also it takes your problem and
fragments it into smaller, more doable parts.
A good example of where this approach is useful is in MatArith from Round 2 of the 2002 topcoder Invitational. The problem requires
you to evaluate an expression involving matrices. You know that in order to get to the numbers you’ll need to parse them (because
they’re in String arrays) and pass those values into an evaluator, change it back into a String array and then you’re done. So you’ll need
a print function, a parse function and a new calc function. Without thinking too hard, if you imaging having all three of these functions
written already the problem could be solved in one line:
public String[] calculate(String[] A, String[] B, String[] C, String eval){
return print(calc(parse(A),parse(B),parse(C),eval));
}
The beauty of this simplest approach tactic is the guidance of your thoughts into a functional hierarchy. You have now fragmented your
work into three steps: making a parse function, a print function, and then a calc function, breaking a tough piece of code into smaller
pieces. If you break down the code fine enough, you won’t have to think hard about the simplest steps, because they’ll become atomic
(more on this below). In fact the rest of this particular problem will fall apart quickly by successive partitioning into functions that
multiply and add the matrices, and one more that reads the eval statement correctly and applies the appropriate functions.
This tactic really works well against recursive problems. The entire idea behind recursive code is that you are breaking the problem into
smaller pieces that look exactly like the original, and since you’re writing the original, you’re almost done. This approach tactic also
plays into the hands of a method of thinking about programs called functional programming. There are several articles on the net and
even a topcoder article written by radeye that talk more about this concept in depth, but the concept is that if properly fragmented, the
code will pass all variable information between functions, and no data needs to be stored between steps, which prevents the possibility
of sideeffects (unintended changes to state variables between steps in code) that are harder to debug.
Plan to Debug
Whenever you use an approach you should always have a plan to debug the code that your approach will create. This is the dark
underbelly of every approach tactic. There is always a way that a solution may fail, and by thinking ahead to the many ways it can
break, you can prevent the bugs in the code before you type them. Furthermore, if you don’t pass examples, you know where to start
looking for problems. Finally, by looking for the stress points in the code’s foundation, it becomes easier to prove to yourself that the
approach is a good one.
In the case of a topdown approach, breaking a problem down allows you to isolate sections of the code where there may be problems,
and it will allow you to group tests that break your code into sections based on the subfunction they seem to exploit the most. There is
also an advantage to breaking your code into functions when you fix a bug, because that bug is fixed in every spot where the code is
used. The alternative to this is when a coder copy/pastes sections of code into every place it is needed, making it harder to propagate a
fix and makes the fix more error prone. Also, when you look for bugs in a topdown approach, you should look for bugs inside the
functions before you look between the calls to each function. These parts make up a debugging strategy: where to look first, how to
test what you think is wrong, how to validate pieces and move on. Only after sufficient practice will a debugging strategy become more
intuitive to your method of attack.
Atomic Code
If you arrive at a section of code that you cannot break down further this is atomic code. Hopefully you know how to code each of these
sections, and these form the most common forms of atomic code. But, don’t be discouraged when you hit a kernel of the problem that
you don’t know how to code; these hardtosolve kernels are in fact what make the problem interesting, and sometimes being able to
see these in advance can make the big difference between solving the problem early with the right approach and heading down the
wrong path with the wrong approach, wasting a lot of time in the process.
The most common type of atomic code you’ll write is in the form of primitives. I’ve always been a proponent of knowing the library of
your language of choice. This is where that knowledge is of utmost importance. What better way to save yourself time is there in both
planning your approach and coding your solution when you know that a possibly difficult section of your code is in fact atomic and
solved using a library function or class?
The second type of atomic code you’ll write are what I call language techniques. These are usually snippets of code committed to
memory that perform a certain operation in the language, like locating the index of the first element in an array with the minimum
value, or parsing a String into tokens separated by whitespace. These techniques are equally essential to planning an approach,
because if you know how to do these fundamental operations intuitively, it makes more tasks in your search for a topdown approach
atomic, thus making the search for the right approach shorter. In addition, it makes the segments of the code in these atomic segments
less error prone. Furthermore, if you are asked to perform a task similar to one that you already know a language technique for, it
makes it much easier to mutate the code to fit the situation (for example: searching for the index of the first maximal element in an
array based on some heuristic is easy if you already know how to type up similar tasks). Looking for these common language
techniques should become an element of your daily practice, and any atomic code should fly off your fingers as soon as you think about
it.
As an aside, I must address the use of code libraries. I know that this is a contested topic, and many successful coders out there make
use of a (sometimes encyclopedic) library as a preinserted segment of code before they start coding. This is totally legal (although
changes to the rules after the 2004 topcoder Open may affect their future legality), and there are obvious advantages to using a library,
mainly through the ability to declare more parts of your topdown approach atomic, and by being able to more quickly construct
bottomup fragments of code (as discussed below). It is my opinion, however, that the disadvantages of using library code outweigh the
advantages. On a small note, library code executed through functions can sometimes slow your coding, because you have to make the
input match the prototype of the code you’re trying to use. Library code is mostly nonmutatable, so if your library is asked to do
something that isn’t expressly defined, you find yourself fumbling over a language technique or algorithm that should already be
internalized. It is also possible that your library code isn’t bugfree, and debugging your library midcompetition is dangerous because
you may have to propagate that change to code you’ve already submitted and also to the template before you open any more
problems. Also, library use is not allowed in onsite competition. Finally, the use of library code (or macros for that manner) get you
used to leaning on your library instead of your instincts of the language, making the use of normal primitives less intuitive and the
understanding of other coder’s solutions during challenge phase not as thorough. If used in moderation your library can be powerful,
but it is not the ultimate weapon for all terrain.
There may be a point where you hit a piece of atomic code that you are unable to fragment. This is when you have to pull out the
thinking cap and start analyzing your current approach. Should I have broken up the tasks differently? Should I store my intermediate
values differently? Or maybe this is the key to the problem that makes the problem hard? All of these things must be considered before
you pound the keys. Even at these points where you realize that you’re stuck, there are ways to manipulate the problem at hand to
come to an insight on how to proceed quickly, and these ways comprise the remaining approach tactics.
Planning an Approach to a Topcoder Problem: Part 2
By leadhyena_inran — topcoder member
Discuss this article in the forums
Bottom Up Programming
This technique is the antithesis to breaking down a program, and should be the first thing you start doing when you get stuck. Bottom
up programming is the process of building up primitive functions into more functional code until the solution becomes as trivial as one
of the primitives. Sometimes you know that you’ll need certain functions to form a solution and if these functions are atomic or easy to
break down, you can start with these functions and build your solution upward instead of breaking it down.
In the case of MatArith, the procedure to add and multiply matrices was given in the problem statement making it easy to follow
directions and get two functions to start with. From there you could make a smaller evalMult function that multiplied matrices together
using a string evaluation and variable names, then a similar evalAdd that treats each term as a block and you have an approach to
solve the problem.
In general, it’s a very good strategy to code up any detailed procedure in the problem statement before tackling the actual problem.
Examples of these are randomizer functions, any data structures you’re asked to simulate, and any operations on mathematical objects
like matrices and complex numbers. You’ll find that by solving these smaller issues and then rereading the problem statement that you
will understand what needs to be done much better. And sometimes, if you’re really stuck, it doesn’t hurt to write a couple atomic
pieces of code that you know that you’ll need in order to convince your mind to break down the problem towards those functions. As
you can see, your path to the right approach need not be linear as long as it follows your train of thought.
Also, in case of a hidden bug, keep in mind that any code that you write using this approach tactic should be scanned for bugs before
your topdown code, because you tend to write this code first and thus at a stage where you understand the problem less than when
you finish the code. This is a good rule of thumb to follow when looking for errors in your code; they usually sit in the older sections of
the code, even if older is only decided by minutes or even seconds.
Brute Force
Any time the solution requires looking for an optimal configuration or a maximal number or any other choice of one of a finite set of
objects, the simplest way to solve the problem is to try all configurations. Any time the solution requires calculating a massive
calculation requiring many steps, the best way to solve it is to do every calculation as asked for in the problem. Any time the problem
asks you to count the number of ways something can be done, the best way to solve it is to try every way and make a tally. In other
words, the first approach to consider in any possibly timeintensive problem is the most obvious one, even if it is horribly inefficient.
This approach tactic, called brute force, is so called because there is no discerning thought about the method of calculation of the return
value. Any time you run into this kind of an optimization problem the first thing you should do is try to figure in your head the worst
possible test cases and if 8 seconds is enough time to solve each one. If so, brute force can be a very speedy and usually less error
prone approach. In order to utilize brute force, you have to know enough about the programming environment to calculate an estimate
how much time any calculation will take. But an estimate is just a guess, and any guess could be wrong. This is where your wisdom is
forced to kick in and make a judgment call. And this particular judgment call has bitten many a coder that didn’t think it could be done
brute force and couldn’t debug a fancier approach, and likewise those that didn’t figure correctly the worst of the cases to be tested for
time.
In general, if you can’t think of a way to solve the problem otherwise, plan to use brute force. If it ends up that you are wrong, and
there is a test case that takes too long, keep the brute force solution around, and while recoding the more elegant solution, use the
bruteforce solution to verify that your elegant code is correct in the smaller cases, knowing that its more direct approach is a good
verification of these cases (being much less errorprone code).
A Place for Algorithms
Wellknown and efficient algorithms exist for many standard problems, much like basic approaches exist for many standard word
problems in math, just like standard responses exist for most common opening moves in chess. While in general it’s a bad idea to lean
heavily upon your knowledge of the standard algorithms (it leads down the path of pattern mining and leaves you vulnerable to more
original problems), it’s a very good idea to know the ones that come up often, especially if you can apply them to either an atomic
section of code or to allow them to break your problem down.
This is not the place to discuss algorithms (there are big books to read and other tutorials in this series to follow that will show you the
important stuff), but rather to discuss how algorithms should be used in determining an approach. It is not sufficient to know how to
use an algorithm in the default sense; always strive to know any algorithms you have memorized inside and out. For example, you may
run into a problem like CityLink (SRM 170 Div I Med), which uses a careful mutation of a basic graph algorithm to solve in time,
whereby just coding the regular algorithm would not suffice. True understanding of how the algorithm works allows the insight needed
to be able to even conceive of the right mutation.
So, when you study algorithms, you need to understand how the code works, how long it will take to run, what parts of the code can be
changed and what effect any changes will have on the algorithm. It’s also extremely important that you know how to code the
algorithm by memory before you try to use it in an approach, because without the experience of implementing an algorithm, it becomes
very hard to tell whether your bugs are being caused by a faulty implementation or faulty input into the implementation. It’s also good
to practice different ways to use the algorithms creatively to solve different problems, to see what works and what doesn’t. Better for
an experiment with code to fall flat on its face in practice than during a competition. This is why broadbased algorithmic techniques
(like divideandconquer, dynamic programming, greedy algorithms) are better to study first before you study your more focused
algorithms because the concepts are easier to manipulate and easier to implement once you understand the procedure involved.
Manipulating the Domain
This situation will become more and more familiar: you find yourself trudging through the planning stages of a problem because of the
sheer amount of work involved in simulating the domain of the problem. This may be due to the inappropriateness of the presented
domain, and there are times when manipulating the domain of the problem to something more convenient will create an easier or more
recognizable problem. The classic example of this is the game of Fifteen (used as a problem in SRM 172). In the game of Fifteen, you
have numbers from 1 to 9 of which you may claim one per turn, and if exactly three of your numbers add to 15 before exactly three of
your opponent’s numbers adds to 15, then you win. For this problem, you can manipulate the domain by placing the numbers in the
configuration of a 3×3 magic square (where every row, column, and diagonal add up to the same sum, in this case 15). Instantly you
realize that the game of Fifteen is just the game TicTacToe in disguise, making the game easier to play and program a solution for,
because the manipulation of the domain transformed the situation of a game where you have no prior knowledge into one where you
have a lot more knowledge. Some mathematicians think of this as the ABA1 approach, because the algebra suggests the proper
process: first you transform the domain, then you perform your action, then (the A1) you reverse your transformation. This approach is
very common in solving complex problems like diagonalizing matrices and solving the Rubik’s Cube.
Most commonly this approach tactic is used to simplify basic calculations. A good example of this type of approach is
HexagonIntersections from SRM 206. In this problem it was needed to find the number of tiled hexagons that touched a given line. The
problem became much easier if you "slanted" the grid by transforming the numbers involved so that the hexagons involved had sides
parallel to the x and y axis and the problem still had the same answer, thereby simplifying calculations.
Extreme care must be taken while debugging if you manipulate the domain. Remember that the correct procedure to domain
manipulation is to first manipulate the domain, then solve the problem, and then correct the domain. When you test the code,
remember that either the domain must be properly reversed by the transformation before the result is returned, or the reversal must
not affect the answer. Also, when looking at values inside the domain manipulation, remember that these are transformed values and
not the real ones. It’s good to leave comment lines around your transformed section of code just to remind yourself of this fact.
Unwinding the Definitions
This approach tactic is an old mathematician’s trick, relating to the incessant stacking of definitions upon definitions, and can be used to
unravel a rather gnarly problem statement to get at the inner intended approach to the problem. The best way to do this is with code.
When you read the definition of something you have never encountered before, try to think how you would code it. If the code asks you
to find the simplest grozmojt in a set of integers, first figure out how your code would verify that something was a grozmojt and then
figure out how to search for it, regardless if you even need to verify that something was a grozmojt in the solution. This is very similar
to the bottomup programming above, but taken at the definition level instead of the procedural one.
Simulation problems fall under similar tactics, and create one of those times when those predisposed to object oriented coding styles
run up the scores. The best way to manage a simulation problem is to create a simulation object that can have actions performed on it
from a main function. That way you don’t worry if you passed enough state into a given function or not; since all of the information in
the simulation is coming along with you, the approach becomes very convenient and reaches atomic code very quickly. This is also the
correct approach to take if an algorithm needs to be simulated to count the steps needed in the algorithm (like MergeSort) or the
number of objects deallocated in the execution of another algorithm (like ImmutableTrees). In these situations, the elegance of the
code is usually sacrificed in the name of correctness and thoroughness, also making the approach easier to plan ahead for.
The Problem is Doable
An old geometry puzzle goes like this: you have a pair of concentric circles and the only length you are given is the length of a chord of
the outer circle (call the chord length x) that is tangent to the inner circle, and you are asked for the area between the circles. You
respond: "Well, if the problem is doable then the inner circle’s radius is irrelevant to the calculation, so I’ll declare it to be 0. Because
the area of the inner circle is 0, or degenerates to the center of the outer circle, the chord of the outer circle passes through the center
and is thus the diameter, and thus the area of the outer circle is Pi(x/2)2." Note that a proper geometric proof of this fact is harder to
do; the sheer fact that a solution exists actually makes the problem easier. Since the writer had to write a solution for the problem, you
know it’s always solvable, and this fact can be used to your advantage in an SRM.
This approach tactic broadens into the concept that the writer is looking for a particular type of solution, and sometimes through edits
of the original problem statement this approach is given away (especially if the original problem is considered too hard for the level it’s
at). Look for lowered constraints like arrays of size 20 (which many a seasoned coder will tell you is almost a codeword that the writer
is looking for a brute force solution), or integers limited to between 1 and 10000 (allowing safe multiplication in ints without overflow).
By leaning on the constraints you are acting similarly to the situation above, by not allowing the complexities of the problem that were
trimmed off by the constraints to complicate your approach.
Sometimes the level of the problem alone will give a hint to what solution is intended. For example, look at FanFailure (from SRM 195
Div I Easy). The problem used the language of subsets and maximal and minimal, so you start to think maybe attack with brute force,
and then you see the constraints opened up to 50 for the array size. 250 distinct subsets rules out brute force (better to find this out in
the approach than in the code, right?) and you could look to fancier algorithms… but then you realize that this is a Div I Easy and
probably isn’t as hard as it looks so you think through the greedy algorithm and decide that it probably works. This choice wouldn’t
have been so obvious had it not been a Div I Easy.
Keep in mind that these invisible cues are not objective and can’t be used to reason why an approach will work or not; they are there
only to suggest what the writer’s mind was thinking. Furthermore, if the writer is evil or particularly tricky, cues of this nature may be
red herrings to throw these tactics astray. As long as you temper this approach tactic with solid analysis before you go on a wild goose
chase, this "circular reasoning" can be used to great advantage.
Case Reduction
Sometimes the simplest problems to state are the ones that provide the most difficulty. With these types of problems it’s not unusual
that the solution requires that you break up the problem not into steps but into cases. By breaking a problem up into cases of different
sets of inputs you can create subproblems that can be much easier to solve. Consider the problem TeamPhoto (SRM 167 Div I Medium).
This problem is simple to state, but abhorrent to solve. If you break up the problem into a series of cases, you find that where the
entire problem couldn’t alone be solved by a greedy algorithm, each of the different cases could be, and you could take the best case
from those optimal configurations to solve the problem.
The most common use of case reduction involves removing the boundary cases so that they don’t mess up a naive solution. A good
example of this is BirthdayOdds (SRM 174 Div I Easy); many people hard coded if(daysInYear==1)return 2; to avoid the possible
problems with the boundary case, even if their solution would have handled it correctly without that statement. By adding that level of
security, it became easier to verify that the approach they chose was correct.
Plans Within Plans
As illustrated above, an approach isn’t easily stated, and is usually glossed over if reduced to a oneword label. Furthermore, there are
many times when there exist levels to a problem, each of which needs to be solved before the full solution comes to light. One clear
example of this is the problem MagicianTour (SRM 191 Div I Hard). There are definitely two delineated steps to this problem: the first
step requires a graph search to find all connected components and their 2coloring, and the second step requires the DP knapsack
algorithm. In cases like this, it’s very helpful to remember that sometimes more than one approach tactic needs to be applied to the
situation to get at the solution. Another great example is TopographicalImage (SRM 209 Div I Hard) which asks for the lowest angle
that places a calculated value based on shortest path under a certain limit. To solve, note that looking for this lowest value can be
approached by binary search, but there are plans within plans, and the inner plan is to apply FloydWarshall’s All Pairs Shortest Paths
algorithm to decide if the angle is satisfactory.
Remember also that an approach isn’t just "Oh, I know how to break this down… Let’s go!" The idea of planning your approach is to
strategically think about: the steps in your code, how an algorithm is to be applied, how the values are to be stored and passed, how
the solution will react to the worst case, where the most probable nesting places are for bugs. The idea is that if the solution is carefully
planned, there is a lower chance of losing it to a challenge or during system tests. For each approach there are steps that contain plans
for their steps.
Tactical Permutation
There is never a right approach for all coders, and there are usually at least two ways to do a problem. Let’s look at an unsavory
Division One Easy called OhamaLow (SRM 206 Div I Easy). One of the more popular ways to approach this problem is to try all
combinations of hands, see if the hand combination is legal, sort the hand combination, and compare it to the best hand so far. This is a
common bruteforce search strategy. But it’s not the entire approach. Remember that there are plans within plans. You have to choose
an approach on how to form each hand (this could be done recursively or using multiple forloops), how to store the hands (int arrays,
Strings, or even a new class), and how to compare them. There are many different ways to do each of these steps, and most of the
ways to proceed with the approach will work. As seen above as well as here, there are many times where more than one approach will
do the job, and these approaches are considered permutable. In fact, one way to permute the toplevel bruteforce search strategy is
to instead of considering all possible constructible hands and picking the best one, you can construct all possible final hands in best to
worst order and stop when you have found one that’s constructible. In other words you take all 5 character substrings of "87654321" in
order and see if the shared hand and player hand can make the chosen hand, and if so return that hand. This approach also requires
substeps (how to decide if the hand can be formed, how to walk the possible hands, and so on) but sometimes (and in this case it is
better) you can break it down faster.
The only way you get to choose between two approaches is if you are able to come up with both of them. A very good way to practice
looking for multiple approaches to problems is to try to solve as many of the problems in the previous SRM using two different
approaches. By doing this, you stretch your mind into looking for these different solutions, increasing your chances of finding the more
elegant solution, or the faster one to type, or even the one that looks easier to debug.
Backtracking from a Flawed Approach
As demonstrated in the previous section, it is very possible for there to exist more than one way to plan an approach to a problem. It
may even hit you in the middle of coding your approach how to more elegantly solve the problem. One of the hardest disciplines to
develop while competing in topcoder Single Round Matches is the facility to stick to the approach you’ve chosen until you can prove
without a shadow of a doubt that you made a mistake in the approach and the solution will not work. Remember that you are not
awarded points for code elegance, or for cleverness, or for optimized code. You are granted points solely on the ability to quickly post
correct code. If you come up with a more elegant solution than the one you’re in the middle of typing up, you have to make the split
second analysis of how much time you’ll lose for changing your current approach, and in most cases it isn’t worth it.
There is no easy answer to planning the right approach the first time. If you code up a solution and you know it is right but has bugs
this is much easier to repair than the sudden realization that you just went the entirely wrong direction. If you get caught in this
situation, whatever you do, don’t erase your code! Relabel your main function and any subfunctions or data structures that may be
affected by further changes. The reason is because while you may desire a clean slate, you must accept that some of your previous
routines may be the same, and retracing your steps by retyping the same code can be counterproductive to your thinking anew.
Furthermore, by keeping the old code you can test against it later looking for cases that will successfully challenge other coders using
the same flawed approach.
Conclusion
Planning an approach is not a science, although there is a lot of rigor in the thought involved. Rather, it is mainly educated guesswork
coupled with successful planning. By being creative, economical, and thorough about your thought process you can become more
successful and confident in your solutions and the time you spend thinking the problem through will save you time later in the coding
and debugging. This ability to plan your code before the fingers hit the keys only develops through lots of practice, but this diligence is
rewarded with an increasing ability to solve problems and eventually a sustained rating increase.
Mentioned in this writeup:
TCI ’02 Round 2 Div I Med – MatArith
SRM 170 Div I Med – CityLink
SRM 172 Div I Med – Fifteen
SRM 206 Div I Hard – HexagonIntersections
SRM 195 Div I Easy – FanFailure
SRM 167 Div I Med – TeamPhoto
SRM 174 Div I Easy – BirthdayOdds
SRM 191 Div I Hard – MagicianTour
SRM 210 Div II Hard – TopographicalImage
SRM 206 Div I Easy – OmahaLow
Mathematics for Topcoders
By dimkadimon — topcoder member
Discuss this article in the forums
Introduction
I have seen a number of competitors complain that they are unfairly disadvantaged because many topcoder problems are too
mathematical. Personally, I love mathematics and thus I am biased in this issue. Nevertheless, I strongly believe that problems should
contain at least some math, because mathematics and computer science often go hand in hand. It is hard to imagine a world where
these two fields could exist without any interaction with each other. These days, a great deal of applied mathematics is performed on
computers such as solving large systems of equations and approximating solutions to differential equations for which no closed formula
exists. Mathematics is widely used in computer science research, as well as being heavily applied to graph algorithms and areas of
computer vision.
This article discusses the theory and practical application to some of the more common mathematical constructs. The topics covered
are: primes, GCD, basic geometry, bases, fractions and complex numbers.
Primes
A number is prime if it is only divisible by 1 and itself. So for example 2, 3, 5, 79, 311 and 1931 are all prime, while 21 is not prime
because it is divisible by 3 and 7. To find if a number n is prime we could simply check if it divides any numbers below it. We can use
the modulus (%) operator to check for divisibility:
for (int i=2; i<n; i++)
if (n%i==0) return false;
return true;
We can make this code run faster by noticing that we only need to check divisibility for values of i that are less or equal to the square
root of n (call this m). If n divides a number that is greater than m then the result of that division will be some number less than m and
thus n will also divide a number less or equal to m. Another optimization is to realize that there are no even primes greater than 2.
Once we’ve checked that n is not even we can safely increment the value of i by 2. We can now write the final method for checking
whether a number is prime:
public boolean isPrime (int n){
if (n<=1) return false;
if (n==2) return true;
if (n%2==0) return false;
int m=Math.sqrt(n);
for (int i=3; i<=m; i+=2)
if (n%i==0)
return false;
return true;
}
Now suppose we wanted to find all the primes from 1 to 100000, then we would have to call the above method 100000 times. This
would be very inefficient since we would be repeating the same calculations over and over again. In this situation it is best to use a
method known as the Sieve of Eratosthenes. The Sieve of Eratosthenes will generate all the primes from 2 to a given number n. It
begins by assuming that all numbers are prime. It then takes the first prime number and removes all of its multiples. It then applies
the same method to the next prime number. This is continued until all numbers have been processed. For example, consider finding
primes in the range 2 to 20. We begin by writing all the numbers down:
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
2 is the first prime. We now cross out all of its multiples, ie every second number:
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
The next noncrossed out number is 3 and thus it is the second prime. We now cross out all the multiples of 3, ie every third number
from 3:
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
All the remaining numbers are prime and we can safely terminate the algorithm. Below is the code for the sieve:
public boolean[] sieve(int n){
boolean[] prime=new boolean[n+1];
Arrays.fill(prime,true);
prime[0]=false;
prime[1]=false;
int m=Math.sqrt(n);
for (int i=2; i<=m; i++)
if (prime[i])
for (int k=i*i; k<=n; k+=i)
prime[k]=false;
return prime;
}
In the above method, we create a boolean array prime which stores the primality of each number less of equal than n. If prime[i] is
true then number i is prime. The outer loop finds the next prime while the inner loop removes all the multiples of the current prime.
GCD
The greatest common divisor (GCD) of two numbers a and b is the greatest number that divides evenly into both a and b. Naively we
could start from the smallest of the two numbers and work our way downwards until we find a number that divides into both of them:
for (int i=Math.min(a,b); i>=1; i‐‐)
if (a%i==0 && b%i==0)
return i;
Although this method is fast enough for most applications, there is a faster method called Euclid’s algorithm. Euclid’s algorithm iterates
over the two numbers until a remainder of 0 is found. For example, suppose we want to find the GCD of 2336 and 1314. We begin by
expressing the larger number (2336) in terms of the smaller number (1314) plus a remainder:
2336 = 1314 x 1 + 1022
We now do the same with 1314 and 1022:
1314 = 1022 x 1 + 292
We continue this process until we reach a remainder of 0:
1022 = 292 x 3 + 146
292 = 146 x 2 + 0
The last nonzero remainder is the GCD. So the GCD of 2336 and 1314 is 146. This algorithm can be easily coded as a recursive
function:
//assume that a and b cannot both be 0
public int GCD(int a, int b){
if (b==0) return a;
return GCD(b,a%b);
}
Using this algorithm we can find the lowest common multiple (LCM) of two numbers. For example the LCM of 6 and 9 is 18 since 18 is
the smallest number that divides both 6 and 9. Here is the code for the LCM method:
public int LCM(int a, int b){
return b*a/GCD(a,b);
}
As a final note, Euclid’s algorithm can be used to solve linear Diophantine equations. These equations have integer coefficients and are
of the form:
ax + by = c
Geometry
Occasionally problems ask us to find the intersection of rectangles. There are a number of ways to represent a rectangle. For the
standard Cartesian plane, a common method is to store the coordinates of the bottomleft and topright corners.
Suppose we have two rectangles R1 and R2. Let (x1, y1) be the location of the bottomleft corner of R1 and (x2, y2) be the location of
its topright corner. Similarly, let (x3, y3) and (x4, y4) be the respective corner locations for R2. The intersection of R1 and R2 will be a
rectangle R3 whose bottomleft corner is at (max(x1, x3), max(y1, y3)) and topright corner at (min(x2, x4), min(y2, y4)). If max(x1,
x3) > min(x2, x4) or max(y1, y3) > min(y2, y4) then R3 does not exist, ie R1 and R2 do not intersect. This method can be extended to
intersection in more than 2 dimensions as seen in CuboidJoin (SRM 191, Div 2 Hard).
Often we have to deal with polygons whose vertices have integer coordinates. Such polygons are called lattice polygons. In his tutorial
on Geometry Concepts, lbackstrom presents a neat way for finding the area of a lattice polygon given its vertices. Now, suppose we do
not know the exact position of the vertices and instead we are given two values:
B = number of lattice points on the boundary of the polygon
I = number of lattice points in the interior of the polygon
Amazingly, the area of this polygon is then given by:
Area = B/2 + I ‐ 1
The above formula is called Pick’s Theorem due to Georg Alexander Pick (1859 – 1943). In order to show that Pick’s theorem holds for
all lattice polygons we have to prove it in 4 separate parts. In the first part we show that the theorem holds for any lattice rectangle
(with sides parallel to axis). Since a rightangled triangle is simply half of a rectangle it is not too difficult to show that the theorem also
holds for any rightangled triangle (with sides parallel to axis). The next step is to consider a general triangle, which can be represented
as a rectangle with some rightangled triangles cut out from its corners. Finally, we can show that if the theorem holds for any two
lattice polygons sharing a common side then it will also hold for the lattice polygon, formed by removing the common side. Combining
the previous result with the fact that every simple polygon is a union of triangles gives us the final version of Pick’s Theorem. Pick’s
theorem is useful when we need to find the number of lattice points inside a large polygon.
Another formula worth remembering is Euler’s Formula for polygonal nets. A polygonal net is a simple polygon divided into smaller
polygons. The smaller polygons are called faces, the sides of the faces are called edges and the vertices of the faces are called vertices.
Euler’s Formula then states:
V ‐ E + F = 2, where
V = number of vertices
E = number of edges
F = number of faces
For example, consider a square with both diagonals drawn. We have V = 5, E = 8 and F = 5 (the outside of the square is also a face)
and so V – E + F = 2.
We can use induction to show that Euler’s formula works. We must begin the induction with V = 2, since every vertex has to be on at
least one edge. If V = 2 then there is only one type of polygonal net possible. It has two vertices connected by E number of edges. This
polygonal net has E faces (E – 1 "in the middle" and 1 "outside"). So V – E + F = 2 – E + E = 2. We now assume that V – E + F = 2 is
true for all 2<=V<=n. Let V = n + 1. Choose any vertex w at random. Now suppose w is joined to the rest of the net by G edges. If we
remove w and all these edges, we have a net with n vertices, E – G edges and F – G + 1 faces. From our assumption, we have:
(n) ‐ (E ‐ G) + (F ‐ G + 1) = 2
thus (n+1) ‐ E + F = 2
Since V = n + 1, we have V – E + F = 2. Hence by the principal of mathematical induction we have proven Euler’s formula.
Bases
A very common problem faced by topcoder competitors during challenges involves converting to and from binary and decimal
representations (amongst others).
So what does the base of the number actually mean? We will begin by working in the standard (decimal) base. Consider the decimal
number 4325. 4325 stands for 5 + 2 x 10 + 3 x 10 x 10 + 4 x 10 x 10 x 10. Notice that the "value" of each consequent digit increases
by a factor of 10 as we go from right to left. Binary numbers work in a similar way. They are composed solely from 0 and 1 and the
"value" of each digit increases by a factor of 2 as we go from right to left. For example, 1011 in binary stands for 1 + 1 x 2 + 0 x 2 x 2
+ 1 x 2 x 2 x 2 = 1 + 2 + 8 = 11 in decimal. We have just converted a binary number to a decimal. The same applies to other bases.
Here is code which converts a number n in base b (2<=b<=10) to a decimal number:
public int toDecimal(int n, int b){
int result=0;
int multiplier=1;
while(n>0){
result+=n%10*multiplier;
multiplier*=b;
n/=10;
}
return result;
}
Java users will be happy to know that the above can be also written as:
return Integer.parseInt(""+n,b);
To convert from a decimal to a binary is just as easy. Suppose we wanted to convert 43 in decimal to binary. At each step of the method
we divide 43 by 2 and memorize the remainder. The final list of remainders is the required binary representation:
43/2 = 21 + remainder 1
21/2 = 10 + remainder 1
10/2 = 5 + remainder 0
5/2 = 2 + remainder 1
2/2 = 1 + remainder 0
1/2 = 0 + remainder 1
So 43 in decimal is 101011 in binary. By swapping all occurrences of 10 with b in our previous method we create a function which
converts from a decimal number n to a number in base b (2<=b<=10):
public int fromDecimal(int n, int b){
int result=0;
int multiplier=1;
while(n>0){
result+=n%b*multiplier;
multiplier*=10;
n/=b;
}
return result;
}
If the base b is above 10 then we must use nonnumeric characters to represent digits that have a value of 10 and more. We can let ‘A’
stand for 10, ‘B’ stand for 11 and so on. The following code will convert from a decimal to any base (up to base 20):
public String fromDecimal2(int n, int b){
String chars="0123456789ABCDEFGHIJ";
String result="";
while(n>0){
result=chars.charAt(n%b) + result;
n/=b;
}
return result;
}
In Java there are some useful shortcuts when converting from decimal to other common representations, such as binary (base 2), octal
(base 8) and hexadecimal (base 16):
Integer.toBinaryString(n);
Integer.toOctalString(n);
Integer.toHexString(n);
Fractions and Complex Numbers
Fractional numbers can be seen in many problems. Perhaps the most difficult aspect of dealing with fractions is finding the right way of
representing them. Although it is possible to create a fractions class containing the required attributes and methods, for most purposes
it is sufficient to represent fractions as 2element arrays (pairs). The idea is that we store the numerator in the first element and the
denominator in the second element. We will begin with multiplication of two fractions a and b:
public int[] multiplyFractions(int[] a, int[] b){
int[] c={a[0]*b[0], a[1]*b[1]};
return c;
}
Adding fractions is slightly more complicated, since only fractions with the same denominator can be added together. First of all we
must find the common denominator of the two fractions and then use multiplication to transform the fractions such that they both have
the common denominator as their denominator. The common denominator is a number which can divide both denominators and is
simply the LCM (defined earlier) of the two denominators. For example lets add 4/9 and 1/6. LCM of 9 and 6 is 18. Thus to transform
the first fraction we need to multiply it by 2/2 and multiply the second one by 3/3:
4/9 + 1/6 = (4*2)/(9 * 2) + (1 * 3)/(6 * 3) = 8/18 + 3/18
Once both fractions have the same denominator, we simply add the numerators to get the final answer of 11/18. Subtraction is very
similar, except we subtract at the last step:
4/9 ‐ 1/6 = 8/18 ‐ 3/18 = 5/18
Here is the code to add two fractions:
public int[] addFractions(int[] a, int[] b){
int denom=LCM(a[1],b[1]);
int[] c={denom/a[1]*a[0] + denom/b[1]*b[0], denom};
return c;
}
Finally it is useful to know how to reduce a fraction to its simplest form. The simplest form of a fraction occurs when the GCD of the
numerator and denominator is equal to 1. We do this like so:
public void reduceFraction(int[] a){
int b=GCD(a[0],a[1]);
a[0]/=b;
a[1]/=b;
}
Using a similar approach we can represent other special numbers, such as complex numbers. In general, a complex number is a
number of the form a + ib, where a and b are reals and i is the square root of 1. For example, to add two complex numbers m = a +
ib and n = c + id we simply group likewise terms:
m + n
= (a + ib) + (c + id)
= (a + c) + i(b + d)
Multiplying two complex numbers is the same as multiplying two real numbers, except we must use the fact that i^2 = 1:
m * n
= (a + ib) * (c + id)
= ac + iad + ibc + (i^2)bd
= (ac ‐ bd) + i(ad + bc)
By storing the real part in the first element and the complex part in the second element of the 2element array we can write code that
performs the above multiplication:
public int[] multiplyComplex(int[] m, int[] n){
int[] prod = {m[0]*n[0] ‐ m[1]*n[1], m[0]*n[1] + m[1]*n[0]};
return prod;
}
Conclusion
In conclusion I want to add that one cannot rise to the top of the topcoder rankings without understanding the mathematical constructs
and algorithms outlined in this article. Perhaps one of the most common topics in mathematical problems is the topic of primes. This is
closely followed by the topic of bases, probably because computers operate in binary and thus one needs to know how to convert from
binary to decimal. The concepts of GCD and LCM are common in both pure mathematics as well as geometrical problems. Finally, I have
included the last topic not so much for its usefulness in topcoder competitions, but more because it demonstrates a means of treating
certain numbers.
Geometry Concepts: Basic Concepts
By lbackstrom — topcoder member
Discuss this article in the forums
Introduction
Vectors
Vector Addition
Dot Product
Cross Product
LinePoint Distance
Polygon Area
Introduction
Many topcoders seem to be mortally afraid of geometry problems. I think it’s safe to say that the majority of them would be in favor of
a ban on topcoder geometry problems. However, geometry is a very important part of most graphics programs, especially computer
games, and geometry problems are here to stay. In this article, I’ll try to take a bit of the edge off of them, and introduce some
concepts that should make geometry problems a little less frightening.
Vectors
Vectors are the basis of a lot of methods for solving geometry problems. Formally, a vector is defined by a direction and a magnitude.
In the case of twodimension geometry, a vector can be represented as pair of numbers, x and y, which gives both a direction and a
magnitude. For example, the line segment from (1,3) to (5,1) can be represented by the vector (4,2). It’s important to understand,
however, that the vector defines only the direction and magnitude of the segment in this case, and does not define the starting or
ending locations of the vector.
Vector Addition
There are a number of mathematical operations that can be performed on vectors. The simplest of these is addition: you can add two
vectors together and the result is a new vector. If you have two vectors (x1, y1) and (x2, y2), then the sum of the two vectors is
simply (x1+x2, y1+y2). The image below shows the sum of four vectors. Note that it doesn’t matter which order you add them up in –
just like regular addition. Throughout these articles, we will use plus and minus signs to denote vector addition and subtraction, where
each is simply the piecewise addition or subtraction of the components of the vector.
Dot Product
The addition of vectors is relatively intuitive; a couple of less obvious vector operations are dot and cross products. The dot product of
two vectors is simply the sum of the products of the corresponding elements. For example, the dot product of (x1, y1) and (x2, y2) is
x1*x2 + y1*y2. Note that this is not a vector, but is simply a single number (called a scalar). The reason this is useful is that the dot
product, A ? B = |A||B|Cos(?), where ? is the angle between the A and B. |A| is called the norm of the vector, and in a 2D geometry
problem is simply the length of the vector, sqrt(x2+y2). Therefore, we can calculate Cos(?) = (A ? B)/(|A||B|). By using the acos
function, we can then find ?. It is useful to recall that Cos(90) = 0 and Cos(0) = 1, as this tells you that a dot product of 0 indicates two
perpendicular lines, and that the dot product is greatest when the lines are parallel. A final note about dot products is that they are not
limited to 2D geometry. We can take dot products of vectors with any number of elements, and the above equality still holds.
Cross Product
An even more useful operation is the cross product. The cross product of two 2D vectors is x1*y2 ‐ y1*x2 Technically, the cross product
is actually a vector, and has the magnitude given above, and is directed in the +z direction. Since we’re only working with 2D
geometry for now, we’ll ignore this fact, and use it like a scalar. Similar to the dot product, A x B = |A||B|Sin(?). However, ? has a
slightly different meaning in this case: |?| is the angle between the two vectors, but ? is negative or positive based on the righthand
rule. In 2D geometry this means that if A is less than 180 degrees clockwise from B, the value is positive. Another useful fact related
to the cross product is that the absolute value of |A||B|Sin(?) is equal to the area of the parallelogram with two of its sides formed by
A and B. Furthermore, the triangle formed by A, B and the red line in the diagram has half of the area of the parallelogram, so we can
calculate its area from the cross product also.
LinePoint Distance
Finding the distance from a point to a line is something that comes up often in geometry problems. Lets say that you are given 3
points, A, B, and C, and you want to find the distance from the point C to the line defined by A and B (recall that a line extends
infinitely in either direction). The first step is to find the two vectors from A to B (AB) and from A to C (AC). Now, take the cross product
AB x AC, and divide by |AB|. This gives you the distance (denoted by the red line) as (AB x AC)/|AB|. The reason this works comes from
some basic high school level geometry. The area of a triangle is found as base*height/2. Now, the area of the triangle formed by A, B
and C is given by (AB x AC)/2. The base of the triangle is formed by AB, and the height of the triangle is the distance from the line to
C. Therefore, what we have done is to find twice the area of the triangle using the cross product, and then divided by the length of the
base. As always with cross products, the value may be negative, in which case the distance is the absolute value.
Things get a little bit trickier if we want to find the distance from a line segment to a point. In this case, the nearest point might be one
of the endpoints of the segment, rather than the closest point on the line. In the diagram above, for example, the closest point to C on
the line defined by A and B is not on the segment AB, so the point closest to C is B. While there are a few different ways to check for
this special case, one way is to apply the dot product. First, check to see if the nearest point on the line AB is beyond B (as in the
example above) by taking AB ? BC. If this value is greater than 0, it means that the angle between AB and BC is between 90 and 90,
exclusive, and therefore the nearest point on the segment AB will be B. Similarly, if BA ? AC is greater than 0, the nearest point is A. If
both dot products are negative, then the nearest point to C is somewhere along the segment. (There is another way to do this, which
I’ll discuss here).
//Compute the dot product AB ? BC
int dot(int[] A, int[] B, int[] C){
AB = new int[2];
BC = new int[2];
AB[0] = B[0]‐A[0];
AB[1] = B[1]‐A[1];
BC[0] = C[0]‐B[0];
BC[1] = C[1]‐B[1];
int dot = AB[0] * BC[0] + AB[1] * BC[1];
return dot;
}
//Compute the cross product AB x AC
int cross(int[] A, int[] B, int[] C){
AB = new int[2];
AC = new int[2];
AB[0] = B[0]‐A[0];
AB[1] = B[1]‐A[1];
AC[0] = C[0]‐A[0];
AC[1] = C[1]‐A[1];
int cross = AB[0] * AC[1] ‐ AB[1] * AC[0];
return cross;
}
//Compute the distance from A to B
double distance(int[] A, int[] B){
int d1 = A[0] ‐ B[0];
int d2 = A[1] ‐ B[1];
return sqrt(d1*d1+d2*d2);
}
//Compute the distance from AB to C
//if isSegment is true, AB is a segment, not a line.
double linePointDist(int[] A, int[] B, int[] C, boolean isSegment){
double dist = cross(A,B,C) / distance(A,B);
if(isSegment){
int dot1 = dot(A,B,C);
if(dot1 > 0)return distance(B,C);
int dot2 = dot(B,A,C);
if(dot2 > 0)return distance(A,C);
}
return abs(dist);
}
That probably seems like a lot of code, but lets see the same thing with a point class and some operator overloading in C++ or C#. The
* operator is the dot product, while ^ is cross product, while + and – do what you would expect.
//Compute the distance from AB to C
//if isSegment is true, AB is a segment, not a line.
double linePointDist(point A, point B, point C, bool isSegment){
double dist = ((B‐A)^(C‐A)) / sqrt((B‐A)*(B‐A));
if(isSegment){
int dot1 = (C‐B)*(B‐A);
if(dot1 > 0)return sqrt((B‐C)*(B‐C));
int dot2 = (C‐A)*(A‐B);
if(dot2 > 0)return sqrt((A‐C)*(A‐C));
}
return abs(dist);
}
Operator overloading is beyond the scope of this article, but I suggest that you look up how to do it if you are a C# or C++ coder, and
write your own 2D point class with some handy operator overloading. It will make a lot of geometry problems a lot simpler.
Polygon Area
Another common task is to find the area of a polygon, given the points around its perimeter. Consider the nonconvex polygon below,
with 5 points. To find its area we are going to start by triangulating it. That is, we are going to divide it up into a number of triangles. In
this polygon, the triangles are ABC, ACD, and ADE. But wait, you protest, not all of those triangles are part of the polygon! We are
going to take advantage of the signed area given by the cross product, which will make everything work out nicely. First, we’ll take the
cross product of AB x AC to find the area of ABC. This will give us a negative value, because of the way in which A, B and C are
oriented. However, we’re still going to add this to our sum, as a negative number. Similarly, we will take the cross product AC x AD to
find the area of triangle ACD, and we will again get a negative number. Finally, we will take the cross product AD x AE and since these
three points are oriented in the opposite direction, we will get a positive number. Adding these three numbers (two negatives and a
positive) we will end up with a negative number, so will take the absolute value, and that will be area of the polygon.
The reason this works is that the positive and negative number cancel each other out by exactly the right amount. The area of ABC and
ACD ended up contributing positively to the final area, while the area of ADE contributed negatively. Looking at the original polygon, it
is obvious that the area of the polygon is the area of ABCD (which is the same as ABC + ABD) minus the area of ADE. One final note, if
the total area we end up with is negative, it means that the points in the polygon were given to us in clockwise order. Now, just to
make this a little more concrete, lets write a little bit of code to find the area of a polygon, given the coordinates as a 2D array, p.
int area = 0;
int N = lengthof(p);
//We will triangulate the polygon
//into triangles with points p[0],p[i],p[i+1]
for(int i = 1; i+1<N; i++){
int x1 = p[i][0] ‐ p[0][0];
int y1 = p[i][1] ‐ p[0][1];
int x2 = p[i+1][0] ‐ p[0][0];
int y2 = p[i+1][1] ‐ p[0][1];
int cross = x1*y2 ‐ x2*y1;
area += cross;
}
return abs(cross/2.0);
Notice that if the coordinates are all integers, then the final area of the polygon is one half of an integer.
Geometry Concepts: Line Intersection and its Applications
By lbackstrom — topcoder member
LineLine Intersection
Finding a Circle From 3 Points
Reflection
Rotation
Convex Hull
In the previous section we saw how to use vectors to solve geometry problems. Now we are going to learn how to use some basic linear
algebra to do line intersection, and then apply line intersection to a couple of other problems.
LineLine Intersection
One of the most common tasks you will find in geometry problems is line intersection. Despite the fact that it is so common, a lot of
coders still have trouble with it. The first question is, what form are we given our lines in, and what form would we like them in? Ideally,
each of our lines will be in the form Ax+By=C, where A, B and C are the numbers which define the line. However, we are rarely given
lines in this format, but we can easily generate such an equation from two points. Say we are given two different points, (x1, y1) and
(x2, y2), and want to find A, B and C for the equation above. We can do so by setting
A = y2‐y1
B = x1‐x2
C = A*x1+B*y1
Regardless of how the lines are specified, you should be able to generate two different points along the line, and then generate A, B and
C. Now, lets say that you have lines, given by the equations:
A1x + B1y = C1
A2x + B2y = C2
To find the point at which the two lines intersect, we simply need to solve the two equations for the two unknowns, x and y.
double det = A1*B2 ‐ A2*B1
if(det == 0){
//Lines are parallel
}else{
double x = (B2*C1 ‐ B1*C2)/det
double y = (A1*C2 ‐ A2*C1)/det
}
To see where this comes from, consider multiplying the top equation by B2, and the bottom equation by B1. This gives you
A1B2x + B1B2y = B2C1
A2B1x + B1B2y = B1C2
Now, subtract the bottom equation from the top equation to get
A1B2x ‐ A2B1x = B2C1 ‐ B1C2
Finally, divide both sides by A1B2 ‐ A2B1, and you get the equation for x. The equation for y can be derived similarly.
This gives you the location of the intersection of two lines, but what if you have line segments, not lines. In this case, you need to make
sure that the point you found is on both of the line segments. If your line segment goes from (x1,y1) to (x2,y2), then to check if (x,y)
is on that segment, you just need to check that min(x1,x2) = x = max(x1,x2), and do the same thing for y. You must be careful about
double precision issues though. If your point is right on the edge of the segment, or if the segment is horizontal or vertical, a simple
comparison might be problematic. In these cases, you can either do your comparisons with some tolerance, or else use a fraction class.
Finding a Circle From 3 Points
Given 3 points which are not colinear (all on the same line) those three points uniquely define a circle. But, how do you find the center
and radius of that circle? This task turns out to be a simple application of line intersection. We want to find the perpendicular bisectors
of XY and YZ, and then find the intersection of those two bisectors. This gives us the center of the circle.
To find the perpendicular bisector of XY, find the line from X to Y, in the form Ax+By=C. A line perpendicular to this line will be given by
the equation ‐Bx+Ay=D, for some D. To find D for the particular line we are interested in, find the midpoint between X and Y by taking
the midpoint of the x and y components independently. Then, substitute those values into the equation to find D. If we do the same
thing for Y and Z, we end up with two equations for two lines, and we can find their intersections as described above.
Reflection
Reflecting a point across a line requires the same techniques as finding a circle from 3 points. First, notice that the distance from X to
the line of reflection is the same as the distance from X’ to the line of reflection. Also note that the line between X and X’ is
perpendicular to the line of reflection. Now, if the line of reflection is given as Ax+By=C, then we already know how to find a line
perpendicular to it: ‐Bx+Ay=D. To find D, we simply plug in the coordinates for X. Now, we can find the intersection of the two lines at Y,
and then find X' = Y ‐ (X ‐ Y).
Rotation
Rotation doesn’t really fit in with line intersection, but I felt that it would be good to group it with reflection. In fact, another way to find
the reflected point is to rotate the original point 180 degrees about Y.
Imagine that we want to rotate one point around another, counterclockwise by ϴ degrees. For simplicity, lets assume that we are
rotating about the origin. In this case, we can find that x' = x Cos(ϴ) ‐ y Sin(ϴ) and y' = x Sin(ϴ) + y Cos(ϴ). If we are rotating
about a point other than the origin, we can account for this by shifting our coordinate system so that the origin is at the point of
rotation, doing the rotation with the above formulas, and then shifting the coordinate system back to where it started.
Convex Hull
A convex hull of a set of points is the smallest convex polygon that contains every one of the points. It is defined by a subset of all the
points in the original set. One way to think about a convex hull is to imagine that each of the points is a peg sticking up out of a board.
Take a rubber band and stretch it around all of the points. The polygon formed by the rubber band is a convex hull. There are many
different algorithms that can be used to find the convex hull of a set of points. In this article, I’m just going to describe one of them,
which is fast enough for most purposes, but is quite slow compared to some of the other algorithms.
First, loop through all of your points and find the leftmost point. If there is a tie, pick the highest point. You know for certain that this
point will be on the convex hull, so we’ll start with it. From here, we are going to move clockwise around the edge of the hull, picking
the points on the hull, one at a time. Eventually, we will get back to the start point. In order to find the next point around the hull, we
will make use of cross products. First, we will pick an unused point, and set the next point, N, to that point. Next, we will iterate
through each unused points, X, and if (X‐P) x (N‐P) (where P is the previous point) is negative, we will set N to X. After we have
iterated through each point, we will end up with the next point on the convex hull. See the diagram below for an illustration of how the
algorithm works. We start with P as the leftmost point. Now, say that we have N and X as shown in the leftmost frame. In this case the
cross product will be negative, so we will set N = X, and there will be no other unused points that make the cross product negative, and
hence we will advance, setting P = N. Now, in the next frame, we will end up setting N = X again, since the cross product here will be
negative. However, we aren’t done yet because there is still another point that will make the cross product negative, as shown in the
final frame.
The basic idea here is that we are using the cross product to find the point which is furthest counterclockwise from our current position
at P. While this may seem fairly straightforward, it becomes a little bit tricky when dealing with colinear points. If you have no colinear
points on the hull, then the code is very straightforward.
convexHull(point[] X){
int N = lengthof(X);
int p = 0;
//First find the leftmost point
for(int i = 1; i<N; i++){
if(X[i] < X[p])
p = i;
}
int start = p;
do{
int n = ‐1;
for(int i = 0; i<N; i++){
//Don't go back to the same point you came from
if(i == p)continue;
//If there is no N yet, set it to i
if(n == ‐1)n = i;
int cross = (X[i] ‐ X[p]) x (X[n] ‐ X[p]);
if(cross < 0){
//As described above, set N=X
n = i;
}
}
p = n;
}while(start!=p);
}
Once we start to deal with colinear points, things get trickier. Right away we have to change our method signature to take a boolean
specifying whether to include all of the colinear points, or only the necessary ones.
//If onEdge is true, use as many points as possible for
//the convex hull, otherwise as few as possible.
convexHull(point[] X, boolean onEdge){
int N = lengthof(X);
int p = 0;
boolean[] used = new boolean[N];
//First find the leftmost point
for(int i = 1; i<N; i++){
if(X[i] < X[p])
p = i;
}
int start = p;
do{
int n = ‐1;
int dist = onEdge?INF:0;
for(int i = 0; i<N; i++){
//X[i] is the X in the discussion
//Don't go back to the same point you came from
if(i==p)continue;
//Don't go to a visited point
if(used[i])continue;
//If there is no N yet, set it to X
if(n == ‐1)n = i;
int cross = (X[i] ‐ X[p]) x (X[n] ‐ X[p]);
//d is the distance from P to X
int d = (X[i] ‐ X[p]) ? (X[i] ‐ X[p]);
if(cross < 0){
//As described above, set N=X
n = i;
dist = d;
}else if(cross == 0){
//In this case, both N and X are in the
//same direction. If onEdge is true, pick the
//closest one, otherwise pick the farthest one.
if(onEdge && d < dist){
dist = d;
n = i;
}else if(!onEdge && d > dist){
dist = d;
n = i;
}
}
}
p = n;
used[p] = true;
}while(start!=p);
}
Geometry Concepts: Using Geometry in Topcoder Problems
By lbackstrom — topcoder member
PointInPolygon
TVTower
Satellites
Further Problems
PointInPolygon (SRM 187)
Requires: LineLine Intersection, LinePoint Distance
First off, we can use our LinePoint Distance code to test for the "BOUNDARY" case. If the distance from any segment to the test point
is 0, then return "BOUNDARY". If you didn’t have that code prewritten, however, it would probably be easier to just check and see if
the test point is between the minimum and maximum x and y values of the segment. Since all of the segments are vertical or
horizontal, this is sufficient, and the more general code is not necessary.
Next we have to check if a point is in the interior or the exterior. Imagine picking a point in the interior and then drawing a ray from
that point out to infinity in some direction. Each time the ray crossed the boundary of the polygon, it would cross from the interior to
the exterior, or vice versa. Therefore, the test point is on the interior if, and only if, the ray crosses the boundary an odd number of
times. In practice, we do not have to draw a raw all the way to infinity. Instead, we can just use a very long line segment from the test
point to a point that is sufficiently far away. If you pick the far away point poorly, you will end up having to deal with cases where the
long segment touches the boundary of the polygon where two edges meet, or runs parallel to an edge of a polygon ? both of which are
tricky cases to deal with. The quick and dirty way around this is to pick two large random numbers for the endpoint of the segment.
While this might not be the most elegant solution to the problem, it works very well in practice. The chance of this segment intersecting
anything but the interior of an edge are so small that you are almost guaranteed to get the right answer. If you are really concerned,
you could pick a few different random points, and take the most common answer.
String testPoint(verts, x, y){
int N = lengthof(verts);
int cnt = 0;
double x2 = random()*1000+1000;
double y2 = random()*1000+1000;
for(int i = 0; i<N; i++){
if(distPointToSegment(verts[i],verts[(i+1)%N],x,y) == 0)
return "BOUNDARY";
if(segmentsIntersect((verts[i],verts[(i+1)%N],{x,y},{x2,y2}))
cnt++;
}
if(cnt%2 == 0)return "EXTERIOR";
else return "INTERIOR";
}
TVTower(SRM 183)
Requires: Finding a Circle From 3 Points
In problems like this, the first thing to figure out is what sort of solutions might work. In this case, we want to know what sort of circles
we should consider. If a circle only has two points on it, then, in most cases, we can make a slightly smaller circle, that still has those
two points on it. The only exception to this is when the two points are exactly opposite each other on the circle. Three points, on the
other hand, uniquely define a circle, so if there are three points on the edge of a circle, we cannot make it slightly smaller, and still have
all three of them on the circle. Therefore, we want to consider two different types of circles: those with two points exactly opposite each
other, and those with three points on the circle. Finding the center of the first type of circle is trivial ? it is simply halfway between the
two points. For the other case, we can use the method for Finding a Circle From 3 Points. Once we find the center of a potential circle, it
is then trivial to find the minimum radius.
int[] x, y;
int N;
double best = 1e9;
void check(double cx, double cy){
double max = 0;
for(int i = 0; i< N; i++){
max = max(max,dist(cx,cy,x[i],y[i]));
}
best = min(best,max);
}
double minRadius(int[] x, int[] y){
this.x = x;
this.y = y;
N = lengthof(x);
if(N==1)return 0;
for(int i = 0; i<N; i++){
for(int j = i+1; j<N; j++){
double cx = (x[i]+x[j])/2.0;
double cy = (y[i]+y[j])/2.0;
check(cx,cy);
for(int k = j+1; k<N; k++){
//center gives the center of the circle with
//(x[i],y[i]), (x[j],y[j]), and (x[k],y[k]) on
//the edge of the circle.
double[] c = center(i,j,k);
check(c[0],c[1]);
}
}
}
return best;
}
Satellites (SRM 180)
Requires: LinePoint Distance
This problem actually requires an extension of the LinePoint Distance method discussed previously. It is the same basic principle, but
the formula for the cross product is a bit different in three dimensions.
The first step here is to convert from spherical coordinates into (x,y,z) triples, where the center of the earth is at the origin.
double x = sin(lng/180*PI)*cos(lat/180*PI)*alt;
double y = cos(lng/180*PI)*cos(lat/180*PI)*alt;
double z = sin(lat/180*PI)*alt;
Now, we want to take the cross product of two 3D vectors. As I mentioned earlier, the cross product of two vectors is actually a vector,
and this comes into play when working in three dimensions. Given vectors (x1,y1,z1) and (x2,y2,z2) the cross product is defined as the
vector (i,j,k) where
i = y1z2 ‐ y2z1;
j = x2z1 ‐ x1z2;
k = x1y2 ‐ x2y1;
Notice that if z1 = z2 = 0, then i and j are 0, and k is equal to the cross product we used earlier. In three dimensions, the cross product
is still related to the area of the parallelogram with two sides from the two vectors. In this case, the area of the parallelogram is the
norm of the vector: sqrt(i*i+j*j+k*k).
Hence, as before, we can determine the distance from a point (the center of the earth) to a line (the line from a satellite to a rocket).
However, the closest point on the line segment between a satellite and a rocket may be one of the end points of the segment, not the
closest point on the line. As before, we can use the dot product to check this. However, there is another way which is somewhat simpler
to code. Say that you have two vectors originating at the origin, S and R, going to the satellite and the rocket, and that |X| represents
the norm of a vector X.
Then, the closest point to the origin is R if |R|2 + |R‐S|2 = |S|2 and it is S if |S|2 + |R‐S|2 = |R|2
Naturally, this trick works in two dimensions also.
Further Problems
Once you think you’ve got a handle on the three problems above, you can give these ones a shot. You should be able to solve all of
them with the methods I’ve outlined, and a little bit of cleverness. I’ve arranged them in what I believe to ascending order of difficulty.
ConvexPolygon (SRM 166)
Requires: Polygon Area
Surveyor (TCCC ’04 Qual 1)
Requires: Polygon Area
Travel (TCI ’02)
Requires: Dot Product
Parachuter (TCI ’01 Round 3)
Requires: Point In Polygon, LineLine Intersection
PuckShot (SRM 186)
Requires: Point In Polygon, LineLine Intersection
ElectronicScarecrows (SRM 173)
Requires: Convex Hull, Dynamic Programming
Mirrors (TCI ’02 Finals)
Requires: Reflection, LineLine Intersection
Symmetry (TCI ’02 Round 4)
Requires: Reflection, LineLine Intersection
Warehouse (SRM 177)
Requires: LinePoint Distance, LineLine Intersection
The following problems all require geometry, and the topics discussed in this article will be useful. However, they all require some
additional skills. If you got stuck on them, the editorials are a good place to look for a bit of help. If you are still stuck, there has yet to
be a problem related question on the round tables that went unanswered.
DogWoods (SRM 201)
ShortCut (SRM 215)
SquarePoints (SRM 192)
Tether (TCCC ’03 W/MW Regional)
TurfRoller (SRM 203)
Watchtower (SRM 176)
Introduction to Graphs and Their Data Structures: Section 1
Section 1: Recognizing and Representing a Graph
By gladius — topcoder member
Discuss this article in the forums
Introduction
Recognizing a graph problem
Representing a graph and key concepts
Singly linked lists
Trees
Graphs
Array representation
Introduction
Graphs are a fundamental data structure in the world of programming, and this is no less so on topcoder. Usually appearing as the hard
problem in Division 2, or the medium or hard problem in Division 1, there are many different forms solving a graph problem can take.
They can range in difficulty from finding a path on a 2D grid from a start location to an end location, to something as hard as finding
the maximum amount of water that you can route through a set of pipes, each of which has a maximum capacity (also known as the
maximumflow minimumcut problem – which we will discuss later). Knowing the correct data structures to use with graph problems is
critical. A problem that appears intractable may prove to be a few lines with the proper data structure, and luckily for us the standard
libraries of the languages used by topcoder help us a great deal here!
Recognizing a graph problem
The first key to solving a graph related problem is recognizing that it is a graph problem. This can be more difficult than it sounds,
because the problem writers don’t usually spell it out for you. Nearly all graph problems will somehow use a grid or network in the
problem, but sometimes these will be well disguised. Secondly, if you are required to find a path of any sort, it is usually a graph
problem as well. Some common keywords associated with graph problems are: vertices, nodes, edges, connections, connectivity, paths,
cycles and direction. An example of a description of a simple problem that exhibits some of these characteristics is:
"Bob has become lost in his neighborhood. He needs to get from his current position back to his home. Bob’s neighborhood is a 2
dimensional grid, that starts at (0, 0) and (width – 1, height – 1). There are empty spaces upon which bob can walk with no difficulty,
and houses, which Bob cannot pass through. Bob may only move horizontally or vertically by one square at a time.
Bob’s initial position will be represented by a ‘B’ and the house location will be represented by an ‘H’. Empty squares on the grid are
represented by ‘.’ and houses are represented by ‘X’. Find the minimum number of steps it takes Bob to get back home, but if it is not
possible for Bob to return home, return 1.
An example of a neighborhood of width 7 and height 5:
...X..B
.X.X.XX
.H.....
...X...
.....X."
Once you have recognized that the problem is a graph problem it is time to start building up your representation of the graph in
memory.
Representing a graph and key concepts
Graphs can represent many different types of systems, from a twodimensional grid (as in the problem above) to a map of the internet
that shows how long it takes data to move from computer A to computer B. We first need to define what components a graph consists
of. In fact there are only two, nodes and edges. A node (or vertex) is a discrete position in the graph. An edge (or connection) is a link
between two vertices that can be either directed or undirected and may have a cost associated with it. An undirected edge means that
there is no restriction on the direction you can travel along the edge. So for example, if there were an undirected edge from A to B you
could move from A to B or from B to A. A directed edge only allows travel in one direction, so if there were a directed edge from A to B
you could travel from A to B, but not from B to A. An easy way to think about edges and vertices is that edges are a function of two
vertices that returns a cost. We will see an example of this methodology in a second.
For those that are used to the mathematical description of graphs, a graph G = {V, E} is defined as a set of vertices, V, and a collection
of edges (which is not necessarily a set), E. An edge can then be defined as (u, v) where u and v are elements of V. There are a few
technical terms that it would be useful to discuss at this point as well:
Order – The number of vertices in a graph Size – The number of edges in a graph
Singly linked lists
An example of one of the simplest types of graphs is a singly linked list! Now we can start to see the power of the graph data structure,
as it can represent very complicated relationships, but also something as simple as a list.
A singly linked list has one "head" node, and each node has a link to the next node. So the structure looks like this:
structure node
[node] link;
[data]
end
node head;
A simple example would be:
node B, C;
head.next = B;
B.next = C;
C.next = null;
This would be represented graphically as head > B > C > null. I’ve used null here to represent the end of a list.
Getting back to the concept of a cost function, our cost function would look as follows:
cost(X, Y) := if (X.link = Y) return 1;
else if (X = Y) return 0;
else "Not possible"
This cost function represents the fact that we can only move directly to the link node from our current node. Get used to seeing cost
functions because anytime that you encounter a graph problem you will be dealing with them in some form or another! A question that
you may be asking at this point is "Wait a second, the cost from A to C would return not possible, but I can get to C from A by stepping
through B!" This is a very valid point, but the cost function simply encodes the *direct* cost from a node to another. We will cover how
to find distances in generic graphs later on.
Now that we have seen an example of the one of the simplest types of graphs, we will move to a more complicated example.
Trees
There will be a whole section written on trees. We are going to cover them very briefly as a steppingstone along the way to a full
fledged graph. In our list example above we are somewhat limited in the type of data we can represent. For example, if you wanted to
start a family tree (a hierarchal organization of children to parents, starting from one child) you would not be able to store more than
one parent per child. So we obviously need a new type of data structure. Our new node structure will look something like this:
structure node
[node] mother, father;
[string] name
end
node originalChild;
With a cost function of:
cost(X, Y) := if ((X.mother = Y) or (X.father = Y)) return 1;
else if (X = Y) return 0;
else "Not possible"
Here we can see that every node has a mother and father. And since node is a recursive structure definition, every mother has mother
and father, and every father has a mother and father, and so on. One of the problems here is that it might be possible to form a loop if
you actually represented this data structure on a computer. And a tree clearly cannot have a loop. A little mind exercise will make this
clear: a father of a child is also the son of that child? It’s starting to make my head hurt already. So you have to be very careful when
constructing a tree to make sure that it is truly a tree structure, and not a more general graph. A more formal definition of a tree is that
it is a connected acyclic graph. This simply means that there are no cycles in the graph and every node is connected to at least one
other node in the graph.
Another thing to note is that we could imagine a situation easily where the tree requires more than two node references, for example in
an organizational hierarchy, you can have a manager who manages many people then the CEO manages many managers. Our example
above was what is known as a binary tree, since it only has two node references. Next we will move onto constructing a data structure
that can represent a general graph!
Graphs
A tree only allows a node to have children, and there cannot be any loops in the tree, with a more general graph we can represent
many different situations. A very common example used is flight paths between cities. If there is a flight between city A and city B
there is an edge between the cities. The cost of the edge can be the length of time that it takes for the flight, or perhaps the amount of
fuel used.
The way that we will represent this is to have a concept of a node (or vertex) that contains links to other nodes, and the data
associated with that node. So for our flight path example we might have the name of the airport as the node data, and for every flight
leaving that city we have an element in neighbors that points to the destination.
structure node
[list of nodes] neighbors
[data]
end
cost(X, Y) := if (X.neighbors contains Y) return X.neighbors[Y];
else "Not possible"
list nodes;
This is a very general way to represent a graph. It allows us to have multiple edges from one node to another and it is a very compact
representation of a graph as well. However the downside is that it is usually more difficult to work with than other representations (such
as the array method discussed below).
Array representation
Representing a graph as a list of nodes is a very flexible method. But usually on topcoder we have limits on the problems that attempt
to make life easier for us. Normally our graphs are relatively small, with a small number of nodes and edges. When this is the case we
can use a different type of data structure that is easier to work with.
The basic concept is to have a 2 dimensional array of integers, where the element in row i, at column j represents the edge cost from
node i to j. If the connection from i to j is not possible, we use some sort of sentinel value (usually a very large or small value, like 1 or
the maximum integer). Another nice thing about this type of structure is that we can represent directed or undirected edges very easily.
So for example, the following connection matrix:
A B C
A 0 1 5
B ‐1 0 1
C ‐1 ‐1 0
Would mean that node A has a 0 weight connection to itself, a 1 weight connection to node B and 5 weight connection to node C. Node
B on the other hand has no connection to node A, a 0 weight connection to itself, and a 1 weight connection to C. Node C is connected
to nobody. This graph would look like this if you were to draw it:
This representation is very convenient for graphs that do not have multiple edges between each node, and allows us to simplify working
with the graph.
Introduction to Graphs and Their Data Structures: Section 2
Section 2: Searching a Graph
By gladius — topcoder member
Basic methods for searching graphs
Introduction
Stack
Depth First Search
Queue
Breadth First Search
Basic methods for searching graphs
Introduction
So far we have learned how to represent our graph in memory, but now we need to start doing something with this information. There
are two methods for searching graphs that are extremely prevalent, and will form the foundations for more advanced algorithms later
on. These two methods are the Depth First Search and the Breadth First Search.
We will begin with the depth first search method, which will utilize a stack. This stack can either by represented explicitly (by a stack
datatype in our language) or implicitly when using recursive functions.
Stack
A stack is one of the simplest data structures available. There are four main operations on a stack:
1. Push – Adds an element to the top of the stack
2. Pop – Removes the top element from the stack
3. Top – Returns the top element on the stack
4. Empty – Tests if the stack is empty or not
In C++, this is done with the STL class stack:
#include <stack>
std::stack<int> myStack;
In Java, we use the Stack class:
import java.util.*;
Stack stack = new Stack();
In C#, we use Stack class:
using System.Collections;
Stack stack = new Stack();
Depth First Search
Now to solve an actual problem using our search! The depth first search is well geared towards problems where we want to find any
solution to the problem (not necessarily the shortest path), or to visit all of the nodes in the graph. A recent topcoder problem was a
classic application of the depth first search, the floodfill. The floodfill operation will be familiar to anyone who has used a graphic
painting application. The concept is to fill a bounded region with a single color, without leaking outside the boundaries.
This concept maps extremely well to a Depth First search. The basic concept is to visit a node, then push all of the nodes to be visited
onto the stack. To find the next node to visit we simply pop a node of the stack, and then push all the nodes connected to that one onto
the stack as well and we continue doing this until all nodes are visited. It is a key property of the Depth First search that we not visit
the same node more than once, otherwise it is quite possible that we will recurse infinitely. We do this by marking the node as we visit
it.
So the basic structure will look something like this:
dfs(node start) {
stack<node> s;
s.push(start);
while (s.empty() == false) {
top = s.top();
s.pop();
if (top is not marked as visited) {
check for termination condition (have we reached the node we want to?)
mark top as visited;
add all of top's neighbors to the stack.
}
}
}
Alternatively we can define the function recursively as follows:
dfs(node current) {
mark current as visited;
visit all of current's unvisited neighbors by calling dfs(neighbor)
}
The problem we will be discussing is grafixMask, a Division 1 500 point problem from SRM 211. This problem essentially asks us to find
the number of discrete regions in a grid that has been filled in with some values already. Dealing with grids as graphs is a very powerful
technique, and in this case makes the problem quite easy.
We will define a graph where each node has 4 connections, one each to the node above, left, right and below. However, we can
represent these connections implicitly within the grid, we need not build out any new data structures. The structure we will use to
represent the grid in grafixMask is a two dimensional array of booleans, where regions that we have already determined to be filled in
will be set to true, and regions that are unfilled are set to false.
To set up this array given the data from the problem is very simple, and looks something like this:
bool fill[600][400];
initialize fills to false;
foreach rectangle in Rectangles
set from (rectangle.left, rectangle.top) to (rectangle.right, retangle.bottom) to true
Now we have an initialized connectivity grid. When we want to move from grid position (x, y) we can either move up, down, left or
right. When we want to move up for example, we simply check the grid position in (x, y1) to see if it is true or false. If the grid
position is false, we can move there, if it is true, we cannot.
Now we need to determine the area of each region that is left. We don’t want to count regions twice, or pixels twice either, so what we
will do is set fill[x][y] to true when we visit the node at (x, y). This will allow us to perform a DepthFirst search to visit all of the nodes
in a connected region and never visit any node twice, which is exactly what the problem wants us to do! So our loop after setting
everything up will be:
int[] result;
for x = 0 to 599
for y = 0 to 399
if (fill[x][y] == false)
result.addToBack(doFill(x,y));
All this code does is check if we have not already filled in the position at (x, y) and then calls doFill() to fill in that region. At this point
we have a choice, we can define doFill recursively (which is usually the quickest and easiest way to do a depth first search), or we can
define it explicitly using the built in stack classes. I will cover the recursive method first, but we will soon see for this problem there are
some serious issues with the recursive method.
We will now define doFill to return the size of the connected area and the start position of the area:
int doFill(int x, int y) {
// Check to ensure that we are within the bounds of the grid, if not, return 0
if (x < 0 || x >= 600) return 0;
// Similar check for y
if (y < 0 || y >= 400) return 0;
// Check that we haven't already visited this position, as we don't want to count it twice
if (fill[x][y]) return 0;
// Record that we have visited this node
fill[x][y] = true;
// Now we know that we have at least one empty square, then we will recursively attempt to
// visit every node adjacent to this node, and add those results together to return.
return 1 + doFill(x ‐ 1, y) + doFill(x + 1, y) + doFill(x, y + 1) + doFill(x, y ‐ 1);
}
This solution should work fine, however there is a limitation due to the architecture of computer programs. Unfortunately, the memory
for the implicit stack, which is what we are using for the recursion above is more limited than the general heap memory. In this
instance, we will probably overflow the maximum size of our stack due to the way the recursion works, so we will next discuss the
explicit method of solving this problem.
Sidenote:
Stack memory is used whenever you call a function; the variables to the function are pushed onto the stack by the compiler for you.
When using a recursive function, the variables keep getting pushed on until the function returns. Also any variables the compiler needs
to save between function calls must be pushed onto the stack as well. This makes it somewhat difficult to predict if you will run into
stack difficulties. I recommend using the explicit Depth First search for every situation you are at least somewhat concerned about
recursion depth.
In this problem we may recurse a maximum of 600 * 400 times (consider the empty grid initially, and what the depth first search will
do, it will first visit 0,0 then 1,0, then 2,0, then 3,0 … until 599, 0. Then it will go to 599, 1 then 598, 1, then 597, 1, etc. until it
reaches 599, 399. This will push 600 * 400 * 2 integers onto the stack in the best case, but depending on what your compiler does it
may in fact be more information. Since an integer takes up 4 bytes we will be pushing 1,920,000 bytes of memory onto the stack,
which is a good sign we may run into trouble.
We can use the same function definition, and the structure of the function will be quite similar, just we won’t use any recursion any
more:
class node { int x, y; }
int doFill(int x, int y) {
int result = 0;
// Declare our stack of nodes, and push our starting node onto the stack
stack<node> s;
s.push(node(x, y));
while (s.empty() == false) {
node top = s.top();
s.pop();
// Check to ensure that we are within the bounds of the grid, if not, continue
if (top.x < 0 || top.x >= 600) continue;
// Similar check for y
if (top.y < 0 || top.y >= 400) continue;
// Check that we haven't already visited this position, as we don't want to count it twice
if (fill[top.x][top.y]) continue;
fill[top.x][top.y] = true; // Record that we have visited this node
// We have found this node to be empty, and part
// of this connected area, so add 1 to the result
result++;
// Now we know that we have at least one empty square, then we will attempt to
// visit every node adjacent to this node.
s.push(node(top.x + 1, top.y));
s.push(node(top.x ‐ 1, top.y));
s.push(node(top.x, top.y + 1));
s.push(node(top.x, top.y ‐ 1));
}
return result;
}
As you can see, this function has a bit more overhead to manage the stack structure explicitly, but the advantage is that we can use the
entire memory space available to our program and in this case, it is necessary to use that much information. However, the structure is
quite similar and if you compare the two implementations they are almost exactly equivalent.
Congratulations, we have solved our first question using a depth first search! Now we will move onto the depthfirst searches close
cousin the Breadth First search.
If you want to practice some DFS based problems, some good ones to look at are:
TCCC 03 Quarterfinals – Marketing – Div 1 500
TCCC 03 Semifinals Room 4 – Circuits – Div 1 275
Queue
A queue is a simple extension of the stack data type. Whereas the stack is a FILO (firstin lastout) data structure the queue is a FIFO
(firstin firstout) data structure. What this means is the first thing that you add to a queue will be the first thing that you get when you
perform a pop().
There are four main operations on a queue:
1. Push – Adds an element to the back of the queue
2. Pop – Removes the front element from the queue
3. Front – Returns the front element on the queue
4. Empty – Tests if the queue is empty or not
In C++, this is done with the STL class queue:
#include <queue>
queue<int> myQueue;
In Java, we unfortunately don’t have a Queue class, so we will approximate it with the LinkedList class. The operations on a linked list
map well to a queue (and in fact, sometimes queues are implemented as linked lists), so this will not be too difficult.
The operations map to the LinkedList class as follows:
1. Push – boolean LinkedList.add(Object o)
2. Pop – Object LinkedList.removeFirst()
3. Front – Object LinkedList.getFirst()
4. Empty – int LinkedList.size()
import java.util.*;
LinkedList myQueue = new LinkedList();
In C#, we use Queue class:
The operations map to the Queue class as follows:
1. Push – void Queue.Enqueue(Object o)
2. Pop – Object Queue.Dequeue()
3. Front – Object Queue.Peek()
4. Empty – int Queue.Count
using System.Collections;
Queue myQueue = new Queue();
Breadth First Search
The Breadth First search is an extremely useful searching technique. It differs from the depthfirst search in that it uses a queue to
perform the search, so the order in which the nodes are visited is quite different. It has the extremely useful property that if all of the
edges in a graph are unweighted (or the same weight) then the first time a node is visited is the shortest path to that node from the
source node. You can verify this by thinking about what using a queue means to the search order. When we visit a node and add all the
neighbors into the queue, then pop the next thing off of the queue, we will get the neighbors of the first node as the first elements in
the queue. This comes about naturally from the FIFO property of the queue and ends up being an extremely useful property. One thing
that we have to be careful about in a Breadth First search is that we do not want to visit the same node twice, or we will lose the
property that when a node is visited it is the quickest path to that node from the source.
The basic structure of a breadth first search will look this:
void bfs(node start) {
queue<node> s;
s.push(start);
mark start as visited
while (s.empty() == false) {
top = s.front();
s.pop();
check for termination condition (have we reached the node we want to?)
add all of top's unvisited neighbors to the queue
mark all of top's unvisited neighbors as visited
}
}
Notice the similarities between this and a depthfirst search, we only differ in the data structure used and we mark a vertex visited as
we push it into the queue, not as we pop it.
The problem we will be discussing in relation to the Breadth First search is a bit harder than the previous example, as we are dealing
with a slightly more complicated search space. The problem is the 1000 from Division 1 in SRM 156, Pathfinding. Once again we will be
dealing in a gridbased problem, so we can represent the graph structure implicitly within the grid.
A quick summary of the problem is that we want to exchange the positions of two players on a grid. There are impassable spaces
represented by ‘X’ and spaces that we can walk in represented by ‘ ‘. Since we have two players our node structure becomes a bit more
complicated, we have to represent the positions of person A and person B. Also, we won’t be able to simply use our array to represent
visited positions any more, we will have an auxiliary data structure to do that. Also, we are allowed to make diagonal movements in this
problem, so we now have 9 choices, we can move in one of 8 directions or simply stay in the same position. Another little trick that we
have to watch for is that the players can not just swap positions in a single turn, so we have to do a little bit of validity checking on the
resulting state.
First, we set up the node structure and visited array:
class node {
int player1X, player1Y, player2X, player2Y;
int steps; // The current number of steps we have taken to reach this step
}
bool visited[20][20][20][20];
Here a node is represented as the (x,y) positions of player 1 and player 2. It also has the current steps that we have taken to reach the
current state, we need this because the problem asks us what the minimum number of steps to switch the two players will be. We are
guaranteed by the properties of the Breadth First search that the first time we visit the end node, it will be as quickly as possible (as all
of our edge costs are 1).
The visited array is simply a direct representation of our node in array form, with the first dimension being player1X, second player1Y,
etc. Note that we don’t need to keep track of steps in the visited array.
Now that we have our basic structure set up, we can solve the problem (note that this code is not compilable):
void pushToQueue(queue<node> q, node v) {
if (visited[v.player1X][v.player1Y][v.player2X][v.player2Y]) return;
q.push(v);
visited[v.player1X][v.player1Y][v.player2X][v.player2Y] = true;
}
int minTurns(String[] board) {
int width = board[0].length;
int height = board.length;
node start;
// Find the initial position of A and B, and save them in start.
queue<node> q;
pushToQueue(q, start)
while (q.empty() == false) {
node top = q.front();
q.pop();
// Check if player 1 or player 2 is out of bounds, or on an X square, if so continue
// Check if player 1 or player 2 is on top of each other, if so continue
// Check if the current positions of A and B are the opposite of what they were in start.
// If they are we have exchanged positions and are finished!
if (top.player1X == start.player2X && top.player1Y == start.player2Y &&
top.player2X == start.player1X && top.player2Y == start.player1Y)
return top.steps;
// Now we need to generate all of the transitions between nodes, we can do this quite easily using some
// nested for loops, one for each direction that it is possible for one player to move. Since we need
// to generate the following deltas: (‐1,‐1), (‐1,0), (‐1,1), (0,‐1), (0,0), (0,1), (1,‐1), (1,0), (1,1)
// we can use a for loop from ‐1 to 1 to do exactly that.
for (int player1XDelta = ‐1; player1XDelta <= 1; player1XDelta++) {
for (int player1YDelta = ‐1; player1YDelta <= 1; player1YDelta++) {
for (int player2XDelta = ‐1; player2XDelta <= 1; player2XDelta++) {
for (int player2YDelta = ‐1; player2YDelta <= 1; player2YDelta++) {
// Careful though! We have to make sure that player 1 and 2 did not swap positions on this turn
if (top.player1X == top.player2X + player2XDelta && top.player1Y == top.player2Y + player2YDelta &&
top.player2X == top.player1X + player1XDelta && top.player2Y == top.player1Y + player1YDelta)
continue;
// Add the new node into the queue
pushToQueue(q, node(top.player1X + player1XDelta, top.player1Y + player1YDelta,
top.player2X + player2XDelta, top.player2Y + player2YDelta,
top.steps + 1));
}
}
}
}
}
// It is not possible to exchange positions, so
// we return ‐1. This is because we have explored
// all the states possible from the starting state,
// and haven't returned an answer yet.
return ‐1;
}
This ended up being quite a bit more complicated than the basic Breadth First search implementation, but you can still see all of the
basic elements in the code. Now, if you want to practice more problems where Breadth First search is applicable, try these:
Inviational 02 Semifinal Room 2 – Div 1 500 – Escape
Introduction to Graphs and Their Data Structures: Section 3
Section 3: Finding the Best Path through a Graph
By gladius — topcoder member
Finding the best path through a graph
Dijkstra (Heap method)
FloydWarshall
Finding the best path through a graph
An extremely common problem on topcoder is to find the shortest path from one position to another. There are a few different ways for
going about this, each of which has different uses. We will be discussing two different methods, Dijkstra using a Heap and the Floyd
Warshall method.
Dijkstra (Heap method)
Dijkstra using a Heap is one of the most powerful techniques to add to your topcoder arsenal. It essentially allows you to write a
Breadth First search, and instead of using a Queue you use a Priority Queue and define a sorting function on the nodes such that the
node with the lowest cost is at the top of the Priority Queue. This allows us to find the best path through a graph in O(m * log(n)) time
where n is the number of vertices and m is the number of edges in the graph.
Sidenote:
If you haven’t seen bigO notation before then I recommend reading this.
First however, an introduction to the Priority Queue/Heap structure is in order. The Heap is a fundamental data structure and is
extremely useful for a variety of tasks. The property we are most interested in though is that it is a semiordered data structure. What I
mean by semiordered is that we define some ordering on elements that are inserted into the structure, then the structure keeps the
smallest (or largest) element at the top. The Heap has the very nice property that inserting an element or removing the top element
takes O(log n) time, where n is the number of elements in the heap. Simply getting the top value is an O(1) operation as well, so the
Heap is perfectly suited for our needs.
The fundamental operations on a Heap are:
1. Add – Inserts an element into the heap, putting the element into the correct ordered location.
2. Pop – Pops the top element from the heap, the top element will either be the highest or lowest element, depending on
implementation.
3. Top – Returns the top element on the heap.
4. Empty – Tests if the heap is empty or not.
Pretty close to the Queue or Stack, so it’s only natural that we apply the same type of searching principle that we have used before,
except substitute the Heap in place of the Queue or Stack. Our basic search routine (remember this one well!) will look something like
this:
void dijkstra(node start) {
priorityQueue s;
s.add(start);
while (s.empty() == false) {
top = s.top();
s.pop();
mark top as visited;
}
}
check for termination condition (have we reached the target node?)
add all of top’s unvisited neighbors to the stack.
Unfortunately, not all of the default language libraries used in topcoder have an easy to use priority queue structure.
C++ users are lucky to have an actual priority_queue<> structure in the STL, which is used as follows:
#include
using namespace std;
priority_queue pq;
1. Add ‐ void pq.push(type)
2. Pop ‐ void pq.pop()
3. Top ‐ type pq.top()
4. Empty ‐ bool pq.empty()
However, you have to be careful as the C++ priority_queue<> returns the *highest* element first, not the lowest. This has been the
cause of many solutions that should be O(m * log(n)) instead ballooning in complexity, or just not working.
To define the ordering on a type, there are a few different methods. The way I find most useful is the following though:
Define your structure:
struct node {
int cost;
int at;
};
And we want to order by cost, so we define the less than operator for this structure as follows:
bool operator<(const node &leftNode, const node &rightNode) {
if (leftNode.cost != rightNode.cost) return leftNode.cost < rightNode.cost;
if (leftNode.at != rightNode.at) return leftNode.at < rightNode.at;
return false;
}
Even though we don’t need to order by the ‘at’ member of the structure, we still do otherwise elements with the same cost but different
‘at’ values may be coalesced into one value. The return false at the end is to ensure that if two duplicate elements are compared the
less than operator will return false.
Java users unfortunately have to do a bit of makeshift work, as there is not a direct implementation of the Heap structure. We can
approximate it with the TreeSet structure which will do full ordering of our dataset. It is less space efficient, but will serve our purposes
fine.
import java.util.*;
TreeSet pq = new TreeSet();
1. Add ‐ boolean add(Object o)
2. Pop ‐ boolean remove(Object o)
In this case, we can remove anything we want, but pop should remove the first element, so we will always call it like
this: pq.remove(pq.first());
3. Top ‐ Object first()
4. Empty ‐ int size()
To define the ordering we do something quite similar to what we use in C++:
class Node implements Comparable {
public int cost, at;
public int CompareTo(Object o) {
Node right = (Node)o;
if (cost < right.cost) return ‐1;
if (cost > right.cost) return 1;
if (at < right.at) return ‐1;
if (at > right.at) return 1;
return 0;
}
}
C# users also have the same problem, so they need to approximate as well, unfortunately the closest thing to what we want that is
currently available is the SortedList class, and it does not have the necessary speed (insertions and deletions are O(n) instead of O(log
n)). Unfortunately there is no suitable builtin class for implementing heap based algorithms in C#, as the HashTable is not suitable
either.
Getting back to the actual algorithm now, the beautiful part is that it applies as well to graphs with weighted edges as the Breadth First
search does to graphs with unweighted edges. So we can now solve much more difficult problems (and more common on topcoder)
than is possible with just the Breadth First search.
There are some extremely nice properties as well, since we are picking the node with the least total cost so far to explore first, the first
time we visit a node is the best path to that node (unless there are negative weight edges in the graph). So we only have to visit each
node once, and the really nice part is if we ever hit the target node, we know that we are done.
For the example here we will be using KiloManX, from SRM 181, the Div 1 1000. This is an excellent example of the application of the
Heap Dijkstra problem to what appears to be a Dynamic Programming question initially. In this problem the edge weight between nodes
changes based on what weapons we have picked up. So in our node we at least need to keep track of what weapons we have picked
up, and the current amount of shots we have taken (which will be our cost). The really nice part is that the weapons that we have
picked up corresponds to the bosses that we have defeated as well, so we can use that as a basis for our visited structure. If we
represent each weapon as a bit in an integer, we will have to store a maximum of 32,768 values (2^15, as there is a maximum of 15
weapons). So we can make our visited array simply be an array of 32,768 booleans. Defining the ordering for our nodes is very easy in
this case, we want to explore nodes that have lower amounts of shots taken first, so given this information we can define our basic
structure to be as follows:
boolean visited[32768];
class node {
int weapons;
int shots;
// Define a comparator that puts nodes with less shots on top appropriate to your language
};
Now we will apply the familiar structure to solve these types of problems.
int leastShots(String[] damageChart, int[] bossHealth) {
priorityQueue pq;
pq.push(node(0, 0));
while (pq.empty() == false) {
node top = pq.top();
pq.pop();
// Make sure we don't visit the same configuration twice
if (visited[top.weapons]) continue;
visited[top.weapons] = true;
// A quick trick to check if we have all the weapons, meaning we defeated all the bosses.
// We use the fact that (2^numWeapons ‐ 1) will have all the numWeapons bits set to 1.
if (top.weapons == (1 << numWeapons) ‐ 1)
return top.shots;
for (int i = 0; i < damageChart.length; i++) {
// Check if we've already visited this boss, then don't bother trying him again
if ((top.weapons >> i) & 1) continue;
// Now figure out what the best amount of time that we can destroy this boss is, given the weapons we have.
// We initialize this value to the boss's health, as that is our default (with our KiloBuster).
int best = bossHealth[i];
for (int j = 0; j < damageChart.length; j++) {
if (i == j) continue;
if (((top.weapons >> j) & 1) && damageChart[j][i] != '0') {
// We have this weapon, so try using it to defeat this boss
int shotsNeeded = bossHealth[i] / (damageChart[j][i] ‐ '0');
if (bossHealth[i] % (damageChart[j][i] ‐ '0') != 0) shotsNeeded++;
best = min(best, shotsNeeded);
}
}
// Add the new node to be searched, showing that we defeated boss i, and we used 'best' shots to defeat him.
pq.add(node(top.weapons | (1 << i), top.shots + best));
}
}
}
There are a huge number of these types of problems on topcoder; here are some excellent ones to try out:
SRM 150 – Div 1 1000 – RoboCourier
SRM 194 – Div 1 1000 – IslandFerries
SRM 198 – Div 1 500 – DungeonEscape
TCCC ’04 Round 4 – 500 – Bombman
FloydWarshall
FloydWarshall is a very powerful technique when the graph is represented by an adjacency matrix. It runs in O(n^3) time, where n is
the number of vertices in the graph. However, in comparison to Dijkstra, which only gives us the shortest path from one source to the
targets, FloydWarshall gives us the shortest paths from all source to all target nodes. There are other uses for FloydWarshall as well;
it can be used to find connectivity in a graph (known as the Transitive Closure of a graph).
First, however we will discuss the Floyd Warshall AllPairs Shortest Path algorithm, which is the most similar to Dijkstra. After running
the algorithm on the adjacency matrix the element at adj[i][j] represents the length of the shortest path from node i to node j. The
pseudocode for the algorithm is given below:
for (k = 1 to n)
for (i = 1 to n)
for (j = 1 to n)
adj[i][j] = min(adj[i][j], adj[i][k] + adj[k][j]);
As you can see, this is extremely simple to remember and type. If the graph is small (less than 100 nodes) then this technique can be
used to great effect for a quick submission.
An excellent problem to test this out on is the Division 2 1000 from SRM 184, TeamBuilder.
Greedy is Good
By supernova — topcoder member
Discuss this article in the forums
John Smith is in trouble! He is a topcoder member and once he learned to master the "Force" of dynamic programming, he began
solving problem after problem. But his once obedient computer acts quite unfriendly today. Following his usual morning ritual, John
woke up at 10 AM, had a cup of coffee and went to solve a problem before breakfast. Something didn’t seem right from the beginning,
but based on his vast newly acquired experience, he wrote the algorithm in a flash. Tired of allocating matrices morning after morning,
the computer complained: "Segmentation fault!". Despite his empty stomach, John has a brilliant idea and gets rid of his beloved
matrix by adding an extra "for cycle". But the computer cries again: "Time limit exceeded!"
Instead of going nuts, John makes a radical decision. Enough programming, he says! He decides to take a vacation as a reward for his
hard work.
Being a very energetic guy, John wants to have the time of his life! With so many things to do, it is unfortunately impossible for him to
enjoy them all. So, as soon as he eats his breakfast, he devises a "Fun Plan" in which he describes a schedule of his upcoming
activities:
ID Scheduled Activity Time Span
Monday, 10:00
1 Debug the room PM – Tuesday,
1:00 AM
Tuesday, 6:00
2 Enjoy a trip to Hawaii AM – Saturday,
10:00 PM
Tuesday, 11:00
Win the Chess
3 AM – Tuesday,
Championship
9:00 PM
Tuesday, 7:00
4 Attend the Rock Concert PM – Tuesday,
11:00 PM
Wednesday,
Win the Starcraft 3:00 PM –
5
Tournament Thursday, 3:00
PM
Thursday,
10:00 AM –
6 Have some paintball fun
Thursday, 4:00
PM
Saturday,
Participate in the
12:00 PM –
7 topcoder Single Round
Saturday, 2:00
Match
PM
Saturday, 8:30
8 Take a shower PM – Saturday
8:45 PM
Saturday, 9:00
Organize a Slumber
9 PM – Sunday,
Party
6:00 AM
Participate in an "All you Saturday, 9:01
10 can eat" and "All you PM – Saturday,
can drink" challenge 11:59 PM
He now wishes to take advantage of as many as he can. Such careful planning requires some cleverness, but his mind has gone on
vacation too. This is John Smith’s problem and he needs our help.
Could we help him have a nice holiday? Maybe we can! But let’s make an assumption first. As John is a meticulous programmer, once
he agrees on something, he sticks to the plan. So, individual activities may either be chosen or not. For each of the two choices
regarding the first activity, we can make another two choices regarding the second. After a short analysis, we find out that we have 2 ^
N possible choices, in our case 1024. Then, we can check each one individually to see whether it abides the time restrictions or not.
From these, finding the choice with the most activities selected should be trivial. There are quite a lot of alternatives, so John would
need to enlist the help of his tired computer. But what happens if we have 50 activities? Even with the most powerful computer in the
world, handling this situation would literally take years. So, this approach is clearly not feasible.
Let’s simply the problem and trust our basic instinct for a moment. A good approach may be to take the chance as the first opportunity
arises. That is, if we have two activities we can follow and they clash, we choose the one that starts earlier in order to save some time.
In this case John will start his first evening by debugging his room. Early the next morning, he has a plane to catch. It is less than a
day, and he has already started the second activity. This is great! Actually, the best choice for now. But what happens next? Spending
5 days in Hawaii is time consuming and by Saturday evening, he will still have only two activities performed. Think of all the activities
he could have done during this five day span! Although very fast and simple, this approach is unfortunately not accurate.
We still don’t want to check for every possible solution, so let’s try another trick. Committing to such a time intensive activity like the
exotic trip to Hawaii can simply be avoided by selecting first the activity which takes the least amount of time and then continuing this
process for the remaining activities that are compatible with those already selected. According to the previous schedule, first of all we
choose the shower. With only 15 minutes consumed, this is by far the best local choice. What we would like to know is whether we
can still keep this "local best" as the other compatible activities are being selected. John’s schedule will look like this:
Take a shower (15 minutes)
Participate in the topcoder Single Round Match (2 hours)
Participate in an "All you can eat" and "All you can drink" challenge (2 hours 58 minutes)
Debug the room (3 hours)
Attend the Rock Concert (4 hours)
Have some paintball fun (6 hours)
Out of the 10 possible activities, we were able to select 6 (which is not so bad). We now run the slow but trustworthy algorithm to see if
this is actually the best choice we can make. And the answer is indeed 6. John is very appreciative for our help, but once he returns
from the holiday, confident in our ingenious approach, he may face a serious problem:
By going for the short date, he misses both the school exam and the match of his favorite team. Being the topcoders that we are, we
must get used to writing reliable programs. A single case which we cannot handle dooms this approach to failure.
What we generally have to do in situations like this is to analyze what might have caused the error in the first place and act accordingly
to avoid it in the future. Let’s look again at the previous scenario. The dating activity clashes with both the exam and the match, while
the other two only clash with the date. So, the idea almost comes from itself. Why not always select the activity that produces the
minimum amount of clashes with the remaining activities? Seems logical – it all makes sense now! We’ll try to prove that this approach
is indeed correct. Suppose we have already selected an activity X and try to check if we could have selected two activities A and B that
clash with X instead. A and B should of course not clash, otherwise the final result will not improve. But now, we are back to the
previous case (X has two clashes, while A and B have only one). If this is the case, A and B are selected from the beginning. The only
way to disprove our assumption is to make A and B clash more, without affecting other activities except X. This is not very intuitive, but
if we think it through we can (unfortunately) build such a case:
The activities represented by the blue lines are the optimal choice given the above schedule. But as the activity in red produces only 2
clashes, it will be chosen first. There are 4 compatible activities left before, but they all clash with each other, so we can only select one.
The same happens for the activities scheduled after, leaving space for only one more choice. This only gives us 3 activities, while the
optimum choice selects 4.
So far, every solution we came up with had a hidden flaw. It seems we have to deal with a devilish problem. Actually, this problem has
quite an elegant and straightforward solution. If we study the figure above more carefully, we see that the blue activity on the bottom
left is the only one which finishes before the "timeline" indicated by the thin vertical bar. So, if we are to choose a single activity,
choosing the one that ends first (at a time t1), will leave all the remaining time interval free for choosing other activities. If we choose
any other activity instead, the remaining time interval will be shorter. This is obvious, because we will end up anyway with only one
activity chosen, but at a time t2 > t1. In the first case we had available all the time span between t1 and finish and that included the
time between t2 and finish. Consequently, there is no disadvantage in choosing the activity that finishes earlier. The advantage may
result in the situation when we are able to insert another activity that starts between t1 and t2 and ends up before the end of any
activity that starts after time t2.
Known as the "Activity Selection", this is a standard problem that can be solved by the Greedy Method. As a greedy man takes as
much as he can as often as he can, in our case we are choosing at every step the activity that finishes first and do so every time there
is no activity in progress. The truth is we all make greedy decisions at some point in our life. When we go shopping or when we drive a
car, we make choices that seem best for the moment. Actually, there are two basic ingredients every greedy algorithm has in common:
Greedy Choice Property: from a local optimum we can reach a global optimum, without having to reconsider the decisions
already taken.
Optimal Substructure Property: the optimal solution to a problem can be determined from the optimal solutions to its
subproblems.
The following pseudo code describes the optimal activity selection given by the "greedy" algorithm proven earlier:
Let N denote the number of activities and
{I} the activity I ( 1 <= I <= N )
For each {I}, consider S[I] and F[I] its starting and finishing time
Sort the activities in the increasing order of their finishing time
‐ that is, for every I < J we must have F [I] <= F [J]
// A denotes the set of the activities that will be selected
A = {1}
// J denotes the last activity selected
J = 1
For I = 2 to N
// we can select activity 'I' only if the last activity
// selected has already been finished
If S [I] >= F [J]
// select activity 'I'
A = A + {I}
// Activity 'I' now becomes the last activity selected
J = I
Endif
Endfor
Return A
After applying the above algorithm, Johnny’s "Fun Plan" would look like this:
Eliminate all the bugs and take some time to rest
Tuesday is for chess, prepare to beat them all
A whole day of Starcraft follows, this should be fun
The next two days are for recovery
As for the final day, get a few rating points on topcoder, take a shower and enjoy the versatile food and the good quality wine
The problem of John Smith is solved, but this is just one example of what Greedy can do. A few examples of real topcoder problems will
help you understand the concept better. But before moving on, you may wish to practice a little bit more what you have read so far on
a problem similar with the Activity Selection, named Boxing.
BioScore
In this problem you are asked to maximize the average homology score for all the pairs in the set. As an optimal solution is required,
this may be a valuable clue in determining the appropriate method we can use. Usually, this kind of problems can be solved by dynamic
programming, but in many cases a Greedy strategy could also be employed.
The first thing we have to do here is to build the frequency matrix. This is an easy task as you just have to compare every pair of
two sequences and count the occurrences of all the combinations of nucleic acids (AA, AC, AG, AT, CA, CC, CG, CT, GA, GC, GG, GT, TA,
TC, TG, TT). Each of these combinations will be an element in the matrix and its value will represent the total number of occurrences.
For example, let’s take the set { "ACTAGAGAC", "AAAAAAAAA", "TAGTCATAC", "GCAGCATTC" } used in Example 2.
In the bottomright part of the figure above, you can see the resulting frequency matrix. Let us denote it by F What we have to do from
now is to find another matrix S such that the sum of the 16 corresponding products of the type F[I,J] * S[I,J] (1 <= I,J <= 4) is
maximized.
Now, let’s look at the matrix restrictions and analyze them one by one:
1) The sum of the 16 entries must be 0.
This is more like a commonsense condition. With all the elements in F positive, the final score tends to increase as we increase the
elements in S. But because the sum must be kept at 0, in order to increase an element, we’ll have to decrease others. The challenge of
this problem resides in finding the optimal distribution.
2) All entries must be integers between 10 and 10 inclusive
Another commonsense condition! Our search space has been drastically reduced, but we are still left with a lot of alternatives.
3) It must be symmetric ( score(x,y) = score(y,x) )
Because of the symmetry, we must attribute the same homology score to combinations like "AC" and "CA". As a result, we can also
count their occurrences together. For the previous example, we have the set of combinations with the following frequencies:
CG + GC: 2 CT + TC: 0
GT + TG: 3
An intuitive approach would be to assign a higher homology score to the combinations that appear more often. But as we must keep
the score sum to 0, another problem arises. Combinations like AA, CC, GG and TT appear only once in the matrix. So, their homology
score contribute less to the total sum.
4) Diagonal entries must be positive ( score(x,x)>0 )
This restriction differentiates the elements on the diagonal from the others even further. Basically, we have two groups: the four
elements on the diagonal (which correspond to the combinations AA, CC, GG and TT) and the six elements not on the diagonal (which
correspond to the combinations AC + CA, AG + GA, AT + TA, CG + GC, CT + TC and GT +TG). Each of these groups can have different
states, depending on the value we assign to their elements.
To make things easier, for each possible state in the first group we wish to find an optimal state for the second group. As all
the elements in the second group have the same property, we will try to find their optimal state by using a Greedy approach. But
because the elements in the first group can take any values between 1 and 10, the sum we wish to obtain for the scores we choose in
the second group has to be recalculated. It’s easy to notice that the sum of the elements in the first group can range anywhere
between 4 and 40. As a result, depending on the choice we make for the first group, we’ll have to obtain a sum between 2 and 20 for
the second (we shall not forget that the symmetrical elements in the matrix have been coupled together, thus they count twice in the
score matrix).
Now, we have finally reached to the problem core. The solution to the entire problem depends on finding the optimal choice for the
scores in the second group. If the problem has indeed the greedy choice property and the optimal substructure property, we’ll be
able to pick one element form the group, assign it the best scenario and proceed with the remaining elements in the same manner.
Claim: If we always give the highest possible score to the combination that has the most occurrences in the group, we’ll
obtain in the end the highest possible score for the entire group.
The first thing we have to do is to sort these six elements in matrix F. Then, we have to actually compute the corresponding score
values in S. As the total score we should obtain is at least 20, one quick insight tells us that the first two elements could be given a
score of 10 (if we assign 10 to all the remaining four elements, 20 can still be achieved). We know as well that the final score is less
than 0. Because we want to maximize the scores for the first elements, the last three elements can only be 10 (in the best case the
score sum of the elements is 2 and then, we assign scores in the following manner: [10, 10, 8, 10, 10, 10]). Finally, the value of the
third element will depend on the choices we make for the first group. From the maximum of 10, we subtract half of the score sum of
the elements in the first group (we should note here that the aforementioned sum must be even).
Now, we have to make sure that our approach is indeed correct. The proof is quite straightforward, as in order keep the sum in S
constant we can only decrease from the score of a combination with more occurrences and increase to the score of a combination with
fewer occurrences. Let f1 and f2 be the frequencies of the two combinations and f1 >= f2. We have f1 * s1 + f2 * s2 = X, where X is
the sum we should maximize. By our greedy assumption, s1 >= s2. As s1 + s2 remains constant, the previous sum changes to: f1*
(s1 – a) + f2*( s2 + a) = Y, where a is strictly greater than 0. We find out that Y – X = a * (f2 – f1). Because f1 >= f2, this difference
will always be less than or equal to 0. It results that Y <= X. As Y was chosen arbitrarily, it can be concluded that the initial greedy
choice always gives the maximum possible score.
We apply the algorithm described above for each state of the elements in the first group and save the best result.
Representation: Instead of using the matrices F and S, we find it more convenient to use arrays for storing both the
combination frequencies and their corresponding score. The first 4 elements of F will denote the frequency of the combinations
AA, CC, GG and TT. The next 6 elements will denote the other possible combinations and are sorted in the decreasing order of their
frequency (F[5] >= F[6] >= F[7] >= F[8] >= F[9] >= F[10]). S will be an array of 10 elements such that S[I] is the score we attribute
to the combination I.
The main algorithm is illustrated in the following pseudo code:
Best = ‐Infinity
For S [1] = 1 to 10
For S [2] = 1 to 10
For S [3] = 1 to 10
For S [4] = 1 to 10
If (S [1] + S [2] + S [3] + S [4]) mod 2 = 0
S [5] = S[6] = 10
S [7] = 10 ‐ (S [1] + S [2] + S [3] + S[4]) / 2
S [8] = S [9] = S [10] = ‐10
// in Best we save the greatest average homology score
Best = max (Best , score (F,S))
// obtained so far.
Endif
Endfor
Endfor
Endfor
Endfor
Return Best
Given the score matrix (in our case the array S), we compute the final result by just making the sum of the products of the form F[I] *
S[I] ( 1 <= I <=10) and divide it by N * (N1) / 2 in order to obtain the average homology score.
GoldMine
We are now going to see how a gold mine can be exploited to its fullest, by being greedy. Whenever we notice the maximum profit is
involved, a greedy switch should activate. In this case, we must allocate all the miners to the available mines, such that the total profit
is maximized. After a short analysis, we realize that we want to know how much money can be earned from a mine in all the possible
cases. And there are not so many cases, as in each mine we can only have between 0 and 6 workers. The table below represents the
possible earnings for the two mines described in the example 0 of the problem statement:
First mine 0 57 87 87 67 47 27
Second mine 0 52 66 75 75 66 48
As we are going to assign workers to different mines, we may be interested in the profit a certain worker can bring to the mine he was
assigned. This can be easily determined, as we compute the difference between the earnings resulted from a mine with the worker and
without. If we only had one worker, the optimal choice would have been to allocate him in the mine where he can bring the best
profit. But as we have more workers, we want to check if assigning them in the same manner would bring the best global profit.
In our example we have 4 workers that must be assigned. The table below shows the profit obtained in the two mines for each
additional worker.
Second mine 52 14 9 0 9 20
We notice that the first mine increases its profit by 57 if we add a worker, while the second by only 52. So, we allocate the first worker
to the first mine.
Second mine 52 14 9 0 9 20
Now, an additional worker assigned to the first mine would only increase its profit by 30. We put him in the second, where the profit
can be increased by 52.
Second mine 52 14 9 0 9 20
The third miner would be more useful to the first mine as he can bring a profit of 30.
Second mine 52 14 9 0 9 20
As for the last miner, we can either place him in the first mine (for a zero profit) or in the second (for a profit of 14). Obviously, we
assign him to the second.
Second mine 52 14 9 0 9 20
In the end two of the workers have been allocated to the first mine and another two to the second. The example shows us that this is
indeed the choice with the best total profit. But will our "greedy" approach always work?
Claim: We obtain the maximum total profit when we assign the workers one by one to the mine where they can bring the
best immediate profit.
Proof: Let A and B be two mines and a1, a2, b1, b2 be defined as below:
a1 – the profit obtained when an additional worker is assigned to mine A
a1 + a2 – the profit obtained when two additional workers are assigned to mine A
b1 – the profit obtained when an additional worker is assigned to mine B
b1 + b2 – the profit obtained when two additional workers are assigned to mine B
Let us now consider that we have two workers to assign and a1 >= b1.
Our greedy algorithm will increase the profit by a1 for the first worker and by max (a2, b1) for the second worker. The total profit in
this case is a1+max(a2,b1). If we were to choose the profit b1 for the first worker instead, the alternatives for the second worker
would be a profit of a1 or a profit of b2.
In the first case, the total profit would be b1+a1 <= a1 + max (a2,b1).
In the second case, the total profit would be b1+b2. We need to prove that b1+b2 <= a1+max(a2,b1). But b2 <= b1 as the profit of
allocating an extra worker to a mine is always higher or equal with the profit of allocating the next extra worker to that
mine.
number of ores > number of workers + 2 60 60
number of ores = number of workers + 2 60 50
number of ores = number of workers + 1 50 20
As b1+b2 <= a1+b2 <= a1+b1 <= a1+max(a2,b1), the greedy choice is indeed the best .
Coding this is not difficult, but one has to take into account the problem constraints (all miners must be placed, there are at most six
workers in a mine and if a worker can be optimally assigned to more than one mine, put him in the mine with the lowest index).
WorldPeace
The greedy algorithms we have seen so far work well in every possible situation as their correction has been proven. But there is
another class of optimization problems where Greedy Algorithms have found their applicability. This category mostly includes NP
complete problems (like the Traveling Salesman Problem) and here, one may prefer to write an heuristic based on a greedy algorithm
than to wait … The solution is not always the best, but for most real purposes, it is good enough. While this problem is not NP, it is an
excellent example of how a simple greedy algorithm can be adapted to fool not only the examples, but also the carefully designed
system tests. Such an algorithm is not very hard to come with and after a short analysis we notice that in order to maximize the total
number of groups it is always optimal to form a group from the k countries that have the highest number of citizens. We
apply this principle at every single step and then sort the sequence again to see which are the next k countries having the highest
number of citizens. This idea is illustrated in the following pseudo code:
Groups = 0
Repeat
//sorts the array in decreasing order
Sort (A)
Min= A[K]
If Min > 0 Groups = Groups + 1
For I = 1 to K
A[I] = A[I] ‐ 1
Endfor
Until Min = 0
Return Groups
Unfortunately, a country can have up to a billion citizens, so we cannot afford to make only one group at a time. Theoretically, for a
given set of k countries, we can make groups until all the citizens in one of these countries have been grouped. And this can be done in
a single step:
Groups = 0
Repeat
// sorts the array in decreasing order
Sort (A)
Min= A[K]
Groups = Groups + Min
For I = 1 to K
A[I] = A[I] ‐ Min
Endfor
Until Min = 0
Return Groups
The execution time is no longer a problem, but it is the algorithm! As we check it on the example 0, our method returns 4 instead of 5.
The result returned for the examples 1, 2 and 3 is correct. As for the last example, instead of making 3983180234 groups, we are able
to make 3983180207. Taking into account the small difference, we may say that our solution is pretty good, so maybe we can refine it
more on this direction.
So far, we have two algorithms:
a first greedy algorithm that is accurate, but not fast enough
a second greedy algorithm that is fast, but not very accurate.
What we want to do is to optimize accuracy as much as we can, without exceeding the execution time limit. Basically, we are looking for
a truce between speed and accuracy. The only difference in the two algorithms described above is the number of groups we select
at a given time. The compromise we will make is to select an arbitrarily large number of groups in the beginning, and as we approach
the end to start being more cautious. When we are left with just a few ungrouped citizens in every country, it makes complete sense to
use the safe brute force approach. In the variable Allowance defined in the algorithm below, we control the number of groups we want
to make at a given moment.
Groups = 0
Repeat
// sorts the array in decreasing order
Sort (A)
Min= A[K]
Allowance = (Min+999) / 1000
Groups = Groups + Allowance
For I = 1 to K
A[I] = A[I] ‐ Allowance
Endfor
Until Min = 0
Return Groups
If this approach is correct indeed, remains to be seen. Despite the fact it escaped both Tomek’s keen eyes and system tests, it is very
likely that the result is not optimal for all the set of possible test cases. This was just an example to show that a carefully chosen
refinement on a simple (but obvious faulty) greedy approach can actually be the "right" way. For more accurate solutions to this
problem, see the Match Editorial.
Conclusion
Greedy algorithms are usually easy to think of, easy to implement and run fast. Proving their correctness may require rigorous
mathematical proofs and is sometimes insidious hard. In addition, greedy algorithms are infamous for being tricky. Missing even a very
small detail can be fatal. But when you have nothing else at your disposal, they may be the only salvation. With backtracking or
dynamic programming you are on a relatively safe ground. With greedy instead, it is more like walking on a mined field. Everything
looks fine on the surface, but the hidden part may backfire on you when you least expect. While there are some standardized problems,
most of the problems solvable by this method call for heuristics. There is no general template on how to apply the greedy method to a
given problem, however the problem specification might give you a good insight. Advanced mathematical concepts such as matroids
may give you a recipe for proving that a class of problems can be solved with greedy, but it ultimately comes down to the keen sense
and experience of the programmer. In some cases there are a lot of greedy assumptions one can make, but only few of them are
correct (see the Activity Selection Problem). In other cases, a hard problem may hide an ingenious greedy shortcut, like there was the
case in the last problem discussed, WorldPeace. And this is actually the whole beauty of greedy algorithms! Needless to say, they can
provide excellent challenge opportunities…
A few final notes
a problem that seems extremely complicated on the surface (see TCSocks) might signal a greedy approach.
problems with a very large input size (such that a n^2 algorithm is not fast enough) are also more likely to be solved by greedy
than by backtracking or dynamic programming.
despite the rigor behind them, you should look to the greedy approaches through the eyes of a detective, not with the glasses of a
mathematician.
in addition, study some of the standard greedy algorithms to grasp the concept better (Fractional Knapsack Problem, Prim
Algorithm, Kruskal Algorithm, Dijkstra Algorithm, Huffman Coding, Optimal Merging, Topological Sort).
Further Problems
Level 1
GroceryBagger – SRM 222
FanFailure – SRM 195
PlayGame – SRM 217
SchoolAssembly – TCO04 Round 2
RockStar – SRM 216
Apothecary – SRM 204
Boxing – TCO04 Round 3
Unblur – TCO04 Semifinal Room 3
Level 2
Crossroads – SRM 217
TCSocks – SRM 207
HeatDeath – TCO04 Round 4
BioScore – TCO04 Semifinal Room 1
Rationalization – SRM 224
Level 3
GoldMine – SRM 169
MLBRecord – TCO04 Round 2
RearrangeFurniture – SRM 220
WorldPeace – SRM 204
Dynamic Programming – From Novice to Advanced
By Dumitru — topcoder member
Discuss this article in the forums
An important part of given problems can be solved with the help of dynamic programming (DP for short). Being able to tackle problems
of this type would greatly increase your skill. I will try to help you in understanding how to solve problems using DP. The article is based
on examples, because a raw theory is very hard to understand.
Note: If you’re bored reading one section and you already know what’s being discussed in it – skip it and go to the next one.
Introduction (Beginner)
What is a dynamic programming, how can it be described?
A DP is an algorithmic technique which is usually based on a recurrent formula and one (or some) starting states. A subsolution of the
problem is constructed from previously found ones. DP solutions have a polynomial complexity which assures a much faster running
time than other techniques like backtracking, bruteforce etc.
Now let’s see the base of DP with the help of an example:
Given a list of N coins, their values (V1, V2, … , VN), and the total sum S. Find the minimum number of coins the sum of which is S (we
can use as many coins of one type as we want), or report that it’s not possible to select coins in such a way that they sum up to S.
Now let’s start constructing a DP solution:
First of all we need to find a state for which an optimal solution is found and with the help of which we can find the optimal solution for
the next state.
What does a "state" stand for?
It’s a way to describe a situation, a subsolution for the problem. For example a state would be the solution for sum i, where i=S. A
smaller state than state i would be the solution for any sum j, where j<i. For finding a state i, we need to first find all smaller states j
(j<i) . Having found the minimum number of coins which sum up to i, we can easily find the next state – the solution for i+1.
How can we find it?
It is simple – for each coin j, Vj=i, look at the minimum number of coins found for the iVjsum (we have already found it previously).
Let this number be m. If m+1 is less than the minimum number of coins already found for current sum i, then we write the new result
for it.
For a better understanding let’s take this example:
Given coins with values 1, 3, and 5.
And the sum S is set to be 11.
First of all we mark that for state 0 (sum 0) we have found a solution with a minimum number of 0 coins. We then go to sum 1. First,
we mark that we haven’t yet found a solution for this one (a value of Infinity would be fine). Then we see that only coin 1 is less than
or equal to the current sum. Analyzing it, we see that for sum 1V1= 0 we have a solution with 0 coins. Because we add one coin to
this solution, we’ll have a solution with 1 coin for sum 1. It’s the only solution yet found for this sum. We write (save) it. Then we
proceed to the next state – sum 2. We again see that the only coin which is less or equal to this sum is the first coin, having a value of
1. The optimal solution found for sum (21) = 1 is coin 1. This coin 1 plus the first coin will sum up to 2, and thus make a sum of 2 with
the help of only 2 coins. This is the best and only solution for sum 2. Now we proceed to sum 3. We now have 2 coins which are to be
analyzed – first and second one, having values of 1 and 3. Let’s see the first one. There exists a solution for sum 2 (3 – 1) and
therefore we can construct from it a solution for sum 3 by adding the first coin to it. Because the best solution for sum 2 that we found
has 2 coins, the new solution for sum 3 will have 3 coins. Now let’s take the second coin with value equal to 3. The sum for which this
coin needs to be added to make 3 , is 0. We know that sum 0 is made up of 0 coins. Thus we can make a sum of 3 with only one coin –
3. We see that it’s better than the previous found solution for sum 3 , which was composed of 3 coins. We update it and mark it as
having only 1 coin. The same we do for sum 4, and get a solution of 2 coins – 1+3. And so on.
Pseudocode:
Set Min[i] equal to Infinity for all of i
Min[0]=0
For i = 1 to S
For j = 0 to N ‐ 1
If (Vj<=i AND Min[i‐Vj]+1<Min[i])
Then Min[i]=Min[i‐Vj]+1
Output Min[S]
Here are the solutions found for all sums:
Coin value added to a smaller sum to
Sum Min. nr. of coins
obtain this sum (it is displayed in brackets)
0 0
1 1 1 (0)
2 2 1 (1)
3 1 3 (0)
4 2 1 (3)
5 1 5 (0)
6 2 3 (3)
7 3 1 (6)
8 2 3 (5)
9 3 1 (8)
10 2 5 (5)
11 3 1 (10)
As a result we have found a solution of 3 coins which sum up to 11.
Additionally, by tracking data about how we got to a certain sum from a previous one, we can find what coins were used in building it.
For example: to sum 11 we got by adding the coin with value 1 to a sum of 10. To sum 10 we got from 5. To 5 – from 0. This way we
find the coins used: 1, 5 and 5.
Having understood the basic way a DP is used, we may now see a slightly different approach to it. It involves the change (update) of
best solution yet found for a sum i, whenever a better solution for this sum was found. In this case the states aren’t calculated
consecutively. Let’s consider the problem above. Start with having a solution of 0 coins for sum 0. Now let’s try to add first coin (with
value 1) to all sums already found. If the resulting sum t will be composed of fewer coins than the one previously found – we’ll update
the solution for it. Then we do the same thing for the second coin, third coin, and so on for the rest of them. For example, we first add
coin 1 to sum 0 and get sum 1. Because we haven’t yet found a possible way to make a sum of 1 – this is the best solution yet found,
and we mark S[1]=1. By adding the same coin to sum 1, we’ll get sum 2, thus making S[2]=2. And so on for the first coin. After the
first coin is processed, take coin 2 (having a value of 3) and consecutively try to add it to each of the sums already found. Adding it to
0, a sum 3 made up of 1 coin will result. Till now, S[3] has been equal to 3, thus the new solution is better than the previously found
one. We update it and mark S[3]=1. After adding the same coin to sum 1, we’ll get a sum 4 composed of 2 coins. Previously we found
a sum of 4 composed of 4 coins; having now found a better solution we update S[4] to 2. The same thing is done for next sums – each
time a better solution is found, the results are updated.
Elementary
To this point, very simple examples have been discussed. Now let’s see how to find a way for passing from one state to another, for
harder problems. For that we will introduce a new term called recurrent relation, which makes a connection between a lower and a
greater state.
Let’s see how it works:
Given a sequence of N numbers – A[1] , A[2] , …, A[N] . Find the length of the longest nondecreasing sequence.
As described above we must first find how to define a "state" which represents a subproblem and thus we have to find a solution for it.
Note that in most cases the states rely on lower states and are independent from greater states.
Let’s define a state i as being the longest nondecreasing sequence which has its last number A[i] . This state carries only data about
the length of this sequence. Note that for i<j the state i is independent from j, i.e. doesn’t change when we calculate state j. Let’s see
now how these states are connected to each other. Having found the solutions for all states lower than i, we may now look for state i.
At first we initialize it with a solution of 1, which consists only of the ith number itself. Now for each j<i let’s see if it’s possible to pass
from it to state i. This is possible only when A[j]=A[i] , thus keeping (assuring) the sequence nondecreasing. So if S[j] (the solution
found for state j) + 1 (number A[i] added to this sequence which ends with number A[j] ) is better than a solution found for i (ie.
S[j]+1>S[i] ), we make S[i]=S[j]+1. This way we consecutively find the best solutions for each i, until last state N.
Let’s see what happens for a randomly generated sequence: 5, 3, 4, 8, 6, 7:
The length of the longest The last sequence i from
I nondecreasing sequence which we "arrived"
of first i numbers to this one
1 1 1 (first number itself)
2 (second number
2 1
itself)
3 2 2
4 3 3
5 3 3
6 4 5
Practice problem:
Given an undirected graph G having N (1<N<=1000) vertices and positive weights. Find the shortest path from vertex 1 to vertex N, or
state that such path doesn’t exist.
Hint: At each step, among the vertices which weren’t yet checked and for which a path from vertex 1 was found, take the one which
has the shortest path, from vertex 1 to it, yet found.
Try to solve the following problems from topcoder competitions:
ZigZag – 2003 TCCC Semifinals 3
BadNeighbors – 2004 TCCC Round 4
FlowerGarden – 2004 TCCC Round 1
Intermediate
Let’s see now how to tackle bidimensional DP problems.
Problem:
A table composed of N x M cells, each having a certain quantity of apples, is given. You start from the upperleft corner. At each step
you can go down or right one cell. Find the maximum number of apples you can collect.
This problem is solved in the same way as other DP problems; there is almost no difference.
First of all we have to find a state. The first thing that must be observed is that there are at most 2 ways we can come to a cell – from
the left (if it’s not situated on the first column) and from the top (if it’s not situated on the most upper row). Thus to find the best
solution for that cell, we have to have already found the best solutions for all of the cells from which we can arrive to the current cell.
From above, a recurrent relation can be easily obtained:
S[i][j]=A[i][j] + max(S[i1][j], if i>0 ; S[i][j1], if j>0) (where i represents the row and j the column of the table , its leftupper
corner having coordinates {0,0} ; and A[i][j] being the number of apples situated in cell i,j).
S[i][j] must be calculated by going first from left to right in each row and process the rows from top to bottom, or by going first from
top to bottom in each column and process the columns from left to right.
Pseudocode:
For i = 0 to N ‐ 1
For j = 0 to M ‐ 1
S[i][j] = A[i][j] +
max(S[i][j‐1], if j>0 ; S[i‐1][j], if i>0 ; 0)
Output S[n‐1][m‐1]
Here are a few problems, from topcoder Competitions, for practicing:
AvoidRoads – 2003 TCO Semifinals 4
ChessMetric – 2003 TCCC Round 4
UpperIntermediate
This section will discuss about dealing DP problems which have an additional condition besides the values that must be calculated.
As a good example would serve the following problem:
Given an undirected graph G having positive weights and N vertices.
You start with having a sum of M money. For passing through a vertex i, you must pay S[i] money. If you don’t have enough money –
you can’t pass through that vertex. Find the shortest path from vertex 1 to vertex N, respecting the above conditions; or state that
such path doesn’t exist. If there exist more than one path having the same length, then output the cheapest one. Restrictions:
1<N<=100 ; 0<=M<=100 ; for each i, 0<=S[i]<=100. As we can see, this is the same as the classical Dijkstra problem (finding the
shortest path between two vertices), with the exception that it has a condition. In the classical Dijkstra problem we would have used a
unidimensional array Min[i] , which marks the length of the shortest path found to vertex i. However in this problem we should also
keep information about the money we have. Thus it would be reasonable to extend the array to something like Min[i][j] , which
represents the length of the shortest path found to vertex i, with j money being left. In this way the problem is reduced to the original
pathfinding algorithm. At each step we find the unmarked state (i,j) for which the shortest path was found. We mark it as visited (not
to use it later), and for each of its neighbors we look if the shortest path to it may be improved. If so – then update it. We repeat this
step until there will remain no unmarked state to which a path was found. The solution will be represented by Min[N1][j] having the
least value (and the greatest j possible among the states having the same value, i.e. the shortest paths to which has the same length).
Pseudocode:
Set states(i,j) as unvisited for all (i,j)
Set Min[i][j] to Infinity for all (i,j)
Min[0][M]=0
While(TRUE)
Among all unvisited states(i,j) find the one for which Min[i][j]
is the smallest. Let this state found be (k,l).
If there wasn't found any state (k,l) for which Min[k][l] is
less than Infinity ‐ exit While loop.
Mark state(k,l) as visited
For All Neighbors p of Vertex k.
If (l‐S[p]>=0 AND
Min[p][l‐S[p]]>Min[k][l]+Dist[k][p])
Then Min[p][l‐S[p]]=Min[k][l]+Dist[k][p]
i.e.
If for state(i,j) there are enough money left for
going to vertex p (l‐S[p] represents the money that
will remain after passing to vertex p), and the
shortest path found for state(p,l‐S[p]) is bigger
than [the shortest path found for
state(k,l)] + [distance from vertex k to vertex p)],
then set the shortest path for state(i,j) to be equal
to this sum.
End For
End While
Find the smallest number among Min[N‐1][j] (for all j, 0<=j<=M);
if there are more than one such states, then take the one with greater
j. If there are no states(N‐1,j) with value less than Infinity ‐ then
such a path doesn't exist.
Here are a few TC problems for practicing:
Jewelry – 2003 TCO Online Round 4
StripePainter – SRM 150 Div 1
QuickSums – SRM 197 Div 2
ShortPalindromes – SRM 165 Div 2
Advanced
The following problems will need some good observations in order to reduce them to a dynamic solution.
Problem StarAdventure – SRM 208 Div 1:
Given a matrix with M rows and N columns (N x M). In each cell there’s a number of apples.
You start from the upperleft corner of the matrix. You can go down or right one cell. You need to arrive to the bottomright corner.
Then you need to go back to the upperleft cell by going each step one cell left or up. Having arrived at this upperleft cell, you need to
go again back to the bottomright cell.
Find the maximum number of apples you can collect.
When you pass through a cell – you collect all the apples left there.
Restrictions: 1 < N, M <= 50 ; each cell contains between 0 and 1000 apples inclusive.
First of all we observe that this problem resembles to the classical one (described in Section 3 of this article), in which you need to go
only once from the topleft cell to the bottomright one, collecting the maximum possible number of apples. It would be better to try to
reduce the problem to this one. Take a good look into the statement of the problem – what can be reduced or modified in a certain way
to make it possible to solve using DP? First observation is that we can consider the second path (going from bottomright cell to the
topleft cell) as a path which goes from topleft to bottomright cell. It makes no difference, because a path passed from bottom to top,
may be passed from top to bottom just in reverse order. In this way we get three paths going from top to bottom. This somehow
decreases the difficulty of the problem. We can consider these 3 paths as left, middle and right. When 2 paths intersect (like in the
figure below)
we may consider them as in the following picture, without affecting the result:
This way we’ll get 3 paths, which we may consider as being one left, one middle and the other – right. More than that, we may see that
for getting an optimal results they must not intersect (except in the leftmost upper corner and rightmost bottom corner). So for each
row y (except first and last), the x coordinates of the lines (x1[y] , x2[y] and respectively x3[y] ) will be : x1[y] < x2[y] < x3[y] .
Having done that – the DP solution now becomes much clearer. Let’s consider the row y. Now suppose that for any configuration of
x1[y1] , x2[y1] and x3[y1] we have already found the paths which collect the maximum number of apples. From them we can
find the optimal solution for row y. We now have to find only the way for passing from one row to the next one. Let Max[i][j][k]
represent the maximum number of apples collected till row y1 inclusive, with three paths finishing at column i, j, and respectively k.
For the next row y, add to each Max[i][j][k] (obtained previously) the number of apples situated in cells (y,i) , (y,j) and (y,k). Thus
we move down at each step. After we made such a move, we must consider that the paths may move in a row to the right. For keeping
the paths out of an intersection, we must first consider the move to the right of the left path, after this of the middle path, and then of
the right path. For a better understanding think about the move to the right of the left path – take every possible pair of, k (where
j<k), and for each i (1 i<j) consider the move from position (i1,j,k) to position (i,j,k). Having done this for the left path, start
processing the middle one, which is done similarly; and then process the right path.
TC problems for practicing:
MiniPaint – SRM 178 Div 1
Additional Note:
When have read the description of a problem and started to solve it, first look at its restrictions. If a polynomialtime algorithm should
be developed, then it’s possible that the solution may be of DP type. In this case try to see if there exist such states (subsolutions)
with the help of which the next states (subsolutions) may be found. Having found that – think about how to pass from one state to
another. If it seems to be a DP problem, but you can’t define such states, then try to reduce the problem to another one (like in the
example above, from Section 5).
Mentioned in this writeup:
TCCC ’03 Semifinals 3 Div I Easy – ZigZag
TCCC ’04 Round 4 Div I Easy – BadNeighbors
TCCC ’04 Round 1 Div I Med – FlowerGarden
TCO ’03 Semifinals 4 Div I Easy – AvoidRoads
TCCC ’03 Round 4 Div I Easy – ChessMetric
TCO ’03 Round 4 Div I Med – Jewelry
SRM 150 Div I Med – StripePainter
SRM 197 Div II Hard – QuickSums
SRM 165 Div II Hard – ShortPalindromes
SRM 208 Div I Hard – StarAdventure
SRM 178 Div I Hard – MiniPaint
Computational Complexity: Section 1
By misof — topcoder member
Discuss this article in the forums
In this article I’ll try to introduce you to the area of computation complexity. The article will be a bit long before we get to the actual
formal definitions because I feel that the rationale behind these definitions needs to be explained as well – and that understanding the
rationale is even more important than the definitions alone.
Why is it important?
Example 1. Suppose you were assigned to write a program to process some records your company receives from time to time. You
implemented two different algorithms and tested them on several sets of test data. The processing times you obtained are in Table 1.
# of
10 20 50 100 1000 5000
records
In praxis, we probably could tell which of the two implementations is better for us (as we usually can estimate the amount of data we
will have to process). For the company this solution may be fine. But from the programmer’s point of view, it would be much better if
he could estimate the values in Table 1 before writing the actual code – then he could only implement the better algorithm.
The same situation occurs during programming challenges: The size of the input data is given in the problem statement. Suppose I
found an algorithm. Questions I have to answer before I start to type should be: Is my algorithm worth implementing? Will it solve the
largest test cases in time? If I know more algorithms solving the problem, which of them shall I implement?
This leads us to the question: How to compare algorithms? Before we answer this question in general, let’s return to our simple
example. If we extrapolate the data in Table 1, we may assume that if the number of processed records is larger than 1000, algorithm
2 will be substantially faster. In other words, if we consider all possible inputs, algorithm 2 will be better for almost all of them.
It turns out that this is almost always the case – given two algorithms, either one of them is almost always better, or they are
approximately the same. Thus, this will be our definition of a better algorithm. Later, as we define everything formally, this will be the
general idea behind the definitions.
A neat trick
If you thing about Example 1 for a while, it shouldn’t be too difficult to see that there is an algorithm with runtimes similar to those in
Table 2:
The idea behind this algorithm: Check the number of records. If it is small enough, run algorithm 1, otherwise run algorithm 2.
Similar ideas are often used in praxis. As an example consider most of the sort() functions provided by various libraries. Often this
function is an implementation of QuickSort with various improvements, such as:
if the number of elements is too small, run InsertSort instead (as InsertSort is faster for small inputs)
if the pivot choices lead to poor results, fall back to MergeSort
What is efficiency?
Example 2. Suppose you have a concrete implementation of some algorithm. (The example code presented below is actually an
implementation of MinSort – a slow but simple sorting algorithm.)
for (int i=0; i<N; i++)
for (int j=i+1; j<N; j++)
if (A[i] > A[j])
swap( A[i], A[j] );
If we are given an input to this algorithm (in our case, the array A and its size N), we can exactly compute the number of steps our
algorithm does on this input. We could even count the processor instructions if we wanted to. However, there are too many possible
inputs for this approach to be practical.
And we still need to answer one important question: What is it exactly we are interested in? Most usually it is the behavior of our
program in the worst possible case – we need to look at the input data and to determine an upper bound on how long will it take if
we run the program.
But then, what is the worst possible case? Surely we can always make the program run longer simply by giving it a larger input. Some
of the more important questions are: What is the worst input with 700 elements? How fast does the maximum runtime grow when we
increase the input size?
Formal notes on the input size
What exactly is this "input size" we started to talk about? In the formal definitions this is the size of the input written in some fixed
finite alphabet (with at least 2 "letters"). For our needs, we may consider this alphabet to be the numbers 0..255. Then the "input size"
turns out to be exactly the size of the input file in bytes.
Usually a part of the input is a number (or several numbers) such that the size of the input is proportional to the number.
E.g. in Example 2 we are given an int N and an array containing N ints. The size of the input file will be roughly 5N (depending on the
OS and architecture, but always linear in N).
In such cases, we may choose that this number will represent the size of the input. Thus when talking about problems on
arrays/strings, the input size is the length of the array/string, when talking about graph problems, the input size depends both on the
number of vertices (N) and the number of edges (M), etc.
We will adopt this approach and use N as the input size in the following parts of the article.
There is one tricky special case you sometimes need to be aware of. To write a (possibly large) number we need only logarithmic space.
(E.g. to write 123456, we need only roughly log10(123456) digits.) This is why the naive primality test does not run in polynomial time
– its runtime is polynomial in the size of the number, but not in its number of digits! If you didn’t understand the part about
polynomial time, don’t worry, we’ll get there later.
How to measure efficiency?
We already mentioned that given an input we are able to count the number of steps an algorithm makes simply by simulating it.
Suppose we do this for all inputs of size at most N and find the worst of these inputs (i.e. the one that causes the algorithm to do the
most steps). Let f (N) be this number of steps. We will call this function the time complexity, or shortly the runtime of our algorithm.
In other words, if we have any input of size N, solving it will require at most f (N) steps.
Let’s return to the algorithm from Example 2. What is the worst case of size N? In other words, what array with N elements will cause
the algorithm to make the most steps? If we take a look at the algorithm, we can easily see that:
the first step is executed exactly N times
the second and third step are executed exactly N(N – 1)/2 times
the fourth step is executed at most N(N – 1)/2 times
Clearly, if the elements in A are in descending order at the beginning, the fourth step will always be executed. Thus in this case the
algorithm makes 3N(N – 1)/2 + N = 1.5N2 – 0.5N steps. Therefore our algorithm has f (N) = 1.5N2 – 0.5N.
As you can see, determining the exact function f for more complicated programs is painful. Moreover, it isn’t even necessary. In our
case, clearly the 0.5N term can be neglected. It will usually be much smaller than the 1.5N2 term and it won’t affect the runtime
significantly. The result "f (N) is roughly equal to 1.5N2" gives us all the information we need. As we will show now, if we want to
compare this algorithm with some other algorithm solving the same problem, even the constant 1.5 is not that important.
Consider two algorithms, one with the runtime N2, the other with the runtime 0.001N3. One can easily see that for N greater than
1 000 the first algorithm is faster – and soon this difference becomes apparent. While the first algorithm is able to solve inputs with N =
20 000 in a matter of seconds, the second one will already need several minutes on current machines.
Clearly this will occur always when one of the runtime functions grows asymptotically faster than the other (i.e. when N grows
beyond all bounds the limit of their quotient is zero or infinity). Regardless of the constant factors, an algorithm with runtime
proportional to N2 will always be better than an algorithm with runtime proportional to N3 on almost all inputs. And this observation
is exactly what we base our formal definition on.
Finally, formal definitions
Let f, g be positive nondecreasing functions defined on positive integers. (Note that all runtime functions satisfy these conditions.) We
say that f (N) is O(g(N)) (read: f is bigoh of g) if for some c and N0 the following condition holds:
N > N0; f (N) < c.g(N)
In human words, f (N) is O(g(N)), if for some c almost the entire graph of the function f is below the graph of the function c.g. Note
that this means that f grows at most as fast as c.g does.
Instead of "f (N) is O(g(N))" we usually write f (N) = O(g(N)). Note that this "equation" is not symmetric – the notion " O(g(N)) = f
(N)" has no sense and " g(N) = O(f (N))" doesn’t have to be true (as we will see later). (If you are not comfortable with this notation,
imagine O(g(N)) to be a set of functions and imagine that there is a instead of =.)
What we defined above is known as the bigoh notation and is conveniently used to specify upper bounds on function growth.
E.g. consider the function f (N) = 3N(N – 1)/2 + N = 1.5N2 – 0.5N from Example 2. We may say that f (N) = O(N2) (one possibility for
the constants is c = 2 and N0 = 0). This means that f doesn’t grow (asymptotically) faster than N2.
Note that even the exact runtime function f doesn’t give an exact answer to the question "How long will the program run on my
machine?" But the important observation in the example case is that the runtime function is quadratic. If we double the input size, the
runtime will increase approximately to four times the current runtime, no matter how fast our computer is.
The f (N) = O(N2) upper bound gives us almost the same – it guarantees that the growth of the runtime function is at most quadratic.
Thus, we will use the Onotation to describe the time (and sometimes also memory) complexity of algorithms. For the algorithm from
Example 2 we would say "The time complexity of this algorithm is O(N2)" or shortly "This algorithm is O(N2)".
In a similar way we defined O we may define and .
We say that f (N) is (g(N)) if g(N) = O(f (N)), in other words if f grows at least as fast as g.
We say that f (N) = (g(N)) if f (N) = O(g(N)) and g(N) = O(f (N)), in other words if both functions have approximately the same rate
of growth.
1.5N2 0.5N = O(N2).
47N log N = O(N2).
N log N + 1 000 047N = (N log N).
All polynomials of order k are O(Nk).
The time complexity of the algorithm in Example 2 is (N2).
If an algorithm is O(N2), it is also O(N5).
Each comparisionbased sorting algorithm is (N log N).
MergeSort run on an array with N elements does roughly N log N comparisions. Thus the time complexity of MergeSort is (N
log N). If we trust the previous statement, this means that MergeSort is an asymptotically optimal general sorting algorithm.
The algorithm in Example 2 uses (N) bytes of memory.
The function giving my number of teeth in time is O(1).
A naive backtracking algorithm trying to solve chess is O(1) as the tre of positions it will examine is finite. (But of course in this
case the constant hidden behind the O(1) is unbelievably large.)
The statement "Time complexity of this algorithm is at least O(N2)" is meaningless. (It says: "Time complexity of this algorithm is
at least at most roughly quadratic." The speaker probably wanted to say: "Time complexity of this algorithm is (N2).")
When speaking about the time/memory complexity of an algorithm, instead of using the formal (f (n))notation we may simply state
the class of functions f belongs to. E.g. if f (N) = (N), we call the algorithm linear. More examples:
f (N) = (log N): logarithmic
f (N) = (N2): quadratic
f (N) = (N3): cubic
f (N) = O(Nk) for some k: polynomial
f (N) = (2N): exponential
For graph problems, the complexity (N + M) is known as "linear in the graph size".
Determining execution time from an asymptotic bound
For most algorithms you may encounter in praxis, the constant hidden behind the O (or ) is usually relatively small. If an algorithm is
(N2), you may expect that the exact time complexity is something like 10N2, not 107N2.
The same observation in other words: if the constant is large, it is usually somehow related to some constant in the problem statement.
In this case it is good practice to give this constant a name and to include it in the asymptotic notation.
An example: The problem is to count occurences of each letter in a string of N letters. A naive algorithm passes through the whole
string once for each possible letter. The size of alphabet is fixed (e.g. at most 255 in C), thus the algorithm is linear in N. Still, it is
better to write that its time complexity is (| S|.N), where S is the alphabet used. (Note that there is a better algorithm solving this
problem in (| S| + N).)
In a topcoder contest, an algorithm doing 1 000 000 000 multiplications runs barely in time. This fact together with the above
observation and some experience with topcoder problems can help us fill the following table:
complexity maximum N
(N) 100 000 000
(N log N) 40 000 000
(N2) 10 000
(N3) 500
(N4) 90
(2N) 20
(N!) 11
Table 3. Approximate maximum problem size solvable in 8 seconds.
A note on algorithm analysis
Usually if we present an algorithm, the best way to present its time complexity is to give a bound. However, it is common practice to
only give an Obound the other bound is usually trivial, O is much easier to type and better known. Still, don’t forget that O
represents only an upper bound. Usually we try to find an Obound that’s as good as possible.
Example 3. Given is a sorted array A. Determine whether it contains two elements with the difference D. Consider the following code
solving this problem:
int j=0;
for (int i=0; i<N; i++) {
while ( (j<N‐1) && (A[i]‐A[j] > D) )
j++;
if (A[i]‐A[j] == D) return 1;
}
It is easy to give an O(N2) bound for the time complexity of this algorithm – the inner whilecycle is called N times, each time we
increase j at most N times. But a more careful analysis shows that in fact we can give an O(N) bound on the time complexity of this
algorithm – it is sufficient to realize that during the whole execution of the algorithm the command "j++;" is executed no more than N
times.
If we said "this algorithm is O(N2)", we would have been right. But by saying "this algorithm is O(N)" we give more information about
the algorithm.
Conclusion
We have shown how to write bounds on the time complexity of algorithms. We have also demonstrated why this way of characterizing
algorithms is natural and (usually moreorless) sufficient.
The next logical step is to show how to estimate the time complexity of a given algorithm. As we have already seen in Example 3,
sometimes this can be messy. It gets really messy when recursion is involved. We will address these issues in the second part of this
article.
Computational Complexity: Section 2
By misof — topcoder member
In this part of the article we will focus on estimating the time complexity for recursive programs. In essence, this will lead to finding the
order of growth for solutions of recurrence equations. Don’t worry if you don’t understand what exactly is a recurrence solution, we will
explain it in the right place at the right time. But first we will consider a simpler case – programs without recursion.
Nested loops
First of all let’s consider simple programs that contain no function calls. The rule of thumb to find an upper bound on the time
complexity of such a program is:
estimate the maximum number of times each loop can be executed,
add these bounds for cycles following each other.
multiply these bounds for nested cycles/parts of code,
Example 1. Estimating the time complexity of a random piece of code.
int result=0; // 1
for (int i=0; i<N; i++){ // 2
for (int j=i; j<N; j++) { // 3
for (int k=0; k<M; k++) { // 4
int x=0; // 5
while (x<N) { result++; x+=3; } // 6
} // 7
for (int k=0; k<2*M; k++) // 8
if (k%7 == 4) result++; // 9
} // 10
} // 11
The time complexity of the whilecycle in line 6 is clearly O(N) – it is executed no more than N/3 + 1 times.
Now consider the forcycle in lines 47. The variable k is clearly incremented O(M) times. Each time the whole whilecycle in line 6 is
executed. Thus the total time complexity of the lines 47 can be bounded by O(MN).
The time complexity of the forcycle in lines 89 is O(M). Thus the execution time of lines 49 is O(MN + M) = O(MN).
This inner part is executed O(N2) times – once for each possible combination of i and j. (Note that there are only N(N + 1)/2 possible
values for [i, j]. Still, O(N2) is a correct upper bound.)
From the facts above follows that the total time complexity of the algorithm in Example 1 is O(N2.MN) = O(MN3).
From now on we will assume that the reader is able to estimate the time complexity of simple parts of code using the method
demonstrated above. We will now consider programs using recursion (i.e. a function occasionally calling itself with different parameters)
and try to analyze the impact of these recursive calls on their time complexity.
Using recursion to generate combinatorial objects
One common use of recursion is to implement a backtracking algorithm to generate all possible solutions of a problem. The general idea
is to generate the solution incrementally and to step back and try another way once all solutions for the current branch have been
exhausted.
This approach is not absolutely universal, there may be problems where it is impossible to generate the solution incrementally.
However, very often the set of all possible solutions of a problem corresponds to the set of all combinatorial objects of some kind. Most
often it is the set of all permutations (of a given size), but other objects (combinations, partitions, etc.) can be seen from time to time.
As a side note, it is always possible to generate all strings of zeroes and ones, check each of them (i.e. check whether it corresponds to
a valid solution) and keep the best found so far. If we can find an upper bound on the size of the best solution, this approach is finite.
However, this approach is everything but fast. Don’t use it if there is any other way.
Example 2. A trivial algorithm to generate all permutations of numbers 0 to N – 1.
vector<int> permutation(N);
vector<int> used(N,0);
void try(int which, int what) {
// try taking the number "what" as the "which"‐th element
permutation[which] = what;
used[what] = 1;
if (which == N‐1)
outputPermutation();
else
// try all possibilities for the next element
for (int next=0; next<N; next++)
if (!used[next])
try(which+1, next);
used[what] = 0;
}
int main() {
// try all possibilities for the first element
for (int first=0; first<N; first++)
try(0,first);
}
In this case a trivial lower bound on the time complexity is the number of possible solutions. Backtracking algorithms are usually used
to solve hard problems – i.e. such that we don’t know whether a significantly more efficient solution exists. Usually the solution space is
quite large and uniform and the algorithm can be implemented so that its time complexity is close to the theoretical lower bound. To
get an upper bound it should be enough to check how much additional (i.e. unnecessary) work the algorithm does.
The number of possible solutions, and thus the time complexity of such algorithms, is usually exponential – or worse.
Divide&conquer using recursion
From the previous example we could get the feeling that recursion is evil and leads to horribly slow programs. The contrary is true.
Recursion can be a very powerful tool in the design of effective algorithms. The usual way to create an effective recursive algorithm is
to apply the divide&conquer paradigm – try to split the problem into several parts, solve each part separately and in the end combine
the results to obtain the result for the original problem. Needless to say, the "solve each part separately" is usually implemented using
recursion – and thus applying the same method again and again, until the problem is sufficiently small to be solved by brute force.
Example 3. The sorting algorithm MergeSort described in pseudocode.
MergeSort(sequence S) {
if (size of S <= 1) return S;
split S into S_1 and S_2 of roughly the same size;
MergeSort(S_1);
MergeSort(S_2);
combine sorted S_1 and sorted S_2 to obtain sorted S;
return sorted S;
}
Clearly O(N) time is enough to split a sequence with N elements into two parts. (Depending on the implementation this may be even
possible in constant time.) Combining the shorter sorted sequences can be done in (N): Start with an empty S. At each moment the
smallest element not yet in S is either at the beginning of S1 or at the beginning of S2. Move this element to the end of S and continue.
Thus the total time to MergeSort a sequence with N elements is (N) plus the time needed to make the two recursive calls.
Let f (N) be the time complexity of MergeSort as defined in the previous part of our article. The discussion above leads us to the
following equation:
where p is a linear function representing the amount of work spent on splitting the sequence and merging the results.
Basically, this is just a recurrence equation. If you don’t know this term, please don’t be afraid. The word "recurrence" stems from the
latin phrase for "to run back". Thus the name just says that the next values of f are defined using the previous (i.e. smaller) values of f.
Well, to be really formal, for the equation to be complete we should specify some initial values – in this case, f (1). This (and knowing
the implementationspecific function p) would enable us to compute the exact values of f.
But as you hopefully understand by now, this is not necessarily our goal. While it is theoretically possible to compute a closedform
formula for f (N), this formula would most probably be really ugly… and we don’t really need it. We only want to find a bound (and
sometimes only an Obound) on the growth of f. Luckily, this can often be done quite easily, if you know some tricks of the trade.
As a consequence, we won’t be interested in the exact form of p, all we need to know is that p(N) = (N). Also, we don’t need to
specify the initial values for the equation. We simply assume that all problem instances with small N can be solved in constant time.
The rationale behind the last simplification: While changing the initial values does change the solution to the recurrence equation, it
usually doesn’t change its asymptotic order of growth. (If your intuition fails you here, try playing with the equation above. For example
fix p and try to compute f (8), f (16) and f (32) for different values of f (1).)
If this would be a formal textbook, at this point we would probably have to develop some theory that would allow us to deal with the
floor and ceiling functions in our equations. Instead we will simply neglect them from now on. (E.g. we can assume that each division
will be integer division, rounded down.)
A reader skilled in math is encouraged to prove that if p is a polynomial (with nonnegative values on N) and q(n) = p(n + 1) then q(n)
= (p(n)). Using this observation we may formally prove that (assuming the f we seek is polynomiallybounded) the right side of each
such equation remains asymptotically the same if we replace each ceiling function by a floor function.
The observations we made allow us to rewrite our example equation in a more simple way:
(1)
Note that this is not an equation in the classical sense. As in the examples in the first part of this article, the equals sign now reads "is
asymptotically equal to". Usually there are lots of different functions that satisfy such an equation. But usually all of them will have the
same order of growth – and this is exactly what we want to determine. Or, more generally, we want to find the smallest upper bound on
the growth of all possible functions that satisfy the given equation.
In the last sections of this article we will discuss various methods of solving these "equations". But before we can do that, we need to
know a bit more about logarithms.
Notes on logarithms
By now, you may have already asked one of the following questions: If the author writes that some complexity is e.g. O(N log N), what
is the base of the logarithm? In some cases, wouldn’t O(N log2N) be a better bound?
The answer: The base of the logarithm does not matter, all logarithmic functions (with base > 1) are asymptotically equal. This is due
to the wellknown equation:
(2)
Note that given two bases a, b, the number 1/logba is just a constant, and thus the function logaN is just a constant multiple of logbN.
To obtain more clean and readable expressions, we always use the notation log N inside bigOh expressions, even if logarithms with a
different base were used in the computation of the bound.
By the way, sadly the meaning of log N differs from country to country. To avoid ambiguity where it may occur: I use log N to denote
the decadic (i.e. base10) logarithm, ln N for the natural (i.e. basee) logarithm, lg N for the binary logarithm and logbN for the general
case.
Now we will show some useful tricks involving logarithms, we will need them later. Suppose a, b are given constants such that a, b > 1.
From (2) we get:
Using this knowledge, we can simplify the term as follows:
(3)
The substitution method
This method can be summarized in one sentence: Guess an asymptotic upper bound on f and (try to) prove it by induction.
As an example, we will prove that if f satisfies the equation (1) then f (N) = O(N log N).
From (1) we know that
for some c. Now we will prove that if we take a large enough (but constant) d then for almost all N we have f (N) dN lg N. We will
start by proving the induction step.
Assume that f (N/2) d (N/2)lg(N/2). Then
In other words, the induction step will hold as long as d > c. We are always able to choose such d.
We are only left with proving the inequality for some initial value N. This gets quite ugly when done formally. The general idea is that if
the d we found so far is not large enough, we can always increase it to cover the initial cases.
Note that for our example equation we won’t be able to prove it for N = 1, because lg 1 = 0. However, by taking f (N) dN lg N, where
d is some fixed constant. Conclusion: from (1) it follows that f (N) = O(N lg N).
The recursion tree
To a beginner, the previous method won’t be very useful. To use it successfully we need to make a good guess – and to make a good
guess we need some insight. The question is, how to gain this insight? Let’s take a closer look at what’s happening, when we try to
evaluate the recurrence (or equivalently, when we run the corresponding recursive program).
We may describe the execution of a recursive program on a given input by a rooted tree. Each node will correspond to some instance of
the problem the program solves. Consider an arbitrary vertex in our tree. If solving its instance requires recursive calls, this vertex will
have children corresponding to the smaller subproblems we solve recursively. The root node of the tree is the input of the program,
leaves represent small problems that are solved by brute force.
Now suppose we label each vertex by the amount of work spent solving the corresponding problem (excluding the recursive calls).
Clearly the runtime is exactly the sum of all labels.
As always, we only want an asymptotic bound. To achieve this, we may "round" the labels to make the summation easier. Again, we will
demonstrate this method on examples.
Example 4. The recursion tree for MergeSort on 5 elements.
The recursion tree for the corresponding recurrence equation. This time, the number inside each vertex represents the number of steps
the algorithm makes there.
Note that in a similar way we may sketch the general form of the recursion tree for any recurrence. Consider our old friend, the
equation (1). Here we know that there is a number c such that the number of operations in each node can be bound by (c times the
current value of N). Thus the tree in the example below is indeed the worst possible case.
Example 5. A worstcase tree for the general case of the recurrence equation (1).
Now, the classical trick from combinatorics is to sum the elements in an order different from the order in which they were created. In
this case, consider an arbitrary level of the tree (i.e. a set of vertices with the same depth). It is not hard to see that the total work on
each of the levels is cN.
Now comes the second question: What is the number of levels? Clearly, the leaves correspond to the trivial cases of the algorithm. Note
that the size of the problem is halved in each step. Clearly after lg N steps we are left with a trivial problem of size 1, thus the number
of levels is (log N).
A side note. If the reader doesn’t trust the simplifications we made when using this method, he is invited to treat this method as a "way
of making a good guess" and then to prove the result using the substitution method. However, with a little effort the application of this
method could also be upgraded to a full formal proof.
More recursion trees
By now you should be asking: Was it really only a coincidence that the total amount of work on each of the levels in Example 5 was the
same?
The answer: No and yes. No, there’s a simple reason why this happened, we’ll discover it later. Yes, because this is not always the case
– as we’ll see in the following two examples.
Example 6. Let’s try to apply our new "recursion tree" method to solve the following recurrence equation:
The recursion tree will look as follows:
Let’s try computing the total work for each of the first few levels. Our results:
level 1 2 3 …
Clearly as we go deeper in the tree, the total amount of work on the current level decreases. The question is, how fast does it
decrease? As we move one level lower, there will be three times that many subproblems. However, their size gets divided by 2, and
thus the time to process each of them decreases to one eighth of the original time. Thus the amount of work is decreased by the factor
3/8.
But this means that the entries in the table above form a geometric progression. For a while assume that this progression is infinite.
Then its sum would be
Thus the total amount of work in our tree is (N3) (summing the infinite sequence gives us an upper bound). But already the first
element of our progression is (N3). It follows that the total amount of work in our tree is (N3) and we are done.
The important generalization of this example: If the amounts of work at subsequent levels of the recursion tree form a decreasing
geometric progression, the total amount of work is asymptotically the same as the amount of work done in the root node.
From this result we can deduce an interesting fact about the (hypothetical) algorithm behind this recurrence equation: The recursive
calls didn’t take much time in this case, the most time consuming part was preparing the recursive calls and/or processing the results.
(I.e. this is the part that should be improved if we need a faster algorithm.)
Example 7. Now let’s try to apply our new "recursion tree" method to solve the following recurrence equation:
The recursion tree will look as follows:
Again, let’s try computing the total work for each of the first few levels. We get:
level 1 2 3 …
work cN cN cN …
This time we have the opposite situation: As we go deeper in the tree, the total amount of work on the current level increases. As we
move one level lower, there will be five times that many subproblems, each of them one third of the previous size, the processing time
is linear in problem size. Thus the amount of work increased by the factor 5/3.
Again, we want to compute the total amount of work. This time it won’t be that easy, because the most work is done on the lowest level
of the tree. We need to know its depth.
The lowest level corresponds to problems of size 1. The size of a problem on level k is N/3k. Solving the equation 1 = N/3k we get k =
log3N. Note that this time we explicitly state the base of the logarithm, as this time it will be important.
Our recursion tree has log3N levels. Each of the levels has five times more vertices than the previous one, thus the last level has
levels. The total work done on this level is then .
Note that using the trick (3) we may rewrite this as .
Now we want to sum the work done on all levels of the tree. Again, this is a geometric progression. But instead of explicitly computing
the sum, we now reverse it. Now we have a decreasing geometric progression…and we are already in the same situation as in the
previous example. Using the same reasoning we can show that the sum is asymptotically equal to the largest element.
It follows that the total amount of work in our tree is and we are done.
Note that the base3 logarithm ends in the exponent, that’s why the base is important. If the base was different, also the result would
be asymptotically different.
The Master Theorem
We already started to see a pattern here. Given a recurrence equation, take the corresponding recurrence tree and compute the
amounts of work done on each level of the tree. You will get a geometric sequence. If it decreases, the total work is proportional to
work done in the root node. If it increases, the total work is proportional to the number of leaves. If it remains the same, the total work
is (the work done on one level) times (the number of levels).
Actually, there are a few ugly cases, but almost often one of these three cases occurs. Moreover, it is possible to prove the statements
from the previous paragraph formally. The formal version of this theorem is known under the name Master Theorem.
For reference, we give the full formal statement of this theorem. (Note that knowing the formal proof is not necessary to apply this
theorem on a given recurrence equation.)
Let a 1 and b > 1 be integer constants. Let p be a nonnegative nondecreasing function. Let f be any solution of the recurrence
equation
Then:
1. If for some > 0 then
2. If , then f (N) = (p(N)log N).
3. If for some > 0, and if ap(N/b) cp(N) for some c < 1 and for almost all N, then f (N) = (p(N)).
Case 1 corresponds to our Example 7. Most of the time is spent making the recursive calls and it’s the number of these calls that
counts.
Case 2 corresponds to our Example 5. The time spent making the calls is roughly equal to the time to prepare the calls and process the
results. On all levels of the recursion tree we do roughly the same amount of work, the depth of the tree is always logarithmic.
Case 3 corresponds to our Example 6. Most of the time is spent on preparing the recursive calls and processing the results. Usually the
result will be asymptotically equal to the time spent in the root node.
Note the word "usually" and the extra condition in Case 3. For this result to hold we need p to be somehow "regular" – in the sense that
for each node in the recursion tree the time spent in the node must be greater than the time spent in its chidren (excluding further
recursive calls). This is nothing to worry about too much, most probably all functions p you will encounter in practice will satisfy this
condition (if they satisfy the first condition of Case 3).
Example 8. Let f (N) be the time Strassen’s fast matrix multiplication algorithm needs to multiply two N x N square matrices. This is a
recursive algorithm, that makes 7 recursive calls, each time multiplying two (N/2) x (N/2) square matrices, and then computes the
answer in (N2) time.
This leads us to the following recurrence equation:
Using the Master Theorem, we see that Case 1 applies. Thus the time complexity of Strassen’s algorithm is . Note
that by implementing the definition of matrix multiplication we get only a (N3) algorithm.
Example 9. Occasionally we may encounter the situation when the problems in the recursive calls are not of the same size. An
example may be the "medianoffive" algorithm to find the kth element of an array. It can be shown that its time complexity satisfies
the recurrence equation
How to solve it? Can the recursion tree be applied also in such asymmetric cases? Is there a more general version of Master Theorem
that handles also these cases? And what should I do with the recurrence f (N) = 4f (N/4) + (N log N), where the Master Theorem
doesn’t apply?
We won’t answer these questions here. This article doesn’t claim to be the one and only reference to computational complexity. If you
are already asking these questions, you understand the basics you need for programming challenges – and if you are interested in
knowing more, there are good books around that can help you.
Thanks for reading this far. If you have any questions, comments, bug reports or any other feedback, please use the Round tables. I’ll
do my best to answer.
Using Regular Expressions
By Dan[Popovici] & mariusmuja — topcoder members
Introduction
A regular expression is a special string that describes a search pattern. Many of you have surely seen and used them already when
typing expressions like ls(or dir) *.txt , to get a list of all the files with the extension txt. Regular expressions are very useful not only
for pattern matching, but also for manipulating text. In SRMs regular expressions can be extremely handy. Many problems that require
some coding can be written using regular expressions on a few lines, making your life much easier.
(Not so) Formal Description of Regular Expressions
A regular expression(regex) is one or more nonempty branches, separated by ‘|’. It matches anything that matches one of the
branches. The following regular expression will match any of the three words “the”,”top”,”coder”(quotes for clarity).
REGEX is : the|top|coder
INPUT is : Marius is one of the topcoders.
Found the text "the" starting at index 17 and ending at index 20.
Found the text "top" starting at index 21 and ending at index 24.
Found the text "coder" starting at index 24 and ending at index 29.
A branch is one or more pieces, concatenated. It matches a match for the first, followed by a match for the second, etc.
A piece is an atom possibly followed by a ‘*’, ‘+’, ‘?’, or bound. An atom followed by ‘*’ matches a sequence of 0 or more matches of the
atom. An atom followed by `+’ matches a sequence of 1 or more matches of the atom. An atom followed by `?’ matches a sequence of
0 or 1 matches of the atom.
The following regular expression matches any successive occurrence of the words ‘top’ and ‘coder’.
REGEX is: (top|coder)+
INPUT is: This regex matches topcoder and also codertop.
Found "topcoder" starting at index 19 and ending at index 27.
Found "codertop" starting at index 37 and ending at index 45.
A bound is ‘{‘ followed by an unsigned decimal integer, possibly followed by ‘,’ possibly followed by another unsigned decimal integer,
always followed by ‘}’.If there are two integers, the first may not exceed the second. An atom followed by a bound containing one
integer i and no comma matches a sequence of exactly i matches of the atom. An atom followed by a bound containing one integer i
and a comma matches a sequence of i or more matches of the atom. An atom followed by a bound containing two integers i and j
matches a sequence of i through j (inclusive) matches of the atom.
The following regular expression matches any sequence made of ’1′s having length 2,3 or 4 .
REGEX is: 1{2,4}
INPUT is: 101 + 10 = 111 , 11111 = 10000 + 1111
Found the text "111" starting at index 11 and ending at index 14.
Found the text "1111" starting at index 17 and ending at index 21.
Found the text "1111" starting at index 33 and ending at index 37.
One should observe that, greedily, the longest possible sequence is being matched and that different matches do not overlap. An atom
is a regular expression enclosed in ‘()’ (matching a match for the regular expression), a bracket expression (see below), ‘.’ (matching
any single character), ‘^’ (matching the null string at the beginning of a line), ‘$’ (matching the null string at the end of a line), a `’
followed by one of the characters `^.[$()|*+?{' (matching that character taken as an ordinary character) or a single character with no
other significance (matching that character). There is one more type of atom, the back reference: `' followed by a nonzero decimal
digit d matches the same sequence of characters matched by the dth parenthesized subexpression (numbering subexpressions by the
positions of their opening parentheses, left to right), so that (e.g.) `([bc])1′ matches `bb’ or `cc’ but not `bc’.
The following regular expression matches a string composed of two lowercase words separated by any character.
Current REGEX is: ([a‐z]+).1
Current INPUT is: top‐topcoder|coder
I found the text "top‐top" starting at index 0 and ending at index 7.
I found the text "coder|coder" starting at index 7 and ending at index 18.
A bracket expression is a list of characters enclosed in ‘[]‘. It normally matches any single character from the list. If the list begins with
‘^’, it matches any single character not from the rest of the list. If two characters in the list are separated by `’, this is shorthand for
the full range of characters between those two inclusive (e.g. ‘[09]‘ matches any decimal digit). With the exception of ‘]’,’^’,’’ all other
special characters, including `’, lose their special significance within a bracket expression.
The following regular expression matches any 3 character words not starting with ‘b’,’c’,’d’ and ending in ‘at’.
Current REGEX is: [^b‐d]at
Current INPUT is: bat
No match found.
Current REGEX is: [^b‐d]at
Current INPUT is: hat
I found the text "hat" starting at index 0 and ending at index 3.
This example combines most concepts presented above. The regex matches a set of open/close pair of html tags.
REGEX is: <([a‐zA‐Z][a‐zA‐Z0‐9]*)(()| [^>]*)>(.*)</1>
INPUT is: <font size="2">Topcoder is the</font> <b>best</b>
Found "<font size="2">Topcoder is the</font>" starting at index 0 and ending at index 37.
Found "<b>best</b>" starting at index 38 and ending at index 49.
([azAZ][azAZ09]*) will match any word that starts with a letter and continues with an arbitrary number of letters or digits.
(()| [^>]*) will match either the empty string or any string which does not contain ‘>’ .
1 will be replaced using backreferencing with the word matched be ([azAZ][azAZ09]*)
The above description is a brief one covering the basics of regular expressions. A regex written following the above rules should work in
both Java(>=1.4) and C++(POSIX EXTENDED). For a more in depth view of the extensions provided by different languages you can see
the links given in the References section.
Using regular expressions
In java
In java(1.4 and above) there is a package “java.util.regex” which allows usage of regular expressions.
This package contains three classes : Pattern, Matcher and PatternSyntaxException.
A Pattern object is a compiled representation of a regular expression. The Pattern class provides no public constructors. To create a
pattern, you must call one of its public static compile methods, both of which will return a Pattern object.
A Matcher object is the engine that interprets the pattern and performs match operations against an input string. Like the Pattern
class, Matcher defines no public constructors. You obtain a Matcher object by calling the public matcher method on a Pattern
object.
A PatternSyntaxException object is an unchecked exception that indicates a syntax error in a regular expression pattern.
Example(adapted from[4]):
Pattern pattern;
Matcher matcher;
pattern = Pattern.compile(<REGEX>);
matcher = pattern.matcher(<INPUT>);
boolean found;
while(matcher.find()) {
System.out.println("Found the text "" + matcher.group() + "" starting at index " + matcher.start() +
" and ending at index " + matcher.end() + ".");
found = true;
}
if(!found){
System.out.println("No match found.");
}
Java also offers the following methods in the String class:
boolean matches(String regex) (returns if the current string matches the regular expression regex)
String replaceAll(String regex, String replacement) (Replaces each substring of this string that matches the given regular
expression with the given replacement.)
String replaceFirst(String regex, String replacement) (Replaces the first substring of this string that matches the given regular
expression with the given replacement.)
String[] split(String regex) (Splits this string around matches of the given regular expression)
In C++
Many Topcoders believe that regular expressions are one of Java’s main strengths over C++ in the arena. C++ programmers don’t
despair, regular expressions can be used in C++ too.
There are several regular expression parsing libraries available for C++, unfortunately they are not very compatible with each other.
Fortunately as a Topcoder in the arena one does not have to cope with all this variety of “not so compatible with one another” libraries.
If you plan to use regular expressions in the arena you have to choose between two flavors of regex APIs: POSIX_regex and
GNU_regex. To use these APIs the header file “regex.h” must be included. Both of these work in two steps – first there is a function call
that compiles the regular expression, and then there is a function call that uses that compiled regular expression to search or match a
string.
Here is a short description of both of these APIs, leaving it up to the coder to choose the one that he likes the most.
POSIX_regex
Includes support for two different regular expression syntaxes, basic and extended. Basic regular expressions are similar to those in ed,
while extended regular expressions are more like those in egrep, adding the ‘|’, ‘+’ and ‘?’ operators and not requiring backslashes on
parenthesized subexpressions or curlybracketed bounds. Basic is the default, but extended is prefered.
With POSIX, you can only search for a given regular expression; you can’t match it. To do this, you must first compile it in a pattern
buffer, using `regcomp’. Once you have compiled the regular expression into a pattern buffer you can search in a null terminated string
using ‘regexec’. If either of the ‘regcomp’ or ‘regexec’ function fail they return an error code. To get an error string corresponding to
these codes you must use ‘regerror’. To free the allocated fields of a pattern buffer you must use ‘regfree’.
For an in depth description of how to use these functions please consult [2] or [3] in the References section.
Example:
Here is a small piece of code showing how these functions can be used:
regex_t reg;
string pattern = "[^tpr]{2,}";
string str = "topcoder";
regmatch_t matches[1];
regcomp(®,pattern.c_str(),REG_EXTENDED|REG_ICASE);
if (regexec(®,str.c_str(),1,matches,0)==0) {
cout << "Match "
cout << str.substr(matches[0].rm_so,matches[0].rm_eo‐matches[0].rm_so)
cout << " found starting at: "
cout << matches[0].rm_so
cout << " and ending at "
cout << matches[0].rm_eo
cout << endl;
} else {
cout << "Match not found."
cout << endl;
}
regfree(®);
GNU_regex
The GNU_regex API has a richer set of functions. With GNU regex functions you can both match a string with a pattern and search a
pattern in a string. The usage of these functions is somehow similar with the usage of the POSIX functions: a pattern must first be
compiled with ‘re_compile_pattern’, and the pattern buffer obtained is used to search and match. The functions used for searching and
matching are ‘re_search’ and ‘re_match’. In case of searching a fastmap can be used in order to optimize search. Without a fastmap the
search algorithm tries to match the pattern at consecutive positions in the string. The fastmap tells the algorithm what the characters
are from which a match can start. The fastmap is constructed by calling the ‘re_compile_fastmap’. The GNU_regex also provides the
functions ‘re_search2′ and ‘re_match2′ for searching and matching with split data. To free the allocated fields of a pattern buffer you
must use ‘regfree’.
For an indepth description of how to use these functions please consult [3].
Example:
string pattern = "([a‐z]+).1";
string str = "top‐topcoder|coder";
re_pattern_buffer buffer;
char map[256];
buffer.translate = 0;
buffer.fastmap = map;
buffer.buffer = 0;
buffer.allocated = 0;
re_set_syntax(RE_SYNTAX_POSIX_EXTENDED);
const char* status = re_compile_pattern(pattern.c_str(),pattern.size(),&buffer);
if (status) {
cout << "Error: " << status << endl;
}
re_compile_fastmap(&buffer);
struct re_registers regs;
int ofs = 0;
if (re_search(&buffer,str.c_str(),str.size(),0,str.size(),®s)!=‐1) {
cout << "Match "
cout << str.substr(regs.start[0],regs.end[0]‐regs.start[0])
cout << " found starting at: "
cout << regs.start[0]
cout << " and ending at "
cout << regs.end[0]
cout << endl;
} else {
cout << "Match not found."
cout << endl;
}
regfree(&buffer);
Real SRMs Examples
The following examples are all written in Java for the sake of clarity. A C++ user can use the POSIX or the GNU regex APIs to construct
functions similar to those available in Java(replace_all, split, matches).
CyberLine (SRM 187 div 1, level 1)
import java.util.*;
public class Cyberline{
public String lastCyberword(String cyberline){
String[]w=cyberline.replaceAll("‐","")
.replaceAll("[^a‐zA‐Z0‐9]"," ")
.split(" ") ;
return w[w.length‐1];
}
}
UnLinker (SRM 203 div 2, level 3)
import java.util.*;
public class UnLinker{
public String clean(String text){
String []m = text.split("((http://)?www[.]|http://)[a‐zA‐Z0‐9.]+[.](com|org|edu|info|tv)",‐1);
String s = m[0] ;
for (int i = 1 ; i < m.length ; ++i)
s = s + "OMIT" + i + m[i] ;
return s ;
}
}
CheatCode (SRM 154 div 1, level 1)
import java.util.*;
public class CheatCode {
public int[] matches(String keyPresses, String[] codes) {
boolean []map = new boolean[codes.length] ;
int count = 0 ;
for (int i=0;i<codes.length; ++i){
String regex = ".*" ;
for (int j=0; j<codes[i].length(); ) {
int k = 1;
while ((j+k)<codes[i].length() && codes[i].charAt(j+k)==codes[i].charAt(j)) k++;
regex = regex + codes[i].charAt(j) + "{"+k+",}";
j+=k;
}
regex = regex + ".*" ;
if (keyPresses.matches(regex)){
map[i] = true ;
count++ ;
}
}
int []res = new int[count] ;
int j=0;
for (int i= 0 ; i < codes.length; ++i)
if(map[i] == true)
res[j++]=i ;
return res ;
}
}
References
1. The regex(7) linux manual page
2. The regex(3) linux manual page
3. http://docs.freebsd.org/info/regex/regex.info.Programming_with_Regex.html
4. http://www.regularexpressions.info/
5. http://java.sun.com/docs/books/tutorial/extra/regex/
Understanding Probabilities
By supernova — topcoder member
Discuss this article in the forums
It has been said that life is a school of probability. A major effect of probability theory on everyday life is in risk assessment. Let’s
suppose you have an exam and you are not so well prepared. There are 20 possible subjects, but you only had time to prepare for 15.
If two subjects are given, what chances do you have to be familiar with both? This is an example of a simple question inspired by the
world in which we live today. Life is a very complex chain of events and almost everything can be imagined in terms of probabilities.
Gambling has become part of our lives and it is an area in which probability theory is obviously involved. Although gambling had existed
since time immemorial, it was not until the seventeenth century that the mathematical foundations finally became established. It all
started with a simple question directed to Blaise Pascal by Chevalier de Mere, a nobleman that gambled frequently to increase his
wealth. The question was whether a double six could be obtained on twentyfour rolls of two dice.
As far as topcoder problems are concerned, they’re inspired by reality. You are presented with many situations, and you are explained
the rules of many games. While it’s easy to recognize a problem that deals with probability computations, the solution may not be
obvious at all. This is partly because probabilities are often overlooked for not being a common theme in programming challenges. But
it is not true and topcoder has plenty of them! Knowing how to approach such problems is a big advantage in topcoder competitions
and this article is to help you prepare for this topic.
Before applying the necessary algorithms to solve these problems, you first need some mathematical understanding. The next chapter
presents the basic principles of probability. If you already have some experience in this area, you might want to skip this part and go to
the following chapter: Step by Step Probability Computation. After that it follows a short discussion on Randomized Algorithms and in
the end there is a list with the available problems on topcoder. This last part is probably the most important. Practice is the key!
Basics
Working with probabilities is much like conducting an experiment. An outcome is the result of an experiment or other situation
involving uncertainty. The set of all possible outcomes of a probability experiment is called a sample space. Each possible result of
such a study is represented by one and only one point in the sample space, which is usually denoted by S. Let’s consider the following
experiments:
Rolling a die once
Sample space S = {1, 2, 3, 4, 5, 6}
Tossing two coins
Sample space S = {(Heads, Heads), (Heads, Tails), (Tails, Heads), (Tails, Tails)}
We define an event as any collection of outcomes of an experiment. Thus, an event is a subset of the sample space S. If we denote an
event by E, we could say that E?S. If an event consists of a single outcome in the sample space, it is called a simple event. Events
which consist of more than one outcome are called compound events.
What we are actually interested in is the probability of a certain event to occur, or P(E). By definition, P(E) is a real number between 0
and 1, where 0 denotes the impossible event and 1 denotes the certain event (or the whole sample space).
As stated earlier, each possible outcome is represented by exactly one point in the sample space. This leads us to the following formula:
That is, the probability of an event to occur is calculated by dividing the number of favorable outcomes (according to the event E) by
the total number of outcomes (according to the sample space S). In order to represent the relationships among events, you can
apply the known principles of set theory. Consider the experiment of rolling a die once. As we have seen previously, the sample space is
S = {1, 2, 3, 4, 5, 6}. Let’s now consider the following events:
Event A = ‘score > 3′ = {4, 5, 6}
Event B = ‘score is odd’ = {1, 3, 5}
Event C = ‘score is 7′ = ?
A?B =’the score is > 3 or odd or both’ = {1, 3, 4, 5, 6}
AnB =’the score is > 3 and odd’ = {5}
A’ = ‘event A does not occur’ = {1, 2, 3}
We have:
P(A?B) = 5/6
P(AnB) = 1/6
P(A’) = 1 – P(A) = 1 – 1/2 = 1/2
P(C) = 0
The first step when trying to solve a probability problem is to be able to recognize the sample space. After that, you basically have to
determine the number of favorable outcomes. This is the classical approach, but the way we implement it may vary from problem to
problem. Let’s take a look at QuizShow (SRM 223, Div 1 – Easy). The key to solving this problem is to take into account all the
possibilities, which are not too many. After a short analysis, we determine the sample space to be the following:
S = { (wager 1 is wrong, wager 2 is wrong, you are wrong),
(wager 1 is wrong, wager 2 is wrong, you are right),
(wager 1 is wrong, wager 2 is right, you are wrong),
(wager 1 is wrong, wager 2 is right, you are right),
(wager 1 is right, wager 2 is wrong, you are wrong),
(wager 1 is right, wager 2 is wrong, you are right),
(wager 1 is right, wager 2 is right, you are wrong),
(wager 1 is right, wager 2 is right, you are right) }
The problem asks you to find a wager that maximizes the number of favorable outcomes. In order to compute the number of favorable
outcomes for a certain wager, we need to determine how many points the three players end with for each of the 8 possible outcomes.
The idea is illustrated in the following program:
int wager (vector scores, int wager1, int wager2){
int best, bet, odds, wage, I, J, K;
best = 0; bet = 0;
for (wage = 0; wage = scores[0]; wage++){
odds = 0;
// in 'odds' we keep the number of favorable outcomes
for (I = ‐1; I = 1; I = I + 2)
for (J = ‐1; J = 1; J = J + 2)
for (K = ‐1; K = 1; K = K + 2)
if (scores[0] + I * wage > scores[1] + J * wager1 &&
scores[0] + I * wage > scores[2] + K * wager2) { odds++; }
if (odds > best) { bet = wage ; best = odds; }
// a better wager has been found
}
return bet;
}
Another good problem to start with is PipeCuts (SRM 233, Div 1 – Easy). This can be solved in a similar manner. There is a finite
number of outcomes and all you need to do is to consider them one by one.
Let’s now consider a series of n independent events: E1, E2, … , En. Two surprisingly common questions that may appear (and many of
you have already encountered) are the following:
1. What is the probability that all events will occur?
2. What is the probability that at least one event will occur?
To answer the first question, we relate to the occurrence of the first event (call it E1). If E1 does not occur, the hypothesis can no longer
be fulfilled. Thus, it must be inferred that E1 occurs with a probability of P(E1). This means there is a P(E1) chance we need to check
for the occurrence of the next event (call it E2). The event E2 occurs with a probability of P(E2) and we can continue this process in the
same manner. Because probability is by definition a real number between 0 and 1, we can synthesize the probability that all events will
occur in the following formula:
The best way to answer the second question is to first determine the probability that no event will occur and then, take the
complement. We have:
These formulae are very useful and you should try to understand them well before you move.
BirthdayOdds
A good example to illustrate the probability concepts discussed earlier is the classical "Birthday Paradox". It has been shown that if
there are at least 23 people in a room, there is a more than 50% chance that at least two of them will share the same birthday. While
this is not a paradox in the real sense of the word, it is a mathematical truth that contradicts common intuition. The topcoder problem
asks you to find the minimum number of people in order to be more than minOdds% sure that at least two of them have the same
birthday. One of the first things to notice about this problem is that it is much easier to solve the complementary problem: "What is the
probability that N randomly selected people have all different birthdays?". The strategy is to start with an empty room and put people in
the room one by one, comparing their birthdays with those of them already in the room:
int minPeople (int minOdds, int days){
int nr;
double target, p;
target = 1 ‐ (double) minOdds / 100;
nr = 1;
p = 1;
while (p > target){
p = p * ( (double) 1 ‐ (double) nr / days);
nr ++;
}
return nr;
}
This so called "Birthday Paradox” has many real world applications and one of them is described in the topcoder problem called Collision
(SRM 153, Div 1 – Medium). The algorithm is practically the same, but one has to be careful about the events that may alter the
sample space.
Sometimes a probability problem can be quite tricky. As we have seen before, the ‘Birthday Paradox’ tends to contradict our common
sense. But the formulas prove to us that the answer is indeed correct. Formulas can help, but to become a master of probabilities you
need one more ingredient: "number sense" . This is partly innate ability and partly learned ability acquired through practice. Take this
quiz to assess your number sense and to also become familiar with some of the common probability misconceptions.
Step by Step Probability Computation
In this chapter we will discuss some real topcoder problems in which the occurrence of an event is influenced by occurrences of
previous events. We can think of it as a graph in which the nodes are events and the edges are dependencies between them. This is a
somewhat forced analogy, but the way we compute the probabilities for different events is similar to the way we traverse the nodes of a
graph. We start from the root, which is the initial state and has a probability of 1. Then, as we consider different scenarios, the
probability is distributed accordingly.
NestedRandomness
This problem looked daunting to some people, but for those who figured it out, it was just a matter of a few lines. For the first step, it is
clear what do we have to do: the function random(N) is called and it returns a random integer uniformly distributed in the range 0 to N
1. Thus, every integer in this interval has a probability of 1/N to occur. If we consider all these outcomes as input for the next step, we
can determine all the outcomes of the random(random(N)) call. To understand this better, let’s work out the case when N = 4.
After the first nesting all
integers have the same
probability to occur, which is 1 /
4.
For the second nesting there is
a 1/4 chance for each of the
following functions to be called:
random(0), random(1),
random(2) and random(3).
Random(0) produces an error,
random(1) returns 0, random (2)
returns 0 or 1 (each with a
probability of 1/2) and
random(3) returns 0, 1 or 2.
As a result, for the third NestedRandomness
nesting, random(0) has a for N = 4
probability of 1/4 + 1/8 + 1/12
of being called, random(1) has a
probability of 1/8 + 1/12 of
being called and random(2) has
a probability of 1/12 of being
called.
Analogously, for the fourth
nesting, the function random(0)
has a probability of 1/4 of being
called, while random(1) has a
probability of 1/24.
As for the fifth nesting, we can
only call random(0), which
produces an error. The whole
process is described in the
picture to the right.
The source code for this problem is given below:
double probability (int N, int nestings, int target){
int I, J, K;
double A[1001], B[2001];
// A[I] represents the probability of number I to appear
for (I = 0; I < N ; I++) A[I] = (double) 1 / N;
for (K = 2; K = nestings; K++){
for (I = 0; I < N; I++) B[I] = 0;
// for each I between 0 and N‐1 we call the function "random(I)"
// as described in the problem statement
for (I = 0; I < N; I++)
for (J = 0; J < I; J++)
B[J] += (double) A[I] / I;
for (I = 0; I < N; I++) A[I] = B[I];
}
return A[target];
}
If you got the taste for this problem, here are another five you may want to try:
ChessKnight – assign each square a probability and for every move check the squares one by one to compute the probabilities for the
next move.
DiceThrows – determine the probability of each possible outcome for both players and then compare the results.
RockSkipping – the same approach, just make sure you got the lake pattern correctly.
PointSystem – represent the event space as a matrix of possible scores (x, y).
VolleyBall – similar to PointSystem, but the scores may go up pretty high.
Let’s now take a look at another topcoder problem, GeneticCrossover, which deals with conditional probability. Here, you are asked
to predict the quality of an animal, based on the genes it inherits from its parents. Considering the problem description, there are two
situations that may occur: a gene does not depend on another gene, or a gene is dependent.
For the first case, consider p the probability that the gene is to be expressed dominantly. There are only 4 cases to consider:
at least one parent has two dominant genes. (p = 1)
each parent has exactly one dominant gene. (p = 0.5)
one parent has one dominant gene and the other has only recessive genes (p = 0.25)
both parents have two recessive genes (p = 0)
Now let’s take the case when a gene is dependent on another. This make things a bit trickier as the "parent" gene may also depend on
another and so on … To determine the probability that a dependent gene is dominant, we take the events that each gene in the chain
(starting with the current gene) is dominant. In order for the current gene to be expressed dominantly, we need all these events to
occur. To do this, we take the product of probabilities for each individual event in the chain. The algorithm works recursively. Here is the
complete source code for this problem:
int n, d[200];
double power[200];
// here we determine the characteristic for each gene (in power[I]
// we keep the probability of gene I to be expressed dominantly)
double detchr (string p1a, string p1b, string p2a, string p2b, int nr){
double p, p1, p2;
p = p1 = p2 = 1.0;
if (p1a[nr] = 'Z') p1 = p1 ‐ 0.5;
// is a dominant gene
if (p1b[nr] = 'Z') p1 = p1 ‐ 0.5;
if (p2a[nr] = 'Z') p2 = p2 ‐ 0.5;
if (p2b[nr] = 'Z') p2 = p2 ‐ 0.5;
p = 1 ‐ p1 * p2;
if (d[nr] != 1) power[nr] = p * detchr (p1a, p1b, p2a, p2b, d[nr]);
// gene 'nr' is dependent on gene d[nr]
else power[nr] = p;
return power[nr];
}
double cross (string p1a, string p1b, string p2a, string p2b,
vector dom, vector rec, vector dependencies){
int I;
double fitness = 0.0;
n = rec.size();
for (I = 0; I < n; I++) d[I] = dependencies[I];
for (I = 0 ;I < n; I++) power[I] = ‐1.0;
for (I = 0; I < n; i++)
if (power[I] == ‐1.0) detchr (p1a, p1b, p2a, p2b, i);
// we check if the dominant character of gene I has
// not already been computed
for (I = 0; I = n; I++)
fitness=fitness+(double) power[I]*dom[I]‐(double) (1‐power[I])*rec[I];
// we compute the expected 'quality' of an animal based on the
// probabilities of each gene to be expressed dominantly
return fitness;
}
See also ProbabilityTree.
Randomized Algorithms
We call randomized algorithms those algorithms that use random numbers to make decisions during their execution. Unlike
deterministic algorithms that for a fixed input always give the same output and the same runningtime, a randomized algorithm
behaves differently from execution to execution. Basically, we distinguish two kind of randomized algorithms:
1. Monte Carlo algorithms: may sometimes produce an incorrect solution – we bound the probability of failure.
2. Las Vegas algorithms: always give the correct solution, the only variation is the running time – we study the distribution of the
running time.
Read these lecture notes from the College of Engineering at UIUC for an example of how these algorithms work.
The main goal of randomized algorithms is to build faster, and perhaps simpler solutions. Being able to tackle "harder" problems is also
a benefit of randomized algorithms. As a result, these algorithms have become a research topic of major interest and have already
been utilized to more easily solve many different problems.
An interesting question is whether such an algorithm may become useful in topcoder competitions. Some problems have many possible
solutions, where a number of which are also optimal. The classical approach is to check them one by one, in an established order. But it
cannot be guaranteed that the optima are uniformly distributed in the solution domain. Thus, a deterministic algorithm may not find
you an optimum quickly enough. The advantage of a randomized algorithm is that there are actually no rules to set about the order in
which the solutions are checked and for the cases when the optima are clustered together, it usually performs much better. See
QueenInterference for a topcoder example.
Randomized algorithms are particularly useful when faced with malicious attackers who deliberately try to feed a bad input to the
algorithm. Such algorithms are widely used in cryptography, but it sometimes makes sense to also use them in topcoder competitions.
It may happen that you have an efficient algorithm, but there are a few degenerate cases for which its running time is significantly
slower. Assuming the algorithm is correct, it has to run fast enough for all inputs. Otherwise, all the points you earned for submitting
that particular problem are lost. This is why here, on topcoder, we are interested in worst case execution time.
To challenge or not to challenge?
Another fierce coding challenge is now over and you have 15 minutes to look for other coders’ bugs. The random call in a competitor’s
submission is likely to draw your attention. This will most likely fall into one of two scenarios:
1. the submission was just a desperate attempt and will most likely fail on many inputs.
2. the algorithm was tested rather thoroughly and the probability to fail (or time out) is virtually null.
The first thing you have to do is to ensure it was not already unsuccessfully challenged (check the coder’s history). If it wasn’t, it may
deserve a closer look. Otherwise, you should ensure that you understand what’s going on before even considering a challenge. Also
take into account other factors such as coder rating, coder submission accuracy, submission time, number of resubmissions or impact
on your ranking.
Will "random" really work?
In most optimizing problems, the ratio between the number of optimal solutions and the total number of solutions is not so obvious. An
easy, but not so clever solution, is to simply try generating different samples and see how the algorithm behaves. Running such a
simulation is usually pretty quick and may also give you some extra clues in how to actually solve the problem.
Max = 1000000; attempt = 0;
while (attempt < Max){
answer = solve_random (...);
if (better (answer, optimum)){
// we found a better solution
optimum = answer;
cout << "Solution " << answer << " found on step " << attempt << endl;
}
attempt ++;
}
Practice Problems
Level 1
PipeCuts – SRM 233
BirthdayOdds – SRM 174
BenfordsLaw – SRM 155
QuizShow – SRM 223
Level 2
Collision – SRM 153
ChessKnight – TCCC05 Round 1
ChipRace – SRM 199
DiceThrows – SRM 242
TopFive – SRM 243
ProbabilityTree – SRM 174
OneArmedBandit – SRM 226
RangeGame – SRM 174
YahtzeeRoll – SRM 222
BagOfDevouring – SRM 184
VolleyBall – TCO04 Round 3
RandomFA – SRM 178
PackageShipping – TCCC05 Round 3
QueenInterference – SRM 208
BaseballLineup – TCO ’03 Finals
Level 3
GeneticCrossover – TCO04 Qual 3
NestedRandomness – TCCC05 Qual 5
RockSkipping – TCCC ’04 Round 1
PointSystem – SRM 174
AntiMatter – SRM 179
TestScores – SRM 226
Hangman42 – SRM 229
KingOfTheCourt – SRM 222
WinningProbability – SRM 218
Disaster – TCCC05 Semi 1
Data Structures
By timmac — topcoder member
Discuss this article in the forums
Even though computers can perform literally millions of mathematical computations per second, when a problem gets large and
complicated, performance can nonetheless be an important consideration. One of the most crucial aspects to how quickly a problem can
be solved is how the data is stored in memory.
To illustrate this point, consider going to the local library to find a book about a specific subject matter. Most likely, you will be able to
use some kind of electronic reference or, in the worst case, a card catalog, to determine the title and author of the book you want.
Since the books are typically shelved by category, and within each category sorted by author’s name, it is a fairly straightforward and
painless process to then physically select your book from the shelves.
Now, suppose instead you came to the library in search of a particular book, but instead of organized shelves, were greeted with large
garbage bags lining both sides of the room, each arbitrarily filled with books that may or may not have anything to do with one another.
It would take hours, or even days, to find the book you needed, a comparative eternity. This is how software runs when data is not
stored in an efficient format appropriate to the application.
Simple Data Structures
The simplest data structures are primitive variables. They hold a single value, and beyond that, are of limited use. When many related
values need to be stored, an array is used. It is assumed that the reader of this article has a solid understanding of variables and
arrays.
A somewhat more difficult concept, though equally primitive, are pointers. Pointers, instead of holding an actual value, simply hold a
memory address that, in theory, contains some useful piece of data. Most seasoned C++ coders have a solid understanding of how to
use pointers, and many of the caveats, while fledgling programmers may find themselves a bit spoiled by more modern "managed"
languages which, for better or worse, handle pointers implicitly. Either way, it should suffice to know that pointers "point" somewhere in
memory, and do not actually store data themselves.
A less abstract way to think about pointers is in how the human mind remembers (or cannot remember) certain things. Many times, a
good engineer may not necessarily know a particular formula/constant/equation, but when asked, they could tell you exactly which
reference to check.
Arrays
Arrays are a very simple data structure, and may be thought of as a list of a fixed length. Arrays are nice because of their simplicity,
and are well suited for situations where the number of data items is known (or can be programmatically determined). Suppose you
need a piece of code to calculate the average of several numbers. An array is a perfect data structure to hold the individual values,
since they have no specific order, and the required computations do not require any special handling other than to iterate through all of
the values. The other big strength of arrays is that they can be accessed randomly, by index. For instance, if you have an array
containing a list of names of students seated in a classroom, where each seat is numbered 1 through n, then studentName[i] is a trivial
way to read or store the name of the student in seat i.
An array might also be thought of as a prebound pad of paper. It has a fixed number of pages, each page holds information, and is in a
predefined location that never changes.
Linked Lists
A linked list is a data structure that can hold an arbitrary number of data items, and can easily change size to add or remove items. A
linked list, at its simplest, is a pointer to a data node. Each data node is then composed of data (possibly a record with several data
values), and a pointer to the next node. At the end of the list, the pointer is set to null.
By nature of its design, a linked list is great for storing data when the number of items is either unknown, or subject to change.
However, it provides no way to access an arbitrary item from the list, short of starting at the beginning and traversing through every
node until you reach the one you want. The same is true if you want to insert a new node at a specific location. It is not difficult to see
the problem of inefficiency.
A typical linked list implementation would have code that defines a node, and looks something like this:
class ListNode {
String data;
ListNode nextNode;
}
ListNode firstNode;
You could then write a method to add new nodes by inserting them at the beginning of the list:
ListNode newNode = new ListNode();
NewNode.nextNode = firstNode;
firstNode = newNode;
Iterating through all of the items in the list is a simple task:
ListNode curNode = firstNode;
while (curNode != null) {
ProcessData(curNode);
curNode = curNode.nextNode;
}
A related data structure, the doubly linked list, helps this problem somewhat. The difference from a typical linked list is that the root
data structure stores a pointer to both the first and last nodes. Each individual node then has a link to both the previous and next node
in the list. This creates a more flexible structure that allows travel in both directions. Even still, however, this is rather limited.
Queues
A queue is a data structure that is best described as "first in, first out". A real world example of a queue is people waiting in line at the
bank. As each person enters the bank, he or she is "enqueued" at the back of the line. When a teller becomes available, they are
"dequeued" at the front of the line.
Perhaps the most common use of a queue within a topcoder problem is to implement a Breadth First Search (BFS). BFS means to first
explore all states that can be reached in one step, then all states that can be reached in two steps, etc. A queue assists in implementing
this solution because it stores a list of all state spaces that have been visited.
A common type of problem might be the shortest path through a maze. Starting with the point of origin, determine all possible
locations that can be reached in a single step, and add them to the queue. Then, dequeue a position, and find all locations that can be
reached in one more step, and enqueue those new positions. Continue this process until either a path is found, or the queue is empty
(in which case there is no path). Whenever a "shortest path" or "least number of moves" is requested, there is a good chance that a
BFS, using a queue, will lead to a successful solution.
Most standard libraries, such the Java API, and the .NET framework, provide a Queue class that provides these two basic interfaces for
adding and removing items from a queue.
BFS type problems appear frequently on challenges; on some problems, successful identification of BFS is simple and immediately,
other times it is not so obvious.
A queue implementation may be as simple as an array, and a pointer to the current position within the array. For instance, if you know
that you are trying to get from point A to point B on a 50×50 grid, and have determined that the direction you are facing (or any other
details) are not relevant, then you know that there are no more than 2,500 "states" to visit. Thus, your queue is programmed like so:
class StateNode {
int xPos;
int yPos;
int moveCount;
}
class MyQueue {
StateNode[] queueData = new StateNode[2500];
int queueFront = 0;
int queueBack = 0;
void Enqueue(StateNode node) {
queueData[queueBack] = node;
queueBack++;
}
StateNode Dequeue() {
StateNode returnValue = null;
if (queueBack > queueFront) {
returnValue = queueData[queueFront];
QueueFront++;
}
return returnValue;
}
boolean isNotEmpty() {
return (queueBack > queueFront);
}
}
Then, the main code of your solution looks something like this. (Note that if our queue runs out of possible states, and we still haven’t
reached our destination, then it must be impossible to get there, hence we return the typical "1" value.)
MyQueue queue = new MyQueue();
queue.Enqueue(initialState);
while (queue.isNotEmpty()) {
StateNode curState = queue.Dequeue();
if (curState == destState)
return curState.moveCount;
for (int dir = 0; dir < 3; dir++) {
if (CanMove(curState, dir))
queue.Enqueue(MoveState(curState, dir));
}
}
Stacks
Stacks are, in a sense, the opposite of queues, in that they are described as "last in, first out". The classic example is the pile of plates
at the local buffet. The workers can continue to add clean plates to the stack indefinitely, but every time, a visitor will remove from the
stack the top plate, which is the last one that was added.
While it may seem that stacks are rarely implemented explicitly, a solid understanding of how they work, and how they are used
implicitly, is worthwhile education. Those who have been programming for a while are intimately familiar with the way the stack is used
every time a subroutine is called from within a program. Any parameters, and usually any local variables, are allocated out of space on
the stack. Then, after the subroutine has finished, the local variables are removed, and the return address is "popped" from the stack,
so that program execution can continue where it left off before calling the subroutine.
An understanding of what this implies becomes more important as functions call other functions, which in turn call other functions. Each
function call increases the "nesting level" (the depth of function calls, if you will) of the execution, and uses increasingly more space on
the stack. Of paramount importance is the case of a recursive function. When a recursive function continually calls itself, stack space is
quickly used as the depth of recursion increases. Nearly every seasoned programmer has made the mistake of writing a recursive
function that never properly returns, and calls itself until the system throws up an "out of stack space" type of error.
Nevertheless, all of this talk about the depth of recursion is important, because stacks, even when not used explicitly, are at the heart
of a depth first search. A depth first search is typical when traversing through a tree, for instance looking for a particular node in an
XML document. The stack is responsible for maintaining, in a sense, a trail of what path was taken to get to the current node, so that
the program can "backtrack" (e.g. return from a recursive function call without having found the desired node) and proceed to the next
adjacent node.
Soma (SRM 198) is an excellent example of a problem solved with this type of approach.
Trees
Trees are a data structure consisting of one or more data nodes. The first node is called the "root", and each node has zero or more
"child nodes". The maximum number of children of a single node, and the maximum depth of children are limited in some cases by the
exact type of data represented by the tree.
One of the most common examples of a tree is an XML document. The toplevel document element is the root node, and each tag
found within that is a child. Each of those tags may have children, and so on. At each node, the type of tag, and any attributes,
constitutes the data for that node. In such a tree, the hierarchy and order of the nodes is well defined, and an important part of the
data itself. Another good example of a tree is a written outline. The entire outline itself is a root node containing each of the toplevel
bullet points, each of which may contain one or more subbullets, and so on. The file storage system on most disks is also a tree
structure.
Corporate structures also lend themselves well to trees. In a classical management hierarchy, a President may have one or more vice
presidents, each of whom is in charge of several managers, each of whom presides over several employees.
PermissionTree (SRM 218) provides an unusual problem on a common file system.
bloggoDocStructure (SRM 214) is another good example of a problem using trees.
Binary Trees
A special type of tree is a binary tree. A binary tree also happens to be one of the most efficient ways to store and read a set of records
that can be indexed by a key value in some way. The idea behind a binary tree is that each node has, at most, two children.
In the most typical implementations, the key value of the left node is less than that of its parent, and the key value of the right node is
greater than that of its parent. Thus, the data stored in a binary tree is always indexed by a key value. When traversing a binary tree, it
is simple to determine which child node to traverse when looking for a given key value.
One might ask why a binary tree is preferable to an array of values that has been sorted. In either case, finding a given key value (by
traversing a binary tree, or by performing a binary search on a sorted array) carries a time complexity of O(log n). However, adding a
new item to a binary tree is an equally simple operation. In contrast, adding an arbitrary item to a sorted array requires some time
consuming reorganization of the existing data in order to maintain the desired ordering.
If you have ever used a field guide to attempt to identify a leaf that you find in the wild, then this is a good way to understand how
data is found in a binary tree. To use a field guide, you start at the beginning, and answer a series of questions like "is the leaf jagged,
or smooth?" that have only two possible answers. Based upon your answer, you are directed to another page, which asks another
question, and so on. After several questions have sufficiently narrowed down the details, you are presented with the name, and perhaps
some further information about your leaf. If one were the editor of such a field guide, newly cataloged species could be added to field
guide in much the same manner, by traversing through the questions, and finally at the end, inserting a new question that differentiates
the new leaf from any other similar leaves. In the case of a computer, the question asked at each node is simply "are you less than or
greater than X?"
Priority Queues
In a typical breadth first search (BFS) algorithm, a simple queue works great for keeping track of what states have been visited. Since
each new state is one more operational step than the current state, adding new locations to the end of the queue is sufficient to insure
that the quickest path is found first. However, the assumption here is that each operation from one state to the next is a single step.
Let us consider another example where you are driving a car, and wish to get to your destination as quickly as possible. A typical
problem statement might say that you can move one block up/down/left/right in one minute. In such a case, a simple queuebased BFS
works perfectly, and is guaranteed to provide a correct result.
But what happens if we say that the car can move forward one block in two minute, but requires three minutes to make a turn and then
move one block (in a direction different from how the car was originally facing)? Depending on what type of move operation we
attempt, a new state is not simply one "step" from the current state, and the "in order" nature of a simple queue is lost.
This is where priority queues come in. Simply put, a priority queue accepts states, and internally stores them in a method such that it
can quickly pull out the state that has the least cost. (Since, by the nature of a "shortest time/path" type of problem, we always want to
explore the states of least cost first.)
A real world example of a priority queue might be waiting to board an airplane. Individuals arriving at their gate earlier will tend to sit
closest to the door, so that they can get in line as soon as they are called. However, those individuals with a "gold card", or who travel
first class, will always be called first, regardless of when they actually arrived.
One very simple implementation of a priority queue is just an array that searches (one by one) for the lowest cost state contained
within, and appends new elements to the end. Such an implementation has a trivial timecomplexity for insertions, but is painfully slow
to pull objects out again.
A special type of binary tree called a heap is typically used for priority queues. In a heap, the root node is always less than (or greater
than, depending on how your value of "priority" is implemented) either of its children. Furthermore, this tree is a "complete tree" from
the left. A very simple definition of a complete tree is one where no branch is n + 1 levels deep until all other branches are n levels
deep. Furthermore, it is always the leftmost node(s) that are filled first.
To extract a value from a heap, the root node (with the lowest cost or highest priority) is pulled. The deepest, rightmost leaf then
becomes the new root node. If the new root node is larger than at at least one of its children, then the root is swapped with its smallest
child, in order to maintain the property that the root is always less than its children. This continues downward as far as necessary.
Adding a value to the heap is the reverse. The new value is added as the next leaf, and swapped upward as many times as necessary to
maintain the heap property.
A convenient property of trees that are complete from the left is that they can be stored very efficiently in a flat array. In general,
element 0 of the array is the root, and elements 2k + 1 and 2k + 2 are the children of element k. The effect here is that adding the
next leaf simply means appending to the array.
Hash Tables
Hash tables are a unique data structure, and are typically used to implement a "dictionary" interface, whereby a set of keys each has
an associated value. The key is used as an index to locate the associated values. This is not unlike a classical dictionary, where
someone can find a definition (value) of a given word (key).
Unfortunately, not every type of data is quite as easy to sort as a simple dictionary word, and this is where the "hash" comes into play.
Hashing is the process of generating a key value (in this case, typically a 32 or 64 bit integer) from a piece of data. This hash value
then becomes a basis for organizing and sorting the data. The hash value might be the first n bits of data, the last n bits of data, a
modulus of the value, or in some cases, a more complicated function. Using the hash value, different "hash buckets" can be set up to
store data. If the hash values are distributed evenly (which is the case for an ideal hash algorithm), then the buckets will tend to fill up
evenly, and in many cases, most buckets will have no more than one or only a few objects in them. This makes the search even faster.
A hash bucket containing more than one value is known as a "collision". The exact nature of collision handling is implementation
specific, and is crucial to the performance of the hash table. One of the simplest methods is to implement a structure like a linked list at
the hash bucket level, so that elements with the same hash value can be chained together at the proper location. Other, more
complicated schemes may involve utilizing adjacent, unused locations in the table, or rehashing the hash value to obtain a new value.
As always, there are good and bad performance considerations (regarding time, size, and complexity) with any approach.
Another good example of a hash table is the Dewey decimal system, used in many libraries. Every book is assigned a number, based
upon its subject matter? the 500′s are all science books, the 700′s are all the arts, etc. Much like a real hash table, the speed at which
a person could find a given book is based upon how well the hash buckets are evenly divided? It will take longer to find a book about
frogs in a library with many science materials than in a library consisting mostly of classical literature.
In applications development, hash tables are a convenient place to store reference data, like state abbreviations that link to full state
names. In problem solving, hash tables are useful for implementing a divideandconquer approach to knapsacktype problems. In
LongPipes, we are asked to find the minimum number of pipes needed to construct a single pipe of a given length, and we have up to
38 pieces of pipe. By dividing this into two sets of 19, and calculating all possible lengths from each set, we create hash tables linking
the length of the pipe to the fewest number of segments used. Then, for each constructed pipe in one set, we can easily look up,
whether or not we constructed a pipe of corresponding length in the other set, such that the two join to form a complete pipe of the
desired length.
Conclusion
The larger picture to be seen from all of this is that data structures are just another set of tools that should be in the kit of a seasoned
programmer. Comprehensive libraries and frameworks available with most languages nowadays preempt the need for a full
understanding of how to implement each of these tools. The result is that developers are able to quickly produce quality solutions that
take advantage of powerful ideas. The challenge lies in knowing which one to select.
Nonetheless, knowing a little about how these tools work should help to make the choices easier. And, when the need arises, perhaps
leave the programmer better equipped to think up a new solution to a new problem? if not while on the job doing work for a client, then
perhaps while contemplating the 1000 point problem 45 minutes into the coding phase of the next SRM.
Sorting
By timmac — TopCoder Member
Discuss this article in the forums
Introduction
Any number of practical applications in computing require things to be in order. Even before we start computing, the importance of
sorting is drilled into us. From group pictures that require the tallest people to stand in the back, to the highest grossing salesman
getting the largest Christmas bonus, the need to put things smallest to largest or first to last cannot be underestimated.
When we query a database, and append an ORDER BY clause, we are sorting. When we look for an entry in the phone book, we are
dealing with a list that has already been sorted. (And imagine if it weren’t!) If you need to search through an array efficiently using a
binary search, it is necessary to first sort the array. When a problem statement dictates that in the case of a tie we should return the
lexicographically first result, well? you get the idea.
General Considerations
Imagine taking a group of people, giving them each a deck of cards that has been shuffled, and requesting that they sort the cards in
ascending rank order. Some people might start making piles, others might spread the cards all over a table, and still others might
juggle the cards around in their hands. For some, the exercise might take a matter of seconds, for others several minutes or longer.
Some might end up with a deck of cards where spades always appear before hearts, in other cases it might be less organized.
Fundamentally, these are all the big bullet points that lead algorithmists to debate the pros and cons of various sorting algorithms.
When comparing various sorting algorithms, there are several things to consider. The first is usually runtime. When dealing with
increasingly large sets of data, inefficient sorting algorithms can become too slow for practical use within an application.
A second consideration is memory space. Faster algorithms that require recursive calls typically involve creating copies of the data to be
sorted. In some environments where memory space may be at a premium (such as an embedded system) certain algorithms may be
impractical. In other cases, it may be possible to modify the algorithm to work “in place”, without creating copies of the data. However,
this modification may also come at the cost of some of the performance advantage.
A third consideration is stability. Stability, simply defined, is what happens to elements that are comparatively the same. In a stable
sort, those elements whose comparison key is the same will remain in the same relative order after sorting as they were before sorting.
In an unstable sort, no guarantee is made as to the relative output order of those elements whose sort key is the same.
Bubble Sort
One of the first sorting algorithms that is taught to students is bubble sort. While it is not fast enough in practice for all but the smallest
data sets, it does serve the purpose of showing how a sorting algorithm works. Typically, it looks something like this:
for (int i = 0; i < data.Length; i++)
for (int j = 0; j < data.Length ‐ 1; j++)
if (data[j] > data[j + 1]){
tmp = data[j];
data[j] = data[j + 1];
data[j + 1] = tmp;
}
The idea is to pass through the data from one end to the other, and swap two adjacent elements whenever the first is greater than the
last. Thus, the smallest elements will “bubble” to the surface. This is O(n?) runtime, and hence is very slow for large data sets. The
single best advantage of a bubble sort, however, is that it is very simple to understand and code from memory. Additionally, it is a
stable sort that requires no additional memory, since all swaps are made in place.
Insertion Sort
Insertion sort is an algorithm that seeks to sort a list one element at a time. With each iteration, it takes the next element waiting to be
sorted, and adds it, in proper location, to those elements that have already been sorted.
for (int i = 0; i <= data.Length; i++) {
int j = i;
while (j > 0 && data[i] < data[j ‐ 1])
j‐‐;
int tmp = data[i];
for (int k = i; k > j; k‐‐)
data[k] = data[k ‐ 1];
data[j] = tmp;
}
The data, as it is processed on each run of the outer loop, might look like this:
{18, 6, 9, 1, 4, 15, 12, 5, 6, 7, 11}
{ 6, 18, 9, 1, 4, 15, 12, 5, 6, 7, 11}
{ 6, 9, 18, 1, 4, 15, 12, 5, 6, 7, 11}
{ 1, 6, 9, 18, 4, 15, 12, 5, 6, 7, 11}
{ 1, 4, 6, 9, 18, 15, 12, 5, 6, 7, 11}
{ 1, 4, 6, 9, 15, 18, 12, 5, 6, 7, 11}
{ 1, 4, 6, 9, 12, 15, 18, 5, 6, 7, 11}
{ 1, 4, 5, 6, 9, 12, 15, 18, 6, 7, 11}
{ 1, 4, 5, 6, 6, 9, 12, 15, 18, 7, 11}
{ 1, 4, 5, 6, 6, 7, 9, 12, 15, 18, 11}
{ 1, 4, 5, 6, 6, 7, 9, 11, 12, 15, 18}
One of the principal advantages of the insertion sort is that it works very efficiently for lists that are nearly sorted initially. Furthermore,
it can also work on data sets that are constantly being added to. For instance, if one wanted to maintain a sorted list of the highest
scores achieved in a game, an insertion sort would work well, since new elements would be added to the data as the game was played.
Merge Sort
A merge sort works recursively. First it divides a data set in half, and sorts each half separately. Next, the first elements from each of
the two lists are compared. The lesser element is then removed from its list and added to the final result list.
int[] mergeSort (int[] data) {
if (data.Length == 1)
return data;
int middle = data.Length / 2;
int[] left = mergeSort(subArray(data, 0, middle ‐ 1));
int[] right = mergeSort(subArray(data, middle, data.Length ‐ 1));
int[] result = new int[data.Length];
int dPtr = 0;
int lPtr = 0;
int rPtr = 0;
while (dPtr < data.Length) {
if (lPtr == left.Length) {
result[dPtr] = right[rPtr];
rPtr++;
} else if (rPtr == right.Length) {
result[dPtr] = left[lPtr];
lPtr++;
} else if (left[lPtr] < right[rPtr]) {
result[dPtr] = left[lPtr];
lPtr++;
} else {
result[dPtr] = right[rPtr];
rPtr++;
}
dPtr++;
}
return result;
}
Each recursive call has O(n) runtime, and a total of O(log n) recursions are required, thus the runtime of this algorithm is O(n * log n).
A merge sort can also be modified for performance on lists that are nearly sorted to begin with. After sorting each half of the data, if
the highest element in one list is less than the lowest element in the other half, then the merge step is unnecessary. (The Java API
implements this particular optimization, for instance.) The data, as the process is called recursively, might look like this:
{18, 6, 9, 1, 4, 15, 12, 5, 6, 7, 11}
{18, 6, 9, 1, 4} {15, 12, 5, 6, 7, 11}
{18, 6} {9, 1, 4} {15, 12, 5} {6, 7, 11}
{18} {6} {9} {1, 4} {15} {12, 5} {6} {7, 11}
{18} {6} {9} {1} {4} {15} {12} {5} {6} {7} {11}
{18} {6} {9} {1, 4} {15} {5, 12} {6} {7, 11}
{6, 18} {1, 4, 9} {5, 12, 15} {6, 7, 11}
{1, 4, 6, 9, 18} {5, 6, 7, 11, 12, 15}
{1, 4, 5, 6, 6, 7, 9, 11, 12, 15, 18}
Apart from being fairly efficient, a merge sort has the advantage that it can be used to solve other problems, such as determining how
“unsorted” a given list is.
Heap Sort
In a heap sort, we create a heap data structure. A heap is a data structure that stores data in a tree such that the smallest (or largest)
element is always the root node. (Heaps, also known as priority queues, were discussed in more detail in Data Structures.) To perform
a heap sort, all data from a list is inserted into a heap, and then the root element is repeatedly removed and stored back into the list.
Since the root element is always the smallest element, the result is a sorted list. If you already have a Heap implementation available
or you utilize the Java PriorityQueue (newly available in version 1.5), performing a heap sort is fairly short to code:
Heap h = new Heap();
for (int i = 0; i < data.Length; i++)
h.Add(data[i]);
int[] result = new int[data.Length];
for (int i = 0; i < data.Length; i++)
data[i] = h.RemoveLowest();
The runtime of a heap sort has an upper bound of O(n * log n). Additionally, its requirement for storage space is only that of storing the
heap; this size is linear in proportion to the size of the list. Heap sort has the disadvantage of not being stable, and is somewhat more
complicated to understand beyond just the basic implementation.
Quick Sort
A quick sort, as the name implies, is intended to be an efficient sorting algorithm. The theory behind it is to sort a list in a way very
similar to how a human might do it. First, divide the data into two groups of “high” values and “low” values. Then, recursively process
the two halves. Finally, reassemble the now sorted list.
Array quickSort(Array data) {
if (Array.Length <= 1)
return;
middle = Array[Array.Length / 2];
Array left = new Array();
Array right = new Array();
for (int i = 0; i < Array.Length; i++)
if (i != Array.Length / 2) {
if (Array[i] <= middle)
left.Add(Array[i]);
else
right.Add(Array[i]);
}
return concatenate(quickSort(left), middle, quickSort(right));
}
The challenge of a quick sort is to determine a reasonable “midpoint” value for dividing the data into two groups. The efficiency of the
algorithm is entirely dependent upon how successfully an accurate midpoint value is selected. In a best case, the runtime is O(n * log
n). In the worst casewhere one of the two groups always has only a single elementthe runtime drops to O(n?). The actual sorting of
the elements might work out to look something like this:
{18, 6, 9, 1, 4, 15, 12, 5, 6, 7, 11}
{6, 9, 1, 4, 12, 5, 6, 7, 11} {15} {18}
{6, 9, 1, 4, 5, 6, 7, 11} {12} {15} {18}
{1, 4} {5} {6, 9, 6, 7, 11} {12} {15} {18}
{1} {4} {5} {6} {6} {9, 7, 11} {12} {15} {18}
{1} {4} {5} {6} {6} {7} {9, 11} {12} {15} {18}
{1} {4} {5} {6} {6} {7} {9} {11} {12} {15} {18}
If it is known that the data to be sorted all fit within a given range, or fit a certain distribution model, this knowledge can be used to
improve the efficiency of the algorithm by choosing midpoint values that are likely to divide the data in half as close to evenly as
possible. A generic algorithm that is designed to work without respect to data types or value ranges may simply select a value from the
unsorted list, or use some random method to determine a midpoint.
Radix Sort
The radix sort was designed originally to sort data without having to directly compare elements to each other. A radix sort first takes
the leastsignificant digit (or several digits, or bits), and places the values into buckets. If we took 4 bits at a time, we would need 16
buckets. We then put the list back together, and have a resulting list that is sorted by the least significant radix. We then do the same
process, this time using the secondleast significant radix. We lather, rinse, and repeat, until we get to the most significant radix, at
which point the final result is a properly sorted list.
For example, let’s look at a list of numbers and do a radix sort using a 1bit radix. Notice that it takes us 4 steps to get our final result,
and that on each step we setup exactly two buckets:
{6, 9, 1, 4, 15, 12, 5, 6, 7, 11}
{6, 4, 12, 6} {9, 1, 15, 5, 7, 11}
{4, 12, 9, 1, 5} {6, 6, 15, 7, 11}
{9, 1, 11} {4, 12, 5, 6, 6, 15, 7}
{1, 4, 5, 6, 6, 7} {9, 11, 12, 15}
Let’s do the same thing, but now using a 2bit radix. Notice that it will only take us two steps to get our result, but each step requires
setting up 4 buckets:
{6, 9, 1, 4, 15, 12, 5, 6, 7, 11}
{4, 12} {9, 1, 5} {6, 6} {15, 7, 11}
{1} {4, 5, 6, 6, 7} {9, 11} {12, 15}
Given the relatively small scope of our example, we could use a 4bit radix and sort our list in a single step with 16 buckets:
{6, 9, 1, 4, 15, 12, 5, 6, 7, 11}
{1} {} {} {4} {5} {6, 6} {7} {} {9} {} {11} {12} {} {} {15}
Notice, however, in the last example, that we have several empty buckets. This illustrates the point that, on a much larger scale, there
is an obvious ceiling to how much we can increase the size of our radix before we start to push the limits of available memory. The
processing time to reassemble a large number of buckets back into a single list would also become an important consideration at some
point.
Because radix sort is qualitatively different than comparison sorting, it is able to perform at greater efficiency in many cases. The
runtime is O(n * k), where k is the size of the key. (32bit integers, taken 4 bits at a time, would have k = 8.) The primary
disadvantage is that some types of data may use very long keys (strings, for instance), or may not easily lend itself to a representation
that can be processed from least significant to mostsignificant. (Negative floatingpoint values are the most commonly cited example.)
Sorting Libraries
Nowadays, most programming platforms include runtime libraries that provide a number of useful and reusable functions for us. The
.NET framework, Java API, and C++ STL all provide some builtin sorting capabilities. Even better, the basic premise behind how they
work is similar from one language to the next.
For standard data types such as scalars, floats, and strings, everything needed to sort an array is already present in the standard
libraries. But what if we have custom data types that require more complicated comparison logic? Fortunately, objectoriented
programming provides the ability for the standard libraries to solve this as well.
In both Java and C# (and VB for that matter), there is an interface called Comparable (IComparable in .NET). By implementing the
IComparable interface on a userdefined class, you add a method int CompareTo (object other), which returns a negative value if less
than, 0 if equal to, or a positive value if greater than the parameter. The library sort functions will then work on arrays of your new data
type.
Additionally, there is another interface called Comparator (IComparer in .NET), which defines a single method int Compare (object
obj1, object obj2), which returns a value indicating the results of comparing the two parameters.
The greatest joy of using the sorting functions provided by the libraries is that it saves a lot of coding time, and requires a lot less
thought and effort. However, even with the heavy lifting already completed, it is still nice to know how things work under the hood.
Maximum Flow: Section 1
By _efer_ — TopCoder Member
Discuss this article in the forums
Introduction
This article covers a problem that often arises in real life situations and, as expected, in programming challenges, with Top Coder being
no exception. It is addressed mostly to coders who are not familiar with the subject, but it may prove useful to the more experienced as
well. Lots of papers have been written, and there are many algorithms known to solve this problem. While they are not the fastest, the
algorithms presented here have the advantage of being simple and efficient, and because of this they are usually preferred during a
challenge setup. The reader is advised to read the article on graph theory first, as the concepts presented there are needed to
understand those presented here.
The Standard Maximum Flow Problem
So, what are we being asked for in a maxflow problem? The simplest form that the statement could take would be something along the
lines of: “A list of pipes is given, with different flowcapacities. These pipes are connected at their endpoints. What is the maximum
amount of water that you can route from a given starting point to a given ending point?” or equivalently “A company owns a factory
located in city X where products are manufactured that need to be transported to the distribution center in city Y. You are given the
oneway roads that connect pairs of cities in the country, and the maximum number of trucks that can drive along each road. What is
the maximum number of trucks that the company can send to the distribution center?”
A first observation is that it makes no sense to send a truck to any other city than Y, so every truck that enters a city other than Y must
also leave it. A second thing to notice is that, because of this, the number of trucks leaving X is equal to the number of trucks arriving
in Y.
Rephrasing the statement in terms of graph theory, we are given a network a directed graph, in which every edge has a certain
capacity c associated with it, a starting vertex (the source, X in the example above), and an ending vertex (the sink). We are asked to
associate another value f satisfying f ? c for each edge such that for every vertex other than the source and the sink, the sum of the
values associated to the edges that enter it must equal the sum of the values associated to the edges that leave it. We will call f the
flow along that edge. Furthermore, we are asked to maximize the sum of the values associated to the arcs leaving the source, which is
the total flow in the network.
The image below shows the optimal solution to an instance of this problem, each edge being labeled with the values f/c associated to it.
How to Solve It
Now how do we actually solve the problem? First, let us define two basic concepts for understanding flow networks: residual networks
and augmenting paths. Consider an arbitrary flow in a network. The residual network has the same vertices as the original network,
and one or two edges for each edge in the original. More specifically, if the flow along the edge xy is less than the capacity there is a
forward edge xy with a capacity equal to the difference between the capacity and the flow (this is called the residual capacity), and if
the flow is positive there is a backward edge yx with a capacity equal to the flow on xy. An augmenting path is simply a path from the
source to the sink in the residual network, whose purpose is to increase the flow in the original one. It is important to understand that
the edges in this path can point the “wrong way” according to the original network. The path capacity of a path is the minimum capacity
of an edge along that path. Let’s take the following example:
By considering the path X_A_C_Y, we can increase the flow by 1 – the edges X_A and A_C have capacity of 3, as in the original
network, but the edge C_Y has capacity 1, and we take the minimum of these values to get the path capacity. Increasing the flow along
this path with 1 yields the flow below:
The value of the current flow is now 2, and as shown in Figure 1, we could do better. So, let’s try to increase the flow. Clearly, there is
no point in considering the directed paths X_A_C_Y or X_B_D_E_Y as the edges C_Y and X_B, respectively, are filled to capacity. As a
matter of fact, there is no directed path in the network shown above, due to the edges mentioned above being filled to capacity. At this
point, the question that naturally comes to mind is: is it possible to increase the flow in this case? And the answer is yes, it is. Let’s
take a look at the residual network:
Let’s consider the only path from X to Y here: X_A_C_B_D_E_Y. Note that this is not a path in the directed graph, because C_B is
walked in the opposite way. We’ll use this path in order to increase the total flow in the original network. We’ll “push” flow on each of
the edges, except for C_B which we will use in order to “cancel” flow on B_C. The amount by which this operation can be performed is
limited by the capacities of all edges along the path (as shown in Figure 3b). Once again we take the minimum, to conclude that this
path also has capacity 1. Updating the path in the way described here yields the flow shown in Figure 1a. We are left with the following
residual network where a path between the source and the sink doesn’t exist:
This example suggests the following algorithm: start with no flow everywhere and increase the total flow in the network while there is
an augmenting path from the source to the sink with no full forward edges or empty backward edges a path in the residual network.
The algorithm (known as the FordFulkerson method) is guaranteed to terminate: due to the capacities and flows of the edges being
integers and the pathcapacity being positive, at each step we get a new flow that is closer to the maximum. As a side note, the
algorithm isn’t guaranteed to even terminate if the capacities are irrationals.
What about the correctness of this algorithm? It is obvious that in a network in which a maximum flow has been found there is no
augmenting path, otherwise we would be able to increase the maximum value of the flow, contradicting our initial assumption. If the
converse of this affirmation is true, so that when there is no augmenting path, the value of the flow has reached its maximum, we can
breathe a sigh of relief, our algo is correct and computes the maximum flow in a network. This is known as the maxflow mincut
theorem and we shall justify its correctness in a few moments.
A cut in a flow network is simply a partition of the vertices in two sets, let’s call them A and B, in such a way that the source vertex is in
A and the sink is in B. The capacity of a cut is the sum of the capacities of the edges that go from a vertex in A to a vertex in B. The
flow of the cut is the difference of the flows that go from A to B (the sum of the flows along the edges that have the starting point in A
and the ending point in B), respectively from B to A, which is exactly the value of the flow in the network, due to the entering flow
equals leaving flow – property, which is true for every vertex other than the source and the sink.
Notice that the flow of the cut is less or equal to the capacity of the cut due to the constraint of the flow being less or equal to the
capacity of every edge. This implies that the maximum flow is less or equal to every cut of the network. This is where the maxflow
mincut theorem comes in and states that the value of the maximum flow through the network is exactly the value of the minimum cut
of the network. Let’s give an intuitive argument for this fact. We will assume that we are in the situation in which no augmenting path
in the network has been found. Let’s color in yellow, like in the figure above, every vertex that is reachable by a path that starts from
the source and consists of nonfull forward edges and of nonempty backward edges. Clearly the sink will be colored in blue, since there
is no augmenting path from the source to the sink. Now take every edge that has a yellow starting point and a blue ending point. This
edge will have the flow equal to the capacity, otherwise we could have added this edge to the path we had at that point and color the
ending point in yellow. Note that if we remove these edges there will be no directed path from the source to the sink in the graph. Now
consider every edge that has a blue starting point and a yellow ending point. The flow on this edge must be 0 since otherwise we could
have added this edge as a backward edge on the current path and color the starting point in yellow. Thus, the value of the flow must
equal the value of the cut, and since every flow is less or equal to every cut, this must be a maximum flow, and the cut is a minimum
cut as well.
In fact, we have solved another problem that at first glance would appear to have nothing to do with maximum flow in a network, ie.
given a weighted directed graph, remove a minimumweighted set of edges in such a way that a given node is unreachable from
another given node. The result is, according to the maxflow mincut theorem, the maximum flow in the graph, with capacities being
the weights given. We are also able to find this set of edges in the way described above: we take every edge with the starting point
marked as reachable in the last traversal of the graph and with an unmarked ending point. This edge is a member of the minimum cut.
AugmentingPath Algorithms
The neat part of the FordFulkerson algorithm described above is that it gets the correct result no matter how we solve (correctly!!) the
subproblem of finding an augmenting path. However, every new path may increase the flow by only 1, hence the number of iterations
of the algorithm could be very large if we carelessly choose the augmenting path algorithm to use. The function max_flow will look like
this, regardless of the actual method we use for finding augmenting paths:
int max_flow()
result = 0
while (true)
// the function find_path returns the path capacity of the
// augmenting path found
path_capacity = find_path()
// no augmenting path found
if (d = 0) exit while
else result += path_capacity
end while
return result
To keep it simple, we will use a 2dimensional array for storing the capacities of the residual network that we are left with after each
step in the algorithm. Initially the residual network is just the original network. We will not store the flows along the edges explicitly,
but it’s easy to figure out how to find them upon the termination of the algorithm: for each edge xy in the original network the flow is
given by the capacity of the backward edge yx in the residual network. Be careful though; if the reversed arc yx also exists in the
original network, this will fail, and it is recommended that the initial capacity of each arc be stored somewhere, and then the flow along
the edge is the difference between the initial and the residual capacity.
We now require an implementation for the function find_path. The first approach that comes to mind is to use a depthfirst search
(DFS), as it probably is the easiest to implement. Unfortunately, its performance is very poor on some networks, and normally is less
preferred to the ones discussed next.
The next best thing in the matter of simplicity is a breadthfirst search (BFS). Recall that this search usually yields the shortest path in
an unweighted graph. Indeed, this also applies here to get the shortest augmenting path from the source to the sink. In the following
pseudocode we will basically: find a shortest path from the source to the sink and compute the minimum capacity of an edge (that
could be a forward or a backward edge) along the path – the path capacity. Then, for each edge along the path we reduce its capacity
and increase the capacity of the reversed edge with the path capacity.
int bfs()
queue Q
push source to Q
mark source as visited
keep an array from with the semnification: from[x] is the
previous vertex visited in the shortest path from the source to x;
initialize from with ‐1 (or any other sentinel value)
while Q is not empty
where = pop from Q
for each vertex next adjacent to where
if next is not visited and capacity[where][next] > 0
push next to Q
mark next as visited
from[next] = where
if next = sink
exit while loop
end for
end while
// we compute the path capacity
where = sink, path_cap = infinity
while from[where] > ‐1
prev = from[where] // the previous vertex
path_cap = min(path_cap, capacity[prev][where])
where = prev
end while
// we update the residual network; if no path is found the while
loop will not be entered
where = sink
while from[where] > ‐1
prev = from[where]
capacity[prev][where] ‐= path_capacity
capacity[where][prev] += path_capacity
where = prev
end while
// if no path is found, path_cap is infinity
if path_cap = infinity
return 0
else return path_cap
As we can see, this is pretty easy to implement. As for its performance, it is guaranteed that this takes at most N * M/2 steps, where N
is the number of vertices and M is the number of edges in the network. This number may seem very large, but it is overestimated for
most networks. For example, in the network we considered 3 augmenting paths are needed which is significantly less than the upper
bound of 28. Due to the O(M) running time of BFS (implemented with adjacency lists) the worstcase running time of the shortest
augmenting path maxflow algorithm is O(N * M²), but usually the algorithm performs much better than this.
Next we will consider an approach that uses a priorityfirst search (PFS), that is very similar to the Dijkstra heap method explained
here. In this method the augmenting path with a maximum path capacity is preferred. Intuitively this would lead to a faster algorithm,
since at each step we increase the flow with the maximum possible amount. However, things are not always so, and the BFS
implementation has better running times on some networks. We assign as a priority to each vertex the minimum capacity of a path (in
the residual network) from the source to that vertex. We process vertices in a greedy manner, as in Dijkstra’s algorithm, in decreasing
order of priorities. When we get to the sink, we are done, since a path with a maximum capacity is found. We would like to implement
this with a data structure that allows us to efficiently find the vertex with the highest priority and increase the priority of a vertex (when
a new better path is found) – this suggests the use of a heap which has a space complexity proportional to the number of vertices. In
TopCoder matches we may find it faster and easier to implement this with a priority queue or some other data structure that
approximates one, even though the space required might grow to being proportional with the number of edges. This is how the
following pseudocode is implemented. We also define a structure node that has the members vertex and priority with the above
significance. Another field from is needed to store the previous vertex on the path.
int pfs()
priority queue PQ
push node(source, infinity, ‐1) to PQ
keep the array from as in bfs()
// if no augmenting path is found, path_cap will remain 0
path_cap = 0
while PQ is not empty
node aux = pop from PQ
where = aux.vertex, cost = aux.priority
if we already visited where continue
from[where] = aux.from
if where = sink
path_cap = cost
exit while loop
mark where as visited
for each vertex next adjacent to where
if capacity[where][next] > 0
new_cost = min(cost, capacity[where][next])
push node(next, new_cost, where) to PQ
end for
end while
// update the residual network
where = sink
while from[where] > ‐1
prev = from[where]
capacity[prev][where] ‐= path_cap
capacity[where][prev] += path_cap
where = prev
end while
return path_cap
The analysis of its performance is pretty complicated, but it may prove worthwhile to remember that with PFS at most 2M1gU steps are
required, where U is the maximum capacity of an edge in the network. As with BFS, this number is a lot larger than the actual number
of steps for most networks. Combine this with the O(M 1g M) complexity of the search to get the worstcase running time of this
algorithm.
Now that we know what these methods are all about, which of them do we choose when we are confronted with a maxflow problem?
The PFS approach seems to have a better worstcase performance, but in practice their performance is pretty much the same. So, the
method that one is more familiar with may prove more adequate. Personally, I prefer the shortestpath method, as I find it easier to
implement during a challenge and less error prone.
Maximum Flow: Section 2
By _efer_ — TopCoder Member
Section 2
MaxFlow/MinCut Related Problems
How to recognize maxflow problems? Often they are hard to detect and usually boil down to maximizing the movement of something
from a location to another. We need to look at the constraints when we think we have a working solution based on maximum flow –
they should suggest at least an O(N?) approach. If the number of locations is large, another algorithm (such as dynamic programming or
greedy), is more appropriate.
The problem description might suggest multiple sources and/or sinks. For example, in the sample statement in the beginning of this
article, the company might own more than one factory and multiple distribution centers. How can we deal with this? We should try to
convert this to a network that has a unique source and sink. In order to accomplish this we will add two “dummy” vertices to our
original network – we will refer to them as supersource and supersink. In addition to this we will add an edge from the supersource
to every ordinary source (a factory). As we don’t have restrictions on the number of trucks that each factory can send, we should
assign to each edge an infinite capacity. Note that if we had such restrictions, we should have assigned to each edge a capacity equal to
the number of trucks each factory could send. Likewise, we add an edge from every ordinary sink (distribution centers) to the super
sink with infinite capacity. A maximum flow in this newbuilt network is the solution to the problem – the sources now become ordinary
vertices, and they are subject to the enteringflow equals leavingflow property. You may want to keep this in your bag of tricks, as it
may prove useful to most problems.
What if we are also given the maximum number of trucks that can drive through each of the cities in the country (other than the cities
where the factory and the distribution center are located)? In other words we have to deal with vertexcapacities too. Intuitively, we
should be able to reduce this to maximumflow, but we must find a way to take the capacities from vertices and put them back on
edges, where they belong. Another nice trick comes into play. We will build a network that has two times more vertices than the initial
one. For each vertex we will have two nodes: an invertex and an outvertex, and we will direct each edge xy from the outvertex of x
to the invertex of y. We can assign them the capacities from the problem statement. Additionally we can add an edge for each vertex
from the in to the outvertex. The capacity this edge will be assigned is obviously the vertexcapacity. Now we just run maxflow on this
network and compute the result.
Maximum flow problems may appear out of nowhere. Let’s take this problem for instance: “You are given the in and out degrees of the
vertices of a directed graph. Your task is to find the edges (assuming that no edge can appear more than once).” First, notice that we
can perform this simple test at the beginning. We can compute the number M of edges by summing the outdegrees or the indegrees
of the vertices. If these numbers are not equal, clearly there is no graph that could be built. This doesn’t solve our problem, though.
There are some greedy approaches that come to mind, but none of them work. We will combine the tricks discussed above to give a
maxflow algorithm that solves this problem. First, build a network that has 2 (in/out) vertices for each initial vertex. Now draw an edge
from every out vertex to every in vertex. Next, add a supersource and draw an edge from it to every outvertex. Add a supersink and
draw an edge from every in vertex to it. We now need some capacities for this to be a flow network. It should be pretty obvious what
the intent with this approach is, so we will assign the following capacities: for each edge drawn from the supersource we assign a
capacity equal to the outdegree of the vertex it points to. As there may be only one arc from a vertex to another, we assign a 1
capacity to each of the edges that go from the outs to the ins. As you can guess, the capacities of the edges that enter the supersink
will be equal to the indegrees of the vertices. If the maximum flow in this network equals M – the number of edges, we have a
solution, and for each edge between the out and in vertices that has a flow along it (which is maximum 1, as the capacity is 1) we can
draw an edge between corresponding vertices in our graph. Note that both xy and yx edges may appear in the solution. This is very
similar to the maximum matching in a bipartite graph that we will discuss later. An example is given below where the outdegrees are
(2, 1, 1, 1) and the indegrees (1, 2, 1, 1).
Some other problems may ask to separate two locations minimally. Some of these problems usually can be reduced to minimumcut in
a network. Two examples will be discussed here, but first let’s take the standard mincut problem and make it sound more like a
TopCoder problem. We learned earlier how to find the value of the mincut and how to find an arbitrary mincut. In addition to this we
will now like to have a minimumcut with the minimum number of edges. An idea would be to try to modify the original network in such
a way that the minimum cut here is the minimum cut with the minimum edges in the original one. Notice what happens if we multiply
each edge capacity with a constant T. Clearly, the value of the maximum flow is multiplied by T, thus the value of the minimum cut is T
times bigger than the original. A minimum cut in the original network is a minimum cut in the modified one as well. Now suppose we
add 1 to the capacity of each edge. Is a minimum cut in the original network a minimum cut in this one? The answer is no, as we can
see in Figure 8 shown below, if we take T = 2.
Why did this happen? Take an arbitrary cut. The value of the cut will be T times the original value of the cut, plus the number of edges
in it. Thus, a nonminimum cut in the first place could become minimum if it contains just a few edges. This is because the constant
might not have been chosen properly in the beginning, as is the case in the example above. We can fix this by choosing T large enough
to neutralize the difference in the number of edges between cuts in the network. In the above example T = 4 would be enough, but to
generalize, we take T = 10, one more than the number of edges in the original network, and one more than the number of edges that
could possibly be in a minimumcut. It is now true that a minimumcut in the new network is minimum in the original network as well.
However the converse is not true, and it is to our advantage. Notice how the difference between minimum cuts is now made by the
number of edges in the cut. So we just find the mincut in this new network to solve the problem correctly.
Let’s illustrate some more the mincut pattern: “An undirected graph is given. What is the minimum number of edges that should be
removed in order to disconnect the graph?” In other words the problem asks us to remove some edges in order for two nodes to be
separated. This should ring a bell – a minimum cut approach might work. So far we have only seen maximum flow in directed graphs,
but now we are facing an undirected one. This should not be a very big problem though, as we can direct the graph by replacing every
(undirected) edge xy with two arcs: xy and yx. In this case the value of the mincut is the number of edges in it, so we assign a 1
capacity to each of them. We are not asked to separate two given vertices, but rather to disconnect optimally any two vertices, so we
must take every pair of vertices and treat them as the source and the sink and keep the best one from these minimumcuts. An
improvement can be made, however. Take one vertex, let’s say vertex numbered 1. Because the graph should be disconnected, there
must be another vertex unreachable from it. So it suffices to treat vertex 1 as the source and iterate through every other vertex and
treat it as the sink.
What if instead of edges we now have to remove a minimum number of vertices to disconnect the graph? Now we are asked for a
different mincut, composed of vertices. We must somehow convert the vertices to edges though. Recall the problem above where we
converted vertexcapacities to edgecapacities. The same trick works here. First “undirect” the graph as in the previous example. Next
double the number of vertices and deal edges the same way: an edge xy is directed from the outx vertex to iny. Then convert the
vertex to an edge by adding a 1capacity arc from the invertex to the outvertex. Now for each two vertices we must solve the sub
problem of minimally separating them. So, just like before take each pair of vertices and treat the outvertex of one of them as the
source and the invertex of the other one as the sink (this is because the only arc leaving the invertex is the one that goes to the out
vertex) and take the lowest value of the maximum flow. This time we can’t improve in the quadratic number of steps needed, because
the first vertex may be in an optimum solution and by always considering it as the source we lose such a case.
Maximum Bipartite Matching
This is one of the most important applications of maximum flow, and a lot of problems can be reduced to it. A matching in a graph is a
set of edges such that no vertex is touched by more than one edge. Obviously, a matching with a maximum cardinality is a maximum
matching. For a general graph, this is a hard problem to deal with.
Let’s direct our attention towards the case where the graph is bipartite – its vertices can be split into two sets such that there is no
edge connecting vertices from the same set. In this case, it may sound like this: “Each of your employees can handle a given set of
jobs. Assign a job to as many of them as you can.”
A bipartite graph can be built in this case: the first set consists of your employees while the second one contains the jobs to be done.
There is an edge from an employee to each of the jobs he could be assigned. An example is given below:
So, Joe can do jobs B, C and D while Mark wouldn’t mind being assigned jobs A, D or E. This is a happy case in which each of your
employees is assigned a job:
In order to solve the problem we first need to build a flow network. Just as we did in the multiplesource multiplesink problem we will
add two “dummy” vertices: a supersource and a supersink, and we will draw an edge from the supersource to each of the vertices in
set A (employees in the example above) and from each vertex in set B to the supersink. In the end, each unit of flow will be equivalent
to a match between an employee and a job, so each edge will be assigned a capacity of 1. If we would have assigned a capacity larger
than 1 to an edge from the supersource, we could have assigned more than one job to an employee. Likewise, if we would have
assigned a capacity larger than 1 to an edge going to the supersink, we could have assigned the same job to more than one employee.
The maximum flow in this network will give us the cardinality of the maximum matching. It is easy to find out whether a vertex in set B
is matched with a vertex x in set A as well. We look at each edge connecting x to a vertex in set B, and if the flow is positive along one
of them, there exists a match. As for the running time, the number of augmenting paths is limited by min(|A|,|B|), where by |X| is denoted
the cardinality of set X, making the running time O(N?M), where N is the number of vertices, and M the number of edges in the graph.
An implementation point of view is in place. We could implement the maximum bipartite matching just like in the pseudocode given
earlier. Usually though, we might want to consider the particularities of the problem before getting to the implementation part, as they
can save time or space. In this case, we could drop the 2dimensional array that stored the residual network and replace it with two
onedimensional arrays: one of them stores the match in set B (or a sentinel value if it doesn’t exist) for each element of set A, while
the other is the other way around. Also, notice that each augmenting path has capacity 1, as it contributes with just a unit of flow. Each
element of set A can be the first (well, the second, after the supersource) in an augmenting path at most once, so we can just iterate
through each of them and try to find a match in set B. If an augmenting path exists, we follow it. This might lead to dematching other
elements along the way, but because we are following an augmenting path, no element will eventually remain unmatched in the
process.
Now let’s solve some TopCoder problems!
RookAttack
Problem Statement
This problem asks us to place a maximum number of rooks on a rows x cols chessboard with some squares cut out. The idea behind this
might be a little hard to spot, but once this is done, we get into a standard maximum bipartitematching problem.
Notice that at most one rook can be placed on each row or column. In other words, each row corresponds at most to one column where
a rook can be placed. This suggests a bipartite matching where set A is composed of elements corresponding to every row of the board,
while set B consists of the columns. For each row add edges to every column if the corresponding square is not cut out of the board.
Now we can just run maximum bipartitematching in this network and compute the result. Since there are at most rows * cols edges, the
time complexity of the algorithm is: O(rows? * cols)
In the C++ code below BFS is used for finding an augmentingpath:
class RookAttack{
// a list of the non‐empty squares for each row
vector lst[300];
// in this arrays we keep matches found to every row and column
int row_match[300], col_match[300];
// we search for an augmenting path starting with row source
bool find_match(int source) {
// from[x] = the row‐vertex that precedes x in the path
int from[300], where, match;
memset(from, ‐1, sizeof(from));
from[source] = source;
deque q;
q.push_back(source);
bool found_path = false;
while (!found_path && !q.empty()) {
// where = current row‐vertex we are in
where = q.front(); q.pop_front();
// we take every uncut square in the current row
for (int i = 0; i < lst[where].size(); ++ i) {
match = lst[where][i];
// next = the row matched with column match
int next = col_match[match];
if (where != next) {
// no row matched with column match thus
// we found an augmenting path
if (next == ‐1) {
found_path = true;
break;
}
// a check whether we already visited
// the row‐vertex next
if (from[next] == ‐1) {
q.push_back(next);
from[next] = where;
}
}
}
}
if (!found_path)
return false;
while (from[where] != where) {
// we de‐match where from its current match (aux)
// and match it with match
int aux = row_match[where];
row_match[where] = match;
col_match[match] = where;
where = from[where];
match = aux;
}
// at this point where = source
row_match[where] = match;
col_match[match] = where;
return true;
}
public:
int howMany(int rows, int cols, vector cutouts) {
// build lst from cutouts; column j should appear in
// row's i list if square (i, j) is present on the board
int ret = 0;
memset(row_match, ‐1, sizeof(row_match));
memset(col_match, ‐1, sizeof(col_match));
// we try to find a match for each row
for (int i = 0; i < rows; ++ i)
ret += find_match(i);
return ret;
}
};
Let’s take a look at the DFS version, too. We can implement the find_match function like this: for each nonempty square in the current
row try to match the row with its corresponding column and call find_match recursively to attempt to find a new match for the current
match (if the current match exists – if not, an augmenting path is found) of this column. If one is found, we can perform the desired
match. Note that to make this run in time we must not visit the same column (or row) twice. Notice the C++ code below is extremely
short:
bool find_match(int where) {
// the previous column was not matched
if (where == ‐1)
return true;
for (int i = 0; i < lst[where].size(); ++ i) {
int match = lst[where][i];
if (visited[match] == false) {
visited[match] = true;
if (find_match(col_match[match])) {
col_match[match] = where;
return true;
}
}
}
return false;
}
This runs in time because the number of augmenting paths is the same for both versions. The only difference is that BFS finds the
shortest augmentingpath while DFS finds a longer one. As implementation speed is an important factor in TopCoder matches, in this
case it would be a good deal to use the slower, but easier DFS version.
The following version of the problem is left as an exercise for the reader: to try and place as many rooks as possible on the board in
such a way that the number of rooks on each row is equal to the number of rooks on each column (it is allowed for two rooks to attack
each other).
Graduation
Problem Statement
In this problem we are given a set of requirements, each stating that a number of classes should be taken from a given set of classes.
Each class may be taken once and fulfills a single requirement. Actually, the last condition is what makes the problem harder, and
excludes the idea of a greedy algorithm. We are also given a set of classes already taken. If it weren’t for this, to ensure the minimality
of the return, the size of the returned string would have been (if a solution existed) the sum of the number of classes for each
requirement. Now as many classes as possible must be used from this set.
At first glance, this would have been a typical bipartitematching problem if every requirement had been fulfilled by taking just a single
class. Set A would have consisted of the classes available (all characters with ASCII code in the range 33126, except for the numeric
characters ’0′’9′), while the set of requirements would have played the role of set B. This can be taken care of easily. Each requirement
will contribute to set B with a number of elements equal to the number of classes that must be taken in order to fulfill it – in other
words, split each requirement into several requirements. At this point, a bipartitematching algorithm can be used, but care should be
allotted to the order in which we iterate through the set of classes and match a class with a requirement.
It is important to understand that any order to iterate through set A can be considered when solving the standard bipartitematching
problem. For example, it doesn’t matter what element from set A we choose to be the first one to be matched. Consider the solution
found by the algorithm containing this element x from A, matched with an element y from B. Also, we should consider any optimal
solution. Clearly, in the optimal, y must be matched with an element z from A, otherwise we can add the pair xy to the matching,
contradicting the fact that the solution is optimal. Then, we can just exchange z with x to come with a solution of the same cardinality,
which completes the proof.
That being said, to gain as much as possible from the classes already taken we first must match each of these with a requirement. If,
after completing this step, all requirements are fulfilled, we just need to return the empty string, as there is no need for taking more
classes. Now we have to deal with the requirement that the return must be the first in lexicographic order. It should be obvious now
that the other classes must be considered in increasing order. If a match is found for a class, that class is added to the return value. In
the end, if not every requirement is fulfilled, we don’t have a solution. The implementation is left as an exercise for the reader.
As a final note, it is possible to speed things up a bit. To achieve this, we will drop the idea of splitting each requirement. Instead we
will modify the capacities of the edges connecting those with the supersink. They will now be equal to the number of classes to be
taken for each requirement. Then we can just go on with the same approach as above.
Parking
Problem Statement
In this problem we have to match each of the cars with a parking spot. Additionally the time it takes for all cars to find a parking spot
must be minimized. Once again we build a bipartite graph: set A is the set that consists of the cars and set B contains the parking
spots. Each edge connecting elements from different sets has as the cost (and not the capacity!) the time required for the car to get to
the parking spot. If the spot is unreachable, we can assign it an infinite cost (or remove it). These costs are determined by running
breadthfirst search.
For solving it, assume that the expected result is less than or equal to a constant D. Then, there exists a matching in which each edge
connecting a car and a parking spot has the cost less than or equal to D. Thus, removing edges with cost greater than D will have no
effect on the solution. This suggests a binary search on D, removing all edges with cost greater than D, and then performing a
maximum bipartitematching algorithm. If a matching exists in which every car can drive to a parking spot, we can decrease D
otherwise we must increase it.
However, there is a faster and more elegant solution using a priorityfirst search. Instead of keeping D fixed as above, we could try to
successively increase D whenever we find that it is too low. We will start with D = 0. Then we iterate through each of the cars and try to
find an augmenting path in which no edge has a cost larger than D. If none exists, we increase D until one path exists. Obviously, we
will increase it with the smallest possible amount. In order to achieve this, we will search for the augmenting path with the smallest
cost – the cost of the path is the maximum cost of an edge on that path. This can be done with a priorityfirst search similar to the PFS
augmentingpath algorithm presented in the first section of the article. C++ code follows:
struct node {
int where, cost, from;
node(int _where, int _cost, int _from): where(_where),
cost(_cost), from(_from) {};
};
bool operator < (node a, node b) {
return a.cost > b.cost;
}
int minTime(vector park){
// build a cost matrix cost[i][j] = cost of getting from car i to
// parking spot j, by doing a BFS
// vertices 0, 1, ..., N ‐ 1 will represent the cars, and
// vertices N, N + 1, ..., N + M ‐ 1 will represent
// the parking spots; N + M will be the super‐sink
int D = 0, sink = N + M;
int car_match[105], park_match[105];
memset(car_match, ‐1, sizeof(car_match));
memset(park_match, ‐1, sizeof(park_match));
for (int source = 0; source < N; ++ source) {
bool visited[210];
memset(visited, false, sizeof(visited));
int from[210];
memset(from, ‐1, sizeof(from));
priority_queue pq;
pq.push(node(source, 0, ‐1));
while (!pq.empty()) {
int cst = pq.top().cost, where = pq.top().where,
_from = pq.top().from;
pq.pop();
if (visited[where]) continue;
visited[where] = true;
from[where] = _from;
// if where is a car try all M parking spots
if (where < N) {
for (int i = 0; i < M; ++ i) {
// if the edge doesn't exist or this car
// is already matched with this parking spot
if (cost[where][i] == infinity ||
car_match[where] == i) continue;
int ncst = max(cst, cost[where][i]);
// the i‐th parking spot is N + i
pq.push(node(N + i, ncst, where));
}
}
else {
// if this parking spot is unmatched we found
the augmenting path with minimum cost
if (park_match[where ‐ N] == ‐1) {
from[sink] = where;
// if D needs to be increased, increase it
D = max(D, cst);
break;
}
// otherwise we follow the backward edge
int next = park_match[where ‐ N];
int ncst = max(cst, cost[next][where]);
pq.push(node(next, ncst, where));
}
}
int where = from[sink];
// if no augmenting path is found we have no solution
if (where == ‐1)
return ‐1;
// follow the augmenting path
while (from[where] > ‐1) {
int prev = from[where];
// if where is a parking spot the edge (prev, where)
// is a forward edge and the match must be updated
if (where >= N) {
car_match[prev] = where;
park_match[where ‐ N] = prev;
}
where = prev;
}
}
return D;
}
Here are some problems to practice:
PlayingCubes – for this one ignore the low constraints and try to find a maxflow algorithm
DataFilter – be warned this is the hard problem from the TCCC 2004 Finals and is tough indeed!
Some other problems from http://acm.uva.es/p/: 563, 753, 820, 10122, 10330, 10511, 10735.
For any questions or comments please use the Forums.
Representation of Integers and Reals: Section 1
By misof — TopCoder Member
Discuss this article in the forums
Choosing the correct data type for your variables can often be the only difference between a faulty solution and a correct one.
Especially when there’s some geometry around, precision problems often cause solutions to fail. To make matters even worse, there are
many (often incorrect) rumors about the reasons of these problems and ways how to solve them.
To be able to avoid these problems, one has to know a bit about how things work inside the computer. In this article we will take a look
at the necessary facts and disprove some false rumors. After reading and understanding it, you should be able to avoid the problems
mentioned above.
This article is in no way intended to be a complete reference, nor to be 100% accurate. Several times, presented things will be a bit
simplified. As the readers of this article are TopCoder (TC) members, we will concentrate on the x86 architecture used by the machines
TC uses to evaluate solutions. For example, we will assume that on our computers a byte consists of 8 bits and that the machines use
32bit integer registers.
While most of this article is general and can be applied on all programming languages used at TC, the article is slightly biased towards
C++ and on some occasions special notes on g++ are included.
We will start by presenting a (somewhat simplified) table of integer data types available in the g++ compiler. You can find this table in
any g++ reference. All of the other compilers used at TC have similar data types and similar tables in their references, look one up if
you don’t know it by heart yet. Below we will explain that all we need to know is the storage size of each of the types, the range of
integers it is able to store can be derived easily.
Table 1: Integer data types in g++.
size in representable
name
bits range
char 8 27to 27 – 1
unsigned char 8 0 to 28 – 1
short 16 215 to 215 – 1
unsigned
short
16 0 to 216 – 1
int 32 231 to 231 – 1
unsigned int 32 0 to 232 – 1
long 32 231 to 231 – 1
unsigned long 32 0 to 232 – 1
long long 64 263 to 263 – 1
unsigned long
long
64 0 to 264 – 1
Notes:
The storage size of an int and an unsigned int is platform dependent. E.g., on machines using 64bit registers, ints in g++ will have
64 bits. The old Borland C compiler used 16bit ints. It is guaranteed that an int will always have at least 16 bits. Similarly, it is
guaranteed that on any system a long will have at least 32 bits.
The type long long is a g++ extension, it is not a part of any C++ standard (yet?). Many other C++ compilers miss this data type
or call it differently. E.g., MSVC++ has __int64 instead.
Rumor: Signed integers are stored using a sign bit and “digit” bits.
Validity: Only partially true.
Most of the current computers, including those used at TC, store the integers in a socalled two’s complement form. It is true that for
nonnegative integers the most significant bit is zero and for negative integers it is one. But this is not exactly a sign bit, we can’t
produce a “negative zero” by flipping it. Negative numbers are stored in a somewhat different way. The negative number n is stored as
a bitwise negation of the nonnegative number (n1).
In Table 2 we present the bit patterns that arise when some small integers are stored in a (signed) char variable. The rightmost bit is
the least significant one.
Table 2: Two’s complement bit patterns for some integers.
two’s
value complement
form
0 00000000
1 00000001
2 00000010
46 00101110
47 00101111
127 01111111
1 11111111
2 11111110
3 11111101
47 11010001
127 10000001
128 10000000
Note that due to the way negative numbers are stored the set of representable numbers is not placed symmetrically around zero. The
largest representable integer in b bits is 2b1 1, the smallest (i.e., most negative) one is 2b1.
A neat way of looking at the two’s complement form is that the bits correspond to digits in base 2 with the exception that the largest
power of two is negative. E.g., the bit pattern 11010001 corresponds to 1 x ( 128) + 1 x 64 + 0 x 32 + 1 x 16 + 0 x 8 + 0 x 4 + 0 x 2
+ 1 x 1 = – 128 + 81 = – 47
Rumor: Unsigned integers are just stored as binary digits of the number.
Validity: True.
In general, the bit pattern consists of base 2 digits of the represented number. E.g., the bit pattern 11010001 corresponds to 1 x 128 +
1 x 64 + 0 x 32 + 1 x 16 + 0 x 8 + 0 x 4 + 0 x 2 + 1 x 1 = 209.
Thus, in a bbit unsigned integer variable, the smallest representable number is zero and the largest is 2b – 1 (corresponding to an all
ones pattern).
Note that if the leftmost (most significant) bit is zero, the pattern corresponds to the same value regardless of whether the variable is
signed or unsigned. If we have a bbit pattern with the leftmost bit set to one, and the represented unsigned integer is x, the same
pattern in a signed variable represents the value x – 2b.
In our previous examples, the pattern 11010001 can represent either 209 (in an unsigned variable) or 47 (in a signed variable).
Rumor: In C++, the code “int A[1000]; memset(A,x,sizeof(A));” stores 1000 copies of x into A.
Validity: False.
The memset() function fills a part of the memory with chars, not ints. Thus for most values of x you would get unexpected results.
However, this does work (and is often used) for two special values of x: 0 and 1. The first case is straightforward. By filling the entire
array with zeroes, all the bits in each of the ints will be zero, thus representing the number 0. Actually, the second case is the same
story: 1 stored in a char is 1111111, thus we fill the entire array with ones, getting an array containing 1s.
(Note that most processors have a special set of instructions to fill a part of memory with a given value. Thus the memset() operation is
usually much faster than filling the array in a cycle.)
When you know what you are doing, memset() can be used to fill the array A with sufficiently large/small values, you just have to supply
a suitable bit pattern as the second argument. E.g., use x = 63 to get really large values ( 1, 061, 109, 567) in A.
Rumor: Bitwise operations can be useful.
Validity: True.
First, they are fast. Second, many useful tricks can be done using just a few bitwise operations.
As an easy example, x is a power of 2 if and only if (x & (x‐1) == 0). (Why? Think how does the bit pattern of a power of 2 look like.)
Note that x=x & (x‐1) clears the least significant set bit. By repeatedly doing this operation (until we get zero) we can easily count the
number of ones in the binary representation of x.
If you are interested in many more such tricks, download the free second chapter of the book Hacker’s Delight and read The Aggregate
Magic Algorithms.
One important trick: unsigned ints can be used to encode subsets of {0, 1,…, 31} in a straightforward way – the ith bit of a variable
will be one if and only if the represented set contains the number i. For example, the number 18 (binary 10010 = 24 +21) represents
the set {1, 4}.
When manipulating the sets, bitwise “and” corresponds to their intersection, bitwise “or” gives their union.
In C++, we may explicitly set the ith bit of x using the command x |= (1<<i), clear it using x &= ~(1<<i) and check whether it is set
using ((x & (1<<i)) != 0). Note that bitset and vector<bool> offer a similar functionality with arbitrarily large sets.
This trick can be used when your program has to compute the answer for all subsets of a given set of things. This concept is quite often
used in SRM problems. We won’t go into more details here, the best way of getting it right is looking at an actual implementation (try
looking at the best solutions for the problems below) and then trying to solve a few such problems on your own.
BorelSets (a simple exercise in set manipulation, generate sets until no new sets appear)
TableSeating
CompanyMessages
ChessMatch (for each subset of your players find the best assignment)
RevolvingDoors (encode your position and the states of all the doors into one integer)
Rumor: Real numbers are represented using a floating point representation.
Validity: True.
The most common way to represent “real” numbers in computers is the floating point representation defined by the IEEE Standard 754.
We will give a brief overview of this representation.
Basically, the words “floating point” mean that the position of the decimal (or more exactly, binary) point is not fixed. This will allow us
to store a large range of numbers than fixed point formats allow.
The numbers will be represented in scientific notation, using a normalized number and an exponent. For example, in base 10 the
number 123.456 could be represented as 1.23456 x 102. As a shorthand, we sometimes use the letter E to denote the phrase “times 10
to the power of”. E.g., the previous expression can be rewritten as 1.23456e2.
Of course, in computers we use binary numbers, thus the number 5.125 (binary 101.001) will be represented as 1.01001 x 22, and the
number 0.125 (binary 0.001) will be represented as 1 x 23.
Note that any (nonzero) real number x can be written in the form ( 1)s x m x 2e, where s ? {0, 1} represents the sign, m ? [1, 2) is
the normalized number and e is the (integer) exponent. This is the general form we are going to use to store real numbers.
What exactly do we need to store? The base is fixed, so the three things to store are the sign bit s, the normalized number (known as
the mantissa) m and the exponent e.
The IEEE Standard 754 defines four types of precision when storing floating point numbers. The two most commonly used are single
and double precision. In most programming languages these are also the names of corresponding data types. You may encounter other
data types (such as float) that are platform dependent and usually map to one of these types. If not sure, stick to these two types.
Single precision floating point numbers use 32 bits (4 bytes) of storage, double precision numbers use 64 bits (8 bytes). These bits are
used as shown in Table 3:
Table 3: Organization of memory in singles and doubles.
single
1 8 23
precision
double
1 11 52
precision
(The bits are given in order. I.e., the sign bit is the most significant bit, 8 or 11 exponent bits and then 23 or 52 mantissa bits follow.)
The sign bit
The sign bit is as simple as it gets. 0 denotes a positive number; 1 denotes a negative number. Inverting this bit changes the sign of
the number.
The exponent
The exponent field needs to represent both positive and negative exponents. To be able to do this, a bias is added to the actual
exponent e. This bias is 127 for single precision and 1023 for double precision. The result is stored as an unsigned integer. (E.g., if e =
– 13 and we use single precision, the actual value stored in memory will be 13 + 127 = 114.)
This would imply that the range of available exponents is 127 to 128 for single and 1023 to 1024 for double precision. This is almost
true. For reasons discussed later, both boundaries are reserved for special numbers. The actual range is then 126 to 127, and 1022 to
1023, respectively.
The mantissa
The mantissa represents the precision bits of the number. If we write the number in binary, these will be the first few digits, regardless
of the position of the binary point. (Note that the position of the binary point is specified by the exponent.)
The fact that we use base 2 allows us to do a simple optimization: We know that for any (nonzero) number the first digit is surely 1.
Thus we don’t have to store this digit. As a result, a bbit mantissa can actually store the b + 1 most significant bits of a number.
Representation of Integers and Reals: Section 2
By misof– TopCoder Member
Rumor: Floating point variables can store not only numbers but also some strange values.
Validity: True.
As stated in the previous answer, the standard reserves both the smallest and the largest possible value of the exponent to store
special numbers. (Note that in memory these values of the exponent are stored as “all zeroes” and “all ones”, respectively.)
Zero
When talking about the signmantissaexponent representation we noted that any nonzero number can be represented in this way.
Zero is not directly representable in this way. To represent zero we will use a special value denoted with both the exponent field and the
mantissa containing all zeroes. Note that 0 and +0 are distinct values, though they both compare as equal.
It is worth noting that if memset() is used to fill an array of floating point variables with zero bytes, the value of the stored numbers will
be zero. Also, global variables in C++ are initialized to a zero bit pattern, thus global floating point variables will be initialized to zero.
Also, note that negative zero is sometimes printed as “0″ or “0.0″. In some programming challenges (with inexperienced
problemsetters) this may cause your otherwise correct solution to fail.
There are quite a few subtle pitfalls concerning the negative zero. For example, the expressions “0.0 ‐ x” and “‐x” are not equivalent –
if x = 0.0, the value of the first expression is 0.0, the second one evaluates to 0.0.
My favorite quote on this topic: Negative zeros can “create the opportunity for an educational experience” when they are printed as
they are often printed as “0″ or “0.0″ (the “educational experience” is the time and effort that you spend learning why you’re getting
these strange values).
Infinities
The values +infinity and infinity correspond to an exponent of all ones and a mantissa of all zeroes. The sign bit distinguishes between
negative infinity and positive infinity. Being able to denote infinity as a specific value is useful because it allows operations to continue
past overflow situations.
Not a Number
The value NaN (Not a Number) is used to represent a value that does not represent a real number. NaNs are represented by a bit
pattern with an exponent of all ones and a nonzero mantissa. There are two categories of NaN: QNaN (Quiet NaN) and SNaN
(Signaling NaN).
A QNaN is a NaN with the most significant bit of the mantissa set. QNaNs propagate freely through most arithmetic operations. These
values pop out of an operation when the result is not mathematically defined. (For example, 3*sqrt(‐1.0) is a QNaN.)
An SNaN is a NaN with the most significant bit of the mantissa clear. It is used to signal an exception when used in operations. SNaNs
can be handy to assign to uninitialized variables to trap premature usage.
If a return value is a QNaN, it means that it is impossible to determine the result of the operation, a SNaN means that the operation is
invalid.
Subnormal numbers
We still didn’t use the case when the exponent is all zeroes and the mantissa is nonzero. We will use these values to store numbers
very close to zero.
These numbers are called subnormal, as they are smaller than the normally representable values. Here we don’t assume we have a
leading 1 before the binary point. If the sign bit is s, the exponent is all zeroes and the mantissa is m, the value of the stored number is
( 1)s x 0.m x 2q, where q is 126 for single and 1022 for double precision.
(Note that zero is just a special case of a subnormal number. Still, we wanted to present it separately.)
Summary of all possible values
In the following table, b is the bias used when storing the exponent, i.e., 127 for single and 1023 for double precision.
00…01 to
0 00…00 0.m x 2b+1
11…11
00…01 to
0 anything 1.m x 2eb
11…10
00…01 to
0 11…11 SNaN
01…11
10…00 to
0 11…11 QNaN
11…11
00…01 to
1 00…00 0.m x 2b+1
11…11
00…01 to
1 anything 1.m x 2eb
11…10
00…01 to
1 11…11 SNaN
01…11
10…00 to
1 11…11 QNaN
11.11
Operations with all the special numbers
All operations with the special numbers presented above are welldefined. This means that your program won’t crash just because one
of the computed values exceeded the representable range. Still, this is usually an unwanted situation and if it may occur, you should
check it in your program and handle the cases when it occurs.
The operations are defined in the probably most intuitive way. Any operation with a NaN yields a NaN as a result. Some other
operations are presented in the table below. (In the table, r is a positive representable number, ? is Infinity, ? is normal floating point
division.) A complete list can be found in the standard or in your compiler’s documentation. Note that even comparison operators are
defined for these values. This topic exceeds the scope of this article, if interested, browse through the references presented at the end
of the article.
operation result
0 ? ?? 0
?r ? ?? 0
(1)s? x
(1)st?
(1)t?
? + ? ?
?r ? 0 ??
0 ? 0 NaN
? – ? NaN
?? ? ?? NaN
?? x 0 NaN
Rumor: Floating point numbers can be compared by comparing the bit patterns in memory.
Validity: True.
Note that we have to handle sign comparison separately. If one of the numbers is negative and the other is positive, the result is clear.
If both numbers are negative, we may compare them by flipping their signs, comparing and returning the opposite answer. From now
on consider nonnegative numbers only.
When comparing the two bit patterns, the first few bits form the exponent. The larger the exponent is, the further is the bit pattern in
lexicographic order. Similarly, patterns with the same exponent are compared according to their mantissa.
Another way of looking at the same thing: when comparing two nonnegative real numbers stored in the form described above, the
result of the comparison is always the same as when comparing integers with the same bit pattern. (Note that this makes the
comparison pretty fast.)
Rumor: Comparing floating point numbers for equality is usually a bad idea.
Validity: True.
Consider the following code:
for (double r=0.0; r!=1.0; r+=0.1) printf("*");
How many stars is it going to print? Ten? Run it and be surprised. The code just keeps on printing the stars until we break it.
Where’s the problem? As we already know, doubles are not infinitely precise. The problem we encountered here is the following: In
binary, the representation of 0.1 is not finite (as it is in base 10). Decimal 0.1 is equivalent to binary 0.0(0011), where the part in the
parentheses is repeated forever. When 0.1 is stored in a double variable, it gets rounded to the closest representable value. Thus if we
add it 10 times the result is not exactly equal to one.
The most common advice is to use some tolerance (usually denoted ?) when comparing two doubles. E.g., you may sometimes hear the
following hint: consider the doubles a and b equal, if fabs(a‐b)<1e‐7. Note that while this is an improvement, it is not the best possible
way. We will show a better way later on.
Rumor: Floating point numbers are not exact, they are rounded.
Validity: Partially true.
Yes, if a number can’t be represented exactly, it has to be rounded. But sometimes an even more important fact is that lots of
important numbers (like zero, the powers of two, etc.) can be stored exactly. And it gets even better. Note that the mantissa of doubles
contains more than 32 bits. Thus all the binary digits of an int fit into the mantissa and the stored value is exact.
This can still be improved. If we note that ? when comparing floating point numbers.
Validity: False.
Often if you visit the Round Tables after a SRM that involved a floating point task you can see people posting messages like “after I
changed the precision from 1e‐12 to 1e‐7 it passed all systests in the practice room”
Examples of such discussions: here, here, here, here and here. (They are worth reading, it is always less painful to learn on the
mistakes of other people made than to learn on your own mistakes.)
We will start our answer by presenting another simple example.
for (double r=0.0; r<1e22; r+=1.0) printf(".");
How many dots will this program print? This time it’s clear, isn’t it? The terminating condition doesn’t use equality testing. The cycle has
to stop after 1022 iterations. Or… has it?
Bad luck, this is again an infinite cycle. Why is it so? Because when the value of r becomes large, the precision of the variable isn’t large
enough to store all decimal digits of r. The last ones become lost. And when we add 1 to such a large number, the result is simply
rounded back to the original number.
Exercise: Try to estimate the largest value of r our cycle will reach. Verify your answer. If your estimate was wrong, find out why.
After making this observation, we will show why the expression fabs(a‐b)<epsilon (with a fixed value of epsilon, usually recommended
between 1e‐7 and 1e‐9) is not ideal for comparing doubles.
Consider the values 123456123456.1234588623046875 and 123456123456.1234741210937500. There’s nothing that special about
them. These are just two values that can be stored in a double without rounding. Their difference is approximately 2e‐5.
Now take a look at the bit patterns of these two values:
first: 01000010 00111100 10111110 10001110 11110010 01000000 00011111 10011011
second: 01000010 00111100 10111110 10001110 11110010 01000000 00011111 10011100
Yes, right. These are two consecutive values that can be stored in a double. Almost any rounding error can change one of them onto the
other one (or even further). And still, they are quite far apart, thus our original test for “equality” fails.
What we really want is to tolerate small precision errors. As we already saw, doubles are able to store approximately 15 most significant
decimal digits. By accumulating precision errors that arise due to rounding, the last few of these digits may become corrupt. But how
exactly shall we implement tolerating such errors?
We won’t use a constant value of ?, but a value relative to the magnitude of the compared numbers. More precisely, if x is a double,
then x*1e‐10 is a number that’s 10 degrees of magnitude smaller than x. Its most significant digit corresponds to x‘s eleventh most
significant digit. This makes it a perfect ? for our needs.
In other words, a better way to compare doubles a and b for “equality” is to check whether a lies between b*(1‐1e‐10) and b*(1+1e‐10).
(Be careful, if b is negative, the first of these two numbers is larger!)
See any problems with doing the comparison this way? Try comparing 1e‐1072 and ‐1e‐1072. Both numbers are almost equal to zero
and to each other, but our test fails to handle this properly. This is why we have to use both the first test (known as testing for an
absolute error) and the second test (known as testing for a relative error).
This is the way TC uses to check whether your return value is correct. Now you know why.
There are even better comparison functions (see one of the references), but it is important to know that in practice you can often get
away with using only the absolute error test. Why? Because the numbers involved in computation come from a limited range. For
example, if the largest number you will ever compare is 9947, you know that a double will be able to store another 11 digits after the
decimal point correctly. Thus if we use epsilon=1e‐8 when doing the absolute error test, we allow the last three significant digits to
become corrupt.
The advantage this approach gives you is clear: checking for an absolute error is much simpler than the advanced tests presented
above.
Elections (a Div2 easy with a success rate of only 57.58%)
Archimedes
SortEstimate (the binary search is quite tricky to get right if you don’t understand precision issues)
PerforatedSheet (beware, huge rounding errors possible)
WatchTower
PackingShapes
Rumor: Computations using floating point variables are as exact as possible.
Validity: True.
Most of the standards require this. To be even more exact: For any arithmetical operation the returned value has to be that
representable value that’s closest to the exact result. Moreover, in C++ the default rounding mode says that if two values are tied for
being the closest, the one that’s more even (i.e., its least significant bit of the mantissa is 0) is returned. (Other standards may have
different rules for this tie breaking.)
As a useful example, note that if an integer n is a square (i.e., n = k2 for some integer k), then sqrt(double(n)) will return the exact
value k. And as we know that k can be stored in a variable of the same type as n, the code int k = int(sqrt(double(n))) is safe, there
will be no rounding errors.
Rumor: If I do the same computation twice, the results I get will be equal.
Validity: Partially true.
Wait, only partially true? Doesn’t this contradict the previous answer? Well, it doesn’t.
In C++ this rumor isn’t always true. The problem is that according to the standard a C++ compiler can sometimes do the internal
calculations using a larger data type. And indeed, g++ sometimes internally uses long doubles instead of doubles to achieve larger
precision. The value stored is only typecast to double when necessary. If the compiler decides that in one instance of your computation
long doubles will be used and in the other just doubles are used internally, the different roundings will influence the results and thus
the final results may differ.
This is one of THE bugs that are almost impossible to find and also one of the most confusing ones. Imagine that you add debug
outputs after each step of the computations. What you unintentionally cause is that after each step each of the intermediate results is
cast to double and output. In other words, you just pushed the compiler to only use doubles internally and suddenly everything works.
Needless to say, after you remove the debug outputs, the program will start to misbehave again.
A workaround is to write your code using long doubles only.
Sadly, this only cures one of the possible problems. The other is that when optimizing your code the compiler is allowed to rearrange
the order in which operations are computed. On different occasions, it may rewrite two identical pieces of C++ code into two different
sets of instructions. And all the precision problems are back.
As an example, the expression x + y z may once be evaluated as x + (y – z) and the other time as (x + y) – z. Try substituting the
values x = 1.0 and y = z = 1030.
Thus even if you have two identical pieces of code, you can’t be sure that they will produce exactly the same result. If you want this
guarantee, wrap the code into a function and call the same function on both occasions.
Further reading
Comparing floating point numbers (a detailed article by Bruce Dawson)
Floatingpoint representation
IEEE Standard 754
Integer Types In C and C++ (an article by Jack Klein)
Java FloatingPoint Number Intricacies (an article by Thomas Wang)
Lecture notes on IEEE754 (by William Kahan)
Lots of referrences about IEEE754
Revision of IEEE754 (note the definition of the operators min and max)
What Every Computer Scientist Should Know About FloatingPoint Arithmetic (a pretty long article by David Goldberg)
Binary Search
By lovro– TopCoder Member
Discuss this article in the forums
Binary search is one of the fundamental algorithms in computer science. In order to explore it, we’ll first build up a theoretical
backbone, then use that to implement the algorithm properly and avoid those nasty offbyone errors everyone’s been talking about.
Finding a value in a sorted sequence
In its simplest form, binary search is used to quickly find a value in a sorted sequence (consider a sequence an ordinary array for now).
We’ll call the sought value the target value for clarity. Binary search maintains a contiguous subsequence of the starting sequence
where the target value is surely located. This is called the search space. The search space is initially the entire sequence. At each step,
the algorithm compares the median value in the search space to the target value. Based on the comparison and because the sequence
is sorted, it can then eliminate half of the search space. By doing this repeatedly, it will eventually be left with a search space consisting
of a single element, the target value.
For example, consider the following sequence of integers sorted in ascending order and say we are looking for the number 55:
0 5 13 19 22 41 55 68 72 81 98
We are interested in the location of the target value in the sequence so we will represent the search space as indices into the sequence.
Initially, the search space contains indices 1 through 11. Since the search space is really an interval, it suffices to store just two
numbers, the low and high indices. As described above, we now choose the median value, which is the value at index 6 (the midpoint
between 1 and 11): this value is 41 and it is smaller than the target value. From this we conclude not only that the element at index 6
is not the target value, but also that no element at indices between 1 and 5 can be the target value, because all elements at these
indices are smaller than 41, which is smaller than the target value. This brings the search space down to indices 7 through 11:
55 68 72 81 98
Proceeding in a similar fashion, we chop off the second half of the search space and are left with:
55 68
Depending on how we choose the median of an even number of elements we will either find 55 in the next step or chop off 68 to get a
search space of only one element. Either way, we conclude that the index where the target value is located is 7.
If the target value was not present in the sequence, binary search would empty the search space entirely. This condition is easy to
check and handle. Here is some code to go with the description:
binary_search(A, target):
lo = 1, hi = size(A)
while lo <= hi:
mid = lo + (hi‐lo)/2
if A[mid] == target:
return mid
else if A[mid] < target:
lo = mid+1
else:
hi = mid‐1
// target was not found
Complexity
Since each comparison binary search uses halves the search space, we can assert and easily prove that binary search will never use
more than (in bigoh notation) O(log N) comparisons to find the target value.
The logarithm is an awfully slowly growing function. In case you’re not aware of just how efficient binary search is, consider looking up
a name in a phone book containing a million names. Binary search lets you systematically find any given name using at most 21
comparisons. If you could manage a list containing all the people in the world sorted by name, you could find any person in less than 35
steps. This may not seem feasible or useful at the moment, but we’ll soon fix that.
Note that this assumes that we have random access to the sequence. Trying to use binary search on a container such as a linked list
makes little sense and it is better use a plain linear search instead.
Binary search in standard libraries
C++’s Standard Template Library implements binary search in algorithms lower_bound, upper_bound, binary_search and equal_range,
depending exactly on what you need to do. Java has a builtin Arrays.binary_search method for arrays and the .NET Framework has
Array.BinarySearch.
You’re best off using library functions whenever possible, since, as you’ll see, implementing binary search on your own can be tricky.
Beyond arrays: the discrete binary search
This is where we start to abstract binary search. A sequence (array) is really just a function which associates integers (indices) with the
corresponding values. However, there is no reason to restrict our usage of binary search to tangible sequences. In fact, we can use the
same algorithm described above on any monotonic function f whose domain is the set of integers. The only difference is that we replace
an array lookup with a function evaluation: we are now looking for some x such that f(x) is equal to the target value. The search space
is now more formally a subinterval of the domain of the function, while the target value is an element of the codomain. The power of
binary search begins to show now: not only do we need at most O(log N) comparisons to find the target value, but we also do not need
to evaluate the function more than that many times. Additionally, in this case we aren’t restricted by practical quantities such as
available memory, as was the case with arrays.
Taking it further: the main theorem
When you encounter a problem which you think could be solved by applying binary search, you need some way of proving it will work. I
will now present another level of abstraction which will allow us to solve more problems, make proving binary search solutions very
easy and also help implement them. This part is a tad formal, but don’t get discouraged, it’s not that bad.
Consider a predicate p defined over some ordered set S (the search space). The search space consists of candidate solutions to the
problem. In this article, a predicate is a function which returns a boolean value, true or false (we’ll also use yes and no as boolean
values). We use the predicate to verify if a candidate solution is legal (does not violate some constraint) according to the definition of
the problem.
What we can call the main theorem states that binary search can be used if and only if for all x in S, p(x) implies p(y) for all y
> x. This property is what we use when we discard the second half of the search space. It is equivalent to saying that ?p(x) implies ?
p(y) for all y < x (the symbol ? denotes the logical not operator), which is what we use when we discard the first half of the search
space. The theorem can easily be proven, although I’ll omit the proof here to reduce clutter.
Behind the cryptic mathematics I am really stating that if you had a yes or no question (the predicate), getting a yes answer for some
potential solution x means that you’d also get a yes answer for any element after x. Similarly, if you got a no answer, you’d get a no
answer for any element before x. As a consequence, if you were to ask the question for each element in the search space (in order),
you would get a series of no answers followed by a series of yes answers.
Careful readers may note that binary search can also be used when a predicate yields a series of yes answers followed by a series of no
answers. This is true and complementing that predicate will satisfy the original condition. For simplicity we’ll deal only with predicates
described in the theorem.
If the condition in the main theorem is satisfied, we can use binary search to find the smallest legal solution, i.e. the smallest x for
which p(x) is true. The first part of devising a solution based on binary search is designing a predicate which can be evaluated and for
which it makes sense to use binary search: we need to choose what the algorithm should find. We can have it find either the first x for
which p(x) is true or the last x for which p(x) is false. The difference between the two is only slight, as you will see, but it is necessary
to settle on one. For starters, let us seek the first yes answer (first option).
The second part is proving that binary search can be applied to the predicate. This is where we use the main theorem, verifying that the
conditions laid out in the theorem are satisfied. The proof doesn’t need to be overly mathematical, you just need to convince yourself
that p(x) implies p(y) for all y > x or that ?p(x) implies ?p(y) for all y < x. This can often be done by applying common sense in a
sentence or two.
When the domain of the predicate are the integers, it suffices to prove that p(x) implies p(x+1) or that ?p(x) implies ?p(x1), the rest
then follows by induction.
These two parts are most often interleaved: when we think a problem can be solved by binary search, we aim to design the predicate
so that it satisfies the condition in the main theorem.
One might wonder why we choose to use this abstraction rather than the simplerlooking algorithm we’ve used so far. This is because
many problems can’t be modeled as searching for a particular value, but it’s possible to define and evaluate a predicate such as “Is
there an assignment which costs x or less?”, when we’re looking for some sort of assignment with the lowest cost. For example, the
usual traveling salesman problem (TSP) looks for the cheapest roundtrip which visits every city exactly once. Here, the target value is
not defined as such, but we can define a predicate “Is there a roundtrip which costs x or less?” and then apply binary search to find
the smallest x which satisfies the predicate. This is called reducing the original problem to a decision (yes/no) problem. Unfortunately,
we know of no way of efficiently evaluating this particular predicate and so the TSP problem isn’t easily solved by binary search, but
many optimization problems are.
Let us now convert the simple binary search on sorted arrays described in the introduction to this abstract definition. First, let’s
rephrase the problem as: “Given an array A and a target value, return the index of the first element in A equal to or greater than the
target value.” Incidentally, this is more or less how lower_bound behaves in C++.
We want to find the index of the target value, thus any index into the array is a candidate solution. The search space S is the set of all
candidate solutions, thus an interval containing all indices. Consider the predicate “Is A[x] greater than or equal to the target value?”. If
we were to find the first x for which the predicate says yes, we’d get exactly what decided we were looking for in the previous
paragraph.
The condition in the main theorem is satisfied because the array is sorted in ascending order: if A[x] is greater than or equal to the
target value, all elements after it are surely also greater than or equal to the target value.
If we take the sample sequence from before:
0 5 13 19 22 41 55 68 72 81 98
With the search space (indices):
1 2 3 4 5 6 7 8 9 10 11
And apply our predicate (with a target value of 55) to it we get:
This is a series of no answers followed by a series of yes answers, as we were expecting. Notice how index 7 (where the target value is
located) is the first for which the predicate yields yes, so this is what our binary search will find.
Implementing the discrete algorithm
One important thing to remember before beginning to code is to settle on what the two numbers you maintain (lower and upper bound)
mean. A likely answer is a closed interval which surely contains the first x for which p(x) is true. All of your code should then be
directed at maintaining this invariant: it tells you how to properly move the bounds, which is where a bug can easily find its way in your
code, if you’re not careful.
Another thing you need to be careful with is how high to set the bounds. By “high” I really mean “wide” since there are two bounds to
worry about. Every so often it happens that a coder concludes during coding that the bounds he or she set are wide enough, only to
find a counterexample during intermission (when it’s too late). Unfortunately, little helpful advice can be given here other than to
always double and triplecheck your bounds! Also, since execution time increases logarithmically with the bounds, you can always set
them higher, as long as it doesn’t break the evaluation of the predicate. Keep your eye out for overflow errors all around, especially in
calculating the median.
Now we finally get to the code which implements binary search as described in this and the previous section:
binary_search(lo, hi, p):
while lo < hi:
mid = lo + (hi‐lo)/2
if p(mid) == true:
hi = mid
else:
lo = mid+1
if p(lo) == false:
complain // p(x) is false for all x in S!
return lo // lo is the least x for which p(x) is true
The two crucial lines are hi = mid and lo = mid+1. When p(mid) is true, we can discard the second half of the search space, since the
predicate is true for all elements in it (by the main theorem). However, we can not discard mid itself, since it may well be the first
element for which p is true. This is why moving the upper bound to mid is as aggressive as we can do without introducing bugs.
In a similar vein, if p(mid) is false, we can discard the first half of the search space, but this time including mid. p(mid) is false so we
don’t need it in our search space. This effectively means we can move the lower bound to mid+1.
If we wanted to find the last x for which p(x) is false, we would devise (using a similar rationale as above) something like:
// warning: there is a nasty bug in this snippet!
binary_search(lo, hi, p):
while lo < hi:
mid = lo + (hi‐lo)/2 // note: division truncates
if p(mid) == true:
hi = mid‐1
else:
lo = mid
if p(lo) == true:
complain // p(x) is true for all x in S!
return lo // lo is the greatest x for which p(x) is false
You can verify that this satisfies our condition that the element we’re looking for always be present in the interval (lo, hi). However,
there is another problem. Consider what happens when you run this code on some search space for which the predicate gives:
no yes
The code will get stuck in a loop. It will always select the first element as mid, but then will not move the lower bound because it wants
to keep the no in its search space. The solution is to change mid = lo + (hilo)/2 to mid = lo + (hilo+1)/2, i.e. so that it rounds up
instead of down. There are other ways of getting around the problem, but this one is possibly the cleanest. Just remember to always
test your code on a twoelement set where the predicate is false for the first element and true for the second.
You may also wonder as to why mid is calculated using mid = lo + (hilo)/2 instead of the usual mid = (lo+hi)/2. This is to avoid
another potential rounding bug: in the first case, we want the division to always round down, towards the lower bound. But division
truncates, so when lo+hi would be negative, it would start rounding towards the higher bound. Coding the calculation this way ensures
that the number divided is always positive and hence always rounds as we want it to. Although the bug doesn’t surface when the search
space consists only of positive integers or real numbers, I’ve decided to code it this way throughout the article for consistency.
Real numbers
Binary search can also be used on monotonic functions whose domain is the set of real numbers. Implementing binary search on reals
is usually easier than on integers, because you don’t need to watch out for how to move bounds:
binary_search(lo, hi, p):
while we choose not to terminate:
mid = lo + (hi‐lo)/2
if p(mid) == true:
hi = mid
else:
lo = mid
return lo // lo is close to the border between no and yes
Since the set of real numbers is dense, it should be clear that we usually won’t be able to find the exact target value. However, we can
quickly find some x such that f(x) is within some tolerance of the border between no and yes. We have two ways of deciding when to
terminate: terminate when the search space gets smaller than some predetermined bound (say 1012) or do a fixed number of
iterations. On TopCoder, your best bet is to just use a few hundred iterations, this will give you the best possible precision without too
much thinking. 100 iterations will reduce the search space to approximately 1030 of its initial size, which should be enough for most (if
not all) problems.
If you need to do as few iterations as possible, you can terminate when the interval gets small, but try to do a relative comparison of
the bounds, not just an absolute one. The reason for this is that doubles can never give you more than 15 decimal digits of precision so
if the search space contains large numbers (say on the order of billions), you can never get an absolute difference of less than 107.
Example
At this point I will show how all this talk can be used to solve a TopCoder problem. For this I have chosen a moderately difficult
problem, FairWorkload, which was the division 1 level 2 problem in SRM 169.
In the problem, a number of workers need to examine a number of filing cabinets. The cabinets are not all of the same size and we are
told for each cabinet how many folders it contains. We are asked to find an assignment such that each worker gets a sequential series
of cabinets to go through and that it minimizes the maximum amount of folders that a worker would have to look through.
After getting familiar with the problem, a touch of creativity is required. Imagine that we have an unlimited number of workers at our
disposal. The crucial observation is that, for some number MAX, we can calculate the minimum number of workers needed so that each
worker has to examine no more than MAX folders (if this is possible). Let’s see how we’d do that. Some worker needs to examine the
first cabinet so we assign any worker to it. But, since the cabinets must be assigned in sequential order (a worker cannot examine
cabinets 1 and 3 without examining 2 as well), it’s always optimal to assign him to the second cabinet as well, if this does not take him
over the limit we introduced (MAX). If it would take him over the limit, we conclude that his work is done and assign a new worker to
the second cabinet. We proceed in a similar manner until all the cabinets have been assigned and assert that we’ve used the minimum
number of workers possible, with the artificial limit we introduced. Note here that the number of workers is inversely proportional to
MAX: the higher we set our limit, the fewer workers we will need.
Now, if you go back and carefully examine what we’re asked for in the problem statement, you can see that we are really asked for the
smallest MAX such that the number of workers required is less than or equal to the number of workers available. With that in mind,
we’re almost done, we just need to connect the dots and see how all of this fits in the frame we’ve laid out for solving problems using
binary search.
With the problem rephrased to fit our needs better, we can now examine the predicate Can the workload be spread so that each worker
has to examine no more than x folders, with the limited number of workers available? We can use the described greedy algorithm to
efficiently evaluate this predicate for any x. This concludes the first part of building a binary search solution, we now just have to prove
that the condition in the main theorem is satisfied. But observe that increasing x actually relaxes the limit on the maximum workload,
so we can only need the same number of workers or fewer, not more. Thus, if the predicate says yes for some x, it will also say yes for
all larger x.
To wrap it up, here’s an STLdriven snippet which solves the problem:
int getMostWork( vector folders, int workers ) {
int n = folders.size();
int lo = *max_element( folders.begin(), folders.end() );
int hi = accumulate( folders.begin(), folders.end(), 0 );
while ( lo < hi ) {
int x = lo + (hi‐lo)/2;
int required = 1, current_load = 0;
for ( int i=0; i<n; ++i ) {
if ( current_load + folders[i] <= x ) {
// the current worker can handle it
current_load += folders[i];
}
else {
// assign next worker
++required;
current_load = folders[i];
}
}
if ( required <= workers )
hi = x;
else
lo = x+1;
}
return lo;
}
Note the carefully chosen lower and upper bounds: you could replace the upper bound with any sufficiently large integer, but the lower
bound must not to be less than the largest cabinet to avoid the situation where a single cabinet would be too large for any worker, a
case which would not be correctly handled by the predicate. An alternative would be to set the lower bound to zero, then handle too
small x’s as a special case in the predicate.
To verify that the solution doesn’t lock up, I used a small no/yes example with folders={1,1} and workers=1.
The overall complexity of the solution is O(n log SIZE), where SIZE is the size of the search space. This is very fast.
As you see, we used a greedy algorithm to evaluate the predicate. In other problems, evaluating the predicate can come down to
anything from a simple math expression to finding a maximum cardinality matching in a bipartite graph.
Conclusion
If you’ve gotten this far without giving up, you should be ready to solve anything that can be solved with binary search. Try to keep a
few things in mind:
Design a predicate which can be efficiently evaluated and so that binary search can be applied
Decide on what you’re looking for and code so that the search space always contains that (if it exists)
If the search space consists only of integers, test your algorithm on a twoelement set to be sure it doesn’t lock up
Verify that the lower and upper bounds are not overly constrained: it’s usually better to relax them as long as it doesn’t break the
predicate
Here are a few problems that can be solved using binary search:
Simple
AutoLoan – SRM 258
SortEstimate SRM 230
Moderate
UnionOfIntervals – SRM 277
Mortgage – SRM 189
FairWorkload SRM 169
HairCuts – SRM 261
Harder
PackingShapes SRM 270
RemoteRover SRM 235
NegativePhotoresist SRM 210
WorldPeace SRM 204
UnitsMoving SRM 278
Parking – SRM 236
SquareFree SRM 190
Flags – SRM 147
A Bit of Fun: Fun with Bits
By bmerry– TopCoder Member
Discuss this article in the forums
Introduction
Most of the optimizations that go into TopCoder challenges are highlevel; that is, they affect the algorithm rather than the
implementation. However, one of the most useful and effective lowlevel optimizations is bit manipulation, or using the bits of an
integer to represent a set. Not only does it produce an orderofmagnitude improvement in both speed and size, it can often simplify
code at the same time.
I’ll start by briefly recapping the basics, before going on to cover more advanced techniques.
The basics
At the heart of bit manipulation are the bitwise operators & (and), | (or), ~ (not) and ^ (xor). The first three you should already be
familiar with in their boolean forms (&&, || and !). As a reminder, here are the truth tables:
0 0 1 0 0 0
0 1 1 0 1 1
1 0 0 0 1 1
1 1 0 1 1 0
The bitwise versions of the operations are the same, except that instead of interpreting their arguments as true or false, they operate
on each bit of the arguments. Thus, if A is 1010 and B is 1100, then
A & B = 1000
A | B = 1110
A ^ B = 0110
~A = 11110101 (the number of 1′s depends on the type of A).
The other two operators we will need are the shift operators a << b and a >> b. The former shifts all the bits in a to the left by b
positions; the latter does the same but shifts right. For nonnegative values (which are the only ones we’re interested in), the newly
exposed bits are filled with zeros. You can think of leftshifting by b as multiplication by 2b and rightshifting as integer division by 2b.
The most common use for shifting is to access a particular bit, for example, 1 << x is a binary number with bit x set and the others
clear (bits are almost always counted from the rightmost/leastsignificant bit, which is numbered 0).
In general, we will use an integer to represent a set on a domain of up to 32 values (or 64, using a 64bit integer), with a 1 bit
representing a member that is present and a 0 bit one that is absent. Then the following operations are quite straightforward, where
ALL_BITS is a number with 1′s for all bits corresponding to the elements of the domain:
Set union
A | B
Set intersection
A & B
Set subtraction
A & ~B
Set negation
ALL_BITS ^ A
Set bit
A |= 1 << bit
Clear bit
A &= ~(1 << bit)
Test bit
(A & 1 << bit) != 0
Extracting every last bit
In this section I’ll consider the problems of finding the highest and lowest 1 bit in a number. These are basic operations for splitting a
set into its elements.
Finding the lowest set bit turns out to be surprisingly easy, with the right combination of bitwise and arithmetic operators. Suppose we
wish to find the lowest set bit of x (which is known to be nonzero). If we subtract 1 from x then this bit is cleared, but all the other one
bits in x remain set. Thus, x & ~(x ‐ 1) consists of only the lowest set bit of x. However, this only tells us the bit value, not the index of
the bit.
If we want the index of the highest or lowest bit, the obvious approach is simply to loop through the bits (upwards or downwards) until
we find one that is set. At first glance this sounds slow, since it does not take advantage of the bitpacking at all. However, if all 2N
subsets of the Nelement domain are equally likely, then the loop will take only two iterations on average, and this is actually the
fastest method.
The 386 introduced CPU instructions for bit scanning: BSF (bit scan forward) and BSR (bit scan reverse). GCC exposes these
instructions through the builtin functions __builtin_ctz (count trailing zeros) and __builtin_clz (count leading zeros). These are the
most convenient way to find bit indices for C++ programmers in TopCoder. Be warned though: the return value is undefined for an
argument of zero.
Finally, there is a portable method that performs well in cases where the looping solution would require many iterations. Use each byte
of the 4 or 8byte integer to index a precomputed 256entry table that stores the index of the highest (lowest) set bit in that byte. The
highest (lowest) bit of the integer is then the maximum (minimum) of the table entries. This method is only mentioned for
completeness, and the performance gain is unlikely to justify its use in a TopCoder match.
Counting out the bits
One can easily check if a number is a power of 2: clear the lowest 1 bit (see above) and check if the result is 0. However, sometimes it
is necessary to know how many bits are set, and this is more difficult.
GCC has a function called __builtin_popcount which does precisely this. However, unlike __builtin_ctz, it does not translate into a
hardware instruction (at least on x86). Instead, it uses a tablebased method similar to the one described above for bit searches. It is
nevertheless quite efficient and also extremely convenient.
Users of other languages do not have this option (although they could reimplement it). If a number is expected to have very few 1
bits, an alternative is to repeatedly extract the lowest 1 bit and clear it.
All the subsets
A big advantage of bit manipulation is that it is trivial to iterate over all the subsets of an Nelement set: every Nbit value represents
some subset. Even better, if A is a subset of B then the number representing A is less than that representing B, which is convenient for
some dynamic programming solutions.
It is also possible to iterate over all the subsets of a particular subset (represented by a bit pattern), provided that you don’t mind
visiting them in reverse order (if this is problematic, put them in a list as they’re generated, then walk the list backwards). The trick is
similar to that for finding the lowest bit in a number. If we subtract 1 from a subset, then the lowest set element is cleared, and every
lower element is set. However, we only want to set those lower elements that are in the superset. So the iteration step is just i = (i ‐
1) & superset.
Even a bit wrong scores zero
There are a few mistakes that are very easy to make when performing bit manipulations. Watch out for them in your code.
1. When executing shift instructions for a << b, the x86 architecture uses only the bottom 5 bits of b (6 for 64bit integers). This
means that shifting left (or right) by 32 does nothing, rather than clearing all the bits. This behaviour is also specified by the Java
and C# language standards; C99 says that shifting by at least the size of the value gives an undefined result. Historical trivia: the
8086 used the full shift register, and the change in behaviour was often used to detect newer processors.
2. The & and | operators have lower precedence than comparison operators. That means that x & 3 == 1 is interpreted as x & (3 ==
1), which is probably not what you want.
3. If you want to write completely portable C/C++ code, be sure to use unsigned types, particularly if you plan to use the topmost
bit. C99 says that shift operations on negative values are undefined. Java only has signed types: >> will signextend values (which
is probably not what you want), but the Javaspecific operator >>> will shift in zeros.
Cute tricks
There are a few other tricks that can be done with bit manipulation. They’re good for amazing your friends, but generally not worth the
effect to use in practice.
Reversing the bits in an integer
x = ((x & 0xaaaaaaaa) >> 1) | ((x & 0x55555555) << 1);
x = ((x & 0xcccccccc) >> 2) | ((x & 0x33333333) << 2);
x = ((x & 0xf0f0f0f0) >> 4) | ((x & 0x0f0f0f0f) << 4);
x = ((x & 0xff00ff00) >> 8) | ((x & 0x00ff00ff) << 8);
x = ((x & 0xffff0000) >> 16) | ((x & 0x0000ffff) << 16);
As an exercise, see if you can adapt this to count the number of bits in a word.
Iterate through all kelement subsets of {0, 1, … N1}
int s = (1 << k) ‐ 1;
while (!(s & 1 << N)){
// do stuff with s
int lo = s & ~(s ‐ 1); // lowest one bit
int lz = (s + lo) & ~s; // lowest zero bit above lo
s |= lz; // add lz to the set
s &= ~(lz ‐ 1); // reset bits below lz
s |= (lz / lo / 2) ‐ 1; // put back right number of bits at end
}
In C, the last line can be written as s |= (lz >> ffs(lo)) ‐ 1 to avoid the division.
Evaluate x ? y : ‐y, where x is 0 or 1
(‐x ^ y) + x
This works on a twoscomplement architecture (which is almost any machine you find today), where negation is done by inverting
all the bits then adding 1. Note that on i686 and above, the original expression can be evaluated just as efficiently (i.e., without
branches) due to the CMOVE (conditional move) instruction.
Sample problems
TCCC 2006, Round 1B Medium
For each city, keep a bitset of the neighbouring cities. Once the partbuilding factories have been chosen (recursively), ANDing
together these bitsets will give a bitset which describes the possible locations of the partassembly factories. If this bitset has k bits,
then there are kCm ways to allocate the partassembly factories.
TCO 2006, Round 1 Easy
The small number of nodes strongly suggests that this is done by considering all possible subsets. For every possible subset we
consider two possibilities: either the smallestnumbered node does not communicate at all, in which case we refer back to the subset
that excludes it, or it communicates with some node, in which case we refer back to the subset that excludes both of these nodes. The
resulting code is extremely short:
static int dp[1 << 18];
int SeparateConnections::howMany(vector <string> mat){
int N = mat.size();
int N2 = 1 << N;
dp[0] = 0;
for (int i = 1; i < N2; i++){
int bot = i & ~(i ‐ 1);
int use = __builtin_ctz(bot);
dp[i] = dp[i ^ bot];
for (int j = use + 1; j < N; j++)
if ((i & (1 << j)) && mat[use][j] == 'Y')
dp[i] = max(dp[i], dp[i ^ bot ^ (1 << j)] + 2);
}
return dp[N2 ‐ 1];
}
SRM 308, Division 1 Medium
The board contains 36 squares and the draughts are indistinguishable, so the possible positions can be encoded into 64bit integers.
The first step is to enumerate all the legal moves. Any legal move can be encoded using three bitfields: a before state, an after state
and a mask, which defines which parts of the before state are significant. The move can be made from the current state if (current &
mask) == before; if it is made, the new state is (current & ~mask) | after.
SRM 320, Division 1 Hard
The constraints tell us that there are at most 8 columns (if there are more, we can swap rows and columns), so it is feasible to consider
every possible way to lay out a row. Once we have this information, we can solve the remainder of the problem (refer to the match
editorial for details). We thus need a list of all nbit integers which do not have two adjacent 1 bits, and we also need to know how
many 1 bits there are in each such row. Here is my code for this:
for (int i = 0; i < (1 << n); i++){
if (i & (i << 1)) continue;
pg.push_back(i);
pgb.push_back(__builtin_popcount(i));
}
Range Minimum Query and Lowest Common Ancestor
By danielp– TopCoder Member
Discuss this article in the forums
Introduction
Notations
Range Minimum Query (RMQ)
Trivial algorithms for RMQ
A <O(N), O(sqrt(N))> solution
Sparse Table (ST) algorithm
Segment Trees
Lowest Common Ancestor (LCA)
A <O(N), O(sqrt(N))> solution
Another easy solution in <O(N logN, O(logN)>
Reduction from LCA to RMQ
From RMQ to LCA
An <O(N), O(1)> algorithm for the restricted RMQ
Conclusion
Introduction
The problem of finding the Lowest Common Ancestor (LCA) of a pair of nodes in a rooted tree has been studied more carefully in the
second part of the 20th century and now is fairly basic in algorithmic graph theory. This problem is interesting not only for the tricky
algorithms that can be used to solve it, but for its numerous applications in string processing and computational biology, for example,
where LCA is used with suffix trees or other treelike structures. Harel and Tarjan were the first to study this problem more attentively
and they showed that after linear preprocessing of the input tree LCA, queries can be answered in constant time. Their work has since
been extended, and this tutorial will present many interesting approaches that can be used in other kinds of problems as well.
Let’s consider a less abstract example of LCA: the tree of life. It’s a wellknown fact that the current habitants of Earth evolved from
other species. This evolving structure can be represented as a tree, in which nodes represent species, and the sons of some node
represent the directly evolved species. Now species with similar characteristics are divided into groups. By finding the LCA of some
nodes in this tree we can actually find the common parent of two species, and we can determine that the similar characteristics they
share are inherited from that parent.
Range Minimum Query (RMQ) is used on arrays to find the position of an element with the minimum value between two specified
indices. We will see later that the LCA problem can be reduced to a restricted version of an RMQ problem, in which consecutive array
elements differ by exactly 1.
However, RMQs are not only used with LCA. They have an important role in string preprocessing, where they are used with suffix arrays
(a new data structure that supports string searches almost as fast as suffix trees, but uses less memory and less coding effort).
In this tutorial we will first talk about RMQ. We will present many approaches that solve the problem — some slower but easier to code,
and others faster. In the second part we will talk about the strong relation between LCA and RMQ. First we will review two easy
approaches for LCA that don’t use RMQ; then show that the RMQ and LCA problems are equivalent; and, at the end, we’ll look at how
the RMQ problem can be reduced to its restricted version, as well as show a fast algorithm for this particular case.
Notations
Suppose that an algorithm has preprocessing time f(n) and query time g(n). The notation for the overall complexity for the algorithm
is <f(n), g(n)>.
We will note the position of the element with the minimum value in some array A between indices i and j with RMQA(i, j).
The furthest node from the root that is an ancestor of both u and v in some rooted tree T is LCAT(u, v).
Range Minimum Query(RMQ)
Given an array A[0, N1] find the position of the element with the minimum value between two given indices.
Trivial algorithms for RMQ
For every pair of indices (i, j) store the value of RMQA(i, j) in a table M[0, N1][0, N1]. Trivial computation will lead us to an
<O(N3), O(1)> complexity. However, by using an easy dynamic programming approach we can reduce the complexity to <O(N2),
O(1)>. The preprocessing function will look something like this:
void process1(int M[MAXN][MAXN], int A[MAXN], int N){
int i, j;
for (i =0; i < N; i++)
M[i][i] = i;
for (i = 0; i < N; i++)
for (j = i + 1; j < N; j++)
if (A[M[i][j ‐ 1]] < A[j])
M[i][j] = M[i][j ‐ 1];
else
M[i][j] = j;
}
This trivial algorithm is quite slow and uses O(N2) memory, so it won’t work for large cases.
An <O(N), O(sqrt(N))> solution
An interesting idea is to split the vector in sqrt(N) pieces. We will keep in a vector M[0, sqrt(N)1] the position for the minimum
value for each section. M can be easily preprocessed in O(N). Here is an example:
Now let’s see how can we compute RMQA(i, j). The idea is to get the overall minimum from the sqrt(N) sections that lie inside the
interval, and from the end and the beginning of the first and the last sections that intersect the bounds of the interval. To get
RMQA(2,7) in the above example we should compare A[2], A[M[1]], A[6] and A[7] and get the position of the minimum value. It’s
easy to see that this algorithm doesn’t make more than 3 * sqrt(N) operations per query.
The main advantages of this approach are that is to quick to code (a plus for TopCoderstyle competitions) and that you can adapt it to
the dynamic version of the problem (where you can change the elements of the array between queries).
Sparse Table (ST) algorithm
A better approach is to preprocess RMQ for sub arrays of length 2k using dynamic programming. We will keep an array M[0, N1][0,
logN] where M[i][j] is the index of the minimum value in the sub array starting at i having length 2j. Here is an example:
For computing M[i][j] we must search for the minimum value in the first and second half of the interval. It’s obvious that the small
pieces have 2j – 1 length, so the recurrence is:
The preprocessing function will look something like this:
void process2(int M[MAXN][LOGMAXN], int A[MAXN], int N){
int i, j;
//initialize M for the intervals with length 1
for (i = 0; i < N; i++)
M[i][0] = i;
//compute values from smaller to bigger intervals
for (j = 1; 1 << j <= N; j++)
for (i = 0; i + (1 << j) ‐ 1 < N; i++)
if (A[M[i][j ‐ 1]] < A[M[i + (1 << (j ‐ 1))][j ‐ 1]])
M[i][j] = M[i][j ‐ 1];
else
M[i][j] = M[i + (1 << (j ‐ 1))][j ‐ 1];
}
Once we have these values preprocessed, let’s show how we can use them to calculate RMQA(i, j). The idea is to select two blocks that
entirely cover the interval [i..j] and find the minimum between them. Let k = [log(j i + 1)]. For computing RMQA(i, j) we can use
the following formula:
So, the overall complexity of the algorithm is <O(N logN), O(1)>.
Segment trees
For solving the RMQ problem we can also use segment trees. A segment tree is a heaplike data structure that can be used for making
update/query operations upon array intervals in logarithmical time. We define the segment tree for the interval [i, j] in the following
recursive manner:
the first node will hold the information for the interval [i, j]
if i<j the left and right son will hold the information for the intervals [i, (i+j)/2] and [(i+j)/2+1, j]
Notice that the height of a segment tree for an interval with N elements is [logN] + 1. Here is how a segment tree for the interval [0,
9] would look like:
The segment tree has the same structure as a heap, so if we have a node numbered x that is not a leaf the left son of x is 2*x and the
right son 2*x+1.
For solving the RMQ problem using segment trees we should use an array M[1, 2 * 2[logN] + 1] where M[i] holds the minimum value
position in the interval assigned to node i. At the beginning all elements in M should be 1. The tree should be initialized with the
following function (b and e are the bounds of the current interval):
void initialize(intnode, int b, int e, int M[MAXIND], int A[MAXN], int N){
if (b == e)
M[node] = b;
else {
//compute the values in the left and right subtrees
initialize(2 * node, b, (b + e) / 2, M, A, N);
initialize(2 * node + 1, (b + e) / 2 + 1, e, M, A, N);
//search for the minimum value in the first and
//second half of the interval
if (A[M[2 * node]] <= A[M[2 * node + 1]])
M[node] = M[2 * node];
else
M[node] = M[2 * node + 1];
}
}
The function above reflects the way the tree is constructed. When calculating the minimum position for some interval we should look at
the values of the sons. You should call the function with node = 1, b = 0 and e = N1.
We can now start making queries. If we want to find the position of the minimum value in some interval [i, j] we should use the next
easy function:
int query(int node, int b, int e, int M[MAXIND], int A[MAXN], int i, int j){
int p1, p2;
//if the current interval doesn't intersect
//the query interval return ‐1
if (i > e || j < b)
return ‐1;
//if the current interval is included in
//the query interval return M[node]
if (b >= i && e <= j)
return M[node];
//compute the minimum position in the
//left and right part of the interval
p1 = query(2 * node, b, (b + e) / 2, M, A, i, j);
p2 = query(2 * node + 1, (b + e) / 2 + 1, e, M, A, i, j);
//return the position where the overall
//minimum is
if (p1 == ‐1)
return M[node] = p2;
if (p2 == ‐1)
return M[node] = p1;
if (A[p1] <= A[p2])
return M[node] = p1;
return M[node] = p2;
}
You should call this function with node = 1, b = 0 and e = N – 1, because the interval assigned to the first node is [0, N1].
It’s easy to see that any query is done in O(log N). Notice that we stop when we reach completely in/out intervals, so our path in the
tree should split only one time.
Using segment trees we get an <O(N), O(logN)> algorithm. Segment trees are very powerful, not only because they can be used for
RMQ. They are a very flexible data structure, can solve even the dynamic version of RMQ problem, and have numerous applications in
range searching problems.
Lowest Common Ancestor (LCA)
Given a rooted tree T and two nodes u and v, find the furthest node from the root that is an ancestor for both u and v. Here is an
example (the root of the tree will be node 1 for all examples in this editorial):
An <O(N), O(sqrt(N))> solution
Dividing our input into equalsized parts proves to be an interesting way to solve the RMQ problem. This method can be adapted for the
LCA problem as well. The idea is to split the tree in sqrt(H) parts, were H is the height of the tree. Thus, the first section will contain
the levels numbered from 0 to sqrt(H) – 1, the second will contain the levels numbered from sqrt(H) to 2 * sqrt(H) – 1, and so on.
Here is how the tree in the example should be divided:
Now, for each node, we should know the ancestor that is situated on the last level of the upper next section. We will preprocess this
values in an array P[1, MAXN]. Here is how P should look like for the tree in the example (for simplity, for every node i in the first
section let P[i] = 1):
Notice that for the nodes situated on the levels that are the first ones in some sections, P[i] = T[i]. We can preprocess P using a depth
first search (T[i] is the father of node i in the tree, nr is [sqrt(H)] and L[i] is the level of the node i):
void dfs(int node, int T[MAXN], int N, int P[MAXN], int L[MAXN], int nr){
int k;
//if node is situated in the first
//section then P[node] = 1
//if node is situated at the beginning
//of some section then P[node] = T[node]
//if none of those two cases occurs, then
//P[node] = P[T[node]]
if (L[node] < nr)
P[node] = 1;
else
if(!(L[node] % nr))
P[node] = T[node];
else
P[node] = P[T[node]];
for each son k of node
dfs(k, T, N, P, L, nr);
}
Now, we can easily make queries. For finding LCA(x, y) we we will first find in what section it lays, and then trivially compute it. Here
is the code:
int LCA(int T[MAXN], int P[MAXN], int L[MAXN], int x, int y){
//as long as the node in the next section of
//x and y is not one common ancestor
//we get the node situated on the smaller lever closer
while (P[x] != P[y])
if (L[x] > L[y])
x = P[x];
else
y = P[y];
//now they are in the same section, so we trivially compute the LCA
while (x != y)
if (L[x] > L[y])
x = T[x];
else
y = T[y];
return x;
}
This function makes at most 2 * sqrt(H) operations. Using this approach we get an <O(N), O(sqrt(H))> algorithm, where H is the
height of the tree. In the worst case H = N, so the overall complexity is <O(N), O(sqrt(N))>. The main advantage of this algorithm is
quick coding (an average Division 1 coder shouldn’t need more than 15 minutes to code it).
Another easy solution in <O(N logN, O(logN)>
If we need a faster solution for this problem we could use dynamic programming. First, let’s compute a table P[1,N][1,logN] where P[i]
[j] is the 2jth ancestor of i. For computing this value we may use the following recursion:
The preprocessing function should look like this:
void process3(int N, int T[MAXN], int P[MAXN][LOGMAXN]){
int i, j;
//we initialize every element in P with ‐1
for (i = 0; i < N; i++)
for (j = 0; 1 << j < N; j++)
P[i][j] = ‐1;
//the first ancestor of every node i is T[i]
for (i = 0; i < N; i++)
P[i][0] = T[i];
//bottom up dynamic programing
for (j = 1; 1 << j < N; j++)
for (i = 0; i < N; i++)
if (P[i][j ‐ 1] != ‐1)
P[i][j] = P[P[i][j ‐ 1]][j ‐ 1];
}
This takes O(N logN) time and space. Now let’s see how we can make queries. Let L[i] be the level of node i in the tree. We must
observe that if p and q are on the same level in the tree we can compute LCA(p, q) using a metabinary search. So, for every power j
of 2 (between log(L[p]) and 0, in descending order), if P[p][j] != P[q][j] then we know that LCA(p, q) is on a higher level and we
will continue searching for LCA(p = P[p][j], q = P[q][j]). At the end, both p and q will have the same father, so return T[p]. Let’s
see what happens if L[p] != L[q]. Assume, without loss of generality, that L[p] < L[q]. We can use the same metabinary search for
finding the ancestor of p situated on the same level with q, and then we can compute the LCA as described below. Here is how the
query function should look:
int query(int N, int P[MAXN][LOGMAXN], int T[MAXN],
int L[MAXN], int p, int q){
int tmp, log, i;
//if p is situated on a higher level than q then we swap them
if (L[p] < L[q])
tmp = p, p = q, q = tmp;
//we compute the value of [log(L[p)]
for (log = 1; 1 << log <= L[p]; log++);
log‐‐;
//we find the ancestor of node p situated on the same level
//with q using the values in P
for (i = log; i >= 0; i‐‐)
if (L[p] ‐ (1 << i) >= L[q])
p = P[p][i];
if (p == q)
return p;
//we compute LCA(p, q) using the values in P
for (i = log; i >= 0; i‐‐)
if (P[p][i] != ‐1 && P[p][i] != P[q][i])
p = P[p][i], q = P[q][i];
return T[p];
}
Now, we can see that this function makes at most 2 * log(H) operations, where H is the height of the tree. In the worst case H = N,
so the overall complexity of this algorithm is <O(N logN), O(logN)>. This solution is easy to code too, and it’s faster than the
previous one.
Reduction from LCA to RMQ
Now, let’s show how we can use RMQ for computing LCA queries. Actually, we will reduce the LCA problem to RMQ in linear time, so
every algorithm that solves the RMQ problem will solve the LCA problem too. Let’s show how this reduction can be done using an
example:
click to enlarge image
Notice that LCAT(u, v) is the closest node from the root encountered between the visits of u and v during a depth first search of T. So,
we can consider all nodes between any two indices of u and v in the Euler Tour of the tree and then find the node situated on the
smallest level between them. For this, we must build three arrays:
E[1, 2*N1] – the nodes visited in an Euler Tour of T; E[i] is the label of ith visited node in the tour
L[1, 2*N1] – the levels of the nodes visited in the Euler Tour; L[i] is the level of node E[i]
H[1, N] – H[i] is the index of the first occurrence of node i in E (any occurrence would be good, so it’s not bad if we consider the
first one)
Assume that H[u] < H[v] (otherwise you must swap u and v). It’s easy to see that the nodes between the first occurrence of u and
the first occurrence of v are E[H[u]…H[v]]. Now, we must find the node situated on the smallest level. For this, we can use RMQ. So,
LCAT(u, v) = E[RMQL(H[u], H[v])] (remember that RMQ returns the index). Here is how E, L and H should look for the example:
click to enlarge image
Notice that consecutive elements in L differ by exactly 1.
From RMQ to LCA
We have shown that the LCA problem can be reduced to RMQ in linear time. Here we will show how we can reduce the RMQ problem to
LCA. This means that we actually can reduce the general RMQ to the restricted version of the problem (where consecutive elements in
the array differ by exactly 1). For this we should use cartesian trees.
A Cartesian Tree of an array A[0, N 1] is a binary tree C(A) whose root is a minimum element of A, labeled with the position i of
this minimum. The left child of the root is the Cartesian Tree of A[0, i 1] if i > 0, otherwise there’s no child. The right child is defined
similary for A[i + 1, N 1]. Note that the Cartesian Tree is not necessarily unique if A contains equal elements. In this tutorial the first
appearance of the minimum value will be used, thus the Cartesian Tree will be unique. It’s easy to see now that RMQA(i, j) = LCAC(i,
j).
Here is an example:
Now we only have to compute C(A) in linear time. This can be done using a stack. At the beginning the stack is empty. We will then
insert the elements of A in the stack. At the ith step A[i] will be added next to the last element in the stack that has a smaller or
equal value to A[i], and all the greater elements will be removed. The element that was in the stack on the position of A[i] before the
insertion was done will become the left son of i, and A[i] will become the right son of the smaller element behind him. At every step
the first element in the stack is the root of the cartesian tree. It’s easier to build the tree if the stack will hold the indexes of the
elements, and not their value.
Here is how the stack will look at each step for the example above:
0 0 0 is the only node in the tree.
1 0 1 1 is added at the end of the stack.
Now, 1 is the right son of 0.
2 0 2 2 is added next to 0, and 1 is
removed (A[2] < A[1]). Now, 2 is
the right son of 0 and the left son
of 2 is 1.
3 3 A[3] is the smallest element in the
vector so far, so all elements in the
stack will be removed and 3 will
become the root of the tree. The
left child of 3 is 0.
4 3 4 4 is added next to 3, and the right
son of 3 is 4.
5 3 4 5 5 is added next to 4, and the right
son of 4 is 5.
6 3 4 5 6 6 is added next to 5, and the right
son of 5 is 6.
7 3 4 5 6 7 is added next to 6, and the right
7 son of 6 is 7.
8 3 8 8 is added next to 3, and all
greater elements are removed. 8
is now the right child of 3 and the
left child of 8 is 4.
9 3 8 9 9 is added next to 8, and the right
son of 8 is 9.
Note that every element in A is only added once and removed at most once, so the complexity of this algorithm is O(N). Here is how
the treeprocessing function will look:
void computeTree(int A[MAXN], int N, int T[MAXN]){
int st[MAXN], i, k, top = ‐1;
//we start with an empty stack
//at step i we insert A[i] in the stack
for (i = 0; i < N; i++){
//compute the position of the first element that is
//equal or smaller than A[i]
k = top;
while (k >= 0 && A[st[k]] > A[i])
k‐‐;
//we modify the tree as explained above
if (k != ‐1)
T[i] = st[k];
if (k < top)
T[st[k + 1]] = i;
//we insert A[i] in the stack and remove
//any bigger elements
st[++k] = i;
top = k;
}
//the first element in the stack is the root of
//the tree, so it has no father
T[st[0]] = ‐1;
}
An<O(N), O(1)> algorithm for the restricted RMQ
Now we know that the general RMQ problem can be reduced to the restricted version using LCA. Here, consecutive elements in the
array differ by exactly 1. We can use this and give a fast <O(N), O(1)> algorithm. From now we will solve the RMQ problem for an
array A[0, N 1] where |A[i] – A[i + 1]| = 1, i = [1, N 1]. We transform A in a binary array with N1 elements, where A[i] =
A[i] – A[i + 1]. It’s obvious that elements in A can be just +1 or 1. Notice that the old value of A[i] is now the sum of A[1], A[2] ..
A[i] plus the old A[0]. However, we won’t need the old values from now on.
To solve this restricted version of the problem we need to partition A into blocks of size l = [(log N) / 2]. Let A’[i] be the minimum
value for the ith block in A and B[i] be the position of this minimum value in A. Both A and B are N/l long. Now, we preprocess A’
using the ST algorithm described in Section1. This will take O(N/l * log(N/l)) = O(N) time and space. After this preprocessing we
can make queries that span over several blocks in O(1). It remains now to show how the inblock queries can be made. Note that the
length of a block is l = [(log N) / 2], which is quite small. Also, note that A is a binary array. The total number of binary arrays of size
l is 2l=sqrt(N). So, for each binary block of size l we need to lock up in a table P the value for RMQ between every pair of indices. This
can be trivially computed in O(sqrt(N)*l2)=O(N) time and space. To index table P, preprocess the type of each block in A and store it
in array T[1, N/l]. The block type is a binary number obtained by replacing 1 with 0 and +1 with 1.
Now, to answer RMQA(i, j) we have two cases:
i and j are in the same block, so we use the value computed in P and T
i and j are in different blocks, so we compute three values: the minimum from i to the end of i’s block using P and T, the
minimum of all blocks between i’s and j‘s block using precomputed queries on A’ and the minimum from the begining of j’s block
to j, again using T and P; finally return the position where the overall minimum is using the three values you just computed.
Conclusion
RMQ and LCA are strongly related problems that can be reduced one to another. Many algorithms can be used to solve them, and they
can be adapted to other kind of problems as well.
Here are some training problems for segment trees, LCA and RMQ:
SRM 310 > Floating Median
http://acm.pku.edu.cn/JudgeOnline/problem?id=1986
http://acm.pku.edu.cn/JudgeOnline/problem?id=2374
http://acmicpclivearchive.uva.es/nuevoportal/data/problem.php?p=2045
http://acm.pku.edu.cn/JudgeOnline/problem?id=2763
http://www.spoj.pl/problems/QTREE2/
http://acm.uva.es/p/v109/10938.html
http://acm.sgu.ru/problem.php?contest=0&problem=155
References
“Theoretical and Practical Improvements on the RMQProblem, with Applications to LCA and LCE” [PDF] by Johannes Fischer and
Volker Heunn
“The LCA Problem Revisited” [PPT] by Michael A.Bender and Martin FarachColton a very good presentation, ideal for quick learning
of some LCA and RMQ aproaches
“Faster algorithms for finding lowest common ancestors in directed acyclic graphs” [PDF] by Artur Czumaj, Miroslav Kowaluk and
Andrzej Lingas
Power up C++ with the Standard Template Library: Part 1
By DmitryKorolev– TopCoder Member
Discuss this article in the forums
Containers
Before we begin
Vector
Pairs
Iterators
Compiling STL Programs
Data manipulation in Vector
String
Set
Map
Notice on Map and Set
More on algorithms
String Streams
Summary
Perhaps you are already using C++ as your main programming language to solve TopCoder problems. This means that you have
already used STL in a simple way, because arrays and strings are passed to your function as STL objects. You may have noticed,
though, that many coders manage to write their code much more quickly and concisely than you.
Or perhaps you are not a C++ programmer, but want to become one because of the great functionality of this language and its libraries
(and, maybe, because of the very short solutions you’ve read in TopCoder practice rooms and competitions).
Regardless of where you’re coming from, this article can help. In it, we will review some of the powerful features of the Standard
Template Library (STL) – a great tool that, sometimes, can save you a lot of time in an algorithm competition.
The simplest way to get familiar with STL is to begin from its containers.
Containers
Any time you need to operate with many elements you require some kind of container. In native C (not C++) there was only one type
of container: the array.
The problem is not that arrays are limited (though, for example, it’s impossible to determine the size of array at runtime). Instead, the
main problem is that many problems require a container with greater functionality.
For example, we may need one or more of the following operations:
Add some string to a container.
Remove a string from a container.
Determine whether a string is present in the container.
Return a number of distinct elements in a container.
Iterate through a container and get a list of added strings in some order.
Of course, one can implement this functionality in an ordinal array. But the trivial implementation would be very inefficient. You can
create the tree of hash structure to solve it in a faster way, but think a bit: does the implementation of such a container depend on
elements we are going to store? Do we have to reimplement the module to make it functional, for example, for points on a plane but
not strings?
If not, we can develop the interface for such a container once, and then use everywhere for data of any type. That, in short, is the idea
of STL containers.
Before we begin
When the program is using STL, it should #include the appropriate standard headers. For most containers the title of standard header
matches the name of the container, and no extension is required. For example, if you are going to use stack, just add the following line
at the beginning of your program:
#include <stack>
Container types (and algorithms, functors and all STL as well) are defined not in global namespace, but in special namespace called
“std.” Add the following line after your includes and before the code begin:
using namespace std;
Another important thing to remember is that the type of a container is the template parameter. Template parameters are specified with
the ‘<’/’>’ “brackets” in code. For example:
vector<int> N;
When making nested constructions, make sure that the “brackets” are not directly following one another – leave a blank between them.
vector< vector<int> > CorrectDefinition;
vector<vector<int>> WrongDefinition; // Wrong: compiler may be confused by 'operator >>'
Vector
The simplest STL container is vector. Vector is just an array with extended functionality. By the way, vector is the only container that is
backwardcompatible to native C code – this means that vector actually IS the array, but with some additional features.
vector<int> v(10);
for(int i = 0; i < 10; i++) {
v[i] = (i+1)*(i+1);
}
for(int i = 9; i > 0; i‐‐) {
v[i] ‐= v[i‐1];
}
Actually, when you type
vector<int> v;
the empty vector is created. Be careful with constructions like this:
vector<int> v[10];
Here we declare ‘v’ as an array of 10 vector<int>’s, which are initially empty. In most cases, this is not that we want. Use parentheses
instead of brackets here. The most frequently used feature of vector is that it can report its size.
int elements_count = v.size();
Two remarks: first, size() is unsigned, which may sometimes cause problems. Accordingly, I usually define macros, something like sz(C)
that returns size of C as ordinal signed int. Second, it’s not a good practice to compare v.size() to zero if you want to know whether the
container is empty. You’re better off using empty() function:
bool is_nonempty_notgood = (v.size() >= 0); // Try to avoid this
bool is_nonempty_ok = !v.empty();
This is because not all the containers can report their size in O(1), and you definitely should not require counting all elements in a
doublelinked list just to ensure that it contains at least one.
Another very popular function to use in vector is push_back. Push_back adds an element to the end of vector, increasing its size by one.
Consider the following example:
vector<int> v;
for(int i = 1; i < 1000000; i *= 2) {
v.push_back(i);
}
int elements_count = v.size();
Don’t worry about memory allocation — vector will not allocate just one element each time. Instead, vector allocates more memory
then it actually needs when adding new elements with push_back. The only thing you should worry about is memory usage, but at
TopCoder this may not matter. (More on vector’s memory policy later.)
When you need to resize vector, use the resize() function:
vector<int> v(20);
for(int i = 0; i < 20; i++) {
v[i] = i+1;
}
v.resize(25);
for(int i = 20; i < 25; i++) {
v[i] = i*2;
}
The resize() function makes vector contain the required number of elements. If you require less elements than vector already contain,
the last ones will be deleted. If you ask vector to grow, it will enlarge its size and fill the newly created elements with zeroes.
Note that if you use push_back() after resize(), it will add elements AFTER the newly allocated size, but not INTO it. In the example
above the size of the resulting vector is 25, while if we use push_back() in a second loop, it would be 30.
vector<int> v(20);
for(int i = 0; i < 20; i++) {
v[i] = i+1;
}
v.resize(25);
for(int i = 20; i < 25; i++) {
v.push_back(i*2); // Writes to elements with indices [25..30), not [20..25) ! <
}
To clear a vector use clear() member function. This function makes vector to contain 0 elements. It does not make elements zeroes
watch out it completely erases the container.
There are many ways to initialize vector. You may create vector from another vector:
vector<int> v1;
// ...
vector<int> v2 = v1;
vector<int> v3(v1);
The initialization of v2 and v3 in the example above are exactly the same.
If you want to create a vector of specific size, use the following constructor:
vector<int> Data(1000);
In the example above, the data will contain 1,000 zeroes after creation. Remember to use parentheses, not brackets. If you want
vector to be initialized with something else, write it in such manner:
vector<string> names(20, “Unknown”);
Remember that you can create vectors of any type.
Multidimensional arrays are very important. The simplest way to create the twodimensional array via vector is to create a vector of
vectors.
vector< vector<int> > Matrix;
It should be clear to you now how to create the twodimensional vector of given size:
int N, M;
// ...
vector< vector<int> > Matrix(N, vector<int>(M, ‐1));
Here we create a matrix of size N*M and fill it with 1.
The simplest way to add data to vector is to use push_back(). But what if we want to add data somewhere other than the end? There is
the insert() member function for this purpose. And there is also the erase() member function to erase elements, as well. But first we
need to say a few words about iterators.
You should remember one more very important thing: When vector is passed as a parameter to some function, a copy of vector is
actually created. It may take a lot of time and memory to create new vectors when they are not really needed. Actually, it’s hard to find
a task where the copying of vector is REALLY needed when passing it as a parameter. So, you should never write:
void some_function(vector<int> v) { // Never do it unless you’re sure what you do!
// ...
}
Instead, use the following construction:
void some_function(const vector<int>& v) { // OK
// ...
}
If you are going to change the contents of vector in the function, just omit the ‘const’ modifier.
int modify_vector(vector<int>& v) { // Correct
V[0]++;
}
Pairs
Before we come to iterators, let me say a few words about pairs. Pairs are widely used in STL. Simple problems, like TopCoder SRM 250
and easy 500point problems, usually require some simple data structure that fits well with pair. STL std::pair is just a pair of elements.
The simplest form would be the following:
template<typename T1, typename T2> struct pair {
T1 first;
T2 second;
};
In general pair<int,int> is a pair of integer values. At a more complex level, pair<string, pair<int, int> > is a pair of string and two
integers. In the second case, the usage may be like this:
pair<string, pair<int,int> > P;
string s = P.first; // extract string
int x = P.second.first; // extract first int
int y = P.second.second; // extract second int
The great advantage of pairs is that they have builtin operations to compare themselves. Pairs are compared firsttosecond element.
If the first elements are not equal, the result will be based on the comparison of the first elements only; the second elements will be
compared only if the first ones are equal. The array (or vector) of pairs can easily be sorted by STL internal functions.
For example, if you want to sort the array of integer points so that they form a polygon, it’s a good idea to put them to the vector<
pair<double, pair<int,int> >, where each element of vector is { polar angle, { x, y } }. One call to the STL sorting function will give
you the desired order of points.
Pairs are also widely used in associative containers, which we will speak about later in this article.
Iterators
What are iterators? In STL iterators are the most general way to access data in containers. Consider the simple problem: Reverse the
array A of N int’s. Let’s begin from a Clike solution:
void reverse_array_simple(int *A, int N) {
int first = 0, last = N‐1; // First and last indices of elements to be swapped
While(first < last) { // Loop while there is something to swap
swap(A[first], A[last]); // swap(a,b) is the standard STL function
first++; // Move first index forward
last‐‐; // Move last index back
}
}
This code should be clear to you. It’s pretty easy to rewrite it in terms of pointers:
void reverse_array(int *A, int N) {
int *first = A, *last = A+N‐1;
while(first < last) {
Swap(*first, *last);
first++;
last‐‐;
}
}
Look at this code, at its main loop. It uses only four distinct operations on pointers ‘first’ and ‘last’:
compare pointers (first < last),
get value by pointer (*first, *last),
increment pointer, and
decrement pointer
Now imagine that you are facing the second problem: Reverse the contents of a doublelinked list, or a part of it. The first code, which
uses indexing, will definitely not work. At least, it will not work in time, because it’s impossible to get element by index in a double
linked list in O(1), only in O(N), so the whole algorithm will work in O(N^2). Errr…
But look: the second code can work for ANY pointerlike object. The only restriction is that that object can perform the operations
described above: take value (unary *), comparison (<), and increment/decrement (++/–). Objects with these properties that are
associated with containers are called iterators. Any STL container may be traversed by means of an iterator. Although not often needed
for vector, it’s very important for other container types.
So, what do we have? An object with syntax very much like a pointer. The following operations are defined for iterators:
get value of an iterator, int x = *it;
increment and decrement iterators it1++, it2–;
compare iterators by ‘!=’ and by ‘<’
add an immediate to iterator it += 20; <=> shift 20 elements forward
get the distance between iterators, int n = it2it1;
But instead of pointers, iterators provide much greater functionality. Not only can they operate on any container, they may also
perform, for example, range checking and profiling of container usage.
And the main advantage of iterators, of course, is that they greatly increase the reuse of code: your own algorithms, based on iterators,
will work on a wide range of containers, and your own containers, which provide iterators, may be passed to a wide range of standard
functions.
Not all types of iterators provide all the potential functionality. In fact, there are socalled “normal iterators” and “random access
iterators”. Simply put, normal iterators may be compared with ‘==’ and ‘!=’, and they may also be incremented and decremented. They
may not be subtracted and we can not add a value to the normal iterator. Basically, it’s impossible to implement the described
operations in O(1) for all container types. In spite of this, the function that reverses array should look like this:
template<typename T> void reverse_array(T *first, T *last) {
if(first != last) {
while(true) {
swap(*first, *last);
first++;
if(first == last) {
break;
}
last‐‐;
if(first == last) {
break;
}
}
}
}
The main difference between this code and the previous one is that we don’t use the “<” comparison on iterators, just the “==” one.
Again, don’t panic if you are surprised by the function prototype: template is just a way to declare a function, which works on any
appropriate parameter types. This function should work perfectly on pointers to any object types and with all normal iterators.
Let’s return to the STL. STL algorithms always use two iterators, called “begin” and “end.” The end iterator is pointing not to the last
object, however, but to the first invalid object, or the object directly following the last one. It’s often very convenient.
Each STL container has member functions begin() and end() that return the begin and end iterators for that container.
Based on these principles, c.begin() == c.end() if and only if c is empty, and c.end() – c.begin() will always be equal to c.size(). (The
last sentence is valid in cases when iterators can be subtracted, i.e. begin() and end() return random access iterators, which is not true
for all kinds of containers. See the prior example of the doublelinked list.)
The STLcompliant reverse function should be written as follows:
template<typename T> void reverse_array_stl_compliant(T *begin, T *end) {
// We should at first decrement 'end'
// But only for non‐empty range
if(begin != end){
end‐‐;
if(begin != end) {
while(true) {
swap(*begin, *end);
begin++;
If(begin == end) {
break;
}
end‐‐;
if(begin == end) {
break;
}
}
}
}
}
Note that this function does the same thing as the standard function std::reverse(T begin, T end) that can be found in algorithms
module (#include <algorithm>).
In addition, any object with enough functionality can be passed as an iterator to STL algorithms and functions. That is where the power
of templates comes in! See the following examples:
vector<int> v;
// ...
vector<int> v2(v);
vector<int> v3(v.begin(), v.end()); // v3 equals to v2
int data[] = { 2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31 };
vector<int> primes(data, data+(sizeof(data) / sizeof(data[0])));
The last line performs a construction of vector from an ordinal C array. The term ‘data’ without index is treated as a pointer to the
beginning of the array. The term ‘data + N’ points to Nth element, so, when N if the size of array, ‘data + N’ points to first element not
in array, so ‘data + length of data’ can be treated as end iterator for array ‘data’. The expression ‘sizeof(data)/sizeof(data[0])’ returns
the size of the array data, but only in a few cases, so don’t use it anywhere except in such constructions. (C programmers will agree
with me!)
Furthermore, we can even use the following constructions:
vector<int> v;
// ...
vector<int> v2(v.begin(), v.begin() + (v.size()/2));
It creates the vector v2 that is equal to the first half of vector v.
Here is an example of reverse() function:
int data[10] = { 1, 3, 5, 7, 9, 11, 13, 15, 17, 19 };
reverse(data+2, data+6); // the range { 5, 7, 9, 11 } is now { 11, 9, 7, 5 };
Each container also has the rbegin()/rend() functions, which return reverse iterators. Reverse iterators are used to traverse the
container in backward order. Thus:
vector<int> v;
vector<int> v2(v.rbegin()+(v.size()/2), v.rend());
will create v2 with first half of v, ordered backtofront.
To create an iterator object, we must specify its type. The type of iterator can be constructed by a type of container by appending
“::iterator”, “::const_iterator”, “::reverse_iterator” or “::const_reverse_iterator” to it. Thus, vector can be traversed in the following
way:
vector<int> v;
// ...
// Traverse all container, from begin() to end()
for(vector<int>::iterator it = v.begin(); it != v.end(); it++) {
*it++; // Increment the value iterator is pointing to
}
I recommend you use ‘!=’ instead of ‘<’, and ‘empty()’ instead of ‘size() != 0′ — for some container types, it’s just very inefficient to
determine which of the iterators precedes another.
Now you know of STL algorithm reverse(). Many STL algorithms are declared in the same way: they get a pair of iterators – the
beginning and end of a range – and return an iterator.
The find() algorithm looks for appropriate elements in an interval. If the element is found, the iterator pointing to the first occurrence of
the element is returned. Otherwise, the return value equals the end of interval. See the code:
vector<int> v;
for(int i = 1; i < 100; i++) {
v.push_back(i*i);
}
if(find(v.begin(), v.end(), 49) != v.end()) {
// ...
}
To get the index of element found, one should subtract the beginning iterator from the result of find():
int i = (find(v.begin(), v.end(), 49) ‐ v.begin();
if(i < v.size()) {
// ...
}
Remember to #include <algorithm> in your source when using STL algorithms.
The min_element and max_element algorithms return an iterator to the respective element. To get the value of min/max element, like
in find(), use *min_element(…) or *max_element(…), to get index in array subtract the begin iterator of a container or range:
int data[5] = { 1, 5, 2, 4, 3 };
vector<int> X(data, data+5);
int v1 = *max_element(X.begin(), X.end()); // Returns value of max element in vector
int i1 = min_element(X.begin(), X.end()) – X.begin; // Returns index of min element in vector
int v2 = *max_element(data, data+5); // Returns value of max element in array
int i3 = min_element(data, data+5) – data; // Returns index of min element in array
Now you may see that the useful macros would be:
#define all(c) c.begin(), c.end()
Don’t put the whole righthand side of these macros into parentheses – that would be wrong!
Another good algorithm is sort(). It’s very easy to use. Consider the following examples:
vector<int> X;
// ...
sort(X.begin(), X.end()); // Sort array in ascending order
sort(all(X)); // Sort array in ascending order, use our #define
sort(X.rbegin(), X.rend()); // Sort array in descending order using with reverse iterators
Compiling STL Programs
One thing worth pointing out here is STL error messages. As the STL is distributed in sources, and it becomes necessary for compilers
to build efficient executables, one of STL’s habits is unreadable error messages.
For example, if you pass a vector<int> as a const reference parameter (as you should do) to some function:
void f(const vector<int>& v) {
for(
vector<int>::iterator it = v.begin(); // hm... where’s the error?..
// ...
// ...
}
The error here is that you are trying to create the nonconst iterator from a const object with the begin() member function (though
identifying that error can be harder than actually correcting it). The right code looks like this:
void f(const vector<int>& v) {
int r = 0;
// Traverse the vector using const_iterator
for(vector<int>::const_iterator it = v.begin(); it != v.end(); it++) {
r += (*it)*(*it);
}
return r;
}
In spite of this, let me tell about very important feature of GNU C++ called ‘typeof’. This operator is replaced to the type of an
expression during the compilation. Consider the following example:
typeof(a+b) x = (a+b);
This will create the variable x of type matching the type of (a+b) expression. Beware that typeof(v.size()) is unsigned for any STL
container type. But the most important application of typeof for TopCoder is traversing a container. Consider the following macros:
#define tr(container, it)
for(typeof(container.begin()) it = container.begin(); it != container.end(); it++)
By using these macros we can traverse every kind of container, not only vector. This will produce const_iterator for const object and
normal iterator for nonconst object, and you will never get an error here.
void f(const vector<int>& v) {
int r = 0;
tr(v, it) {
r += (*it)*(*it);
}
return r;
}
Note: I did not put additional parentheses on the #define line in order to improve its readability. See this article below for more correct
#define statements that you can experiment with in practice rooms.
Traversing macros is not really necessary for vectors, but it’s very convenient for more complex data types, where indexing is not
supported and iterators are the only way to access data. We will speak about this later in this article.
Data manipulation in vector
One can insert an element to vector by using the insert() function:
vector<int> v;
// ...
v.insert(1, 42); // Insert value 42 after the first
All elements from second (index 1) to the last will be shifted right one element to leave a place for a new element. If you are planning
to add many elements, it’s not good to do many shifts – you’re better off calling insert() one time. So, insert() has an interval form:
vector<int> v;
vector<int> v2;
// ..
// Shift all elements from second to last to the appropriate number of elements.
// Then copy the contents of v2 into v.
v.insert(1, all(v2));
Vector also has a member function erase, which has two forms. Guess what they are:
erase(iterator);
erase(begin iterator, end iterator);
At first case, single element of vector is deleted. At second case, the interval, specified by two iterators, is erased from vector.
The insert/erase technique is common, but not identical for all STL containers.
String
There is a special container to manipulate with strings. The string container has a few differences from vector<char>. Most of the
differences come down to string manipulation functions and memory management policy.
String has a substring function without iterators, just indices:
string s = "hello";
string
s1 = s.substr(0, 3), // "hel"
s2 = s.substr(1, 3), // "ell"
s3 = s.substr(0, s.length()‐1), "hell"
s4 = s.substr(1); // "ello"
Beware of (s.length()1) on empty string because s.length() is unsigned and unsigned(0) – 1 is definitely not what you are expecting!
Set
It’s always hard to decide which kind of container to describe first – set or map. My opinion is that, if the reader has a basic knowledge
of algorithms, beginning from ‘set’ should be easier to understand.
Consider we need a container with the following features:
add an element, but do not allow duples [duplicates?]
remove elements
get count of elements (distinct elements)
check whether elements are present in set
This is quite a frequently used task. STL provides the special container for it – set. Set can add, remove and check the presence of
particular element in O(log N), where N is the count of objects in the set. While adding elements to set, the duples [duplicates?] are
discarded. A count of the elements in the set, N, is returned in O(1). We will speak of the algorithmic implementation of set and map
later — for now, let’s investigate its interface:
set<int> s;
for(int i = 1; i <= 100; i++) {
s.insert(i); // Insert 100 elements, [1..100]
}
s.insert(42); // does nothing, 42 already exists in set
for(int i = 2; i <= 100; i += 2) {
s.erase(i); // Erase even values
}
int n = int(s.size()); // n will be 50
The push_back() member may not be used with set. It make sense: since the order of elements in set does not matter, push_back() is
not applicable here.
Since set is not a linear container, it’s impossible to take the element in set by index. Therefore, the only way to traverse the elements
of set is to use iterators.
// Calculate the sum of elements in set
set<int> S;
// ...
int r = 0;
for(set<int>::const_iterator it = S.begin(); it != S.end(); it++) {
r += *it;
}
It’s more elegant to use traversing macros here. Why? Imagine you have a set< pair<string, pair< int, vector<int> > >. How to
traverse it? Write down the iterator type name? Oh, no. Use our traverse macros instead.
set< pair<string, pair< int, vector<int> > > SS;
int total = 0;
tr(SS, it) {
total += it‐>second.first;
}
Notice the ‘it>second.first’ syntax. Since ‘it’ is an iterator, we need to take an object from ‘it’ before operating. So, the correct syntax
would be ‘(*it).second.first’. However, it’s easier to write ‘something>’ than ‘(*something)’. The full explanation will be quite long –just
remember that, for iterators, both syntaxes are allowed.
To determine whether some element is present in set use ‘find()’ member function. Don’t be confused, though: there are several ‘find()’
’s in STL. There is a global algorithm ‘find()’, which takes two iterators, element, and works for O(N). It is possible to use it for
searching for element in set, but why use an O(N) algorithm while there exists an O(log N) one? While searching in set and map (and
also in multiset/multimap, hash_map/hash_set, etc.) do not use global find – instead, use member function ‘set::find()’. As ‘ordinal’
find, set::find will return an iterator, either to the element found, or to ‘end()’. So, the element presence check looks like this:
set<int> s;
// ...
if(s.find(42) != s.end()) {
// 42 presents in set
}
else {
// 42 not presents in set
}
Another algorithm that works for O(log N) while called as member function is count. Some people think that
if(s.count(42) != 0) {
// …
}
or even
if(s.count(42)) {
// …
}
is easier to write. Personally, I don’t think so. Using count() in set/map is nonsense: the element either presents or not. As for me, I
prefer to use the following two macros:
#define present(container, element) (container.find(element) != container.end())
#define cpresent(container, element) (find(all(container),element) != container.end())
(Remember that all(c) stands for “c.begin(), c.end()”)
Here, ‘present()’ returns whether the element presents in the container with member function ‘find()’ (i.e. set/map, etc.) while
‘cpresent’ is for vector.
To erase an element from set use the erase() function.
set<int> s;
// …
s.insert(54);
s.erase(29);
The erase() function also has the interval form:
set<int> s;
// ..
set<int>::iterator it1, it2;
it1 = s.find(10);
it2 = s.find(100);
// Will work if it1 and it2 are valid iterators, i.e. values 10 and 100 present in set.
s.erase(it1, it2); // Note that 10 will be deleted, but 100 will remain in the container
Set has an interval constructor:
int data[5] = { 5, 1, 4, 2, 3 };
set<int> S(data, data+5);
It gives us a simple way to get rid of duplicates in vector, and sort it:
vector<int> v;
// …
set<int> s(all(v));
vector<int> v2(all(s));
Here ‘v2′ will contain the same elements as ‘v’ but sorted in ascending order and with duplicates removed.
Any comparable elements can be stored in set. This will be described later.
Map
There are two explanation of map. The simple explanation is the following:
map<string, int> M;
M["Top"] = 1;
M["Coder"] = 2;
M["SRM"] = 10;
int x = M["Top"] + M["Coder"];
if(M.find("SRM") != M.end()) {
M.erase(M.find("SRM")); // or even M.erase("SRM")
}
Very simple, isn’t it?
Actually map is very much like set, except it contains not just values but pairs <key, value>. Map ensures that at most one pair with
specific key exists. Another quite pleasant thing is that map has operator [] defined.
Traversing map is easy with our ‘tr()’ macros. Notice that iterator will be an std::pair of key and value. So, to get the value use it
>second. The example follows:
map<string, int> M;
// …
int r = 0;
tr(M, it) {
r += it‐>second;
}
Don’t change the key of map element by iterator, because it may break the integrity of map internal data structure (see below).
There is one important difference between map::find() and map::operator []. While map::find() will never change the contents of map,
operator [] will create an element if it does not exist. In some cases this could be very convenient, but it’s definitly a bad idea to use
operator [] many times in a loop, when you do not want to add new elements. That’s why operator [] may not be used if map is passed
as a const reference parameter to some function:
void f(const map<string, int>& M) {
if(M["the meaning"] == 42) { // Error! Cannot use [] on const map objects!
}
if(M.find("the meaning") != M.end() && M.find("the meaning")‐>second == 42) { // Correct
cout << "Don't Panic!" << endl;
}
}
Notice on Map and Set
Internally map and set are almost always stored as redblack trees. We do not need to worry about the internal structure, the thing to
remember is that the elements of map and set are always sorted in ascending order while traversing these containers. And that’s why
it’s strongly not recommended to change the key value while traversing map or set: If you make the modification that breaks the order,
it will lead to improper functionality of container’s algorithms, at least.
But the fact that the elements of map and set are always ordered can be practically used while solving TopCoder problems.
Another important thing is that operators ++ and — are defined on iterators in map and set. Thus, if the value 42 presents in set, and
it’s not the first and the last one, than the following code will work:
set<int> S;
// ...
set<int>::iterator it = S.find(42);
set<int>::iterator it1 = it, it2 = it;
it1‐‐;
it2++;
int a = *it1, b = *it2;
Here ‘a’ will contain the first neighbor of 42 to the left and ‘b’ the first one to the right.
More on algorithms
It’s time to speak about algorithms a bit more deeply. Most algorithms are declared in the #include <algorithm> standard header. At
first, STL provides three very simple algorithms: min(a,b), max(a,b), swap(a,b). Here min(a,b) and max(a,b) returns the minimum and
maximum of two elements, while swap(a,b) swaps two elements.
Algorithm sort() is also widely used. The call to sort(begin, end) sorts an interval in ascending order. Notice that sort() requires random
access iterators, so it will not work on all containers. However, you probably won’t ever call sort() on set, which is already ordered.
You’ve already heard of algorithm find(). The call to find(begin, end, element) returns the iterator where ‘element’ first occurs, or end if
the element is not found. Instead of find(…), count(begin, end, element) returns the number of occurrences of an element in a
container or a part of a container. Remember that set and map have the member functions find() and count(), which works in O(log N),
while std::find() and std::count() take O(N).
Other useful algorithms are next_permutation() and prev_permutation(). Let’s speak about next_permutation. The call to
next_permutation(begin, end) makes the interval [begin, end) hold the next permutation of the same elements, or returns false if the
current permutation is the last one. Accordingly, next_permutation makes many tasks quite easy. If you want to check all permutations,
just write:
vector<int> v;
for(int i = 0; i < 10; i++) {
v.push_back(i);
}
do {
Solve(..., v);
} while(next_permutation(all(v));
Don’t forget to ensure that the elements in a container are sorted before your first call to next_permutation(…). Their initial state
should form the very first permutation; otherwise, some permutations will not be checked.
String Streams
You often need to do some string processing/input/output. C++ provides two interesting objects for it: ‘istringstream’ and
‘ostringstream’. They are both declared in #include <sstream>.
Object istringstream allows you to read from a string like you do from a standard input. It’s better to view source:
void f(const string& s) {
// Construct an object to parse strings
istringstream is(s);
// Vector to store data
vector<int> v;
// Read integer while possible and add it to the vector
int tmp;
while(is >> tmp) {
v.push_back(tmp);
}
}
The ostringstream object is used to do formatting output. Here is the code:
string f(const vector<int>& v) {
// Constucvt an object to do formatted output
ostringstream os;
// Copy all elements from vector<int> to string stream as text
tr(v, it) {
os << ' ' << *it;
}
// Get string from string stream
string s = os.str();
// Remove first space character
if(!s.empty()) { // Beware of empty string here
s = s.substr(1);
}
return s;
}
Summary
To go on with STL, I would like to summarize the list of templates to be used. This will simplify the reading of code samples and, I hope,
improve your TopCoder skills. The short list of templates and macros follows:
typedef vector<int> vi;
typedef vector<vi> vvi;
typedef pair<int,int> ii;
#define sz(a) int((a).size())
#define pb push_back
#defile all(c) (c).begin(),(c).end()
#define tr(c,i) for(typeof((c).begin() i = (c).begin(); i != (c).end(); i++)
#define present(c,x) ((c).find(x) != (c).end())
#define cpresent(c,x) (find(all(c),x) != (c).end())
The container vector<int> is here because it’s really very popular. Actually, I found it convenient to have short aliases to many
containers (especially for vector<string>, vector<ii>, vector< pair<double, ii> >). But this list only includes the macros that are
required to understand the following text.
Another note to keep in mind: When a token from the lefthand side of #define appears in the righthand side, it should be placed in
braces to avoid many nontrivial problems.
Power up C++ with the Standard Template Library: Part 2
By DmitryKorolev– TopCoder Member
Creating Vector from Map
Copying Data Between Containers
Merging Lists
Calculating Algorithms
Nontrivial Sorting
Using Your Own Objects in Maps and Sets
Memory Management in Vectors
Implementing Real Algorithms with STL
Depthfirst Search (DFS)
A word on other container types and their usage
Queue
Breadthfirst Search (BFS)
Priority_Queue
Dijkstra
Dijkstra priority_queue
Dijkstra by set
What Is Not Included in STL
Creating Vector from Map
As you already know, map actually contains pairs of element. So you can write it in like this:
map<string, int> M;
// ...
vector< pair<string, int> > V(all(M)); // remember all(c) stands for
(c).begin(),(c).end()
Now vector will contain the same elements as map. Of course, vector will be sorted, as is map. This feature may be useful if you are not
planning to change elements in map any more but want to use indices of elements in a way that is impossible in map.
Copying data between containers
Let’s take a look at the copy(…) algorithm. The prototype is the following:
copy(from_begin, from_end, to_begin);
This algorithm copies elements from the first interval to the second one. The second interval should have enough space available. See
the following code:
vector<int> v1;
vector<int> v2;
// ...
// Now copy v2 to the end of v1
v1.resize(v1.size() + v2.size());
// Ensure v1 have enough space
copy(all(v2), v1.end() ‐ v2.size());
// Copy v2 elements right after v1 ones
Another good feature to use in conjunction with copy is inserters. I will not describe it here due to limited space but look at the code:
vector<int> v;
// ...
set<int> s;
// add some elements to set
copy(all(v), inserter(s));
The last line means:
tr(v, it) {
// remember traversing macros from Part I
s.insert(*it);
}
But why use our own macros (which work only in gcc) when there is a standard function? It’s a good STL practice to use standard
algorithms like copy, because it will be easy to others to understand your code.
To insert elemements to vector with push_back use back_inserter, or front_inserter is available for deque container. And in some cases
it is useful to remember that the first two arguments for ‘copy’ may be not only begin/end, but also rbegin/rend, which copy data in
reverse order.
Merging lists
Another common task is to operate with sorted lists of elements. Imagine you have two lists of elements — A and B, both ordered. You
want to get a new list from these two. There are four common operations here:
‘union’ the lists, R = A+B
intersect the lists, R = A*B
set difference, R = A*(~B) or R = AB
set symmetric difference, R = A XOR B
STL provides four algorithms for these tasks: set_union(…), set_intersection(…), set_difference(…) and set_symmetric_difference(…).
They all have the same calling conventions, so let’s look at set_intersection. A freestyled prototype would look like this:
end_result = set_intersection(begin1, end1, begin2, end2, begin_result);
Here [begin1,end1) and [begin2,end2) are the input lists. The 'begin_result' is the iterator from where the result will be written. But the
size of the result is unknown, so this function returns the end iterator of output (which determines how many elements are in the
result). See the example for usage details:
int data1[] = { 1, 2, 5, 6, 8, 9, 10 };
int data2[] = { 0, 2, 3, 4, 7, 8, 10 };
vector<int> v1(data1, data1+sizeof(data1)/sizeof(data1[0]));
vector<int> v2(data2, data2+sizeof(data2)/sizeof(data2[0]));
vector<int> tmp(max(v1.size(), v2.size());
vector<int> res = vector<int> (tmp.begin(), set_intersection(all(v1), all(v2), tmp.begin());
Look at the last line. We construct a new vector named 'res'. It is constructed via interval constructor, and the beginning of the interval
will be the beginning of tmp. The end of the interval is the result of the set_intersection algorithm. This algorithm will intersect v1 and
v2 and write the result to the output iterator, starting from 'tmp.begin()'. Its return value will actually be the end of the interval that
forms the resulting dataset.
One comment that might help you understand it better: If you would like to just get the number of elements in set intersection, use int
cnt = set_intersection(all(v1), all(v2), tmp.begin()) – tmp.begin();
Actually, I would never use a construction like ' vector<int> tmp'. I don't think it's a good idea to allocate memory for each set_***
algorithm invoking. Instead, I define the global or static variable of appropriate type and enough size. See below:
set<int> s1, s2;
for(int i = 0; i < 500; i++) {
s1.insert(i*(i+1) % 1000);
s2.insert(i*i*i % 1000);
}
static int temp[5000]; // greater than we need
vector<int> res = vi(temp, set_symmetric_difference(all(s1), all(s2), temp));
int cnt = set_symmetric_difference(all(s1), all(s2), temp) – temp;
Here 'res' will contain the symmetric difference of the input datasets.
Remember, input datasets need to be sorted to use these algorithms. So, another important thing to remember is that, because sets
are always ordered, we can use sets (and even maps, if you are not scared by pairs) as parameters for these algorithms.
These algorithms work in single pass, in O(N1+N2), when N1 and N2 are sizes of input datasets.
Calculating Algorithms
Yet another interesting algorithm is accumulate(...). If called for a vector of ints and third parameter zero, accumulate(...) will return
the sum of elements in vector:
vector<int> v;
// ...
int sum = accumulate(all(v), 0);
The result of accumulate() call always has the type of its third argument. So, if you are not sure that the sum fits in integer, specify the
third parameter's type directly:
vector<int> v;
// ...
long long sum = accumulate(all(v), (long long)0);
Accumulate can even calculate the product of values. The fourth parameter holds the predicate to use in calculations. So, if you want
the product:
vector<int> v;
// ...
double product = accumulate(all(v), double(1), multiplies<double>());
// don’t forget to start with 1 !
Another interesting algorithm is inner_product(...). It calculates the scalar product of two intervals. For example:
vector<int> v1;
vector<int> v2;
for(int i = 0; i < 3; i++) {
v1.push_back(10‐i);
v2.push_back(i+1);
}
int r = inner_product(all(v1), v2.begin(), 0);
'r' will hold (v1[0]*v2[0] + v1[1]*v2[1] + v1[2]*v2[2]), or (10*1+9*2+8*3), which is 52.
As for ‘accumulate’ the type of return value for inner_product is defined by the last parameter. The last parameter is the initial value for
the result. So, you may use inner_product for the hyperplane object in multidimensional space: just write inner_product(all(normal),
point.begin(), shift).
It should be clear to you now that inner_product requires only increment operation from iterators, so queues and sets can also be used
as parameters. Convolution filter, for calculating the nontrivial median value, could look like this:
set<int> values_ordered_data(all(data));
int n = sz(data); // int n = int(data.size());
vector<int> convolution_kernel(n);
for(int i = 0; i < n; i++) {
convolution_kernel[i] = (i+1)*(n‐i);
}
double result = double(inner_product(all(ordered_data), convolution_kernel.begin(), 0)) / accumulate(all(convolution_kernel), 0
Of course, this code is just an example practically speaking, it would be faster to copy values to another vector and sort it.
It's also possible to write a construction like this:
vector<int> v;
// ...
int r = inner_product(all(v), v.rbegin(), 0);
This will evaluate V[0]*V[N1] + V[1]+V[N2] + ... + V[N1]*V[0] where N is the number of elements in 'v'.
Nontrivial Sorting
Actually, sort(...) uses the same technique as all STL:
all comparison is based on 'operator <'
This means that you only need to override 'operator <'. Sample code follows:
struct fraction {
int n, d; // (n/d)
// ...
bool operator < (const fraction& f) const {
if(false) {
return (double(n)/d) < (double(f.n)/f.d);
// Try to avoid this, you're the TopCoder!
}
else {
return n*f.d < f.n*d;
}
}
};
// ...
vector<fraction> v;
// ...
sort(all(v));
In cases of nontrivial fields, your object should have default and copy constructor (and, maybe, assignment operator but this
comment is not for TopCoders).
Remember the prototype of 'operator <' : return type bool, const modifier, parameter const reference.
Another possibility is to create the comparison functor. Special comparison predicate may be passed to the sort(...) algorithm as a third
parameter. Example: sort points (that are pair<double,double>) by polar angle.
typedef pair<double, double> dd;
const double epsilon = 1e‐6;
struct sort_by_polar_angle {
dd center;
// Constuctor of any type
// Just find and store the center
template<typename T> sort_by_polar_angle(T b, T e) {
int count = 0;
center = dd(0,0);
while(b != e) {
center.first += b‐>first;
center.second += b‐>second;
b++;
count++;
}
double k = count ? (1.0/count) : 0;
center.first *= k;
center.second *= k;
}
// Compare two points, return true if the first one is earlier
// than the second one looking by polar angle
// Remember, that when writing comparator, you should
// override not ‘operator <’ but ‘operator ()’
bool operator () (const dd& a, const dd& b) const {
double p1 = atan2(a.second‐center.second, a.first‐center.first);
double p2 = atan2(b.second‐center.second, b.first‐center.first);
return p1 + epsilon < p2;
}
};
// ...
vector<dd> points;
// ...
sort(all(points), sort_by_polar_angle(all(points)));
This code example is complex enough, but it does demonstrate the abilities of STL. I should point out that, in this sample, all code will
be inlined during compilation, so it's actually really fast.
Also remember that 'operator <' should always return false for equal objects. It's very important – for the reason why, see the next
section.
Using your own objects in Maps and Sets
Elements in set and map are ordered. It's the general rule. So, if you want to enable using of your objects in set or map you should
make them comparable. You already know the rule of comparisons in STL:
| * all comparison is based on 'operator <'
Again, you should understand it in this way: "I only need to implement operator < for objects to be stored in set/map."
Imagine you are going to make the 'struct point' (or 'class point'). We want to intersect some line segments and make a set of
intersection points (sound familiar?). Due to finite computer precision, some points will be the same while their coordinates differ a bit.
That's what you should write:
const double epsilon = 1e‐7;
struct point {
double x, y;
// ...
// Declare operator < taking precision into account
bool operator < (const point& p) const {
if(x < p.x ‐ epsilon) return true;
if(x > p.x + epsilon) return false;
if(y < p.y ‐ epsilon) return true;
if(y > p.y + epsilon) return false;
return false;
}
};
Now you can use set<point> or map<point, string>, for example, to look up whether some point is already present in the list of
intersections. An even more advanced approach: use map<point, vector<int> > and list the list of indices of segments that intersect at
this point.
It's an interesting concept that for STL 'equal' does not mean 'the same', but we will not delve into it here.
Memory management in Vectors
As has been said, vector does not reallocate memory on each push_back(). Indeed, when push_back() is invoked, vector really
allocates more memory than is needed for one additional element. Most STL implementations of vector double in size when
push_back() is invoked and memory is not allocated. This may not be good in practical purposes, because your program may eat up
twice as much memory as you need. There are two easy ways to deal with it, and one complex way to solve it.
The first approach is to use the reserve() member function of vector. This function orders vector to allocate additional memory. Vector
will not enlarge on push_back() operations until the size specified by reserve() will be reached.
Consider the following example. You have a vector of 1,000 elements and its allocated size is 1024. You are going to add 50 elements
to it. If you call push_back() 50 times, the allocated size of vector will be 2048 after this operation. But if you write
v.reserve(1050);
before the series of push_back(), vector will have an allocated size of exactly 1050 elements.
If you are a rapid user of push_back(), then reserve() is your friend.
By the way, it’s a good pattern to use v.reserve() followed by copy(…, back_inserter(v)) for vectors.
Another situation: after some manipulations with vector you have decided that no more adding will occur to it. How do you get rid of
the potential allocation of additional memory? The solution follows:
vector<int> v;
// ...
vector<int>(all(v)).swap(v);
This construction means the following: create a temporary vector with the same content as v, and then swap this temporary vector with
'v'. After the swap the original oversized v will be disposed. But, most likely, you won’t need this during SRMs.
The proper and complex solution is to develop your own allocator for the vector, but that's definitely not a topic for a TopCoder STL
tutorial.
Implementing real algorithms with STL
Armed with STL, let's go on to the most interesting part of this tutorial: how to implement real algorithms efficiently.
Depthfirst search (DFS)
I will not explain the theory of DFS here – instead, read this section of gladius's Introduction to Graphs and Data Structures tutorial –
but I will show you how STL can help.
At first, imagine we have an undirected graph. The simplest way to store a graph in STL is to use the lists of vertices adjacent to each
vertex. This leads to the vector< vector<int> > W structure, where W[i] is a list of vertices adjacent to i. Let’s verify our graph is
connected via DFS:
/*
Reminder from Part 1:
typedef vector<int> vi;
typedef vector<vi> vvi;
*/
int N; // number of vertices
vvi W; // graph
vi V; // V is a visited flag
void dfs(int i) {
if(!V[i]) {
V[i] = true;
for_each(all(W[i]), dfs);
}
}
bool check_graph_connected_dfs() {
int start_vertex = 0;
V = vi(N, false);
dfs(start_vertex);
return (find(all(V), 0) == V.end());
}
That’s all. STL algorithm 'for_each' calls the specified function, 'dfs', for each element in range. In check_graph_connected() function
we first make the Visited array (of correct size and filled with zeroes). After DFS we have either visited all vertices, or not – this is easy
to determine by searching for at least one zero in V, by means of a single call to find().
Notice on for_each: the last argument of this algorithm can be almost anything that “can be called like a function”. It may be not only
global function, but also adapters, standard algorithms, and even member functions. In the last case, you will need mem_fun or
mem_fun_ref adapters, but we will not touch on those now.
One note on this code: I don't recommend the use of vector<bool>. Although in this particular case it’s quite safe, you're better off not
to use it. Use the predefined ‘vi’ (vector<int>). It’s quite OK to assign true and false to int’s in vi. Of course, it requires
8*sizeof(int)=8*4=32 times more memory, but it works well in most cases and is quite fast on TopCoder.
A word on other container types and their usage
Vector is so popular because it's the simplest array container. In most cases you only require the functionality of an array from vector –
but, sometimes, you may need a more advanced container.
It is not good practice to begin investigating the full functionality of some STL container during the heat of a Single Round Match. If you
are not familiar with the container you are about to use, you'd be better off using vector or map/set. For example, stack can always be
implemented via vector, and it’s much faster to act this way if you don’t remember the syntax of stack container.
STL provides the following containers: list, stack, queue, deque, priority_queue. I’ve found list and deque quite useless in SRMs
(except, probably, for very special tasks based on these containers). But queue and priority_queue are worth saying a few words about.
Queue
Queue is a data type that has three operations, all in O(1) amortized: add an element to front (to “head”) remove an element from
back (from “tail”) get the first unfetched element (“tail”) In other words, queue is the FIFO buffer.
Breadthfirst search (BFS)
Again, if you are not familiar with the BFS algorithm, please refer back to this TopCoder tutorial first. Queue is very convenient to use in
BFS, as shown below:
/*
Graph is considered to be stored as adjacent vertices list.
Also we considered graph undirected.
vvi is vector< vector<int> >
W[v] is the list of vertices adjacent to v
*/
int N; // number of vertices
vvi W; // lists of adjacent vertices
bool check_graph_connected_bfs() {
int start_vertex = 0;
vi V(N, false);
queue<int> Q;
Q.push(start_vertex);
V[start_vertex] = true;
while(!Q.empty()) {
int i = Q.front();
// get the tail element from queue
Q.pop();
tr(W[i], it) {
if(!V[*it]) {
V[*it] = true;
Q.push(*it);
}
}
}
return (find(all(V), 0) == V.end());
}
More precisely, queue supports front(), back(), push() (== push_back()), pop (== pop_front()). If you also need push_front() and
pop_back(), use deque. Deque provides the listed operations in O(1) amortized.
There is an interesting application of queue and map when implementing a shortest path search via BFS in a complex graph. Imagine
that we have the graph, vertices of which are referenced by some complex object, like:
pair< pair<int,int>, pair< string, vector< pair<int, int> > > >
(this case is quite usual: complex data structure may define the position in
some game, Rubik’s cube situation, etc…)
Consider we know that the path we are looking for is quite short, and the total number of positions is also small. If all edges of this
graph have the same length of 1, we could use BFS to find a way in this graph. A section of pseudocode follows:
// Some very hard data structure
typedef pair< pair<int,int>, pair< string, vector< pair<int, int> > > > POS;
// ...
int find_shortest_path_length(POS start, POS finish) {
map<POS, int> D;
// shortest path length to this position
queue<POS> Q;
D[start] = 0; // start from here
Q.push(start);
while(!Q.empty()) {
POS current = Q.front();
// Peek the front element
Q.pop(); // remove it from queue
int current_length = D[current];
if(current == finish) {
return D[current];
// shortest path is found, return its length
}
tr(all possible paths from 'current', it) {
if(!D.count(*it)) {
// same as if(D.find(*it) == D.end), see Part I
// This location was not visited yet
D[*it] = current_length + 1;
}
}
}
// Path was not found
return ‐1;
}
// ...
If the edges have different lengths, however, BFS will not work. We should use Dijkstra instead. It's possible to implement such a
Dijkstra via priority_queue see below.
Priority_Queue
Priority queue is the binary heap. It's the data structure, that can perform three operations:
push any element (push)
view top element (top)
pop top element (pop)
For the application of STL's priority_queue see the TrainRobber problem from SRM 307.
Dijkstra
In the last part of this tutorial I’ll describe how to efficiently implement Dijktra’s algorithm in sparse graph using STL containers. Please
look through this tutorial for information on Dijkstra’s algoritm.
Consider we have a weighted directed graph that is stored as vector< vector< pair<int,int> > > G, where
G.size() is the number of vertices in our graph
G[i].size() is the number of vertices directly reachable from vertex with index i
G[i][j].first is the index of jth vertex reachable from vertex i
G[i][j].second is the length of the edge heading from vertex i to vertex G[i][j].first
We assume this, as defined in the following two code snippets:
typedef pair<int,int> ii;
typedef vector<ii> vii;
typedef vector<vii> vvii;
Dijstra via priority_queue
Many thanks to misof for spending the time to explain to me why the complexity of this algorithm is good despite not removing
deprecated entries from the queue.
vi D(N, 987654321);
// distance from start vertex to each vertex
priority_queue<ii,vector<ii>, greater<ii> > Q;
// priority_queue with reverse comparison operator,
// so top() will return the least distance
// initialize the start vertex, suppose it’s zero
D[0] = 0;
Q.push(ii(0,0));
// iterate while queue is not empty
while(!Q.empty()) {
// fetch the nearest element
ii top = Q.top();
Q.pop();
// v is vertex index, d is the distance
int v = top.second, d = top.first;
// this check is very important
// we analyze each vertex only once
// the other occurrences of it on queue (added earlier)
// will have greater distance
if(d <= D[v]) {
// iterate through all outcoming edges from v
tr(G[v], it) {
int v2 = it‐>first, cost = it‐>second;
if(D[v2] > D[v] + cost) {
// update distance if possible
D[v2] = D[v] + cost;
// add the vertex to queue
Q.push(ii(D[v2], v2));
}
}
}
}
I will not comment on the algorithm itself in this tutorial, but you should notice the priority_queue object definition. Normally,
priority_queue<ii> will work, but the top() member function will return the largest element, not the smallest. Yes, one of the easy
solutions I often use is just to store not distance but (distance) in the first element of a pair. But if you want to implement it in the
“proper” way, you need to reverse the comparison operation of priority_queue to reverse one. Comparison function is the third template
parameter of priority_queue while the second paramerer is the storage type for container. So, you should write priority_queue<ii,
vector<ii>, greater<ii> >.
Dijkstra via set
Petr gave me this idea when I asked him about efficient Dijkstra implementation in C#. While implementing Dijkstra we use the
priority_queue to add elements to the “vertices being analyzed” queue in O(logN) and fetch in O(log N). But there is a container besides
priority_queue that can provide us with this functionality it’s ‘set’! I’ve experimented a lot and found that the performance of Dijkstra
based on priority_queue and set is the same.
So, here’s the code:
vi D(N, 987654321);
// start vertex
set<ii> Q;
D[0] = 0;
Q.insert(ii(0,0));
while(!Q.empty()) {
// again, fetch the closest to start element
// from “queue” organized via set
ii top = *Q.begin();
Q.erase(Q.begin());
int v = top.second, d = top.first;
// here we do not need to check whether the distance
// is perfect, because new vertices will always
// add up in proper way in this implementation
tr(G[v], it) {
int v2 = it‐>first, cost = it‐>second;
if(D[v2] > D[v] + cost) {
// this operation can not be done with priority_queue,
// because it does not support DECREASE_KEY
if(D[v2] != 987654321) {
Q.erase(Q.find(ii(D[v2],v2)));
}
D[v2] = D[v] + cost;
Q.insert(ii(D[v2], v2));
}
}
}
One more important thing: STL’s priority_queue does not support the DECREASE_KEY operation. If you will need this operation, ‘set’
may be your best bet.
I’ve spent a lot of time to understand why the code that removes elements from queue (with set) works as fast as the first one.
These two implementations have the same complexity and work in the same time. Also, I’ve set up practical experiments and the
performance is exactly the same (the difference is about ~%0.1 of time).
As for me, I prefer to implement Dijkstra via ‘set’ because with ‘set’ the logic is simpler to understand, and we don’t need to remember
about ‘greater<int>’ predicate overriding.
What is not included in STL
If you have made it this far in the tutorial, I hope you have seen that STL is a very powerful tool, especially for TopCoder SRMs. But
before you embrace STL wholeheartedly, keep in mind what is NOT included in it.
First, STL does not have BigIntegers. If a task in an SRM calls for huge calculations, especially multiplication and division, you have
three options:
use a prewritten template
use Java, if you know it well
say “Well, it was definitely not my SRM!”
I would recommend option number one.
Nearly the same issue arises with the geometry library. STL does not have geometry support, so you have those same three options
again.
The last thing – and sometimes a very annoying thing – is that STL does not have a builtin string splitting function. This is especially
annoying, given that this function is included in the default template for C++ in the ExampleBuilder plugin! But actually I’ve found that
the use of istringstream(s) in trivial cases and sscanf(s.c_str(), …) in complex cases is sufficient.
Those caveats aside, though, I hope you have found this tutorial useful, and I hope you find the STL a useful addition to your use of
C++. Best of luck to you in the Arena!
Note from the author: In both parts of this tutorial I recommend the use of some templates to minimize the time required to implement
something. I must say that this suggestion should always be up to the coder. Aside from whether templates are a good or bad tactic for
SRMs, in everyday life they can become annoying for other people who are trying to understand your code. While I did rely on them for
some time, ultimately I reached the decision to stop. I encourage you to weigh the pros and cons of templates and to consider this
decision for yourself.
Prime Numbers, Factorization and Euler Function
By medv– TopCoder Member
Discuss this article in the forums
In addition to being a TopCoder member, medv is a lecturer in Kiev National University’s cybernetics faculty.
Prime numbers and their properties were extensively studied by the ancient Greek mathematicians. Thousands of years later, we
commonly use the different properties of integers that they discovered to solve problems. In this article we’ll review some definitions,
wellknown theorems, and number properties, and look at some problems associated with them.
A prime number is a positive integer, which is divisible on 1 and itself. The other integers, greater than 1, are composite. Coprime
integers are a set of integers that have no common divisor other than 1 or 1.
The fundamental theorem of arithmetic:
Any positive integer can be divided in primes in essentially only one way. The phrase ‘essentially one way’ means that we do not
consider the order of the factors important.
One is neither a prime nor composite number. One is not composite because it doesn’t have two distinct divisors. If one is prime, then
number 6, for example, has two different representations as a product of prime numbers: 6 = 2 * 3 and 6 = 1 * 2 * 3. This would
contradict the fundamental theorem of arithmetic.
Euclid’s theorem:
There is no largest prime number.
To prove this, let’s consider only n prime numbers: p1, p2, …, pn. But no prime pi divides the number
N = p1 * p2 * … * pn + 1,
so N cannot be composite. This contradicts the fact that the set of primes is finite.
Exercise 1. Sequence an is defined recursively:
Prove that ai and aj, i ¹ j are relatively prime.
Hint: Prove that an+1 = a1a2…an + 1 and use Euclid’s theorem.
Exercise 2. Ferma numbers Fn (n = 0) are positive integers of the form
Prove that Fi and Fj, i ? j are relatively prime.
Hint: Prove that Fn +1 = F0F1F2…Fn + 2 and use Euclid’s theorem.
Dirichlet’s theorem about arithmetic progressions:
For any two positive coprime integers a and b there are infinitely many primes of the form a + n*b, where n > 0.
Trial division:
Trial division is the simplest of all factorization techniques. It represents a bruteforce method, in which we are trying to divide n by
every number i not greater than the square root of n. (Why don’t we need to test values larger than the square root of n?) The
procedure factor prints the factorization of number n. The factors will be printed in a line, separated with one space. The number n can
contain no more than one factor, greater than n.
void factor(int n) {
int i;
for(i=2;i<=(int)sqrt(n);i++) {
while(n % i == 0) {
printf("%d ",i);
n /= i;
}
}
if (n > 1) printf("%d",n);
printf("\n");
}
Consider a problem that asks you to find the factorization of integer g(231 < g <231) in the form
g = f1 x f2 x … x fn or g = 1 x f1 x f2 x … x fn
where fi is a prime greater than 1 and fi = fj for i < j.
For example, for g = 192 the answer is 192 = 1 x 2 x 2 x 2 x 2 x 2 x 2 x 3.
To solve the problem, it is enough to use trial division as shown in function factor.
Sieve of Eratosthenes:
The most efficient way to find all small primes was proposed by the Greek mathematician Eratosthenes. His idea was to make a list of
positive integers not greater than n and sequentially strike out the multiples of primes less than or equal to the square root of n. After
this procedure only primes are left in the list.
The procedure of finding prime numbers gen_primes will use an array primes[MAX] as a list of integers. The elements of this array will
be filled so that
At the beginning we mark all numbers as prime. Then for each prime number i (i = 2), not greater than vMAX, we mark all numbers i*i,
i*(i + 1), … as composite.
void gen_primes() {
int i,j;
for(i=0;i<MAX;i++) primes[i] = 1;
for(i=2;i<=(int)sqrt(MAX);i++)
if (primes[i])
for(j=i;j*i<MAX;j++) primes[i*j] = 0;
}
For example, if MAX = 16, then after calling gen_primes, the array ‘primes’ will contain next values:
i 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
primes[i] 1 1 1 1 0 1 0 1 0 0 0 1 0 1 0 0
Goldbach’s Conjecture:
For any integer n (n = 4) there exist two prime numbers p1 and p2 such that p1 + p2 = n. In a problem we might need to find the
number of essentially different pairs (p1, p2), satisfying the condition in the conjecture for a given even number n (4 = n = 2 15). (The
word ‘essentially’ means that for each pair (p1, p2) we have p1 =p2.)
For example, for n = 10 we have two such pairs: 10 = 5 + 5 and 10 = 3 + 7.
To solve this,as n = 215 = 32768, we’ll fill an array primes[32768] using function gen_primes. We are interested in primes, not greater
than 32768.
The function FindSol(n) finds the number of different pairs (p1, p2), for which n = p1 + p2. As p1 = p2, we have p1 = n/2. So to solve
the problem we need to find the number of pairs (i, n – i), such that i and n – i are prime numbers and 2 = i = n/2.
int FindSol(int n) {
int i,res=0;
for(i=2;i<=n/2;i++)
if (primes[i] && primes[n‐i]) res++;
return res;
}
Euler’s totient function
The number of positive integers, not greater than n, and relatively prime with n, equals to Euler’s totient function f (n). In symbols we
can state that
f (n) ={a Î N: 1 = a = n, gcd(a, n) = 1}
This function has the following properties:
1. If p is prime, then f (p) = p – 1 and f (pa) = p a * (1 – 1/p) for any a.
2. If m and n are coprime, then f (m * n) = f (m) * f (n).
3. If n = , then Euler function can be found using formula:
f (n) = n * (1 – 1/p 1) * (1 – 1/p 2) * … * (1 – 1/p k)
The function fi(n) finds the value of f(n):
int fi(int n) {
int result = n;
for(int i=2;i*i <= n;i++) {
if (n % i == 0) result ‐= result / i;
while (n % i == 0) n /= i;
}
if (n > 1) result ‐= result / n;
return result;
}
For example, to find f(616) we need to factorize the argument: 616 = 23 * 7 * 11. Then, using the formula, we’ll get:
f(616) = 616 * (1 – 1/2) * (1 – 1/7) * (1 – 1/11) = 616 * 1/2 * 6/7 * 10/11 = 240.
Say you’ve got a problem that, for a given integer n (0 < n = 109), asks you to find the number of positive integers less than n and
relatively prime to n. For example, for n = 12 we have 4 such numbers: 1, 5, 7 and 11.
The solution: The number of positive integers less than n and relatively prime to n equals to f(n). In this problem, then, we need do
nothing more than to evaluate Euler’s totient function.
Or consider a scenario where you are asked to calculate a function Answer(x, y), with x and y both integers in the range [1, n], 1 = n =
50000. If you know Answer(x, y), then you can easily derive Answer(k*x, k*y) for any integer k. In this situation you want to know
how many values of Answer(x, y) you need to precalculate. The function Answer is not symmetric.
For example, if n = 4, you need to precalculate 11 values: Answer(1, 1), Answer(1, 2), Answer(2, 1), Answer(1, 3), Answer(2, 3),
Answer(3, 2), Answer(3, 1), Answer(1, 4), Answer(3, 4), Answer(4, 3) and Answer(4, 1).
The solution here is to let res(i) be the minimum number of Answer(x, y) to precalculate, where x, y Î{1, …, i}. It is obvious that res(1)
= 1, because if n = 1, it is enough to know Answer(1, 1). Let we know res(i). So for n = i + 1 we need to find Answer(1, i + 1),
Answer(2, i + 1), … , Answer(i + 1, i + 1), Answer(i + 1, 1), Answer(i + 1, 2), … , Answer(i + 1, i).
The values Answer(j, i + 1) and Answer(i + 1, j), j Î{1, …, i + 1}, can be found from known values if GCD(j, i + 1) > 1, i.e. if the
numbers j and i + 1 are not common primes. So we must know all the values Answer(j, i + 1) and Answer(i + 1, j) for which j and i +
1 are coprime. The number of such values equals to 2 * f (i + 1), where f is an Euler’s totient function. So we have a recursion to solve
a problem:
res(1) = 1,
res(i + 1) = res(i) + 2 * j (i + 1), i > 1
Euler’s totient theorem:
If n is a positive integer and a is coprime to n, then a f (n) º 1 (mod n).
Fermat’s little theorem:
If p is a prime number, then for any integer a that is coprime to n, we have
a p = a (mod p)
This theorem can also be stated as: If p is a prime number and a is coprime to p, then
a p 1 = 1 (mod p)
Fermat’s little theorem is a special case of Euler’s totient theorem when n is prime.
The number of divisors:
If n = , then the number of its positive divisors equals to
(a1 + 1) * (a2 + 1) * … * (ak + 1)
For a proof, let A i be the set of divisors . Any divisor of number n can be represented as a product x1 * x2 * …
* x k , where xi Î Ai. As |Ai| = ai + 1, we have
(a1 + 1) * (a2 + 1) * … * (ak + 1)
possibilities to get different products x1 * x2 * … * xk.
For example, to find the number of divisors for 36, we need to factorize it first: 36 = 2² * 3². Using the formula above, we’ll get the
divisors amount for 36. It equals to (2 + 1) * (2 + 1) = 3 * 3 = 9. There are 9 divisors for 36: 1, 2, 3, 4, 6, 9, 12, 18 and 36.
Here’s another problem to think about: For a given positive integer n (0 < n < 231) we need to find the number of such m that 1 = m
= n, GCD(m, n) ? 1 and GCD(m, n) ? m. For example, for n = 6 we have only one such number m = 4.
The solution is to subtract from n the amount of numbers, coprime with it (its amount equals to f(n)) and the amount of its divisors.
But the number 1 simultaneously is coprime with n and is a divisor of n. So to obtain the difference we must add 1. If n = is
a factorization of n, the number n has (a1 + 1) * (a2 + 1) * … * (ak + 1) divisors. So the answer to the problem for a given n equals to
n – f(n) – (a1 + 1) * (a2 + 1) * … * (ak + 1) + 1
Practice Room:
Want to put some of these theories into practice? Try out these problems, from the TopCoder Archive:
Refactoring (SRM 216)
PrimeAnagrams (SRM 223)
DivisibilityCriteria (SRM 239)
PrimePolynom (SRM 259)
DivisorInc (SRM 302)
PrimePalindromic (SRM 303)
RugSizes (SRM 304)
PowerCollector (SRM 305)
PreprimeNumbers (SRM 307)
EngineersPrimes (SRM 181)
SquareFree (SRM 190)
An Introduction to Recursion, Part 1
By jmzero– TopCoder Member
Discuss this article in the forums
Recursion is a wonderful programming tool. It provides a simple, powerful way of approaching a variety of problems. It is often hard,
however, to see how a problem can be approached recursively; it can be hard to “think” recursively. It is also easy to write a recursive
program that either takes too long to run or doesn’t properly terminate at all. In this article we’ll go over the basics of recursion and
hopefully help you develop, or refine, a very important programming skill.
What is Recursion?
In order to say exactly what recursion is, we first have to answer “What is recursion?” Basically, a function is said to be recursive if it
calls itself. Below is pseudocode for a recursive function that prints the phrase “Hello World” a total of count times:
function HelloWorld(count){
if(count<1)return
print("Hello World!")
HelloWorld(count ‐ 1)
}
It might not be immediately clear what we’re doing here – so let’s follow through what happens if we call our function with count set to
10. Since count is not less than 1, we do nothing on the first line. On the next, we print “Hello World!” once. At this point we need to
print our phrase 9 more times. Since we now have a HelloWorld function that can do just that, we simply call HelloWorld (this time with
count set to 9) to print the remaining copies. That copy of HelloWorld will print the phrase once, and then call another copy of
HelloWorld to print the remaining 8. This will continue until finally we call HelloWorld with count set to zero. HelloWorld(0) does nothing;
it just returns. Once HelloWorld(0) has finished, HelloWorld(1) is done too, and it returns. This continues all the way back to our original
call of HelloWorld(10), which finishes executing having printed out a total of 10 “Hello World!”s.
You may be thinking this is not terribly exciting, but this function demonstrates some key considerations in designing a recursive
algorithm:
1. It handles a simple “base case” without using recursion.
In this example, the base case is “HelloWorld(0)”; if the function is asked to print zero times then it returns without spawning any
more “HelloWorld”s.
2. It avoids cycles.
Imagine if “HelloWorld(10)” called “HelloWorld(10)” which called “HelloWorld(10).” You’d end up with an infinite cycle of calls, and
this usually would result in a “stack overflow” error while running. In many recursive programs, you can avoid cycles by having
each function call be for a problem that is somehow smaller or simpler than the original problem. In this case, for example, count
will be smaller and smaller with each call. As the problem gets simpler and simpler (in this case, we’ll consider it “simpler” to print
something zero times rather than printing it 5 times) eventually it will arrive at the “base case” and stop recursing. There are
many ways to avoid infinite cycles, but making sure that we’re dealing with progressively smaller or simpler problems is a good
rule of thumb.
3. Each call of the function represents a complete handling of the given task.
Sometimes recursion can seem kind of magical in the way it breaks down big problems. However, there is no such thing as a free
lunch. When our function is given an argument of 10, we print “Hello World!” once and then we print it 9 more times. We can pass
a part of the job along to a recursive call, but the original function still has to account for all 10 copies somehow.
Why use Recursion?
The problem we illustrated above is simple, and the solution we wrote works, but we probably would have been better off just using a
loop instead of bothering with recursion. Where recursion tends to shine is in situations where the problem is a little more complex.
Recursion can be applied to pretty much any problem, but there are certain scenarios for which you’ll find it’s particularly helpful. In the
remainder of this article we’ll discuss a few of these scenarios and, along the way, we’ll discuss a few more core ideas to keep in mind
when using recursion.
Scenario #1: Hierarchies, Networks, or Graphs
In algorithm discussion, when we talk about a graph we’re generally not talking about a chart showing the relationship between
variables (like your TopCoder ratings graph, which shows the relationship between time and your rating). Rather, we’re usually talking
about a network of things, people, or concepts that are connected to each other in various ways. For example, a road map could be
thought of as a graph that shows cities and how they’re connected by roads. Graphs can be large, complex, and awkward to deal with
programatically. They’re also very common in algorithm theory and algorithm competitions. Luckily, working with graphs can be made
much simpler using recursion. One common type of a graph is a hierarchy, an example of which is a business’s organization chart:
Name Manager
Betty Sam
Bob Sally
Dilbert Nathan
Joseph Sally
Nathan Veronica
Sally Veronica
Sam Joseph
Susan Bob
Veronica
In this graph, the objects are people, and the connections in the graph show who reports to whom in the company. An upward line on
our graph says that the person lower on the graph reports to the person above them. To the right we see how this structure could be
represented in a database. For each employee we record their name and the name of their manager (and from this information we
could rebuild the whole hierarchy if required – do you see how?).
Now suppose we are given the task of writing a function that looks like “countEmployeesUnder(employeeName)”. This function is
intended to tell us how many employees report (directly or indirectly) to the person named by employeeName. For example, suppose
we’re calling “countEmployeesUnder(‘Sally’)” to find out how many employees report to Sally.
To start off, it’s simple enough to count how many people work directly under her. To do this, we loop through each database record,
and for each employee whose manager is Sally we increment a counter variable. Implementing this approach, our function would return
a count of 2: Bob and Joseph. This is a start, but we also want to count people like Susan or Betty who are lower in the hierarchy but
report to Sally indirectly. This is awkward because when looking at the individual record for Susan, for example, it’s not immediately
clear how Sally is involved.
A good solution, as you might have guessed, is to use recursion. For example, when we encounter Bob’s record in the database we
don’t just increment the counter by one. Instead, we increment by one (to count Bob) and then increment it by the number of people
who report to Bob. How do we find out how many people report to Bob? We use a recursive call to the function we’re writing:
“countEmployeesUnder(‘Bob’)”. Here’s pseudocode for this approach:
function countEmployeesUnder(employeeName){
declare variable counter
counter = 0
for each person in employeeDatabase{
if(person.manager == employeeName){
counter = counter + 1
counter = counter + countEmployeesUnder(person.name)
}
}
return counter
}
If that’s not terribly clear, your best bet is to try following it through linebyline a few times mentally. Remember that each time you
make a recursive call, you get a new copy of all your local variables. This means that there will be a separate copy of counter for each
call. If that wasn’t the case, we’d really mess things up when we set counter to zero at the beginning of the function. As an exercise,
consider how we could change the function to increment a global variable instead. Hint: if we were incrementing a global variable, our
function wouldn’t need to return a value.
Mission Statements
A very important thing to consider when writing a recursive algorithm is to have a clear idea of our function’s “mission statement.” For
example, in this case I’ve assumed that a person shouldn’t be counted as reporting to him or herself. This means
“countEmployeesUnder(‘Betty’)” will return zero. Our function’s mission statment might thus be “Return the count of people who report,
directly or indirectly, to the person named in employeeName – not including the person named employeeName.”
Let’s think through what would have to change in order to make it so a person did count as reporting to him or herself. First off, we’d
need to make it so that if there are no people who report to someone we return one instead of zero. This is simple — we just change
the line “counter = 0″ to “counter = 1″ at the beginning of the function. This makes sense, as our function has to return a value 1
higher than it did before. A call to “countEmployeesUnder(‘Betty’)” will now return 1.
However, we have to be very careful here. We’ve changed our function’s mission statement, and when working with recursion that
means taking a close look at how we’re using the call recursively. For example, “countEmployeesUnder(‘Sam’)” would now give an
incorrect answer of 3. To see why, follow through the code: First, we’ll count Sam as 1 by setting counter to 1. Then when we encounter
Betty we’ll count her as 1. Then we’ll count the employees who report to Betty — and that will return 1 now as well.
It’s clear we’re double counting Betty; our function’s “mission statement” no longer matches how we’re using it. We need to get rid of
the line “counter = counter + 1″, recognizing that the recursive call will now count Betty as “someone who reports to Betty” (and thus
we don’t need to count her before the recursive call).
As our functions get more and more complex, problems with ambiguous “mission statements” become more and more apparent. In
order to make recursion work, we must have a very clear specification of what each function call is doing or else we can end up with
some very difficult to debug errors. Even if time is tight it’s often worth starting out by writing a comment detailing exactly what the
function is supposed to do. Having a clear “mission statement” means that we can be confident our recursive calls will behave as we
expect and the whole picture will come together correctly.
In Part 2, we’ll look at how recursion works with multiple related decisions, such as navigating a maze, and with explicit recursive
relationships.
An Introduction to Recursion: Part 2
By jmzero– TopCoder Member
Scenario #2: Multiple Related Decisions
When our program only has to make one decision, our approach can be fairly simple. We loop through each of the options for our
decision, evaluate each one, and pick the best. If we have two decisions, we can have nest one loop inside the other so that we try
each possible combination of decisions. However, if we have a lot of decisions to make (possibly we don’t even know how many
decisions we’ll need to make), this approach doesn’t hold up.
For example, one very common use of recursion is to solve mazes. In a good maze we have multiple options for which way to go. Each
of those options may lead to new choices, which in turn may lead to new choices as the path continues to branch. In the process of
getting from start to finish, we may have to make a number of related decisions on which way to turn. Instead of making all of these
decisions at once, we can instead make just one decision. For each option we try for the first decision, we then make a recursive call to
try each possibility for all of the remaining decisions. Suppose we have a maze like this:
For this maze, we want to determine the following: is it possible to get from the ‘S’ to the ‘E’ without passing through any ‘*’
characters. The function call we’ll be handling is something like this: “isMazeSolveable(maze[ ][ ])”. Our maze is represented as a 2
dimensional array of characters, looking something like the grid above. Now naturally we’re looking for a recursive solution, and indeed
we see our basic “multiple related decision” pattern here. To solve our maze we’ll try each possible initial decision (in this case we start
at B3, and can go to B2 or B4), and then use recursion to continue exploring each of those initial paths. As we keep recursing we’ll
explore further and further from the start. If the maze is solveable, at some point we’ll reach the ‘E’ at G7. That’s one of our base
cases: if we are asked “can we get from G7 to the end”, we’ll see that we’re already at the end and return true without further
recursion. Alternatively, if we can’t get to the end from either B2 or B4, we’ll know that we can’t get to the end from B3 (our initial
starting point) and thus we’ll return false.
Our first challenge here is the nature of the input we’re dealing with. When we make our recursive call, we’re going to want an easy
way to specify where to start exploring from – but the only parameter we’ve been passed is the maze itself. We could try moving the ‘S’
character around in the maze in order to tell each recursive call where to start. That would work, but would be very slow because in
each call we’d have to first look through the entire maze to find where the ‘S’ is. A better idea would be to find the ‘S’ once, and then
pass around our starting point in separate variables. This happens fairly often when using recursion: we have to use a “starter” function
that will initialize any data and get the parameters in a form that will be easy to work with. Once things are ready, the “starter” function
calls the recursive function that will do the rest of the work. Our starter function here might look something like this:
function isMazeSolveable(maze[][]){
declare variables x,y,startX,startY
startX=‐1
startY=‐1
// Look through grid to find our starting point
for each x from A to H{
for each y from 1 to 8{
if maze[x][y]=='S' then {
startX=x
startY=y
}
}
}
// If we didn't find starting point, maze isn't solveable
if startX==‐1 then return false
// If we did find starting point, start exploring from that point
return exploreMaze(maze[][],startX,startY)
}
We’re now free to write our recursive function exploreMaze. Our mission statement for the function will be “Starting at the position
(X,Y), is it possible to reach the ‘E’ character in the given maze. If the position (x,y) is not traversable, then return false.” Here’s a first
stab at the code:
function exploreMaze(maze[][],x,y){
// If the current position is off the grid, then
// we can't keep going on this path
if y>8 or y<1 or x<'A' or x>'H' then return false
// If the current position is a '*', then we
// can't continue down this path
if maze[x][y]=='*' then return false
// If the current position is an 'E', then
// we're at the end, so the maze is solveable.
if maze[x][y]=='E' then return true
// Otherwise, keep exploring by trying each possible
// next decision from this point. If any of the options
// allow us to solve the maze, then return true. We don't
// have to worry about going off the grid or through a wall ‐
// we can trust our recursive call to handle those possibilities
// correctly.
if exploreMaze(maze,x,y‐1) then return true // search up
if exploreMaze(maze,x,y+1) then return true // search down
if exploreMaze(maze,x‐1,y) then return true // search left
if exploreMaze(maze,x+1,y) then return true // search right
// None of the options worked, so we can't solve the maze
// using this path.
return false
}
Avoiding Cycles
If you’re keen eyed, you likely noticed a flaw in our code above. Consider what happens when we’re exploring from our initial position
of B3. From B3, we’ll try going up first, leading us to explore B2. From there, we’ll try up again and go to B1. B1 won’t work (there’s a
‘*’ there), so that will return false and we’ll be back considering B2. Since up didn’t work, we’ll try down, and thus we’ll consider B3.
And from B3, we’ll consider B2 again. This will continue on until we error out: there’s an infinite cycle.
We’ve forgotten one of our rules of thumb: we need to make sure the problem we’re considering is somehow getting smaller or simpler
with each recursive call. In this case, testing whether we can reach the end from B2 is no simpler than considering whether we can
reach the end from B3. Here we can get a clue from reallife mazes: if you feel like you’ve seen this place before, then you may be
going in circles. We need to revise our mission statement to include “avoid exploring from any position we’ve already considered”. As
the number of places we’ve considered grows, the problem gets simpler and simpler because each decision will have less valid options.
The remaining problem is, then, “how do we keep track of places we’ve already considered?”. A good solution would be to pass around
another 2 dimensional array of true/false values that would contain a “true” for each grid cell we’ve already been to. A quickerand
dirtier way would be to change maze itself, replacing the current position with a ‘*’ just before we make any recursive calls. This way,
when any future path comes back to the point we’re considering, it’ll know that it went in a circle and doesn’t need to continue
exploring. Either way, we need to make sure we mark the current point as visited before we make the recursive calls, as otherwise we
won’t avoid the infinite cycle.
Scenario #3: Explicit Recursive Relationships
You may have heard of the Fibonacci number sequence. This sequence looks like this: 0, 1, 1, 2, 3, 5, 8, 13… After the first two values,
each successive number is the sum of the previous two numbers. We can define the Fibonacci sequence like this:
Fibonacci[0] = 0
Fibonacci[1] = 1
Fibonacci[n] = Fibonacci[n‐2] + Fibonacci[n‐1]
This definition already looks a lot like a recursive function. 0 and 1 are clearly the base cases, and the other possible values can be
handled with recursion. Our function might look like this:
function fib(n){
if(n<1)return 0
if(n==1)return 1
return fib(n‐2) + fib(n‐1)
}
This kind of relationship is very common in mathematics and computer science – and using recursion in your software is a very natural
way to model this kind of relationship or sequence. Looking at the above function, our base cases (0 and 1) are clear, and it’s also clear
that n gets smaller with each call (and thus we shouldn’t have problems with infinite cycles this time).
Using a Memo to Avoid Repetitious Calculation
The above function returns correct answers, but in practice it is extremely slow. To see why, look at what happens if we called “fib(5)”.
To calculate “fib(5)”, we’ll need to calculate “fib(4)” and “fib(3)”. Each of these two calls will make two recursive calls each – and they in
turn will spawn more calls. The whole execution tree will look like this:
The above tree grows exponentially for higher values of n because of the way calls tend to split – and because of the tendency we have
to keep recalculating the same values. In calculating “fib(5)”, we ended up calculating “fib(2)” 3 times. Naturally, it would be better to
only calculate that value once – and then remember that value so that it doesn’t need to be calculated again next time it is asked for.
This is the basic idea of memoization. When we calculate an answer, we’ll store it in an array (named memo for this example) so we can
reuse that answer later. When the function is called, we’ll first check to see if we’ve already got the answer stored in memo, and if we
do we’ll return that value immediately instead of recalculating it.
To start off, we’ll initialize all the values in memo to 1 to mark that they have not been calculated yet. It’s convenient to do this by
making a “starter” function and a recursive function like we did before:
function fib(n){
declare variable i,memo[n]
for each i from 0 to n{
memo[i]=‐1
}
memo[0]=0
memo[1]=1
return calcFibonacci(n,memo)
}
function calcFibonacci(n,memo){
// If we've got the answer in our memo, no need to recalculate
if memo[n]!=‐1 then return memo[n]
// Otherwise, calculate the answer and store it in memo
memo[n] = calcFibonacci(n‐2,memo) + calcFibonacci(n‐1,memo)
// We still need to return the answer we calculated
return memo[n]
}
The execution tree is now much smaller because values that have been calculated already no longer spawn more recursive calls. The
result is that our program will run much faster for larger values of n. If our program is going to calculate a lot of Fibonacci numbers, it
might be best to keep memo somewhere more persistent; that would save us even more calculations on future calls. Also, you might
have noticed another small trick in the above code. Instead of worrying about the base cases inside calcFibonacci, we preloaded values
for those cases into the memo. Preloading base values – especially if there’s a lot of them – can make our recursive functions faster by
allowing us to check base cases and the memo at the same time. The difference is especially noticeable in situations where the base
cases are more numerous or hard to distinguish.
This basic memoization pattern can be one of our best friends in solving TopCoder algorithm problems. Often, using a memo is as
simple as looking at the input parameters, creating a memo array that corresponds to those input parameters, storing calculated values
at the end of the function, and checking the memo as the function starts. Sometimes the input parameters won’t be simple integers
that map easily to a memo array – but by using other objects (like a hash table) for the memo we can continue with the same general
pattern. In general, if you find a recursive solution for a problem, but find that the solution runs too slowly, then the solution is often
memoization.
Conclusion
Recursion is a fundamental programming tool that can serve you well both in TopCoder competitions and “real world” programming. It’s
a subject that many experienced programmers still find threatening, but practice using recursion in TopCoder situations will give you a
great start in thinking recursively, and using recursion to solve complicated programming problems.
An Introduction to Binary Search and RedBlack Trees
By cpphamza– TopCoder Member
Discuss this article in the forums
As a programmer, you’ll frequently come across tasks that deal with a number of objects — numbers, strings, people, and so forth —
and that require you to store and process those objects. If you need to maintain a list of objects that are sorted and unique, and if you
need to be able to quickly insert and retrieve objects to and from this list, the ideal data structure will be a tree set (or a tree map, if
you consider each object a key and associate another object called a value to it).
Many programming languages provide builtin support for treebased sets and maps — for instance, Java’s TreeSet/TreeMap classes and
the C++ Standard Template Library’s set and map classes — but because of their common use, it’s easy to misunderstand what actually
happens behind the scenes, and how the underlying data structures actually work. That’s what this article is all about.
We’ll start off by looking at some of the general terms and concepts used in dealing with trees We’ll then focus on binary search trees
(BST), a special type of tree that keeps elements sorted (but doesn’t guarantee efficient insertion and retrieval). Finally we’ll look at
redblack trees, a variation of binary search trees that overcome BST’s limitations through a logarithmic bound on insertion and
retrieval.
Trees terminology
A tree is a data structure that represents data in a hierarchical manner. It associates every object to a node in the tree and maintains
the parent/child relationships between those nodes. Each tree must have exactly one node, called the root, from which all nodes of the
tree extend (and which has no parent of its own). The other end of the tree – the last level down — contains the leaf nodes of the tree.
The number of lines you pass through when you travel from the root until you reach a particular node is the depth of that node in the
tree (node G in the figure above has a depth of 2). The height of the tree is the maximum depth of any node in the tree (the tree in
Figure 1 has a height of 3). The number of children emanating from a given node is referred to as its degree — for example, node A
above has a degree of 3 and node H has a degree of 1.
Binary Search Tree (BST)
A binary search tree is a tree with one additional constraint — it keeps the elements in the tree in a particular order. Formally each node
in the BST has two children (if any are missing we consider it a nil node), a left child and a right child. Nodes are rooted in place based
on their values, with the smallest on the left and largest on the right.
Traversing BST
A common requirement when using sets and maps is to go through the elements in order. With binary search trees, traversing from left
to right is known as inordertree traversal. In a tree where each node has a value and two pointers to the left and right children,
inorder tree traversal can be thought of as:
Procedure Inorder_traversal(Node n)
if(n == nil)
return;
Inorder_traversal(n.left_subtree);
Print(n.value);
Inorder_traversal(n.right_subtree);
…
Inorder_traversal(root);
Operations on BST (insertion, deletion and retrieval)
BST insertion:
When adding a new node to a binary search tree, note that the new node will always be a leaf in the tree. To insert a new node, all we
will do is navigate the BST starting from the root. If the new node value is smaller than the current node value, we go left – if it is
larger, we go right. When we reach a leaf node, the last step is to attach the new node as a child to this leaf node in a way that
preserves the BST constraint.
For example, consider we want to add a new node with value 4 to the BST in Figure 1. Here are the steps we will go through:
Let the current node = root = 5.
As the new node is smaller than the current node (4 < 5), we will go left and set current node to 2.
As the new node is greater than current node (4 > 2), we will go right and set the current node to 3.
We’ve reached a leaf, so the last step is to attach the new node to the right of the current node. Here is how the new BST looks:
BST Deletion:
Deleting a node from BST is a little more subtle. Formally there are three cases for deleting node n from a BST:
If n has no children, we only have to remove n from the tree.
If n has a single child, we remove n and connect its parent to its child.
If n has two children, we need to :
Find the smallest node that is larger than n, call it m.
Remove m from the tree (if you have reached this case then m will always have no left child, though I’ll leave the proof to the
reader), so we apply one or the other of the above cases to do this.
Replace the value of n with m.
Figure 4 shows these three cases in action.
BST Retrieval:
Retrieving an element from binary search trees requires simple navigation, starting from the root and going left, if the current node is
larger than the node we are looking for, or going right otherwise.
Any of these primitive operations on BST run in O(h) time, where h is the tree height, so the smaller the tree height the better running
time operations will achieve. The problem with BST is that, depending on the order of inserting elements in the tree, the tree shape can
vary. In the worst cases (such as inserting elements in order) the tree will look like a linked list in which each node has only a right
child. This yields O(n) for primitive operations on the BST, with n the number of nodes in the tree.
To solve this problem many variations of binary search trees exist. Of these variations, redblack trees provide a wellbalanced BST that
guarantees a logarithmic bound on primitive operations.
Redblack Trees
Redblack trees are an evolution of binary search trees that aim to keep the tree balanced without affecting the complexity of the
primitive operations. This is done by coloring each node in the tree with either red or black and preserving a set of properties that
guarantee that the deepest path in the tree is not longer than twice the shortest one.
A redblack tree is a binary search tree with the following properties:
1. Every node is colored with either red or black.
2. All leaf (nil) nodes are colored with black; if a node’s child is missing then we will assume that it has a nil child in that place and
this nil child is always colored black.
3. Both children of a red node must be black nodes.
4. Every path from a node n to a descendent leaf has the same number of black nodes (not counting node n). We call this number
the black height of n, which is denoted by bh(n).
Figure 5 shows an example of a redblack tree.
Using these properties, we can show in two steps that a redblack tree which contains n nodes has a height of O(log n), thus all
primitive operations on the tree will be of O(log n) since their order is a function of tree height.
1. First, notice that for a redblack tree with height h, bh(root) is at least h/2 by property 3 above (as each red node strictly requires
black children).
2. The next step is to use the following lemma:
Lemma: A subtree rooted at node v has at least 2^bh(v) – 1 internal nodes
Proof by induction: The basis is when h(v) = 0, which means that v is a leaf node and therefore bh(v) = 0 and the subtree
rooted at node v has 2^bh(v)1 = 2^01 = 11 = 0 nodes.
Inductive hypothesis: if node v1 with height x has 2^bh(v1)1 internal nodes then node v2 with height x+1 has 2^bh(v2)1
For any nonleaf node v (height > 0) we can see that the black height of any of its two children is at least equal to bh(v)1 — if the
child is black, that is, otherwise it is equal to bh(v) . By applying the hypothesis above we conclude that each child has at least
2^[bh(v)1]1 internal nodes, accordingly node v has at least
2^[bh(v)1]1 + 2^[bh(v)1]1 + 1 = 2^bh(v)1
internal nodes, which ends the proof.
By applying the lemma to the root node (with bh of at least h/2, as shown above) we get
n >= 2^(h/2) – 1
where n is the number of internal nodes of a redblack tree (the subtree rooted at the root). Playing with the equation a little bit we get
h <= 2 log (n+1), which guarantees the logarithmic bound of redblack trees.
Rotations
How does inserting or deleting nodes affect a redblack tree? To ensure that its color scheme and properties don’t get thrown off, red
black trees employ a key operation known as rotation. Rotation is a binary operation, between a parent node and one of its children,
that swaps nodes and modifys their pointers while preserving the inorder traversal of the tree (so that elements are still sorted).
There are two types of rotations: left rotation and right rotation. Left rotation swaps the parent node with its right child, while right
rotation swaps the parent node with its left child. Here are the steps involved in for left rotation (for right rotations just change “left” to
“right” below):
Assume node x is the parent and node y is a nonleaf right child.
Let y be the parent and x be its left child.
Let y’s left child be x’s right child.
Operations on redblack tree (insertion, deletion and retrieval)
Redblack tree operations are a modified version of BST operations, with the modifications aiming to preserve the properties of red
black trees while keeping the operations complexity a function of tree height.
Redblack tree insertion:
Inserting a node in a redblack tree is a two step process:
1. A BST insertion, which takes O(log n) as shown before.
2. Fixing any violations to redblack tree properties that may occur after applying step 1. This step is O(log n) also, as we start by
fixing the newly inserted node, continuing up along the path to the root node and fixing nodes along that path. Fixing a node is
done in constant time and involves recoloring some nodes and doing rotations.
Accordingly the total running time of the insertion process is O(log n). Figure 7 shows the redblack tree in figure 5 before and after
insertion of a node with value 4. You can see how the swap operations modified the tree structure to keep it balanced.
Redblack tree deletion:
The same concept behind redblack tree insertions applies here. Removing a node from a redblack tree makes use of the BST deletion
procedure and then restores the redblack tree properties in O(log n). The total running time for the deletion process takes O(log n)
time, then, which meets the complexity requirements for the primitive operations.
Redblack tree retrieval:
Retrieving a node from a redblack tree doesn’t require more than the use of the BST procedure, which takes O(log n) time.
Conclusion
Although you may never need to implement your own set or map classes, thanks to their common builtin support, understanding how
these data structures work should help you better assess the performance of your applications and give you more insight into what
structure is right for a given task. For more practice with these concepts, check out these problems from the TopCoder archive that
involve trees:
MonomorphicTyper (SRM 286)
PendingTasks (TCHS SRM 8)
RedBlack (SRM 155)
DirectoryTree (SRM 168)
EncodingTrees (SRM 261)
AntiChess (SRM 266)
IncompleteBST (SRM 319)
References
“Data Structures via C++: Objects by Evolution,” by A. Michael Berman
“Fundamentals in Data Structures in C++, Second Edition,” by Ellis Horowitz, Sartaj Sahni and Dinesh P. Mehta
“Introduction to Algorithms, Second Edition,” by Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford Stein.
Line Sweep Algorithms
By bmerry– TopCoder Member
Discuss this article in the forums
A previous series of articles covered the basic tools of computational geometry. In this article I’ll explore some more advanced
algorithms that can be built from these basic tools. They are all based on the simple but powerful idea of a sweep line: a vertical line
that is conceptually “swept” across the plane. In practice, of course, we cannot simulate all points in time and so we consider only some
discrete points.
In several places I’ll refer to the Euclidean and Manhattan distances. The Euclidean distance is the normal, everyday distance given by
Pythagoras’ Theorem. The Manhattan distance between points (x1, y1) and (x2, y2) is the distance that must be travelled while moving
only horizontally or vertically, namely |x1 – x2| + |y1 y2|. It is called the Manhattan distance because the roads in Manhattan are laid
out in a grid and so the Manhattan distance is the distance that must be travelled by road (it is also called the “taxicab distance,” or
more formally the L1 metric).
In addition, a balanced binary tree is used in some of the algorithms. Generally you can just use a set in C++ or a TreeSet in Java, but in
some cases this is insufficient because it is necessary to store extra information in the internal nodes.
Closest pair Given a set of points, find the pair that is closest (with either metric). Of course, this can be solved in O(N2) time by
considering all the pairs, but a line sweep can reduce this to O(N log N).
Suppose that we have processed points 1 to N 1 (ordered by X) and the shortest distance we have found so far is h. We now process
point N and try to find a point closer to it than h. We maintain a set of all alreadyprocessed points whose X coordinates are within h of
point N, as shown in the light grey rectangle. As each point is processed, it is added to the set, and when we move on to the next point
or when h is decreased, points are removed from the set. The set is ordered by y coordinate. A balanced binary tree is suitable for this,
and accounts for the log N factor.
To search for points closer than h to point N, we need only consider points in the active set, and furthermore we need only consider
points whose y coordinates are in the range yN h to yN + h (those in the dark grey rectangle). This range can be extracted from the
sorted set in O(log N) time, but more importantly the number of elements is O(1) (the exact maximum will depend on the metric used),
because the separation between any two points in the set is at least h. It follows that the search for each point requires O(log N) time,
giving a total of O(N log N).
Line segment intersections
We’ll start by considering the problem of returning all intersections in a set of horizontal and vertical line segments. Since horizontal
lines don’t have a single X coordinate, we have to abandon the idea of sorting objects by X. Instead, we have the idea of an event: an X
coordinate at which something interesting happens. In this case, the three types of events are: start of a horizontal line, end of a
horizontal line, and a vertical line. As the sweep line moves, we’ll keep an active set of horizontal lines cut by the sweep line, sorted by
Y value (the red lines in the figure).
To handle either of the horizontal line events, we simply need to add or remove an element from the set. Again, we can use a balanced
binary tree to guarantee O(log N) time for these operations. When we hit a vertical line, a range search immediately gives all the
horizontal lines that it cuts. If horizontal or vertical segments can overlap there is some extra work required, and we must also consider
whether lines with coincident endpoints are considered to intersect, but none of this affects the computational complexity.
If the intersections themselves are required, this takes O(N log N + I) time for I intersections. By augmenting the binary tree structure
(specifically, by storing the size of each subtree in the root of that subtree), it is possible to count the intersections in O(N log N) time.
In the more general case, lines need not be horizontal or vertical, so lines in the active set can exchange places when they intersect.
Instead of having all the events presorted, we have to use a priority queue and dynamically add and remove intersection events. At
any point in time, the priority queue contains events for the endpoints of linesegments, but also for the intersection points of adjacent
elements of the active set (providing they are in the future). Since there are O(N + I) events that will be reached, and each requires
O(log N) time to update the active set and the priority queue, this algorithm takes O(N log N + I log N) time. The figure below shows
the future events in the priority queue (blue dots); note that not all future intersections are in the queue, either because one of the
lines isn’t yet active, or because the two lines are not yet adjacent in the active list.
Area of the union of rectangles
Given a set of axisaligned rectangles, what is the area of their union? Like the lineintersection problem, we can handle this by dealing
with events and active sets. Each rectangle has two events: left edge and right edge. When we cross the left edge, the rectangle is
added to the active set. When we cross the right edge, it is removed from the active set.
We now know which rectangles are cut by the sweep line (red in the diagram), but we actually want to know the length of sweep line
that is cut (the total length of the solid blue segments). Multiplying this length by the horizontal distance between events gives the area
swept out between those two events.
We can determine the cut length by running the same algorithm in an inner loop, but rotated 90 degrees. Ignore the inactive
rectangles, and consider a horizontal sweep line that moves topdown. The events are now the horizontal edges of the active
rectangles, and every time we cross one, we can simply increment or decrement a counter that says how many rectangles overlap at
the current point. The cut length increases as long as the counter is nonzero. Of course, we do not increase it continuously, but rather
while moving from one event to the next.
With the right data structures, this can be implemented in O(N2) time (hint: use a boolean array to store the active set rather than a
balanced binary tree, and presort the entire set of horizontal edges). In fact the inner line sweep can be replaced by some clever
binary tree manipulation to reduce the overall time to O(N log N), but that is more a problem in data structures than in geometry, and
is left as an exercise for the reader. The algorithm can also be adapted to answer similar questions, such as the total perimeter length
of the union or the maximum number of rectangles that overlap at any point.
Convex hull
The convex hull of a set of points is the smallest convex polygon that surrounds the entire set, and has a number of practical
applications. An efficient method that is often used in challenges is the Graham scan [2], which requires a sort by angle. This isn’t as
easy as it looks at first, since computing the actual angles is expensive and introduces problems with numeric error. A simpler yet
equally efficient algorithm is due to Andrew [1], and requires only a sort by X for a line sweep (although Andrew’s original paper sorts
by Y and has a few optimizations I won’t discuss here).
Andrew’s algorithm splits the convex hull into two parts, the upper and lower hull. Usually these meet at the ends, but if more than one
points has minimal (or maximal) X coordinate, then they are joined by a vertical line segment. We’ll describe just how to construct the
upper hull; the lower hull can be constructed in similar fashion, and in fact can be built in the same loop.
To build the upper hull, we start with the point with minimal X coordinate, breaking ties by taking the largest Y coordinate. After this,
points are added in order of X coordinate (always taking the largest Y value when multiple points have the same X value). Of course,
sometimes this will cause the hull to become concave instead of convex:
The black path shows the current hull. After adding point 7, we check whether the last triangle (5, 6, 7) is convex. In this case it isn’t,
so we delete the secondlast point, namely 6. The process is repeated until a convex triangle is found. In this case we also examine (4,
5, 7) and delete 5 before examining (1, 4, 7) and finding that it is convex, before proceeding to the next point. This is essentially the
same procedure that is used in the Graham scan, but proceeding in order of X coordinate rather than in order of the angle made with
the starting point. It may at first appear that this process is O(N2) because of the inner backtracking loop, but since no point can be
deleted more than once it is in fact O(N). The algorithm overall is O(N log N), because the points must initially be sorted by X
coordinate.
Manhattan minimum spanning tree We can create even more powerful algorithms by combining a line sweep with a divideand
conquer algorithm. One example is computing the minimum spanning tree of a set of points, where the distance between any pair of
points is the Manhattan distance. This is essentially the algorithm presented by Guibas and Stolfi [3].
We first break this down into a simpler problem. Standard MST algorithms for general graphs (e.g., Prim’s algorithm) can compute the
MST in O((E + N) log N) time for E edges. If we can exploit geometric properties to reduce the number of edges to O(N), then this is
merely O(N log N). In fact we can consider, for each point P, only its nearest neighbors in each of the 8 octants of the plane (see the
figure below). The figure shows the situation in just one of the octants, the WestNorthwest one. Q is the closest neighbour (with the
dashed line indicating points at the same Manhattan distance as Q), and R is some other point in the octant. If PR is an edge in a
spanning tree, then it can be removed and replaced by either PQ or QR to produce a better spanning tree, because the shape of the
octant guarantees that |QR| = |PR|. Thus, we do not need to consider PR when building the spanning tree.
This reduces the problem to that of finding the nearest neighbour in each octant. We’ll just consider the octant shown; the others are
no different and can be handled by symmetry. It should be clear that within this octant, finding the nearest neighbour is equivalent to
just finding the point with the largest value of x y, subject to an upper bound on x + y and a lower bound on y, and this is the form in
which we’ll consider the problem.
Now imagine for the moment that the lower bound on y did not exist. In this case we could solve the problem for every P quite easily:
sweep through the points in increasing order of x + y, and Q will be the point with the largest x y value of those seen so far. This is
where the divideandconquer principle comes into play: we partition the point set into two halves with a horizontal line, and recursively
solve the problem for each half. For points P in the upper half, nothing further needs to be done, because points in the bottom half
cannot play Q to their P. For the bottom half, we have to consider that by ignoring the upper half so far we may have missed some
closer points. However, we can take these points into account in a similar manner as before: walk through all the points in x + y order,
keeping track of the best point in the top half (largest x y value), and for each point in the bottom half, checking whether this best
tophalf point is better than the current neighbour.
So far I have blithely assumed that any set of points can be efficiently partitioned on Y and also walked in x + y order without saying
how this should be done. In fact, one of the most beautiful aspects of this class of divideandconquer plus linesweep algorithms is that
it has essentially the same structure as a merge sort, to the point that a mergesort by x + y can be folded into the algorithm in such a
way that each subset is sorted on x + y just when this is needed (the points initially all being sorted on Y). This gives the algorithm a
running time of O(N log N).
The idea of finding the closest point within an angle range can also be used to solve the Euclidean MST problem, but the O(N log N)
running time is no longer guaranteed in the worst cases, because the distance is no longer a linear equation. It is actually possible to
compute the Euclidean MST in O(N log N) time, because it is a subset of the Delaunay triangulation.
Sample problems
BoxUnion
This is the union of area of rectangles problem above. In this instance there are at most three rectangles which makes simpler
solutions feasible, but you can still use this to practice.
CultureGrowth
While written in a misleading fashion, the task is just to compute the area of the convex hull of a set of points.
PowerSupply
For each power line orientation, sweep the power line in the perpendicular direction. Consumers are added D units ahead of the
sweep and dropped D units behind the sweep. In fact, the low constraints mean that the connected set can be computed from
scratch for each event.
ConvexPolygons
The events of interest are the vertices of the two polygons, and the intersection points of their edges. Between consecutive events,
the section cut by the sweep line varies linearly. Thus, we can sample the cut area at the midpoint X value of each of these
regions to get the average for the whole region. Sampling at these midpoints also eliminates a lot of specialcase handling,
because the sweep line is guaranteed not to pass anywhere near a vertex. Unlike the solution proposed in the match editorial, the
only geometric tool required is lineline intersection.
Conclusion
Like dynamic programming, the sweep line is an extremely powerful tool in an algorithm competitor’s toolkit because it is not simply an
algorithm: it is an algorithm pattern that can be tailored to solve a wide variety of problems, including other textbooks problems that I
have not discussed here (Delaunay triangulations, for example), but also novel problems that may have been created specifically for a
contest. In Topcoder the small constraints often mean that one can take shortcuts (such as processing each event from scratch rather
than incrementally, and in arbitrary order), but the concept of the sweep line is still useful in finding a solution.
References
1. A. M. Andrew. 1979. Another efficient algorithm for convex hulls in two dimensions. Information Processing Letters 9(5) pp 216
219.
2. R. L. Graham. 1972. An efficient algorithm for determining the convex hull of a finite planar set. Information Processing Letters
1(4) pp 132133.
3. Leonidas J. Guibas and Jorge Stolfi. 1983. On computing all northeast nearest neighbors in the L1 metric. Information Processing
Letters 17(4) pp 219223.
Minimum Cost Flow: Part One – Key Concepts
By Zealint– TopCoder Member
Read this article in the forums
This article covers the socalled “mincost flow” problem, which has many applications for both TopCoder competitors and professional
programmers. The article is targeted to readers who are not familiar with the subject, with a focus more on providing a general
understanding of the ideas involved rather than heavy theory or technical details; for a more indepth look at this topic, check out the
references at the end of this article, in particular [1]. In addition, readers of this article should be familiar with the basic concepts of
graph theory – including shortest paths [4], paths with negative cost arcs, negative cycles [1] — and maximum flow theory’s basic
algorithms [3].
The article is divided into three parts. In Part 1, we’ll look at the problem itself. The next part will describe three basic algorithms, and
Part 3 some applications to the problem will be covered in Part 3.
Statement of the Problem
What is the minimum cost flow problem? Let’s begin with some important terminology.
We associate with each vertex a number bi. This value represents supply/demand of the vertex. If bi > 0, node i is a supply
node; if bi < 0, node i is a demand node (its demand is equal to bi). We call vertex i a transshipment if bi is zero.
For simplification, let’s call G a transportation network and write in case we want to show all the network parameters
explicitly.
Figure 1. An example of the transportation network. In this we have 2 supply vertexes (with supply values 5 and 2), 3
demand vertexes (with demand values 1, 4 and 2), and 1 transshipment node. Each edge has two numbers, capacity and
cost, divided by comma.
Representing the flow on arc by xij, we can obtain the optimization model for the minimum cost flow problem:
subject to
The first constraint states that the total outflow of a node minus the total inflow of the node must be equal to mass balance
(supply/demand value) of this node. This is known as the mass balance constraints. Next, the flow bound constraints model
physical capacities or restrictions imposed on the flow’s range. As you can see, this optimization model describes a typical relationship
between warehouses and shops, for example, in a case where we have only one kind of product. We need to satisfy the demand of each
shop by transferring goods from the subset of warehouses, while minimizing the expenses on transportation.
This problem could be solved using simplexmethod, but in this article we concentrate on some other ideas related to network flow
theory. Before we move on to the three basic algorithms used to solve the minimum cost flow problem, let’s review the necessary
theoretical base.
Finding a solution
When does the minimum cost flow problem have a feasible (though not necessarily optimal) solution? How do we determine whether it
is possible to translate the goods or not?
If then the problem has no solution, because either the supply or the demand dominates in the network and the mass
balance constraints come into play.
We can easily avoid this situation, however, if we add a special node r with the supply/demand value . Now we have two options:
If (supply dominates) then for each node with bi > 0 we add an arc with infinite capacity and zero cost; otherwise
(demand dominates), for each node with bi < 0, we add an arc with the same properties. Now we have a new network with
— and it is easy to prove that this network has the same optimal value as the objective function.
Consider the vertex r as a rubbish or scrap dump. If the shops demand is less than what the warehouse supplies, then we have to eject
the useless goods as rubbish. Otherwise, we take the missing goods from the dump. This would be considered shady in real life, of
course, but for our purposes it is very convenient. Keep in mind that, in this case, we cannot say what exactly the “solution” of the
corrected (with scrap) problem is. And it is up to the reader how to classify the emergency uses of the “dump.” For example, we can
suggest that goods remain in the warehouses or some of the shop’s demands remain unsatisfied.
Even if we have we are not sure that the edge’s capacities allow us to transfer enough flow from supply vertexes to demand ones.
To determine if the network has a feasible flow, we want to find any transfer way what will satisfy all the problem’s constraints. Of
course, this feasible solution is not necessarily optimal, but if it is absent we cannot solve the problem.
Figure 2. Maximum flow in the transformed network. For simplicity we are ignoring the costs.
The new network is called a transformed network. Next, we solve a maximum flow problem from s to t (ignoring costs, see fig.2). If
the maximum flow saturates all the source and sink arcs, then the problem has a feasible solution; otherwise, it is infeasible. As for why
this approach works, we’ll leave its proof to the reader.
Having found a maximum flow, we can now remove source, sink, and all adjacent arcs and obtain a feasible flow in G. How do we
detect whether the flow is optimal or not? Does this flow minimize costs of the objective function z? We usually verify “optimality
conditions” for the answer to these questions, but let us put them on hold for a moment and discuss some assumptions.
Now, suppose that we have a network that has a feasible solution. Does it have an optimal solution? If our network contains the
negative cost cycle of infinite capacity then the objective function will be unbounded. However, in some tasks, we are able to assign
finite capacity to each uncapacitated edge escaping such a situation.
So, from the theoretical point of view, for any minimum cost flow problem we have to check some conditions: The supply/demand
balance, the existence of a feasible solution, and the last situation with uncapacitated negative cycles. These are necessary conditions
for resolving the problem. But from the practical point of view, we can check the conditions while the solution is being found.
Assumptions
In understanding the basics of network flow theory it helps to make some assumptions, although sometimes they can lead to a loss of
generality. Of course, we could solve the problems without these assumptions, but the solutions would rapidly become too complex.
Fortunately, these assumptions are not as restrictive as they might seem.
Assumption 1. All data (uij, cij, bi) are integral.
As we have to deal with a computer, which works with rational numbers, this assumption is not restrictive in practice. We can convert
rational numbers to integers by multiplying by a suitable large number.
Assumption 2. The network is directed.
If the network were undirected we would transform it into a directed one. Unfortunately, this transformation requires the edge’s cost to
be nonnegative. Let’s validate this assumption.
To transform an undirected case to a directed one, we replace each undirected edge connecting vertexes i and j by two directed edges
and , both with the capacity and cost of the replaced arc. To establish the correctness of this transformation, first we note
that for undirected arc we have constraint and the term in the objective function. Given that we
see that in some optimal solution either xij or xji will be zero. We call such a solution nonoverlapping. Now it is easy to make sure (and
we leave it to the reader) that every nonoverlapping flow in the original network has an associated flow in the transformed network
with the same cost, and vise versa.
Assumption 3. All costs associated with edges are nonnegative.
This assumption imposes a loss of generality. We will show below that if a network with negative costs had no negative cycle it would be
possible to transform it into one with nonnegative costs. However, one of the algorithms (namely cyclecanceling algorithm) which we
are going to discuss is able to work without this assumption.
For other values of we obtain following result:
For a fixed , the difference is constant. Therefore, a flow that minimizes also minimizes z(x) and vice versa. We
have proved:
The following result contains very useful properties of reduced costs.
Suppose W is a directed cycle. Then for any node potential
Remember this reduced cost technique, since it appears in many applications and other algorithms (for example, Johnson’s algorithm
for all pair shortest path in sparse networks uses it [2]).
Assumption 4. The supply/demand at the vertexes satisfy the condition and the minimum cost flow problem has a feasible
solution.
This assumption is a consequence of the “Finding a Solution” section of this article. If the network doesn’t satisfy the first part of this
assumption, we can either say that the problem has no solution or make corresponding transformation according to the steps outlined
in that section. If the second part of the assumption isn’t met then the solution doesn’t exist.
By making these assumptions we do transform our original transportation network. However, many problems are often given in such a
way which satisfies all the assumptions.
Now that all the preparations are behind us, we can start to discuss the algorithms in Part 2.
References
[1] Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin. Network Flows: Theory, Algorithms, and Applications.
[2] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest. Introduction to Algorithms.
[3] _efer_. Algorithm Tutorial: Maximum Flow.
[4] gladius. Algorithm Tutorial: Introduction to graphs and their data structures: Section 3.
Minimum Cost Flow: Part Two – Algorithms
By Zealint– TopCoder Member
In Part 1, we looked at the basics of minimum cost flow. In this section, we’ll look at three algorithms that can be applied to minimum
cost flow problems.
Working with Residual Networks
Let’s consider the concept of residual networks from the perspective of mincost flow theory. You should be familiar with this concept
thanks to maximum flow theory, so we’ll just extend it to minimum cost flow theory.
We start with the following intuitive idea. Let G be a network and x be a feasible solution of the minimum cost flow problem. Suppose
that an edge (i,j) in E carries xij units of flow. We define the residual capacity of the edge (i,j) as rij = uij – xij. This means that we can
send an additional rij units of flow from vertex i to vertex j. We can also cancel the existing flow xij on the arc if we send up xij units of
flow from j to i over the arc (i,j). Now note that sending a unit of flow from i to j along the arc (i,j) increases the objective function by
cij, while sending a unit of flow from j to i on the same arc decreases the flow cost by cij.
Figure 1. The transportation network from Part 1. (a) A feasible solution. (b) The residual network with respect to the found
feasible solution.
Based on these ideas we define the residual network with respect to the given flow x as follows. Suppose we have a transportation
network G = (V,E). A feasible solution x engenders a new (residual) transportation network, which we are used to defining by Gx =
(V,Ex), where Ex is a set of residual edges corresponding to the feasible solution x.
What is Ex? We replace each arc (i,j) in E by two arcs (i,j), (j,i): the arc (i,j) has cost cij and (residual) capacity rij = uij – xij, and the
arc (j,i) has cost cij and (residual) capacity rji=xij. Then we construct the set Ex from the new edges with a positive residual capacity.
Look at Figure 1 to make sure that you understand the construction of the residual network.
You can notice immediately that such a definition of the residual network has some technical difficulties. Let’s sum them up:
If G contains both the edges (i,j) and (j,i) (remember assumption 2) the residual network may contain four edges between i and j
(two parallel arcs from i to j and two contrary). To avoid this situation we have two options. First, transform the original network to
one in which the network contains either edge (i,j) or edge (j,i), but not both, by splitting the vertexes i and j. Second, represent
our network by the adjacency list, which is handling parallel arcs. We could even use two adjacency matrixes if it were more
convenient.
Let’s imagine now that we have a lot of parallel edges from i to j with different costs. Unfortunately, we can’t merge them by
summarizing their capacities, as we could do while we were finding the maximum flow. So, we need to keep each of the parallel
edges in our data structure separate.
The proof of the fact that there is a onetoone correspondence between the original and residual networks is out the scope of this
article, but you could prove all the necessary theorems as it was done within the maximum flow theory, or by reading [1].
Cyclecanceling Algorithm
This section describes the negative cycle optimality conditions and, as a consequence, cyclecanceling algorithm. We are starting with
this important theorem:
Theorem 1 (Solution Existence). Let G be a transportation network. Suppose that G contains no uncapacitated negative cost cycle
and there exists a feasible solution of the minimum cost flow problem. Then the optimal solution exists.
Proof. One can see that the minimum cost flow problem is a special case of the linear programming problem. The latter is well known to
have an optimal solution if it has a feasible solution and its objective function is bounded. Evidently, if G doesn’t contain an
uncapacitated negative cycle then the objective function of the minimum cost flow problem is bounded from below — therefore, the
assertion of the theorem follows forthwith.
We will use the following theorem without proof, because we don’t want our article to be overloaded with difficult theory, but you can
read the proof in [1].
Theorem 2 (Negative Cycle Optimality Conditions). Let x* be a feasible solution of a minimum cost flow problem. Then x* is an
optimal solution if and only if the residual network Gx* contains no negative cost (directed) cycle.
Figure 2. CycleCanceling Algorithm, example of the network from Figure 1. (a) We have a feasible solution of cost 54. (b) A
negative cycle 1231 is detected in the residual network. Its cost is 1 and capacity is 1. (c) The residual network after
augmentation along the cycle. (d) Another negative cost cycle 3453 is detected. It has cost 2 and capacity 3. (e) The
residual network after augmentation. It doesn’t contain negative cycles. (f) Optimal flow cost value is equal to 47.
This theorem gives the cyclecanceling algorithm for solving the minimum cost flow problem. First, we use any maximum flow algorithm
[3] to establish a feasible flow in the network (remember assumption 4). Then the algorithm attempts to improve the objective function
by finding negative cost cycles in the residual network and augmenting the flow on these cycles. Let us specify a program in pseudo
code like it is done in [1].
Cycle‐Canceling
1 Establish a feasible flow x in the network
2 while ( Gx contains a negative cycle ) do
3 identify a negative cycle W
4
How many iterations does the algorithm perform? First, note that due to assumption 1 all the data is integral. After line 1 of the
program we have an integral feasible solution x. It implies the integrality of Gx. In each iteration of the cycle in line 2 the algorithm
finds the minimum residual capacity in the found negative cycle. In the first iteration will be an integer. Therefore, the modified
residual capacities will be integers, too. And in all subsequent iterations the residual capacities will be integers again. This reasoning
implies:
Theorem 3 (Integrality Property). If all edge capacities and supplies/demands on vertexes are integers, then the minimum cost
flow problem always has an integer solution.
The cyclecanceling algorithm works in cases when the minimum cost flow problem has an optimal solution and all the data is integral
and we don’t need any other assumptions.
Now let us denote the maximum capacity of an arc by U and its maximum absolute value of cost by C. Suppose that m denotes the
number of edges in G and n denotes the number of vertexes. For a minimum cost flow problem, the absolute value of the objective
function is bounded by mCU. Any cycle canceling decreases the objective function by a strictly positive amount. Since we are assuming
that all data is integral, the algorithm terminates within O(mCU) iterations. One can use O(nm) algorithm for identifying a negative
cycle (for instance, BellmanFord’s algorithm or label correcting algorithm [1]), and obtain complexity O(nm2CU) of the algorithm.
Successive Shortest Path Algorithm
The previous algorithm solves the maximum flow problem as a subtask. The successive shortest path algorithm searches for the
maximum flow and optimizes the objective function simultaneously. It solves the socalled maxflowmincost problem by using the
following idea.
Suppose we have a transportation network G and we have to find an optimal flow across it. As it is described in the “Finding a Solution”
section we transform the network by adding two vertexes s and t (source and sink) and some edges as follows. For each node i in V
with bi > 0, we add a source arc (s,i) with capacity bi and cost 0. For each node i in V with bi < 0, we add a sink arc (i,t) with capacity
bi and cost 0.
Then, instead of searching for the maximum flow as usual, we send flow from s to t along the shortest path (with respect to arc costs).
Next we update the residual network, find another shortest path and augment the flow again, etc. The algorithm terminates when the
residual network contains no path from s to t (the flow is maximal). Since the flow is maximal, it corresponds to a feasible solution of
the original minimum cost flow problem. Moreover, this solution will be optimal (and we are going to explain why).
The successive shortest path algorithm can be used when G contains no negative cost cycles. Otherwise, we cannot say exactly what
“the shortest path” means. Now let us justify the successive shortest path approach. When the current flow has zero value, the
transportation network G doesn’t contain a negative cost cycle (by hypothesis). Suppose that after some augmenting steps we have
flow x and Gx still contains no negative cycles. If x is maximal then it is optimal, according to theorem 2. Otherwise, let us denote the
next successfully found shortest path in Gx by P.
Figure 3. How could a negative cycle appear in a residual network?
Suppose that after augmenting the current flow x along path P a negative cost cycle W turned up in the residual network. Before
augmenting there were no negative cycles. This means that there was an edge (i,j) in P (or subpath (i,…,j) in P) the reversal of which
(j,i) closed cycle W after the augmentation. Evidently, we could choose another path from s to t, which goes from s to i then from i to j
along edges of W then from j to t. Moreover, the cost of this path is less than the cost of P. We have a contradiction to the supposition
that P is the shortest.
What do we have? After the last step we have a feasible solution and the residual network contains no negative cycle. The latter is the
criterion of optimality.
A simple analysis shows that the algorithm performs at most O(nB) augmentations, where B is assigned to an upper bound on the
largest supply of any node. Really, each augmentation strictly decreases the residual capacity of a source arc (which is equal to the
supply of the corresponding node). Thanks to the integrality property it decreases by at least one unit. By using an O(nm) algorithm for
finding a shortest path (there may be negative edges), we achieve an O(n2mB) complexity of the successive shortest path algorithm.
Successive Shortest Path
1 Transform network G by adding source and sink
2 Initial flow x is zero
3 while ( Gx contains a path from s to t ) do
4 Find any shortest path P from s to t
5 Augment current flow x along P
6 update Gx
Let us reveal the meaning of node potentials from assumption 3. As it is said within assumption 3, we are able to make all edge costs
nonnegative by using, for instance, BellmanFord’s algorithm. Since working with residual costs doesn’t change shortest paths (by
theorem 2, part 1) we can work with the transformed network and use Dijkstra’s algorithm to find the successive shortest path more
efficiently. However, we need to keep the edge costs nonnegative on each iteration — for this purpose, we update node potentials and
reduce costs right after the shortest path has been found. The reduce cost function could be written in the following manner:
Reduce Cost ( )
1 For each (i,j) in Ex do
2
3
Having found the successive shortest path we need to update node potentials. For each i in V the potential is equal to the length of
the shortest paths from s to t. After having reduced the cost of each arc, we will see that along the shortest path from s to i arcs will
have zero cost while the arcs which lie out of any shortest path to any vertex will have a positive cost. That is why we assign zero cost
to any reversal arc (crev(i,j)) in the Reduce Cost Procedure in line 3. The augmentation (along the found path) adds reversal arc (j,i)
and due to the fact that (reduced) cost cij = 0 we make (crev(i,j)) = 0 beforehand.
Why have we denoted cost of reversal arc by (crev(i,j)) instead of cji? Because the network may contain both arcs (i,j) and (j,i)
(remember assumption 2 and “Working with Residual Networks” section). For other arcs (which lie out of the augmenting path) this
forcible assignment does nothing, because its reversal arcs will not appear in the residual network. Now we propose a pseudocode
program:
Successive Shortest Path with potentials
1 Transform network G by adding source and sink
2 Initial flow x is zero
3 Use Bellman‐Ford's algorithm to establish potentials
4 Reduce Cost ( )
5 while ( Gx contains a path from s to t ) do
6 Find any shortest path P from s to t
7 Reduce Cost ( )
8 Augment current flow x along P
9 update Gx
Before starting the cycle in line 5 we calculate node potentials and obtain all costs to be nonnegative. We use the same massif of
costs c when reducing. In line 6 we use Dijkstra’s algorithm to establish a shortest path with respect to the reduced costs. Then we
reduce costs and augment flow along the path. After the augmentation all costs will remain nonnegative and in the next iteration
Dijkstra’s algorithm will work correctly.
Figure 4. The Successive shortest Path Algorithm. (a) Initial task. (b) Node potentials are calculated after line 3 of the
program. (c) Reduced costs after line 4. (d) The first augmenting path s1234t of capacity 2 is found and new node
potentials are calculated. (e) The residual network with reduced costs. (f) The second augmenting path s134t of capacity
1 is found. (g) The residual network with reduced costs. (h) The third shortest augmenting path s135t and new node
potentials are found. (i) The residual network contains no augmenting paths. (j) The reconstructed transportation network.
Optimal flow has cost 12.
We use BellmanFord’s algorithm only once to avoid negative costs on edges. It takes O(nm) time. Then O(nB) times we use Dijkstra
algorithm, which takes either O(n2) (simple realization) or O(mlogn) (heap realization for sparse network, [4]) time. Summing up, we
receive O(n3B) estimate working time for simple realization and O(nmBlogn) if using heap. One could even use Fibonacci Heaps to
obtain O(nlogn+m) complexity of Dijkstra’s shortest path algorithm; however I wouldn’t recommend doing so because this case works
badly in practice.
PrimalDual Algorithm
The primaldual algorithm for the minimum cost flow problem is similar to the successive shortest path algorithm in the sense that it
also uses node potentials and shortest path algorithm to calculate them. Instead of augmenting the flow along one shortest path,
however, this algorithm increases flow along all the shortest paths at once. For this purpose in each step it uses any maximum flow
algorithm to find the maximum flow through the so called admissible network, which contains only those arcs in Gx with a zero
reduced cost. We represent the admissible residual network with respect to flow x as . Let’s explain the idea by using a pseudocode
program.
Primal‐Dual
1 Transform network G by adding source and sink
2 Initial flow x is zero
3 Use Bellman‐Ford's algorithm to establish potentials
4 Reduce Cost ( )
5 while ( Gx contains a path from s to t ) do
6 Calculate node potential using Dijkstra's algorithm
7 Reduce Cost ( )
8 Establish a maximum flow y from s to t in
9 x x + y
10 update Gx
For a better illustration look at Figure 5.
Figure 5. PrimalDual algorithm. (a) Example network. (b) Node potentials are calculated. (c) The maximum flow in the
admissible network. (d) Residual network and new node potentials. (e) The maximum flow in the admissible network. (f)
Residual network with no augmenting paths. (g) The optimal solution.
As mentioned above, the primaldual algorithm sends flow along all shortest paths at once; therefore, proof of correctness is similar to
the successive shortest path one.
First, the primaldual algorithm guarantees that the number of iterations doesn’t exceed O(nB) as well as the successive shortest path
algorithm. Moreover, since we established a maximum flow in , the residual network Gx contains no directed path from vertex s to
vertex t consisting entirely of arcs of zero costs. Consequently, the distance between s and t increases by at least one unit. These
observations give a bound of min{nB,nC} on the number of iterations which the primaldual algorithm performs. Keep in mind, though,
that the algorithm incurs the additional expense of solving a maximum flow problem at every iteration. However, in practice both the
successive shortest path and the primaldual algorithm work fast enough within the constraint of 50 vertexes and reasonable
supply/demand values and costs.
In the next section, we’ll discuss some applications of the minimum cost flow problem.
References
[1] Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin. Network Flows: Theory, Algorithms, and Applications.
[2] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest. Introduction to Algorithms.
[3] _efer_. Algorithm Tutorial: Maximum Flow.
[4] gladius. Algorithm Tutorial: Introduction to graphs and their data structures: Section 3.
Minimum Cost Flow: Part Three – Applications
By Zealint– TopCoder Member
The last part of the article introduces some well known applications of the minimum cost flow problem. Some of the applications are
described according to [1].
The Assignment Problem
There are a number of agents and a number of tasks. Any agent can be assigned to perform any task, incurring some cost that may
vary depending on the agenttask assignment. We have to get all tasks performed by assigning exactly one agent to each task in such
a way that the total cost of the assignment is minimal with respect to all such assignments.
In other words, consider we have a square matrix with n rows and n columns. Each cell of the matrix contains a number. Let’s denote
by cij the number which lays on the intersection of ith row and jth column of the matrix. The task is to choose a subset of the
numbers from the matrix in such a way that each row and each column has exactly one number chosen and sum of the chosen
numbers is as minimal as possible. For example, assume we had a matrix like this:
In this case, we would chose numbers 3, 4, and 3 with sum 10. In other words, we have to find an integral solution of the following
linear programming problem:
subject to
If binary variable xij = 1 we will choose the number from cell (i,j) of the given matrix. Constraints guarantee that each row and each
column of the matrix will have only one number chosen. Evidently, the problem has a feasible solution (one can choose all diagonal
numbers). To find the optimal solution of the problem we construct the bipartite transportation network as it is drawn in Figure 1. Each
edge (i,j’) of the graph has unit capacity and cost cij. All supplies and demands are equal to 1 and 1 respectively. Implicitly, minimum
cost flow solution corresponds to the optimal assignment and vise versa. Thanks to lefttoright directed edges the network contains no
negative cycles and one is able to solve it with complexity of O(n3). Why? Hint: use the successive shortest path algorithm.
Figure 1. Full weighted bipartite network for the assignment problem. Each edge has capacity 1 and cost according to the
number in the given matrix.
The assignment problem can also be represented as weight matching in a weighted bipartite graph. The problem allows some
extensions:
Suppose that there is a different number of supply and demand nodes. The objective might be to find a maximum matching with a
minimum weight.
Suppose that we have to choose not one but k numbers in each row and each column. We could easily solve this task if we
considered supplies and demands to be equal to k and k (instead of 1 and 1) respectively.
However, we should point out that, due to the specialty of the assignment problem, there are more effective algorithms to solve it. For
instance, the Hungarian algorithm has complexity of O(n3), but it works much more quickly in practice.
Discrete Location Problems
Suppose we have n building sites and we have to build n new facilities on these sites. The new facilities interact with m existing
facilities. The objective is to assign each new facility i to the available building site j in such a way that minimizes the total
transportation cost between the new and existing facilities. One example is the location of hospitals, fire stations etc. in the city; in this
case we can treat population concentrations as the existing facilities.
Let’s denote by dkj the distance between existing facility k and site j; and the total transportation cost per unit distance between the
new facility i and the existing one k by wik. Let’s denote the assignment by binary variable xij. Given an assignment x we can get a
corresponding transportation cost between the new facility i and the existing facility k:
Thus the total transportation cost is given by
Note, that is the cost of locating the new facility i at site j. Appending necessary conditions, we obtain another instance
of the assignment problem.
The Transportation Problem
A minimum cost flow problem is well known to be a transportation problem in the statement of network. But there is a special case of
transportation problem which is called the transportation problem in statement of matrix. We can obtain the optimization model for this
case as follows.
subject to
For example, suppose that we have a set of m warehouses and a set of n shops. Each warehouse i has nonnegative supply value bi
while each shop j has nonnegative demand value dj. We are able to transfer goods from a warehouse i directly to a shop j by the cost
cij per unit of flow.
Figure 2. Formulating the transportation problem as a minimum cost flow problem. Each edge connecting a vertex i and a
vertex j’ has capacity uij and cost cij.
There is an upper bound to the amount of flow between each warehouse i and each shop j denoted by uij. Minimizing total
transportation cost is the object. Representing the flow from a warehouse i to a shop j by xij we obtain the model above. Evidently, the
assignment problem is a special case of the transportation problem in the statement of matrix, which in turn is a special case of the
minimum cost flow problem.
Optimal Loading of a Hopping Airplane
We took this application from [1]. A small commuter airline uses a plane with the capacity to carry at most p passengers on a “hopping
flight.” The hopping flight visits the cities 1, 2, …, n, in a fixed sequence. The plane can pick up passengers at any node and drop them
off at any other node.
Let bij denote the number of passengers available at node i who want to go to node j, and let fij denote the fare per passenger from
node i to node j.
The airline would like to determine the number of passengers that the plane should carry between the various origins and destinations
in order to maximize the total fare per trip while never exceeding the plane capacity.
Figure 3. Formulating the hopping plane flight problem as a minimum cost flow problem.
Figure 3 shows a minimum cost flow formulation of this hopping plane flight problem. The network contains data for only those arcs
with nonzero costs and with finite capacities: Any arc without an associated cost has a zero cost; any arc without an associated capacity
has an infinite capacity.
Consider, for example, node 1. Three types of passengers are available at node 1, those whose destination is node 2, node 3, or node
4. We represent these three types of passengers by the nodes 12, 13, and 14 with supplies b12, b13, and b14. A passenger available
at any such node, say 13, either boards the plane at its origin node by flowing though the arc (13,1) and thus incurring a cost of f13
units, or never boards the plane which we represent by the flow through the arc (13,3).
We invite the reader to establish onetoone correspondence between feasible passenger routings and feasible flows in the minimum
cost flow formulation of the problem.
Dynamic Lot Sizing
Here’s another application that was first outlined in [1]. In the dynamic lotsize problem, we wish to meet prescribed demand dj for
each of K periods j = 1, 2, …, K by either producing an amount aj in period j and/or by drawing upon the inventory Ij1 carried from the
previous period. Figure 4 shows the network for modeling this problem.
The network has K+1 vertexes: The jth vertex, for j = 1, 2, …, K, represents the jth planning period; node 0 represents the “source”
of all production. The flow on the “production arc” (0,j) prescribes the production level aj in period j, and the flow on “inventory
carrying arc” (j,j+1) prescribes the inventory level Ij to be carried from period j to period j+1.
Figure 4. Network flow model of the dynamic lotsize problem.
The mass balance equation for each period j models the basic accounting equation: Incoming inventory plus production in that period
must equal the period’s demand plus the final inventory at the end of the period. The mass balance equation for vertex 0 indicates that
during the planning periods 1, 2, …, K, we must produce all of the demand (we are assuming zero beginning and zero final inventory
over the planning horizon).
If we impose capacities on the production and inventory in each period and suppose that the costs are linear, the problem becomes a
minimum cost flow model.
References
[1] Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin. Network Flows: Theory, Algorithms, and Applications.
[2] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest. Introduction to Algorithms.
[3] _efer_. Algorithm Tutorial: Maximum Flow.
[4] gladius. Algorithm Tutorial: Introduction to graphs and their data structures: Section 3.
Algorithm Games
By rasto6sk– TopCoder Member
Discuss this article in the forums
Introduction
The games we will talk about are twoperson games with perfect information, no chance moves, and a winorlose outcome. In these
games, players usually alternate moves until they reach a terminal position. After that, one player is declared the winner and the other
the loser. Most card games don’t fit this category, for example, because we do not have information about what cards our opponent
has.
First we will look at the basic division of positions to winning and losing. After that we will master the most important game — the
Game of Nim — and see how understanding it will help us to play composite games. We will not be able to play many of the games
without decomposing them to smaller parts (subgames), precomputing some values for them, and then obtaining the result by
combining these values.
The Basics
A simple example is the following game, played by two players who take turns moving. At the beginning there are n coins. When it is a
player’s turn he can take away 1, 3 or 4 coins. The player who takes the last one away is declared the winner (in other words, the
player who can not make a move is the loser). The question is: For what n will the first player win if they both play optimally?
We can see that n = 1, 3, 4 are winning positions for the first player, because he can simply take all the coins. For n=0 there are no
possible moves — the game is finished — so it is the losing position for the first player, because he can not make a move from it. If n=2
the first player has only one option, to remove 1 coin. If n=5 or 6 a player can move to 2 (by removing 3 or 4 coins), and he is in a
winning position. If n=7 a player can move only to 3, 4, 6, but from all of them his opponent can win…
Positions have the following properties:
All terminal positions are losing.
If a player is able to move to a losing position then he is in a winning position.
If a player is able to move only to the winning positions then he is in a losing position.
These properties could be used to create a simple recursive algorithm WLAlgorithm:
boolean isWinning(position pos) {
moves[] = possible positions to which I can move from the
position pos;
for (all x in moves)
if (!isWinning(x)) return true;
return false;
}
Table 1: Game with 11 coins and subtraction set {1, 3, 4}:
n 0 1 2 3 4 5 6 7 8 9 10 11
position L W L W W W W L W L W W
This game could be played also with a rule (usually called the misere play rule) that the player who takes away the last coin is declared
the loser. You need to change only the behavior for the terminal positions in WLAlgorithm. Table 1 will change to this:
n 0 1 2 3 4 5 6 7 8 9 10 11
position W L W L W W W W L W L W
It can be seen that whether a position is winning or losing depends only on the last k positions, where k is the maximum number of
coins we can take away. While there are only 2^k possible values for the sequences of the length k, our sequence will become periodic.
You can try to use this observation to solve the following problem:
SRM 330: LongLongNim
The Game of Nim
The most famous mathematical game is probably the Game of Nim. This is the game that you will probably encounter the most times
and there are many variations on it, as well as games that can be solved by using the knowledge of how to play the game. Usually you
will meet them as Division I 1000 pointers (though hopefully your next encounter will seem much easier). Although these problems
often require a clever idea, they are usually very easy to code.
Rules of the Game of Nim: There are n piles of coins. When it is a player’s turn he chooses one pile and takes at least one coin from
it. If someone is unable to move he loses (so the one who removes the last coin is the winner).
Let n1, n2, … nk, be the sizes of the piles. It is a losing position for the player whose turn it is if and only if n1 xor n2 xor .. xor nk =
0.
How is xor being computed?
6 = (110)2 1 1 0
9 = (1001)2 1 0 0 1
3 = (11)2 1 1
‐‐‐‐‐‐‐‐
1 1 0 0
xor of two logic values is true if and only if one of them is true and the second is false
when computing xor of integers, first write them as binary numbers and then apply xor on columns.
so xor of even number of 1s is 0 and xor of odd number of 1s is 1
Why does it work?
From the losing positions we can move only to the winning ones:
– if xor of the sizes of the piles is 0 then it will be changed after our move (at least one 1 will be changed to 0, so in that column
will be odd number of 1s).
From the winning positions it is possible to move to at least one losing:
– if xor of the sizes of the piles is not 0 we can change it to 0 by finding the left most column where the number of 1s is odd,
changing one of them to 0 and then by changing 0s or 1s on the right side of it to gain even number of 1s in every column.
Examples:
Position (1, 2, 3) is losing because 1 xor 2 xor 3 = (1)2 xor (10)2 xor (11)2 = 0
Position (7, 4, 1) is winning because 7 xor 4 xor 1 = (111)2 xor (10)2 xor (1)2 = (10)2 = 2
Example problems:
SRM 338: CakeParty
SRM 309: StoneGameStrategist
The last one example problem is harder, because it is not so easy to identify where the sizes of piles are hidden. Small hint: Notice the
differences between the sizes of piles. If you would not be able to figure it out you can find the solution in the SRM 309 Problem set &
Analysis.
Composite games – Grundy numbers
Example game: N x N chessboard with K knights on it. Unlike a knight in a traditional game of chess, these can move only as shown
in the picture below (so the sum of coordinates is decreased in every move). There can be more than one knight on the same square at
the same time. Two players take turns moving and, when it is a player’s, turn he chooses one of the knights and moves it. A player who
is not able to make a move is declared the loser.
This is the same as if we had K chessboards with exactly one knight on every chessboard. This is the ordinary sum of K games and it
can be solved by using the grundy numbers. We assign grundy number to every subgame according to which size of the pile in the
Game of Nim it is equivalent to. When we know how to play Nim we will be able to play this game as well.
int grundyNumber(position pos) {
moves[] = possible positions to which I can move from pos
set s;
for (all x in moves) insert into s grundyNumber(x);
//return the smallest non‐negative integer not in the set s;
int ret=0;
while (s.contains(ret)) ret++;
return ret;
}
The following table shows grundy numbers for an 8 x 8 board:
We could try to solve the original problem with our WLAlgorithm, but it would time out because of the large number of possible
positions.
A better approach is to compute grundy numbers for an N x N chessboard in O(n2) time and then xor these K (one for every horse)
values. If their xor is 0 then we are in a losing position, otherwise we are in a winning position.
Why is the pile of Nim equivalent to the subgame if its size is equal to the grundy number of that subgame?
If we decrease the size of the pile in Nim from A to B, we can move also in the subgame to the position with the grundy number B.
(Our current position had grundy number A so it means we could move to positions with all smaller grundy numbers, otherwise the
grundy number of our position would not be A.)
If we are in the subgame at a position with a grundy number higher than 0, by moving in it and decreasing its grundy number we
can also decrease the size of pile in the Nim.
If we are in the subgame at the position with grundy number 0, by moving from that we will get to a position with a grundy
number higher than 0. Because of that, from such a position it is possible to move back to 0. By doing that we can nullify every
move from the position from grundy number 0.
Example problems:
SRM 216: Roxor
Other composite games
It doesn’t happen often, but you can occasionally encounter games with a slightly different set of rules. For example, you might see the
following changes:
1. When it is a player’s move he can choose some of the horses (at least one) and move with all the chosen ones.
Solution: You are in a losing position if and only if every horse is in a losing position on his own chessboard (so the grundy number for
every square, where the horse is, is 0).
2. When it is a player’s move he can choose some of the horses (at least one), but not all of them, and move with all chosen ones.
Solution: You are in a losing position if and only if the grundy numbers of all the positions, where horses are, are the same.
You can verify correctness of both solutions by verifying the basic properties (from a winning position it is possible to move to a losing
one and from a losing position it is possible to move only to the winning ones). Of course, everything works for all other composite
games with these rules (not only for horse games).
Homework: What would be changed if a player had to move with every horse and would lose if he were not able to do so?
Conclusion
Don’t worry if you see a game problem during SRM — it might be similar to one the games described above, or it could be reduced to
one of them. If not, just think about it on concrete examples. Once you figure it out the coding part is usually very simple and
straightforward. Good luck and have fun.
Other resources:
Winning ways for your mathematical plays by Elwyn R. Berlekamp, John H. Conway, Richard K. Guy
Binary Indexed Trees
By boba5551– TopCoder Member
Discuss this article in the forums
Introduction
Notation
Basic idea
Isolating the last digit
Read cumulative frequency
Change frequency at some position and update tree
Read the actual frequency at a position
Scaling the entire tree by a constant factor
Find index with given cumulative frequency
2D BIT
Sample problem
Conclusion
References
Introduction
We often need some sort of data structure to make our algorithms faster. In this article we will discuss the Binary Indexed Trees
structure. According to Peter M. Fenwick, this structure was first used for data compression. Now it is often used for storing frequencies
and manipulating cumulative frequency tables.
Let’s define the following problem: We have n boxes. Possible queries are
1. add marble to box i
2. sum marbles from box k to box l
The naive solution has time complexity of O(1) for query 1 and O(n) for query 2. Suppose we make m queries. The worst case (when
all queries are 2) has time complexity O(n * m). Using some data structure (i.e. RMQ) we can solve this problem with the worst case
time complexity of O(m log n). Another approach is to use Binary Indexed Tree data structure, also with the worst time complexity O(m
log n) — but Binary Indexed Trees are much easier to code, and require less memory space, than RMQ.
Notation
BIT – Binary Indexed Tree
MaxVal – maximum value which will have nonzero frequency
f[i] – frequency of value with index i, i = 1 .. MaxVal
c[i] – cumulative frequency for index i (f[1] + f[2] + … + f[i])
tree[i] – sum of frequencies stored in BIT with index i (latter will be described what index means); sometimes we will write tree
frequency instead sum of frequencies stored in BIT
num?– complement of integer num (integer where each binary digit is inverted: 0 > 1; 1 > 0 )
NOTE: Often we put f[0] = 0, c[0] = 0, tree[0] = 0, so sometimes I will just ignore index 0.
Basic idea
Each integer can be represented as sum of powers of two. In the same way, cumulative frequency can be represented as sum of sets of
subfrequencies. In our case, each set contains some successive number of nonoverlapping frequencies.
idx is some index of BIT. r is a position in idx of the last digit 1 (from left to right) in binary notation. tree[idx] is sum of frequencies
from index (idx – 2^r + 1) to index idx (look at the Table 1.1 for clarification). We also write that idx is responsible for indexes from
(idx 2^r + 1) to idx (note that responsibility is the key in our algorithm and is the way of manipulating the tree).
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
f 1 0 2 1 1 3 0 4 2 5 2 2 3 1 0 2
c 1 1 3 4 5 8 8 12 14 19 21 23 26 27 27 29
tree 1 1 2 4 1 4 0 12 2 7 2 11 3 4 0 29
Table 1.1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
tree 1 1..2 3 1..4 5 5..6 7 1..8 9 9..10 11 9..12 13 13..14 15 1..16
Table 1.2 – table of responsibility
Image 1.3 – tree of responsibility for indexes (bar shows range of frequencies accumulated in top element)
Image 1.4 – tree with tree frequencies
Suppose we are looking for cumulative frequency of index 13 (for the first 13 elements). In binary notation, 13 is equal to 1101.
Accordingly, we will calculate c[1101] = tree[1101] + tree[1100] + tree[1000] (more about this later).
Isolating the last digit
NOTE: Instead of “the last nonzero digit,” it will write only “the last digit.”
There are times when we need to get just the last digit from a binary number, so we need an efficient way to do that. Let num be the
integer whose last digit we want to isolate. In binary notation num can be represented as a1b, where a represents binary digits before
the last digit and b represents zeroes after the last digit.
Integer num is equal to (a1b)¯ + 1 = a¯0b¯ + 1. b consists of all zeroes, so b¯ consists of all ones. Finally we have
num = (a1b)¯ + 1 = a¯0b¯ + 1 = a¯0(0…0)¯ + 1 = a¯0(1…1) + 1 = a¯1(0…0) = a¯1b.
Now, we can easily isolate the last digit, using bitwise operator AND (in C++, Java it is &) with num and num:
a1b
& a¯1b
——————–
= (0…0)1(0…0)
Read cumulative frequency
If we want to read cumulative frequency for some integer idx, we add to sum tree[idx], substract last bit of idx from itself (also we
can write – remove the last digit; change the last digit to zero) and repeat this while idx is greater than zero. We can use next function
(written in C++)
int read(int idx){
int sum = 0;
while (idx > 0){
sum += tree[idx];
idx ‐= (idx & ‐idx);
}
return sum;
}
Example for idx = 13; sum = 0:
1 13 = 1101 0 1 (2 ^0) 3
2 12 = 1100 2 4 (2 ^2) 14
3 8 = 1000 3 8 (2 ^3) 26
4 0 = 0 — — —
Image 1.5 – arrows show path from index to zero which we use to get sum (image shows example for index 13)
So, our result is 26. The number of iterations in this function is number if bits in idx, which is at most log MaxVal.
Time complexity: O(log MaxVal).
Code length: Up to ten lines.
Change frequency at some position and update tree
The concept is to update tree frequency at all indexes which are responsible for frequency whose value we are changing. In reading
cumulative frequency at some index, we were removing the last bit and going on. In changing some frequency val in tree, we should
increment value at the current index (the starting index is always the one whose frequency is changed) for val, add the last digit to
index and go on while the index is less than or equal to MaxVal. Function in C++:
void update(int idx ,int val){
while (idx <= MaxVal){
tree[idx] += val;
idx += (idx & ‐idx);
}
}
Let’s show example for idx = 5:
1 5 = 101 0 1 (2 ^0)
2 6 = 110 1 2 (2 ^1)
3 8 = 1000 3 8 (2 ^3)
4 16 = 10000 4 16 (2 ^4)
5 32 = 100000 — —
Image 1.6 – Updating tree (in brackets are tree frequencies before updating); arrows show path while we update tree from index to
MaxVal (image shows example for index 5)
Using algorithm from above or following arrows shown in Image 1.6 we can update BIT.
Time complexity: O(log MaxVal).
Code length: Up to ten lines.
Read the actual frequency at a position
We’ve described how we can read cumulative frequency for an index. It is obvious that we can not read just tree[idx] to get the actual
frequency for value at index idx. One approach is to have one additional array, in which we will separately store frequencies for values.
Both reading and storing take O(1); memory space is linear. Sometimes it is more important to save memory, so we will show how you
can get actual frequency for some value without using additional structures.
Probably everyone can see that the actual frequency at a position idx can be calculated by calling function read twice – f[idx] =
read(idx) – read(idx – 1) — just by taking the difference of two adjacent cumulative frequencies. This procedure always works in 2 *
O(log n) time. If we write a new function, we can get a bit faster algorithm, with smaller const.
If two paths from two indexes to root have the same part of path, then we can calculate the sum until the paths meet, substract stored
sums and we get a sum of frequencies between that two indexes. It is pretty simple to calculate sum of frequencies between adjacent
indexes, or read the actual frequency at a given index.
Mark given index with x, its predecessor with y. We can represent (binary notation) y as a0b, where b consists of all ones. Then, x will
be a1b¯ (note that b¯ consists all zeros). Using our algorithm for getting sum of some index, let it be x, in first iteration we remove
the last digit, so after the first iteration x will be a0b¯, mark a new value with z.
Repeat the same process with y. Using our function for reading sum we will remove the last digits from the number (one by one). After
several steps, our y will become (just to remind, it was a0b) a0b¯, which is the same as z. Now, we can write our algorithm. Note that
the only exception is when x is equal to 0. Function in C++:
int readSingle(int idx){
int sum = tree[idx]; // sum will be decreased
if (idx > 0){ // special case
int z = idx ‐ (idx & ‐idx); // make z first
idx‐‐; // idx is no important any more, so instead y, you can use idx
while (idx != z){ // at some iteration idx (y) will become z
sum ‐= tree[idx];
// substruct tree frequency which is between y and "the same path"
idx ‐= (idx & ‐idx);
}
}
return sum;
}
Here’s an example for getting the actual frequency for index 12:
First, we will calculate z = 12 – (12 & 12) = 8, sum = 11
1 11 = 1011 0 1 (2 ^0) 9
2 10 = 1010 1 2 (2 ^1) 2
3 8 = 1000 — — —
Image 1.7 – read actual frequency at some index in BIT
(image shows example for index 12)
Let’s compare algorithm for reading actual frequency at some index when we twice use function read and the algorithm written above.
Note that for each odd number, the algorithm will work in const time O(1), without any iteration. For almost every even number idx, it
will work in c * O(log idx), where c is strictly less than 1, compare to read(idx) – read(idx – 1), which will work in c1 * O(log idx),
where c1 is always greater than 1.
Time complexity: c * O(log MaxVal), where c is less than 1.
Code length: Up to fifteen lines.
Scaling the entire tree by a constant factor
Sometimes we want to scale our tree by some factor. With the procedures described above it is very simple. If we want to scale by
some factor c, then each index idx should be updated by (c – 1) * readSingle(idx) / c (because f[idx] – (c – 1) * f[idx] / c =
f[idx] / c). Simple function in C++:
void scale(int c){
for (int i = 1 ; i <= MaxVal ; i++)
update(‐(c ‐ 1) * readSingle(i) / c , i);
}
This can also be done more quickly. Factor is linear operation. Each tree frequency is a linear composition of some frequencies. If we
scale each frequency for some factor, we also scaled tree frequency for the same factor. Instead of rewriting the procedure above,
which has time complexity O(MaxVal * log MaxVal), we can achieve time complexity of O(MaxVal):
void scale(int c){
for (int i = 1 ; i <= MaxVal ; i++)
tree[i] = tree[i] / c;
}
Time complexity: O(MaxVal).
Code length: Just a few lines.
Find index with given cumulative frequency
The naive and most simple solution for finding an index with a given cumultive frequency is just simply iterating through all indexes,
calculating cumulative frequency, and checking if it’s equal to the given value. In case of negative frequencies it is the only solution.
However, if we have only nonnegative frequencies in our tree (that means cumulative frequencies for greater indexes are not smaller)
we can figure out logarithmic algorithm, which is modification of binary search. We go through all bits (starting with the highest one),
make the index, compare the cumulative frequency of the current index and given value and, according to the outcome, take the lower
or higher half of the interval (just like in binary search). Function in C++:
// if in tree exists more than one index with a same
// cumulative frequency, this procedure will return
// some of them (we do not know which one)
// bitMask ‐ initialy, it is the greatest bit of MaxVal
// bitMask store interval which should be searched
int find(int cumFre){
int idx = 0; // this var is result of function
while ((bitMask != 0) && (idx < MaxVal)){ // nobody likes overflow :)
int tIdx = idx + bitMask; // we make midpoint of interval
if (cumFre == tree[tIdx]) // if it is equal, we just return idx
return tIdx;
else if (cumFre > tree[tIdx]){
// if tree frequency "can fit" into cumFre,
// then include it
idx = tIdx; // update index
cumFre ‐= tree[tIdx]; // set frequency for next loop
}
bitMask >>= 1; // half current interval
}
if (cumFre != 0) // maybe given cumulative frequency doesn't exist
return ‐1;
else
return idx;
}
// if in tree exists more than one index with a same
// cumulative frequency, this procedure will return
// the greatest one
int findG(int cumFre){
int idx = 0;
while ((bitMask != 0) && (idx < MaxVal)){
int tIdx = idx + bitMask;
if (cumFre >= tree[tIdx]){
// if current cumulative frequency is equal to cumFre,
// we are still looking for higher index (if exists)
idx = tIdx;
cumFre ‐= tree[tIdx];
}
bitMask >>= 1;
}
if (cumFre != 0)
return ‐1;
else
return idx;
}
Example for cumulative frequency 21 and function find:
First
tIdx is 16; tree[16] is greater than 21; half bitMask and continue
iteration
tIdx is 8; tree[8] is less than 21, so we should include first 8 indexes in result,
remember idx because we surely know it is part of result; subtract tree[8] of
Second
cumFre (we do not want to look for the same cumulative frequency again – we are
iteration
looking for another cumulative frequency in the rest/another part of tree); half
bitMask and contiue
tIdx is 12; tree[12] is greater than 9 (there is no way to overlap interval 18, in
Third
this example, with some further intervals, because only interval 116 can overlap);
iteration
half bitMask and continue
Forth tIdx is 10; tree[10] is less than 9, so we should update values; half bitMask and
iteration continue
Fifth
tIdx is 11; tree[11] is equal to 2; return index (tIdx)
iteration
Time complexity: O(log MaxVal).
Code length: Up to twenty lines.
2D BIT
BIT can be used as a multidimensional data structure. Suppose you have a plane with dots (with nonnegative coordinates). You make
three queries:
1. set dot at (x , y)
2. remove dot from (x , y)
3. count number of dots in rectangle (0 , 0), (x , y) – where (0 , 0) if downleft corner, (x , y) is upright corner and sides are parallel
to xaxis and yaxis.
If m is the number of queries, max_x is maximum x coordinate, and max_y is maximum y coordinate, then the problem should be
solved in O(m * log (max_x) * log (max_y)). In this case, each element of the tree will contain array – (tree[max_x][max_y]).
Updating indexes of xcoordinate is the same as before. For example, suppose we are setting/removing dot (a , b). We will call
update(a , b , 1)/update(a , b , 1), where update is:
void update(int x , int y , int val){
while (x <= max_x){
updatey(x , y , val);
// this function should update array tree[x]
x += (x & ‐x);
}
}
The function updatey is the “same” as function update:
void updatey(int x , int y , int val){
while (y <= max_y){
tree[x][y] += val;
y += (y & ‐y);
}
}
It can be written in one function/procedure:
void update(int x , int y , int val){
int y1;
while (x <= max_x){
y1 = y;
while (y1 <= max_y){
tree[x][y1] += val;
y1 += (y1 & ‐y1);
}
x += (x & ‐x);
}
}
Image 1.8 – BIT is array of arrays, so this is twodimensional BIT (size 16 x 8).
Blue fields are fields which we should update when we are updating index (5 , 3).
The modification for other functions is very similar. Also, note that BIT can be used as an ndimensional data structure.
Sample problem
SRM 310 – FloatingMedian
Problem 2:
Statement:
There is an array of n cards. Each card is putted face down on table. You have two queries:
1. T i j (turn cards from index i to index j, include ith and jth card – card which was face down will be face up; card which was
face up will be face down)
2. Q i (answer 0 if ith card is face down else answer 1)
Solution:
This has solution for each query (and 1 and 2) has time complexity O(log n). In array f (of length n + 1) we will store each query
T (i , j) – we set f[i]++ and f[j + 1]–. For each card k between i and j (include i and j) sum f[1] + f[2] + … + f[k] will be
increased for 1, for all others will be same as before (look at the image 2.0 for clarification), so our solution will be described sum
(which is same as cumulative frequency) module 2.
Image 2.0
Use BIT to store (increase/decrease) frequency and read cumulative frequency.
Conclusion
Binary Indexed Trees are very easy to code.
Each query on Binary Indexed Tree takes constant or logarithmic time.
Binary Indexeds Tree require linear memory space.
You can use it as an ndimensional data structure.
References
[1] RMQ
[2] Binary Search
[3] Peter M. Fenwick
Introduction to String Searching Algorithms
RabinKarp and KnuthMorrisPratt Algorithms
By TheLlama– TopCoder Member
Discuss the article in the forums
The fundamental string searching (matching) problem is defined as follows: given two strings – a text and a pattern, determine
whether the pattern appears in the text. The problem is also known as “the needle in a haystack problem.”
The “Naive” Method
Its idea is straightforward — for every position in the text, consider it a starting position of the pattern and see if you get a match.
function brute_force(text[], pattern[]){
// let n be the size of the text and m the size of the
// pattern
for(i = 0; i < n; i++) {
for(j = 0; j < m && i + j < n; j++)
if(text[i + j] != pattern[j]) break;
// mismatch found, break the inner loop
if(j == m) // match found
}
}
The “naive” approach is easy to understand and implement but it can be too slow in some cases. If the length of the text is n and the
length of the pattern m, in the worst case it may take as much as (n * m) iterations to complete the task.
It should be noted though, that for most practical purposes, which deal with texts based on human languages, this approach is much
faster since the inner loop usually quickly finds a mismatch and breaks. A problem arises when we are faced with different kinds of
“texts,” such as the genetic code.
RabinKarp Algorithm (RK)
This is actually the “naive” approach augmented with a powerful programming technique – the hash function.
Every string s[] of length m can be seen as a number H written in a positional numeral system in base B (B >= size of the alphabet
used in the string):
H = s[0] * B(m – 1) + s[1] * B(m – 2) + … + s[m 2] * B1 + s[m 1] * B0
If we calculate the number H (the hash value) for the pattern and the same number for every substring of length m of the text than the
inner loop of the “naive” method will disappear – instead of comparing two strings character by character we will have just to compare
two integers.
A problem arises when m and B are big enough and the number H becomes too large to fit into the standard integer types. To
overcome this, instead of the number H itself we use its remainder when divided by some other number M. To get the remainder we do
not have to calculate H. Applying the basic rules of modular arithmetic to the above expression:
A + B = C => (A % M + B % M) % M = C % M
A * B = C => ((A % M) * (B % M)) % M = C % M
We get:
H % M = (((s[0] % M) * (B(m – 1) % M)) % M + ((s[1] % M) * (B(m – 2) % M)) % M +…
…+ ((s[m 2] % M) * (B1 % M)) % M + ((s[m 1] % M) * (B0 % M)) % M) % M
The drawback of using remainders is that it may turn out that two different strings map to the same number (it is called a collision).
This is less likely to happen if M is sufficiently large and B and M are prime numbers. Still this does not allow us to entirely skip the
inner loop of the “naive” method. However, its usage is significantly limited. We have to compare the “candidate” substring of the text
with the pattern character by character only when their hash values are equal.
Obviously the approach described so far would be absolutely useless if we were not able to calculate the hash value for every substring
of length m in the text in just one pass through the entire text. At first glance to do these calculations we will again need two nested
loops: an outer one – to iterate through all possible starting positions — and an inner one – to calculate the hash function for every
starting position. Fortunately, this is not the case. Let’s consider a string s[], and let’s suppose we are to calculate the hash value for
every substring in s[] with length say m = 3. It is easy to see that:
H0 = Hs[0]…s[2] = s[0] * B2 + s[1] * B + s[2]
H1 = Hs[1]..s[3] = s[1] * B2 + s[2] * B + s[3]
H1 = (H0 – s[0] * B2 ) * B + s[3]
In general:
Hi = ( Hi – 1 – s[i 1] * Bm 1 ) * B + s[i + m 1]
Applying again the rules of modular arithmetic, we get:
Hi % M = (((( Hi – 1 % M – ((s[i 1] % M) * (Bm – 1 % M)) % M ) % M) * (B % M)) % M +
+ s[i + m 1] % M) % M
Obviously the value of (Hi – 1 – s[i 1] * Bm 1) may be negative. Again, the rules of modular arithmetic come into play:
A – B = C => (A % M – B % M + k * M) % M = C % M
Since the absolute value of (Hi – 1 – s[i 1] * Bm 1) is between 0 and (M – 1), we can safely use a value of 1 for k.
Pseudocode for RK follows:
// correctly calculates a mod b even if a < 0
function int_mod(int a, int b){
return (a % b + b) % b;
}
function Rabin_Karp(text[], pattern[]){
// let n be the size of the text, m the size of the
// pattern, B ‐ the base of the numeral system,
// and M ‐ a big enough prime number
if(n < m) return; // no match is possible
// calculate the hash value of the pattern
hp = 0;
for(i = 0; i < m; i++)
hp = int_mod(hp * B + pattern[i], M);
// calculate the hash value of the first segment
// of the text of length m
ht = 0;
for(i = 0; i < m; i++)
ht = int_mod(ht * B + text[i], M);
if(ht == hp) check character by character if the first
segment of the text matches the pattern;
// start the "rolling hash" ‐ for every next character in
// the text calculate the hash value of the new segment
// of length m; E = (Bm‐1) modulo M
for(i = m; i < n; i++) {
ht = int_mod(ht ‐ int_mod(text[i ‐ m] * E, M), M);
ht = int_mod(ht * B, M);
ht = int_mod(ht + text[i], M);
if(ht == hp) check character by character if the
current segment of the text matches
the pattern;
}
}
Unfortunately, there are still cases when we will have to run the entire inner loop of the “naive” method for every starting position in
the text – for example, when searching for the pattern “aaa” in the string “aaaaaaaaaaaaaaaaaaaaaaaaa” — so in the worst case we
will still need (n * m) iterations. How do we overcome this?
Let’s go back to the basic idea of the method — to replace the string comparison character by character by a comparison of two
integers. In order to keep those integers small enough we have to use modular arithmetic. This causes a “side effect” — the mapping
between strings and integers ceases to be unique. So now whenever the two integers are equal we still have to “confirm” that the two
strings are identical by running characterbycharacter comparison. It can become a kind of vicious circle...
The way to solve this problem is “rational gambling,” or the so called “double hash” technique. We “gamble” — whenever the hash
values of two strings are equal, we assume that the strings are identical, and do not compare them character by character. To make the
likelihood of a “mistake” negligibly small we compute for every string not one but two independent hash values based on different
numbers B and M. If both are equal, we assume that the strings are identical. Sometimes even a “triple hash” is used, but this is rarely
justifiable from a practical point of view.
The “pure” form of “the needle in a haystack problem” is considered too straightforward and is rarely seen in programming challenges.
However, the “rolling hash” technique used in RK is an important weapon. It is especially useful in problems where we have to look at
all substrings of fixed length of a given text. An example is “the longest common substring problem”: given two strings find the longest
string that is a substring of both. In this case, the combination of binary search (BS) and “rolling hash” works quite well. The important
point that allows us to use BS is the fact that if the given strings have a common substring of length n, they also have at least one
common substring of any length m < n. And if the two strings do not have a common substring of length n they do not have a common
substring of any length m > n. So all we need is to run a BS on the length of the string we are looking for. For every substring of the
first string of the length fixed in the BS we insert it in a hash table using one hash value as an index and a second hash value (“double
hash”) is inserted in the table. For every substring of the fixed length of the second string, we calculate the corresponding two hash
values and check in the table to see if they have been already seen in the first string. A hash table based on open addressing is very
suitable for this task.
Of course in “real life” (real challenges) the number of the given strings may be greater than two, and the longest substring we are
looking for should not necessarily be present in all the given strings. This does not change the general approach.
Another type of problems where the “rolling hash” technique is the key to the solution are those that ask us to find the most frequent
substring of a fixed length in a given text. Since the length is already fixed we do not need any BS. We just use a hash table and keep
track of the frequencies.
KnuthMorrisPratt Algorithm (KMP)
In some sense, the “naive” method and its extension RK reflect the standard approach of human logic to “the needle in a haystack
problem”. The basic idea behind KMP is a bit different. Let’s suppose that we are able, after one pass through the text, to identify all
positions where an existing match with the pattern ends. Obviously, this will solve our problem. Since we know the length of the
pattern, we can easily identify the starting position of every match.
Is this approach feasible? It turns out that it is, when we apply the concept of the automaton. We can think of an automaton as of a
kind of abstract object, which can be in a finite number of states. At each step some information is presented to it. Depending on this
information and its current state the automaton goes to a new state, uniquely determined by a set of internal rules. One of the states is
considered as “final”. Every time we reach this “final” state we have found an end position of a match.
The automaton used in KMP is just an array of “pointers” (which represents the “internal rules”) and a separate “external” pointer to
some index of that array (which represents the “current state”). When the next character from the text is presented to the automaton,
the position of the “external” pointer changes according to the incoming character, the current position, and the set of “rules” contained
in the array. Eventually a “final” state is reached and we can declare that we have found a match.
The general idea behind the automaton is relatively simple. Let us consider the string
A B A B A C
as a pattern, and let’s list all its prefixes:
0 /the empty string/
1 A
2 A B
3 A B A
4 A B A B
5 A B A B A
6 A B A B A C
Let us now consider for each such listed string (prefix) the longest proper suffix (a suffix different from the string itself), which is at the
same time a prefix of it:
0 /the empty string/
1 /the empty string/
2 /the empty string/
3 A
4 A B
5 A B A
6 /the empty string/
It’s easy to see that if we have at some point a partial match up to say the prefix (A B A B A) we also have a partial match up to the
prefixes (A B A), and (A) – which are both prefixes of the initial string and suffix/prefixes of the current match. Depending on the next
“incoming” character from the text, three cases arise:
1. The next character is C. We can “expand” the match at the level of the prefix (A B A B A). In this particular case this leads to a full
match and we just notice this fact.
2. The next character is B. The partial match for the prefix (A B A B A) cannot be “expanded”. The best we can do is to return to the
largest different partial match we have so far – the prefix (A B A) and try to “expand” it. Now B “fits” so we continue with the next
character from the text and our current “best” partial match will become the string (A B A B) from our “list of prefixes”.
3. The “incoming” character is, for example, D. The “journey” back to (A B A) is obviously insufficient to “expand” the match. In this
case we have to go further back to the second largest partial match (the second largest proper suffix of the initial match that is at
the same time a prefix of it) – that is (A) and finally to the empty string (the third largest proper suffix in our case). Since it turns
out that there is no way to “expand” even the empty string using the character D, we skip D and go to the next character from the
text. But now our “best” partial match so far will be the empty string.
In order to build the KMP automaton (or the so called KMP “failure function”) we have to initialize an integer array F[]. The indexes
(from 0 to m – the length of the pattern) represent the numbers under which the consecutive prefixes of the pattern are listed in our
“list of prefixes” above. Under each index is a “pointer” – that identifies the index of the longest proper suffix, which is at the same time
a prefix of the given string (or in other words F[i] is the index of next best partial match for the string under index i). In our case (the
string A B A B A C) the array F[] will look as follows:
F[0] = 0
F[1] = 0
F[2] = 0
F[3] = 1
F[4] = 2
F[5] = 3
F[6] = 0
Notice that after initialization F[i] contains information not only about the largest next partial match for the string under index i but also
about every partial match of it. F[i] is the first best partial match, F[F[i]] – is the second best, F[F[F[i]]] – the third, and so on. Using
this information we can calculate F[i] if we know the values F[k] for all k < i. The best next partial match of string i will be the largest
partial match of string i – 1 whose character that “expands” it is equal to the last character of string i. So all we need to do is to check
every partial match of string i – 1 in descending order of length and see if the last character of string i “expands” the match at this
level. If no partial match can be “expanded” than F[i] is the empty string. Otherwise F[i] is the largest “expanded” partial match (after
its “expansion”).
In terms of pseudocode the initialization of the array F[] (the “failure function”) may look like this:
// Pay attention!
// the prefix under index i in the table above is
// is the string from pattern[0] to pattern[i ‐ 1]
// inclusive, so the last character of the string under
// index i is pattern[i ‐ 1]
function build_failure_function(pattern[]){
// let m be the length of the pattern
F[0] = F[1] = 0; // always true
for(i = 2; i <= m; i++) {
// j is the index of the largest next partial match
// (the largest suffix/prefix) of the string under
// index i ‐ 1
j = F[i ‐ 1];
for( ; ; ) {
// check to see if the last character of string i ‐
// ‐ pattern[i ‐ 1] "expands" the current "candidate"
// best partial match ‐ the prefix under index j
if(pattern[j] == pattern[i ‐ 1]) {
F[i] = j + 1; break;
}
// if we cannot "expand" even the empty string
if(j == 0) { F[i] = 0; break; }
// else go to the next best "candidate" partial match
j = F[j];
}
}
}
The automaton consists of the initialized array F[] (“internal rules”) and a pointer to the index of the prefix of the pattern that is the
best (largest) partial match that ends at the current position in the text (“current state”). The use of the automaton is almost identical
to what we did in order to build the “failure function”. We take the next character from the text and try to “expand” the current partial
match. If we fail, we go to the next best partial match of the current partial match and so on. According to the index where this
procedure leads us, the “current state” of the automaton is changed. If we are unable to “expand” even the empty string we just skip
this character, go to the next one in the text, and the “current state” becomes zero.
function Knuth_Morris_Pratt(text[], pattern[]){
// let n be the size of the text, m the
// size of the pattern, and F[] ‐ the
// "failure function"
build_failure_function(pattern[]);
i = 0; // the initial state of the automaton is
// the empty string
j = 0; // the first character of the text
for( ; ; ) {
if(j == n) break; // we reached the end of the text
// if the current character of the text "expands" the
// current match
if(text[j] == pattern[i]) {
i++; // change the state of the automaton
j++; // get the next character from the text
if(i == m) // match found
}
// if the current state is not zero (we have not
// reached the empty string yet) we try to
// "expand" the next best (largest) match
else if(i > 0) i = F[i];
// if we reached the empty string and failed to
// "expand" even it; we go to the next
// character from the text, the state of the
// automaton remains zero
else j++;
}
}
Many problems in programming challenges focus more on the properties of KMP’s “failure function,” rather than on its use for string
matching. An example is: given a string (a quite long one), find all its proper suffixes that are also prefixes of it. All we have to do is
just to calculate the “failure function” of the given string and using the information stored in it to print the answer.
A typical problem seen quite often is: given a string find its shortest substring, such that the concatenation of one or more copies of it
results in the original string. Again the problem can be reduced to the properties of the failure function. Let’s consider the string
A B A B A B
and all its proper suffix/prefixes in descending order:
1 A B A B
2 A B
3 /the empty string/
Every such suffix/prefix uniquely defines a string, which after being “inserted” in front of the given suffix/prefix gives the initial string.
In our case:
1 A B
2 A B A B
3 A B A B A B
Every such “augmenting” string is a potential “candidate” for a string, the concatenation of several copies of which results in the initial
string. This follows from the fact that it is not only a prefix of the initial string but also a prefix of the suffix/prefix it “augments”. But
that means that now the suffix/prefix contains at least two copies of the “augmenting” string as a prefix (since it’s also a prefix of the
initial string) and so on. Of course if the suffix/prefix under question is long enough. In other words, the length of a successful
“candidate” must divide with no remainder the length of the initial string.
So all we have to do in order to solve the given problem is to iterate through all proper suffixes/prefixes of the initial string in
descending order. This is just what the “failure function” is designed for. We iterate until we find an “augmenting” string of the desired
length (its length divides with no remainder the length of the initial string) or get to the empty string, in which case the “augmenting”
string that meets the above requirement will be the initial string itself.
RabinKarp and KnuthMorrisPratt at TopCoder
In the problem types mentioned above, we are dealing with relatively “pure” forms of RK, KMP and the techniques that are the essence
of these algorithms. While you’re unlikely to encounter these pure situations in a TopCoder SRM, the drive towards ever more
challenging TopCoder problems can lead to situations where these algorithms appear as one level in complex, “multilayer” problems.
The specific input size limitations favor this trend, since we will not be presented as input with multimillion character strings, but rather
with a “generator”, which may be by itself algorithmic in nature. A good example is “InfiniteSoup,” Division 1 – Level Three, SRM 286.
Maximum Flow: Augmenting Path Algorithms Comparison
By Zealint– TopCoder Member
Discuss this article in the forums
With this article, we’ll revisit the socalled “maxflow” problem, with the goal of making some practical analysis of the most famous
augmenting path algorithms. We will discuss several algorithms with different complexity from O(nm2) to O(nmlogU) and reveal the
most efficient one in practice. As we will see, theoretical complexity is not a good indicator of the actual value of an algorithm.
The article is targeted to the readers who are familiar with the basics of network flow theory. If not, I’ll refer them to check out [1], [2]
or algorithm tutorial [5].
In the first section we remind some necessary definitions and statements of the maximum flow theory. Other sections discuss the
augmenting path algorithms themselves. The last section shows results of a practical analysis and highlights the best in practice
algorithm. Also we give a simple implementation of one of the algorithms.
Statement of the Problem
Suppose we have a directed network G = (V, E) defined by a set V of nodes (or vertexes) and a set E of arcs (or edges). Each arc (i,j)
in E has an associated nonnegative capacity uij. Also we distinguish two special nodes in G: a source node s and a sink node t. For each
i in V we denote by E(i) all the arcs emanating from node i. Let U = max uij by (i,j) in E. Let us also denote the number of vertexes by n
and the number of edges by m.
We wish to find the maximum flow from the source node s to the sink node t that satisfies the arc capacities and mass balance
constraints at all nodes. Representing the flow on arc (i,j) in E by xij we can obtain the optimization model for the maximum flow
problem:
Vector (xij) which satisfies all constraints is called a feasible solution or, a flow (it is not necessary maximal). Given a flow x we are able
to construct the residual network with respect to this flow according to the following intuitive idea. Suppose that an edge (i,j) in E
carries xij units of flow. We define the residual capacity of the edge (i,j) as rij = uij – xij. This means that we can send an additional rij
units of flow from vertex i to vertex j. We can also cancel the existing flow xij on the arc if we send up xij units of flow from j to i over
the arc (i,j).
So, given a feasible flow x we define the residual network with respect to the flow x as follows. Suppose we have a network G = (V, E).
A feasible solution x engenders a new (residual) network, which we define by Gx = (V, Ex), where Ex is a set of residual edges
corresponding to the feasible solution x.
What is Ex? We replace each arc (i,j) in E by two arcs (i,j), (j,i): the arc (i,j) has (residual) capacity rij = uij – xij, and the arc (j,i) has
(residual) capacity rji=xij. Then we construct the set Ex from the new edges with a positive residual capacity.
Augmenting Path Algorithms as a whole
In this section we describe one method on which all augmenting path algorithms are being based. This method was developed by Ford
and Fulkerson in 1956 [3]. We start with some important definitions.
Augmenting path is a directed path from a source node s to a sink node t in the residual network. The residual capacity of an
augmenting path is the minimum residual capacity of any arc in the path. Obviously, we can send additional flow from the source to the
sink along an augmenting path.
All augmenting path algorithms are being constructed on the following basic idea known as augmenting path theorem:
Theorem 1 (Augmenting Path Theorem). A flow x* is a maximum flow if and only if the residual network Gx* contains no
augmenting path.
According to the theorem we obtain a method of finding a maximal flow. The method proceeds by identifying augmenting paths and
augmenting flows on these paths until the network contains no such path. All algorithms that we wish to discuss differ only in the way
of finding augmenting paths.
We consider the maximum flow problem subject to the following assumptions.
Assumption 1. The network is directed.
Assumption 2. All capacities are nonnegative integers.
This assumption is not necessary for some algorithms, but the algorithms whose complexity bounds involve U assume the integrality of
the data.
Assumption 3. The problem has a bounded optimal solution.
This assumption in particular means that there are no uncapacitated paths from the source to the sink.
Assumption 4. The network does not contain parallel arcs.
This assumption imposes no loss of generality, because one can summarize capacities of all parallel arcs.
As to why these assumptions are correct we leave the proof to the reader.
It is easy to determine that the method described above works correctly. Under assumption 2, on each augmenting step we increase
the flow value by at least one unit. We (usually) start with zero flow. The maximum flow value is bounded from above, according to
assumption 3. This reasoning implies the finiteness of the method.
With those preparations behind us, we are ready to begin discussing the algorithms.
Shortest Augmenting Path Algorithm, O(n2m)
In 1972 Edmonds and Karp — and, in 1970, Dinic — independently proved that if each augmenting path is shortest one, the algorithm
will perform O(nm) augmentation steps. The shortest path (length of each edge is equal to one) can be found with the help of breadth
first search (BFS) algorithm [2], [6]. Shortest Augmenting Path Algorithm is well known and widely discussed in many books and
articles, including [5], which is why we will not describe it in great detail. Let’s review the idea using a kind of pseudocode:
In line 5, current flow x is being increased by some positive amount.
The algorithm was said to perform O(nm) steps of finding an augmenting path. Using BFS, which requires O(m) operation in the worst
case, one can obtain O(nm2) complexity of the algorithm itself. If m ~ n2 then one must use BFS procedure O(n3) times in worst case.
There are some networks on which this numbers of augmentation steps is being achieved. We will show one simple example below.
Improved Shortest Augmenting Path Algorithm, O(n2m)
As mentioned earlier, the natural approach for finding any shortest augmenting path would be to look for paths by performing a
breadthfirst search in the residual network. It requires O(m) operations in the worst case and imposes O(nm2) complexity of the
maximum flow algorithm. Ahuja and Orlin improved the shortest augmenting path algorithm in 1987 [1]. They exploited the fact that
the minimum distance from any node i to the sink node t is monotonically nondecreasing over all augmentations and reduced the
average time per augmentation to O(n). The improved version of the augmenting path algorithm, then, runs in O(n2m) time. We can
now start discussing it according to [1].
Definition 1. Distance function d: V_ Z+ with respect to the residual capacities rij is a function from the set of nodes to nonnegative
integers. Let’s say that distance function is valid if it is satisfies the following conditions:
d(t)=0;
d(i) ≤ d(j) + 1, for every (i,j) in E with rij>0.
Informally (and it is easy to prove), valid distance label of node i, represented by d(i), is a lower bound on the length of the shortest
path from i to t in the residual network Gx. We call distance function exact if each i in V d(i) equals the length of the shortest path from
i to t in the residual network. It is also easy to prove that if d(s) ≥ n then the residual network contains no path from the source to the
sink.
An arc (i,j) in E is called admissible if d(i) = d(j) + 1. We call other arcs inadmissible. If a path from s to t consists of admissible arcs
then the path is admissible. Evidently, an admissible path is the shortest path from the source to the sink. As far as every arc in an
admissible path satisfies condition rij>0, the path is augmenting.
So, the improved shortest augmenting path algorithm consists of four steps (procedures): main cycle, advance, retreat and augment.
The algorithm maintains a partial admissible path, i.e., a path from s to some node i, consisting of admissible arcs. It performs advance
or retreat steps from the last node of the partial admissible path (such node is called current node). If there is some admissible arc (i,j)
from current node i, then the algorithm performs the advance step and adds the arc to the partial admissible path. Otherwise, it
performs the retreat step, which increases distance label of i and backtracks by one arc.
If the partial admissible path reaches the sink, we perform an augmentation. The algorithm stops when d(s) ≥ n. Let’s describe these
steps in pseudocode [1]. We denoted residual (with respect to flow x) arcs emanating from node i by Ex(i). More formally, Ex(i) = {
(i,j) in E(i): rij > 0 }.
In line 1 of retreat procedure if Ex(i) is empty, then suppose d(i) equals n.
Ahuja and Orlin suggest the following data structure for this algorithm [1]. We maintain the arc list E(i) which contains all the arcs
emanating from node i. We arrange the arcs in this list in any fixed order. Each node i has a current arc, which is an arc in E(i) and is
the next candidate for admissibility testing. Initially, the current arc of node i is the first arc in E(i). In line 5 the algorithm tests whether
the node’s current arc is admissible. If not, it designates the next arc in the list as the current arc. The algorithm repeats this process
until either it finds an admissible arc or reaches the end of the arc list. In the latter case the algorithm declares that E(i) contains no
admissible arc; it again designates the first arc in E(i) as the current arc of node i and performs the relabel operation by calling the
retreat procedure (line 10).
Now we outline a proof that the algorithm runs in O(n2m) time.
Lemma 1. The algorithm maintains distance labels at each step. Moreover, each relabel (or, retreat) step strictly increases the distance
label of a node.
Sketch to proof. Perform induction on the number of relabel operation and augmentations.
Lemma 2. Distance label of each node increases at most n times. Consecutively, relabel operation performs at most n2 times.
Proof. This lemma is consequence of lemma 1 and the fact that if d(s) ≥ n then the residual network contains no augmenting path.
Since the improved shortest augmenting path algorithm makes augmentations along the shortest paths (like unimproved one), the total
number of augmentations is the same O(nm). Each retreat step relabels a node, that is why number of retreat steps is O(n2) (according
to lemma 2). Time to perform retreat/relabel steps is O( n ∑i in V |E(i)| ) = O(nm). Since one augmentation requires O(n) time, total
augmentation time is O(n2m). The total time of advance steps is bounded by the augmentation time plus the retreat/relabel time and it
is again O(n2m). We obtain the following result:
Theorem 2. The improved shortest augmenting path algorithm runs in O(n2m) time.
Ahuja and Orlin [1] suggest one very useful practical improvement of the algorithm. Since the algorithm performs many useless relabel
operations while the maximum flow has been found, it will be better to give an additional criteria of terminating. Let’s introduce (n+1)
dimensional additional array, numbs, whose indices vary from 0 to n. The value numbs(k) is the number of nodes whose distance label
equals k. The algorithm initializes this array while computing the initial distance labels using BFS. At this point, the positive entries in
the array numbs are consecutive (i.e., numbs(0), numbs(1), …, numbs(l) will be positive up to some index l and the remaining entries
will all be zero).
When the algorithm increases a distance label of a node from x to y, it subtracts 1 from numbs(x), adds 1 to numbs(y) and checks
whether numbs(x) = 0. If it does equal 0, the algorithm terminates.
This approach is some kind of heuristic, but it is really good in practice. As to why this approach works we leave the proof to the reader
(hint: show that the nodes i with d(i) > x and nodes j with d(j) < x engender a cut and use maximumflowminimumcut theorem).
Comparison of Improved and Unimproved versions
In this section we identify the worst case for both shortest augmenting path algorithms with the purpose of comparing their running
times.
In the worst case both improved and unimproved algorithms will perform O(n3) augmentations, if m ~ n2. Norman Zadeh [4]
developed some examples on which this running time is based. Using his ideas we compose a somewhat simpler network on which the
algorithms have to perform O(n3) augmentations and which is not dependent on a choice of next path.
Figure 1. Worst case example for the shortest augmenting path algorithm.
All vertexes except s and t are divided into four subsets: S={s1,…,sk}, T={t1,…,tk}, U={u1,…,u2p} and V={v1,…,v2p}. Both sets S
and T contain k nodes while both sets U and V contain 2p nodes. k and p are fixed integers. Each bold arc (connecting S and T) has unit
capacity. Each dotted arc has an infinite capacity. Other arcs (which are solid and not straight) have capacity k.
First, the shortest augmenting path algorithm has to augment flow k2 time along paths (s, S, T, t) which have length equal to 3. The
capacities of these paths are unit. After that the residual network will contain reversal arcs (T, S) and the algorithm will chose another
k2 augmenting paths (s, u1, u2, T, S, v2, v1, t) of length 7. Then the algorithm will have to choose paths (s, u1, u2, u3, u4, S, T, v4,
v3, v2, v1, t) of length 11 and so on…
Now let’s calculate the parameters of our network. The number of vertexes is n = 2k + 4p + 2. The number of edges is m = k2 + 2pk +
2k + 4p. As it easy to see, the number of augmentations is a = k2 (p+1).
Consider that p = k – 1. In this case n = 6k – 2 and a = k3. So, one can verify that a ~ n3 / 216. In [4] Zadeh presents examples of
networks that require n3 / 27 and n3 / 12 augmentations, but these examples are dependent on a choice of the shortest path.
We made 5 worstcase tests with 100, 148, 202, 250 and 298 vertexes and compared the running times of the improved version of the
algorithm against the unimproved one. As you can see on figure 2, the improved algorithm is much faster. On the network with 298
nodes it works 23 times faster. Practice analysis shows us that, in general, the improved algorithm works n / 14 times faster.
Figure 2. Xaxis is the number of nodes. Yaxis is working time in milliseconds.
Blue colour indicates the shortest augmenting path algorithm and red does it improved version.
However, our comparison in not definitive, because we used only one kind of networks. We just wanted to justify that the O(n2m)
algorithm works O(n) times faster than the O(nm2) on a dense network. A more revealing comparison is waiting for us at the end of the
article.
Maximum Capacity Path Algorithm, O(n2mlognU) / O(m2 lognU logn) / O(m2 lognU logU)
In 1972 Edmonds and Karp developed another way to find an augmenting path. At each step they tried to increase the flow with the
maximum possible amount. Another name of this algorithm is “gradient modification of the FordFulkerson method.” Instead of using
BFS to identify a shortest path, this modification uses Dijkstra’s algorithm to establish a path with the maximal possible capacity. After
augmentation, the algorithm finds another such path in the residual network, augments flow along it, and repeats these steps until the
flow is maximal.
There’s no doubt that the algorithm is correct in case of integral capacity. However, there are tests with nonintegral arc’s capacities on
which the algorithm may fail to terminate.
Let’s get the algorithm’s running time bound by starting with one lemma. To understand the proof one should remember that the value
of any flow is less than or equal to the capacity of any cut in a network, or read this proof in [1], [2]. Let’s denote capacity of a cut
(S,T) by c(S,T).
Lemma 3. Let F be the maximum flow’s value, then G contains augmenting path with capacity not less than F/m.
Proof. Suppose G contains no such path. Let’s construct a set E’={ (i,j) in E: uij ≥ F/m }. Consider network G’ = (V, E’) which has no
path from s to t. Let S be a set of nodes obtainable from s in G and T = V S. Evidently, (S, T) is a cut and c(S, T) ≥ F. But cut (S, T)
intersects only those edges (i,j) in E which have uij < F/m. So, it is clear that
c(S,T) < (F/m) _ m = F,
and we got a contradiction with the fact that c(S,T) ≥ F.
Theorem 3. The maximum capacity path algorithm performs O(m log (nU)) augmentations.
Sketch to proof. Suppose that the algorithm terminates after k augmentations. Let’s denote by f1 the capacity of the first found
augmentation path, by f2 the capacity of the second one, and so on. fk will be the capacity of the latter kth augmenting path.
Consider, Fi = f1 + f2 +…+ fi. Let F* be the maximum flow’s value. Under lemma 3 one can justify that
fi ≥ (F*Fi1) / m.
Now we can estimate the difference between the value of the maximal flow and the flow after i consecutive augmentations:
F* – Fi = F* Fi1 – fi ≤ F* – Fi1 – (F* Fi1) / m = (1 – 1 / m) (F* – Fi1) ≤ … ≤ (1 – 1 / m)i _ F*
We have to find such an integer i, which gives (1 – 1 / m)i _ F* < 1. One can check that
i log m / (m+1) F* = O(m _ log F*) = O(m _ log(nU))
And the latter inequality proves the theorem.
To find a path with the maximal capacity we use Dijkstra’s algorithm, which incurs additional expense at every iteration. Since a simple
realization of Dijkstras’s algorithm [2] incurs O(n2) complexity, the total running time of the maximum capacity path algorithm is
O(n2mlog(nU)).
Using a heap implementation of Dijkstra’s algorithm for sparse network [7] with running time O(mlogn), one can obtain an O(m2 logn
log(nU)) algorithm for finding the maximum flow. It seems to be better that the improved EdmondsKarp algorithm. However, this
estimate is very deceptive.
There is another variant to find the maximum capacity path. One can use binary search to establish such a path. Let’s start by finding
the maximum capacity path on piece [0,U]. If there is some path with capacity U/2, then we continue finding the path on piece [U/2,
U]; otherwise, we try to find the path on [0,U/21]. This approach incurs additional O(mlogU) expense and gives O(m2log(nU)logU)
time bound to the maximum flow algorithm. However, it works really poorly in practice.
Capacity Scaling Algorithm, O(m2logU)
In 1985 Gabow described the socalled “bitscaling” algorithm. The similar capacity scaling algorithm described in this section is due to
Ahuja and Orlin [1].
Informally, the main idea of the algorithm is to augment the flow along paths with sufficient large capacities, instead of augmenting
along maximal capacities. More formally, let’s introduce a parameter Delta. First, Delta is quite a large number that, for instance, equals
U. The algorithm tries to find an augmenting path with capacity not less that Delta, then augments flow along this path and repeats this
procedure while any such Deltapath exists in the residual network.
The algorithm either establishes a maximum flow or reduces Delta by a factor of 2 and continues finding paths and augmenting flow
with the new Delta. The phase of the algorithm that augments flow along paths with capacities at least Delta is called Deltascaling
phase or, Deltaphase. Delta is an integral value and, evidently, algorithm performs O(logU) Deltaphases. When Delta is equal to 1
there is no difference between the capacity scaling algorithm and the EdmondsKarp algorithm, which is why the algorithm works
correctly.
We can obtain a path with the capacity of at least Delta fairly easily in O(m) time (by using BFS). At the first phase we can set Delta
to equal either U or the largest power of 2 that doesn’t exceed U.
The proof of the following lemma is left to the reader.
Lemma 4. At every Deltaphase the algorithm performs O(m) augmentations in worst case.
Sketch to proof. Use the induction by Delta to justify that the minimum cut at each Deltascaling phase less that 2m Delta.
Applying lemma 4 yields the following result:
Theorem 4. Running time of the capacity scaling algorithm is O(m2logU).
Keep in mind that there is no difference between using breadthfirst search and depthfirst search when finding an augmenting path.
However, in practice, there is a big difference, and we will see it.
Improved Capacity Scaling Algorithm, O(nmlogU)
In the previous section we described an O(m2logU) algorithm for finding the maximum flow. We are going to improve the running time
of this algorithm to O(nmlogU) [1].
Now let’s look at each Deltaphase independently. Recall from the preceding section that a Deltascaling phase contains O(m)
augmentations. Now we use the similar technique at each Deltaphase that we used when describing the improved variant of the
shortest augmenting path algorithm. At every phase we have to find the “maximum” flow by using only paths with capacities equal to
at least Delta. The complexity analysis of the improved shortest augmenting path algorithm implies that if the algorithm is guaranteed
to perform O(m) augmentations, it would run in O(nm) time because the time for augmentations reduces from O(n2m) to O(nm) and
all other operations, as before, require O(nm) time. These reasoning instantly yield a bound of O(nmlogU) on the running time of the
improved capacity scaling algorithm.
Unfortunately, this improvement hardly decreases the running time of the algorithm in practice.
Practical Analysis and Comparison
Now let’s have some fun. In this section we compare all described algorithms from practical point of view. With this purpose I have
made some test cases with the help of [8] and divided them into three groups by density. The first group of tests is made of networks
with m ≤ n1.4 – some kind of sparse networks. The second one consists of middle density tests with n1.6 ≤ m ≤ n1.7. And the last
group represents some kinds of almost full graphs (including full acyclic networks) with m ≥ n1.85.
I have simple implementations of all algorithms described before. I realized these algorithms without any kind of tricks and with no
preference towards any of them. All implementations use adjacency list for representing a network.
Let’s start with the first group of tests. These are 564 sparse networks with number of vertexes limited by 2000 (otherwise, all
algorithms work too fast). All working times are given in milliseconds.
Figure 3. Comparison on sparse networks. 564 test cases. m ≤ n1.4.
As you can see, it was a big mistake to try Dijkstra’s without heap implementation of the maximum capacity path algorithm on sparse
networks (and it’s not surprising); however, its heap implementation works rather faster than expected. Both the capacity scaling
algorithms (with using DFS and BFS) work in approximately the same time, while the improved implementation is almost 2 times faster.
Surprisingly, the improved shortest path algorithm turned out to be the fastest on sparse networks.
Now let’s look at the second group of test cases. It is made of 184 tests with middle density. All networks are limited to 400 nodes.
Figure 4. Comparison on middle density networks. 184 test cases. n1.6 ≤ m ≤ n1.7.
On the middle density networks the binary search implementation of the maximum capacity path algorithm leaves much to be desired,
but the heap implementation still works faster than the simple (without heap) one. The BFS realization of the capacity scaling algorithm
is faster than the DFS one. The improved scaling algorithm and the improved shortest augmenting path algorithm are both very good in
this case.
It is very interesting to see how these algorithms run on dense networks. Let’s take a look — the third group is made up of 200 dense
networks limited by 400 vertexes.
Figure 5. Comparison on dense networks. 200 test cases. m ≥ n1.85.
Now we see the difference between the BFS and DFS versions of the capacity scaling algorithm. It is interesting that the improved
realization works in approximately the same time as the unimproved one. Unexpectedly, the Dijkstra’s with heap implementation of the
maximum capacity path algorithm turned out to be faster than one without heap.
Without any doubt, the improved implementation of EdmondsKarp algorithm wins the game. Second place is taken by the improved
scaling capacity algorithm. And the scaling capacity with BFS got bronze.
As to maximum capacity path, it is better to use one variant with heap; on sparse networks it gives very good results. Other algorithms
are really only good for theoretical interest.
As you can see, the O(nmlogU) algorithm isn’t so fast. It is even slower than the O(n2m) algorithm. The O(nm2) algorithm (it is the
most popular) has worse time bounds, but it works much faster than most of the other algorithms with better time bounds.
My recommendation: Always use the scaling capacity path algorithm with BFS, because it is very easy to implement. The improved
shortest augmenting path algorithm is rather easy, too, but you need to be very careful to write the program correctly. During the
challenge it is very easy to miss a bug.
I would like to finish the article with the full implementation of the improved shortest augmenting path algorithm. To maintain a
network I use the adjacency matrix with purpose to providing best understanding of the algorithm. It is not the same realization what
was used during our practical analysis. With the “help” of the matrix it works a little slower than one that uses adjacency list. However,
it works faster on dense networks, and it is up to the reader which data structure is best for them.
#include
#define N 2007 // Number of nodes
#define oo 1000000000 // Infinity
// Nodes, Arcs, the source node and the sink node
int n, m, source, sink;
// Matrixes for maintaining
// Graph and Flow
int G[N][N], F[N][N];
int pi[N]; // predecessor list
int CurrentNode[N]; // Current edge for each node
int queue[N]; // Queue for reverse BFS
int d[N]; // Distance function
int numbs[N]; // numbs[k] is the number of nodes i with d[i]==k
// Reverse breadth‐first search
// to establish distance function d
int rev_BFS() {
int i, j, head(0), tail(0);
// Initially, all d[i]=n
for (i = 1; i <= n; i++)
numbs[d[i] = n] ++;
// Start from the sink
numbs[n]‐‐;
d[sink] = 0;
numbs[0]++;
queue[++tail] = sink;
// While queue is not empty
while (head != tail) {
i = queue[++head]; // Get the next node
// Check all adjacent nodes
for (j = 1; j <= n; j++) {
// If it was reached before or there is no edge
// then continue
if (d[j] < n || G[j][i] == 0) continue;
// j is reached first time
// put it into queue
queue[++tail] = j;
// Update distance function
numbs[n]‐‐;
d[j] = d[i] + 1;
numbs[d[j]]++;
}
}
return 0;
}
// Augmenting the flow using predecessor list pi[]
int Augment() {
int i, j, tmp, width(oo);
// Find the capacity of the path
for (i = sink, j = pi[i]; i != source; i = j, j = pi[j]) {
tmp = G[j][i];
if (tmp < width) width = tmp;
}
// Augmentation itself
for (i = sink, j = pi[i]; i != source; i = j, j = pi[j]) {
G[j][i] ‐= width; F[j][i] += width;
G[i][j] += width; F[i][j] ‐= width;
}
return width;
}
// Relabel and backtrack
int Retreat(int &i) {
int tmp;
int j, mind(n ‐ 1);
// Check all adjacent edges
// to find nearest
for (j = 1; j <= n; j++)
// If there is an arc
// and j is "nearer"
if (G[i][j] > 0 & #038;& d[j] < mind)
mind = d[j];
tmp = d[i]; // Save previous distance
// Relabel procedure itself
numbs[d[i]]‐‐;
d[i] = 1 + mind;
numbs[d[i]]++;
// Backtrack, if possible (i is not a local variable! )
if (i != source) i = pi[i];
// If numbs[ tmp ] is zero, algorithm will stop
return numbs[tmp];
}
// Main procedure
int find_max_flow() {
int flow(0), i, j;
rev_BFS(); // Establish exact distance function
// For each node current arc is the first arc
for (i = 1; i <= n; i++) CurrentNode[i] = 1;
// Begin searching from the source
i = source;
// The main cycle (while the source is not "far" from the sink)
for (; d[source] < n; ) {
// Start searching an admissible arc from the current arc
for (j = CurrentNode[i]; j <= n; j++)
// If the arc exists in the residual network
// and if it is an admissible
if (G[i][j] > 0 & #038;& d[i] == d[j] + 1)
// Then finish searhing
break;
// If the admissible arc is found
if (j <= n) {
CurrentNode[i] = j; // Mark the arc as "current"
pi[j] = i; // j is reachable from i
i = j; // Go forward
// If we found an augmenting path
if (i == sink) {
flow += Augment(); // Augment the flow
i = source; // Begin from the source again
}
}
// If no an admissible arc found
else {
CurrentNode[i] = 1; // Current arc is the first arc again
// If numbs[ d[i] ] == 0 then the flow is the maximal
if (Retreat(i) == 0)
break;
}
} // End of the main cycle
// We return flow value
return flow;
}
// The main function
// Graph is represented in input as triples
// No comments here
int main() {
int i, p, q, r;
scanf("%d %d %d %d", &n, &m, &source, &sink);
for (i = 0; i < m; i++) {
scanf("%d %d %d", &p, &q, &r);
G[p][q] += r;
}
printf("%d", find_max_flow());
return 0;
}
References
[1] Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin. Network Flows: Theory, Algorithms, and Applications.
[2] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest. Introduction to Algorithms.
[3] Ford, L. R., and D. R. Fulkerson. Maximal flow through a network.
[4] Norman Zadeh. Theoretical Efficiency of the EdmondsKarp Algorithm for Computing Maximal Flows.
[5] _efer_. Algorithm Tutorial: MaximumFlow.
[6] gladius. Algorithm Tutorial: Introduction to graphs and their data structures: Section 1.
[7] gladius. Algorithm Tutorial: Introduction to graphs and their data structures: Section 3.
[8] http://elib.zib.de/pub/mptestdata/generators/index.html A number of generators for network flow problems.
Basics of Combinatorics
By xray– TopCoder Member
Discuss this article in the forums
Introduction
Counting the objects that satisfy some criteria is a very common task in both TopCoder problems and in reallife situations. The myriad
ways of counting the number of elements in a set is one of the main tasks in combinatorics, and I’ll try to describe some basic aspects
of it in this tutorial. These methods are used in a range of applications, from discrete math and probability theory to statistics, physics,
biology, and more.
Combinatorial primitives
Let’s begin with a quick overview of the basic rules and objects that we will reference later.
The rule of sum
The rule of product
For example, if we have three towns — A, B and C — and there are 3 roads from A to B and 5 roads from B to C, then we can get from
A to C through B in 3*5=15 different ways.
These rules can be used for a finite collections of sets.
Permutation without repetition
When we choose k objects from nelement set in such a way that the order matters and each object can be chosen only once:
For example, suppose we are planning the next 3 challenges and we have a set of 10 easy problems to choose from. We will only use
one easy problem in each contest, so we can choose our problems in different ways.
Permutation (variation) with repetition
The number of possible choices of k objects from a set of n objects when order is important and one object can be chosen more than
once:
nk
For example, if we have 10 different prizes that need to be divided among 5 people, we can do so in 510 ways.
Permutation with repetition
The number of different permutations of n objects, where there are n1 indistinguishable objects of type 1, n2 indistinguishable objects
of type 2,…, and nk indistinguishable objects of type k (n1+n2+…+nk=n), is:
For example, if we have 97 coders and want to assign them to 5 rooms (rooms 14 have 20 coders in each, while the 5th room has 17),
then there are possible ways to do it.
Combinations without repetition
In combinations we choose a set of elements (rather than an arrangement, as in permutations) so the order doesn’t matter. The
number of different kelement subsets (when each element can be chosen only once) of nelement set is:
For example, if we have 7 different colored balls, we can choose any 3 of them in different ways.
Combination with repetition
Let’s say we choose k elements from an nelement set, the order doesn’t matter and each element can be chosen more than once. In
that case, the number of different combinations is:
For example, let’s say we have 11 identical balls and 3 different pockets, and we need to calculate the number of different divisions of
these balls to the pockets. There would be different combinations.
It is useful to know that is also the number of integer solutions to this equation:
Why? It’s easy to prove. Consider a vector (1, 1, …, 1) consisting of (n+k‐1) ones, in which we want to substitute n‐1 ones for zeroes in
such way that we’ll get n groups of ones (some of which may be empty) and the number of ones in the ith group will be the value of xi:
The sum of xi will be k, because k ones are left after substitution.
The Basics
Binary vectors
Some problems, and challenge problems are no exception, can be reformulated in terms of binary vectors. Accordingly, some
knowledge of the basic combinatorial properties of binary vectors is rather important. Let’s have a look at some simple things
associated with them:
1. Number of binary vectors of length n: 2n.
2. Number of binary vectors of length n and with k ‘1’ is
We just choose k positions for our ‘1’s.
3. The number of ordered pairs (a, b) of binary vectors, such that the distance between them (k) can be calculated as follows: .
The distance between a and b is the number of components that differs in a and b — for example, the distance between (0, 0, 1, 0) and
(1, 0, 1, 1) is 2).
Let a = (a1, a2, …an), b = (b1, b2, …bn) and distance between them is k. Next, let’s look at the sequence of pairs (a1, b1), (a2, b2), …
(an, bn). There are exactly k indices i in which ai ≠ bi. They can be (0,1) or (1,0), so there are 2 variants, and nk can be either (0,0) or
(1,1), for another 2 variants. To calculate the answer we can choose k indices in which vectors differs in ways, then we choose
components that differs in 2k ways and components that are equal in 2nk ways (all of which use the permutation with repetition
formula), and in the end we just multiply all these numbers and get .
Delving deeper
Now let’s take a look at a very interesting and useful formula called the inclusionexclusion principle (also known as the sieve principle):
This formula is a generalization of:
There are many different problems that can be solved using the sieve principle, so let’s focus our attention on one of them. This
problem is best known as “Derangements”. A derangement of the finite set X is a bijection from X into X that doesn’t have fixed points.
A small example: For set X = {1, 2, 3} bijection {(1,1), (2,3), (3,2)} is not derangement, because of (1,1), but bijection {(1,2), (2,3),
(3,1)} is derangement. So let’s turn back to the problem, the goal of which is to find the number of derangements of nelement set.
We have X = {1, 2, 3,…, n}. Let:
A be the set of all bijections from X into X, |A|=n!,
A0 be the set of all derangements of X,
Ai ( i ∈ X ) be the set of bijections from X into X that have (i,i),
AI (I ⊆ X) be the set of bijections from X into X that have (i,i) ∀i⊆I, so and |AI|=(n|AI|)!.
(because there are exactly ielement subsets of X).
Now we just put that result into the sieve principle’s formula:
And now the last step, from we’ll have the answer:
And the last remark:
This problem may not look very practical for use in TopCoder problems, but the thinking behind it is rather important, and these ideas
can be widely applied.
Another interesting method in combinatorics — and one of my favorites, because of its elegance — is called method of paths (or
trajectories). The main idea is to find a geometrical interpretation for the problem in which we should calculate the number of paths of
a special type. More precisely, if we have two points A, B on a plane with integer coordinates, then we will operate only with the
shortest paths between A and B that pass only through the lines of the integer grid and that can be done only in horizontal or vertical
movements with length equal to 1. For example:
All paths between A and B have the same length equal to n+m (where n is the difference between xcoordinates and m is the difference
between ycoordinates). We can easily calculate the number of all the paths between A and B as follows:
or .
Let’s solve a famous problem using this method. The goal is to find the number of Dyck words with a length of 2n. What is a Dyck
word? It’s a string consisting only of n X’s and n Y’s, and matching this criteria: each prefix of this string has more X’s than Y’s. For
example, “XXYY” and “XYXY” are Dyck words, but “XYYX” and “YYXX” are not.
Let’s start the calculation process. We are going to build a geometrical analog of this problem, so let’s consider paths that go from point
A(0, 0) to point B(n, n) and do not cross segment AB, but can touch it (see examples for n=4).
It is obvious that these two problems are equivalent; we can just build a bijection in such a way: step right – ‘X’, step up – ‘Y’.
Here’s the main idea of the solution: Find the number of paths from A to B that cross segment AB, and call them “incorrect”. If path is
“incorrect” it has points on the segment CD, where C = (0, 1), D = (n1, n). Let E be the point nearest to A that belongs to CD (and to
the path). Let’s symmetrize AE, part of our “incorrect” path with respect to the line CD. After this operation we’ll get a path from F =
(1, 1) to B.
It should be easy to see that, for each path from F to B, we can build only one “incorrect” path from A to B, so we’ve got a bijection.
Thus, the number of “incorrect” paths from A to B is . Now we can easily get the answer, by subtracting the number of “incorrect”
paths from all paths:
This number is also known as n’s Catalan number: Cn is the number of Dyck words with length 2n. These numbers also appear in many
other problems, for example, Cn counts the number of correct arrangements of n pairs of parentheses in the expression, and Cn is also
the number of the possible triangulations of a polygon with (n+2) vertices, and so on.
Using recurrence relations
Recurrence relations probably deserves their own separate article, but I should mention that they play a great role in combinatorics.
Particularly with regard to TopCoder, most calculation problems seem to require coders to use some recurrence relation and find the
solution for the values of parameters.
If you’d like to learn more, check out these tutorials: An Introduction to Recursion, Recursion, Part 2, and Dynamic Programming: From
novice to advanced. Done reading? Let’s take a look at some examples.
ChristmasTree (SRM 331 Division Two – Level Three):
We’ll use DP to solve this — it may not be the best way to tackle this problem, but it’s the easiest to understand. Let cnt[lev][r][g][b]
be the number of possible ways to decorate the first lev levels of tree using r red, g green and b blue baubles. To make a recurrent
step calculating cnt[lev][r][g][b] we consider 3 variants:
cnt[lev][r][g][b]=
1) we fill the last level with one color (red, green or blue), so just:
= cnt [lev1][rlev][g][b]+ cnt[lev1][r][glev][b]+ cnt[lev1][r][g][blev]+ ;
2) if (lev%2 == 0) we fill the last level with two colors (red+green, green+blue or red+blue), then we calculate the number of possible
decorations using the Permutation with repetition formula. We’ll get possible variants for each two colors, so just
(cnt[lev1][rlev/2][glev/2][b]+…+cnt[lev1][r][glev/2][blev/2])+;
3) if (lev%3 == 0) we fill the last level with three colors and, again using the Permutation with repetition formula, we’ll get
possible variants, so we’ll get:
(all cnt[l][i][j][k] with negative indices are 0).
DiceGames (SRM 349 Division One – Level Two):
First we should do some preparation for the main part of the solution, by sorting sides array in increasing order and calculating only
the formations where the numbers on the dice go in nondecreasing order. This preparation saves us from calculating the same
formations several times (see SRM 349 – Problem Set & Analysis for additional explanation). Now we will only need to build the
recurrence relation, since the implementation is rather straightforward. Let ret[i][j] be the number of different formations of the first i
dice, with the last dice equal to j (so , where n is the number of elements in sides). Now we can simply write
the recurrence relation:
The answer will be .
Conclusion
As this article was written for novices in combinatorics, it focused mainly on the basic aspects and methods of counting. To learn more,
I recommend you check out the reference works listed below, and keep practicing – both in TopCoder SRMs and pure mathematical
problems. Good luck!
References:
1. Hall M. “Combinatorial theory”
2. Cameron P.J. “Combinatorics: Topics, Techniques, Algorithms”
3. en.wikipedia.org :)
PushRelabel Approach to the Maximum Flow Problem
By NilayVaish– TopCoder Member
Discuss this article in the forums
Introduction
This article presents a new approach for computing maximum flow in a graph. Previous articles had concentrated on finding maximum
flow by finding augmenting paths. FordFulkerson and EdmondsKarp algorithms belong to that class. The approach presented in
this article is called pushrelabel, which is a separate class of algorithms. We’ll look at an algorithm first described by Andrew V.
Goldberg and Robert E. Tarjan, which is not very hard to code and, for dense graphs, is much faster than the augmenting path
algorithms. If you haven’t yet done so, I’d advise you to review the articles on graph theory and maximum flow using augmenting
paths for a better understanding of the basic concepts on the two topics.
The Standard Maximum Flow Problem
Let G = (V,E) be a directed graph with vertex set V and edge set E. Size of set V is n and size of set E is m. G has two distinguished
vertices, a source s and a sink t. Each edge (u,v) ε E has a capacity c(u,v). For all edges (u,v) ∉E, we define c(u,v) = 0. A flow f on
G is a real valued function satisfying the following constraints:
1. Capacity: f(v,w) ≤ c(v,w) ∨ (v,w) ∊ V × V
2. Antisymmetry: f(v,w) = – f(w,v) ∨ (v,w) ∊ V × V
3. Flow Conservation: ∑ u ∊ V f(u,v) = 0 ∨ v ∊ V – {s,t}
The value of a flow f is the net flow into the sink i.e. | f | = ∑ u ∊ V f(u,t) . Figure 1 below shows a flow network with the edges
marked with their capacities. Using this network we will illustrate the steps in the algorithm.
Figure 1 : Flow Network with Capacities
Intuitive Approach Towards the Algorithm
Assume that we have a network of water tanks connected with pipes in which we want to find the maximum flow of water from the
source tank to the sink tank. Each of these water tanks are arbitrarily large and will be used for accumulating water. A tank is said to be
overflowing if it has water in it. Tanks are at a height from the ground level. The edges can be assumed to be pipes connecting these
water tanks. The natural action of water is to flow from a higher level to a lower level. The same holds for this algorithm. The height of
the water tank determines the direction in which water will flow. We can push new flow from a tank to another tank that is downhill
from the first tank, i.e. to tanks that are at a lesser height than the first tank. We need to note one thing here, however:The flow from
a lower tank to a higher tank might be positive. Present height of a tank only determines the direction of new flow.
We fix the initial height of the source s at n and that of sink t at 0. All other tanks have initial height 0, which increases with time. Now
send as much as possible flow from the source toward the sink. Each outgoing pipe from the source s is filled to capacity. We will now
examine the tanks other than s and t. The flow from overflowing tanks is pushed downhill. If an overflowing tank is at the same level or
below the tanks to which it can push flow, then this tank is raised just enough so that it can push more flow. If the sink t is not
reachable from an overflowing tank, we then send this excess water back to the source. This is done by raising the overflowing tank the
fixed height n of the source. Eventually all the tanks except the source s and the sink t stop overflowing. At this point the flow from the
source to the sink is actually the maxflow.
Mathematical Functions
In this section we’ll examine the mathematical notations required for better understanding of the algorithm.
1. Preflow – Intuitively preflow function gives the amount of water flowing through a pipe. It is similar to the flow function. Like
flow, it is a function f: V × V → R. It also follows the capacity and antisymmetry constraints. But for the preflow function, the
conservation constraint is weakened.
∑ u ∊ V f(u,v) ≥ 0 ∨ v ∊ V {s,t}
That is the total net flow at a vertex other than the source that is nonnegative. During the course of the algorithm, we will
manipulate this function to achieve the maximum flow.
2. Excess Flow – We define the excess flow e as e(u) = f(V,u), the net flow into u. A vertex u ∊ V{s,t} is overflowing / active if
e(u) > 0.
3. Residual Capacity – We define residual capacity as function cf: V × V → R where
cf(u,v) = c(u,v) – f(u,v)
If cf(u,v) > 0, then (u,v) is called a residual edge. Readers would have come across this concept in augmenting path algorithms
as well.
4. Height – This function is defined as h: V → N. It denotes the height of a water tank. It has the following properties
h(s) = n
h(t) = 0
h(u) ≤ h(v) + 1 for every residual edge (u,v).
We will prove the last property while discussing the correctness of the algorithm.
Basic Operations
Following are the three basic functions that constitute the algorithm:
1. Initialization – This is carried out once during the beginning of the algorithm. The height for all the vertices is set to zero except
the source for which the height is kept at n. The preflow and the excess flow functions for all vertices are set to zero. Then all the
edges coming out the source are filled to capacity. Figure 2 shows the network after initialization.
Figure 2 : Network after Initialization
Code for Initialization:
void initiate_preflow() {
memset(h, 0, sizeof(int)*n);
h[s] = n;
memset(e, 0, sizeof(int)*n);
for (int i = n ‐ 1; i >= 0; ‐‐i)
memeset(f[i], 0, sizeof(int)*n);
for (int i = 0; i < G[s].size(); i++) {
int v = G[s][i];
f[s][v] = c[s][v];
f[v][s] = ‐c[s][v];
e[v] = c[s][v];
e[s] ‐= c[s][v];
cf[s][v] = c[s][v] ‐ f[s][v];
cf[v][s] = c[v][s] ‐ f[v][s];
}
}
2. push(u,v) – This operation pushes flow from an overflowing tank to a tank to which it has a pipe that can take in more flow and
the second tank is at a lower height than the first one. In other words, if vertex u has a positive excess flow, cf(u,v) > 0 and
h(u) > h(v), then flow can be pushed onto the residual edge (u,v). The amount of the flow pushed is given by
min(e(u),cf(u,v)).
Figure 4 shows a vertex B that has an excess flow of 10. It makes two pushes. In the first push, 7 units of flow are pushed to C.
In the second push, 3 units of flow are pushed to E. Figure 5 illustrates the final result.
Figure 4 : Network before pushing flow from B
Figure 5 : Network pushing flow from B
Code for Push subroutine:
void push(int u, int v) {
temp = min(e(u), cf[u][v]);
f[u][v] = f[u][v] + temp;
f[v][u] = ‐f[u][v];
e[u] = e[u] ‐ temp;
e[v] = e[v] + temp;
cf[u][v] = c[u][v] ‐ f[u][v];
cf[v][u] = c[v][u] ‐ f[v][u];
}
3. relabel(u) – This operation raises the height of an overflowing tank that has no other tanks downhill from it. It applies if u is
overflowing and h(u) ≤ h(v) ∨ residual edges (u,v) i.e. on all the residual edges from u, flow cannot be pushed. The height of
the vertex u is increased by 1 more than the minimum height of its neighbor to which u has a residual edge.
Figure 7 : Network after relabeling D
In Figure 4, pick up vertex D for applying the push operation. We find that it has no vertex that is downhill to it and the edge
from D to that vertex has excess capacity. So we relabel D as shown in Figure 5.
Code for Relabel subroutine:
void relabel(int u) {
int temp = ‐1;
for (int i = 0; i < G[u].size(); i++) {
v = G[u][i];
if (cf[u][v] > 0) {
if (temp == ‐1 || temp > h[v])
temp = h[v];
}
}
h[u] = 1 + temp;
}
Generic Algorithm
The generic algorithm is as follows:
void maxflow() {
initialize_preflow();
while(there is an operatoin that can be carried out)
Select an operation and perform it.
}
The value e(t) will represent the maximum flow. We now try to see why this algorithm would work. This would need two observations
that hold at all times during and after the execution of the algorithm.
1. A residual edge from u to v implies h(u) ≤ h(v) + 1 We had earlier introduced this as a property of the height function. Now
we make use of induction for proving this property.
Initially residual edges are from vertices of height 0 to the source that is at height n.
Consider a vertex u getting relabeled and v is the vertex of minimum height to which u has a residual edge. After relabeling,
the property holds for the residual edge (u,v). For any other residual edge (u,w), h(v) ≤h(w). So after relabeling the
property will hold for residual edge (u,w). For a residual edge, (w,u), since u’s height only goes up, therefore the property
will continue to hold trivially.
Consider a push operation from u to v. After the operation, edge (v,u) will be present in the residual graph. Now h(u) > h(v)
∧ h(u) ≤ h(v) + 1
→ h(u) = h(v) + 1
→ h(v) = h(u) − 1
→ h(v) _ h(u) + 1
2. There is no path for source to sink in the residual graph – Let there be a simple path {v0 , v1 , . . . , vk−1, vk } from v0 = s
to vk = t. Then, as per the above observation,
h(v i ) ≤ h(v i+1 ) + 1
→ h(s) ≤ h(t) + k
→ n ≤ k
→ n < k + 1
This violates the simple path condition as the path has k+1 vertices. Hence there is no path from source to sink.
When the algorithm ends, no vertex is overflowing except the source and the sink. There is also no path from source to sink. This is the
same situation in which an augmenting path algorithm ends. Hence, we have maximum flow at the end of the algorithm.
Analysis
The analysis requires proving certain properties relating to the algorithm.
1. s is reachable from all the overflowing vertices in the residual graph – Consider u as an overflowing and S as the set of all
the vertices that are reachable from u in the residual graph. Suppose s ∉ 2S . Then for every vertex pair (v,w) such that v ∊ S
and w ∊ VS , f(w,v) ≤ 0. Because if f(w,v) > 0, then cf(v,w) > 0 which implies w is reachable from u. Thus, since e(v) ≤ 0,
for all v ∊ S, e(v) = 0. In particular, e(u) = 0, which is a contradiction.
2. The height of a vertex never decreases – The height of a vertex changes only during the relabeling operation. Suppose we
relabel vertex u. Then for all vertices v such that (u,v) is a residual edge, we have h(v) ≤ h(u), which clearly means
h(v)+1>h(u). So the height of a vertex never decreases.
3. The height of a vertex can be at maximum 2n1 – This holds for s and t since their heights never change. Consider a vertex u
such that e(u) > 0. Then there exists a simple path from u to s in the residual graph. Let this path be u = v0, v1 , . . . , vk−1, vk
= s. Then k can be at most n1. Now, as per the definition of h, h(vi) ≤ h(vi+1) + 1. This would yield h(u) ≤ h(s) + n – 1 =
2n – 1.
Now we are in a position to count the number of operations that are carried out during the execution of the code.
Relabeling operations – The height for each vertex other than s and t can change from 0 to 2n1. It can only increase, so there
can be utmost 2n1 relabelings for each vertex. In total there can be a maximum of (2n1)(n2) relabelings. Each relabeling
requires at most degree(vertex) operations. Summing this up over all the vertices and over all the relabelings, the total time
spent in relabeling operations is O(nm).
Saturated Pushes – A saturating push from u to v, results in f(u,v) = c(u,v). In order to push flow from u to v again, flow must
be pushed from v to u first. Since h(u) = h(v) + 1, so v’s height must increase by at least 2. Similarly h(u) should increase by
at least 2 for the next push. Combining this with the earlier result on the maximum height of avertex, the total number of
saturating pushes is at most 2n1 per edge. So that total overall the edges is at most (2n1)m < 2nm.
Nonsaturated Pushes – Let φ = ∑ u ∊ V, u is active h(u). Each nonsaturating push from a vertex u to any other vertex v
causes φ to decrease by at least 1 since h(u) > h(v). Each saturating push can increase φ by at most 2n1 since it could be that
v was not active before. The total increase in φ due to saturating pushes is at most (2n1)(2nm). The total increase in φ due to
relabeling operation is at most (2n1)(n2). Therefore, the total number of nonsaturating pushes is at most (2n1)(2nm) +
(2n1)(n2)≤ 4n2m.
Thus the generic algorithm takes O(n2m) operations in total. Since each push operation requires O(1) time, hence the running time of
the algorithm will also be O(n2m) if the condition given in the while loop can be tested in O(1) time. The next section provides an
implementation for doing the same. In fact, by ordering the operations in a particular manner, a more rigorous analysis proves that a
better time bound can be achieved.
Firstin Firstout Algorithm
We will make use of a firstin, firstout strategy for selecting vertices on which push/relabel operation will be performed. This is done by
creating a queue of vertices that are overflowing. Initially all the overflowing vertices are put into the queue in any order. We will run
the algorithm till the queue becomes empty.
In every iteration the vertex at the front of the queue is chosen for carrying out the operations. Let this vertex be u. Consider all the
edges of u, both those that are incident on u and those that are incident on other vertices from u. These are edges along which u can
potentially push more flow. Go through these edges one by one, pushing flow along edges that have excess capacity. This is done until
either u becomes inactive or all the edges have been explored. If during the iteration any new vertex starts to overflow, then add that
vertex to the end of the queue. If at the end of the iteration u is still overflowing, then it means u needs a relabeling. Relabel u and
start the next iteration with u at the front of the queue. If any time during the process or at end of it u becomes inactive, remove u
from the queue. This process of pushing excess flow from a vertex until it becomes inactive is called discharging a vertex.
Code for FirstIn FirstOut Algorithm:
void maxflow() {
initialize_perflow();
queue q;
char *l = new char[n];
int u, v, m;
memset(l, 0, sizeof(char)*n);
for (int i = 0; i < G[s].size(); i++) {
if (G[s][i] != t) {
q.push(G[s][i]);
l[G[s][i]] = 1;
}
}
while (q.size() != 0) {
u = q.front();
m = ‐1;
for (int i = 0; i < G[u].size() && e[u] > 0; i++) {
v = G[u][i];
if (cf[u][v] > 0) {
if (h[u] > h[v]) {
push(u, v);
if (l[v] == 0 && v != s && v != t) {
l[v] = 1;
q.push(v);
}
}
else if (m == ‐1)m = h[v];
else m = min(m, h[v]);
}
}
if (e[u] != 0) h[u] = 1 + m;
else {
l[u] = 0;
q.pop();
}
}
}
Analysis of FirstIn FirstOut Algorithm
To analyze the strategy presented above, the concept of a pass over a queue needs to be defined. Pass 1 consists of discharging the
vertices added to the queue during initialization. Pass i + 1 consists of discharging vertices added during the Pass i.
The number of passes over the queue is at most 4n2 – Let φ = max {h(u) |u is active|}. If no heights changes during a
given pass, each vertex has its excess moved to vertices that are lower in height. Hence φ decreases during the pass. If φ does
not change, then there is at least one vertex whose height increases by at least 1. If φ increases during a pass, then some
vertex’s height must have increased by as much as the increase φ. Using the proof about the maximum height of the vertex, the
maximum number of passes in which φ increases or remains same is 2n2. The total number of passes in which φ decreases is
also utmost 2n2.Thus the total number of passes is utmost 4n2.
The number of nonsaturating pushes is at most 4n3 – There can be only one nonsaturating push per vertex in a single
pass. So the total number of nonsaturating pushes is at most 4n3.
On combining all the assertions, the total time required for running the firstin firstout algorithm is O(nm + n3) which is O(n3).
Related Problems
In general any problem that has a solution using maxflow can be solved using “pushrelabel” approach. Taxi and Quest4 are good
problems for practice. More problems can be found in the article on max flow using augmenting paths. In problems where the time
limits are very strict, pushrelabel is more likely to succeed as compared to augmenting path approach. Moreover, the code size is not
very significant. The code provided is of 61 lines (16 + 36 + 9), and I believe it can be shortened.
In places where only the maximum flow value is required and the actual flow need not be reported, the algorithm can be changed
slightly to work faster. The vertices that have heights ≥n may not be added to the queue. This is because these vertices can never push
more flow towards the sink. This may improve the running time by a factor of 2. Note that this will not change the asymptotic order of
the algorithm. This change can also be applied in places where the mincut is required. Let (S,T) be the mincut. Then, T contains the
vertices that reachable from t in the residual graph.
References
1. Andrew V. Goldberg and Robert E. Tarjan, A new approach to the maximumflow problem.
2. Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest. Introduction to Algorithms.
3. _efer_. Algorithm Tutorial: MaximumFlow
4. gladius. Algorithm Tutorial: Introduction to graphs and their data structures
The author would like to acknowledge the assistance of zenithankit, who helped review and improve this article prior to publication.
Disjointset Data Structures
By vlad_D– TopCoder Member
Discuss this article in the forums
Introduction
Many times the efficiency of an algorithm depends on the data structures used in the algorithm. A wise choice in the structure you use
in solving a problem can reduce the time of execution, the time to implement the algorithm and the amount of memory used. During
SRM competitions we are limited to a time limit of 2 seconds and 64 MB of memory, so the right data structure can help you remain in
competition. While some Data Structures have been covered before, in this article we’ll focus on data structures for disjoint sets.
The problem
Let’s consider the following problem: In a room are N persons, and we will define two persons are friends if they are directly or
indirectly friends. If A is a friend with B, and B is a friend with C, then A is a friend of C too. A group of friends is a group of persons
where any two persons in the group are friends. Given the list of persons that are directly friends find the number of groups of friends
and the number of persons in each group. For example N = 5 and the list of friends is: 12, 54, and 51. Here is the figure of the
graph that represents the groups of friends. 1 and 2 are friends, then 5 and 4 are friends, and then 5 and 1 are friends, but 1 is friend
with 2; therefore 5 and 2 are friends, etc.
In the end there are 2 groups of friends: one group is {1, 2, 4, 5}, the other is {3}.
The solution
This problem can be solved using BFS, but let’s see how to solve this kind of problem using data structures for disjoint sets. First of all:
a disjointset data structure is a structure that maintains a collection S1, S2, S3, …, Sn of dynamic disjoint sets. Two sets are disjoint if
their intersection is null. For example set {1, 2, 3} and set {1, 5, 6} aren’t disjoint because they have in common {1}, but the sets {1,
2, 3} and {5, 6} are disjoint because their intersection is null. In a data structure of disjoint sets every set contains a representative,
which is one member of the set.
Let’s see how things will work with sets for the example of the problem. The groups will be represented by sets, and the representative
of each group is the person with the biggest index. At the beginning there are 5 groups (sets): {1}, {2}, {3}, {4}, {5}. Nobody is
anybody’s friend and everyone is the representative of his or her own group.
The next step is that 1 and 2 become friends, this means the group containing 1 and the group with 2 will become one group. This will
give us these groups: {1, 2} , {3}, {4}, {5}, and the representative of the first group will become 2. Next, 5 and 4 become friends.
The groups will be {1,2}, {3}, {4, 5}. And in the last step 5 and 1 become friends and the groups will be {1, 2, 4, 5}, {3}. The
representative of the first group will be 5 and the representative for second group will be 3. (We will see why we need representatives
later). At the end we have 2 sets, the first set with 4 elements and the second with one, and this is the answer for the problem
example: 2 groups, 1 group of 4 and one group of one.
Perhaps now you are wondering how you can check if 2 persons are in the same group. This is where the use of the representative
elements comes in. Let’s say we want to check if 3 and 2 are in the same group, we will know this if the representative of the set that
contains 3 is the same as the representative of the set that contains 2. One representative is 5 and the other one is 3; therefore 3 and
2 aren’t in same groups of friends.
Some operations
Let’s define the following operations:
CREATESET(x) – creates a new set with one element {x}.
MERGESETS(x, y) – merge into one set the set that contains element x and the set that contains element y (x and y are in
different sets). The original sets will be destroyed.
FINDSET(x) – returns the representative or a pointer to the representative of the set that contains element x.
The solution using these operations
Let’s see the solution for our problem using these operations:
Read N;
for (each person x from 1 to N) CREATE‐SET(x)
for (each pair of friends (x y) ) if (FIND‐SET(x) != FIND‐SET(y)) MERGE‐SETS(x, y)
Now if we want to see if 2 persons (x, y) are in same group we check if FINDSET(x) == FINDSET(y).
We will analyze the running time of the disjointset data structure in terms of N and M, where N is the number of times that CREATE
SET(x) is called and M is the total number of times that CREATESET(x), MERGESETS(x, y) and FINDSET(x) are called. Since the sets
are disjoint, each time MERGESETS(x, y) is called one set will be created and two will be destroyed, giving us one less set. If there are
n sets after n1 calls of MERGESETS(x,y) there will remain only one set. That’s why the number of MERGESETS(x,y) calls is less than
or equal to the number of CREATESET(x) operations.
Implementation with linked lists
One way to implement disjoint set data structures is to represent each set by a linked list. Each element (object) will be in a linked list
and will contain a pointer to the next element in the set and another pointer to the representative of the set. Here is a figure of how the
example of the problem will look like after all operations are made. The blue arrows are the pointers to the representatives and the
black arrows are the pointers to the next element in the sets. Representing sets with linked lists we will obtain a complexity of O(1) for
CREATESET(x) and FINDSET(x). CREATESET(x) will just create a new linked list whose only element (object) is x, the operation
FINDSET(x) just returns the pointer to the representative of the set that contains element (object) x.
Now let’s see how to implement the MERGESETS(x, y) operations. The easy way is to append x’s list onto the end of y’s list. The
representative of the new set is the representative of the original set that contained y. We must update the pointer to the
representative for each element (object) originally on x’s list, which takes linear time in terms of the length of x’s list. It’s easy to prove
that, in the worst case, the complexity of the algorithm will be O(M^2) where M is the number of operations MERGESETS(x, y). With
this implementation the complexity will average O(N) per operation where N represents the number of elements in all sets.
The “weighted union heuristic”
Let’s see how a heuristic will make the algorithm more efficient. The heuristic is called “a weightedunion heuristic.” In this case, let’s
say that the representative of a set contains information about how many objects (elements) are in that set as well. The optimization is
to always append the smaller list onto the longer and, in case of ties, append arbitrarily. This will bring the complexity of the algorithm
to O(M + NlogN) where M is the number of operations (FINDSET, MERGESETS, CREATESETS) and N is the number of operations
CREATESETS. I will not prove why the complexity is this, but if you are interested you can find the proof in the resources mentioned at
the end of the article.
So far we reach an algorithm to solve the problem in O(M + NlogN) where N is the number of persons and M is the number of
friendships and a memory of O(N). Still the BFS will solve the problem in O(M + N) and memory of O(N + M). We can see that we have
optimized the memory but not the execution time.
Next step: root trees
The next step is to see what we can do for a faster implementation of disjoint set data structures. Let’s represent sets by rooted trees,
with each node containing one element and each tree representing one set. Each element will point only to its parent and the root of
each tree is the representative of that set and its own parent. Let’s see, in steps, how the trees will look for the example from the
problem above.
Step 1: nobody is anybody friend
We have 5 trees and each tree has a single element, which is the root and the representative of that tree.
Step 2: 1 and 2 are friends, MERGESETS(1, 2):
The operation made is MERGESETS(1, 2). We have 4 trees one tree contain 2 elements and have the root 1. The other trees have a
single element.
Step 3: 5 and 4 are friends, MERGESETS(5, 4)
The operation made is MERGESETS(5, 4). Now we have 3 trees, 2 trees with 2 elements and one tree with one element.
Step 4: 5 and 1 are friends, MERGESETS(5, 1)
The operation made is MERGESETS(5, 1). Now we have 2 trees, one tree has 4 elements and the other one has only one element.
As we see so far the algorithm using rooted trees is no faster than the algorithm using linked lists.
Two heuristics
Next we will see how, by using two heuristics, we will achieve the asymptotically fastest disjoint set data structure known so far, which
is almost linear in terms of the number of operations made. These two heuristics are called “union by rank” and “path compression.”
The idea in the first heuristic “union by rank” is to make the root of the tree with fewer nodes point to the root of the tree with more
nodes. For each node, we maintain a rank that approximates the logarithm of the subtree size and is also an upper bound on the
height of the node. When MERGESETS(x, y) is called, the root with smaller rank is made to point to the root with larger rank. The idea
in the second heuristic “path compression,” which is used for operation FINDSET(x), is to make each node on the find path point
directly to the root. This will not change any ranks.
To implement a disjoint set forest with these heuristics, we must keep track of ranks. With each node x, we keep the integer value
rank[x], which is bigger than or equal to the number of edges in the longest path between node x and a subleaf. When CREATESET(x)
is called the initial rank[x] will be 0. When a MERGESETS(x, y) operation is made then the root of higher rank will become the parent
of the root of lower rank – or, in case of tie, we arbitrarily choose one of the roots as the parent and increment its rank.
Let’s see how the algorithm will look.
Let P[x] = the parent of node x.
CREATE‐SET(x)
P[x] = x
rank[x] = 0
MERGE‐SETS(x, y)
PX = FIND‐SET(X)
PY =FIND‐SET(Y)
If (rank[PX] > rank[PY]) P[PY] = PX
Else P[PX] = PY
If (rank[PX] == rank[PY]) rank[PY] = rank[PY] + 1
And the last operation looks like:
FIND‐SET(x)
If (x != P[x]) p[x] = FIND‐SET(P[X])
Return P[X]
Now let’s see how the heuristics helped the running time. If we use only the first heuristic “union by rank” then we will get the same
running time we achieved with the weighted union heuristic when we used lists for representation. When we use both “union by rank”
and “path compression,” the worst running time is O( m α(m,n)), where α(m,n) is the very slowly growing inverse of Ackermann’s
function. In application α(m,n) <= 4 that’s why we can say that the running time is linear in terms of m, in practical situations. (For
more details on Ackermann’s function or complexity see the references below.)
Back to the problem
The problem from the beginning of the article is solvable in O(N + M) and O(N) for memory using disjointset data structure. The
difference for time execution is not big if the problem is solved with BFS, but we don’t need to keep in memory the vertices of the
graph. Let’s see if the problem was like: In a room are N persons and you had to handle Q queries. A query is of the form “x y 1,”
meaning that x is friends with y, or “x y 2” meaning that we are asked to output if x and y are in same group of friends at that moment
in time. In this case the solution with disjointset data structure is the fastest, giving a complexity of O(N + Q).
Practice
Disjointset data structures are a helpful tool for use in different algorithms, or even for solving problems in an SRM. They are efficient
and use small amount of memory. They are useful in applications like “Computing the shorelines of a terrain,” “Classifying a set of
atoms into molecules or fragments,” “Connected component labeling in image analysis,” and others.
To practice what you’ve learned, try to solve GrafixMask – the Division 1 500 from SRM211. The idea is to keep track of all the blocks
and consider each grid point as a node. Next, take all the nodes that aren’t blocked and let (x, y) be the coordinate of the left, right,
down or up node, and if (x, y) is not blocked then you do the operation MERGESETS(node, node2). You should also try to determine
how disjointset data structures can be used in the solution of RoadReconstruction from SRM 356. Disjointset data structures can also
be used in TopographicalImage from SRM 210 and PathFinding, from SRM 156.
I hope you find this data structure to be useful. Good luck in the Arena!
References:
Thomas H. Cormen, Introduction to Algorithms
en.wikipedia.org/wiki/Disjointset_data_structure
en.wikipedia.org/wiki/Ackermann_function
Using Tries
By luison9999– TopCoder Member
Discuss this article
Introduction
There are many algorithms and data structures to index and search strings inside a text, some of them are included in the standard
libraries, but not all of them; the trie data structure is a good example of one that isn’t.
Let word be a single string and let dictionary be a large set of words. If we have a dictionary, and we need to know if a single word is
inside of the dictionary the tries are a data structure that can help us. But you may be asking yourself, “Why use tries if set <string>
and hash tables can do the same?” There are two main reasons:
The tries can insert and find strings in O(L) time (where L represent the length of a single word). This is much faster than set , but
is it a bit faster than a hash table.
The set <string> and the hash tables can only find in a dictionary words that match exactly with the single word that we are
finding; the trie allow us to find words that have a single character different, a prefix in common, a character missing, etc.
The tries can be useful in TopCoder problems, but also have a great amount of applications in software engineering. For example,
consider a web browser. Do you know how the web browser can auto complete your text or show you many possibilities of the text that
you could be writing? Yes, with the trie you can do it very fast. Do you know how an orthographic corrector can check that every word
that you type is in a dictionary? Again a trie. You can also use a trie for suggested corrections of the words that are present in the text
but not in the dictionary.
What is a Tree?
You may read about how wonderful the tries are, but maybe you don’t know yet what the tries are and why the tries have this name.
The word trie is an infix of the word “retrieval” because the trie can find a single word in a dictionary with only a prefix of the word. The
main idea of the trie data structure consists of the following parts:
The trie is a tree where each vertex represents a single word or a prefix.
The root represents an empty string (“”), the vertexes that are direct sons of the root represent prefixes of length 1, the vertexes
that are 2 edges of distance from the root represent prefixes of length 2, the vertexes that are 3 edges of distance from the root
represent prefixes of length 3 and so on. In other words, a vertex that are k edges of distance of the root have an associated
prefix of length k.
Let v and w be two vertexes of the trie, and assume that v is a direct father of w, then v must have an associated prefix of w.
The next figure shows a trie with the words “tree”, “trie”, “algo”, “assoc”, “all”, and “also.”
Note that every vertex of the tree does not store entire prefixes or entire words. The idea is that the program should remember the
word that represents each vertex while lower in the tree.
Coding a Trie
The tries can be implemented in many ways, some of them can be used to find a set of words in the dictionary where every word can
be a little different than the target word, and other implementations of the tries can provide us with only words that match exactly with
the target word. The implementation of the trie that will be exposed here will consist only of finding words that match exactly and
counting the words that have some prefix. This implementation will be pseudo code because different coders can use different
programming languages.
We will code these 4 functions:
addWord. This function will add a single string word to the dictionary.
countPreffixes. This function will count the number of words in the dictionary that have a string prefix as prefix.
countWords. This function will count the number of words in the dictionary that match exactly with a given string word.
Our trie will only support letters of the English alphabet.
We need to also code a structure with some fields that indicate the values stored in each vertex. As we want to know the number of
words that match with a given string, every vertex should have a field to indicate that this vertex represents a complete word or only a
prefix (for simplicity, a complete word is considered also a prefix) and how many words in the dictionary are represented by that prefix
(there can be repeated words in the dictionary). This task can be done with only one integer field words.
Because we want to know the number of words that have like prefix a given string, we need another integer field prefixes that
indicates how many words have the prefix of the vertex. Also, each vertex must have references to all his possible sons (26
references). Knowing all these details, our structure should have the following members:
structure Trie
integer words;
integer prefixes;
reference edges[26];
And we also need the following functions:
initialize(vertex)
addWord(vertex, word);
integer countPrefixes(vertex, prefix);
integer countWords(vertex, word);
First of all, we have to initialize the vertexes with the following function:
initialize(vertex)
vertex.words=0
vertex.prefixes=0
for i=0 to 26
edges[i]=NoEdge
The addWord function consists of two parameters, the vertex of the insertion and the word that will be added. The idea is that when a
string word is added to a vertex vertex, we will add word to the corresponding branch of vertex cutting the leftmost character of
word. If the needed branch does not exist, we will have to create it. All the TopCoder languages can simulate the process of cutting a
character in constant time instead of creating a copy of the original string or moving the other characters.
addWord(vertex, word)
if isEmpty(word)
vertex.words=vertex.words+1
else
vertex.prefixes=vertex.prefixes+1
k=firstCharacter(word)
if(notExists(edges[k]))
edges[k]=createEdge()
initialize(edges[k])
cutLeftmostCharacter(word)
addWord(edges[k], word)
The functions countWords and countPrefixes are very similar. If we are finding an empty string we only have to return the number of
words/prefixes associated with the vertex. If we are finding a nonempty string, we should to find in the corresponding branch of the
tree, but if the branch does not exist, we have to return 0.
countWords(vertex, word)
k=firstCharacter(word)
if isEmpty(word)
return vertex.words
else if notExists(edges[k])
return 0
else
cutLeftmostCharacter(word)
return countWords(edges[k], word);
countPrefixes(vertex, prefix)
k=firstCharacter(prefix)
if isEmpty(word)
return vertex.prefixes
else if notExists(edges[k])
return 0
else
cutLeftmostCharacter(prefix)
return countWords(edges[k], prefix)
Complexity Analysis
In the introduction you may read that the complexity of finding and inserting a trie is linear, but we have not done the analysis yet. In
the insertion and finding notice that lowering a single level in the tree is done in constant time, and every time that the program lowers
a single level in the tree, a single character is cut from the string; we can conclude that every function lowers L levels on the tree and
every time that the function lowers a level on the tree, it is done in constant time, then the insertion and finding of a word in a trie can
be done in O(L) time. The memory used in the tries depends on the methods to store the edges and how many words have prefixes in
common.
Other Kinds of Tries
We used the tries to store words with lowercase letters, but the tries can be used to store many other things. We can use bits or bytes
instead of lowercase letters and every data type can be stored in the tree, not only strings. Let flow your imagination using tries! For
example, suppose that you want to find a word in a dictionary but a single letter was deleted from the word. You can modify the
countWords function:
countWords(vertex, word, missingLetters)
k=firstCharacter(word)
if isEmpty(word)
return vertex.word
else if notExists(edges[k]) and missingLetters=0
return 0
else if notExists(edges[k])
cutLeftmostCharacter(word)
return countWords(vertex, word, missingLetters‐1)
Here we cut a character but we don't go lower in the tree
else
We are adding the two possibilities: the first
character has been deleted plus the first character is present
r=countWords(vertex, word, missingLetters‐1)
cutLeftmostCharacter(word)
r=r+countWords(edges[k], word, missingLetters)
return r
The complexity of this function may be larger than the original, but it is faster than checking all the subsets of characters of a word.
Problems to Practice
WordFind SRM 232
SearchBox SRM 361
CyclicWords SRM 358
TagalogDictionary SRM 342
JoinedString SRM 302
CmpdWords SRM 268
Scruffle 2007 TCCC Marathon Online Round 1
An Introduction to Multidimensional Databases
By dcp– Topcoder Member
Discuss this article in the forums
In this article, we’ll explore the exciting topic of multidimensional databases, which we’ll refer to as MDDBs from now on. MDDBs are a
very popular technology in the business intelligence arena, and they allow a company to perform indepth, strategic analysis on a
variety of factors affecting their company. In addition, MDDBs allow analysts in the company to leverage tools with which they’re
already familiar, such as Microsoft® Excel, to work with and analyze data from MDDBs in a “slice and dice” fashion. The ability to slice
and dice the data puts tremendous power in the hands of the business user, and that will become more apparent as we delve into more
of the details of MDDBs.
The format of this article will be primarily focused on MDDB concepts over specific MDDB vendors; however we’ll look at Hyperion®
Essbase in several parts of the article to give you some concrete examples of MDDB technology in practice. In addition, we’ll look at
some real world examples of MDDBs so that you can get a feel of how they are used in the business arena and the impact they can
make.
Having knowledge of relational databases will be helpful to someone reading this article, but it’s not required. However, we’ll assume
that you at least know what a relational database is, because we’ll be comparing relational and multidimensional databases. If you need
an introduction to relational databases, you can refer to the article here.
Introduction
First, just what in the world are we talking about when we say “multidimensional database”? Well, before we get to that, there are a
few terms we need to understand. The first term is online transaction processing, commonly referred to as OLTP. As the name implies,
OLTP consists of transactionbased systems, which frequently use a relational database as the back end data store. OLTP systems are
usually focused on quick response times and servicing needs immediately. A good example of an OLTP system is the automatic teller
machine (ATM) at your local bank. When you go to make a withdrawal from the ATM, you aren’t really interested in analyzing a bunch
of data, or at least the people in line behind you hope you aren’t! In other words, you are there to make a withdrawal, deposit, etc. and
move along so that the next person waiting in line can complete their transaction. The OLTP system needs to expedite that process.
This leads us to another term we need to define, OLAP, which is stands for online analytical processing. For you history buffs, the
term OLAP was coined by E.F. Codd & Associates in a white paper in 1994[1]. The OLAP approach is focused on the analysis of data,
and typically that data originates from OLTP systems. That’s not to say that OLAP technology is “slower” than OLTP technology. But the
two are focused on entirely different things. OLTP is focused on “getting the job done”, and OLAP is focused on “analyzing the results of
the job”.
So it’s important to understand that there’s quite a paradigm shift when we go from relational databases to MDDBs. To give you another
more concrete example, think of the order processing system running at a sales warehouse. All day long that system is storing orders,
shipping information, inventory, and a lot of other information required for running a sales operation. Moreover, there are many people
at the warehouse who are interacting with that OLTP system, recording orders, shipping product, managing late orders, etc. These
individuals are concerned with making sure that the operation is running smoothly and that customer needs are met, so it’s critical that
the OLTP system respond as quickly as possible.
But a sales analyst might ask the question, “How are we doing on our sales this year vs. our budget (i.e. our sales plan)?” The people
in the sales warehouse probably aren’t going to be able to answer that question. However, it’s their OLTP system that holds the data
which needs to be analyzed to come up with the answer. The OLTP system has all the orders for all product sold for the year, which is
exactly what the sales analyst needs so they can compare it to the budget (probably stored in another system). But there might be
billions of records in the OLTP system, so trying to run reports off of the OLTP data is bad for a number of reasons. First and foremost,
OLTP systems are usually missioncritical to the business. When these systems slow down or experience failure, the entire business
suffers. Have you ever tried to order something off a website and gotten some kind of message like “Our site is currency experiencing
problems, please try again later”? Few things are more frustrating. So it’s very important that OLTP systems stay online and function at
top performance levels, and we want to avoid putting any unnecessary load on those systems.
The sales analyst still needs an answer, so what should we do? Well, that’s where OLAP technology comes in. We can take the data from
the OLTP system and load it to our OLAP system. Note that when we say “OLAP system”, we’re really referring to the MDDB. The two
terms mean essentially the same thing. Another common term you’ll hear used for an MDDB is a “cube”, or an “OLAP cube”. Really, the
terms MDDB, cube, and OLAP cube are used pretty interchangeably, but we will discuss the cube concept shortly. Let’s get back to our
sales analyst’s issue. We would typically perform the load of OLTP data to the MDDB in the early morning hours when no users are
using the OLTP system, so as to minimize any kind of potential business impact. In systems that are available 24 hours a day 7 days a
week, we should load the OLTP data during nonpeak times. The load process is usually a lot more involved and may require
“massaging” the data to get it into an analytical form so it can be put into the MDDB. The term for this load process is called ETL
(extract/transform/load), and there are many ETL tools out on the market today. But ETL is really beyond the scope of our discussion
here. So once the nightly ETL load runs, the OLAP database is loaded and the sales analyst can come in the next morning and find the
answer to her question. The beauty of it is that she can do that without impacting the OLTP system one bit. The folks in the sales
warehouse can continue taking orders and running the business, and the sales analyst can do critical analysis of important data which
can help determine how the business is performing.
One thing to keep in mind when discussing OLAP databases is that there is usually data latency in an OLAP cube. In other words, the
data is usually from some point in time, like maybe the snapshot of data from the night before. But from an analytical standpoint, this
data latency is usually just fine because users are doing analysis like comparing sales trends, etc. and it’s not really critical to have up
totheminute data from the OLTP system. Note that some of the newer OLAP technology offered by Microsoft Analysis Services allows
for nearly real time analysis of business intelligence data[3]. In other words, they have mechanisms where you can keep the OLAP data
in sync with the relational data.
So to wrap up this section, we’ve looked at OLTP and OLAP and the differences between the two. Make sure you have a good grasp of
these concepts before moving to the next section, where we’ll begin to explore the “cube” concept.
The “Cube”
In the last section, we introduced the term “OLAP cube”. The reason the term “cube” is used when describing a MDDB database is
because of the way the data is stored. From a conceptual standpoint, you can think of the data stored in an MDDB as being stored in
multidimensional arrays, where we’re usually dealing with much more than 2 dimensions. This is much different from a relational
database, because in a relational database table you are really only working in two dimensions, namely, row and column. So if I have a
table called coder in a relational database, and the columns in that table are handle and rating, the handle and rating columns
represent a single dimension, and the actual rows of data in the table represent another single dimension (refer to figure 1).
Figure 1
However, when we talk about cubes, we’re typically dealing with many dimensions, not just two. When you design a cube, one of the
first things you must do is determine the dimensions you want in the cube. It’s very important to spend time really thinking through
your cube design before implementing it, because if you don’t choose your dimensions wisely it can cause the cube to perform poorly
and, even worse, you won’t meet the business needs. It’s important to understand that the dimensions are really the architectural
blueprint of your cube.
So going back to our sales warehouse example, we might have a cube with a Time dimension, a Scenarios dimension, and a Measures
dimension (see Figure 2 below). The Scenarios dimension might contain members for plan and actual, which would correspond to
planned (i.e. budgeted) values and actual (i.e. real sales) values respectively. Our Measures dimension could contain members for
things like number of total orders taken, number of orders which were shipped or cancelled, and the order fulfillment rate, which is the
ratio of orders shipped vs. orders placed. Finally, our Time dimension would contain members representing the fiscal calendar for the
company, like fiscal week number, month of the year, fiscal quarter, etc. So conceptually, our cube might look like figure 2, shown
below.
Figure 2
From a coder perspective, you could think of this cube as a threedimensional array, in which the aforementioned dimension names
correspond to dimensions in the array. And to extend that analogy further, if you wanted to know how many actual orders were shipped
for September, you could look at cube[Actual][Sep][shipped orders], which would correspond to the “data cell” in the cube that
contained that piece of data.
The real power of an OLAP cube stems from the fact that as we access those data cells in the multidimensional “array”, the access time
is almost instantaneous. Now, let’s stop for a minute and think about how we would determine how many actual orders were shipped
for September from a relational database. We would most likely write a SQL query to extract the answer from the relational database
tables. However, as we mentioned earlier we obviously wouldn’t want to run that type of query on our OLTP tables as that could slow
down the performance of the OLTP database. But not only that, we would have to make sure that the relational database tables had the
appropriate indexes on the tables we were querying so that the query would run in a reasonable amount of time. Remember, OLTP
database tables can contain billions of records.
In summary, one of the primary advantages offered by MDDBs is the fast access they offer to data.
Dimensions
We introduced dimensions in the previous section, but now we’ll take a more formal look at them. Dimensions, as their name imply,
provide a way to categorize data into different groups. For example, let’s pretend that we’re building an OLAP cube to store TopCoder
data about coders so we can analyze it. We need to think about the different types of data we have and see how we might categorize it,
so let’s make a list of some items in which we might be interested:
Coder Handle
Algorithm Rating
Design Rating
Development Rating
Algorithm Rating Color
Design Rating Color
Development Rating Color
Continent
Country
Join Date
School
Number of Algorithm Rated Events
Number of SRM Wins
Earnings
Next, we need to come up with some logical groupings for this data, and those logical groupings will become our dimensions. So let’s
start with the “measures”, which are typically numeric values or percentages. We’ve already seen an example of a Measures dimension
in the sales cube from the previous section. In the above list, we could make measures out of the numeric rating fields and the number
of algorithm rated events field, so our measures dimension could look look like this:
Measures
Algorithm Rating
Development Rating
Design Rating
Number of Algorithm Rated Events
Number of SRM Wins
Earnings
The way to think about measures is to think of them as the “data values” you are storing in the multidimensional array. All the other
dimensions are focused around analyzing these measures. In other words, measures are really a special dimension. Rating is clearly a
measurement in the case of our coder cube, because it has a numeric value, and the same is true for the number of algorithm rated
events, the number of SRM wins, and earnings.
Next, let’s look at coder handle. We have some options here, and one of them would be to combine coder handle, country and
geographic region into one dimension. In other words, within each dimension, we can have a hierarchy. So we could do something like
this for our coder dimension:
Figure 3
Coder
Asia
China
Coder1
Coder2
Coder3
.
.
North Korea
Coder4
Coder5
Coder6
.
.
North America
United States
Coder7
Coder8
.
.
.
.
Europe
France
Coder9
Coder10
.
.
.
.
In this example, we’ve started our hierarchy with the continent, and underneath each continent we have country, and underneath each
country we have the actual coders for that country. What we have to decide here is whether this hierarchy will meet the needs of our
users. The key thing to notice is that the more data you combine into a dimension, the less flexibility you have later. For example, what
if we want to run some analysis by coder country? We can do it with this hierarchy, but it’s not as easy as it could be if we had country
as a separate dimension. In other words, to run analysis by country with the hierarchy above, we have to drill into each continent and
find all the countries for that continent. If, on the other hand, we had country as its own dimension, then we could run the analysis by
country much more easily since we wouldn’t have to drill in by continent to find each country. For the purposes of our discussion, we’ll
assume that the above hierarchy is sufficient and we go with that for our coder dimension.
This brings us to another concept regarding cubes, namely, aggregation. The idea behind aggregation is that we typically load values
to the lowest level in each dimension in the cube, and then sum up, or “aggregate” the values to the upper levels in each dimension.
With Hyperion® Essbase, level 0 represents the leaf level members in each dimension. For example, in figure 3 above, the coder
members of the dimension would be level 0. The parents of level 0 members are level 1 members, so the countries would be level 1.
The parents of level 1 members are level 2, and we continue this numbering scheme on up until we get to the root member of the
dimension. So when we are loading our cube with coder data, we would load it to the lowest level in our coder dimension. In other
words, we would load it at the coder level. And after we load that data, we can then “aggregate” or roll it up to the upper levels. This is
one of the main reasons OLAP cubes are so powerful. Since those values are aggregated, we don’t have to do any summing when
retrieving the values from the cube. The sum is already precalculated! You can think of a cube as sort of a “power array”, in that we
can look at any combination of members from the dimensions in the cube and get the answer in which we’re interested. So
conceptually, we could look at cube[Coder1][Earnings] to get Coder1′s total earnings, or cube[China][Earnings] to get the total
earnings for Chinese coders. And again, the real power here is that this answer is already precalculated in the OLAP cube so access
time is minimal.
Again, let’s contrast how we would get this information with a relational database. In a relational database, if you wanted to know the
total earnings for coders from China, you would have to run a SQL statement and sum up the earnings (assuming that you hadn’t built
some kind of summary table in the relational database already). This may not seem like much when you’re only talking about a few
thousand coder records, and a SQL query would probably do just fine in this example. But if you consider having to run a SQL query to
sum up values for 800 million transaction records from a sales database, you can see where OLAP cubes offer a huge advantage by
having the values of interest already aggregated.
Some of you savvy folks out there might have noticed that aggregation doesn’t necessarily make sense for all measures. For example,
summing up two coders’ ratings doesn’t provide a meaningful value. So for these types of values, we wouldn’t want to aggregate them
to upper levels. Most MDDB vendors have a way to control this when you set up the cube. For example, with Hyperion® Essbase you
can specify the type of aggregation to use, such as addition, subtraction, etc., but you can give the dimension member the “no
consolidation” property, which means its values won’t be rolled up to its parents.
So now that we’ve defined our coder and measures dimension, let’s look at the other dimensions we need. We obviously want to store
information about a coder’s colors, and that information is in the algorithm rating color, development rating color, and design rating
color values. We could make members in our measures dimension for these items, and just store red, yellow, etc. for the actual data
value in the cube. And that may be fine, depending on the needs and types of analysis that are going to be performed on the cube. The
other choice we could make here would be to make these fields into three cube dimensions, as shown below in figure 4:
Figure 4
Algorithm Rating Color
Algo‐Red
Algo‐Yellow
Algo‐Blue
Algo‐Green
Algo‐Gray
Algo‐Not Rated
Design Rating Color
Design‐Red
Design‐Yellow
Design‐Blue
Design‐Green
Design‐Gray
Design‐Not Rated
Development Rating Color
Dev‐Red
Dev‐Yellow
Dev‐Blue
Dev‐Green
Dev‐Gray
Dev‐Not Rated
Notice that the color in each dimension is prefixed with Algo, Design, or Dev respectively. The reason for this prefix is that with some
MDDB vendors (depending on version), each member name must be unique across every dimension in the cube. Later versions of
Hyperion® Essbase offer the ability to have duplicate member names, but earlier versions did not.
Now, what advantage does having the three rating color values as separate dimensions offer us over having them as members in the
measures dimension? Well, it gives us the ability to “drillin” to the cube to find the total earnings for all coders from North America
who are AlgoRed and DevBlue, for example (i.e. cube[North America][AlgoRed][DevBlue]). Note that we could also drill into the
“North America” member and find out the actual coders. The way this works is that when you drill into the cube, if no data is present
for a member then it won’t bring the member back (depending on the addin settings). So if you drill down into North America, but
there are no members that are both AlgoRed and DevBlue, then no members would be brought back. Different MDDB vendors have
various tools for interfacing with a cube, but most of the more popular ones use some kind of Microsoft® Excel AddIn because
spreadsheets work extremely well for working with cubes and drilling in and out of dimensions. With Hyperion® Essbase, for example,
if we were to connect to the cube we have defined so far, we might be presented with a sheet like figure 5 below:
Figure 5
From there, you can “doubleclick” any dimension, and the addin will drill into the dimension. So if I doubleclicked on Coder, I’d end
up something like this (obviously there would be many more countries but this is just for illustrative purposes):
Figure 6
I could keep drilling into dimensions, and drill into the measures I’m interested in to see the actual values. So we could drill into
Measures, and also navigate down to the AlgoRed, DesignBlue, DevNot Rated members and end up with this configuration:
Figure 5
From there, you can “doubleclick” any dimension, and the addin will drill into the dimension. So if I doubleclicked on Coder, I’d end
up something like this (obviously there would be many more countries but this is just for illustrative purposes):
Figure 6
I could keep drilling into dimensions, and drill into the measures I’m interested in to see the actual values. So we could drill into
Measures, and also navigate down to the AlgoRed, DesignBlue, DevNot Rated members and end up with this configuration:
Figure 7 (illustrative purposes only, not actual data)
As you can see in figure 7, we now know the total earnings by continent for coders who are red in algorithms, blue in design, and not
rated in development.
I hope you can see how this “drilling” ability gives the user virtually endless possibilities to slice and dice data into various formats to do
the analysis they need. They can look at the dimensions at various levels (ex. Continent, country, etc.) to get the summary information
in which they’re interested. And the beauty of it all is that once you build a cube for them, they can do this type of analysis themselves
and you rarely have to write custom reports (that use relational database data) for them anymore! It truly puts the power in the
analyst’s hands, and we’ll look at a real world example of that a bit later.
We’ve made great progress on our cube! There are only two more values left to map, namely, school and join date. We’ll go ahead and
make school its own dimension since it probably makes sense to be able to do analysis at school level.
As for join date, usually with date values it’s best to split date into two dimensions: one for year and one for calendar. We usually call
the calendar dimension “Time”. Here’s an example:
Figure 8
Year
2008
2007
2006
2005
2004
2003
2002
2001
2000
Time
Qtr_1
Jan
Jan_1st
Jan_2nd
Jan_3rd
.
.
Feb
Feb_1st
Feb_2nd
Feb_3rd
.
.
Mar Mar_1st
Mar_2nd
Mar_3rd
.
.
Qtr_2
Apr
Apr_1st
Apr_2nd
Apr_3rd
.
.
May
May_1st
May_2nd
May_3rd
.
.
Jun
Jun_1st
Jun_2nd
Jun_3rd
.
.
Qtr_3
.
.
Qtr_4
.
.
So we essentially create a time hierarchy so that if we want to know which coders joined in the first quarter of 2004, we can drill into
the above dimensions and find that information. And by using this dimension structure, we don’t ever have to change our Time
dimension; we just add a new member to our Year dimension as we reach a new year, or we could just preadd as many as we need.
Besides standard dimensions, another type of dimension you can use in a cube is an “attribute dimension”. Different MDDB vendors
implement attribute dimensions in various ways, but the way to think of attribute dimensions is that they are “characteristics” of
members of another dimension. For example, if join date wasn’t something that we needed to be able to do heavy analysis on, we
could make it an attribute dimension associated with the coder dimension, since each coder has a join date. The thing to remember
about attribute dimensions is that they don’t give you quite the analysis power that a normal dimension does, so you have to ensure
that they are going to meet your needs. For example, you can only associate one member of an attribute dimension with a member of
a standard dimension. So let’s say, for example, that we wanted to track coder’s algo rating over time, but we wanted to make algo
rating an attribute dimension. That would not work, because you could only have one rating value associated with a coder since it’s an
attribute dimension. In that case, you’d have to use two standard dimensions, one for rating time and one for rating. The other thing to
remember about attribute dimensions is that the values aren’t preaggregated in the cube like they are with standard dimensions, so
the aggregation happens at data retrieval time. This can be a performance hit if you have many members in the attribute dimension, so
you want to be smart in you considerations of where to use them or a standard dimension.
One word of caution about dimensions: the more dimensions you have, typically the longer it will take to aggregate the data and drill
into dimensions. In other words, you don’t want to just keep adding dimensions arbitrarily, because eventually if you get too many
dimensions the cube is going to suffer performancewise.
So at this point, we have designed our first OLAP cube! Give yourself a pat on the back!
Building Dimensions & Loading Data to the Cube
Obviously, a cube is worthless without data. We already alluded to ETL earlier in the article, and that is the way we get the data in the
relational database into the form we need for loading into the cube. Sometimes, nothing has to be done here (i.e. the data can be
loaded directly from the relational tables “as is” without any ETL). In either case, we must get the data into the cube. And note that
many MDDB vendors don’t restrict you to using a relational database as the data source for loading cube data. You can use
spreadsheets, CSV files, and many other types of data source.
Before we can load the data into the cube, however, we must make sure we have all the dimension members present for the data which
we wish to load. In other words, we can’t load Coder14 data until we’ve actually added the Coder14 member to the coder dimension.
Again, how you build dimensions varies with MDDB vendors. With Hyperion® Essbase, for example, you use what’s known as a “load
rule”. With a load rule, you typically write an SQL statement (assuming you’re using a relational database as the data source) to query
for the dimension members you wish to load. You can load an individual dimension, or multiple dimensions in the same load rule. You
can also specify if members are to be removed if they aren’t found in the data source, so that is one way you can remove members
from the cube (inactive coders, for example).
One the dimension members have been added, we’re ready to load the actual cube data. The load rule for loading data to the cube
must specify a member value for each dimension. If you think about it, that makes sense because if we leave out a dimension, then the
loader won’t know where to put the data in the cube. As mentioned earlier, it’s usually best to load data to the lowest level of each
dimension. Note that if you accidentally load a data value to a higher level, then when you “calculate” the cube (covered in the next
section), you will wipe out the values because the aggregation will roll up the lower level value, which is nothing, to the higher level. As
an example, if I loaded earnings values to the country level instead of the coder level, when I did the aggregation the country values
would be overwritten because it would aggregate the values at coder level upwards.
Load rules are a big subject, and we’ve just touched on them briefly here. There are many caveats to deal with when loading data to a
cube, such as what to do with rejected records, and how to handle duplicate member names. It’s important to have a good strategy in
place for handling these types of situations whenever you are going to build cubes. The strategy may vary from cube to cube, because
with some cubes you may not care as much if records are rejected, but in other cubes it may be a critical event.
Calculations
So now that we’ve loaded our data to the OLAP cube, we need to aggregate it up to the higher levels. To do that, we must run a
“calculation” step, which does just that. Note that we are not limited to just aggregating (i.e. rolling up) data, we can also have
complex calculation formulas on data members. The industry standard for writing these types of calculations is MDX, which is a query
language for OLAP databases. It’s like SQL in a lot of ways, but it’s sort of a paradigm shift because you have to think a bit differently
when working with MDX because you are working in many dimensions, not just two.
MDX Calculation scripts can do other things as well, such as allocating a sales adjustment across several products over a time period.
For example, let’s say a department store has a 6 million dollar sales adjustment. This dollar amount needs to be distributed across
fiscal year 2008, 1st quarter sales. The MDX script could take that 6 million dollars and “spread” it down to the lowest levels in each
dimension, so that each product would contain a portion of the adjustment. These types of adjustments are quite common in sales
cubes, where things like kickbacks and other types of dollar adjustments occur on a frequent basis (quarterly, for example) and have to
be distributed across products.
MDX calculations also allow us to do simple calculations, such as creating a formula to add two values together and subtract another.
For example, we might have the following formula for net sales:
Net_Sales = Gross_Sales ‐ Cost of Goods Sold + Sales Adjustments
We could do that calculation ourselves in the load rule using a SQL expression, or we could let the cube calculate it for us. If you do the
calculation in the SQL expression though, then you are aggregating computed values, which can cause a loss of precision as those
computed values are aggregated to higher levels. So it’s usually a good strategy to load the raw values and put the calculations in the
cube itself.
Storage Options
In this section, we’ll deal with some specifics regarding Hyperion® Essbase storage options because it’s important to understand the
options you have available. So far, we’ve been discussing the Block Storage Option (BSO), which requires us to calculate the data to
get it to higher levels. Obviously, this aggregation step can take substantial time if you have a lot of data and/or dimensions in the
cube. One thing you can do with BSO cubes to reduce aggregation time is specify that a member is a Dynamic Calc member. This
means that the value isn’t calculated until retrieval time. The advantage here obviously is that your aggregation time is reduced.
However, you must realize that it can slow down your retrieval time, because now the aggregation happens at fetch time. So it’s a “pay
me now, pay me later” situation. Generally, you want to limit dynamic calcs in a cube because you are usually going to load the cube in
a nightly batch process, when time isn’t a scarce resource. So a longer calc time doesn’t matter as much since it’s happening at night.
But when an impatient financial analyst is fetching a value from the cube, you want that result to come back quickly.
However, Hyperion® Essbase also offers another type of storage option called Aggregate Storage Option (ASO). With the ASO
technology, no additional calculation step is required to aggregate the data! The values are aggregated dynamically at retrieval time. So
basically, you load your data and you’re done. ASO cubes are usually much smaller than their BSO counterparts, so that is one reason
why the values can be aggregated at runtime and you can still get good retrieval performance. However, you can also precalculate
values in an ASO cube and store them in aggregations, which is useful when the cube gets extremely big.
So which option should you choose, BSO or ASO? That depends. ASO cubes don’t allow you to run calculation scripts, since all data
retrieval is dynamic (except for the precalculated values we just mentioned). So in a financial cube, for example, where you are going
to be doing a lot of allocation calculation scripts, BSO is probably the best choice.
However, in a cube where you want to have a lot of detailed data and want to be able to load and retrieve it quickly, ASO is a great
option to consider.
Some Real World Examples
In this section, I’ll share some real world examples of how Hyperion® Essbase OLAP technology has made a huge impact in one of the
places I’ve worked.
In the shop where I worked, we were literally getting new report requests every 23 days for some new report for our forecasting
system. The users wanted to see forecast totals by style, color, and size (it was an apparel manufacturing business), then somebody
else would want to see totals by style only. Another user just wanted a report that showed total sales for last year for our “Big 3″
retailers. In short, we couldn’t finish one report before there were 2 to 3 more user requests being added to our queue! That’s when we
decided to replace the entire forecasting system with an OLAP cube. One thing I haven’t mentioned until now is that you can also load
data to the cube through the Excel AddIn. So we made the OLAP cube the data repository for the forecast, so users would use the
Excel AddIn to “lock and send” their forecast data to the cube using Excel. They also had the ability to calculate the cube (it was a BSO
cube) so their forecast would be aggregated. We also loaded our actual sales data to the cube on a weekly basis so the users could
compare actual to forecast to see how the company was doing, and also so they could plan the forecast accurately by looking at prior
year sales.
Once we gave the forecast group this OLAP cube, amazingly, the weekly report requests suddenly stopped! We found that the users
were able to meet all their reporting needs with the Excel AddIn and OLAP cube, and we no longer had to write custom reports. It was
a huge win for the company, and the users absolutely loved it!
In another area of the company, we were having major problems with excess inventory and we needed a way to track and locate this
inventory at the plants and get rid of it by selling it to our wholesale distributor. We had a short timeframe to do this, because after a
selling season ended, the inventory was basically useless and the wholesaler would no longer buy it. So we developed an OLAP cube to
track the inventory and compare it against the forecast (the same forecasting system I just described). Doing this allowed our financial
planning group to see ahead of time which of our inventory was not going to be sold by subtracting forecast from inventory and seeing
what was left. The financial planning department could then take a proactive approach to selling the inventory to our wholesaler before
it became obsolete, and we literally saved millions of dollars in the process!
Conclusion
I hope by this point in the article you can see the advantages that MDDB’s offer. The learning curve is a bit steep when you’re getting
acclimated with building cubes, but the payoff is tremendous, as you can use the technologies to make a huge impact in a company.
References
[1] – http://en.wikipedia.org/wiki/OLAP_cube
[2] – HYPERION® SYSTEM™ 9 BI +™ ANALYTIC SERVICES™ RELEASE 9.3 Database Administrators Guide
[3] – http://www.microsoft.com/technet/prodtechnol/sql/2005/rtbissas.mspx
The Best Questions for Wouldbe C++ Programmers: Part 1
By zmij– Topcoder Member
Discuss this article in the forums
It seems that an almost obligatory and very important part of the recruitment process is “the test.” “The test” can provide information
both for the interviewer and the candidate. The interviewer is provided with a means to test the candidate’s practical knowhow and
particular programming language understanding; the candidate can get an indication of the technologies used for the job and the level
of proficiency the company expects and even decide if he still wants (and is ready for) the job in question.
I’ve had my fair share of interviews, more or less successful, and I would like to share with you my experience regarding some
questions I had to face. I also asked for feedback from three of the top rated TopCoder members: bmerry, kyky and sql_lall , who
were kind enough to correct me where I was not accurate (or plain wrong) and suggest some other questions (that I tried to answer to
my best knowledge). Of course every question might have alternative answers that I did not manage to cover, but I tried to maintain
the dynamics of an interview and to let you also dwell on them (certainly I’m not the man that knows all the answers). So, pencils up
everyone and “let’s identify potential C++ programmers or C++ programmers with potential.”
1. What is a class?
A class is a way of encapsulating data, defining abstract data types along with initialization conditions and operations allowed
on that data; a way of hiding the implementation (hiding the guts & exposing the skin); a way of sharing behavior and
characteristics
2. What are the differences between a C struct and a C++ struct?
A C struct is just a way of combining data together; it only has characteristics (the data) and does not include behavior
(functions may use the structure but are not tied up to it)
Typedefed names are not automatically generated for C structure tags; e.g.,:
// a C struct
struct my_struct {
int someInt;
char* someString;
};
// you declare a variable of type my_struct in C
struct my_struct someStructure;
// in C you have to typedef the name to easily
// declare the variable
typedef my_struct MyStruct;
MyStruct someOtherStuct;
// a C++ struct
struct MyCppStruct {
int someInt;
char* someString;
};
// you declare a variable of type MyCppStruct in C++
MyCppStruct someCppStruct;
// as you can see the name is automatically typedefed
But what’s more important is that a C struct does not provide enablement for OOP concepts like encapsulation or
polymorphism. Also “C structs can’t have static members or member functions”, [bmerry]. A C++ struct is actually a
class, the difference being that the default member and base class access specifiers are different: class defaults to private
whereas struct defaults to public.
3. What does the keyword const mean and what are its advantages over #define?
In short and by far not complete, const means “readonly”! A named constant (declared with const) it’s like a normal
variable, except that its value cannot be changed. Any data type, userdefined or builtin, may be defined as a const, e.g.,:
// myInt is a constant (read‐only) integer
const int myInt = 26;
// same as the above (just to illustrate const is
// right and also left associative)
int const myInt = 26;
// a pointer to a constant instance of custom
// type MyClass
const MyClass* myObject = new MyObject();
// a constant pointer to an instance of custom
// type MyClass
MyClass* const myObject = new MyObject();
// myInt is a constant pointer to a constant integer
const int someInt = 26;
const int* const myInt = &someInt;
#define is error prone as it is not enforced by the compiler like const is. It merely declares a substitution that the
preprocessor will perform without any checks; that is const ensures the correct type is used, whereas #define does not.
“Defines” are harder to debug as they are not placed in the symbol table.
A constant has a scope in C++, just like a regular variable, as opposed to “defined” names that are globally available and
may clash. A constant must also be defined at the point of declaration (must have a value) whereas “defines” can be “empty.”
Code that uses const is inherently protected by the compiler against inadvertent changes: e.g., to a class’ internal state
(const member variables cannot be altered, const member functions do not alter the class state); to parameters being used
in methods (const arguments do not have their values changed within methods) [sql_lall]. A named constant is also subject
for compiler optimizations.
In conclusion, you will have fewer bugs and headaches by preferring const to #define.
4. Can you explain the private, public and protected access specifiers?
public: member variables and methods with this access specifier can be directly accessed from outside the class
private: member variables and methods with this access specifier cannot be directly accessed from outside the class
protected: member variables and methods with this access specifier cannot be directly accessed from outside the class with
the exception of child classes
These access specifiers are also used in inheritance (that’s a whole other story, see next question). You can inherit publicly,
privately or protected (though I must confess, I cannot see the benefits of the latter).
5. Could you explain public and private inheritance?[kyky, sql_lall]
Public inheritance is the “default” inheritance mechanism in C++ and it is realized by specifying the public keyword before the
base class
class B : public A{
};
Private inheritance is realized by specifying the private keyword before the base class or omitting it completely, as private is
the default specifier in C++
class B : private A{
};
or
class B : A{
};
The public keyword in the inheritance syntax means that the publicly/protected/privately accessible members inherited from
the base class stay public/protected/private in the derived class; in other words, the members maintain their access
specifiers. The private keyword in the inheritance syntax means that all the base class members, regardless of their access
specifiers, become private in the derived class; in other words, private inheritance degrades the access of the base class’
members – you won’t be able to access public members of the base class through the derived one (in other languages, e.g.,
Java, the compiler won’t let you do such a thing).
From the relationship between the base and derived class point of view,
class B : public A {}; B "is a" A but class B : private A {};
means B “is implemented in terms of” A.
Public inheritance creates subtypes of the base type. If we have class B : public A {}; then any B object is substituteable by
its base calls object (through means of pointers and references) so you can safely write
A* aPointer = new B();
Private inheritance, on the other hand, class B : private A {};, does not create subtypes making the base type inaccessible
and is a form of object composition. The following illustrates that:
class A {
public:
A();
~A();
void doSomething();
};
void A :: doSomething(){
}
class B : private A {
public:
B();
~B();
};
B* beePointer = new B();
// ERROR! compiler complains that the
// method is not accessible
beePointer‐>doSomething();
// ERROR! compiler complains that the
// conversion from B* to A* exists but
// is not accessible
A* aPointer = new B();
// ! for the following two the standard
// stipulates the behavior as undefined;
// the compiler should generate an error at least
// for the first one saying that B is not a
// polymorphic type
A* aPointer2 = dynamic_cast<A*>(beePointer);
A* aPointer3 = reinterpret_cast<A*>(beePointer);
6. Is the “friend” keyword really your friend?[sql_lall]
The friend keyword provides a convenient way to let specific nonmember functions or classes to access the private members
of a class
friends are part of the class interface and may appear anywhere in the class (class access specifiers do not apply to friends);
friends must be explicitly declared in the declaration of the class; e.g., :
class Friendly;
// class that will be the friend
class Friend {
public:
void doTheDew(Friendly& obj);
};
class Friendly {
// friends: class and function; may appear
// anywhere but it's
// better to group them toghether;
// the default private access specifier does
// not affect friends
friend class Friend;
friend void friendAction(Friendly& obj);
public:
Friendly(){ };
~Friendly(){ };
private:
int friendlyInt;
};
// the methods in this class can access
// private members of the class that declared
// to accept this one as a friend
void Friend :: doTheDew(Friendly& obj) {
obj.friendlyInt = 1;
}
// notice how the friend function is defined
// as any regular function
void friendAction(Friendly& obj){
// access the private member
if(1 == obj.friendlyInt){
obj.friendlyInt++;
} else {
obj.friendlyInt = 1;
}
}
“friendship isn’t inherited, transitive or reciprocal,” that is, your father’s best friend isn’t your friend; your best friend’s friend
isn’t necessarily yours; if you consider me your friend, I do not necessarily consider you my friend.
Friends provide some degree of freedom in a class’ interface design. Also in some situations friends are syntactically better,
e.g., operator overloading – binary infix arithmetic operators, a function that implements a set of calculations (same
algorithm) for two related classes, depending on both (instead of duplicating the code, the function is declared a friend of
both; classic example is Matrix * Vector multiplication).
And to really answer the question, yes, friend keyword is indeed our friend but always “prefer member functions over
nonmembers for operations that need access to the representation.”[Stroustrup]
7. For a class MyFancyClass { }; what default methods will the compiler generate?
The default constructor with no parameters
The destructor
The copy constructor and the assignment operator
All of those generated methods will be generated with the public access specifier
E.g. MyFancyClass{ }; would be equivalent to the following :
class MyFancyClass{
public:
// default constructor
MyFancyClass();
// copy constructor
MyFancyClass(const MyFancyClass&);
// destructor
~MyFancyClass();
// assignment operator
MyFancyClass& operator=(const MyFancyClass&);
};
All of these methods are generated only if needed
The default copy constructor and assignment operator perform memberwise copy construction/assignment of the nonstatic
data members of the class; if references or constant data members are involved in the definition of the class the assignment
operator is not generated (you would have to define and declare your own, if you want your objects to be assignable)
I was living under the impression that the unary & (address of operator) is as any builtin operator – works for builtin types;
why should the builtin operator know how to take the address of your homebrewed type? I thought that there’s no
coincidence that the “&” operator is also available for overloading (as are +, , >, < etc.) and it’s true is not so common to
overload it, as you can live with the one generated by the compiler that looks like the following:
inline SomeClass* SomeClass::operator&() {
return this;
}
Thanks to bmerry for making me doubt what seemed the obvious. I found out the following:
From the ISO C++ standard:
Clause 13.5/6 [over.oper] states that operator =, (unary) & and , (comma) have a predefined meaning for each type. Clause
5.3.1/2 [expr.unary.op] describes the meaning of the addressof operator. No special provisions are mode for classtype objects
(unlike in the description of the assignment expression). Clause 12/1 [special] lists all the special member functions, and states
that these will be implicitly declared if needed. The addressof operator is not in the list.
From Stroustrup’s The C++ Programming Language – Special 3rd Edition:
“Because of historical accident, the operators = (assignment), & (addressof), and , (sequencing) have predefined meanings when
applied to class objects. These predefined meanings can be made inaccessible to general users by making them private:…
Alternatively, they can be given new meanings by suitable definitions.”
From the second edition of Meyer’s Effective C++:
“A class declaring no operator& function(s) does NOT have them implicitly declared. Rather, compilers use the builtin addressof
operator whenever “&” is applied to an object of that type. This behavior, in turn, is technically not an application of a global
operator& function. Rather, it is a use of a builtin operator.” In the errata http://www.aristeia.com/BookErrata/ec++2e
errata_frames.html
8. How can you force the compiler not to generate the above mentioned methods?
Declare and define them yourself – the ones that make sense in your class’ context. The default noparameters constructor
will not be generated if the class has at least one constructor with parameters declared and defined.
Declare them private – disallow calls from the outside of the class and DO NOT define them (do not provide method bodies for
them) – disallow calls from member and friend functions; such a call will generate a linker error.
9. What is a constructor initialization list?
A special initialization point in a constructor of a class (initially developed for use in inheritance).
Occurs only in the definition of the constructor and is a list of constructor calls separated by commas.
The initialization the constructor initialization list performs occurs before any of the constructor’s code is executed – very
important point, as you’ll have access to fully constructed member variables in the constructor!
For example:
// a simple base class just for illustration purposes
class SimpleBase{
public:
SimpleBase(string&);
~SimpleBase();
private:
string& m_name;
};
// example of initializer list with a call to the
// data member constructor
SimpleBase :: SimpleBase(string& name) : m_name(name){
}
// a class publicly derived from SimpleBase just for
// illustration purposes
class MoreComplex : public SimpleBase{
public:
MoreComplex(string&, vector<int>*, long);
~MoreComplex();
private:
vector<int>* m_data;
const long m_counter;
};
// example of initializer list with calls to the base
// class constructor and data member constructor;
// you can see that built‐in types can also be
// constructed here
MoreComplex :: MoreComplex(string &name,
vector<int>* someData, long counter) :
SimpleBase(name), m_data(someData),
m_counter(counter){
}
As you saw in the above example, builtin types can also be constructed as part of the constructor initialization list.
Of course you do not have to use the initialization list all the time (see the next question for situations that absolutely require
an initialization list) and there are situations that are not suitable for that: e.g., you have to test one of the constructor’s
arguments before assigning it to your internal member variable and throw if not appropriate.
It is recommended that the initialization list has a consistent form: first the call to the base class(es) constructor(s), and then
calls to constructors of data members in the order they were specified in the class’ declaration . Note that this is just a matter
of coding style: you declare your member variables in a certain order and it will look good and consistent to initialize them in
the same order in the initialization list.
10. When “must” you use a constructor initialization list?
Constant and reference data members of a class may only be initialized, never assigned, so you must use a constructor
initialization list to properly construct (initialize) those members.
In inheritance, when the base class does not have a default constructor or you want to change a default argument in a default
constructor, you have to explicitly call the base class’ constructor in the initialization list.
For reasons of correctness – any calls you make to member functions of subobjects (used in composition) go to initialized
objects.
For reasons of efficiency. Looking at the previous question example we could rewrite the SimpleBase constructor as follows:
SimpleBase :: SimpleBase(string &name){
m_name = name;
}
The above will generate a call to the default string constructor to construct the class member m_name and then the assignment
operator of the string class to assign the name argument to the m_name member. So you will end up with two calls before the
data member m_name is fully constructed and initialized.
SimpleBase :: SimpleBase(string &name) : m_name(name){
}
The above will only generate a single call, which is to the copy constructor of the string class, thus being more efficient.
That’s it for the first part of this installment. Stay tuned for the second one, as we’re going to go deeper into the language features.
Good luck on those interviews!
References
[1] – Bjarne Stroustrup – The C++ Programming Language Special 3rd Edition
[2] – Stanley B. Lippman, Josee Lajoie, Barbara E. Moo – C++ Primer
[3] – C++ FAQ Lite
[4] Herb Sutter – Exceptional C++: 47 Engineering Puzzles, Programming Problems, and Solutions
[5] Scott Meyers – Effective C++: 55 Specific Ways to Improve Your Programs and Designs
[6] Scott Meyers – More Effective C++: 35 New Ways to Improve Your Programs and Designs
[7] Bruce Eckel – Thinking in C++, Volume 1: Introduction to Standard C++
The Best Questions for Wouldbe C++ Programmers: Part 2
By zmij– Topcoder Member
In the second part of this installment we’ll tackle some questions regarding more advanced features of the language (the experienced
C++ programmers will consider some of these more on the basic side). So let’s get to it and work on the second part of this “interview”.
1. What are virtual functions?
Virtual functions represent the mechanism through which C++ implements the OO concept of polymorphism. Virtual functions
allow the programmer to redefine in each derived class functions from the base class with altered behavior so that you can
call the right function for the right object (allow to perform the right operation for an object through only a pointer/reference
to that object’s base class)
A member function is declared virtual by preceding its declaration (not the definition) with the virtual keyword
class Shape{
public:
...
//a shape can be drawn in many ways
virtual void draw(){ };
};
“A virtual function must be defined for the class in which it is first declared …” [Stroustrup]. The redefinition of a virtual
function in a derived class is called overriding (complete rewrite) or augmentation (rewrite but with a call to the base class
function)
class Rectangle : public Shape {
public:
...
void draw() { };
};
class Square : public Rectangle{
public:
...
void draw() { };
};
...
Shape* theShape = new Square();
// with the help of virtual functions
// a Square will be drawn and not
// a Rectangle or any other Shape
theShape‐>draw();
Through virtual functions C++ achieves what is called late binding (dynamic binding or runtime binding), that is actually
connecting a function call to a function body at runtime based on the type of the object and not at compilation time (static
binding) (**)
2. What is a virtual destructor and when would you use one?
A virtual destructor is a class’ destructor conforming to the C++’s polymorphism mechanism; by declaring the destructor
virtual you ensure it is placed in the VTABLE of the class and it will be called at proper times
You make a class’ destructor virtual to ensure proper cleanup when the class is supposed to be subclassed to form a
hierarchy and you want to delete a derived object thorough a pointer to it (the base class)
E.g. :
#include <vector>
#include <iostream>
using namespace std;
class Base{
public:
Base(const char* name);
// warning! the destructor should be virtual
~Base();
virtual void doStuff();
private:
const char* m_name;
};
Base :: Base(const char* name) : m_name(name){
}
Base :: ~Base(){
}
void Base :: doStuff(){
cout << "Doing stuff in Base" << endl;
}
class Derived : public Base{
public:
Derived(const char* name);
~Derived();
virtual void doStuff();
private:
vector<int>* m_charCodes;
};
Derived :: Derived(const char* name) : Base(name){
m_charCodes = new vector<int>;
}
Derived :: ~Derived(){
delete m_charCodes;
}
void Derived :: doStuff() {
cout << "Doing stuff in Derived" << endl;
}
int main(int argc, char* argv[]){
// assign the derived class object pointer to
// the base class pointer
char* theName = "Some fancy name";
Base* b = new Derived(theName);
// do some computations and then delete the
// pointer
delete b;
return 0;
}
What will happen in our rather lengthy example? Everything seems OK and most of the available C++ compilers will not
complain about anything (*). Nevertheless there is something pretty wrong here. The C++ standard is clear on this topic:
when you want to delete a derived class object through a base class pointer and the destructor of the base class
is not virtual the result is undefined. That means you’re on your own from there and the compiler won’t help you! What is
the most often behavior in such situations is that the derived class’ destructor is never called and parts of your derived object
are left undestroyed. In the example above you will leave behind a memory leak, the m_charCodes member will not be
destroyed because the destructor ~Derived() will not be called
A thing to notice is that declaring all destructors virtual is also pretty inefficient and not advisable. That makes sense
(declaring the destructor virtual) only if your class is supposed to be part of a hierarchy as a base class, otherwise you’ll just
waste memory with the class’ vtable generated only for the destructor. So declare a virtual destructor in a class “if and only
if that class is part of a class hierarchy, containing at least one virtual function. In other words, it is not
necessary for the class itself to have that virtual function – it is sufficient for one of its descendents to have
one.”[kyky]
3. How do you implement something along the lines of Java interfaces in C++?[kyky]
C++ as a language does not support the concept of “interfaces” (as opposed to other languages like Java or D for example),
but it achieves something similar through Abstract Classes
You obtain an abstract class in C++ by declaring at least one pure virtual function in that class. A virtual function is
transformed in a pure virtual with the help of the initializer “= 0″. A pure virtual function does not need a definition. An
abstract class cannot be instantiated but only used as a base in a hierarchy
class MySillyAbstract{
public:
// just declared not defined
virtual void beSilly() = 0;
};
A derivation from an abstract class must implement all the pure virtuals, otherwise it transforms itself into an abstract class
You can obtain an “interface” in C++ by declaring an abstract class with all the functions pure virtual functions and public
and no member variables – only behavior and no data
class IOInterface{
public:
virtual int open(int opt) = 0;
virtual int close(int opt) = 0;
virtual int read(char* p, int n) = 0;
virtual int write(const char* p, int n) = 0;
};
[adapted after an example found in Stroustup The C++ Programming Language 3rd Edition]
In this way you can specify and manipulate a variety of IO devices through the interface.
4. Could you point out some differences between pointers and references?
A reference must always be initialized because the object it refers to already exists; a pointer can be left uninitialized (though
is not recommended)
There’s no such thing a s a “NULL reference” because a reference must always refer to some object, so the “no object”
concept makes no sense; pointers, as we all know, can be NULL
References can be more efficient, because there’s no need to test the validity of a reference before using it (see above
comment); pointers most often have to be tested against NULL to ensure there are valid objects behind them
Pointers may be reassigned to point to different objects (except constant pointers, of course), but references cannot be
(references are “like” constant pointers that are automatically dereferenced by the compiler)
References are tied to someone else’s storage (memory) while pointers have their own storage they account for
One would use the dot operator “.” to access members of references to objects, however to access members of pointers to
objects one uses the arrow “>”[sql_lall]
5. When would you use a reference?
You should use a reference when you certainly know you have something to refer to, when you never want to refer to
anything else and when implementing operators whose syntactic requirements make the use of pointers undesirable; in all
other cases, “stick with pointers”
Do not use references just to reduce typing. That (and that being the sole reason) is not an appropriate usage of the
reference concept in C++; using references having in mind just the reason of reduced typing would lead you to a “reference
spree” – it must be clear in one’s mind when to use references and when to use pointers; overusing any of the two is an
inefficient path
6. Can you point out some differences between new & malloc?
“new” is an operator builtin into the C++ language, “malloc” is a function of the C standard library
“new” is aware of constructors/destructors, “malloc” is not; e.g. :
string* array1 = static_cast<string*>(malloc(10 * sizeof(string)));
free(array1);
array1 in the above example points to enough memory to hold 10 strings but no objects have been constructed and there’s
no easy and clean (proper) way from OO point of view to initialize them (see the question about placement new – in most
day to day programming tasks there’s no need to use such techniques). The call to free() deallocates the memory but does
not destroy the objects (supposing you managed to initialize them).
string* array2 = new string[10];
delete[] array2;
on the other hand array2 points to 10 fully constructed objects (they have not been properly initialized but they are
constructed), because “new” allocates memory and also calls the string default constructor for each object. The call to the
delete operator deallocates the memory and also destroys the objects
You got to remember to always use free() to release memory allocated with malloc() and delete (or the array
correspondent delete[]) to release memory allocated with new (or the array correspondent new[])
7. What are the differences between “operator new” and the “new” operator?
“new” is an operator built into the language and it’s meaning cannot be changed; “operator new” is a function and
manages how the “new” operator allocates memory its signature being: void* operator new(size_t size)
The “new” operator is allowed to call a constructor, because new has 2 major steps in achieving its goals : in step 1 it
allocates enough memory using “operator new” and then in step 2 calls the constructor(s) to construct the object(s) in the
memory that was allocated
“operator new” can be overridden meaning that you can change the way the “new” operator allocates memory, that is the
mechanism, but not the way the “new” operator behaves, that is it’s policy(semantics) , because what “new” does is fixed
by the language
8. What is “placement new”?
A special form of constructing an object in a given allocated zone of memory
The caller already knows what the pointer to the memory should be, because it knows where is supposed to be placed.
“placement new” returns the pointer that’s passed into it
Usage of “placement new” implies an explicit call to the object’s destructor when the object is to be deleted, because the
memory was allocated/obtained by other means than the standard “new” operator allocation
E.g. :
// supposing a "buffer" of memory large enough for
// the object we want to construct was
// previously allocated using malloc
MyClass* myObject = new (buffer) MyClass(string& name);
// !!ERROR
delete myObject;
// the correct way is
myObject‐>~MyClass();
// then the "buffer" must also be properly
// deallocated
free(buffer);
9. What is a “virtual constructor”?[kyky]
There is no such thing as a virtual constructor in C++ simply because you need to know the exact type of the object you want
to create and virtual represent the exact opposite concept (***)
But using an indirect way to create objects represents what is known as “Virtual Constructor Idiom”. For example you
could implement a clone() function as an indirect copy constructor or a create() member function as an indirect default
constructor (C++ FAQ Lite)
The GoF calls a variant of this idiom the Factory Method Pattern – “define an interface for creating an object, but let
subclasses decide which class to instantiate. Factory Method lets a class defer instantiation to subclasses”. A concrete example
will speak for itself:
[Created using the TopCoder UML Tool]
// Product
class Page{
};
// ConcreteProduct
class SkillsPage : public Page{
};
// ConcreteProduct
class ExperiencePage : public Page{
};
// ConcreteProduct
class IntroductionPage : public Page{
};
// ConcreteProduct
class TableOfContentsPage : public Page{
};
// Creator
class Document{
// Constructor calls abstract Factory method
public:
Document();
// Factory Method
virtual void CreatePages() { };
protected:
std::list<Page*> thePageList;
};
Document :: Document(){
CreatePages();
};
// ConcreteCreator
class Resume : public Document{
public:
// Factory Method implementation
void CreatePages();
};
// Factory Method implementation
void Resume :: CreatePages(){
thePageList.push_back(new SkillsPage());
thePageList.push_back(new ExperiencePage());
}
// ConcreteCreator
class Report : public Document{
public:
// Factory Method implementation
void CreatePages();
};
// Factory Method implementation
void Report :: CreatePages(){
thePageList.push_back(new TableOfContentsPage());
thePageList.push_back(new IntroductionPage());
}
int main(int argc, char* argv[]){
// Note: constructors call Factory Method
vector<Document*> documents(2);
documents[0] = new Resume();
documents[1] = new Report();
return 0;
}
10. What is RAII?
RAII – Resource Acquisition Is Initialization – is a C++ technique (but not limited to the C++ language) that combines
acquisition and release of resources with initialization and uninitialization of variables
E.g. :
// this is a hypothetic LogFile class using an
// hypothetic File class just for the illustration
// of the technique
class LogFile{
public:
LogFile(const char*);
~LogFile();
void write(const char*);
private:
File* m_file;
};
LogFile :: LogFile(const char* fileName) :
// ! acquisition and initialization
m_file(OpenFile(fileName)){
if(NULL == m_file){
throw FailedOpenException();
}
}
LogFile :: ~LogFile(){
// ! release and uninitialization
CloseFile(m_file);
}
void LogFile :: write(const char* logLine){
WriteFile(m_file, logLine);
}
// a hypothetical usage example
void SomeClass :: someMethod(){
LogFile log("log.tx");
log.write("I've been logged!");
// !exceptions can be thrown without
// worrying about closing the log file
// or leaking the file resource
if(...){
throw SomeException();
}
}
Without RAII each usage of the LogFile class would be also combined with the explicit management of the File resource. Also
in the presence of exceptions you would have to be careful and cleanup after yourself, thing that is taken care of with the
proper usage of RAII as illustrated in the example above
RAII is best used with languages that call the destructor for local objects when they go out of scope (implicit support of the
technique) like C++. In other languages, like Java & C#, that rely on the garbage collector to destruct local objects, you need
finalization routines (e.g. tryfinally blocks) to properly use RAII
Real usage examples of the technique are the C++ Standard Library’s file streams classes and STL’s auto_ptr class (to name
just very, very few)
That was it, folks! I hope that even if those questions did not pose any challenges, you still had fun doing/reading this quiz and
refreshing your memory on some aspects of the C++ language. Good luck on those interviews!
Notes
(*) bmerry suggested that my claim is not accurate but I’ve tested the example on Windows XP: Visual Studio 2005 Professional
Edition (the evaluation one that you can get from the Microsoft site ) did not warn, not even after setting the warnings level to Level 4
(Level 3 is the default one); Mingw compiler based on GCC (that comes with the Bloodshed DevCpp version 4.9.9.2) also did not warn
(the compiler settings from within the IDE are minimalist; tried to pass pedantic and Wextra to the compiler command line but still no
success); Digital Mars C++ compiler (dmc) also did not warn with all warnings turned on; Code Warrior Professional Edition 9 does not
warn also (this is pretty old, but Metrowerks compilers were renowned for the robustness and standard conformance). So, unless you
start digging through the documentation of those compilers to find that right command line switch or start writing the right code, you’re
in the harms way at least with the “out of the box” installations of these compilers.
(**) The compiler does all the magic: first, for each class that contains virtual functions (base and derived), the compiler creates a
static table called the VTABLE. Each virtual function will have a corresponding entry in that table (a function pointer); for the derived
classes the entries will contain the overridden virtual functions’ pointers. For each base class (it’s not static, each object will have it)
the compiler adds a hidden pointer called the VPTR, that will be initialized to point to the beginning of the VTABLE – in the derived
classes the (same) VPTR will be initialized to point to the beginning of the derived class’ VTABLE. So when “you make a virtual
function call through a base class pointer (that is, when you make a polymorphic call), the compiler quietly inserts code to fetch the
VPTR and look up the function address in the VTABLE, thus calling the correct function”. This might seem overly complicated but on a
typical machine it does not take much space and it’s very, very fast as a smart man said once “fetch, fetch call”.
(***) For that and other fine C++ gems go to Stroustrup.
References
[1] Bjarne Stroustrup – The C++ Programming Language Special 3rd Edition
[2] Stanley B. Lippman, Josee Lajoie, Barbara E. Moo – C++ Primer
[3] C++ FAQ Lite
[4] Gamma, Helm, Johnson, Vlissides (GoF) – Design Patterns Elements of Reusable ObjectOriented Software
[5] Herb Sutter – Exceptional C++: 47 Engineering Puzzles, Programming Problems, and Solutions
[6] Scott Meyers – Effective C++: 55 Specific Ways to Improve Your Programs and Designs
[7] Scott Meyers – More Effective C++: 35 New Ways to Improve Your Programs and Designs
[8] Bruce Eckel – Thinking in C++, Volume 1: Introduction to Standard C++
Primality Testing : Nondeterministic Algorithms
By innocentboy– TopCoder Member
Discuss this article in the forums
Introduction
Primality testing of a number is perhaps the most common problem concerning number theory that topcoders deal with. A prime
number is a natural number which has exactly two distinct natural number divisors: 1 and itself. Some basic algorithms and details
regarding primality testing and factorization can be found here.
The problem of detecting whether a given number is a prime number has been studied extensively but nonetheless, it turns out that all
the deterministic algorithms for this problem are too slow to be used in real life situations and the better ones amongst them are
tedious to code. But, there are some probabilistic methods which are very fast and very easy to code. Moreover, the probability of
getting a wrong result with these algorithms is so low that it can be neglected in normal situations.
This article discusses some of the popular probabilistic methods such as Fermat’s test, RabinMiller test, SolovayStrassen test.
Modular Exponentiation
All the algorithms which we are going to discuss will require you to efficiently compute (ab)%c ( where a,b,c are nonnegative integers
). A straightforward algorithm to do the task can be to iteratively multiply the result with ‘a’ and take the remainder with ‘c’ at each
step.
/* a function to compute (ab)%c */
int modulo(int a,int b,int c){
// res is kept as long long because intermediate results might overflow in "int"
long long res = 1;
for(int i=0;i<b;i++){
res *= a;
res %= c; // this step is valid because (a*b)%c = ((a%c)*(b%c))%c
}
return res%c;
}
However, as you can clearly see, this algorithm takes O(b) time and is not very useful in practice. We can do it in O( log(b) ) by using
what is called as exponentiation by squaring. The idea is very simple:
(a2)(b/2) if b is even and b > 0
ab = a*(a2)((b‐1)/2) if b is odd
1 if b = 0
This idea can be implemented very easily as shown below:
/* This function calculates (ab)%c */
int modulo(int a,int b,int c){
long long x=1,y=a; // long long is taken to avoid overflow of intermediate results
while(b > 0){
if(b%2 == 1){
x=(x*y)%c;
}
y = (y*y)%c; // squaring the base
b /= 2;
}
return x%c;
}
i
Notice that after i iterations, b becomes b/(2i), and y becomes (y(2 ))%c. Multiplying x with y is equivalent to adding 2i to the overall
power. We do this if the ith bit from right in the binary representation of b is 1. Let us take an example by computing (7107)%9. If we
use the above code, the variables after each iteration of the loop would look like this: ( a = 7, c = 9 )
iterations b x y
0 107 1 7
1 53 7 4
2 26 1 7
3 13 1 4
4 6 4 7
5 3 4 4
6 1 7 7
7 0 4 4
Now b becomes 0 and the return value of the function is 4. Hence (7107)%9 = 4.
The above code could only work for a,b,c in the range of type “int” or the intermediate results will run out of the range of “long long”. To
write a function for numbers up to 10^18, we need to compute (a*b)%c when computing a*b directly can grow larger than what a long
long can handle. We can use a similar idea to do that:
(2*a)*(b/2) if b is even and b > 0
a*b = a + (2*a)*((b‐1)/2) if b is odd
0 if b = 0
Here is some code which uses the idea described above ( you can notice that its the same code as exponentiation, just changing a
couple of lines ):
/* this function calculates (a*b)%c taking into account that a*b might overflow */
long long mulmod(long long a,long long b,long long c){
long long x = 0,y=a%c;
while(b > 0){
if(b%2 == 1){
x = (x+y)%c;
}
y = (y*2)%c;
b /= 2;
}
return x%c;
}
We could replace x=(x*y)%c with x = mulmod(x,y,c) and y = (y*y)%c with y = mulmod(y,y,c) in the original function for calculating
(ab)%c. This function requires that 2*c should be in the range of long long. For numbers larger than this, we could write our own BigInt
class ( java has an inbuilt one ) with addition, multiplication and modulus operations and use them.
This method for exponentiation could be further improved by using Montgomery Multiplication. Montgomery Multiplication algorithm is a
quick method to compute (a*b)%c, but since it requires some preprocessing, it doesn’t help much if you are just going to compute
one modular multiplication. But while doing exponentiation, we need to do the preprocessing for ‘c’ just once, that makes it a better
choice if you are expecting very high speed. You can read about it at the links mentioned in the reference section.
Similar technique can be used to compute (ab)%c in O(n3 * log(b)), where a is a square matrix of size n x n. All we need to do in this
case is manipulate all the operations as matrix operations. Matrix exponentiation is a very handy tool for your algorithm library and you
can see problems involving this every now and then.
Fermat Primality Test
Fermat’s Little Theorem
According to Fermat’s Little Theorem if p is a prime number and a is a positive integer less than p, then
ap = a ( mod p )
or alternatively:
a(p‐1) = 1 ( mod p )
Algorithm of the test
If p is the number which we want to test for primality, then we could randomly choose a, such that a < p and then calculate (a(p1))%p.
If the result is not 1, then by Fermat’s Little Theorem p cannot be prime. What if that is not the case? We can choose another a and
then do the same test again. We could stop after some number of iterations and if the result is always 1 in each of them, then we can
state with very high probability that p is prime. The more iterations we do, the higher is the probability that our result is correct. You
can notice that if the method returns composite, then the number is sure to be composite, otherwise it will be probably prime.
Given below is a simple function implementing Fermat’s primality test:
/* Fermat's test for checking primality, the more iterations the more is accuracy */
bool Fermat(long long p,int iterations){
if(p == 1){ // 1 isn't prime
return false;
}
for(int i=0;i<iterations;i++){
// choose a random integer between 1 and p‐1 ( inclusive )
long long a = rand()%(p‐1)+1;
// modulo is the function we developed above for modular exponentiation.
if(modulo(a,p‐1,p) != 1){
return false; /* p is definitely composite */
}
}
return true; /* p is probably prime */
}
More iterations of the function will result in higher accuracy, but will take more time. You can choose the number of iterations
depending upon the application.
Though Fermat is highly accurate in practice there are certain composite numbers p known as Carmichael numbers for which all values
of a<p for which gcd(a,p)=1, (a(p1))%p = 1. If we apply Fermat’s test on a Carmichael number the probability of choosing an a such
that gcd(a,p) != 1 is very low ( based on the nature of Carmichael numbers ), and in that case, the Fermat’s test will return a wrong
result with very high probability. Although Carmichael numbers are very rare ( there are about 250,000 of them less than 1016 ), but
that by no way means that the result you get is always correct. Someone could easily challenge you if you were to use Fermat’s test :).
Out of the Carmichael numbers less than 1016, about 95% of them are divisible by primes < 1000. This suggests that apart from
applying Fermat’s test, you may also test the number for divisibility with small prime numbers and this will further reduce the
probability of failing. However, there are other improved primality tests which don’t have this flaw as Fermat’s. We will discuss some of
them now.
MillerRabin Primality Test
Key Ideas and Concepts
1. Fermat’s Little Theorem.
2. If p is prime and x2 = 1 ( mod p ), then x = +1 or 1 ( mod p ). We could prove this as follows:
x2 = 1 ( mod p )
x2 ‐ 1 = 0 ( mod p )
(x‐1)(x+1) = 0 ( mod p )
Now if p does not divide both (x1) and (x+1) and it divides their product, then it cannot be a prime, which is a contradiction. Hence, p
will either divide (x1) or it will divide (x+1), so x = +1 or 1 ( mod p ).
Let us assume that p – 1 = 2d * s where s is odd and d >= 0. If p is prime, then either as = 1 ( mod p ) as in this case, repeated
r
squaring from as will always yield 1, so (a(p1))%p will be 1; or a(s*(2 )) = 1 ( mod p ) for some r such that 0 <= r < d, as repeated
squaring from it will always yield 1 and finally a(p1) = 1 ( mod p ). If none of these hold true, a(p1) will not be 1 for any prime number
a ( otherwise there will be a contradiction with fact #2 ).
Algorithm
Let p be the given number which we have to test for primality. First we rewrite p1 as (2d)*s. Now we pick some a in range [1,n1] and
r
then check whether as = 1 ( mod p ) or a(s*(2 )) = 1 ( mod p ). If both of them fail, then p is definitely composite. Otherwise p is
probably prime. We can choose another a and repeat the same test. We can stop after some fixed number of iterations and claim that
either p is definitely composite, or it is probably prime.
A small procedure realizing the above algorithm is given below:
/* Miller‐Rabin primality test, iteration signifies the accuracy of the test */
bool Miller(long long p,int iteration){
if(p<2){
return false;
}
if(p!=2 && p%2==0){
return false;
}
long long s=p‐1;
while(s%2==0){
s/=2;
}
for(int i=0;i<iteration;i++){
long long a=rand()%(p‐1)+1,temp=s;
long long mod=modulo(a,temp,p);
while(temp!=p‐1 && mod!=1 && mod!=p‐1){
mod=mulmod(mod,mod,p);
temp *= 2;
}
if(mod!=p‐1 && temp%2==0){
return false;
}
}
return true;
}
It can be shown that for any composite number p, at least (3/4) of the numbers less than p will witness p to be composite when chosen
as ‘a’ in the above test. Which means that if we do 1 iteration, probability that a composite number is returned as prime is (1/4). With k
iterations the probability of test failing is (1/4)k or 4(k). This test is comparatively slower compared to Fermat’s test but it doesn’t
break down for any specific composite numbers and 1820 iterations is a quite good choice for most applications.
SolovayStrassen Primality Test
Key Ideas and Concepts
1. Legendre Symbol: This symbol is defined for a pair of integers a and p such that p is prime. It is denoted by (a/p) and calculated
as:
= 0 if a%p = 0
(a/p) = 1 if there exists an integer k such that k2 = a ( mod p )
=‐1 otherwise.
It is proved by Euler that:
(a/p) = (a((p‐1)/2)) % p
So we can also say that:
(ab/p) = (ab((p‐1)/2)) % p = (a((p‐1)/2))%p * (b((p‐1)/2))%p = (a/p)*(b/p)
2. Jacobian Symbol: This symbol is a generalization of Legendre Symbol as it does not require ‘p’ to be prime. Let a and n be two
positive integers, and n = p1k1 * .. * pnkn, then Jacobian symbol is defined as:
(a/n) = ((a/p1)k1) * ((a/p2)k2) * ..... * ((a/pn)kn)
So you can see that if n is prime, the Jacobian symbol and Legendre symbol are equal.
There are some properties of these symbols which we can exploit to quickly calculate them:
2. (a/n) = 0 if gcd(a,n) != 1, Hence (0/n) = 0. This is because if gcd(a,n) != 1, then there must be some prime pi such that pi
divides both a and n. In that case (a/pi) = 0 [ by definition of Legendre Symbol ].
3. (ab/n) = (a/n) * (b/n). It can be easily derived from the fact (ab/p) = (a/p)(b/p) ( here (a/p) is the Legendry Symbol ).
4. if a is even, than (a/n) = (2/n)*((a/2)/n). It can be shown that:
= 1 if n = 1 ( mod 8 ) or n = 7 ( mod 8 )
(2/n) = ‐1 if n = 3 ( mod 8 ) or n = 5 ( mod 8 )
= 0 otherwise
5. (a/n) = (n/a)*(1((a1)(n1)/4)) if a and n are both odd.
The algorithm for the test is really simple. We can pick up a random a and compute (a/n). If n is a prime then (a/n) should be equal to
(a((n1)/2))%n [ as proved by Euler ]. If they are not equal then n is composite, and we can stop. Otherwise we can choose more
random values for a and repeat the test. We can declare n to be probably prime after some iterations.
Note that we are not interested in calculating Jacobi Symbol (a/n) if n is an even integer because we can trivially see that n isn’t prime,
except 2 of course.
Let us write a little code to compute Jacobian Symbol (a/n) and for the test:
//calculates Jacobian(a/n) n>0 and n is odd
int calculateJacobian(long long a,long long n){
if(!a) return 0; // (0/n) = 0
int ans=1;
long long temp;
if(a<0){
a=‐a; // (a/n) = (‐a/n)*(‐1/n)
if(n%4==3) ans=‐ans; // (‐1/n) = ‐1 if n = 3 ( mod 4 )
}
if(a==1) return ans; // (1/n) = 1
while(a){
if(a<0){
a=‐a; // (a/n) = (‐a/n)*(‐1/n)
if(n%4==3) ans=‐ans; // (‐1/n) = ‐1 if n = 3 ( mod 4 )
}
while(a%2==0){
a=a/2; // Property (iii)
if(n%8==3||n%8==5) ans=‐ans;
}
swap(a,n); // Property (iv)
if(a%4==3 && n%4==3) ans=‐ans; // Property (iv)
a=a%n; // because (a/p) = (a%p / p ) and a%pi = (a%n)%pi if n % pi = 0
if(a>n/2) a=a‐n;
}
if(n==1) return ans;
return 0;
}
/* Iterations determine the accuracy of the test */
bool Solovoy(long long p,int iteration){
if(p<2) return false;
if(p!=2 && p%2==0) return false;
for(int i=0;i<iteration;i++){
long long a=rand()%(p‐1)+1;
long long jacobian=(p+calculateJacobian(a,p))%p;
long long mod=modulo(a,(p‐1)/2,p);
if(!jacobian || mod!=jacobian){
return false;
}
}
return true;
}
It is shown that for any composite n, at least half of the a will result in n being declared as composite according to SolovayStrassen
test. This shows that the probability of getting a wrong result after k iterations is (1/2)k. However, it is generally less preferred than
RabinMiller test in practice because it gives poorer performance.
The various routines provided in the article can be highly optimized just by using bitwise operators instead of them. For example /= 2
can be replaced by “>>= 1″, “%2″ can be replaced by “&1″ and “*= 2″ can be replaced by “<<=1″. Inline Assembly can also be used
to optimize them further.
Practice Problems
Problems involving nondeterministic primality tests are not very suitable for the SRM format. But problems involving modular
exponentiation and matrix exponentiation are common. Here are some of the problems where you can apply the methods studied
above:
PowerDigit ( TCO 06 Online Round 2 )
MarbleMachine ( SRM 376 )
DrivingAround ( SRM 342 )
PON
PRIC
SOLSTRAS
DIVSUM2 [ this one also involves Pollard's Rho Algorithm ]
References and Further Reading
http://www.cse.iitk.ac.in/users/manindra/algebra/primality_v6.pdf
http://security.ece.orst.edu/papers/j37acmon.pdf
http://icpc.baylor.edu/Past/icpc2004/RegReport/guan.cse.nsysu.edu.tw/data/montg.pdf
Assignment Problem and Hungarian Algorithm
By xray– TopCoder Member
Discuss this article in the forums
Introduction
Are you familiar with the following situation? You open the Div I Medium and don’t know how to approach it, while a lot of people in
your room submitted it in less than 10 minutes. Then, after the contest, you find out in the editorial that this problem can be simply
reduced to a classical one. If yes, then this tutorial will surely be useful for you.
Problem statement
In this article we’ll deal with one optimization problem, which can be informally defined as:
Assume that we have N workers and N jobs that should be done. For each pair (worker, job) we know salary that should be paid to
worker for him to perform the job. Our goal is to complete all jobs minimizing total inputs, while assigning each worker to exactly one
job and vice versa.
Converting this problem to a formal mathematical definition we can form the following equations:
– cost matrix, where cij – cost of worker i to perform job j.
– resulting binary matrix, where xij = 1 if and only if ith worker is assigned to jth job.
– one worker to one job assignment.
– one job to one worker assignment.
– total cost function.
We can also rephrase this problem in terms of graph theory. Let’s look at the job and workers as if they were a bipartite graph, where
each edge between the ith worker and jth job has weight of cij. Then our task is to find minimumweight matching in the graph (the
matching will consists of N edges, because our bipartite graph is complete).
Small example just to make things clearer:
General description of the algorithm
This problem is known as the assignment problem. The assignment problem is a special case of the transportation problem, which in
turn is a special case of the mincost flow problem, so it can be solved using algorithms that solve the more general cases. Also, our
problem is a special case of binary integer linear programming problem (which is NPhard). But, due to the specifics of the problem,
there are more efficient algorithms to solve it. We’ll handle the assignment problem with the Hungarian algorithm (or KuhnMunkres
algorithm). I’ll illustrate two different implementations of this algorithm, both graph theoretic, one easy and fast to implement with
O(n4) complexity, and the other one with O(n3) complexity, but harder to implement.
There are also implementations of Hungarian algorithm that do not use graph theory. Rather, they just operate with cost matrix, making
different transformation of it (see [1] for clear explanation). We’ll not touch these approaches, because it’s less practical for TopCoder
needs.
O(n4) algorithm explanation
As mentioned above, we are dealing with a bipartite graph. The main idea of the method is the following: consider we’ve found the
perfect matching using only edges of weight 0 (hereinafter called “0weight edges”). Obviously, these edges will be the solution of the
assignment problem. If we can’t find perfect matching on the current step, then the Hungarian algorithm changes weights of the
available edges in such a way that the new 0weight edges appear and these changes do not influence the optimal solution.
To clarify, let’s look at the stepbystep overview:
Step 0)
A. For each vertex from left part (workers) find the minimal outgoing edge and subtract its weight from all weights connected with this
vertex. This will introduce 0weight edges (at least one).
B. Apply the same procedure for the vertices in the right part (jobs).
Actually, this step is not necessary, but it decreases the number of main cycle iterations.
Step 1)
A. Find the maximum matching using only 0weight edges (for this purpose you can use maxflow algorithm, augmenting path
algorithm, etc.).
B. If it is perfect, then the problem is solved. Otherwise find the minimum vertex cover V (for the subgraph with 0weight edges only),
the best way to do this is to use Köning’s graph theorem.
Step 2) Let and adjust the weights using the following rule:
Step 3) Repeat Step 1 until solved.
But there is a nuance here; finding the maximum matching in step 1 on each iteration will cause the algorithm to become O(n5). In
order to avoid this, on each step we can just modify the matching from the previous step, which only takes O(n2) operations.
It’s easy to see that no more than n2 iterations will occur, because every time at least one edge becomes 0weight. Therefore, the
overall complexity is O(n4).
O(n3) algorithm explanation
Warning! In this section we will deal with the maximumweighted matching problem. It’s obviously easy to transform minimum problem
to the maximum one, just by setting:
or
.
Before discussing the algorithm, let’s take a look at some of the theoretical ideas. Let’s start off by considering we have a complete
bipartite graph G=(V,E) where and , w(x,y) – weight of edge (x,y).
Vertex and set neighborhood
Vertex labeling
This is simply a function (for each vertex we assign some number called a label). Let’s call this labeling feasible if it satisfies
the following condition: . In other words, the sum of the labels of the vertices on both sides of a given
edge are greater than or equal to the weight of that edge.
Equality subgraph
Let Gl=(V,El) be a spanning subgraph of G (in other words, it includes all vertices from G). If G only those edges (x,y) which satisfy
the following condition: , then it is an equality subgraph. In other words, it only includes
those edges from the bipartite matching which allow the vertices to be perfectly feasible.
Now we’re ready for the theorem which provides the connection between equality subgraphs and maximumweighted matching:
If M* is a perfect matching in the equality subgraph Gl, then M* is a maximumweighted matching in G.
The proof is rather straightforward, but if you want you can do it for practice. Let’s continue with a few final definitions:
Alternating path and alternating tree
Consider we have a matching M ( ).
(In the diagram below, W1, W2, W3, J1, J3, J4 are matched, W4, J2 are exposed)
Path P is called alternating if its edges alternate between M and EM. (For example, (W4, J4, W3, J3, W2, J2) and (W4, J1, W1) are
alternating paths)
If the first and last vertices in alternating path are exposed, it is called augmenting (because we can increment the size of the matching
by inverting edges along this path, therefore matching unmatched edges and vice versa). ((W4, J4, W3, J3, W2, J2) – augmenting
alternating path)
A tree which has a root in some exposed vertex, and a property that every path starting in the root is alternating, is called an
alternating tree. (Example on the picture above, with root in W4)
That’s all for the theory, now let’s look at the algorithm:
First let’s have a look on the scheme of the Hungarian algorithm:
Step 0. Find some initial feasible vertex labeling and some initial matching.
(1)
and replace existing labeling with the next one:
(2)
Now replace with
Step 3. Find some vertex . If y is exposed then an alternating path from x (root of the tree) to y exists, augment
matching along this path and go to step 1. If y is matched in M with some vertex z add (z,y) to the alternating tree and set
, go to step 2.
And now let’s illustrate these steps by considering an example and writing some code.
As an example we’ll use the previous one, but first let’s transform it to the maximumweighted matching problem, using the second
method from the two described above. (See Picture 1)
Picture 1
Here are the global variables that will be used in the code:
#define N 55 //max number of vertices in one part
#define INF 100000000 //just infinity
int cost[N][N]; //cost matrix
int n, max_match; //n workers and n jobs
int lx[N], ly[N]; //labels of X and Y parts
int xy[N]; //xy[x] ‐ vertex that is matched with x,
int yx[N]; //yx[y] ‐ vertex that is matched with y
bool S[N], T[N]; //sets S and T in algorithm
int slack[N]; //as in the algorithm description
int slackx[N]; //slackx[y] such a vertex, that
// l(slackx[y]) + l(y) ‐ w(slackx[y],y) = slack[y]
int prev[N];//array for memorizing alternating paths
Step 0:
It’s easy to see that next initial labeling will be feasible:
And as an initial matching we’ll use an empty one. So we’ll get equality subgraph as on Picture 2. The code for initializing is quite easy,
but I’ll paste it for completeness:
void init_labels(){
memset(lx, 0, sizeof(lx));
memset(ly, 0, sizeof(ly));
for (int x = 0; x < n; x++)
for (int y = 0; y < n; y++)
lx[x] = max(lx[x], cost[x][y]);
}
The next three steps will be implemented in one function, which will correspond to a single iteration of the algorithm. When the
algorithm halts, we will have a perfect matching, that's why we'll have n iterations of the algorithm and therefore (n+1) calls of the
function.
Step 1
According to this step we need to check whether the matching is already perfect, if the answer is positive we just stop algorithm,
otherwise we need to clear S, T and alternating tree and then find some exposed vertex from the X part. Also, in this step we are
initializing a slack array, I'll describe it on the next step.
void augment(){ //main function of the algorithm
if (max_match == n) return; //check wether matching is already perfect
int x, y, root; //just counters and root vertex
int q[N], wr = 0, rd = 0; //q ‐ queue for bfs, wr,rd ‐ write and read
//pos in queue
memset(S, false, sizeof(S)); //init set S
memset(T, false, sizeof(T)); //init set T
memset(prev, ‐1, sizeof(prev)); //init set prev ‐ for the alternating tree
for (x = 0; x < n; x++) //finding root of the tree
if (xy[x] == ‐1){
q[wr++] = root = x;
prev[x] = ‐2;
S[x] = true;
break;
}
for (y = 0; y < n; y++){//initializing slack array
slack[y] = lx[root] + ly[y] ‐ cost[root][y];
slackx[y] = root;
}
Step 2
On this step, the alternating tree is completely built for the current labeling, but the augmenting path hasn't been found yet, so we
need to improve the labeling. It will add new edges to the equality subgraph, giving an opportunity to expand the alternating tree. This
is the main idea of the method; we are improving the labeling until we find an augmenting path in the equality graph corresponding to
the current labeling. Let's turn back to step 2. There we just change labels using formulas (1) and (2), but using them in an obvious
manner will cause the algorithm to have O(n4) time. So, in order to avoid this we use a slack array initialized in O(n) time because we
only augment the array created in step 1:
Then we just need O(n) to calculate a delta Δ (see (1)):
Updating slack:
1) On step 3, when vertex x moves from XS to S, this takes O(n).
2) On step 2, when updating labeling, it's also takes O(n), because:
So we get O(n) instead of O(n2) as in the straightforward approach.
Here's code for the label updating function:
void update_labels(){
int x, y, delta = INF; //init delta as infinity
for (y = 0; y < n; y++) //calculate delta using slack
if (!T[y])
delta = min(delta, slack[y]);
for (x = 0; x < n; x++) //update X labels
if (S[x]) lx[x] ‐= delta;
for (y = 0; y < n; y++) //update Y labels
if (T[y]) ly[y] += delta;
for (y = 0; y < n; y++) //update slack array
if (!T[y])
slack[y] ‐= delta;
}
Step 3
In step 3, first we build an alternating tree starting from some exposed vertex, chosen at the beginning of each iteration. We will do
this using breadthfirst search algorithm. If on some step we meet an exposed vertex from the Y part, then finally we can augment our
path, finishing up with a call to the main function of the algorithm. So the code will be the following:
1) Here's the function that adds new edges to the alternating tree:
void add_to_tree(int x, int prevx){
//x ‐ current vertex,prevx ‐ vertex from X before x in the alternating path,
//so we add edges (prevx, xy[x]), (xy[x], x)
S[x] = true; //add x to S
prev[x] = prevx; //we need this when augmenting
for (int y = 0; y < n; y++) //update slacks, because we add new vertex to S
if (lx[x] + ly[y] ‐ cost[x][y] < slack[y]){
slack[y] = lx[x] + ly[y] ‐ cost[x][y];
slackx[y] = x;
}
}
3) And now, the end of the augment() function:
//second part of augment() function
while (true){ //main cycle
while (rd < wr){ //building tree with bfs cycle
x = q[rd++]; //current vertex from X part
for (y = 0; y < n; y++) //iterate through all edges in equality graph
if (cost[x][y] == lx[x] + ly[y] && !T[y]){
if (yx[y] == ‐1)
break;
//an exposed vertex in Y found, so
//augmenting path exists!
T[y] = true;
//else just add y to T,
q[wr++] = yx[y]; //add vertex yx[y], which is matched
//with y, to the queue
add_to_tree(yx[y], x); //add edges (x,y) and (y,yx[y]) to the tree
}
if (y < n) break; //augmenting path found!
}
if (y < n)
break; //augmenting path found!
update_labels(); //augmenting path not found, so improve labeling
wr = rd = 0;
for (y = 0; y < n; y++)
//in this cycle we add edges that were added to the equality graph as a
//result of improving the labeling, we add edge (slackx[y], y) to the tree if
//and only if !T[y] && slack[y] == 0, also with this edge we add another one
//(y, yx[y]) or augment the matching, if y was exposed
if (!T[y] && slack[y] == 0){
if (yx[y] == ‐1){ //exposed vertex in Y found ‐ augmenting path exists!
x = slackx[y];
break;
}
else{
T[y] = true; //else just add y to T,
if (!S[yx[y]]) {
q[wr++] = yx[y]; //add vertex yx[y], which is matched with
//y, to the queue
add_to_tree(yx[y], slackx[y]); //and add edges (x,y) and (y,
//yx[y]) to the tree
}
}
}
if (y < n) break; //augmenting path found!
}
if (y < n){ //we found augmenting path!
max_match++; //increment matching
//in this cycle we inverse edges along augmenting path
for (int cx = x, cy = y, ty; cx != ‐2; cx = prev[cx], cy = ty){
ty = xy[cx];
yx[cy] = cx;
xy[cx] = cy;
}
augment(); //recall function, go to step 1 of the algorithm
}
}//end of augment() function
The only thing in code that hasn't been explained yet is the procedure that goes after labels are updated. Say we've updated labels and
now we need to complete our alternating tree; to do this and to keep algorithm in O(n3) time (it's only possible if we use each edge no
more than one time per iteration) we need to know what edges should be added without iterating through all of them, and the answer
for this question is to use BFS to add edges only from those vertices in Y, that are not in T and for which slack[y] = 0 (it's easy to
prove that in such way we'll add all edges and keep algorithm to be O(n3)). See picture below for explanation:
At last, here's the function that implements Hungarian algorithm:
int hungarian(){
int ret = 0; //weight of the optimal matching
max_match = 0; //number of vertices in current matching
memset(xy, ‐1, sizeof(xy));
memset(yx, ‐1, sizeof(yx));
init_labels(); //step 0
augment(); //steps 1‐3
for (int x = 0; x < n; x++) //forming answer there
ret += cost[x][xy[x]];
return ret;
}
To see all this in practice let's complete the example started on step 0.
→ → →
→ → →
→ → →
→ → →
Build Augmenting
alternating tree path found
Optimal matching found
→ →
Finally, let's talk about the complexity of this algorithm. On each iteration we increment matching so we have n iterations. On each
iterations each edge of the graph is used no more than one time when finding augmenting path, so we've got O(n2) operations.
Concerning labeling we update slack array each time when we insert vertex from X into S, so this happens no more than n times per
iteration, updating slack takes O(n) operations, so again we've got O(n2). Updating labels happens no more than n time per iterations
(because we add at least one vertex from Y to T per iteration), it takes O(n) operations again O(n2). So total complexity of this
implementation is O(n3).
Some practice
For practice let's consider the medium problem from SRM 371 (div. 1). It's obvious we need to find the maximumweighted matching in
graph, where the X part is our players, the Y part is the opposing club players, and the weight of each edge is:
Though this problem has a much simpler solution, this one is obvious and fast coding can bring more points.
Try this one for more practice. I hope this article has increased the wealth of your knowledge in classical algorithms… Good luck and
have fun!
References
1. Mike Dawes "The Optimal Assignment Problem"
2. Mordecaj J. Golin "Bipartite Matching and the Hungarian Method"
3. Samir Khuller "Design and Analysis of Algorithms: Course Notes"
4. Lawler E.L. "Combinatorial Optimization: Networks and Matroids"