A3c5 PDF

Alexs Anthology of Algorithms
Common Code for Contests in Concise C++

(Draft, December 2015)
Compiled, Edited, and Written by

Alex Li
University of Toronto
Email: [email protected]
This page is intentionally left blank.
Preface
Note: Visit http://github.com/Alextrovert/Algorithm-Anthology for the most up-to-date digital

version of this codebook. The version you are reading is currently being reviewed, revised, and rewritten.
0.1 Introduction
This anthology started as a personal project to implement common algorithms in the most concise and
vanilla way possible so that theyre easily adaptable for use in algorithm competitions. To that end,
several properties of the algorithm implementations should be satisfied, not limited to the following:
Implementations must be clear. There is no time to write rigorous documentation within contests.
This makes it all the more important to make class and variable names reflexive of what they
represent. Clarity must also be carefully balanced with not making them too long-winded, since it
can be just as time-consuming to type out long identifiers.
Implementations must be generic. The more code that must be changed during the contest, the
more room there is for mistakes. Thus, it should be easy to apply implementations to different
purposes. C++ templates are often used to accomplish this at the slight cost of readability.
Implementations must be portable. Different contest environments use different versions of C++
(though almost all of them use GCC), so in order to make programs as compatible as possible,
non-standard features should be avoided. This is also why no features from C++0x or above are
used, since many constest systems remain stuck on older versions of the language. Refer to the
Portability section below for more information.
Implementations must be efficient. The code cannot simply demonstrate an idea, it should also
have the correct running time and a reasonably low constant overhead. This is sometimes
challenging if concision is to be preserved. However, contest problem setters will often be
understanding and set time limits liberally. If an implementation from here does not pass in time,
chances are you are choosing the wrong algorithm.
Implementations must be concise. During timed contests, code chunks are often moved around the
file. To minimize the amount of scrolling, code design and formatting conventions should ensure as
much code fits on the screen as possible (while not excessively sacrificing readability). Its a given
that each algorithm should be placed within singleton files. Nearly all contest environments
demand submissions to be contained within a single file.
A good trade-off between clarity, genericness, portability, efficiency, and concision is what comprises the
ultimate goal of adaptability.
i
ii Preface
0.2 Portability
All programs are tested with version 4.7.3 of the GNU Compiler Collection (GCC) compiled for a 32-bit
target system.
That means the following assumptions are made:
bool and char are 8-bit
int and float are 32-bit
double and long long are 64-bit
long double is 96-bit
Programs are highly portable (ISO C++ 1998 compliant), except in the following regards:
Usage of long long and related features [-Wlong-long] (such as LLONG MIN in <climits>), which
are compliant in C99/C++0x or later. 64-bit integers are a must for many programming contest
problems, so it is necessary to include these.
Usage of variable sized arrays [-Wvla] (an easy fix using vectors, but I chose to keep it because it
is simpler and because dynamic memory is generally good to avoid in contests)
Usage of GCCs built-in functions like builtin popcount() and builtin clz(). These can be
extremely convenient, and are easily implemented if theyre not available. See here for a reference:
https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html
Usage of compound-literals, e.g. vec.push back((mystruct){a, b, c}). This is used in the
anthology because it makes code much more concise by not having to define a constructor (which
is trivial to do).
Ad-hoc cases where bitwise hacks are intentionally used, such as functions for getting the signbit
with type-puned pointers. If you are looking for these features, chances are you dont care about
portability anyway.
0.3 Usage Notes
The primary purpose of this project is not to better your understanding of algorithms. To take
advantage of this anthology, you must have prior understanding of the algorithms in question. In each
source code file, you will find brief descriptions and simple examples to clarify how the functions and
classes should be used (not so much how they work). This is why if you actually want to learn
algorithms, you are better off researching the idea and trying to implement it independently. Directly
using the code found here should be considered a last resort during the pressures of an actual contest.
All information from the comments (descriptions, complexities, etc.) come from Wikipedia and other
online sources. Some programs here are direct implementations of pseudocode found online, while others
are adaptated and translated from informatics books and journals. If references for a program are not
listed in its comments, you may assume that I have written them from scratch. You are free to use,
modify, and distribute these programs in accordance to the license, but please first examine any
corresponding references of each program for more details on usage and authorship.
Cheers and hope you enjoy!
Alex Li
December, 2015
Contents
Preface i
0.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
0.2 Portability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
0.3 Usage Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
1 Elementary Algorithms 1
1.1 Array Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Sorting Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Array Rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.3 Counting Inversions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.4 Coordinate Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.1.5 Selection (Quickselect) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Array Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.1 Longest Increasing Subsequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.2 Maximal Subarray Sum (Kadanes) . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.3 Majority Element (Boyer-Moore) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.4 Subset Sum (Meet-in-the-Middle) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.2.5 Maximal Zero Submatrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3 Searching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.1 Discrete Binary Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.2 Ternary Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.3 Hill Climbing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.3.4 Convex Hull Trick (Semi-Dynamic) . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3.5 Convex Hull Trick (Fully Dynamic) . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4 Cycle Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.4.1 Floyds Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
iii
iv Contents
1.4.2 Brents Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

1.5 Binary Exponentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2 Graph Theory 37
2.1 Depth-First Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.1.1 Graph Class and Depth-First Search . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.1.2 Topological Sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.1.3 Eulerian Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.1.4 Unweighted Tree Centers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2 Shortest Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2.1 Breadth First Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2.2 Dijkstras Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.2.3 Bellman-Ford Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2.4 Floyd-Warshall Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.3 Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.3.1 Strongly Connected Components (Kosarajus Algorithm) . . . . . . . . . . . . . . 51
2.3.2 Strongly Connected Components (Tarjans Algorithm) . . . . . . . . . . . . . . . . 53
2.3.3 Bridges, Cut-points, and Biconnectivity . . . . . . . . . . . . . . . . . . . . . . . . 55
2.4 Minimal Spanning Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.4.1 Prims Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.4.2 Kruskals Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.5 Maximum Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.5.1 Ford-Fulkerson Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.5.2 Edmonds-Karp Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.5.3 Dinics Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.5.4 Push-Relabel Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.6 Backtracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.6.1 Max Clique (Bron-Kerbosch Algorithm) . . . . . . . . . . . . . . . . . . . . . . . . 68
2.6.2 Graph Coloring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.7 Maximum Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.7.1 Maximum Bipartite Matching (Kuhns Algorithm) . . . . . . . . . . . . . . . . . . 72
2.7.2 Maximum Bipartite Matching (Hopcroft-Karp Algorithm) . . . . . . . . . . . . . . 73
2.7.3 Maximum Graph Matching (Edmondss Algorithm) . . . . . . . . . . . . . . . . . 75
2.8 Hamiltonian Path and Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.8.1 Shortest Hamiltonian Cycle (Travelling Salesman) . . . . . . . . . . . . . . . . . . 77
Contents v
2.8.2 Shortest Hamiltonian Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3 Data Structures 81
3.1 Disjoint Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.1.1 Disjoint Set Forest (Simple) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.1.2 Disjoint Set Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.2 Fenwick Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.2.1 Simple Fenwick Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.2.2 Fenwick Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2.3 Fenwick Tree (Point Query) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.2.4 Fenwick Tree (Range Update) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.2.5 Fenwick Tree (Map) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.2.6 2D Fenwick Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.2.7 2D Fenwick Tree (Range Update) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.3 1D Range Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.3.1 Simple Segment Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.3.2 Segment Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.3.3 Segment Tree (Range Updates) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3.4 Segment Tree (Fast, Non-recursive) . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.3.5 Implicit Treap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.3.6 Sparse Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.3.7 Square Root Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.3.8 Interval Tree (Augmented Treap) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.4 2D Range Queries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.4.1 Quadtree (Simple) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.4.2 Quadtree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.4.3 2D Segment Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.4.4 K-d Tree (2D Range Query) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.4.5 K-d Tree (Nearest Neighbor) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.4.6 R-Tree (Nearest Segment) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.4.7 2D Range Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.5 Search Trees and Alternatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.5.1 Binary Search Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.5.2 Treap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.5.3 Size Balanced Tree (Order Statistics) . . . . . . . . . . . . . . . . . . . . . . . . . . 133
vi Contents
3.5.4 Hashmap (Chaining) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

3.5.5 Skip List (Probabilistic) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
3.6 Tree Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
3.6.1 Heavy-Light Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
3.6.2 Link-Cut Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
3.7 Lowest Common Ancestor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.7.1 Sparse Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.7.2 Segment Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4 Mathematics 155
4.1 Mathematics Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
4.2 Combinatorics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4.2.1 Combinatorial Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4.2.2 Enumerating Arrangements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.2.3 Enumerating Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.2.4 Enumerating Combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
4.2.5 Enumerating Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
4.2.6 Enumerating Generic Combinatorial Sequences . . . . . . . . . . . . . . . . . . . . 180
4.3 Number Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
4.3.1 GCD, LCM, Mod Inverse, Chinese Remainder . . . . . . . . . . . . . . . . . . . . 183
4.3.2 Generating Primes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
4.3.3 Primality Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
4.3.4 Integer Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
4.3.5 Eulers Totient Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
4.4 Arbitrary Precision Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
4.4.1 Big Integers (Simple) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
4.4.2 Big Integer and Rational Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
4.4.3 FFT and Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
4.5 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
4.5.1 Matrix Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
4.5.2 Determinant (Gauss) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
4.5.3 Gaussian Elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
4.5.4 LU Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
4.5.5 Simplex Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
4.6 Root-Finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Contents vii
4.6.1 Real Root Finding (Differentiation) . . . . . . . . . . . . . . . . . . . . . . . . . . 228

4.6.2 Complex Root Finding (Laguerres) . . . . . . . . . . . . . . . . . . . . . . . . . . 230
4.6.3 Complex Root Finding (RPOLY) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
4.7 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
4.7.1 Simpsons Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
5 Geometry 242
5.1 Geometric Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
5.1.1 Point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
5.1.2 Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
5.1.3 Circle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
5.2 Geometric Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5.2.1 Angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5.2.2 Distances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
5.2.3 Line Intersections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
5.2.4 Circle Intersections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
5.3 Common Geometric Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
5.3.1 Polygon Sorting and Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
5.3.2 Point in Polygon Query . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
5.3.3 Convex Hull . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
5.3.4 Minimum Enclosing Circle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
5.3.5 Diameter of Point Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
5.3.6 Closest Point Pair . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
5.3.7 Segment Intersection Finding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
5.4 Advanced Geometric Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
5.4.1 Convex Polygon Cut . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
5.4.2 Polygon Union and Intersection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
5.4.3 Delaunay Triangulation (Simple) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
5.4.4 Delaunay Triangulation (Fast) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
6 Strings 292
6.1 Strings Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
6.2 Expression Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
6.2.1 Recursive Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
6.2.2 Recursive Descent (Simple) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
viii Contents
6.2.3 Shunting Yard Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299

6.3 String Searching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
6.3.1 Longest Common Substring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
6.3.2 Longest Common Subsequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
6.3.3 Edit Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
6.4 Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
6.4.1 Longest Common Substring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
6.4.2 Longest Common Subsequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
6.4.3 Edit Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
6.5 Suffix Array and LCP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
6.5.1 O(N log 2 N ) Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
6.5.2 O(N log N ) Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
6.5.3 O(N log N ) Construction (DC3/Skew) . . . . . . . . . . . . . . . . . . . . . . . . . 312
6.6 String Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
6.5.1 Simple Trie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
6.5.2 Radix Trie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
6.5.3 Suffix Trie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
6.5.4 Suffix Automaton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
Chapter 1
Elementary Algorithms
1.1 Array Transformations
1.1.1 Sorting Algorithms
1 /*
2
3 The following functions are to be used like std::sort(), taking two
4 RandomAccessIterators as the range to be sorted, and optionally a
5 comparison function object to replace the default < operator.
6
7 They are not intended to compete with the standard library sorting
8 functions in terms of speed, but are merely demonstrations of how to
9 implement common sorting algorithms concisely in C++.
10
11 */
12
13 #include <algorithm> /* std::copy(), std::swap() */
14 #include <functional> /* std::less */
15 #include <iterator> /* std::iterator_traits */
16
17 /*
18
19 Quicksort
20
21 Quicksort repeatedly selects a pivot and "partitions" the range so that
22 all values comparing less than the pivot come before it, and all values
23 comparing greater comes after it. Divide and conquer is then applied to
24 both sides of the pivot until the original range is sorted. Despite
25 having a worst case of O(n^2), quicksort is faster in practice than
26 merge sort and heapsort, which each has a worst case of O(n log n).
27
28 The pivot chosen in this implementation is always the middle element
29 of the range to be sorted. To reduce the likelihood of encountering the
30 worst case, the algorithm should be modified to select a random pivot,
31 or use the "median of three" method.
32
33 Time Complexity (Average): O(n log n)
34 Time Complexity (Worst): O(n^2)
35 Space Complexity: O(log n) auxiliary.
1
2 Chapter 1. Elementary Algorithms
36 Stable?: No
37
38 */
39
40 template<class It, class Compare>
41 void quicksort(It lo, It hi, Compare comp) {
42 if (hi - lo < 2) return;
43 typedef typename std::iterator_traits<It>::value_type T;
44 T pivot = *(lo + (hi - lo) / 2);
45 It i, j;
46 for (i = lo, j = hi - 1; ; ++i, --j) {
47 while (comp(*i, pivot))
48 ++i;
49 while (comp(pivot, *j))
50 --j;
51 if (i >= j)
52 break;
53 std::swap(*i, *j);
54 }
55 quicksort(lo, i, comp);
56 quicksort(i, hi, comp);
57 }
58
59 template<class It> void quicksort(It lo, It hi) {
61 quicksort(lo, hi, std::less<T>());
62 }
63
64 /*
65
66 Merge Sort
67
68 Merge sort works by first dividing a list into n sublists, each with
69 one element, then recursively merging sublists to produce new sorted
70 sublists until only a single sorted sublist remains. Merge sort has a
71 better worse case than quicksort, and is also stable, meaning that it
72 will preserve the relative ordering of elements considered equal by
73 the < operator or comparator (a < b and b < a both return false).
74
75 While std::stable_sort() is a corresponding function in the standard
76 library, the implementation below differs in that it will simply fail
77 if extra memory is not available. Meanwhile, std::stable_sort() will
78 not fail, but instead fall back to a time complexity of O(n log^2 n).
79
81 Time Complexity (Worst): O(n log n)
82 Space Complexity: O(n) auxiliary.
83 Stable?: Yes
84
85 */
86
88 void mergesort(It lo, It hi, Compare comp) {
89 if (hi - lo < 2) return;
90 It mid = lo + (hi - lo - 1) / 2, a = lo, c = mid + 1;
91 mergesort(lo, mid + 1, comp);
92 mergesort(mid + 1, hi, comp);
94 T *buf = new T[hi - lo], *b = buf;
1.1. Array Transformations 3
95 while (a <= mid && c < hi)

96 *(b++) = comp(*c, *a) ? *(c++) : *(a++);
97 if (a > mid) {
98 for (It k = c; k < hi; k++)
99 *(b++) = *k;
100 } else {
101 for (It k = a; k <= mid; k++)
102 *(b++) = *k;
103 }
104 for (int i = hi - lo - 1; i >= 0; i--)
105 *(lo + i) = buf[i];
106 delete[] buf;
107 }
108
109 template<class It> void mergesort(It lo, It hi) {
111 mergesort(lo, hi, std::less<T>());
112 }
113
114 /*
115
116 Heapsort
117
118 Heapsort first rearranges an array to satisfy the heap property, and
119 then the max element of the heap is repeated removed and added to the
120 end of the resulting sorted list. A heapified array has the root node
121 at index 0. The two children of the node at index n are respectively
122 located at indices 2n + 1 and 2n + 2. Each node is greater than both
123 of its children. This leads to a structure that takes O(log n) to
124 insert any element or remove the max element. Heapsort has a better
125 worst case complexity than quicksort, but a better space complexity
126 complexity than merge sort.
127
128 The standard library equivalent is calling std::make_heap(), followed
129 by std::sort_heap() on the input range.
130
132 Time Complexity (Worst): O(n log n)
133 Space Complexity: O(1) auxiliary.
134 Stable?: No
135
136 */
137
139 void heapsort(It lo, It hi, Compare comp) {
140 typename std::iterator_traits<It>::value_type t;
141 It i = lo + (hi - lo) / 2, j = hi, parent, child;
142 for (;;) {
143 if (i <= lo) {
144 if (--j == lo)
145 return;
146 t = *j;
147 *j = *lo;
148 } else {
149 t = *(--i);
150 }
151 parent = i;
152 child = lo + 2 * (i - lo) + 1;
153 while (child < j) {
154 if (child + 1 < j && comp(*child, *(child + 1)))

155 child++;
156 if (!comp(t, *child))
157 break;
158 *parent = *child;
159 parent = child;
160 child = lo + 2 * (parent - lo) + 1;
161 }
162 *(lo + (parent - lo)) = t;
163 }
164 }
165
166 template<class It> void heapsort(It lo, It hi) {
168 heapsort(lo, hi, std::less<T>());
169 }
170
171 /*
172
173 Comb Sort
174
175 Comb sort is an improved bubble sort. While bubble sort increases the
176 gap between swapped elements for every inner loop iteration, comb sort
177 uses a fixed gap for the inner loop and decreases the gap size by a
178 shrink factor for every iteration of the outer loop.
179
180 Even though the average time complexity is theoretically O(n^2), if the
181 increments (gap sizes) are relatively prime and the shrink factor is
182 sensible (1.3 is empirically determined to be the best), then it will
183 require astronomically large n to make the algorithm exceed O(n log n)
184 steps. In practice, comb sort is only 2-3 times slower than merge sort.
185
186 Time Complexity (Average): O(n^2 / 2^p) for p increments.
187 Time Complexity (Worst): O(n^2)
189 Stable?: No
190
191 */
192
194 void combsort(It lo, It hi, Compare comp) {
195 int gap = hi - lo;
196 bool swapped = true;
197 while (gap > 1 || swapped) {
198 if (gap > 1)
199 gap = (int)((float)gap / 1.3f);
200 swapped = false;
201 for (It i = lo; i + gap < hi; i++)
202 if (comp(*(i + gap), *i)) {
203 std::swap(*i, *(i + gap));
204 swapped = true;
205 }
206 }
207 }
208
209 template<class It> void combsort(It lo, It hi) {
211 combsort(lo, hi, std::less<T>());
212 }
213
214 /*
215
216 Radix Sort
217
218 Radix sort can be used to sort integer keys with a constant number of
219 bits in linear time. The keys are grouped by the individual digits of
220 a particular base which share the same significant position and value.
221
222 The implementation below only works on ranges pointing to unsigned
223 integer primitives (but can be modified to also work on signed values).
224 Note that the input range need not strictly be "unsigned" types, as
225 long as the values are all technically non-negative. A power of two is
226 chosen to be the base of the sort since bitwise operations may be used
227 to extract digits (instead of modulos and powers, which are much less
228 efficient). In practice, its been demonstrated that 2^8 is the best
229 choice for sorting 32-bit integers (roughly 5 times faster than using
230 std::sort and 2 to 4 times faster than any other chosen power of two).
231
232 This implementation was adapted from: http://qr.ae/RbdDTa
233 Explanation of base 2^8 choice: http://qr.ae/RbdDcG
234
235 Time Complexity: O(n * w) for n integers of w bits.
236 Space Complexity: O(n + w) auxiliary.
237 Stable?: Yes
238
239 */
240
241 template<class UnsignedIt>
242 void radix_sort(UnsignedIt lo, UnsignedIt hi) {
243 if (hi - lo < 2)
244 return;
245 const int radix_bits = 8;
246 const int radix_base = 1 << radix_bits; //e.g. 2^8 = 256
247 const int radix_mask = radix_base - 1; //e.g. 2^8 - 1 = 0xFF
248 int num_bits = 8 * sizeof(*lo); //8 bits per byte
249 typedef typename std::iterator_traits<UnsignedIt>::value_type T;
250 T *l = new T[hi - lo];
251 for (int pos = 0; pos < num_bits; pos += radix_bits) {
252 int count[radix_base] = {0};
253 for (UnsignedIt it = lo; it != hi; it++)
254 count[(*it >> pos) & radix_mask]++;
255 T *bucket[radix_base], *curr = l;
256 for (int i = 0; i < radix_base; curr += count[i++])
257 bucket[i] = curr;
258 for (UnsignedIt it = lo; it != hi; it++)
259 *bucket[(*it >> pos) & radix_mask]++ = *it;
260 std::copy(l, l + (hi - lo), lo);
261 }
262 delete[] l;
263 }
264
265 /*** Example Usage
266
267 Sample Output:
268
269 mergesort() with default comparisons: 1.32 1.41 1.62 1.73 2.58 2.72 3.14 4.67
270 mergesort() with compare_as_ints(): 1.41 1.73 1.32 1.62 2.72 2.58 3.14 4.67
271 ------
272 Sorting five million integers...

273 std::sort(): 0.429s
274 quicksort(): 0.498s
275 mergesort(): 1.437s
276 heapsort(): 1.179s
277 combsort(): 1.023s
278 radix_sort(): 0.078s
279
280 ***/
281
282 #include <cassert>
283 #include <cstdlib>
284 #include <ctime>
285 #include <iomanip>
286 #include <iostream>
287 #include <vector>
288 using namespace std;
289
290 template<class It> void print_range(It lo, It hi) {
291 while (lo != hi)
292 cout << *(lo++) << " ";
293 cout << endl;
294 }
295
296 template<class It> bool is_sorted(It lo, It hi) {
297 while (++lo != hi)
298 if (*(lo - 1) > *lo)
299 return false;
300 return true;
301 }
302
303 bool compare_as_ints(double i, double j) {
304 return (int)i < (int)j;
305 }
306
307 int main () {
308 { //can be used to sort arrays like std::sort()
309 int a[] = {32, 71, 12, 45, 26, 80, 53, 33};
310 quicksort(a, a + 8);
311 assert(is_sorted(a, a + 8));
312 }
313 { //STL containers work too
314 int a[] = {32, 71, 12, 45, 26, 80, 53, 33};
315 vector<int> v(a, a + 8);
316 quicksort(v.begin(), v.end());
317 assert(is_sorted(v.begin(), v.end()));
318 }
319 { //reverse iterators work as expected
320 int a[] = {32, 71, 12, 45, 26, 80, 53, 33};
322 heapsort(v.rbegin(), v.rend());
323 assert(is_sorted(v.rbegin(), v.rend()));
324 }
325 { //doubles are also fine
326 double a[] = {1.1, -5.0, 6.23, 4.123, 155.2};
327 vector<double> v(a, a + 5);
328 combsort(v.begin(), v.end());
329 assert(is_sorted(v.begin(), v.end()));
330 }
331 { //only unsigned ints work for radix_sort (but reverse works!)
332 int a[] = {32, 71, 12, 45, 26, 80, 53, 33};
334 radix_sort(v.rbegin(), v.rend());
335 assert(is_sorted(v.rbegin(), v.rend()));
336 }
337
338 //example from http://www.cplusplus.com/reference/algorithm/stable_sort
339 double a[] = {3.14, 1.41, 2.72, 4.67, 1.73, 1.32, 1.62, 2.58};
340 {
342 cout << "mergesort() with default comparisons: ";
343 mergesort(v.begin(), v.end());
344 print_range(v.begin(), v.end());
345 }
346 {
348 cout << "mergesort() with compare_as_ints(): ";
349 mergesort(v.begin(), v.end(), compare_as_ints);
350 print_range(v.begin(), v.end());
351 }
352 cout << "------" << endl;
353
354 vector<int> v, v2;
355 for (int i = 0; i < 5000000; i++)
356 v.push_back((rand() & 0x7fff) | ((rand() & 0x7fff) << 15));
357 v2 = v;
358 cout << "Sorting five million integers..." << endl;
359 cout.precision(3);
360
361 #define test(sortfunc) { \
362 clock_t start = clock(); \
363 sortfunc(v.begin(), v.end()); \
364 double t = (double)(clock() - start) / CLOCKS_PER_SEC; \
365 cout << setw(14) << left << #sortfunc "(): "; \
366 cout << fixed << t << "s" << endl; \
367 assert(is_sorted(v.begin(), v.end())); \
368 v = v2; \
369 }
370 test(std::sort);
371 test(quicksort);
372 test(mergesort);
373 test(heapsort);
374 test(combsort);
375 test(radix_sort);
376
377 return 0;
378 }
1.1.2 Array Rotation
1 /*
2
3 The following functions are equivalent to std::rotate(), taking three
4 iterators lo, mid, hi, and swapping the elements in the range [lo, hi)
5 in such a way that the element at mid becomes the first element of the
6 new range and the element at mid - 1 becomes the last element.
7
8 All three versions achieve the same result using no temporary arrays.
9 Version 1 uses a straightforward swapping algorithm listed on many C++
10 reference sites, requiring only forward iterators. Version 2 requires
11 bidirectional iterators, employing the well-known technique of three
12 simple reversals. Version 3 applies a "juggling" algorithm which first
13 divides the range into gcd(n, k) sets (n = hi - lo and k = mid - lo)
14 and then rotates the corresponding elements in each set. This version
15 requires random access iterators.
16
17 Time Complexity: O(n) on the distance between lo and hi.
19
20 */
21
22 #include <algorithm> /* std::reverse(), std::rotate(), std::swap() */
23
24 template<class It> void rotate1(It lo, It mid, It hi) {
25 It next = mid;
26 while (lo != next) {
27 std::swap(*lo++, *next++);
28 if (next == hi)
29 next = mid;
30 else if (lo == mid)
31 mid = next;
32 }
33 }
34
36 std::reverse(lo, mid);
37 std::reverse(mid, hi);
38 std::reverse(lo, hi);
39 }
40
41 int gcd(int a, int b) {
42 return b == 0 ? a : gcd(b, a % b);
43 }
44
46 int n = hi - lo, jump = mid - lo;
47 int g = gcd(jump, n), cycle = n / g;
48 for (int i = 0; i < g; i++) {
49 int curr = i, next;
50 for (int j = 0; j < cycle - 1; j++) {
51 next = curr + jump;
52 if (next >= n)
53 next -= n;
54 std::swap(*(lo + curr), *(lo + next));
55 curr = next;
56 }
57 }
58 }
59
61
62 Sample Output:
63
64 before sort: 2 4 2 0 5 10 7 3 7 1
65 after sort: 0 1 2 2 3 4 5 7 7 10
66 rotate left: 1 2 2 3 4 5 7 7 10 0
67 rotate right: 0 1 2 2 3 4 5 7 7 10
68
69 ***/
70
71 #include <algorithm>
76
77 int main() {
78 std::vector<int> v0, v1, v2, v3;
79 for (int i = 0; i < 10000; i++)
80 v0.push_back(i);
81 v1 = v2 = v3 = v0;
82 int mid = 5678;
83 std::rotate(v0.begin(), v0.begin() + mid, v0.end());
84 rotate1(v1.begin(), v1.begin() + mid, v1.end());
87 assert(v0 == v1 && v0 == v2 && v0 == v3);
88
89 //example from: http://en.cppreference.com/w/cpp/algorithm/rotate
90 int a[] = {2, 4, 2, 0, 5, 10, 7, 3, 7, 1};
92 cout << "before sort: ";
93 for (int i = 0; i < (int)v.size(); i++)
94 cout << v[i] << ;
95 cout << endl;
96
97 //insertion sort
98 for (vector<int>::iterator i = v.begin(); i != v.end(); ++i)
99 rotate1(std::upper_bound(v.begin(), i, *i), i, i + 1);
100 cout << "after sort: ";
102 cout << v[i] << ;
103 cout << endl;
104
105 //simple rotation to the left
106 rotate2(v.begin(), v.begin() + 1, v.end());
107 cout << "rotate left: ";
109 cout << v[i] << ;
110 cout << endl;
111
112 //simple rotation to the right
113 rotate3(v.rbegin(), v.rbegin() + 1, v.rend());
114 cout << "rotate right: ";
116 cout << v[i] << ;
117 cout << endl;
118 return 0;
119 }
1.1.3 Counting Inversions

1 /*
2
3 The number of inversions in an array a[] is the number of ordered pairs
4 (i, j) such that i < j and a[i] > a[j]. This is roughly how "close" an
5 array is to being sorted, but is *not* the same as the minimum number
6 of swaps required to sort the array. If the array is sorted then the
7 inversion count is 0. If the array is sorted in decreasing order, then
8 the inversion count is maximal. The following are two methods of
9 efficiently counting the number of inversions.
10
11 */
12
13 #include <algorithm> /* std::fill(), std::max() */
15
16 /*
17
18 Version 1: Merge sort
19
20 The input range [lo, hi) will become sorted after the function call,
21 and then the number of inversions will be returned. The iterators
22 value type must have the less than < operator defined appropriately.
23
24 Explanation: http://www.geeksforgeeks.org/counting-inversions
25
26 Time Complexity: O(n log n) on the distance between lo and hi.
27 Space Complexty: O(n) auxiliary.
28
29 */
30
31 template<class It> long long inversions(It lo, It hi) {
32 if (hi - lo < 2) return 0;
33 It mid = lo + (hi - lo - 1) / 2, a = lo, c = mid + 1;
34 long long res = 0;
35 res += inversions(lo, mid + 1);
36 res += inversions(mid + 1, hi);
38 T *buf = new T[hi - lo], *ptr = buf;
39 while (a <= mid && c < hi) {
40 if (*c < *a) {
41 *(ptr++) = *(c++);
42 res += (mid - a) + 1;
43 } else {
44 *(ptr++) = *(a++);
45 }
46 }
47 if (a > mid) {
48 for (It k = c; k < hi; k++)
49 *(ptr++) = *k;
50 } else {
51 for (It k = a; k <= mid; k++)
52 *(ptr++) = *k;
53 }
54 for (int i = hi - lo - 1; i >= 0; i--)
55 *(lo + i) = buf[i];
56 delete[] buf;
57 return res;
58 }
59
60 /*
61
62 Version 2: Magic
63
64 The following magic is courtesy of misof, and works for any array of
65 nonnegative integers.
66
67 Explanation: http://codeforces.com/blog/entry/17881?#comment-232099
68
69 The complexity depends on the magnitude of the maximum value in a[].
70 Coordinate compression should be applied on the values of a[] so that
71 they are strictly integers with magnitudes up to n for best results.
72 Note that after calling the function, a[] will be entirely set to 0.
73
74 Time Complexity: O(m log m), where m is maximum value in the array.
75 Space Complexity: O(m) auxiliary.
76
77 */
78
79 long long inversions(int n, int a[]) {
80 int mx = 0;
81 for (int i = 0; i < n; i++)
82 mx = std::max(mx, a[i]);
83 int *cnt = new int[mx];
85 while (mx > 0) {
86 std::fill(cnt, cnt + mx, 0);
87 for (int i = 0; i < n; i++) {
88 if (a[i] % 2 == 0)
89 res += cnt[a[i] / 2];
90 else
91 cnt[a[i] / 2]++;
92 }
93 mx = 0;
94 for (int i = 0; i < n; i++)
95 mx = std::max(mx, a[i] /= 2);
96 }
97 delete[] cnt;
98 return res;
99 }
100
101 /*** Example Usage ***/
102
104
105 int main() {
106 {
107 int a[] = {6, 9, 1, 14, 8, 12, 3, 2};
108 assert(inversions(a, a + 8) == 16);
109 }
110 {
111 int a[] = {6, 9, 1, 14, 8, 12, 3, 2};
112 assert(inversions(8, a) == 16);
113 }
114 return 0;
115 }
1.1.4 Coordinate Compression
1 /*
2
3 Given an array a[] of size n, reassign integers to each value of a[]
4 such that the magnitude of each new value is no more than n, while the
5 relative order of each value as they were in the original array is
6 preserved. That is, if a[] is the original array and b[] is the result
7 array, then for every pair (i, j), the result of comparing a[i] < a[j]
8 will be exactly the same as the result of b[i] < b[j]. Furthermore,
9 no value of b[] will exceed the *number* of distinct values in a[].
10
11 In the following implementations, values in the range [lo, hi) will be
12 converted to integers in the range [0, d), where d is the number of
13 distinct values in the original range. lo and hi must be random access
14 iterators pointing to a numerical type that int can be assigned to.
15
16 Time Complexity: O(n log n) on the distance between lo and hi.
18
19 */
20
21 #include <algorithm> /* std::lower_bound(), std::sort(), std::unique() */
23 #include <map>
24
25 //version 1 - using std::sort(), std::unique() and std::lower_bound()
26 template<class It> void compress1(It lo, It hi) {
28 T *a = new T[hi - lo];
29 int n = 0;
30 for (It it = lo; it != hi; ++it)
31 a[n++] = *it;
32 std::sort(a, a + n);
33 int n2 = std::unique(a, a + n) - a;
35 *it = (int)(std::lower_bound(a, a + n2, *it) - a);
36 delete[] a;
37 }
38
39 //version 2 - using std::map
40 template<class It> void compress2(It lo, It hi) {
42 std::map<T, int> m;
44 m[*it] = 0;
45 typename std::map<T, int>::iterator x = m.begin();
46 for (int i = 0; x != m.end(); x++)
47 x->second = i++;
49 *it = m[*it];
50 }
51
53
54 Sample Output:
55
56 0 4 4 1 3 2 5 5
57 0 4 4 1 3 2 5 5
58 1 0 2 0 3 1
59
60 ***/
61
64
66 while (lo != hi)
67 cout << *(lo++) << " ";
68 cout << endl;
69 }
70
71 int main() {
72 {
73 int a[] = {1, 30, 30, 7, 9, 8, 99, 99};
74 compress1(a, a + 8);
75 print_range(a, a + 8);
76 }
77 {
78 int a[] = {1, 30, 30, 7, 9, 8, 99, 99};
81 }
82 { //works on doubles too
83 double a[] = {0.5, -1.0, 3, -1.0, 20, 0.5};
86 }
87 return 0;
88 }
1.1.5 Selection (Quickselect)
1 /*
2
3 Quickselect (also known as Hoares algorithm) is a selection algorithm
4 which rearranges the elements in a sequence such that the element at
5 the nth position is the element that would be there if the sequence
6 were sorted. The other elements in the sequence are partioned around
7 the nth element. That is, they are left in no particular order, except
8 that no element before the nth element is is greater than it, and no
9 element after it is less.
10
11 The following implementation is equivalent to std::nth_element(),
12 taking in two random access iterators as the range and performing the
13 described operation in expected linear time.
14
15 Time Complexity (Average): O(n) on the distance between lo and hi.
16 Time Complexity (Worst): O(n^2), although this *almost never* occurs.
18
19 */
20
21 #include <algorithm> /* std::swap() */
22 #include <cstdlib> /* rand() */

24
25 int rand32() {
26 return (rand() & 0x7fff) | ((rand() & 0x7fff) << 15);
27 }
28
29 template<class It> It rand_partition(It lo, It hi) {
30 std::swap(*(lo + rand32() % (hi - lo)), *(hi - 1));
31 typename std::iterator_traits<It>::value_type mid = *(hi - 1);
32 It i = lo - 1;
33 for (It j = lo; j != hi; ++j)
34 if (*j <= mid)
35 std::swap(*(++i), *j);
36 return i;
37 }
38
39 template<class It> void nth_element2(It lo, It n, It hi) {
40 for (;;) {
41 It k = rand_partition(lo, hi);
42 if (n < k)
43 hi = k;
44 else if (n > k)
45 lo = k + 1;
46 else
47 return;
48 }
49 }
50
52
53 Sample Output:
54 2 3 1 5 4 6 8 7 9
55
56 ***/
57
60
62 while (lo != hi)
63 cout << *(lo++) << " ";
64 cout << endl;
65 }
66
67 int main () {
68 int a[] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
69 random_shuffle(a, a + 9);
70 nth_element2(a, a + 5, a + 9);
72 return 0;
73 }
1.2 Array Queries
1.2.1 Longest Increasing Subsequence

1.2. Array Queries 15
1 /*
2
3 Given an array a[] of size n, determine a longest subsequence of a[]
4 such that all of its elements are in ascending order. This subsequence
5 is not necessarily contiguous or unique, so only one such answer will
6 be found. The problem is efficiently solved using dynamic programming
7 and binary searching, since it has the following optimal substructure
8 with respect to the i-th position in the array:
9
10 LIS[i] = 1 + max(LIS[j] for all j < i and a[j] < a[i])
11 Otherwise if such a j does not exist, then LIS[i] = 1.
12
13 Explanation: https://en.wikipedia.org/wiki/Longest_increasing_subsequence
14
15 Time Complexity: O(n log n) on the size of the array.
17
18 */
19
21
22 std::vector<int> tail, prev;
23
24 template<class T> int lower_bound(int len, T a[], int key) {
25 int lo = -1, hi = len;
26 while (hi - lo > 1) {
27 int mid = (lo + hi) / 2;
28 if (a[tail[mid]] < key)
29 lo = mid;
30 else
31 hi = mid;
32 }
33 return hi;
34 }
35
36 template<class T> std::vector<T> lis(int n, T a[]) {
37 tail.resize(n);
38 prev.resize(n);
39 int len = 0;
40 for (int i = 0; i < n; i++) {
41 int pos = lower_bound(len, a, a[i]);
42 if (len < pos + 1)
43 len = pos + 1;
44 prev[i] = pos > 0 ? tail[pos - 1] : -1;
45 tail[pos] = i;
46 }
47 std::vector<T> res(len);
48 for (int i = tail[len - 1]; i != -1; i = prev[i])
49 res[--len] = a[i];
50 return res;
51 }
52
54
55 Sample Output:
56 -5 1 9 10 11 13
57
58 ***/
59
62
64 while (lo != hi)
65 cout << *(lo++) << " ";
66 cout << endl;
67 }
68
69 int main () {
70 int a[] = {-2, -5, 1, 9, 10, 8, 11, 10, 13, 11};
71 vector<int> res = lis(10, a);
72 print_range(res.begin(), res.end());
73 return 0;
74 }
1.2.2 Maximal Subarray Sum (Kadanes)
1 /*
2
3 Given a sequence of numbers (with at least one positive number), find
4 the maximum possible sum of any contiguous subarray. Kadanes algorithm
5 scans through the array, computing at each index the maximum (positive
6 sum) subarray ending at that position. This subarray is either empty
7 (in which case its sum is zero) or consists of one more element than
8 the maximum subarray ending at the previous position.
9
10 */
11
12 #include <algorithm> /* std::fill() */
14 #include <limits> /* std::numeric_limits */
16
17 /*
18
19 The following implementation takes two random access iterators as the
20 range of values to be considered. Optionally, two pointers to integers
21 may be passed to have the positions of the begin and end indices of
22 the maximal sum subarray stored. begin_idx will be inclusive while
23 end_idx will be exclusive (i.e. (lo + begin_idx) will reference the
24 first element of the max sum subarray and (lo + end_idx) will reference
25 the index just past the last element of the subarray. Note that the
26 following version does not allow empty subarrays to be returned, so the
27 the max element will simply be returned if the array is all negative.
28
29 Time Complexity: O(n) on the distance between lo and hi.
30 Space Complexty: O(1) auxiliary.
31
32 */
33
34 template<class It> typename std::iterator_traits<It>::value_type
35 max_subarray_sum(It lo, It hi, int *begin_idx = 0, int *end_idx = 0) {
37 int curr_begin = 0, begin = 0, end = -1;
38 T sum = 0, max_sum = std::numeric_limits<T>::min();
39 for (It it = lo; it != hi; ++it) {
40 sum += *it;
41 if (sum < 0) {
42 sum = 0;
43 curr_begin = (it - lo) + 1;
44 } else if (max_sum < sum) {
45 max_sum = sum;
46 begin = curr_begin;
47 end = (it - lo) + 1;
48 }
49 }
50 if (end == -1) { //all negative, just return the max value
52 if (max_sum < *it) {
53 max_sum = *it;
54 begin = it - lo;
55 end = begin + 1;
56 }
57 }
58 }
59 if (begin_idx != 0 && end_idx != 0) {
60 *begin_idx = begin;
61 *end_idx = end;
62 }
63 return max_sum;
64 }
65
66 /*
67
68 Maximal Submatrix Sum
69
70 In the 2-dimensional version of the problem, the largest sum of any
71 rectangular submatrix must be found for a matrix n rows by m columns.
72 Kadanes algorithm is applied to each interval [lcol, hcol] of columns
73 in the matrix, for an overall cubic time solution. The input must be a
74 two dimensional vector, where the outer vector must contain n vectors
75 each with m elements. Optionally, four int pointers begin_row, end_row,
76 begin_col, and end_col may be passed. If so, then their dereferenced
77 values will be set to the boundary indices of the max sum submatrix.
78 Note that begin_row and begin_col are inclusive indices, while end_row
79 and end_col are exclusive (referring to the index just past the end).
80
81 Time Complexity: O(m^2 * n) for a matrix with m columns and n rows.
83
84 */
85
86 template<class T>
87 T max_submatrix_sum(const std::vector< std::vector<T> > & mat,
88 int *begin_row = 0, int *end_row = 0,
89 int *begin_col = 0, int *end_col = 0) {
90 int n = mat.size(), m = mat[0].size();
91 std::vector<T> sums(n);
92 T sum, max_sum = std::numeric_limits<T>::min();
93 for (int lcol = 0; lcol < m; lcol++) {
94 std::fill(sums.begin(), sums.end(), 0);
95 for (int hcol = lcol; hcol < m; hcol++) {
96 for (int i = 0; i < n; i++)
97 sums[i] += mat[i][hcol];
98 int begin, end;
99 sum = max_subarray_sum(sums.begin(), sums.end(), &begin, &end);

100 if (sum > max_sum) {
101 max_sum = sum;
102 if (begin_row != 0) {
103 *begin_row = begin;
104 *end_row = end;
105 *begin_col = lcol;
106 *end_col = hcol + 1;
107 }
108 }
109 }
110 }
111 return max_sum;
112 }
113
115
116 Sample Output:
117 1D example - the max sum subarray is
118 4 -1 2 1
119 2D example - the max sum submatrix is
120 9 2
121 -4 1
122 -1 8
123
124 ***/
125
129
130 int main() {
131 {
132 int a[] = {-2, -1, -3, 4, -1, 2, 1, -5, 4};
133 int begin, end;
134 assert(max_subarray_sum(a, a + 3) == -1);
135 assert(max_subarray_sum(a, a + 9, &begin, &end) == 6);
136 cout << "1D example - the max sum subarray is" << endl;
137 for (int i = begin; i < end; i++)
138 cout << a[i] << " ";
139 cout << endl;
140 }
141 {
142 const int n = 4, m = 5;
143 int a[n][m] = {{ 0, -2, -7, 0, 5},
144 { 9, 2, -6, 2, -4},
145 {-4, 1, -4, 1, 0},
146 {-1, 8, 0, -2, 3}};
147 vector< vector<int> > mat(n);
148 for (int i = 0; i < n; i++)
149 mat[i] = vector<int>(a[i], a[i] + m);
150 int lrow, hrow, lcol, hcol;
151 assert(max_submatrix_sum(mat, &lrow, &hrow, &lcol, &hcol) == 15);
152 cout << "2D example - The max sum submatrix is" << endl;
153 for (int i = lrow; i < hrow; i++) {
154 for (int j = lcol; j < hcol; j++)
155 cout << mat[i][j] << " ";
156 cout << endl;
157 }
158 }
159 return 0;
160 }
1.2.3 Majority Element (Boyer-Moore)
1 /*
2
3 Given a sequence of n elements, the majority vote problem asks to find
4 an element that occurs more frequently than all others, or determine
5 that no such element exists. Formally, a value must occur strictly
6 greater than floor(n/2) times to be considered the majority element.
7 Boyer-Moore majority vote algorithm scans through the sequence and
8 keeps track of a running counter for the most likely candidate so far.
9 Whenever a value is equal to the current candidate, the counter is
10 incremented, otherwise the counter is decremented. When the counter is
11 zero, the candidate is eliminated and a new candidate is considered.
12
13 The following implementation takes two random access iterators as the
14 sequence [lo, hi) of elements and returns an iterator pointing to one
15 instance of the majority element if it exists, or the iterator hi if
16 there is no majority.
17
18 Time Complexity: O(n) on the size of the array.
19 Space Complexty: O(1) auxiliary.
20
21 */
22
23 template<class It> It majority(It lo, It hi) {
24 int cnt = 0;
25 It candidate = lo;
27 if (cnt == 0) {
28 candidate = it;
29 cnt = 1;
30 } else if (*it == *candidate) {
31 cnt++;
32 } else {
33 cnt--;
34 }
35 }
36 cnt = 0;
38 if (*it == *candidate)
39 cnt++;
40 }
41 if (cnt <= (hi - lo) / 2)
42 return hi;
43 return candidate;
44 }
45
47
49
50 int main() {
51 int a[] = {3, 2, 3, 1, 3};
52 assert(*majority(a, a + 5) == 3);
53 int b[] = {2, 3, 3, 3, 2, 1};
54 assert(majority(b, b + 6) == b + 6);
55 return 0;
56 }
1.2.4 Subset Sum (Meet-in-the-Middle)
1 /*
2
3 Given a sequence of n (not necessarily unique) integers and a number v,
4 determine the minimum possible sum of any subset of the given sequence
5 that is not less than v. This is a generalization of a more well-known
6 version of the subset sum problem which asks whether a subset summing
7 to 0 exists (equivalent here to seeing if v = 0 yields an answer of 0).
8 Both problems are NP-complete. A meet-in-the-middle algorithm divides
9 the array in two equal parts. All possible sums of the lower and higher
10 parts are precomputed and sorted in a table. Finally, the table is
11 searched to find the lower bound.
12
13 The following implementation accepts two random access iterators as the
14 sequence [lo, hi) of integers, and the number v. Note that since the
15 sums can get large, 64-bit integers are necessary to avoid overflow.
16
17 Time Complexity: O(n * 2^(n/2)) on the distance between lo and hi.
19
20 */
21
22 #include <algorithm> /* std::max(), std::sort() */
24
25 template<class It>
26 long long sum_lower_bound(It lo, It hi, long long v) {
27 int n = hi - lo;
28 int llen = 1 << (n / 2);
29 int hlen = 1 << (n - n / 2);
30 long long *lsum = new long long[llen];
31 long long *hsum = new long long[hlen];
32 std::fill(lsum, lsum + llen, 0);
33 std::fill(hsum, hsum + hlen, 0);
34 for (int mask = 0; mask < llen; mask++) {
35 for (int i = 0; i < n / 2; i++) {
36 if ((mask >> i) & 1)
37 lsum[mask] += *(lo + i);
38 }
39 }
40 for (int mask = 0; mask < hlen; mask++) {
41 for (int i = 0; i < n - n / 2; i++) {
42 if ((mask >> i) & 1)
43 hsum[mask] += *(lo + i + n / 2);
44 }
45 }
46 std::sort(lsum, lsum + llen);
47 std::sort(hsum, hsum + llen);
48 int l = 0, r = hlen - 1;
49 long long curr = std::numeric_limits<long long>::min();
50 while (l < llen && r >= 0) {

51 if (lsum[l] + hsum[r] <= v) {
52 curr = std::max(curr, lsum[l] + hsum[r]);
53 l++;
54 } else {
55 r--;
56 }
57 }
58 delete[] lsum;
59 delete[] hsum;
60 return curr;
61 }
62
64
66
67 int main() {
68 int a[] = {9, 1, 5, 0, 1, 11, 5};
69 assert(sum_lower_bound(a, a + 7, 8) == 7);
70 int b[] = {-7, -3, -2, 5, 8};
71 assert(sum_lower_bound(b, b + 5, 0) == 0);
72 return 0;
73 }
1.2.5 Maximal Zero Submatrix
1 /*
2
3 Given an n by m rectangular matrix of 0s and 1s, determine the area
4 of the largest rectangular submatrix which contains only 0s. This can
5 be reduced the problem of finding the maximum rectangular area under a
6 histogram, which can be efficiently solved using a stack. The following
7 implementation accepts a 2-dimensional vector of bools and returns the
8 area of the maximum zero submatrix.
9
10 Explanation: http://stackoverflow.com/a/13657337
11
12 Time Complexity: O(n * m) for a matrix n rows by m columns.
13 Space Complexity: O(m) auxiliary.
14
15 */
16
17 #include <algorithm> /* std::max() */
19
20 int max_zero_submatrix(const std::vector< std::vector<bool> > & mat) {
21 int n = mat.size(), m = mat[0].size(), res = 0;
22 std::vector<int> d(m, -1), d1(m), d2(m), stack;
23 for (int r = 0; r < n; r++) {
24 for (int c = 0; c < m; c++) {
25 if (mat[r][c])
26 d[c] = r;
27 }
28 stack.clear();
29 for (int c = 0; c < m; c++) {
30 while (!stack.empty() && d[stack.back()] <= d[c])
31 stack.pop_back();
32 d1[c] = stack.empty() ? -1 : stack.back();
33 stack.push_back(c);
34 }
35 stack.clear();
36 for (int c = m - 1; c >= 0; c--) {
37 while (!stack.empty() && d[stack.back()] <= d[c])
39 d2[c] = stack.empty() ? m : stack.back();
40 stack.push_back(c);
41 }
42 for (int j = 0; j < m; j++)
43 res = std::max(res, (r - d[j]) * (d2[j] - d1[j] - 1));
44 }
45 return res;
46 }
47
49
52
53 int main() {
54 const int n = 5, m = 6;
55 bool a[n][m] = {{1, 0, 1, 1, 0, 0},
56 {1, 0, 0, 1, 0, 0},
57 {0, 0, 0, 0, 0, 1},
58 {1, 0, 0, 1, 0, 0},
59 {1, 0, 1, 0, 0, 1}};
60 std::vector< std::vector<bool> > mat(n);
61 for (int i = 0; i < n; i++)
62 mat[i] = vector<bool>(a[i], a[i] + m);
63 assert(max_zero_submatrix(mat) == 6);
64 return 0;
65 }
1.3 Searching
1.3.1 Discrete Binary Search
1 /*
2
3 Not only can binary search be used to find the position of a given
4 element in a sorted array, it can also be used to find the input value
5 corresponding to any output value of a monotonic (either strictly
6 non-increasing or strictly non-decreasing) function in O(log n) running
7 time with respect to the domain. This is a special case of finding
8 the exact point at which any given monotonic Boolean function changes
9 from true to false (or vice versa). Unlike searching through an array,
10 discrete binary search is not restricted by available memory, which is
11 especially important while handling infinitely large search spaces such
12 as the real numbers.
13
14 binary_search_first_true() takes two integers lo and hi as boundaries
15 for the search space [lo, hi) (i.e. including lo, but excluding hi),
16 and returns the least integer k (lo <= k < hi) for which the Boolean
1.3. Searching 23
17 predicate pred(k) tests true. This function is correct if and only if

18 there exists a constant k where the return value of pred(x) is false
19 for all x < k and true for all x >= k.
20
21 binary_search_last_true() takes two integers lo and hi as boundaries
22 for the search space [lo, hi) (i.e. including lo, but excluding hi),
23 and returns the greatest integer k (lo <= k < hi) for which the Boolean
24 predicate pred(k) tests true. This function is correct if and only if
25 there exists a constant k where the return value of pred(x) is true
26 for all x <= k and false for all x > k.
27
28 Time Complexity: At most O(log n) calls to pred(), where n is the
29 distance between lo and hi.
30
32
33 */
34
35 //000[1]11
36 template<class Int, class IntPredicate>
37 Int binary_search_first_true(Int lo, Int hi, IntPredicate pred) {
38 Int mid, _hi = hi;
39 while (lo < hi) {
40 mid = lo + (hi - lo) / 2;
41 if (pred(mid))
42 hi = mid;
43 else
44 lo = mid + 1;
45 }
46 if (!pred(lo)) return _hi; //all false
47 return lo;
48 }
49
50 //11[1]000
51 template<class Int, class IntPredicate>
52 Int binary_search_last_true(Int lo, Int hi, IntPredicate pred) {
53 Int mid, _hi = hi;
55 mid = lo + (hi - lo + 1) / 2;
56 if (pred(mid))
57 lo = mid;
58 else
59 hi = mid - 1;
60 }
61 if (!pred(lo)) return _hi; //all true
62 return lo;
63 }
64
65 /*
66
67 fbinary_search() is the equivalent of binary_search_first_true() on
68 floating point predicates. Since any given range of reals numbers is
69 dense, it is clear that the exact target cannot be found. Instead, the
70 function will return a value that is very close to the border between
71 false and true. The precision of the answer depends on the number of
72 repetitions the function uses. Since each repetition bisects the search
73 space, for r repetitions, the absolute error of the answer will be
74 1/(2^r) times the distance between lo and hi. Although its possible to
75 control the error by looping while hi - lo is greater than an arbitrary
76 epsilon, it is much simpler to let the loop run for a sizable number of
77 iterations until floating point arithmetic breaks down. 100 iterations
78 is typically sufficient, reducing the search space to 2^-100 ~ 10^-30
79 times its original size.
80
81 Note that the function can be modified to find the "last true" point
82 in the range by interchanging lo and hi in the if-else statement.
83
84 Time Complexity: At most O(log n) calls to pred(), where n is the
85 distance between lo and hi divided by the desired absolute error.
86
88
89 */
90
91 //000[1]11
92 template<class DoublePredicate>
93 double fbinary_search(double lo, double hi, DoublePredicate pred) {
94 double mid;
95 for (int reps = 0; reps < 100; reps++) {
96 mid = (lo + hi) / 2.0;
97 if (pred(mid))
98 hi = mid;
99 else
100 lo = mid;
101 }
102 return lo;
103 }
104
106
108 #include <cmath>
109
110 //Simple predicate examples:
111 bool pred1(int x) { return x >= 3; }
112 bool pred2(int x) { return false; }
113 bool pred3(int x) { return x <= 5; }
114 bool pred4(int x) { return true; }
115 bool pred5(double x) { return x >= 1.2345; }
116
117 int main() {
118 assert(binary_search_first_true(0, 7, pred1) == 3);
119 assert(binary_search_first_true(0, 7, pred2) == 7);
120 assert(binary_search_last_true(0, 7, pred3) == 5);
121 assert(binary_search_last_true(0, 7, pred4) == 7);
122 assert(fabs(fbinary_search(-10.0, 10.0, pred5) - 1.2345) < 1e-15);
123 return 0;
124 }
1.3.2 Ternary Search
1 /*
2
3 Given a unimodal function f(x), find its maximum or minimum point to a
4 an arbitrarily specified absolute error.
5
1.3. Searching 25
6 ternary_search_min() takes the domain [lo, hi] of a continuous function

7 f(x) and returns a number x such that f is strictly decreasing on the
8 interval [lo, x] and strictly increasing on the interval [x, hi]. For
9 ternary search to work, this x must exist and be unique.
10
11 ternary_search_max() takes the domain [lo, hi] of a continuous function
12 f(x) and returns a number x such that f is strictly increasing on the
13 interval [lo, x] and strictly decreasing on the interval [x, hi]. For
14 ternary search to work, this x must exist and be unique.
15
16 Time Complexity: At most O(log n) calls to f, where n is the distance
17 between lo and hi divided by the desired absolute error (epsilon).
18
20
21 */
22
23 template<class UnimodalFunction>
24 double ternary_search_min(double lo, double hi, UnimodalFunction f) {
25 static const double EPS = 1e-9;
26 double lthird, hthird;
27 while (hi - lo > EPS) {
28 lthird = lo + (hi - lo) / 3;
29 hthird = hi - (hi - lo) / 3;
30 if (f(lthird) < f(hthird))
31 hi = hthird;
32 else
33 lo = lthird;
34 }
35 return lo;
36 }
37
38 template<class UnimodalFunction>
39 double ternary_search_max(double lo, double hi, UnimodalFunction f) {
40 static const double EPS = 1e-9;
41 double lthird, hthird;
42 while (hi - lo > EPS) {
43 lthird = lo + (hi - lo) / 3;
44 hthird = hi - (hi - lo) / 3;
45 if (f(lthird) < f(hthird))
46 lo = lthird;
47 else
48 hi = hthird;
49 }
50 return hi;
51 }
52
54
55 #include <cmath>
57
58 bool eq(double a, double b) {
59 return fabs(a - b) < 1e-9;
60 }
61
62 //parabola opening up with vertex at (-2, -24)
63 double f1(double x) {
64 return 3*x*x + 12*x - 12;
65 }
66
67 //parabola opening down with vertex at (2/19, 8366/95)
69 return -5.7*x*x + 1.2*x + 88;
70 }
71
72 //absolute value function shifted to the right by 30 units
74 return fabs(x - 30);
75 }
76
77 int main() {
78 assert(eq(ternary_search_min(-1000, 1000, f1), -2));
79 assert(eq(ternary_search_max(-1000, 1000, f2), 2.0 / 19));
80 assert(eq(ternary_search_min(-1000, 1000, f3), 30));
81 return 0;
82 }
1.3.3 Hill Climbing
1 /*
2
3 Given a continuous function f on two real numbers, hill climbing is a
4 technique that can be used to find the local maximum or minimum point
5 based on some (possibly random) initial guess. Then, the algorithm
6 considers taking a single step in each of a fixed number of directions.
7 The direction with the best result is selected and steps are further
8 taken there until the answer no longer improves. When this happens, the
9 step size is reduced and the process repeats until a desired absolute
10 error is reached. The result is not necessarily the global extrema, and
11 the algorithms success will heavily depend on the initial guess.
12
13 The following function find_min() takes the function f, any starting
14 guess (x0, y0), and optionally two pointers to double used for storing
15 the answer coordinates. find_min() returns a local minimum point near
16 the initial guess, and if the two pointers are given, then coordinates
17 will be stored into the variables pointed to by x_ans and y_ans.
18
19 Time Complexity: At most O(d log n) calls to f, where d is the number
20 of directions considered at each position and n is the search space,
21 roughly proportional to the largest possible step size divided by the
22 smallest possible step size.
23
24 */
25
26 #include <cmath>
29
30 template<class BinaryFunction>
31 double find_min(BinaryFunction f, double x0, double y0,
32 double *x_ans = 0, double *y_ans = 0) {
33 static const double PI = acos(-1.0);
34 static const double STEP_MAX = 1000000;
35 static const double STEP_MIN = 1e-9;
36 static const int DIRECTIONS = 6;
1.3. Searching 27
37 double x = x0, y = y0, res = f(x0, y0);

38 for (double step = STEP_MAX; step > STEP_MIN; ) {
39 double best = res, best_x = x, best_y = y;
40 bool found = false;
41 for (int i = 0; i < DIRECTIONS; i++) {
42 double a = 2.0 * PI * i / DIRECTIONS;
43 double x2 = x + step * cos(a);
44 double y2 = y + step * sin(a);
45 double val = f(x2, y2);
46 if (best > val) {
47 best_x = x2;
48 best_y = y2;
49 best = val;
50 found = true;
51 }
52 }
53 if (!found) {
54 step /= 2.0;
55 } else {
56 x = best_x;
57 y = best_y;
58 res = best;
59 }
60 }
61 if (x_ans != 0 && y_ans != 0) {
62 *x_ans = x;
63 *y_ans = y;
64 }
65 return res;
66 }
67
69
71 #include <cmath>
72
73 bool eq(double a, double b) {
74 return fabs(a - b) < 1e-8;
75 }
76
77 //minimized at f(2, 3) = 0
78 double f(double x, double y) {
79 return (x - 2)*(x - 2) + (y - 3)*(y - 3);
80 }
81
82 int main() {
83 double x, y;
84 assert(eq(find_min(f, 0, 0, &x, &y), 0));
85 assert(eq(x, 2) && eq(y, 3));
86 return 0;
87 }
1.3.4 Convex Hull Trick (Semi-Dynamic)
1 /*
2
3 Given a set of pairs (m, b) describing lines of the form y = mx + b,
4 process a set of x-coordinate queries each asking to find the minimum

5 y-value of any of the given lines when evaluated at the specified x.
6 The convex hull optimization technique first ignores all lines which
7 never take on the maximum at any x value, then sorts the rest in order
8 of descending slope. The intersection points of adjacent lines in this
9 sorted list form the upper envelope of a convex hull, and line segments
10 connecting these points always take on the minimum y-value. The result
11 can be split up into x-intervals each mapped to the line which takes on
12 the minimum in that interval. The intervals can be binary searched to
13 solve each query in O(log n) time on the number of lines.
14
15 Explanation: http://wcipeg.com/wiki/Convex_hull_trick
16
17 The following implementation is a concise, semi-dynamic version which
18 supports an an interlaced series of add line and query operations.
19 However, two key preconditions are that each call to add_line(m, b)
20 must have m as the minimum slope of all lines added so far, and each
21 call to get_min(x) must have x as the maximum x of all queries so far.
22 As a result, pre-sorting the lines and queries may be necessary (in
23 which case the running time will be that of the sorting algorithm).
24
25 Time Complexity: O(n) on the number of calls to add_line(). Since the
26 number of steps taken by add_line() and get_min() are both bounded by
27 the number of lines added so far, their running times are respectively
28 O(1) amortized.
29
30 Space Complexity: O(n) auxiliary on the number of calls to add_line().
31
32 */
33
35
36 std::vector<long long> M, B;
37 int ptr = 0;
38
39 void add_line(long long m, long long b) {
40 int len = M.size();
41 while (len > 1 && (B[len - 2] - B[len - 1]) * (m - M[len - 1]) >=
42 (B[len - 1] - b) * (M[len - 1] - M[len - 2])) {
43 len--;
44 }
45 M.resize(len);
46 B.resize(len);
47 M.push_back(m);
48 B.push_back(b);
49 }
50
51 long long get_min(long long x) {
52 if (ptr >= (int)M.size())
53 ptr = (int)M.size() - 1;
54 while (ptr + 1 < (int)M.size() && M[ptr + 1] * x + B[ptr + 1] <=
55 M[ptr] * x + B[ptr]) {
56 ptr++;
57 }
58 return M[ptr] * x + B[ptr];
59 }
60
62
1.3. Searching 29
64
65 int main() {
66 add_line(3, 0);
67 add_line(2, 1);
68 add_line(1, 2);
69 add_line(0, 6);
70 assert(get_min(0) == 0);
74 return 0;
75 }
1.3.5 Convex Hull Trick (Fully Dynamic)
1 /*
2
3 Given a set of pairs (m, b) describing lines of the form y = mx + b,
4 process a set of x-coordinate queries each asking to find the minimum
5 y-value of any of the given lines when evaluated at the specified x.
6 The convex hull optimization technique first ignores all lines which
7 never take on the maximum at any x value, then sorts the rest in order
8 of descending slope. The intersection points of adjacent lines in this
9 sorted list form the upper envelope of a convex hull, and line segments
10 connecting these points always take on the minimum y-value. The result
11 can be split up into x-intervals each mapped to the line which takes on
12 the minimum in that interval. The intervals can be binary search to
13 solve each query in O(log n) time on the number of lines.
14
15 Explanation: http://wcipeg.com/wiki/Convex_hull_trick
16
17 The following implementation is a fully dynamic version, using a
18 self-balancing binary search tree (std::set) to support calling line
19 addition and query operations in any desired order. In addition, one
20 may instead optimize for maximum y by setting QUERY_MAX to true.
21
22 Time Complexity: O(n log n) for n calls to add_line(), where each call
23 is O(log n) amortized on the number of lines added so far. Each call to
24 get_best() runs in O(log n) on the number of lines added so far.
25
26 Space Complexity: O(n) auxiliary on the number of calls to add_line().
27
28 */
29
30 #include <set>
31
32 const bool QUERY_MAX = false;
33 const double INF = 1e30;
34
35 struct line {
36 long long m, b, val;
37 double xlo;
38 bool is_query;
39
40 line(long long m, long long b) {
41 this->m = m;
42 this->b = b;
43 val = 0;
44 xlo = -INF;
45 is_query = false;
46 }
47
48 long long evaluate(long long x) const {
49 return m * x + b;
50 }
51
52 bool parallel(const line & l) const {
53 return m == l.m;
54 }
55
56 double intersect(const line & l) const {
57 if (parallel(l))
58 return INF;
59 return (double)(l.b - b)/(m - l.m);
60 }
61
62 bool operator < (const line & l) const {
63 if (l.is_query)
64 return QUERY_MAX ? xlo < l.val : l.val < xlo;
65 return m < l.m;
66 }
67 };
68
69 std::set<line> hull;
70
71 typedef std::set<line>::iterator hulliter;
72
73 bool has_prev(hulliter it) {
74 return it != hull.begin();
75 }
76
77 bool has_next(hulliter it) {
78 return it != hull.end() && ++it != hull.end();
79 }
80
81 bool irrelevant(hulliter it) {
82 if (!has_prev(it) || !has_next(it))
83 return false;
84 hulliter prev = it; --prev;
85 hulliter next = it; ++next;
86 return QUERY_MAX ?
87 prev->intersect(*next) <= prev->intersect(*it) :
88 next->intersect(*prev) <= next->intersect(*it);
89 }
90
91 hulliter update_left_border(hulliter it) {
92 if ((QUERY_MAX && !has_prev(it)) || (!QUERY_MAX && !has_next(it)))
93 return it;
94 hulliter it2 = it;
95 double val = it->intersect(QUERY_MAX ? *--it2 : *++it2);
96 line buf(*it);
97 buf.xlo = val;
98 hull.erase(it++);
99 return hull.insert(it, buf);
100 }
1.4. Cycle Detection 31
101
102 void add_line(long long m, long long b) {
103 line l(m, b);
104 hulliter it = hull.lower_bound(l);
105 if (it != hull.end() && it->parallel(l)) {
106 if ((QUERY_MAX && it->b < b) || (!QUERY_MAX && b < it->b))
108 else
109 return;
110 }
111 it = hull.insert(it, l);
112 if (irrelevant(it)) {
113 hull.erase(it);
114 return;
115 }
116 while (has_prev(it) && irrelevant(--it))
118 while (has_next(it) && irrelevant(++it))
119 hull.erase(it--);
120 it = update_left_border(it);
121 if (has_prev(it))
122 update_left_border(--it);
123 if (has_next(++it))
124 update_left_border(++it);
125 }
126
127 long long get_best(long long x) {
128 line q(0, 0);
129 q.val = x;
130 q.is_query = true;
131 hulliter it = hull.lower_bound(q);
132 if (QUERY_MAX)
133 --it;
134 return it->evaluate(x);
135 }
136
138
140
141 int main() {
142 add_line(3, 0);
143 add_line(0, 6);
144 add_line(1, 2);
145 add_line(2, 1);
146 assert(get_best(0) == 0);
150 return 0;
151 }
1.4 Cycle Detection
1.4.1 Floyds Algorithm

1 /*
2
3 For a function f which maps a finite set S to itself and any initial
4 value x[0] in S, the same value must occur twice in the sequence below:
5
6 x[0], x[1] = f(x[0]), x[2] = f(x[1]), ..., x[i] = f(x[i - 1])
7
8 That is, there must exist numbers i, j (i < j) such that x[i] = x[j].
9 Once this happens, the sequence will continue periodically by repeating
10 the same sequence of values from x[i] to x[j 1]. Cycle detection asks
11 to find i and j, given the function f and initial value x[0]. This is
12 also analogous to the problem of detecting a cycle in a linked list,
13 which will make it degenerate.
14
15 Floyds cycle-finding algorithm, a.k.a. the "tortoise and the hare
16 algorithm", is a space-efficient algorithm that moves two pointers
17 through the sequence at different speeds. Each step in the algorithm
18 moves the "tortoise" one step forward and the "hare" two steps forward
19 in the sequence, comparing the sequence values at each step. The first
20 value which is simultaneously pointed to by both pointers is the start
21 of the sequence.
22
23 Time Complexity: O(mu + lambda), where mu is the smallest index of the
24 sequence on which a cycle starts, and lambda is the cycles length.
25
27
28 */
29
30 #include <utility> /* std::pair */
31
32 template<class IntFunction>
33 std::pair<int, int> find_cycle(IntFunction f, int x0) {
34 int tortoise = f(x0), hare = f(f(x0));
35 while (tortoise != hare) {
36 tortoise = f(tortoise);
37 hare = f(f(hare));
38 }
39 int start = 0;
40 tortoise = x0;
43 hare = f(hare);
44 start++;
45 }
46 int length = 1;
47 hare = f(tortoise);
49 hare = f(hare);
50 length++;
51 }
52 return std::make_pair(start, length);
53 }
54
56
58 #include <set>
1.4. Cycle Detection 33

61
62 const int x0 = 0;
63
64 int f(int x) {
65 return (123 * x * x + 4567890) % 1337;
66 }
67
68 void verify(int x0, int start, int length) {
69 set<int> s;
70 int x = x0;
71 for (int i = 0; i < start; i++) {
72 assert(!s.count(x));
73 s.insert(x);
74 x = f(x);
75 }
76 int startx = x;
77 s.clear();
78 for (int i = 0; i < length; i++) {
80 s.insert(x);
81 x = f(x);
82 }
83 assert(startx == x);
84 }
85
86 int main () {
87 pair<int, int> res = find_cycle(f, x0);
88 assert(res == make_pair(4, 2));
89 verify(x0, res.first, res.second);
90 return 0;
91 }
1.4.2 Brents Algorithm
1 /*
2
3 For a function f which maps a finite set S to itself and any initial
4 value x[0] in S, the same value must occur twice in the sequence below:
5
6 x[0], x[1] = f(x[0]), x[2] = f(x[1]), ..., x[i] = f(x[i - 1])
7
8 That is, there must exist numbers i, j (i < j) such that x[i] = x[j].
9 Once this happens, the sequence will continue periodically by repeating
10 the same sequence of values from x[i] to x[j 1]. Cycle detection asks
11 to find i and j, given the function f and initial value x[0]. This is
12 also analogous to the problem of detecting a cycle in a linked list,
13 which will make it degenerate.
14
15 While Floyds cycle-finding algorithm finds cycles by simultaneously
16 moving two pointers at different speeds, Brents algorithm keeps the
17 tortoise pointer stationary and "teleports" it to the hare pointer
18 every power of two. The smallest power of two for which they meet is
19 the start of the first cycle. This improves upon the constant factor
20 of Floyds algorithm by reducing the number of function calls.
21
22 Time Complexity: O(mu + lambda), where mu is the smallest index of the
23 sequence on which a cycle starts, and lambda is the cycles length.

24
26
27 */
28
29
31
32 template<class IntFunction>
33 std::pair<int, int> find_cycle(IntFunction f, int x0) {
34 int power = 1, length = 1;
35 int tortoise = x0, hare = f(x0);
37 if (power == length) {
38 tortoise = hare;
39 power *= 2;
40 length = 0;
41 }
42 hare = f(hare);
43 length++;
44 }
45 hare = x0;
46 for (int i = 0; i < length; i++)
47 hare = f(hare);
48 int start = 0;
49 tortoise = x0;
52 hare = f(hare);
53 start++;
54 }
55 return std::make_pair(start, length);
56 }
57
59
61 #include <set>
63
64 const int x0 = 0;
65
66 int f(int x) {
67 return (123 * x * x + 4567890) % 1337;
68 }
69
70 void verify(int x0, int start, int length) {
71 set<int> s;
72 int x = x0;
73 for (int i = 0; i < start; i++) {
75 s.insert(x);
76 x = f(x);
77 }
78 int startx = x;
79 s.clear();
80 for (int i = 0; i < length; i++) {
1.5. Binary Exponentiation 35
82 s.insert(x);
83 x = f(x);
84 }
85 assert(startx == x);
86 }
87
88 int main () {
89 pair<int, int> res = find_cycle(f, x0);
90 assert(res == make_pair(4, 2));
91 verify(x0, res.first, res.second);
92 return 0;
93 }
1.5 Binary Exponentiation
1 /*
2
3 Given three positive, signed 64-bit integers, powmod() efficiently
4 computes the power of the first two integers, modulo the third integer.
5 Binary exponentiation, also known as "exponentiation by squaring,"
6 decomposes the computation with the observation that the exponent is
7 reduced by half whenever the base is squared. Odd-numbered exponents
8 can be dealt with by subtracting one and multiplying the overall
9 expression by the base of the power. This yields a logarithmic number
10 of multiplications while avoiding overflow. To further prevent overflow
11 in intermediate multiplications, multiplication can be done using the
12 similar principle of multiplication by adding. Despite using unsigned
13 64-bit integers for intermediate calculations and as parameter types,
14 each argument to powmod() must not exceed 2^63 - 1, the maximum value
15 of a signed 64-bit integer.
16
17 Time Complexity: O(log n) on the exponent of the power.
19
20 */
21
22 typedef unsigned long long int64;
23
24 int64 mulmod(int64 a, int64 b, int64 m) {
25 int64 x = 0, y = a % m;
26 for (; b > 0; b >>= 1) {
27 if (b & 1)
28 x = (x + y) % m;
29 y = (y << 1) % m;
30 }
31 return x % m;
32 }
33
34 int64 powmod(int64 a, int64 b, int64 m) {
35 int64 x = 1, y = a;
36 for (; b > 0; b >>= 1) {
37 if (b & 1)
38 x = mulmod(x, y, m);
39 y = mulmod(y, y, m);
40 }
41 return x % m;
42 }
43
45
47
48 int main() {
49 assert(powmod(2, 10, 1000000007) == 1024);
50 assert(powmod(2, 62, 1000000) == 387904);
51 assert(powmod(10001, 10001, 100000) == 10001);
52 return 0;
53 }
Chapter 2
Graph Theory
2.1 Depth-First Search
2.1.1 Graph Class and Depth-First Search
1 /*
2
3 A graph can be represented as a set of objects (a.k.a. vertices, or
4 nodes) and connections (a.k.a. edges) between pairs of objects. It can
5 also be stored as an adjacency matrix or adjacency list, the latter of
6 which is more space efficient but less time efficient for particular
7 operations such as checking whether a connection exists. A fundamental
8 task to perform on graphs is traversal, where all reachable vertices
9 are visited and actions are performed. Given any arbitrary starting
10 node, depth-first search (DFS) recursively explores each "branch" from
11 the current node as deep as possible before backtracking and following
12 other branches. Depth-first search has many applications, including
13 detecting cycles and solving generic puzzles.
14
15 The following implements a simple graph class using adjacency lists,
16 along with with depth-first search and a few applications. The nodes of
17 the graph are identified by integers indices numbered consecutively
18 starting from 0. The total number nodes will automatically increase
19 based upon the maximum argument ever passed to add_edge().
20
21 Time Complexity:
22 - add_edge() is O(1) amortized per call, or O(n) for n calls where each
23 node index added is at most n.
24 - dfs(), has_cycle(), is_tree(), and is_dag() are each O(n) per call on
25 the number of edges added so far.
26 - All other public member functions are O(1).
27
28 Space Complexity:
29 - O(n) to store a graph of n edges.
30 - dfs(), has_cycle(), is_tree(), and is_dag() each require O(n)
31 auxiliary on the number of edges.
32 - All other public member functions require O(1) auxiliary.
33
34 */
35
37
38 Chapter 2. Graph Theory
36 #include <algorithm> /* std::max */

37 #include <cstddef> /* size_t */
39
40 class graph {
41 std::vector< std::vector<int> > adj;
42 bool _is_directed;
43
44 template<class Action>
45 void dfs(int n, std::vector<bool> & vis, Action act);
46
47 bool has_cycle(int n, int prev, std::vector<bool> & vis,
48 std::vector<bool> & onstack);
49
50 public:
51 graph(bool is_directed = true) {
52 this->_is_directed = is_directed;
53 }
54
55 bool is_directed() const {
56 return _is_directed;
57 }
58
59 size_t nodes() const {
60 return adj.size();
61 }
62
63 std::vector<int>& operator [] (int n) {
64 return adj[n];
65 }
66
67 void add_edge(int u, int v);
68 template<class Action> void dfs(int start, Action act);
69 bool has_cycle();
70 bool is_tree();
71 bool is_dag();
72 };
73
74 void graph::add_edge(int u, int v) {
75 if (u >= (int)adj.size() || v >= (int)adj.size())
76 adj.resize(std::max(u, v) + 1);
77 adj[u].push_back(v);
78 if (!is_directed())
79 adj[v].push_back(u);
80 }
81
82 template<class Action>
83 void graph::dfs(int n, std::vector<bool> & vis, Action act) {
84 act(n);
85 vis[n] = true;
86 std::vector<int>::iterator it;
87 for (it = adj[n].begin(); it != adj[n].end(); ++it) {
88 if (!vis[*it])
89 dfs(*it, vis, act);
90 }
91 }
92
93 template<class Action> void graph::dfs(int start, Action act) {
94 std::vector<bool> vis(nodes(), false);
2.1. Depth-First Search 39
95 dfs(start, vis, act);

96 }
97
98 bool graph::has_cycle(int n, int prev, std::vector<bool> & vis,
99 std::vector<bool> & onstack) {
100 vis[n] = true;
101 onstack[n] = true;
102 std::vector<int>::iterator it;
103 for (it = adj[n].begin(); it != adj[n].end(); ++it) {
104 if (is_directed() && onstack[*it])
105 return true;
106 if (!is_directed() && vis[*it] && *it != prev)
107 return true;
108 if (!vis[*it] && has_cycle(*it, n, vis, onstack))
109 return true;
110 }
111 onstack[n] = false;
112 return false;
113 }
114
115 bool graph::has_cycle() {
116 std::vector<bool> vis(nodes(), false), onstack(nodes(), false);
117 for (int i = 0; i < (int)adj.size(); i++)
118 if (!vis[i] && has_cycle(i, -1, vis, onstack))
119 return true;
120 return false;
121 }
122
123 bool graph::is_tree() {
124 return !is_directed() && !has_cycle();
125 }
126
127 bool graph::is_dag() {
128 return is_directed() && !has_cycle();
129 }
130
132
133 Sample Output:
134 DFS order: 0 1 2 3 4 5 6 7 8 9 10 11
135
136 ***/
137
141
142 void print_node(int n) {
143 cout << n << " ";
144 }
145
146 int main() {
147 {
148 graph g;
149 g.add_edge(0, 1);
150 g.add_edge(0, 6);
151 g.add_edge(0, 7);
152 g.add_edge(1, 2);
153 g.add_edge(1, 5);
154 g.add_edge(2, 3);

155 g.add_edge(2, 4);
156 g.add_edge(7, 8);
157 g.add_edge(7, 11);
158 g.add_edge(8, 9);
159 g.add_edge(8, 10);
160 cout << "DFS order: ";
161 g.dfs(0, print_node);
162 cout << endl;
163 assert(g[0].size() == 3);
164 assert(!g.has_cycle());
165 }
166 {
167 graph tree(false);
168 tree.add_edge(0, 1);
172 assert(tree.is_tree());
174 assert(!tree.is_tree());
175 }
176 return 0;
177 }
2.1.2 Topological Sorting
1 /*
2
3 Description: Given a directed acyclic graph (DAG), order the nodes
4 such that for every edge from a to b, a precedes b in the ordering.
5 Usually, there is more than one possible valid ordering. The
6 following program uses DFS to produce one possible ordering.
7 This can also be used to detect whether the graph is a DAG.
8 Note that the DFS algorithm here produces a reversed topological
9 ordering, so the output must be printed backwards. The graph is
10 stored in an adjacency list.
11
12 Complexity: O(V+E) on the number of vertices and edges.
13
14 =~=~=~=~= Sample Input =~=~=~=~=
15 8 9
16 0 3
17 0 4
18 1 3
19 2 4
20 2 7
21 3 5
22 3 6
23 3 7
24 4 6
25
26 =~=~=~=~= Sample Output =~=~=~=~=
27 The topological order: 2 1 0 4 3 7 6 5
28
29 */
30
31 #include <algorithm> /* std::fill(), std::reverse() */

33 #include <stdexcept> /* std::runtime_error() */
36
37 const int MAXN = 100;
38 vector<bool> vis(MAXN), done(MAXN);
39 vector<int> adj[MAXN], sorted;
40
41 void dfs(int u) {
42 if (vis[u])
43 throw std::runtime_error("Not a DAG.");
44 if (done[u]) return;
45 vis[u] = true;
46 for (int j = 0; j < (int)adj[u].size(); j++)
47 dfs(adj[u][j]);
48 vis[u] = false;
49 done[u] = true;
50 sorted.push_back(u);
51 }
52
53 void toposort(int nodes) {
54 fill(vis.begin(), vis.end(), false);
55 fill(done.begin(), done.end(), false);
56 sorted.clear();
57 for (int i = 0; i < nodes; i++)
58 if (!done[i]) dfs(i);
59 reverse(sorted.begin(), sorted.end());
60 }
61
62 int main() {
63 int nodes, edges, u, v;
64 cin >> nodes >> edges;
65 for (int i = 0; i < edges; i++) {
66 cin >> u >> v;
68 }
69 toposort(nodes);
70 cout << "The topological order:";
71 for (int i = 0; i < (int)sorted.size(); i++)
72 cout << " " << sorted[i];
73 cout << "\n";
74 return 0;
75 }
2.1.3 Eulerian Cycles
1 /*
2
3 Description: A Eulerian trail is a trail in a graph which
4 visits every edge exactly once. Similarly, an Eulerian circuit
5 or Eulerian cycle is an Eulerian trail which starts and ends
6 on the same vertex.
7
8 An undirected graph has an Eulerian cycle if and only if every
9 vertex has even degree, and all of its vertices with nonzero
10 degree belong to a single connected component.

11
12 A directed graph has an Eulerian cycle if and only if every
13 vertex has equal in degree and out degree, and all of its
14 vertices with nonzero degree belong to a single strongly
15 connected component.
16
18
19 =~=~=~=~= Sample Input =~=~=~=~=
20 5 6
21 0 1
22 1 2
23 2 0
24 1 3
25 3 4
26 4 1
27
28 =~=~=~=~= Sample Output =~=~=~=~=
29 Eulerian cycle from 0 (directed): 0 1 3 4 1 2 0
30 Eulerian cycle from 2 (undirected): 2 1 3 4 1 0 2
31
32 */
33
34 #include <algorithm> /* std::reverse() */
38
40
41 vector<int> euler_cycle_directed(vector<int> adj[], int u) {
42 vector<int> stack, res, cur_edge(MAXN);
43 stack.push_back(u);
44 while (!stack.empty()) {
45 u = stack.back();
47 while (cur_edge[u] < (int)adj[u].size()) {
49 u = adj[u][cur_edge[u]++];
50 }
51 res.push_back(u);
52 }
53 reverse(res.begin(), res.end());
54 return res;
55 }
56
57 vector<int> euler_cycle_undirected(vector<int> adj[], int u) {
58 vector<vector<bool> > used(MAXN, vector<bool>(MAXN, false));
59 vector<int> stack, res, cur_edge(MAXN);
61 while (!stack.empty()) {
62 u = stack.back();
64 while (cur_edge[u] < (int)adj[u].size()) {
65 int v = adj[u][cur_edge[u]++];
66 if (!used[min(u, v)][max(u, v)]) {
67 used[min(u, v)][max(u, v)] = 1;
69 u = v;
70 }
71 }
73 }
74 reverse(res.begin(), res.end());
75 return res;
76 }
77
78 int main() {
80 vector<int> g1[5], g2[5], cycle;
81
83 for (int i = 0; i < edges; i++) {
84 cin >> u >> v;
85 g1[u].push_back(v);
86 g2[u].push_back(v);
87 g2[v].push_back(u);
88 }
89
90 cycle = euler_cycle_directed(g1, 0);
91 cout << "Eulerian cycle from 0 (directed):";
92 for (int i = 0; i < (int)cycle.size(); i++)
93 cout << " " << cycle[i];
94 cout << "\n";
95
96 cycle = euler_cycle_undirected(g2, 2);
97 cout << "Eulerian cycle from 2 (undirected):";
98 for (int i = 0; i < (int)cycle.size(); i++)
99 cout << " " << cycle[i];
100 cout << "\n";
101 return 0;
102 }
2.1.4 Unweighted Tree Centers
1 /*
2
3 The following applies to unweighted, undirected trees only.
4
5 find_centers(): Returns 1 or 2 tree centers. The center
6 (or Jordan center) of a graph is the set of all vertices of
7 minimum eccentricity, that is, the set of all vertices A
8 where the max distance d(A,B) to other vertices B is minimal.
9
10 find_centroid(): Returns a vertex where all of its subtrees
11 have size <= N/2, where N is the number of nodes in the tree.
12
13 diameter(): The diameter of a tree is the greatest distance
14 d(A,B) between any two of the nodes in the tree.
15
16 Complexity: All three functions are O(V) on the number of
17 vertices in the tree.
18
19 =~=~=~=~= Sample Input =~=~=~=~=
20 6
21 0 1
22 1 2
23 1 4
24 3 4
25 4 5
26
27 =~=~=~=~= Sample Output =~=~=~=~=
28 Center(s): 1 4
29 Centroid: 4
30 Diameter: 3
31
32 */
33
37
39 vector<int> adj[MAXN];
40
41 vector<int> find_centers(int n) {
42 vector<int> leaves, degree(n);
43 for (int i = 0; i < n; i++) {
44 degree[i] = adj[i].size();
45 if (degree[i] <= 1) leaves.push_back(i);
46 }
47 int removed = leaves.size();
48 while (removed < n) {
49 vector<int> nleaves;
50 for (int i = 0; i < (int)leaves.size(); i++) {
51 int u = leaves[i];
52 for (int j = 0; j < (int)adj[u].size(); j++) {
53 int v = adj[u][j];
54 if (--degree[v] == 1)
55 nleaves.push_back(v);
56 }
57 }
58 leaves = nleaves;
59 removed += leaves.size();
60 }
61 return leaves;
62 }
63
64 int find_centroid(int n, int u = 0, int p = -1) {
65 int cnt = 1, v;
66 bool good_center = true;
68 if ((v = adj[u][j]) == p) continue;
69 int res = find_centroid(n, v, u);
70 if (res >= 0) return res;
71 int size = -res;
72 good_center &= (size <= n / 2);
73 cnt += size;
74 }
75 good_center &= (n - cnt <= n / 2);
76 return good_center ? u : -cnt;
77 }
78
79 pair<int, int> dfs(int u, int p, int depth) {
2.2. Shortest Paths 45
80 pair<int, int> res = make_pair(depth, u);

81 for (int j = 0; j < (int)adj[u].size(); j++)
82 if (adj[u][j] != p)
83 res = max(res, dfs(adj[u][j], u, depth + 1));
84 return res;
85 }
86
87 int diameter() {
88 int furthest_vertex = dfs(0, -1, 0).second;
89 return dfs(furthest_vertex, -1, 0).first;
90 }
91
92 int main() {
93 int nodes, u, v;
94 cin >> nodes;
95 for (int i = 0; i < nodes - 1; i++) {
96 cin >> u >> v;
99 }
100 vector<int> centers = find_centers(nodes);
101 cout << "Center(s):";
102 for (int i = 0; i < (int)centers.size(); i++)
103 cout << " " << centers[i];
104 cout << "\nCentroid: " << find_centroid(nodes);
105 cout << "\nDiameter: " << diameter() << "\n";
106 return 0;
107 }
2.2 Shortest Paths
2.2.1 Breadth First Search
1 /*
2
3 Description: Given an unweighted graph, traverse all reachable
4 nodes from a source node and determine the shortest path.
5
7
8 Note: The line "for (q.push(start); !q.empty(); q.pop())"
9 is simply a mnemonic for looping a BFS with a FIFO queue.
10 This will not work as intended with a priority queue, such as in
11 Dijkstras algorithm for solving weighted shortest paths
12
13 =~=~=~=~= Sample Input =~=~=~=~=
14 4 5
15 0 1
16 0 3
17 1 2
18 1 3
19 2 3
20 0 3
21
22 =~=~=~=~= Sample Output =~=~=~=~=
23 The shortest distance from 0 to 3 is 2.
24 Take the path: 0->1->3.

25
26 */
27
29 #include <queue>
32
33 const int MAXN = 100, INF = 0x3f3f3f3f;
34 int dist[MAXN], pred[MAXN];
36
37 void bfs(int nodes, int start) {
38 vector<bool> vis(nodes, false);
39 for (int i = 0; i < nodes; i++) {
40 dist[i] = INF;
41 pred[i] = -1;
42 }
43 int u, v, d;
44 queue<pair<int, int> > q;
45 q.push(make_pair(start, 0));
46 while (!q.empty()) {
47 u = q.front().first;
48 d = q.front().second;
49 q.pop();
50 vis[u] = true;
52 if (vis[v = adj[u][j]]) continue;
53 dist[v] = d + 1;
54 pred[v] = u;
55 q.push(make_pair(v, d + 1));
56 }
57 }
58 }
59
60 //Use the precomputed pred[] array to print the path
61 void print_path(int dest) {
62 int i = 0, j = dest, path[MAXN];
63 while (pred[j] != -1) j = path[++i] = pred[j];
64 cout << "Take the path: ";
65 while (i > 0) cout << path[i--] << "->";
66 cout << dest << ".\n";
67 }
68
69 int main() {
70 int nodes, edges, u, v, start, dest;
72 for (int i = 0; i < edges; i++) {
73 cin >> u >> v;
75 }
76 cin >> start >> dest;
77 bfs(nodes, start);
78 cout << "The shortest distance from " << start;
79 cout << " to " << dest << " is " << dist[dest] << ".\n";
80 print_path(dest);
81 return 0;
82 }
2.2.2 Dijkstras Algorithm
1 /*
2
3 Description: Given a directed graph with positive weights only, find
4 the shortest distance to all nodes from a single starting node.
5
6 Implementation Notes: The graph is stored using an adjacency list.
7 This implementation negates distances before adding them to the
8 priority queue, since the container is a max-heap by default. This
9 method is suggested in contests because it is easier than defining
10 special comparators. An alternative would be declaring the queue
11 with template parameters (clearly, this way is very verbose and ugly):
12 priority_queue< pair<int, int>, vector<pair<int, int> >,
13 greater<pair<int, int> > > pq;
14 If only the path between a single pair of nodes is needed, for speed,
15 we may break out of the loop as soon as the destination is reached
16 by inserting the line "if (a == dest) break;" after the line "pq.pop();"
17
18 Complexity: This version uses an adjacency list and priority queue
19 (internally a binary heap) and has a complexity of O((E+V) log V) =
20 O(E log V). The priority queue and adjacency list improves the
21 simplest O(V^2) version of the algorithm, which uses looping and
22 an adjacency matrix. If the priority queue is implemented as a more
23 sophisticated Fibonacci heap, the complexity becomes O(E + V log V).
24
25 Modification to Shortest Path Faster Algorithm: The code for Dijkstras
26 algorithm here can be easily modified to become the Shortest Path Faster
27 Algorithm (SPFA) by simply commenting out "visit[a] = true;" and changing
28 the priority queue to a FIFO queue like in BFS. SPFA is a faster version
29 of the Bellman-Ford algorithm, working on negative path lengths (whereas
30 Dijkstras cannot). Certain graphs can be constructed to make SPFA slow.
31
32 =~=~=~=~= Sample Input =~=~=~=~=
33 4 5
34 0 1 2
35 0 3 8
36 1 2 2
37 1 3 4
38 2 3 1
39 0 3
40
41 =~=~=~=~= Sample Output =~=~=~=~=
43 Take the path: 0->1->2->3.
44
45 */
46
48 #include <queue>
51
54 vector<pair<int, int> > adj[MAXN];
55
56 void dijkstra(int nodes, int start) {
57 vector<bool> vis(nodes, false);

58 for (int i = 0; i < nodes; i++) {
59 dist[i] = INF;
60 pred[i] = -1;
61 }
62 int u, v;
63 dist[start] = 0;
64 priority_queue<pair<int, int> > pq;
65 pq.push(make_pair(0, start));
66 while (!pq.empty()) {
67 u = pq.top().second;
68 pq.pop();
69 vis[u] = true;
71 if (vis[v = adj[u][j].first]) continue;
72 if (dist[v] > dist[u] + adj[u][j].second) {
73 dist[v] = dist[u] + adj[u][j].second;
74 pred[v] = u;
75 pq.push(make_pair(-dist[v], v));
76 }
77 }
78 }
79 }
80
87 cout << dest << ".\n";
88 }
89
90 int main() {
91 int nodes, edges, u, v, w, start, dest;
93 for (int i = 0; i < edges; i++) {
94 cin >> u >> v >> w;
95 adj[u].push_back(make_pair(v, w));
96 }
98 dijkstra(nodes, start);
102 return 0;
103 }
2.2.3 Bellman-Ford Algorithm
1 /*
2
3 Description: Given a directed graph with positive or negative weights
4 but no negative cycles, find the shortest distance to all nodes from
5 a single starting node. The input graph is stored using an edge list.
6
7 Complexity: O(V*E) on the number of vertices and edges, respectively.
8
9 =~=~=~=~= Sample Input =~=~=~=~=
10 3 3
11 0 1 1
12 1 2 2
13 0 2 5
14 0 2
15
16 =~=~=~=~= Sample Output =~=~=~=~=
19
20 */
21
23 #include <stdexcept>
26
27 struct edge { int u, v, w; };
28
31 vector<edge> e;
32
33 void bellman_ford(int nodes, int start) {
34 for (int i = 0; i < nodes; i++) {
35 dist[i] = INF;
36 pred[i] = -1;
37 }
38 dist[start] = 0;
39 for (int i = 0; i < nodes; i++) {
40 for (int j = 0; j < (int)e.size(); j++) {
41 if (dist[e[j].v] > dist[e[j].u] + e[j].w) {
42 dist[e[j].v] = dist[e[j].u] + e[j].w;
43 pred[e[j].v] = e[j].u;
44 }
45 }
46 }
47 //optional: report negative-weight cycles
48 for (int i = 0; i < (int)e.size(); i++)
49 if (dist[e[i].v] > dist[e[i].u] + e[i].w)
50 throw std::runtime_error("Negative-weight found");
51 }
52
59 cout << dest << ".\n";
60 }
61
62 int main() {
65 for (int i = 0; i < edges; i++) {
66 cin >> u >> v >> w;
67 e.push_back((edge){u, v, w});
68 }
70 bellman_ford(nodes, start);
74 return 0;
75 }
2.2.4 Floyd-Warshall Algorithm
1 /*
2
3 Description: Given a directed graph with positive or negative
4 weights but no negative cycles, find the shortest distance
5 between all pairs of nodes. The input graph is stored using
6 an adjacency matrix. Note that the input adjacency matrix
7 is converted to the distance matrix afterwards. If you still
8 need the adjacencies afterwards, back it up at the beginning.
9
10 Complexity: O(V^3) on the number of vertices.
11
12 =~=~=~=~= Sample Input =~=~=~=~=
13 3 3
14 0 1 1
15 1 2 2
16 0 2 5
17 0 2
18
19 =~=~=~=~= Sample Output =~=~=~=~=
22
23 */
24
27
29 int dist[MAXN][MAXN], next[MAXN][MAXN];
30
31 void initialize(int nodes) {
33 for (int j = 0; j < nodes; j++) {
34 dist[i][j] = (i == j) ? 0 : INF;
35 next[i][j] = -1;
36 }
37 }
38
39 void floyd_warshall(int nodes) {
40 for (int k = 0; k < nodes; k++)
42 for (int j = 0; j < nodes; j++)
43 if (dist[i][j] > dist[i][k] + dist[k][j]) {
44 dist[i][j] = dist[i][k] + dist[k][j];
45 next[i][j] = k;
2.3. Connectivity 51
46 }
47 }
48
49 void print_path(int u, int v) {
50 if (next[u][v] != -1) {
51 print_path(u, next[u][v]);
52 cout << next[u][v];
53 print_path(next[u][v], v);
54 } else cout << "->";
55 }
56
57 int main() {
60 initialize(nodes);
61 for (int i = 0; i < edges; i++) {
62 cin >> u >> v >> w;
63 dist[u][v] = w;
64 }
66 floyd_warshall(nodes);
68 cout << " to " << dest << " is ";
69 cout << dist[start][dest] << ".\n";
70
71 //Use next[][] to recursively print the path
72 cout << "Take the path " << start;
73 print_path(start, dest);
74 cout << dest << ".\n";
75 return 0;
76 }
2.3 Connectivity
2.3.1 Strongly Connected Components (Kosarajus Algorithm)
1 /*
2
3 Description: Determines the strongly connected components (SCC)
4 from a given directed graph. Given a directed graph, its SCCs
5 are its maximal strongly connected sub-graphs. A graph is
6 strongly connected if there is a path from each node to every
7 other node. Condensing the strongly connected components of a
8 graph into single nodes will result in a directed acyclic graph.
9 The input is stored in an adjacency list.
10
12
13 Comparison with other SCC algorithms:
14 The strongly connected components of a graph can be efficiently
15 computed using Kosarajus algorithm, Tarjans algorithm, or the
16 path-based strong component algorithm. Tarjans algorithm can
17 be seen as an improved version of Kosarajus because it performs
18 a single DFS rather than two. Though they both have the same
19 complexity, Tarjans algorithm is much more efficient in
20 practice. However, Kosarajus algorithm is conceptually simpler.
21
22 =~=~=~=~= Sample Input =~=~=~=~=
23 8 14
24 0 1
25 1 2
26 1 4
27 1 5
28 2 3
29 2 6
30 3 2
31 3 7
32 4 0
33 4 5
34 5 6
35 6 5
36 7 3
37 7 6
38
39 =~=~=~=~= Sample Output =~=~=~=~=
40 Component: 1 4 0
41 Component: 7 3 2
42 Component: 5 6
43
44 */
45
46 #include <algorithm> /* std::fill(), std::reverse() */
50
52 vector<bool> vis(MAXN);
53 vector<int> adj[MAXN], order;
54 vector<vector<int> > scc;
55
56 void dfs(vector<int> graph[], vector<int> & res, int u) {
57 vis[u] = true;
58 for (int j = 0; j < (int)graph[u].size(); j++)
59 if (!vis[graph[u][j]])
60 dfs(graph, res, graph[u][j]);
62 }
63
64 void kosaraju(int nodes) {
65 scc.clear();
66 order.clear();
67 vector<int> rev[nodes];
70 if (!vis[i]) dfs(adj, order, i);
72 for (int j = 0; j < (int)adj[i].size(); j++)
73 rev[adj[i][j]].push_back(i);
75 reverse(order.begin(), order.end());
76 for (int i = 0; i < (int)order.size(); i++) {
77 if (vis[order[i]]) continue;
78 vector<int> component;
79 dfs(rev, component, order[i]);
80 scc.push_back(component);
81 }
82 }
83
84 int main() {
87 for (int i = 0; i < edges; i++) {
88 cin >> u >> v;
90 }
91 kosaraju(nodes);
92 for (int i = 0; i < (int)scc.size(); i++) {
93 cout << "Component:";
94 for (int j = 0; j < (int)scc[i].size(); j++)
95 cout << " " << scc[i][j];
96 cout << "\n";
97 }
98 return 0;
99 }
2.3.2 Strongly Connected Components (Tarjans Algorithm)
1 /*
2
3 Description: Determines the strongly connected components (SCC)
4 from a given directed graph. Given a directed graph, its SCCs
5 are its maximal strongly connected sub-graphs. A graph is
6 strongly connected if there is a path from each node to every
7 other node. Condensing the strongly connected components of a
8 graph into single nodes will result in a directed acyclic graph.
9 The input is stored in an adjacency list.
10
11 In this implementation, a vector is used to emulate a stack
12 for the sake of simplicity. One useful property of Tarjans
13 algorithm is that, while there is nothing special about the
14 ordering of nodes within each component, the resulting DAG
15 is produced in reverse topological order.
16
18
19 Comparison with other SCC algorithms:
20 The strongly connected components of a graph can be efficiently
21 computed using Kosarajus algorithm, Tarjans algorithm, or the
22 path-based strong component algorithm. Tarjans algorithm can
23 be seen as an improved version of Kosarajus because it performs
24 a single DFS rather than two. Though they both have the same
25 complexity, Tarjans algorithm is much more efficient in
26 practice. However, Kosarajus algorithm is conceptually simpler.
27
28 =~=~=~=~= Sample Input =~=~=~=~=
29 8 14
30 0 1
31 1 2
32 1 4
33 1 5
34 2 3
35 2 6
36 3 2
37 3 7
38 4 0
39 4 5
40 5 6
41 6 5
42 7 3
43 7 6
44
45 =~=~=~=~= Sample Output =~=~=~=~=
46 Component 1: 5 6
47 Component 2: 7 3 2
49
50 */
51
56
58 int timer, lowlink[MAXN];
60 vector<int> adj[MAXN], stack;
61 vector<vector<int> > scc;
62
63 void dfs(int u) {
64 lowlink[u] = timer++;
65 vis[u] = true;
67 bool is_component_root = true;
68 int v;
70 if (!vis[v = adj[u][j]]) dfs(v);
71 if (lowlink[u] > lowlink[v]) {
72 lowlink[u] = lowlink[v];
73 is_component_root = false;
74 }
75 }
76 if (!is_component_root) return;
78 do {
79 vis[v = stack.back()] = true;
81 lowlink[v] = INF;
82 component.push_back(v);
83 } while (u != v);
84 scc.push_back(component);
85 }
86
87 void tarjan(int nodes) {
88 scc.clear();
89 stack.clear();
90 fill(lowlink, lowlink + nodes, 0);
92 timer = 0;
94 if (!vis[i]) dfs(i);
95 }
96
97 int main() {
100 for (int i = 0; i < edges; i++) {
101 cin >> u >> v;
103 }
104 tarjan(nodes);
105 for (int i = 0; i < (int)scc.size(); i++) {
107 for (int j = 0; j < (int)scc[i].size(); j++)
108 cout << " " << scc[i][j];
109 cout << "\n";
110 }
111 return 0;
112 }
2.3.3 Bridges, Cut-points, and Biconnectivity
1 /*
2
3 Description: The following operations apply to undirected graphs.
4
5 A bridge is an edge, when deleted, increases the number of
6 connected components. An edge is a bridge if and only ifit is not
7 contained in any cycle.
8
9 A cut-point (i.e. cut-vertex or articulation point) is any vertex
10 whose removal increases the number of connected components.
11
12 A biconnected component of a graph is a maximally biconnected
13 subgraph. A biconnected graph is a connected and "nonseparable"
14 graph, meaning that if any vertex were to be removed, the graph
15 will remain connected. Therefore, a biconnected graph has no
16 articulation vertices.
17
18 Any connected graph decomposes into a tree of biconnected
19 components called the "block tree" of the graph. An unconnected
20 graph will thus decompose into a "block forest."
21
22 See: http://en.wikipedia.org/wiki/Biconnected_component
23
25
26 =~=~=~=~= Sample Input =~=~=~=~=
27 8 6
28 0 1
29 0 5
30 1 2
31 1 5
32 3 7
33 4 5
34
35 =~=~=~=~= Sample Output =~=~=~=~=
36 Cut Points: 5 1
37 Bridges:
38 1 2
39 5 4
40 3 7
41 Edge-Biconnected Components:
42 Component 1: 2
43 Component 2: 4
45 Component 4: 7
46 Component 5: 3
47 Component 6: 6
48 Adjacency List for Block Forest:
49 0 => 2
50 1 => 2
51 2 => 0 1
52 3 => 4
53 4 => 3
54 5 =>
55
56 */
57
58 #include <algorithm> /* std::fill(), std::min() */
62
64 int timer, lowlink[MAXN], tin[MAXN], comp[MAXN];
66 vector<int> adj[MAXN], bcc_forest[MAXN];
67 vector<int> stack, cutpoints;
68 vector<vector<int> > bcc;
69 vector<pair<int, int> > bridges;
70
71 void dfs(int u, int p) {
72 vis[u] = true;
73 lowlink[u] = tin[u] = timer++;
75 int v, children = 0;
76 bool cutpoint = false;
79 if (vis[v]) {
80 //lowlink[u] = min(lowlink[u], lowlink[v]);
81 lowlink[u] = min(lowlink[u], tin[v]);
82 } else {
83 dfs(v, u);
84 lowlink[u] = min(lowlink[u], lowlink[v]);
85 cutpoint |= (lowlink[v] >= tin[u]);
86 if (lowlink[v] > tin[u])
87 bridges.push_back(make_pair(u, v));
88 children++;
89 }
90 }
91 if (p == -1) cutpoint = (children >= 2);
92 if (cutpoint) cutpoints.push_back(u);
93 if (lowlink[u] == tin[u]) {
95 do {
96 v = stack.back();
98 component.push_back(v);
99 } while (u != v);
100 bcc.push_back(component);
101 }
102 }
103
104 void tarjan(int nodes) {
105 bcc.clear();
106 bridges.clear();
107 cutpoints.clear();
108 stack.clear();
109 fill(lowlink, lowlink + nodes, 0);
110 fill(tin, tin + nodes, 0);
112 timer = 0;
114 if (!vis[i]) dfs(i, -1);
115 }
116
117 //condenses each bcc to a node and generates a tree
118 //global variables adj and bcc must be set beforehand
119 void get_block_tree(int nodes) {
120 fill(comp, comp + nodes, 0);
121 for (int i = 0; i < nodes; i++) bcc_forest[i].clear();
122 for (int i = 0; i < (int)bcc.size(); i++)
123 for (int j = 0; j < (int)bcc[i].size(); j++)
124 comp[bcc[i][j]] = i;
127 if (comp[i] != comp[adj[i][j]])
128 bcc_forest[comp[i]].push_back(comp[adj[i][j]]);
129 }
130
131 int main() {
134 for (int i = 0; i < edges; i++) {
135 cin >> u >> v;
138 }
139 tarjan(nodes);
140 cout << "Cut-points:";
141 for (int i = 0; i < (int)cutpoints.size(); i++)
142 cout << " " << cutpoints[i];
143 cout << "\nBridges:\n";
144 for (int i = 0; i < (int)bridges.size(); i++)
145 cout << bridges[i].first << " " << bridges[i].second << "\n";
146 cout << "Edge-Biconnected Components:\n";
147 for (int i = 0; i < (int)bcc.size(); i++) {
149 for (int j = 0; j < (int)bcc[i].size(); j++)
150 cout << " " << bcc[i][j];
151 cout << "\n";
152 }
153 get_block_tree(nodes);
154 cout << "Adjacency List for Block Forest:\n";

155 for (int i = 0; i < (int)bcc.size(); i++) {
156 cout << i << " =>";
157 for (int j = 0; j < (int)bcc_forest[i].size(); j++)
158 cout << " " << bcc_forest[i][j];
159 cout << "\n";
160 }
161 return 0;
162 }
2.4 Minimal Spanning Trees
2.4.1 Prims Algorithm
1 /*
2
3 Description: Given an undirected graph, its minimum spanning
4 tree (MST) is a tree connecting all nodes with a subset of its
5 edges such that their total weight is minimized. Prims algorithm
6 greedily selects edges from a priority queue, and is similar to
7 Dijkstras algorithm, where instead of processing nodes, we
8 process individual edges. If the graph is not connected, Prims
9 algorithm will produce the minimum spanning forest. The input
10 graph is stored in an adjacency list.
11
12 Note that the concept of the minimum spanning tree makes Prims
13 algorithm work with negative weights. In fact, a big positive
14 constant added to all of the edge weights of the graph will not
15 change the resulting spanning tree.
16
17 Implementation Notes: Similar to the implementation of Dijkstras
18 algorithm in the previous section, weights are negated before they
19 are added to the priority queue (and negated once again when they
20 are retrieved). To find the maximum spanning tree, simply skip the
21 two negation steps and the max weighted edges will be prioritized.
22
23 Complexity: This version uses an adjacency list and priority queue
24 (internally a binary heap) and has a complexity of O((E+V) log V) =
25 O(E log V). The priority queue and adjacency list improves the
26 simplest O(V^2) version of the algorithm, which uses looping and
27 an adjacency matrix. If the priority queue is implemented as a more
28 sophisticated Fibonacci heap, the complexity becomes O(E + V log V).
29
30 =~=~=~=~= Sample Input =~=~=~=~=
31 7 7
32 0 1 4
33 1 2 6
34 2 0 3
35 3 4 1
36 4 5 2
37 5 6 3
38 6 4 4
39
40 =~=~=~=~= Sample Output =~=~=~=~=
41 Total distance: 13
42 0<->2
2.4. Minimal Spanning Trees 59
43 0<->1
44 3<->4
45 4<->5
46 5<->6
47
48 */
49
52 #include <queue>
55
57 vector<pair<int, int> > adj[MAXN], mst;
58
59 int prim(int nodes) {
60 mst.clear();
61 vector<bool> vis(nodes);
62 int u, v, w, total_dist = 0;
63 for (int i = 0; i < nodes; i++) {
64 if (vis[i]) continue;
65 vis[i] = true;
66 priority_queue<pair<int, pair<int, int> > > pq;
68 pq.push(make_pair(-adj[i][j].second,
69 make_pair(i, adj[i][j].first)));
70 while (!pq.empty()) {
71 w = -pq.top().first;
72 u = pq.top().second.first;
73 v = pq.top().second.second;
74 pq.pop();
75 if (vis[u] && !vis[v]) {
76 vis[v] = true;
77 if (v != i) {
78 mst.push_back(make_pair(u, v));
79 total_dist += w;
80 }
81 for (int j = 0; j < (int)adj[v].size(); j++)
82 pq.push(make_pair(-adj[v][j].second,
83 make_pair(v, adj[v][j].first)));
84 }
85 }
86 }
87 return total_dist;
88 }
89
90 int main() {
91 int nodes, edges, u, v, w;
93 for (int i = 0; i < edges; i++) {
94 cin >> u >> v >> w;
95 adj[u].push_back(make_pair(v, w));
96 adj[v].push_back(make_pair(u, w));
97 }
98 cout << "Total distance: " << prim(nodes) << "\n";
99 for (int i = 0; i < (int)mst.size(); i++)
100 cout << mst[i].first << "<->" << mst[i].second << "\n";
101 return 0;
102 }
2.4.2 Kruskals Algorithm
1 /*
2
3 Description: Given an undirected graph, its minimum spanning
4 tree (MST) is a tree connecting all nodes with a subset of its
5 edges such that their total weight is minimized. If the graph
6 is not connected, Kruskals algorithm will produce the minimum
7 spanning forest. The input graph is stored in an edge list.
8
9 Complexity: O(E log V) on the number of edges and vertices.
10
11 =~=~=~=~= Sample Input =~=~=~=~=
12 7 7
13 0 1 4
14 1 2 6
15 2 0 3
16 3 4 1
17 4 5 2
18 5 6 3
19 6 4 4
20
21 =~=~=~=~= Sample Output =~=~=~=~=
22 Total distance: 13
23 3<->4
24 4<->5
25 2<->0
26 5<->6
27 0<->1
28
29 Note: If you already have a disjoint set data structure,
30 then the middle section of the program can be replaced by:
31
32 disjoint_set_forest<int> dsf;
33 for (int i = 0; i < nodes; i++) dsf.make_set(i);
34 for (int i = 0; i < E.size(); i++) {
35 a = E[i].second.first;
36 b = E[i].second.second;
37 if (!dsf.is_united(a, b)) {
38 ...
39 dsf.unite(a, b);
40 }
41 }
42
43 */
44
45 #include <algorithm> /* std::sort() */
49
51 int root[MAXN];
52 vector<pair<int, pair<int, int> > > E;
53 vector<pair<int, int> > mst;
2.5. Maximum Flow 61
54
55 int find_root(int x) {
56 if (root[x] != x)
57 root[x] = find_root(root[x]);
58 return root[x];
59 }
60
61 int kruskal(int nodes) {
62 mst.clear();
63 sort(E.begin(), E.end());
64 int u, v, total_dist = 0;
65 for (int i = 0; i < nodes; i++) root[i] = i;
66 for (int i = 0; i < (int)E.size(); i++) {
67 u = find_root(E[i].second.first);
68 v = find_root(E[i].second.second);
69 if (u != v) {
70 mst.push_back(E[i].second);
71 total_dist += E[i].first;
72 root[u] = root[v];
73 }
74 }
75 return total_dist;
76 }
77
78 int main() {
81 for (int i = 0; i < edges; i++) {
82 cin >> u >> v >> w;
83 E.push_back(make_pair(w, make_pair(u, v)));
84 }
85 cout << "Total distance: " << kruskal(nodes) << "\n";
86 for (int i = 0; i < (int)mst.size(); i++)
87 cout << mst[i].first << "<->" << mst[i].second << "\n";
88 return 0;
89 }
2.5 Maximum Flow
2.5.1 Ford-Fulkerson Algorithm
1 /*
2
3 Description: Given a flow network, find a flow from a single
4 source node to a single sink node that is maximized. Note
5 that in this implementation, the adjacency matrix cap[][]
6 will be modified by the function ford_fulkerson() after its
7 been called. Make a back-up if you require it afterwards.
8
9 Complexity: O(V^2*|F|), where V is the number of
10 vertices and |F| is the magnitude of the max flow.
11
12 Real-valued capacities:
13 The Ford-Fulkerson algorithm is only optimal on graphs with
14 integer capacities; there exists certain real capacity inputs
15 for which it will never terminate. The Edmonds-Karp algorithm
16 is an improvement using BFS, supporting real number capacities.

17
18 =~=~=~=~= Sample Input =~=~=~=~=
19 6 8
20 0 1 3
21 0 2 3
22 1 2 2
23 1 3 3
24 2 4 2
25 3 4 1
26 3 5 2
27 4 5 3
28 0 5
29
30 =~=~=~=~= Sample Output =~=~=~=~=
31 5
32
33 */
34
39
41 int nodes, source, sink, cap[MAXN][MAXN];
43
44 int dfs(int u, int f) {
45 if (u == sink) return f;
46 vis[u] = true;
47 for (int v = 0; v < nodes; v++) {
48 if (!vis[v] && cap[u][v] > 0) {
49 int df = dfs(v, min(f, cap[u][v]));
50 if (df > 0) {
51 cap[u][v] -= df;
52 cap[v][u] += df;
53 return df;
54 }
55 }
56 }
57 return 0;
58 }
59
60 int ford_fulkerson() {
61 int max_flow = 0;
62 for (;;) {
64 int df = dfs(source, INF);
65 if (df == 0) break;
66 max_flow += df;
67 }
68 return max_flow;
69 }
70
71 int main() {
72 int edges, u, v, capacity;
74 for (int i = 0; i < edges; i++) {
75 cin >> u >> v >> capacity;

76 cap[u][v] = capacity;
77 }
78 cin >> source >> sink;
79 cout << ford_fulkerson() << "\n";
80 return 0;
81 }
2.5.2 Edmonds-Karp Algorithm
1 /*
2
5 that in this implementation, the adjacency list adj[] will
6 be modified by the function edmonds_karp() after its been called.
7
8 Complexity: O(min(V*E^2, E*|F|)), where V is the number of
9 vertices, E is the number of edges, and |F| is the magnitude of
10 the max flow. This improves the original Ford-Fulkerson algorithm,
11 which runs in O(E*|F|). As the Edmonds-Karp algorithm is also
12 bounded by O(E*|F|), it is guaranteed to be at least as fast as
13 Ford-Fulkerson. For an even faster algorithm, see Dinics
14 algorithm in the next section, which runs in O(V^2*E).
15
16 Real-valued capacities:
17 Although the Ford-Fulkerson algorithm is only optimal on graphs
18 with integer capacities, the Edmonds-Karp algorithm also works
19 correctly on real-valued capacities.
20
21 =~=~=~=~= Sample Input =~=~=~=~=
22 6 8
23 0 1 3
24 0 2 3
25 1 2 2
26 1 3 3
27 2 4 2
28 3 4 1
29 3 5 2
30 4 5 3
31 0 5
32
33 =~=~=~=~= Sample Output =~=~=~=~=
34 5
35
36 */
37
42
43 struct edge { int s, t, rev, cap, f; };
44
46 vector<edge> adj[MAXN];
47
48 void add_edge(int s, int t, int cap) {

49 adj[s].push_back((edge){s, t, (int)adj[t].size(), cap, 0});
50 adj[t].push_back((edge){t, s, (int)adj[s].size() - 1, 0, 0});
51 }
52
53 int edmonds_karp(int nodes, int source, int sink) {
54 static int q[MAXN];
56 for (;;) {
57 int qt = 0;
58 q[qt++] = source;
59 edge * pred[nodes];
60 fill(pred, pred + nodes, (edge*)0);
61 for (int qh = 0; qh < qt && !pred[sink]; qh++) {
62 int u = q[qh];
64 edge * e = &adj[u][j];
65 if (!pred[e->t] && e->cap > e->f) {
66 pred[e->t] = e;
67 q[qt++] = e->t;
68 }
69 }
70 }
71 if (!pred[sink]) break;
72 int df = INF;
73 for (int u = sink; u != source; u = pred[u]->s)
74 df = min(df, pred[u]->cap - pred[u]->f);
75 for (int u = sink; u != source; u = pred[u]->s) {
76 pred[u]->f += df;
77 adj[pred[u]->t][pred[u]->rev].f -= df;
78 }
79 max_flow += df;
80 }
81 return max_flow;
82 }
83
84 int main() {
85 int nodes, edges, u, v, capacity, source, sink;
87 for (int i = 0; i < edges; i++) {
89 add_edge(u, v, capacity);
90 }
92 cout << edmonds_karp(nodes, source, sink) << "\n";
93 return 0;
94 }
2.5.3 Dinics Algorithm
1 /*
2
5 that in this implementation, the adjacency list adj[] will
6 be modified by the function dinic() after its been called.
7
8 Complexity: O(V^2*E) on the number of vertices and edges.

9
10 Comparison with Edmonds-Karp Algorithm:
11 Dinics is similar to the Edmonds-Karp algorithm in that it
12 uses the shortest augmenting path. The introduction of the
13 concepts of the level graph and blocking flow enable Dinics
14 algorithm to achieve its better performance. Hence, Dinics
15 algorithm is also called Dinics blocking flow algorithm.
16
17 =~=~=~=~= Sample Input =~=~=~=~=
18 6 8
19 0 1 3
20 0 2 3
21 1 2 2
22 1 3 3
23 2 4 2
24 3 4 1
25 3 5 2
26 4 5 3
27 0 5
28
29 =~=~=~=~= Sample Output =~=~=~=~=
30 5
31
32 */
33
38
39 struct edge { int to, rev, cap, f; };
40
42 int dist[MAXN], ptr[MAXN];
43 vector<edge> adj[MAXN];
44
45 void add_edge(int s, int t, int cap) {
46 adj[s].push_back((edge){t, (int)adj[t].size(), cap, 0});
47 adj[t].push_back((edge){s, (int)adj[s].size() - 1, 0, 0});
48 }
49
50 bool dinic_bfs(int nodes, int source, int sink) {
51 fill(dist, dist + nodes, -1);
52 dist[source] = 0;
53 int q[nodes], qh = 0, qt = 0;
54 q[qt++] = source;
55 while (qh < qt) {
56 int u = q[qh++];
58 edge & e = adj[u][j];
59 if (dist[e.to] < 0 && e.f < e.cap) {
60 dist[e.to] = dist[u] + 1;
61 q[qt++] = e.to;
62 }
63 }
64 }
65 return dist[sink] >= 0;
66 }
67
68 int dinic_dfs(int u, int f, int sink) {
69 if (u == sink) return f;
70 for (; ptr[u] < (int)adj[u].size(); ptr[u]++) {
71 edge &e = adj[u][ptr[u]];
72 if (dist[e.to] == dist[u] + 1 && e.f < e.cap) {
73 int df = dinic_dfs(e.to, min(f, e.cap - e.f), sink);
74 if (df > 0) {
75 e.f += df;
76 adj[e.to][e.rev].f -= df;
77 return df;
78 }
79 }
80 }
81 return 0;
82 }
83
84 int dinic(int nodes, int source, int sink) {
85 int max_flow = 0, delta;
86 while (dinic_bfs(nodes, source, sink)) {
87 fill(ptr, ptr + nodes, 0);
88 while ((delta = dinic_dfs(source, INF, sink)) != 0)
89 max_flow += delta;
90 }
91 return max_flow;
92 }
93
94 int main() {
97 for (int i = 0; i < edges; i++) {
99 add_edge(u, v, capacity);
100 }
102 cout << dinic(nodes, source, sink) << "\n";
103 return 0;
104 }
2.5.4 Push-Relabel Algorithm
1 /*
2
4 source node to a single sink node that is maximized. The push-
5 relabel algorithm is considered one of the most efficient
6 maximum flow algorithms. However, unlike the Ford-Fulkerson or
7 Edmonds-Karp algorithms, it cannot take advantage of the fact
8 if max flow itself has a small magnitude.
9
11
12 =~=~=~=~= Sample Input =~=~=~=~=
13 6 8
14 0 1 3
15 0 2 3
16 1 2 2
17 1 3 3
18 2 4 2
19 3 4 1
20 3 5 2
21 4 5 3
22 0 5
23
24 =~=~=~=~= Sample Output =~=~=~=~=
25 5
26
27 */
28
32
33 const int MAXN = 100, INF = 0x3F3F3F3F;
34 int cap[MAXN][MAXN], f[MAXN][MAXN];
35
36 int push_relabel(int nodes, int source, int sink) {
37 int e[nodes], h[nodes], maxh[nodes];
38 fill(e, e + nodes, 0);
39 fill(h, h + nodes, 0);
40 fill(maxh, maxh + nodes, 0);
42 fill(f[i], f[i] + nodes, 0);
43 h[source] = nodes - 1;
44 for (int i = 0; i < nodes; i++) {
45 f[source][i] = cap[source][i];
46 f[i][source] = -f[source][i];
47 e[i] = cap[source][i];
48 }
49 int sz = 0;
50 for (;;) {
51 if (sz == 0) {
53 if (i != source && i != sink && e[i] > 0) {
54 if (sz != 0 && h[i] > h[maxh[0]]) sz = 0;
55 maxh[sz++] = i;
56 }
57 }
58 if (sz == 0) break;
59 while (sz != 0) {
60 int i = maxh[sz - 1];
61 bool pushed = false;
62 for (int j = 0; j < nodes && e[i] != 0; j++) {
63 if (h[i] == h[j] + 1 && cap[i][j] - f[i][j] > 0) {
64 int df = min(cap[i][j] - f[i][j], e[i]);
65 f[i][j] += df;
66 f[j][i] -= df;
67 e[i] -= df;
68 e[j] += df;
69 if (e[i] == 0) sz--;
70 pushed = true;
71 }
72 }
73 if (!pushed) {
74 h[i] = INF;
76 if (h[i] > h[j] + 1 && cap[i][j] - f[i][j] > 0)

77 h[i] = h[j] + 1;
78 if (h[i] > h[maxh[0]]) {
79 sz = 0;
80 break;
81 }
82 }
83 }
84 }
87 max_flow += f[source][i];
88 return max_flow;
89 }
90
91 int main() {
94 for (int i = 0; i < edges; i++) {
96 cap[u][v] = capacity;
97 }
99 cout << push_relabel(nodes, source, sink) << "\n";
100 return 0;
101 }
2.6 Backtracking
2.6.1 Max Clique (Bron-Kerbosch Algorithm)
1 /*
2
3 Description: Given an undirected graph, determine a subset of
4 the graphs vertices such that every pair of vertices in the
5 subset are connected by an edge, and that the subset is as
6 large as possible. For the weighted version, each vertex is
7 assigned a weight and the objective is to find the clique in
8 the graph that has maximum total weight.
9
10 Complexity: O(3^(V/3)) where V is the number of vertices.
11
12 =~=~=~=~= Sample Input =~=~=~=~=
13 5 8
14 0 1
15 0 2
16 0 3
17 1 2
18 1 3
19 2 3
20 3 4
21 4 2
22 10 20 30 40 50
23
24 =~=~=~=~= Sample Output =~=~=~=~=
25 Max unweighted clique: 4
2.6. Backtracking 69
26 Max weighted clique: 120

27
28 */
29
31 #include <bitset>
35
37 typedef bitset<MAXN> bits;
38 typedef unsigned long long ull;
39
40 int w[MAXN];
41 bool adj[MAXN][MAXN];
42
43 int rec(int nodes, bits & curr, bits & pool, bits & excl) {
44 if (pool.none() && excl.none()) return curr.count();
45 int ans = 0, u = 0;
46 for (int v = 0; v < nodes; v++)
47 if (pool[v] || excl[v]) u = v;
48 for (int v = 0; v < nodes; v++) {
49 if (!pool[v] || adj[u][v]) continue;
50 bits ncurr, npool, nexcl;
51 for (int i = 0; i < nodes; i++) ncurr[i] = curr[i];
52 ncurr[v] = true;
53 for (int j = 0; j < nodes; j++) {
54 npool[j] = pool[j] && adj[v][j];
55 nexcl[j] = excl[j] && adj[v][j];
56 }
57 ans = max(ans, rec(nodes, ncurr, npool, nexcl));
58 pool[v] = false;
59 excl[v] = true;
60 }
61 return ans;
62 }
63
64 int bron_kerbosch(int nodes) {
65 bits curr, excl, pool;
66 pool.flip();
67 return rec(nodes, curr, pool, excl);
68 }
69
70 //This is a fast implementation using bitmasks.
71 //Precondition: the number of nodes must be less than 64.
72 int bron_kerbosch_weighted(int nodes, ull g[], ull curr, ull pool, ull excl) {
73 if (pool == 0 && excl == 0) {
74 int res = 0, u = __builtin_ctzll(curr);
75 while (u < nodes) {
76 res += w[u];
77 u += __builtin_ctzll(curr >> (u + 1)) + 1;
78 }
79 return res;
80 }
81 if (pool == 0) return -1;
82 int res = -1, pivot = __builtin_ctzll(pool | excl);
83 ull z = pool & ~g[pivot];
84 int u = __builtin_ctzll(z);
85 while (u < nodes) {

86 res = max(res, bron_kerbosch_weighted(nodes, g, curr | (1LL << u),
87 pool & g[u], excl & g[u]));
88 pool ^= 1LL << u;
89 excl |= 1LL << u;
90 u += __builtin_ctzll(z >> (u + 1)) + 1;
91 }
92 return res;
93 }
94
95 int bron_kerbosch_weighted(int nodes) {
96 ull g[nodes];
97 for (int i = 0; i < nodes; i++) {
98 g[i] = 0;
100 if (adj[i][j]) g[i] |= 1LL << j;
101 }
102 return bron_kerbosch_weighted(nodes, g, 0, (1LL << nodes) - 1, 0);
103 }
104
105 int main() {
108 for (int i = 0; i < edges; i++) {
109 cin >> u >> v;
110 adj[u][v] = adj[v][u] = true;
111 }
112 for (int i = 0; i < nodes; i++) cin >> w[i];
113 cout << "Max unweighted clique: ";
114 cout << bron_kerbosch(nodes) << "\n";
115 cout << "Max weighted clique: ";
116 cout << bron_kerbosch_weighted(nodes) << "\n";
117 return 0;
118 }
2.6.2 Graph Coloring
1 /*
2
3 Description: Given an undirected graph, assign colors to each
4 of the vertices such that no pair of adjacent vertices have the
5 same color. Furthermore, do so using the minimum # of colors.
6
7 Complexity: Exponential on the number of vertices. The exact
8 running time is difficult to calculate due to several pruning
9 optimizations used here.
10
11 =~=~=~=~= Sample Input =~=~=~=~=
12 5 7
13 0 1
14 0 4
15 1 3
16 1 4
17 2 3
18 2 4
19 3 4
20
2.6. Backtracking 71
21 =~=~=~=~= Sample Output =~=~=~=~=

22 Colored using 3 color(s). The colorings are:
23 Color 1: 0 3
24 Color 2: 1 2
25 Color 3: 4
26
27 */
28
33
35 int cols[MAXN], adj[MAXN][MAXN];
36 int id[MAXN + 1], deg[MAXN + 1];
37 int min_cols, best_cols[MAXN];
38
39 void dfs(int from, int to, int cur, int used_cols) {
40 if (used_cols >= min_cols) return;
41 if (cur == to) {
42 for (int i = from; i < to; i++)
43 best_cols[id[i]] = cols[i];
44 min_cols = used_cols;
45 return;
46 }
47 vector<bool> used(used_cols + 1);
48 for (int i = 0; i < cur; i++)
49 if (adj[id[cur]][id[i]]) used[cols[i]] = true;
50 for (int i = 0; i <= used_cols; i++) {
51 if (!used[i]) {
52 int tmp = cols[cur];
53 cols[cur] = i;
54 dfs(from, to, cur + 1, max(used_cols, i + 1));
55 cols[cur] = tmp;
56 }
57 }
58 }
59
60 int color_graph(int nodes) {
61 for (int i = 0; i <= nodes; i++) {
62 id[i] = i;
63 deg[i] = 0;
64 }
65 int res = 1;
66 for (int from = 0, to = 1; to <= nodes; to++) {
67 int best = to;
68 for (int i = to; i < nodes; i++) {
69 if (adj[id[to - 1]][id[i]]) deg[id[i]]++;
70 if (deg[id[best]] < deg[id[i]]) best = i;
71 }
72 int tmp = id[to];
73 id[to] = id[best];
74 id[best] = tmp;
75 if (deg[id[to]] == 0) {
76 min_cols = nodes + 1;
77 fill(cols, cols + nodes, 0);
78 dfs(from, to, from, 0);
79 from = to;
80 res = max(res, min_cols);

81 }
82 }
83 return res;
84 }
85
86 int main() {
89 for (int i = 0; i < edges; i++) {
90 cin >> u >> v;
91 adj[u][v] = adj[v][u] = true;
92 }
93 cout << "Colored using " << color_graph(nodes);
94 cout << " color(s). The colorings are:\n";
95 for (int i = 0; i < min_cols; i++) {
96 cout << "Color " << i + 1 << ":";
98 if (best_cols[j] == i) cout << " " << j;
99 cout << "\n";
100 }
101 return 0;
102 }
2.7 Maximum Matching
2.7.1 Maximum Bipartite Matching (Kuhns Algorithm)
1 /*
2
3 Description: Given two sets of vertices A = {0, 1, ..., n1}
4 and B = {0, 1, ..., n2} as well as a set of edges E mapping
5 nodes from set A to set B, determine the largest possible
6 subset of E such that no pair of edges in the subset share
7 a common vertex. Precondition: n2 >= n1.
8
9 Complexity: O(V*E) on the number of vertices and edges.
10
11 =~=~=~=~= Sample Input =~=~=~=~=
12 3 4 6
13 0 1
14 1 0
15 1 1
16 1 2
17 2 2
18 2 3
19
20 =~=~=~=~= Sample Output =~=~=~=~=
21 Matched 3 pairs. Matchings are:
22 1 0
23 0 1
24 2 2
25
26 */
27
2.7. Maximum Matching 73
32
34 int match[MAXN];
37
38 bool dfs(int u) {
39 vis[u] = true;
41 int v = match[adj[u][j]];
42 if (v == -1 || (!vis[v] && dfs(v))) {
43 match[adj[u][j]] = u;
44 return true;
45 }
46 }
47 return false;
48 }
49
50 int kuhn(int n1, int n2) {
52 fill(match, match + n2, -1);
53 int matches = 0;
54 for (int i = 0; i < n1; i++) {
55 for (int j = 0; j < n1; j++) vis[j] = 0;
56 if (dfs(i)) matches++;
57 }
58 return matches;
59 }
60
61 int main() {
62 int n1, n2, edges, u, v;
63 cin >> n1 >> n2 >> edges;
64 for (int i = 0; i < edges; i++) {
65 cin >> u >> v;
67 }
68 cout << "Matched " << kuhn(n1, n2);
69 cout << " pair(s). Matchings are:\n";
70 for (int i = 0; i < n2; i++) {
71 if (match[i] == -1) continue;
72 cout << match[i] << " " << i << "\n";
73 }
74 return 0;
75 }
2.7.2 Maximum Bipartite Matching (Hopcroft-Karp Algorithm)
1 /*
2
3 Description: Given two sets of vertices A = {0, 1, ..., n1}
4 and B = {0, 1, ..., n2} as well as a set of edges E mapping
5 nodes from set A to set B, determine the largest possible
6 subset of E such that no pair of edges in the subset share
7 a common vertex. Precondition: n2 >= n1.
8
9 Complexity: O(E sqrt V) on the number of edges and vertices.
10
11 =~=~=~=~= Sample Input =~=~=~=~=
12 3 4 6
13 0 1
14 1 0
15 1 1
16 1 2
17 2 2
18 2 3
19
20 =~=~=~=~= Sample Output =~=~=~=~=
21 Matched 3 pairs. Matchings are:
22 1 0
23 0 1
24 2 2
25
26 */
27
32
34 int match[MAXN], dist[MAXN];
35 vector<bool> used(MAXN), vis(MAXN);
37
38 void bfs(int n1, int n2) {
39 fill(dist, dist + n1, -1);
40 int q[n2], qb = 0;
41 for (int u = 0; u < n1; ++u) {
42 if (!used[u]) {
43 q[qb++] = u;
44 dist[u] = 0;
45 }
46 }
47 for (int i = 0; i < qb; i++) {
48 int u = q[i];
51 if (v >= 0 && dist[v] < 0) {
52 dist[v] = dist[u] + 1;
53 q[qb++] = v;
54 }
55 }
56 }
57 }
58
59 bool dfs(int u) {
60 vis[u] = true;
63 if (v < 0 || (!vis[v] && dist[v] == dist[u] + 1 && dfs(v))) {
64 match[adj[u][j]] = u;
65 used[u] = true;
66 return true;
2.7. Maximum Matching 75
67 }
68 }
69 return false;
70 }
71
72 int hopcroft_karp(int n1, int n2) {
73 fill(match, match + n2, -1);
74 fill(used.begin(), used.end(), false);
75 int res = 0;
76 for (;;) {
77 bfs(n1, n2);
79 int f = 0;
80 for (int u = 0; u < n1; ++u)
81 if (!used[u] && dfs(u)) f++;
82 if (!f) return res;
83 res += f;
84 }
85 return res;
86 }
87
88 int main() {
89 int n1, n2, edges, u, v;
90 cin >> n1 >> n2 >> edges;
91 for (int i = 0; i < edges; i++) {
92 cin >> u >> v;
94 }
95 cout << "Matched " << hopcroft_karp(n1, n2);
97 for (int i = 0; i < n2; i++) {
98 if (match[i] == -1) continue;
99 cout << match[i] << " " << i << "\n";
100 }
101 return 0;
102 }
2.7.3 Maximum Graph Matching (Edmondss Algorithm)
1 /*
2
3 Description: Given a general directed graph, determine a maximal
4 subset of the edges such that no vertex is repeated in the subset.
5
7
8 =~=~=~=~= Sample Input =~=~=~=~=
9 4 8
10 0 1
11 1 0
12 1 2
13 2 1
14 2 3
15 3 2
16 3 0
17 0 3
18
19 =~=~=~=~= Sample Output =~=~=~=~=

20 Matched 2 pair(s). Matchings are:
21 0 1
22 2 3
23
24 */
25
29
31 int p[MAXN], base[MAXN], match[MAXN];
33
34 int lca(int nodes, int a, int b) {
35 vector<bool> used(nodes);
36 for (;;) {
37 a = base[a];
38 used[a] = true;
39 if (match[a] == -1) break;
40 a = p[match[a]];
41 }
42 for (;;) {
43 b = base[b];
44 if (used[b]) return b;
45 b = p[match[b]];
46 }
47 }
48
49 void mark_path(vector<bool> & blossom, int v, int b, int children) {
50 for (; base[v] != b; v = p[match[v]]) {
51 blossom[base[v]] = blossom[base[match[v]]] = true;
52 p[v] = children;
53 children = match[v];
54 }
55 }
56
57 int find_path(int nodes, int root) {
58 vector<bool> used(nodes);
59 for (int i = 0; i < nodes; ++i) {
60 p[i] = -1;
61 base[i] = i;
62 }
63 used[root] = true;
64 int q[nodes], qh = 0, qt = 0;
65 q[qt++] = root;
66 while (qh < qt) {
67 int v = q[qh++];
68 for (int j = 0, to; j < (int)adj[v].size(); j++) {
69 to = adj[v][j];
70 if (base[v] == base[to] || match[v] == to) continue;
71 if (to == root || (match[to] != -1 && p[match[to]] != -1)) {
72 int curbase = lca(nodes, v, to);
73 vector<bool> blossom(nodes);
74 mark_path(blossom, v, curbase, to);
75 mark_path(blossom, to, curbase, v);
77 if (blossom[base[i]]) {
2.8. Hamiltonian Path and Cycle 77
78 base[i] = curbase;
79 if (!used[i]) {
80 used[i] = true;
81 q[qt++] = i;
82 }
83 }
84 } else if (p[to] == -1) {
85 p[to] = v;
86 if (match[to] == -1) return to;
87 to = match[to];
88 used[to] = true;
89 q[qt++] = to;
90 }
91 }
92 }
93 return -1;
94 }
95
96 int edmonds(int nodes) {
97 for (int i = 0; i < nodes; i++) match[i] = -1;
98 for (int i = 0; i < nodes; i++) {
99 if (match[i] == -1) {
100 int v, pv, ppv;
101 for (v = find_path(nodes, i); v != -1; v = ppv) {
102 ppv = match[pv = p[v]];
103 match[v] = pv;
104 match[pv] = v;
105 }
106 }
107 }
108 int matches = 0;
110 if (match[i] != -1) matches++;
111 return matches / 2;
112 }
113
114 int main() {
117 for (int i = 0; i < edges; i++) {
118 cin >> u >> v;
120 }
121 cout << "Matched " << edmonds(nodes);
123 for (int i = 0; i < nodes; i++) {
124 if (match[i] != -1 && i < match[i])
125 cout << i << " " << match[i] << "\n";
126 }
127 return 0;
128 }
2.8 Hamiltonian Path and Cycle
2.8.1 Shortest Hamiltonian Cycle (Travelling Salesman)

1 /*
2
3 Description: Given a weighted, directed graph, the shortest
4 hamiltonian cycle is a cycle of minimum distance that visits
5 each vertex exactly once and returns to the original vertex.
6 This is also known as the traveling salesman problem (TSP).
7 Since this is a bitmasking solution with 32-bit integers,
8 the number of vertices must be less than 32.
9
10 Complexity: O(2^V * V^2) on the number of vertices.
11
12 =~=~=~=~= Sample Input =~=~=~=~=
13 5 10
14 0 1 1
15 0 2 10
16 0 3 1
17 0 4 10
18 1 2 10
19 1 3 10
20 1 4 1
21 2 3 1
22 2 4 1
23 3 4 10
24
25 =~=~=~=~= Sample Output =~=~=~=~=
26 The shortest hamiltonian cycle has length 5.
27 Take the path: 0->3->2->4->1->0
28
29 */
30
34
36 int adj[MAXN][MAXN], order[MAXN];
37
38 int shortest_hamiltonian_cycle(int nodes) {
39 int dp[1 << nodes][nodes];
40 for (int i = 0; i < (1 << nodes); i++)
41 fill(dp[i], dp[i] + nodes, INF);
42 dp[1][0] = 0;
43 for (int mask = 1; mask < (1 << nodes); mask += 2) {
45 if ((mask & 1 << i) != 0)
47 if ((mask & 1 << j) != 0)
48 dp[mask][i] = min(dp[mask][i], dp[mask ^ (1 << i)][j] + adj[j][i]);
49 }
50 int res = INF + INF;
52 res = min(res, dp[(1 << nodes) - 1][i] + adj[i][0]);
53 int cur = (1 << nodes) - 1, last = 0;
54 for (int i = nodes - 1; i >= 1; i--) {
55 int bj = -1;
56 for (int j = 1; j < nodes; j++) {
57 if ((cur & 1 << j) != 0 && (bj == -1 ||
58 dp[cur][bj] + adj[bj][last] > dp[cur][j] + adj[j][last])) {
59 bj = j;
2.8. Hamiltonian Path and Cycle 79
60 }
61 }
62 order[i] = bj;
63 cur ^= 1 << bj;
64 last = bj;
65 }
66 return res;
67 }
68
69 int main() {
72 for (int i = 0; i < edges; i++) {
73 cin >> u >> v >> w;
74 adj[u][v] = adj[v][u] = w; //only set adj[u][v] if directed edges
75 }
76 cout << "The shortest hamiltonian cycle has length ";
77 cout << shortest_hamiltonian_cycle(nodes) << ".\n";
79 for (int i = 0; i < nodes; i++) cout << order[i] << "->";
80 cout << order[0] << "\n";
81 return 0;
82 }
2.8.2 Shortest Hamiltonian Path
1 /*
2
3 Description: Given a weighted, directed graph, the shortest
4 hamiltonian path is a path of minimum distance that visits
5 each vertex exactly once. Unlike the travelling salesman
6 problem, we dont have to return to the starting vertex.
7 Since this is a bitmasking solution with 32-bit integers,
8 the number of vertices must be less than 32.
9
10 Complexity: O(2^V * V^2) on the number of vertices.
11
12 =~=~=~=~= Sample Input =~=~=~=~=
13 3 6
14 0 1 1
15 0 2 1
16 1 0 7
17 1 2 2
18 2 0 3
19 2 1 5
20
21 =~=~=~=~= Sample Output =~=~=~=~=
22 The shortest hamiltonian path has length 3.
23 Take the path: 0->1->2
24
25 */
26
30
32
33 int adj[MAXN][MAXN], order[MAXN];
34
35 int shortest_hamiltonian_path(int nodes) {
36 int dp[1 << nodes][nodes];
37 for (int i = 0; i < (1 << nodes); i++)
38 fill(dp[i], dp[i] + nodes, INF);
39 for (int i = 0; i < nodes; i++) dp[1 << i][i] = 0;
40 for (int mask = 1; mask < (1 << nodes); mask += 2) {
42 if ((mask & 1 << i) != 0)
44 if ((mask & 1 << j) != 0)
45 dp[mask][i] = min(dp[mask][i], dp[mask ^ (1 << i)][j] + adj[j][i]);
46 }
47 int res = INF + INF;
49 res = min(res, dp[(1 << nodes) - 1][i]);
50 int cur = (1 << nodes) - 1, last = -1;
51 for (int i = nodes - 1; i >= 0; i--) {
52 int bj = -1;
53 for (int j = 0; j < nodes; j++) {
54 if ((cur & 1 << j) != 0 && (bj == -1 ||
55 dp[cur][bj] + (last == -1 ? 0 : adj[bj][last]) >
56 dp[cur][j] + (last == -1 ? 0 : adj[j][last]))) {
57 bj = j;
58 }
59 }
60 order[i] = bj;
61 cur ^= 1 << bj;
62 last = bj;
63 }
64 return res;
65 }
66
67 int main() {
70 for (int i = 0; i < edges; i++) {
71 cin >> u >> v >> w;
72 adj[u][v] = w;
73 }
74 cout << "The shortest hamiltonian path has length ";
75 cout << shortest_hamiltonian_path(nodes) << ".\n";
76 cout << "Take the path: " << order[0];
77 for (int i = 1; i < nodes; i++) cout << "->" << order[i];
78 return 0;
79 }
Chapter 3
Data Structures
3.1 Disjoint Sets
3.1.1 Disjoint Set Forest (Simple)
1 /*
2
3 Description: This data structure dynamically keeps track
4 of items partitioned into non-overlapping sets (a disjoint
5 set forest). It is also known as a union-find data structure.
6
7 Time Complexity: Every function below is O(a(N)) amortized
8 on the number of items in the set due to the optimizations
9 of union by rank and path compression. Here, a(N) is the
10 extremely slow growing inverse of the Ackermann function.
11 For all practical values of n, a(n) is less than 5.
12
13 Space Complexity: O(N) total.
14
15 */
16
18 int num_sets = 0, root[MAXN+1], rank[MAXN+1];
19
21 if (root[x] != x) root[x] = find_root(root[x]);
22 return root[x];
23 }
24
25 void make_set(int x) {
26 root[x] = x;
27 rank[x] = 0;
28 num_sets++;
29 }
30
31 bool is_united(int x, int y) {
32 return find_root(x) == find_root(y);
33 }
34
35 void unite(int x, int y) {
81
82 Chapter 3. Data Structures
36 int X = find_root(x), Y = find_root(y);

37 if (X == Y) return;
38 num_sets--;
39 if (rank[X] < rank[Y]) root[X] = Y;
40 else if (rank[X] > rank[Y]) root[Y] = X;
41 else rank[root[Y] = X]++;
42 }
43
45
49
50 int main() {
51 for (char c = a; c <= g; c++) make_set(c);
52 unite(a, b);
53 unite(b, f);
54 unite(d, e);
55 unite(e, g);
56 assert(num_sets == 3);
57 assert(is_united(a, b));
58 assert(!is_united(a, c));
59 assert(!is_united(b, g));
60 assert(is_united(d, g));
61 return 0;
62 }
3.1.2 Disjoint Set Forest
1 /*
2
3 Description: This data structure dynamically keeps track
4 of items partitioned into non-overlapping sets (a disjoint
5 set forest). It is also known as a union-find data structure.
6 This particular templatized version employs an std::map for
7 built in storage and coordinate compression. That is, the
8 magnitude of values inserted is not limited.
9
10 Time Complexity: make_set(), unite() and is_united() are
11 O(a(N) + log N) = O(log N) on the number of elements in the
12 disjoint set forest. get_all_sets() is O(N). find() is is
13 O(a(N)) amortized on the number of items in the set due to
14 the optimizations of union by rank and path compression.
15 Here, a(N) is the extremely slow growing inverse of the
16 Ackermann function. For all practical values of n, a(n) is
17 less than 5.
18
19 Space Complexity: O(N) storage and auxiliary.
20
21 =~=~=~=~= Sample Output =~=~=~=~=
22 Elements: 7, Sets: 3
23 [[a,b,f],[c],[d,e,g]]
24
25 */
26
27 #include <map>
3.1. Disjoint Sets 83
29
30 template<class T> class disjoint_set_forest {
31 int num_elements, num_sets;
32 std::map<T, int> ID;
33 std::vector<int> root, rank;
34
36 if (root[x] != x) root[x] = find_root(root[x]);
37 return root[x];
38 }
39
40 public:
41 disjoint_set_forest(): num_elements(0), num_sets(0) {}
42 int elements() { return num_elements; }
43 int sets() { return num_sets; }
44
45 bool is_united(const T & x, const T & y) {
46 return find_root(ID[x]) == find_root(ID[y]);
47 }
48
49 void make_set(const T & x) {
50 if (ID.find(x) != ID.end()) return;
51 root.push_back(ID[x] = num_elements++);
52 rank.push_back(0);
53 num_sets++;
54 }
55
56 void unite(const T & x, const T & y) {
57 int X = find_root(ID[x]), Y = find_root(ID[y]);
58 if (X == Y) return;
59 num_sets--;
60 if (rank[X] < rank[Y]) root[X] = Y;
61 else if (rank[X] > rank[Y]) root[Y] = X;
62 else rank[root[Y] = X]++;
63 }
64
65 std::vector<std::vector<T> > get_all_sets() {
66 std::map<int, std::vector<T> > tmp;
67 for (typename std::map<T, int>::iterator
68 it = ID.begin(); it != ID.end(); it++)
69 tmp[find_root(it->second)].push_back(it->first);
70 std::vector<std::vector<T> > ret;
71 for (typename std::map<int, std::vector<T> >::
72 iterator it = tmp.begin(); it != tmp.end(); it++)
73 ret.push_back(it->second);
74 return ret;
75 }
76 };
77
79
82
83 int main() {
84 disjoint_set_forest<char> d;
85 for (char c = a; c <= g; c++) d.make_set(c);
86 d.unite(a, b);
87 d.unite(b, f);
88 d.unite(d, e);
89 d.unite(e, g);
90 cout << "Elements: " << d.elements();
91 cout << ", Sets: " << d.sets() << endl;
92 vector<vector<char> > s = d.get_all_sets();
93 cout << "[";
94 for (int i = 0; i < (int)s.size(); i++) {
95 cout << (i > 0 ? ",[" : "[");
96 for (int j = 0; j < (int)s[i].size(); j++)
97 cout << (j > 0 ? "," : "") << s[i][j];
98 cout << "]";
99 }
100 cout << "]\n";
101 return 0;
102 }
3.2 Fenwick Trees
3.2.1 Simple Fenwick Tree
1 /*
2
3 Description: A Fenwick tree (a.k.a. binary indexed tree) is a
4 data structure that allows for the sum of an arbitrary range
5 of values in an array to be dynamically queried in logarithmic
6 time. Note that unlike the object-oriented versions of this
7 data structure found in later sections, the operations here
8 work on 1-based indices (i.e. between 1 and MAXN, inclusive).
9 The array a[] is always synchronized with the bit[] array and
10 should not be modified outside of the functions below.
11
12 Time Complexity: All functions are O(log MAXN).
13 Space Complexity: O(MAXN) storage and auxiliary.
14
15 */
16
18 int a[MAXN + 1], bit[MAXN + 1];
19
20 //a[i] += v
21 void add(int i, int v) {
22 a[i] += v;
23 for (; i <= MAXN; i += i & -i)
24 bit[i] += v;
25 }
26
27 //a[i] = v
28 void set(int i, int v) {
29 int inc = v - a[i];
30 add(i, inc);
31 }
32
33 //returns sum(a[i] for i = 1..hi inclusive)
34 int sum(int hi) {
35 int ret = 0;
3.2. Fenwick Trees 85
36 for (; hi > 0; hi -= hi & -hi)

37 ret += bit[hi];
38 return ret;
39 }
40
41 //returns sum(a[i] for i = lo..hi inclusive)
42 int sum(int lo, int hi) {
43 return sum(hi) - sum(lo - 1);
44 }
45
47
50
51 int main() {
52 for (int i = 1; i <= 5; i++) set(i, i);
53 add(4, -5);
54 cout << "BIT values: ";
55 for (int i = 1; i <= 5; i++)
56 cout << a[i] << " "; //1 2 3 -1 5
57 cout << "\nSum of range [1,3] is ";
58 cout << sum(1, 3) << ".\n"; //6
59 return 0;
60 }
3.2.2 Fenwick Tree
1 /*
2
6 time. All methods below work on 0-based indices (i.e. indices
7 in the range from 0 to size() - 1, inclusive, are valid).
8
9 Time Complexity: add(), set(), and sum() are all O(log N) on
10 the length of the array. size() and at() are O(1).
11
12 Space Complexity: O(N) storage and O(N) auxiliary on size().
13
14 */
15
17
18 template<class T> class fenwick_tree {
19 int len;
20 std::vector<int> a, bit;
21
22 public:
23 fenwick_tree(int n): len(n),
24 a(n + 1), bit(n + 1) {}
25
26 //a[i] += v
27 void add(int i, const T & v) {
28 a[++i] += v;
29 for (; i <= len; i += i & -i)
30 bit[i] += v;
31 }
32
33 //a[i] = v
34 void set(int i, const T & v) {
35 T inc = v - a[i + 1];
36 add(i, inc);
37 }
38
40 T sum(int hi) {
41 T res = 0;
42 for (hi++; hi > 0; hi -= hi & -hi)
43 res += bit[hi];
44 return res;
45 }
46
48 T sum(int lo, int hi) {
50 }
51
52 inline int size() { return len; }
53 inline T at(int i) { return a[i + 1]; }
54 };
55
57
60
61 int main() {
62 int a[] = {10, 1, 2, 3, 4};
63 fenwick_tree<int> t(5);
64 for (int i = 0; i < 5; i++) t.set(i, a[i]);
65 t.add(0, -5);
67 for (int i = 0; i < t.size(); i++)
68 cout << t.at(i) << " "; //5 1 2 3 4
69 cout << "\nSum of range [1, 3] is ";
70 cout << t.sum(1, 3) << ".\n"; //6
71 return 0;
72 }
3.2.3 Fenwick Tree (Point Query)
1 /*
2
6 time. Range updating in a Fenwick tree can only increment
7 values in a range, not set them all to the same value. This
8 version is a very concise version if only point queries are
9 needed. The functions below work on 1-based indices (between
10 1 and MAXN, inclusive).
11
12 Time Complexity: add() and at() are O(log MAXN).

13 Space Complexity: O(N).
14
15 */
16
18 int bit[MAXN + 1];
19
20 //a[i] += v
21 void add(int i, int v) {
22 for (i++; i <= MAXN; i += i & -i) bit[i] += v;
23 }
24
25 //a[i] += v for i = lo..hi, inclusive
26 void add(int lo, int hi, int v) {
27 add(lo, v);
28 add(hi + 1, -v);
29 }
30
31 //returns a[i]
32 int at(int i) {
33 int sum = 0;
34 for (i++; i > 0; i -= i & -i) sum += bit[i];
35 return sum;
36 }
37
39
42
43 int main() {
44 add(1, 2, 5);
45 add(2, 3, 5);
46 add(3, 5, 10);
47 cout << "BIT values: "; //5 10 15 10 10
48 for (int i = 1; i <= 5; i++)
49 cout << at(i) << " ";
50 cout << "\n";
51 return 0;
52 }
3.2.4 Fenwick Tree (Range Update)
1 /*
2
3 Description: Using two arrays, a Fenwick tree can be made to
4 support range updates and range queries simultaneously. However,
5 the range updates can only be used to add an increment to all
6 values in a range, not set them to the same value. The latter
7 problem may be solved using a segment tree + lazy propagation.
8 All methods below operate 0-based indices (i.e. indices in the
9 range from 0 to size() - 1, inclusive, are valid).
10
11 Time Complexity: add(), set(), at(), and sum() are all O(log N)
12 on the length of the array. size() is O(1).
13
14 Space Complexity: O(N) storage and auxiliary.

15
16 =~=~=~=~= Sample Output =~=~=~=~=
17 BIT values: 15 6 7 -5 4
18 Sum of range [0, 4] is 27.
19
20 */
21
23
24 template<class T> class fenwick_tree {
25 int len;
26 std::vector<T> b1, b2;
27
28 T sum(const std::vector<T> & b, int i) {
29 T res = 0;
30 for (; i != 0; i -= i & -i) res += b[i];
31 return res;
32 }
33
34 void add(std::vector<T> & b, int i, const T & v) {
35 for (; i <= len; i += i & -i) b[i] += v;
36 }
37
38 public:
39 fenwick_tree(int n):
40 len(n + 1), b1(n + 2), b2(n + 2) {}
41
42 //a[i] += v for i = lo..hi, inclusive
43 void add(int lo, int hi, const T & v) {
44 lo++, hi++;
45 add(b1, lo, v);
46 add(b1, hi + 1, -v);
47 add(b2, lo, v * (lo - 1));
48 add(b2, hi + 1, -v * hi);
49 }
50
51 //a[i] = v
52 void set(int i, const T & v) { add(i, i, v - at(i)); }
53
55 T sum(int hi) { return sum(b1, hi)*hi - sum(b2, hi); }
56
58 T sum(int lo, int hi) { return sum(hi + 1) - sum(lo); }
59
60 inline int size() const { return len - 1; }
61 inline T at(int i) { return sum(i, i); }
62 };
63
65
68
69 int main() {
70 int a[] = {10, 1, 2, 3, 4};
71 fenwick_tree<int> t(5);
72 for (int i = 0; i < 5; i++) t.set(i, a[i]);
73 t.add(0, 2, 5); //15 6 7 3 4

74 t.set(3, -5); //15 6 7 -5 4
76 for (int i = 0; i < t.size(); i++)
77 cout << t.at(i) << " ";
78 cout << "\nSum of range [0, 4] is ";
79 cout << t.sum(0, 4) << ".\n"; //27
80 return 0;
81 }
3.2.5 Fenwick Tree (Map)
1 /*
2
3 Description: Using two std::maps to represent the Fenwick tree,
4 there no longer needs to be a restriction on the magnitude of
5 queried indices. All indices in range [0, MAXN] are valid.
6
7 Time Complexity: All functions are O(log^2 MAXN). If the
8 std::map is replaced with an std::unordered_map, then the
9 running time will become O(log MAXN) amortized.
10
11 Space Complexity: O(n) on the number of indices accessed.
12
13 */
14
15 #include <map>
16
17 const int MAXN = 1000000000;
18 std::map<int, int> tmul, tadd;
19
20 void _add(int at, int mul, int add) {
21 for (int i = at; i <= MAXN; i = (i | (i+1))) {
22 tmul[i] += mul;
23 tadd[i] += add;
24 }
25 }
26
27 //a[i] += v for all i = lo..hi, inclusive
28 void add(int lo, int hi, int v) {
29 _add(lo, v, -v * (lo - 1));
30 _add(hi, -v, v * hi);
31 }
32
34 int sum(int hi) {
35 int mul = 0, add = 0, start = hi;
36 for (int i = hi; i >= 0; i = (i & (i + 1)) - 1) {
37 if (tmul.find(i) != tmul.end())
38 mul += tmul[i];
39 if (tadd.find(i) != tadd.end())
40 add += tadd[i];
41 }
42 return mul*start + add;
43 }
44
46 int sum(int lo, int hi) {

48 }
49
50 //a[i] = v
51 void set(int i, int v) {
52 add(i, i, v - sum(i, i));
53 }
54
56
59
60 int main() {
61 add(500000001, 500000010, 3);
62 add(500000011, 500000015, 5);
63 set(500000000, 10);
64 cout << sum(500000000, 500000015) << "\n"; //65
65 return 0;
66 }
3.2.6 2D Fenwick Tree
1 /*
2
3 Description: A 2D Fenwick tree is abstractly a 2D array which also
4 supports efficient queries for the sum of values in the rectangle
5 with top-left (1, 1) and bottom-right (r, c). The implementation
6 below has indices accessible in the range [1...xmax][1...ymax].
7
8 Time Complexity: All functions are O(log(xmax)*log(ymax)).
9 Space Complexity: O(xmax*ymax) storage and auxiliary.
10
11 */
12
13 const int xmax = 100, ymax = 100;
14
15 int a[xmax+1][ymax+1], bit[xmax+1][ymax+1];
16
17 //a[x][y] += v
18 void add(int x, int y, int v) {
19 a[x][y] += v;
20 for (int i = x; i <= xmax; i += i & -i)
21 for (int j = y; j <= ymax; j += j & -j)
22 bit[i][j] += v;
23 }
24
25 //a[x][y] = v
26 void set(int x, int y, int v) {
27 int inc = v - a[x][y];
28 add(x, y, inc);
29 }
30
31 //returns sum(data[1..x][1..y], all inclusive)
32 int sum(int x, int y) {
33 int ret = 0;
34 for (int i = x; i > 0; i -= i & -i)

35 for (int j = y; j > 0; j -= j & -j)
36 ret += bit[i][j];
37 return ret;
38 }
39
40 //returns sum(data[x1..x2][y1..y2], all inclusive)
41 int sum(int x1, int y1, int x2, int y2) {
42 return sum(x2, y2) + sum(x1 - 1, y1 - 1) -
43 sum(x1 - 1, y2) - sum(x2, y1 - 1);
44 }
45
47
51
52 int main() {
53 set(1, 1, 5);
54 set(1, 2, 6);
55 set(2, 1, 7);
56 add(3, 3, 9);
57 add(2, 1, -4);
58 /*
59 5 6 0
60 3 0 0
61 0 0 9
62 */
63 cout << "2D BIT values:\n";
64 for (int i = 1; i <= 3; i++) {
65 for (int j = 1; j <= 3; j++)
66 cout << a[i][j] << " ";
67 cout << "\n";
68 }
69 assert(sum(1, 1, 1, 2) == 11);
70 assert(sum(1, 1, 2, 1) == 8);
71 assert(sum(1, 1, 3, 3) == 23);
72 return 0;
73 }
3.2.7 2D Fenwick Tree (Range Update)
1 /*
2
3 Description: A 2D Fenwick tree is abstractly a 2D array which also
4 supports efficient queries for the sum of values in the rectangle
5 with top-left (1, 1) and bottom-right (r, c). The implementation
6 below has indices accessible in the range [0..xmax][0...ymax].
7
8 Time Complexity: All functions are O(log(xmax)*log(ymax)*log(N))
9 where N is the number of indices operated on so far. Use an array
10 or an unordered_map instead of a map to remove the log(N) factor.
11
12 Space Complexity: O(xmax*ymax) storage and auxiliary.
13
14 */
15
16 #include <map>
17 #include <utility>
18
19 template<class T> class fenwick_tree_2d {
20 static const int xmax = 1000000000;
21 static const int ymax = 1000000000;
22
23 std::map<std::pair<int, int>, T> t1, t2, t3, t4;
24
25 template<class Tree>
26 void add(Tree & t, int x, int y, const T & v) {
27 for (int i = x; i <= xmax; i += i & -i)
28 for (int j = y; j <= ymax; j += j & -j)
29 t[std::make_pair(i, j)] += v;
30 }
31
32 //a[i][j] += v for i = [1,x], j = [1,y]
33 void add_pre(int x, int y, const T & v) {
34 add(t1, 1, 1, v);
35
36 add(t1, 1, y + 1, -v);
37 add(t2, 1, y + 1, v * y);
38
39 add(t1, x + 1, 1, -v);
40 add(t3, x + 1, 1, v * x);
41
42 add(t1, x + 1, y + 1, v);
43 add(t2, x + 1, y + 1, -v * y);
44 add(t3, x + 1, y + 1, -v * x);
45 add(t4, x + 1, y + 1, v * x * y);
46 }
47
48 public:
49 //a[i][j] += v for i = [x1,x2], j = [y1,y2]
50 void add(int x1, int y1, int x2, int y2, const T & v) {
51 x1++; y1++; x2++; y2++;
52 add_pre(x2, y2, v);
53 add_pre(x1 - 1, y2, -v);
54 add_pre(x2, y1 - 1, -v);
55 add_pre(x1 - 1, y1 - 1, v);
56 }
57
58 //a[x][y] += v
59 void add(int x, int y, const T & v) {
60 add(x, y, x, y, v);
61 }
62
63 //a[x][y] = v
64 void set(int x, int y, const T & v) {
65 add(x, y, v - at(x, y));
66 }
67
68 //returns sum(a[i][j] for i = [1,x], j = [1,y])
69 T sum(int x, int y) {
70 x++; y++;
71 T s1 = 0, s2 = 0, s3 = 0, s4 = 0;
72 for (int i = x; i > 0; i -= i & -i)
73 for (int j = y; j > 0; j -= j & -j) {
3.3. 1D Range Queries 93
74 s1 += t1[std::make_pair(i, j)];
78 }
79 return s1 * x * y + s2 * x + s3 * y + s4;
80 }
81
82 //returns sum(a[i][j] for i = [x1,x2], j = [y1,y2])
83 T sum(int x1, int y1, int x2, int y2) {
84 return sum(x2, y2) + sum(x1 - 1, y1 - 1) -
85 sum(x1 - 1, y2) - sum(x2, y1 - 1);
86 }
87
88 T at(int x, int y) { return sum(x, y, x, y); }
89 };
90
92
96
97 int main() {
98 fenwick_tree_2d<long long> t;
99 t.set(0, 0, 5);
100 t.set(0, 1, 6);
101 t.set(1, 0, 7);
102 t.add(2, 2, 9);
103 t.add(1, 0, -4);
104 t.add(1, 1, 2, 2, 5);
105 /*
106 5 6 0
107 3 5 5
108 0 5 14
109 */
110 cout << "2D BIT values:\n";
111 for (int i = 0; i < 3; i++) {
112 for (int j = 0; j < 3; j++)
113 cout << t.at(i, j) << " ";
114 cout << "\n";
115 }
116 assert(t.sum(0, 0, 0, 1) == 11);
117 assert(t.sum(0, 0, 1, 0) == 8);
118 assert(t.sum(1, 1, 2, 2) == 29);
119 return 0;
120 }
3.3 1D Range Queries
3.3.1 Simple Segment Tree
1 /*
2
3 Description: A segment tree is a data structure used for
4 solving the dynamic range query problem, which asks to
5 determine the minimum (or maximum) value in any given

6 range in an array that is constantly being updated.
7
8 Time Complexity: Assuming merge() is O(1), build is O(n)
9 while query() and update() are O(log n). If merge() is
10 not O(1), then all running times are multiplied by a
11 factor of whatever complexity merge() runs in.
12
13 Space Complexity: O(MAXN). Note that a segment tree with
14 N leaves requires 2^(log2(N) - 1) = 4*N total nodes.
15
16 Note: This implementation is 0-based, meaning that all
17 indices from 0 to MAXN - 1, inclusive, are accessible.
18
19 =~=~=~=~= Sample Input =~=~=~=~=
20 5 10
21 35232
22 390942
23 649675
24 224475
25 18709
26 Q 1 3
27 M 4 475689
28 Q 2 3
29 Q 1 3
30 Q 1 2
31 Q 3 3
32 Q 2 3
33 M 2 645514
34 M 2 680746
35 Q 0 4
36
37 =~=~=~=~= Sample Output =~=~=~=~=
38 224475
39 224475
40 224475
41 390942
42 224475
43 224475
44 35232
45
46 */
47
49 int N, M, a[MAXN], t[4*MAXN];
50
51 //define your custom nullv and merge() below.
52 //merge(x, nullv) must return x for all x
53
54 const int nullv = 1 << 30;
55
56 inline int merge(int a, int b) { return a < b ? a : b; }
57
58 void build(int n, int lo, int hi) {
59 if (lo == hi) {
60 t[n] = a[lo];
61 return;
62 }
63 build(2*n + 1, lo, (lo + hi)/2);
64 build(2*n + 2, (lo + hi)/2 + 1, hi);

65 t[n] = merge(t[2*n + 1], t[2*n + 2]);
66 }
67
68 //x and y must be manually set before each call to the
69 //functions below. For query(), [x, y] is the range to
70 //be considered. For update(), a[x] is to be set to y.
71 int x, y;
72
73 //merge(a[i] for i = x..y, inclusive)
74 int query(int n, int lo, int hi) {
75 if (hi < x || lo > y) return nullv;
76 if (lo >= x && hi <= y) return t[n];
77 return merge(query(2*n + 1, lo, (lo + hi) / 2),
78 query(2*n + 2, (lo + hi) / 2 + 1, hi));
79 }
80
81 //a[x] = y
82 void update(int n, int lo, int hi) {
83 if (hi < x || lo > x) return;
84 if (lo == hi) {
85 t[n] = y;
86 return;
87 }
88 update(2*n + 1, lo, (lo + hi)/2);
89 update(2*n + 2, (lo + hi)/2 + 1, hi);
90 t[n] = merge(t[2*n + 1], t[2*n + 2]);
91 }
92
93 /*** Example Usage (wcipeg.com/problem/segtree) ***/
94
95 #include <cstdio>
96
97 int main() {
98 scanf("%d%d", &N, &M);
99 for (int i = 0; i < N; i++) scanf("%d", &a[i]);
100 build(0, 0, N - 1);
101 char op;
102 for (int i = 0; i < M; i++) {
103 scanf(" %c%d%d", &op, &x, &y);
104 if (op == Q) {
105 printf("%d\n", query(0, 0, N - 1));
106 } else if (op == M) {
107 update(0, 0, N - 1);
108 }
109 }
110 return 0;
111 }
3.3.2 Segment Tree
1 /*
2
7
8 Time Complexity: Assuming merge() is O(1), query(),
9 update(), and at() are O(log N). size() is O(1). If
10 merge() is not O(1), then all logarithmic running times
11 are multiplied by a factor of the complexity of merge().
12
13 Space Complexity: O(MAXN). Note that a segment tree with
14 N leaves requires 2^(log2(N) - 1) = 4*N total nodes.
15
17 indices from 0 to N - 1, inclusive, are accessible.
18
19 */
20
21 #include <limits> /* std::numeric_limits<T>::min() */
23
24 template<class T> class segment_tree {
25 int len, x, y;
26 std::vector<T> t;
27 T val, *init;
28
29 //define the following yourself. merge(x, nullv) must return x for all x
30 static inline T nullv() { return std::numeric_limits<T>::min(); }
31 static inline T merge(const T & a, const T & b) { return a > b ? a : b; }
32
34 if (lo == hi) {
35 t[n] = init[lo];
36 return;
37 }
38 build(n * 2 + 1, lo, (lo + hi) / 2);
39 build(n * 2 + 2, (lo + hi) / 2 + 1, hi);
40 t[n] = merge(t[n * 2 + 1], t[n * 2 + 2]);
41 }
42
43 void update(int n, int lo, int hi) {
44 if (x < lo || x > hi) return;
45 if (lo == hi) {
46 t[n] = val;
47 return;
48 }
49 update(n * 2 + 1, lo, (lo + hi) / 2);
50 update(n * 2 + 2, (lo + hi) / 2 + 1, hi);
51 t[n] = merge(t[n * 2 + 1], t[n * 2 + 2]);
52 }
53
54 T query(int n, int lo, int hi) {
55 if (hi < x || lo > y) return nullv();
56 if (lo >= x && hi <= y) return t[n];
57 return merge(query(n * 2 + 1, lo, (lo + hi) / 2),
58 query(n * 2 + 2, (lo + hi) / 2 + 1, hi));
59 }
60
61 public:
62 segment_tree(int n, T * a = 0): len(n), t(4 * n, nullv()) {
63 if (a != 0) {
64 init = a;
65 build(0, 0, len - 1);
66 }
67 }
68
69 //a[i] = v
70 void update(int i, const T & v) {
71 x = i;
72 val = v;
73 update(0, 0, len - 1);
74 }
75
76 //merge(a[i] for i = lo..hi, inclusive)
77 T query(int lo, int hi) {
78 x = lo;
79 y = hi;
80 return query(0, 0, len - 1);
81 }
82
84 inline T at(int i) { return query(i, i); }
85 };
86
88
91
92 int main() {
93 int arr[5] = {6, -2, 1, 8, 10};
94 segment_tree<int> T(5, arr);
95 T.update(1, 4);
96 cout << "Array contains:";
97 for (int i = 0; i < T.size(); i++)
98 cout << " " << T.at(i);
99 cout << "\nThe max value in the range [0, 3] is ";
100 cout << T.query(0, 3) << ".\n"; //8
101 return 0;
102 }
3.3.3 Segment Tree (Range Updates)
1 /*
2
7 Lazy propagation is a technique applied to segment trees that
8 allows range updates to be carried out in O(log N) time. The
9 range updating mechanism is less versatile than the one
10 implemented in the next section.
11
12 Time Complexity: Assuming merge() is O(1), query(), update(),
13 at() are O(log(N)). If merge() is not constant time, then all
14 running times are multiplied by whatever complexity the merge
15 function runs in.
16
17 Space Complexity: O(N) on the size of the array. A segment
18 tree for an array of size N needs 2^(log2(N)-1) = 4N nodes.

19
21 indices from 0 to size() - 1, inclusive, are accessible.
22
23 */
24
27
29 int len, x, y;
30 std::vector<T> tree, lazy;
31 T val, *init;
32
33 //define the following yourself. merge(x, nullv) must return x for all valid x
36
38 if (lo == hi) {
39 tree[n] = init[lo];
40 return;
41 }
42 build(n * 2 + 1, lo, (lo + hi) / 2);
43 build(n * 2 + 2, (lo + hi) / 2 + 1, hi);
44 tree[n] = merge(tree[n * 2 + 1], tree[n * 2 + 2]);
45 }
46
47 T query(int n, int lo, int hi) {
48 if (x > hi || y < lo) return nullv();
49 if (x <= lo && hi <= y) {
50 if (lazy[n] == nullv()) return tree[n];
51 return tree[n] = lazy[n];
52 }
53 int lchild = n * 2 + 1, rchild = n * 2 + 2;
54 if (lazy[n] != nullv()) {
55 lazy[lchild] = lazy[rchild] = lazy[n];
56 lazy[n] = nullv();
57 }
58 return merge(query(lchild, lo, (lo + hi)/2),
59 query(rchild, (lo + hi)/2 + 1, hi));
60 }
61
62 void _update(int n, int lo, int hi) {
63 if (x > hi || y < lo) return;
64 if (lo == hi) {
65 tree[n] = val;
66 return;
67 }
68 if (x <= lo && hi <= y) {
69 tree[n] = lazy[n] = merge(lazy[n], val);
70 return;
71 }
72 int lchild = n * 2 + 1, rchild = n * 2 + 2;
73 if (lazy[n] != nullv()) {
74 lazy[lchild] = lazy[rchild] = lazy[n];
75 lazy[n] = nullv();
76 }
77 _update(lchild, lo, (lo + hi) / 2);

78 _update(rchild, (lo + hi) / 2 + 1, hi);
79 tree[n] = merge(tree[lchild], tree[rchild]);
80 }
81
82 public:
83 segment_tree(int n, T * a = 0):
84 len(n), tree(4 * n, nullv()), lazy(4 * n, nullv()) {
85 if (a != 0) {
86 init = a;
87 build(0, 0, len - 1);
88 }
89 }
90
91 void update(int i, const T & v) {
92 x = y = i;
93 val = v;
94 _update(0, 0, len - 1);
95 }
96
97 //a[i] = v for i = lo..hi, inclusive
98 void update(int lo, int hi, const T & v) {
99 x = lo; y = hi;
100 val = v;
101 _update(0, 0, len - 1);
102 }
103
104 //returns merge(a[i] for i = lo..hi, inclusive)
106 x = lo;
107 y = hi;
108 return query(0, 0, len - 1);
109 }
110
113 };
114
116
119
120 int main() {
121 int arr[5] = {6, 4, 1, 8, 10};
122 segment_tree<int> T(5, arr);
123 cout << "Array contains:"; //6 4 1 8 10
124 for (int i = 0; i < T.size(); i++)
125 cout << " " << T.at(i);
126 cout << "\n";
127 T.update(2, 4, 12);
128 cout << "Array contains:"; //6 4 12 12 12
129 for (int i = 0; i < T.size(); i++)
130 cout << " " << T.at(i);
131 cout << "\nThe max value in the range [0, 3] is ";
132 cout << T.query(0, 3) << ".\n"; //12
133 return 0;
134 }
3.3.4 Segment Tree (Fast, Non-recursive)
1 /*
2
7 Lazy propagation is a technique applied to segment trees that
8 allows range updates to be carried out in O(log N) time.
9
10 Time Complexity: Assuming merge() is O(1), query(), update(),
11 at() are O(log(N)). If merge() is not constant time, then all
12 running times are multiplied by whatever complexity the merge
13 function runs in.
14
15 Space Complexity: O(N) on the size of the array.
16
18 indices from 0 to T.size() - 1, inclusive, are accessible.
19
20 */
21
23 #include <stdexcept> /* std::runtime_error */
25
27 //Modify the following 5 methods to implement your custom
28 //operations on the tree. This implements the Add/Max operations.
29 //Operations like Add/Sum, Set/Max can also be implemented.
30 static inline T modify_op(const T & x, const T & y) {
31 return x + y;
32 }
33
34 static inline T query_op(const T & x, const T & y) {
35 return std::max(x, y);
36 }
37
38 static inline T delta_on_segment(const T & delta, int seglen) {
39 if (delta == nullv()) return nullv();
40 //Here you must write a fast equivalent of following slow code:
41 // T result = delta;
42 // for (int i = 1; i < seglen; i++) result = query_op(result, delta);
43 // return result;
44 return delta;
45 }
46
47 static inline T nullv() { return 0; }
48 static inline T initv() { return 0; }
49
50 int length;
51 std::vector<T> value, delta;
52 std::vector<int> len;
53
54 static T join_value_with_delta(const T & val, const T & delta) {
55 return delta == nullv() ? val : modify_op(val, delta);
56 }
57
58 static T join_deltas(const T & delta1, const T & delta2) {
59 if (delta1 == nullv()) return delta2;
60 if (delta2 == nullv()) return delta1;
61 return modify_op(delta1, delta2);
62 }
63
64 T join_value_with_delta(int i) {
65 return join_value_with_delta(value[i], delta_on_segment(delta[i], len[i]));
66 }
67
68 void push_delta(int i) {
69 int d = 0;
70 while ((i >> d) > 0) d++;
71 for (d -= 2; d >= 0; d--) {
72 int x = i >> d;
73 value[x >> 1] = join_value_with_delta(x >> 1);
74 delta[x] = join_deltas(delta[x], delta[x >> 1]);
75 delta[x ^ 1] = join_deltas(delta[x ^ 1], delta[x >> 1]);
76 delta[x >> 1] = nullv();
77 }
78 }
79
80 public:
81 segment_tree(int n):
82 length(n), value(2 * n), delta(2 * n, nullv()), len(2 * n) {
83 std::fill(len.begin() + n, len.end(), 1);
84 for (int i = 0; i < n; i++) value[i + n] = initv();
85 for (int i = 2 * n - 1; i > 1; i -= 2) {
86 value[i >> 1] = query_op(value[i], value[i ^ 1]);
87 len[i >> 1] = len[i] + len[i ^ 1];
88 }
89 }
90
92 if (lo < 0 || hi >= length || lo > hi)
93 throw std::runtime_error("Invalid query range.");
94 push_delta(lo += length);
95 push_delta(hi += length);
96 T res = 0;
97 bool found = false;
98 for (; lo <= hi; lo = (lo + 1) >> 1, hi = (hi - 1) >> 1) {
99 if ((lo & 1) != 0) {
100 res = found ? query_op(res, join_value_with_delta(lo)) :
101 join_value_with_delta(lo);
102 found = true;
103 }
104 if ((hi & 1) == 0) {
105 res = found ? query_op(res, join_value_with_delta(hi)) :
106 join_value_with_delta(hi);
107 found = true;
108 }
109 }
110 if (!found) throw std::runtime_error("Not found.");
111 return res;
112 }
113
114 void modify(int lo, int hi, const T & delta) {
115 if (lo < 0 || hi >= length || lo > hi)
116 throw std::runtime_error("Invalid modify range.");

117 push_delta(lo += length);
118 push_delta(hi += length);
119 int ta = -1, tb = -1;
120 for (; lo <= hi; lo = (lo + 1) >> 1, hi = (hi - 1) >> 1) {
121 if ((lo & 1) != 0) {
122 this->delta[lo] = join_deltas(this->delta[lo], delta);
123 if (ta == -1) ta = lo;
124 }
125 if ((hi & 1) == 0) {
126 this->delta[hi] = join_deltas(this->delta[hi], delta);
127 if (tb == -1) tb = hi;
128 }
129 }
130 for (int i = ta; i > 1; i >>= 1)
131 value[i >> 1] = query_op(join_value_with_delta(i),
132 join_value_with_delta(i ^ 1));
133 for (int i = tb; i > 1; i >>= 1)
134 value[i >> 1] = query_op(join_value_with_delta(i),
135 join_value_with_delta(i ^ 1));
136 }
137
138 inline int size() { return length; }
140 };
141
143
146
147 int main() {
148 segment_tree<int> T(10);
149 T.modify(0, 0, 10);
150 T.modify(1, 1, 5);
151 T.modify(1, 1, 4);
152 T.modify(2, 2, 7);
153 T.modify(3, 3, 8);
154 cout << T.query(0, 3) << "\n"; //10
155 cout << T.query(1, 3) << "\n"; //9
156 T.modify(0, 9, 5);
157 cout << T.query(0, 9) << "\n"; //15
158 cout << "Array contains:"; //15 14 12 13 5 5 5 5 5 5
159 for (int i = 0; i < T.size(); i++)
160 cout << " " << T.at(i);
161 cout << "\n";
162 return 0;
163 }
3.3.5 Implicit Treap
1 /*
2
3 Description: A treap is a self-balancing binary search tree that
4 uses randomization to maintain a low height. In this version,
5 it is used emulate the operations of an std::vector with a tradeoff
6 of increasing the running time of push_back() and at() from O(1) to
7 O(log N), while decreasing the running time of insert() and erase()
8 from O(N) to O(log N). Furthermore, this version supports the same
9 operations as a segment tree with lazy propagation, allowing range
10 updates and queries to be performed in O(log N).
11
12 Time Complexity: Assuming the join functions have constant complexity:
13 insert(), push_back(), erase(), at(), modify(), and query() are all
14 O(log N), while walk() is O(N).
15
16 Space Complexity: O(N) on the size of the array.
17
19 indices from 0 to size() - 1, inclusive, are accessible.
20
21 */
22
23 #include <climits> /* INT_MIN */
24 #include <cstdlib> /* srand(), rand() */
25 #include <ctime> /* time() */
26
27 template<class T> class implicit_treap {
28 //Modify the following 5 functions to implement your custom
31 static inline T join_values(const T & a, const T & b) {
32 return a > b ? a : b;
33 }
34
35 static inline T join_deltas(const T & d1, const T & d2) {
36 return d1 + d2;
37 }
38
39 static inline T join_value_with_delta(const T & v, const T & d, int len) {
40 return v + d;
41 }
42
43 static inline T null_delta() { return 0; }
44 static inline T null_value() { return INT_MIN; }
45
46 struct node_t {
47 static inline int rand32() {
49 }
50
51 T value, subtree_value, delta;
52 int count, priority;
53 node_t *L, *R;
54
55 node_t(const T & val) {
56 value = subtree_value = val;
57 delta = null_delta();
58 count = 1;
59 L = R = 0;
60 priority = rand32();
61 }
62 } *root;
63
64 static int count(node_t * n) {
65 return n ? n->count : 0;
66 }
67
68 static T subtree_value(node_t * n) {
69 return n ? n->subtree_value : null_value();
70 }
71
72 static void update(node_t * n) {
73 if (n == 0) return;
74 n->subtree_value = join_values(join_values(subtree_value(n->L), n->value),
75 subtree_value(n->R));
76 n->count = 1 + count(n->L) + count(n->R);
77 }
78
79 static void apply_delta(node_t * n, const T & delta) {
81 n->delta = join_deltas(n->delta, delta);
82 n->value = join_value_with_delta(n->value, delta, 1);
83 n->subtree_value = join_value_with_delta(n->subtree_value, delta, n->count);
84 }
85
86 static void push_delta(node_t * n) {
88 apply_delta(n->L, n->delta);
89 apply_delta(n->R, n->delta);
90 n->delta = null_delta();
91 }
92
93 static void merge(node_t *& n, node_t * L, node_t * R) {
94 push_delta(L);
95 push_delta(R);
96 if (L == 0) n = R;
97 else if (R == 0) n = L;
98 else if (L->priority < R->priority)
99 merge(L->R, L->R, R), n = L;
100 else
101 merge(R->L, L, R->L), n = R;
102 update(n);
103 }
104
105 static void split(node_t * n, node_t *& L, node_t *& R, int key) {
106 push_delta(n);
107 if (n == 0) L = R = 0;
108 else if (key <= count(n->L))
109 split(n->L, L, n->L, key), R = n;
110 else
111 split(n->R, n->R, R, key - count(n->L) - 1), L = n;
112 update(n);
113 }
114
115 static void insert(node_t *& n, node_t * item, int idx) {
116 push_delta(n);
117 if (n == 0) n = item;
118 else if (item->priority < n->priority)
119 split(n, item->L, item->R, idx), n = item;
120 else if (idx <= count(n->L))
121 insert(n->L, item, idx);
122 else
123 insert(n->R, item, idx - count(n->L) - 1);
124 update(n);
125 }
126
127 static T get(node_t * n, int idx) {
128 push_delta(n);
129 if (idx < count(n->L))
130 return get(n->L, idx);
131 else if (idx > count(n->L))
132 return get(n->R, idx - count(n->L) - 1);
133 return n->value;
134 }
135
136 static void erase(node_t *& n, int idx) {
137 push_delta(n);
138 if (idx == count(n->L)) {
139 delete n;
140 merge(n, n->L, n->R);
141 } else if (idx < count(n->L)) {
142 erase(n->L, idx);
143 } else {
144 erase(n->R, idx - count(n->L) - 1);
145 }
146 }
147
148 template<class UnaryFunction>
149 void walk(node_t * n, UnaryFunction f) {
150 if (n == 0) return;
151 push_delta(n);
152 if (n->L) walk(n->L, f);
153 f(n->value);
154 if (n->R) walk(n->R, f);
155 }
156
157 void clean_up(node_t *& n) {
158 if (n == 0) return;
159 clean_up(n->L);
160 clean_up(n->R);
161 delete n;
162 }
163
164 public:
165 implicit_treap(): root(0) { srand(time(0)); }
166 ~implicit_treap() { clean_up(root); }
167
168 int size() const { return count(root); }
169 bool empty() const { return root == 0; }
170
171 //list.insert(list.begin() + idx, val)
172 void insert(int idx, const T & val) {
173 if (idx < 0 || idx > size()) return;
174 node_t * item = new node_t(val);
175 insert(root, item, idx);
176 }
177
178 void push_back(const T & val) {
179 insert(size(), val);
180 }
181
182 //list.erase(list.begin() + idx)
183 void erase(int idx) {
184 if (idx < 0 || idx >= size()) return;

185 erase(root, idx);
186 }
187
188 T at(int idx) {
189 if (root == 0 || idx < 0 || idx >= size())
190 return null_value();
191 return get(root, idx);
192 }
193
194 template<class UnaryFunction> void walk(UnaryFunction f) {
195 walk(root, f);
196 }
197
198 //for (i = a; i <= b; i++)
199 // list[i] = join_value_with_delta(list[i], delta)
200 void modify(int a, int b, const T & delta) {
201 if (a < 0 || b < 0 || a >= size() || b >= size() || a > b)
202 return;
203 node_t *l1, *r1;
204 split(root, l1, r1, b + 1);
205 node_t *l2, *r2;
206 split(l1, l2, r2, a);
207 apply_delta(r2, delta);
208 node_t *t;
209 merge(t, l2, r2);
210 merge(root, t, r1);
211 }
212
213 //return join_values(list[a..b])
214 T query(int a, int b) {
215 if (a < 0 || b < 0 || a >= size() || b >= size() || a > b)
216 return null_value();
217 node_t *l1, *r1;
218 split(root, l1, r1, b + 1);
219 node_t *l2, *r2;
220 split(l1, l2, r2, a);
221 int res = subtree_value(r2);
222 node_t *t;
223 merge(t, l2, r2);
224 merge(root, t, r1);
225 return res;
226 }
227 };
228
230
233
234 void print(int x) { cout << x << " "; }
235
236 int main() {
237 implicit_treap<int> T;
238 T.push_back(7);
239 T.push_back(8);
240 T.push_back(9);
241 T.insert(1, 5);
242 T.erase(3);
243 T.walk(print); cout << "\n"; //7 5 8

244 T.modify(0, 2, 2);
245 T.walk(print); cout << "\n"; //9 7 10
246 cout << T.at(1) << "\n"; //7
247 cout << T.query(0, 2) << "\n"; //10
248 cout << T.size() << "\n"; //3
249 return 0;
250 }
3.3.6 Sparse Table
1 /*
2
3 Description: The static range minimum query problem can be solved
4 using a sparse table data structure. The RMQ for sub arrays of
5 length 2^k is pre-processed using dynamic programming with formula:
6
7 dp[i][j] = dp[i][j-1], if A[dp[i][j-1]] <= A[dp[i+2^(j-1)-1][j-1]]
8 dp[i+2^(j-1)-1][j-1], otherwise
9
10 where dp[i][j] is the index of the minimum value in the sub array
11 starting at i having length 2^j.
12
13 Time Complexity: O(N log N) for build() and O(1) for min_idx()
14 Space Complexity: O(N log N) on the size of the array.
15
17 indices from 0 to N - 1, inclusive, are valid.
18
19 */
20
22
24 std::vector<int> logtable, dp[MAXN];
25
26 void build(int n, int a[]) {
27 logtable.resize(n + 1);
28 for (int i = 2; i <= n; i++)
29 logtable[i] = logtable[i >> 1] + 1;
30 for (int i = 0; i < n; i++) {
31 dp[i].resize(logtable[n] + 1);
32 dp[i][0] = i;
33 }
34 for (int k = 1; (1 << k) < n; k++) {
35 for (int i = 0; i + (1 << k) <= n; i++) {
36 int x = dp[i][k - 1];
37 int y = dp[i + (1 << (k - 1))][k - 1];
38 dp[i][k] = a[x] <= a[y] ? x : y;
39 }
40 }
41 }
42
43 //returns index of min element in [lo, hi]
44 int min_idx(int a[], int lo, int hi) {
45 int k = logtable[hi - lo];
46 int x = dp[lo][k];
47 int y = dp[hi - (1 << k) + 1][k];

48 return a[x] <= a[y] ? x : y;
49 }
50
52
55
56 int main() {
57 int a[] = {7, -10, 5, 20};
58 build(4, a);
59 cout << min_idx(a, 0, 3) << "\n"; //1
60 return 0;
61 }
3.3.7 Square Root Decomposition
1 /*
2
3 Description: To solve the dynamic range query problem using
4 square root decomposition, we split an array of size N into
5 sqrt(N) buckets, each bucket of size sqrt(N). As a result,
6 each query and update operation will be sqrt(N) in running time.
7
8 Time Complexity: O(N*sqrt(N)) to construct the initial
9 decomposition. After, query() and update() are O(sqrt N)/call.
10
11 Space Complexity: O(N) for the array. O(sqrt N) for the buckets.
12
14 indices from 0 to N - 1, inclusive, are accessible.
15
16 =~=~=~=~= Sample Input =~=~=~=~=
17 5 10
18 35232
19 390942
20 649675
21 224475
22 18709
23 Q 1 3
24 M 4 475689
25 Q 2 3
26 Q 1 3
27 Q 1 2
28 Q 3 3
29 Q 2 3
30 M 2 645514
31 M 2 680746
32 Q 0 4
33
34 =~=~=~=~= Sample Output =~=~=~=~=
35 224475
36 224475
37 224475
38 390942
39 224475
40 224475
41 35232
42
43 */
44
45 #include <cmath> /* sqrt() */
46 #include <limits> /* std::numeric_limits<T>::max() */
48
49 template<class T> class sqrt_decomp {
50 //define the following yourself. merge(x, nullv) must return x for all x
51 static inline T nullv() { return std::numeric_limits<T>::max(); }
52 static inline T merge(const T & a, const T & b) { return a < b ? a : b; }
53
54 int len, blocklen, blocks;
55 std::vector<T> array, block;
56
57 public:
58 sqrt_decomp(int n, T * a = 0): len(n), array(n) {
59 blocklen = (int)sqrt(n);
60 blocks = (n + blocklen - 1) / blocklen;
61 block.resize(blocks);
62 for (int i = 0; i < n; i++)
63 array[i] = a ? a[i] : nullv();
64 for (int i = 0; i < blocks; i++) {
65 int h = (i + 1) * blocklen;
66 if (h > n) h = n;
67 block[i] = nullv();
68 for (int j = i * blocklen; j < h; j++)
69 block[i] = merge(block[i], array[j]);
70 }
71 }
72
73 void update(int idx, const T & val) {
74 array[idx] = val;
75 int b = idx / blocklen;
76 int h = (b + 1) * blocklen;
77 if (h > len) h = len;
78 block[b] = nullv();
79 for (int i = b * blocklen; i < h; i++)
80 block[b] = merge(block[b], array[i]);
81 }
82
84 T ret = nullv();
85 int lb = ceil((double)lo / blocklen);
86 int hb = (hi + 1) / blocklen - 1;
87 if (lb > hb) {
88 for (int i = lo; i <= hi; i++)
89 ret = merge(ret, array[i]);
90 } else {
91 int l = lb * blocklen - 1;
92 int h = (hb + 1) * blocklen;
93 for (int i = lo; i <= l; i++)
95 for (int i = lb; i <= hb; i++)
96 ret = merge(ret, block[i]);
97 for (int i = h; i <= hi; i++)
99 }
100 return ret;
101 }
102
104 inline int at(int idx) { return array[idx]; }
105 };
106
107 /*** Example Usage (wcipeg.com/problem/segtree) ***/
108
110
111 int N, M, A, B, init[100005];
112
113 int main() {
114 scanf("%d%d", &N, &M);
115 for (int i = 0; i < N; i++) scanf("%d", &init[i]);
116 sqrt_decomp<int> a(N, init);
117 char op;
118 for (int i = 0; i < M; i++) {
119 scanf(" %c%d%d", &op, &A, &B);
120 if (op == Q) {
121 printf("%d\n", a.query(A, B));
122 } else if (op == M) {
123 a.update(A, B);
124 }
125 }
126 return 0;
127 }
3.3.8 Interval Tree (Augmented Treap)
1 /*
2
3 Description: An interval tree is structure used to store and efficiently
4 query intervals. An interval may be dynamically inserted, and range
5 queries of [lo, hi] may be performed to have the tree report all intervals
6 that intersect with the queried interval. Augmented trees, described in
7 CLRS (2009, Section 14.3: pp. 348354), is one way to represent these
8 intervals. This implementation uses a treap to maintain balance.
9 See: http://en.wikipedia.org/wiki/Interval_tree#Augmented_tree
10
11 Time Complexity: On average O(log N) for insert() and O(k) for query(),
12 where N is the number of intervals in the tree and k is the number of
13 intervals that will be reported by each query().
14
15 Space Complexity: O(N) on the number of intervals in the tree.
16
17 */
18
19 #include <cstdlib> /* srand() */
21 #include <utility> /* std:pair */
22
23 class interval_tree {
24 typedef std::pair<int, int> interval;
25
26 static bool overlap(const interval & a, const interval & b) {

27 return a.first <= b.second && b.first <= a.second;
28 }
29
30 struct node_t {
33 }
34
35 interval i;
36 int maxh, priority;
37 node_t *L, *R;
38
39 node_t(const interval & i) {
40 this->i = i;
41 maxh = i.second;
42 L = R = 0;
44 }
45
46 void update() {
47 maxh = i.second;
48 if (L != 0 && L->maxh > maxh) maxh = L->maxh;
49 if (R != 0 && R->maxh > maxh) maxh = R->maxh;
50 }
51 } *root;
52
53 static void rotate_l(node_t *& k2) {
54 node_t *k1 = k2->R;
55 k2->R = k1->L;
56 k1->L = k2;
57 k2 = k1;
58 k2->update();
59 k1->update();
60 }
61
62 static void rotate_r(node_t *& k2) {
63 node_t *k1 = k2->L;
64 k2->L = k1->R;
65 k1->R = k2;
66 k2 = k1;
67 k2->update();
68 k1->update();
69 }
70
71 interval i; //temporary
72
73 void insert(node_t *& n) {
74 if (n == 0) { n = new node_t(i); return; }
75 if (i.first < (n->i).first) {
76 insert(n->L);
77 if (n->L->priority < n->priority) rotate_r(n);
78 } else {
79 insert(n->R);
80 if (n->R->priority < n->priority) rotate_l(n);
81 }
82 n->update();
83 }
84
85 template<class ReportFunction>
86 void query(node_t * n, ReportFunction f) {
87 if (n == 0 || n->maxh < i.first) return;
88 if (overlap(n->i, i)) f(n->i.first, n->i.second);
89 query(n->L, f);
90 query(n->R, f);
91 }
92
93 static void clean_up(node_t * n) {
95 clean_up(n->L);
96 clean_up(n->R);
97 delete n;
98 }
99
100 public:
101 interval_tree(): root(0) { srand(time(0)); }
102 ~interval_tree() { clean_up(root); }
103
104 void insert(int lo, int hi) {
105 i = interval(lo, hi);
106 insert(root);
107 }
108
110 void query(int lo, int hi, ReportFunction f) {
111 i = interval(lo, hi);
112 query(root, f);
113 }
114 };
115
117
121
122 void print(int lo, int hi) {
123 cout << "[" << lo << "," << hi << "] ";
124 }
125
126 int cnt;
127 void count(int lo, int hi) { cnt++; }
128
129 int main() {
130 int N = 6;
131 int intv[6][2] = {{15, 20}, {10, 30}, {17, 19}, {5, 20}, {12, 15}, {30, 40}};
132 interval_tree T;
133 for (int i = 0; i < N; i++) {
134 T.insert(intv[i][0], intv[i][1]);
135 }
136 T.query(10, 20, print); cout << "\n"; //[15,20] [10,30] [5,20] [12,15] [17,19]
137 T.query(0, 5, print); cout << "\n"; //[5,20]
138 T.query(25, 45, print); cout << "\n"; //[10,30] [30,40]
139 //check correctness
140 for (int l = 0; l <= 50; l++) {
141 for (int h = l; h <= 50; h++) {
142 cnt = 0;
143 T.query(l, h, count);
144 int cnt2 = 0;

145 for (int i = 0; i < N; i++)
146 if (intv[i][0] <= h && l <= intv[i][1])
147 cnt2++;
148 assert(cnt == cnt2);
149 }
150 }
151 return 0;
152 }
3.4 2D Range Queries
3.4.1 Quadtree (Simple)
1 /*
2
3 Description: A quadtree can be used to dynamically query values
4 of rectangles in a 2D array. In a quadtree, every node has exactly
5 4 children. The following uses a statically allocated array to
6 store the nodes. This is less efficient than a 2D segment tree.
7
8 Time Complexity: For update(), query() and at(): O(log(N*M)) on
9 average and O(sqrt(N*M)) in the worst case, where N is the number
10 of rows and M is the number of columns in the 2D array.
11
12 Space Complexity: O(N*M)
13
14 Note: This implementation is 0-based. Valid indices for
15 all operations are [0..xmax][0..ymax]
16
17 */
18
20
21 const int xmax = 100, ymax = 100;
22 int tree[4 * xmax * ymax];
23 int X, Y, X1, X2, Y1, Y2, V; //temporary value to speed up recursion
24
26 inline int nullv() { return INT_MIN; }
27 inline int merge(int a, int b) { return a > b ? a : b; }
28
29 void update(int n, int x1, int x2, int y1, int y2) {
30 if (X < x1 || X > x2 || Y < y1 || Y > y2) return;
31 if (x1 == x2 && y1 == y2) {
32 tree[n] = V;
33 return;
34 }
35 update(n * 4 + 1, x1, (x1 + x2) / 2, y1, (y1 + y2) / 2);
36 update(n * 4 + 2, x1, (x1 + x2) / 2, (y1 + y2) / 2 + 1, y2);
37 update(n * 4 + 3, (x1 + x2) / 2 + 1, x2, y1, (y1 + y2) / 2);
38 update(n * 4 + 4, (x1 + x2) / 2 + 1, x2, (y1 + y2) / 2 + 1, y2);
39 tree[n] = merge(merge(tree[n * 4 + 1], tree[n * 4 + 2]),
40 merge(tree[n * 4 + 3], tree[n * 4 + 4]));
41 }
42
43 void query(int n, int x1, int x2, int y1, int y2) {
44 if (x1 > X2 || x2 < X1 || y2 < Y1 || y1 > Y2 || merge(tree[n], V) == V)
45 return;
46 if (x1 >= X1 && x2 <= X2 && y1 >= Y1 && y2 <= Y2) {
47 V = merge(tree[n], V);
48 return;
49 }
50 query(n * 4 + 1, x1, (x1 + x2) / 2, y1, (y1 + y2) / 2);
51 query(n * 4 + 2, x1, (x1 + x2) / 2, (y1 + y2) / 2 + 1, y2);
52 query(n * 4 + 3, (x1 + x2) / 2 + 1, x2, y1, (y1 + y2) / 2);
53 query(n * 4 + 4, (x1 + x2) / 2 + 1, x2, (y1 + y2) / 2 + 1, y2);
54 }
55
56 void update(int x, int y, int v) {
57 X = x;
58 Y = y;
59 V = v;
60 update(0, 0, xmax - 1, 0, ymax - 1);
61 }
62
63 int query(int x1, int y1, int x2, int y2) {
64 X1 = x1;
65 X2 = x2;
66 Y1 = y1;
67 Y2 = y2;
68 V = nullv();
69 query(0, 0, xmax - 1, 0, ymax - 1);
70 return V;
71 }
72
74
77
78 int main() {
79 int arr[5][5] = {{1, 2, 3, 4, 5},
80 {5, 4, 3, 2, 1},
81 {6, 7, 8, 0, 0},
82 {0, 1, 2, 3, 4},
83 {5, 9, 9, 1, 2}};
84 for (int r = 0; r < 5; r++)
85 for (int c = 0; c < 5; c++)
86 update(r, c, arr[r][c]);
87 cout << "The maximum value in the rectangle with ";
88 cout << "upper left (0,2) and lower right (3,4) is ";
89 cout << query(0, 2, 3, 4) << ".\n"; //8
90 return 0;
91 }
3.4.2 Quadtree
1 /*
2
3 Description: A quadtree can be used to dynamically query values
4 of rectangles in a 2D array. In a quadtree, every node has exactly
5 4 children. The following uses dynamically allocated memory to
6 store the nodes, which allows arbitrarily large indices to exist

7 without affecting the performance of operations.
8
9 Time Complexity: For update(), query() and at(): O(log(N*M)) on
10 average and O(sqrt(N*M)) in the worst case, where N is the number
11 of rows and M is the number of columns in the 2D array.
12
13 Space Complexity: O(N*M)
14
16 all operations are [0..XMAX][0..YMAX]
17
18 */
19
20 #include <algorithm> /* std::max(), std::min() */
22
23 template<class T> class quadtree {
24 //these can be set to large values without affecting your memory usage!
27
31
32 int X, Y, X1, X2, Y1, Y2; T V; //temp vals for speed
33
34 struct node_t {
35 node_t * child[4];
36 int x1, x2, y1, y2;
37 T value;
38
39 node_t(int x, int y) {
40 x1 = x2 = x;
41 y1 = y2 = y;
42 child[0] = child[1] = child[2] = child[3] = 0;
43 value = nullv();
44 }
45 } *root;
46
47 void update(node_t *& n, int x1, int x2, int y1, int y2) {
48 if (X < x1 || X > x2 || Y < y1 || Y > y2) return;
49 if (n == 0) n = new node_t(X, Y);
50 if (x1 == x2 && y1 == y2) {
51 n->value = V;
52 return;
53 }
54 int xmid = (x1 + x2)/2, ymid = (y1 + y2)/2;
55 update(n->child[0], x1, xmid, y1, ymid);
56 update(n->child[1], xmid + 1, x2, y1, ymid);
57 update(n->child[2], x1, xmid, ymid + 1, y2);
58 update(n->child[3], xmid + 1, x2, ymid + 1, y2);
59 for (int i = 0; i < 4; i++) {
60 if (!n->child[i] || n->child[i]->value == nullv()) continue;
61 n->x1 = std::min(n->x1, n->child[i]->x1);
62 n->x2 = std::max(n->x2, n->child[i]->x2);
63 n->y1 = std::min(n->y1, n->child[i]->y1);
64 n->y2 = std::max(n->y2, n->child[i]->y2);
65 n->value = merge(n->value, n->child[i]->value);

66 }
67 }
68
69 void query(node_t * n) {
70 if (n == 0 || n->x1 > X2 || n->x2 < X1 || n->y2 < Y1 || n->y1 > Y2 ||
71 merge(n->value, V) == V)
72 return;
73 if (n->x1 >= X1 && n->y1 >= Y1 && n->x2 <= X2 && n->y2 <= Y2) {
74 V = merge(V, n->value);
75 return;
76 }
77 for (int i = 0; i < 4; i++) query(n->child[i]);
78 }
79
82 for (int i = 0; i < 4; i++) clean_up(n->child[i]);
83 delete n;
84 }
85
86 public:
87 quadtree() { root = 0; }
88 ~quadtree() { clean_up(root); }
89
90 void update(int x, int y, const T & v) {
91 X = x;
92 Y = y;
93 V = v;
94 update(root, 0, xmax - 1, 0, ymax - 1);
95 }
96
97 T query(int x1, int y1, int x2, int y2) {
98 X1 = x1;
99 X2 = x2;
100 Y1 = y1;
101 Y2 = y2;
102 V = nullv();
103 query(root);
104 return V;
105 }
106
107 T at(int x, int y) {
108 return query(x, y, x, y);
109 }
110 };
111
113
116
117 int main() {
118 int arr[5][5] = {{1, 2, 3, 4, 5},
119 {5, 4, 3, 2, 1},
120 {6, 7, 8, 0, 0},
121 {0, 1, 2, 3, 4},
122 {5, 9, 9, 1, 2}};
123 quadtree<int> T;
124 for (int r = 0; r < 5; r++)

125 for (int c = 0; c < 5; c++)
126 T.update(r, c, arr[r][c]);
129 cout << T.query(0, 2, 3, 4) << ".\n"; //8
130 return 0;
131 }
3.4.3 2D Segment Tree
1 /*
2
3 Description: A quadtree is a segment tree but with 4 children
4 per node, making its running time proportional to the square
5 root of the number of leaves. However, a 2D segment tree is a
6 segment tree of segment trees, making its running time
7 proportional to the log of its size. The following implementation
8 is a highly optimized implementation with features such as
9 coordinate compression and path compression.
10
11 Time Complexity: O(log(xmax)*log(ymax)) for update(), query(),
12 and at() operations. size() is O(1).
13
14 Space Complexity: Left as an exercise for the reader.
15
17 all operations are [0..xmax][0..ymax]
18
19 */
20
22
23 template<class T> class segment_tree_2d {
24 //these can be set to large values without affecting your memory usage!
27
31
32 struct layer2_node {
33 int lo, hi;
34 layer2_node *L, *R;
35 T value;
36 layer2_node(int l, int h) : lo(l), hi(h), L(0), R(0) {}
37 };
38
39 struct layer1_node {
40 layer1_node *L, *R;
41 layer2_node l2;
42 layer1_node() : L(0), R(0), l2(0, ymax) {}
43 } *root;
44
45 void update2(layer2_node * node, int Q, const T & v) {
46 int lo = node->lo, hi = node->hi, mid = (lo + hi)/2;
47 if (lo + 1 == hi) {
48 node->value = v;
49 return;
50 }
51 layer2_node *& tgt = Q < mid ? node->L : node->R;
52 if (tgt == 0) {
53 tgt = new layer2_node(Q, Q + 1);
54 tgt->value = v;
55 } else if (tgt->lo <= Q && Q < tgt->hi) {
56 update2(tgt, Q, v);
57 } else {
58 do {
59 (Q < mid ? hi : lo) = mid;
60 mid = (lo + hi)/2;
61 } while ((Q < mid) == (tgt->lo < mid));
62 layer2_node *nnode = new layer2_node(lo, hi);
63 (tgt->lo < mid ? nnode->L : nnode->R) = tgt;
64 tgt = nnode;
65 update2(nnode, Q, v);
66 }
67 node->value = merge(node->L ? node->L->value : nullv(),
68 node->R ? node->R->value : nullv());
69 }
70
71 T query2(layer2_node * nd, int A, int B) {
72 if (nd == 0 || B <= nd->lo || nd->hi <= A) return nullv();
73 if (A <= nd->lo && nd->hi <= B) return nd->value;
74 return merge(query2(nd->L, A, B), query2(nd->R, A, B));
75 }
76
77 void update1(layer1_node * node, int lo, int hi, int x, int y, const T & v) {
78 if (lo + 1 == hi) update2(&node->l2, y, v);
79 else {
80 int mid = (lo + hi)/2;
81 layer1_node *& nnode = x < mid ? node->L : node->R;
82 (x < mid ? hi : lo) = mid;
83 if (nnode == 0) nnode = new layer1_node();
84 update1(nnode, lo, hi, x, y, v);
85 update2(&node->l2, y, merge(
86 node->L ? query2(&node->L->l2, y, y + 1) : nullv(),
87 node->R ? query2(&node->R->l2, y, y + 1) : nullv())
88 );
89 }
90 }
91
92 T query1(layer1_node * nd, int lo, int hi, int A1, int B1, int A2, int B2) {
93 if (nd == 0 || B1 <= lo || hi <= A1) return nullv();
94 if (A1 <= lo && hi <= B1) return query2(&nd->l2, A2, B2);
95 int mid = (lo + hi) / 2;
96 return merge(query1(nd->L, lo, mid, A1, B1, A2, B2),
97 query1(nd->R, mid, hi, A1, B1, A2, B2));
98 }
99
100 void clean_up2(layer2_node * n) {
101 if (n == 0) return;
102 clean_up2(n->L);
103 clean_up2(n->R);
104 delete n;
105 }
106
107 void clean_up1(layer1_node * n) {
108 if (n == 0) return;
109 clean_up2(n->l2.L);
110 clean_up2(n->l2.R);
111 clean_up1(n->L);
112 clean_up1(n->R);
113 delete n;
114 }
115
116 public:
117 segment_tree_2d() { root = new layer1_node(); }
118 ~segment_tree_2d() { clean_up1(root); }
119
120 void update(int x, int y, const T & v) {
121 update1(root, 0, xmax, x, y, v);
122 }
123
124 T query(int x1, int y1, int x2, int y2) {
125 return query1(root, 0, xmax, x1, x2 + 1, y1, y2 + 1);
126 }
127
128 T at(int x, int y) {
129 return query(x, y, x, y);
130 }
131 };
132
134
137
138 int main() {
139 int arr[5][5] = {{1, 2, 3, 4, 5},
140 {5, 4, 3, 2, 1},
141 {6, 7, 8, 0, 0},
142 {0, 1, 2, 3, 4},
143 {5, 9, 9, 1, 2}};
144 segment_tree_2d<int> T;
145 for (int r = 0; r < 5; r++)
146 for (int c = 0; c < 5; c++)
147 T.update(r, c, arr[r][c]);
150 cout << T.query(0, 2, 3, 4) << ".\n"; //8
151 return 0;
152 }
3.4.4 K-d Tree (2D Range Query)
1 /*
2
3 Description: k-d tree (short for k-dimensional tree) is a space-
4 partitioning data structure for organizing points in a k-
5 dimensional space. The following implementation supports
6 counting the number of points in rectangular ranges after the
7 tree has been build.
8
9 Time Complexity: O(N log N) for build(), where N is the number
10 of points in the tree. count() is O(sqrt N).
11
12 Space Complexity: O(N) on the number of points.
13
14 */
15
16 #include <algorithm> /* nth_element(), max(), min() */
17 #include <climits> /* INT_MIN, INT_MAX */
20
21 class kd_tree {
22 typedef std::pair<int, int> point;
23
24 static inline bool cmp_x(const point & a, const point & b) {
25 return a.first < b.first;
26 }
27
28 static inline bool cmp_y(const point & a, const point & b) {
29 return a.second < b.second;
30 }
31
32 std::vector<int> tx, ty, cnt, minx, miny, maxx, maxy;
33 int x1, y1, x2, y2; //temporary values to speed up recursion
34
35 void build(int lo, int hi, bool div_x, point P[]) {
36 if (lo >= hi) return;
37 int mid = (lo + hi) >> 1;
38 std::nth_element(P + lo, P + mid, P + hi, div_x ? cmp_x : cmp_y);
39 tx[mid] = P[mid].first;
40 ty[mid] = P[mid].second;
41 cnt[mid] = hi - lo;
42 minx[mid] = INT_MAX; miny[mid] = INT_MAX;
43 maxx[mid] = INT_MIN; maxy[mid] = INT_MIN;
44 for (int i = lo; i < hi; i++) {
45 minx[mid] = std::min(minx[mid], P[i].first);
46 maxx[mid] = std::max(maxx[mid], P[i].first);
47 miny[mid] = std::min(miny[mid], P[i].second);
48 maxy[mid] = std::max(maxy[mid], P[i].second);
49 }
50 build(lo, mid, !div_x, P);
51 build(mid + 1, hi, !div_x, P);
52 }
53
54 int count(int lo, int hi) {
55 if (lo >= hi) return 0;
56 int mid = (lo + hi) >> 1;
57 int ax = minx[mid], ay = miny[mid];
58 int bx = maxx[mid], by = maxy[mid];
59 if (ax > x2 || x1 > bx || ay > y2 || y1 > by) return 0;
60 if (x1 <= ax && bx <= x2 && y1 <= ay && by <= y2) return cnt[mid];
61 int res = count(lo, mid) + count(mid + 1, hi);
62 res += (x1 <= tx[mid] && tx[mid] <= x2 && y1 <= ty[mid] && ty[mid] <= y2);
63 return res;
64 }
65
66 public:
67 kd_tree(int n, point P[]): tx(n), ty(n), cnt(n),

68 minx(n), miny(n), maxx(n), maxy(n) {
69 build(0, n, true, P);
70 }
71
72 int count(int x1, int y1, int x2, int y2) {
73 this->x1 = x1;
74 this->y1 = y1;
75 this->x2 = x2;
76 this->y2 = y2;
77 return count(0, tx.size());
78 }
79 };
80
82
85
86 int main() {
87 pair<int, int> P[4];
88 P[0] = make_pair(0, 0);
89 P[1] = make_pair(10, 10);
90 P[2] = make_pair(0, 10);
91 P[3] = make_pair(10, 0);
92 kd_tree t(4, P);
93 assert(t.count(0, 0, 10, 9) == 2);
94 assert(t.count(0, 0, 10, 10) == 4);
95 return 0;
96 }
3.4.5 K-d Tree (Nearest Neighbor)
1 /*
2
3 Description: k-d tree (short for k-dimensional tree) is a space-
4 partitioning data structure for organizing points in a k-
5 dimensional space. The following implementation supports
6 querying the nearest neighboring point to (x, y) in terms of
7 Euclidean distance after the tree has been build. Note that
8 a point is not considered its own neighbour if it already exists
9 in the tree.
10
11 Time Complexity: O(N log N) for build(), where N is the number of
12 points in the tree. nearest_neighbor_id() is O(log(N)) on average.
13
15
16 */
17
18 #include <algorithm> /* nth_element(), max(), min(), swap() */
22
23 class kd_tree {
25
26 static inline bool cmp_x(const point & a, const point & b) {
27 return a.first < b.first;
28 }
29
30 static inline bool cmp_y(const point & a, const point & b) {
32 }
33
34 std::vector<int> tx, ty;
35 std::vector<bool> div_x;
36
37 void build(int lo, int hi, point P[]) {
39 int mid = (lo + hi) >> 1;
40 int minx = INT_MAX, maxx = INT_MIN;
41 int miny = INT_MAX, maxy = INT_MIN;
42 for (int i = lo; i < hi; i++) {
43 minx = std::min(minx, P[i].first);
44 maxx = std::max(maxx, P[i].first);
45 miny = std::min(miny, P[i].second);
46 maxy = std::max(maxy, P[i].second);
47 }
48 div_x[mid] = (maxx - minx) >= (maxy - miny);
49 std::nth_element(P + lo, P + mid, P + hi, div_x[mid] ? cmp_x : cmp_y);
50 tx[mid] = P[mid].first;
51 ty[mid] = P[mid].second;
52 if (lo + 1 == hi) return;
53 build(lo, mid, P);
54 build(mid + 1, hi, P);
55 }
56
57 long long min_dist;
58 int min_dist_id, x, y;
59
60 void nearest_neighbor(int lo, int hi) {
62 int mid = (lo + hi) >> 1;
63 int dx = x - tx[mid], dy = y - ty[mid];
64 long long d = dx*(long long)dx + dy*(long long)dy;
65 if (min_dist > d && d) {
66 min_dist = d;
67 min_dist_id = mid;
68 }
69 if (lo + 1 == hi) return;
70 int delta = div_x[mid] ? dx : dy;
71 long long delta2 = delta*(long long)delta;
72 int l1 = lo, r1 = mid, l2 = mid + 1, r2 = hi;
73 if (delta > 0) std::swap(l1, l2), std::swap(r1, r2);
74 nearest_neighbor(l1, r1);
75 if (delta2 < min_dist) nearest_neighbor(l2, r2);
76 }
77
78 public:
79 kd_tree(int N, point P[]) {
80 tx.resize(N);
81 ty.resize(N);
82 div_x.resize(N);
83 build(0, N, P);
84 }
85
86 int nearest_neighbor_id(int x, int y) {
87 this->x = x; this->y = y;
88 min_dist = LLONG_MAX;
89 nearest_neighbor(0, tx.size());
90 return min_dist_id;
91 }
92 };
93
95
98
99 int main() {
100 pair<int, int> P[3];
101 P[0] = make_pair(0, 2);
102 P[1] = make_pair(0, 3);
103 P[2] = make_pair(-1, 0);
104 kd_tree T(3, P);
105 int res = T.nearest_neighbor_id(0, 0);
106 cout << P[res].first << " " << P[res].second << "\n"; //-1, 0
107 return 0;
108 }
3.4.6 R-Tree (Nearest Segment)
1 /*
2
3 Description: R-trees are tree data structures used for spatial
4 access methods, i.e., for indexing multi-dimensional information
5 such as geographical coordinates, rectangles or polygons. The
6 following implementation supports querying of the nearing line
7 segment to a point after a tree of line segments have been built.
8
9 Time Complexity: O(N log N) for build(), where N is the number of
10 points in the tree. nearest_neighbor_id() is O(log(N)) on average.
11
13
14 */
15
16 #include <algorithm> /* nth_element(), max(), min(), swap() */
17 #include <cfloat> /* DBL_MAX */
20
21 struct segment { int x1, y1, x2, y2; };
22
23 class r_tree {
24
25 static inline bool cmp_x(const segment & a, const segment & b) {
26 return a.x1 + a.x2 < b.x1 + b.x2;
27 }
28
29 static inline bool cmp_y(const segment & a, const segment & b) {
30 return a.y1 + a.y2 < b.y1 + b.y2;

31 }
32
33 std::vector<segment> s;
34 std::vector<int> minx, maxx, miny, maxy;
35
36 void build(int lo, int hi, bool div_x, segment s[]) {
38 int mid = (lo + hi) >> 1;
39 std::nth_element(s + lo, s + mid, s + hi, div_x ? cmp_x : cmp_y);
40 this->s[mid] = s[mid];
41 for (int i = lo; i < hi; i++) {
42 minx[mid] = std::min(minx[mid], std::min(s[i].x1, s[i].x2));
43 miny[mid] = std::min(miny[mid], std::min(s[i].y1, s[i].y2));
44 maxx[mid] = std::max(maxx[mid], std::max(s[i].x1, s[i].x2));
45 maxy[mid] = std::max(maxy[mid], std::max(s[i].y1, s[i].y2));
46 }
47 build(lo, mid, !div_x, s);
48 build(mid + 1, hi, !div_x, s);
49 }
50
51 double min_dist;
52 int min_dist_id, x, y;
53
54 void nearest_neighbor(int lo, int hi, bool div_x) {
56 int mid = (lo + hi) >> 1;
57 double pdist = point_to_segment_squared(x, y, s[mid]);
58 if (min_dist > pdist) {
59 min_dist = pdist;
60 min_dist_id = mid;
61 }
62 long long delta = div_x ? 2*x - s[mid].x1 - s[mid].x2 :
63 2*y - s[mid].y1 - s[mid].y2;
64 if (delta <= 0) {
65 nearest_neighbor(lo, mid, !div_x);
66 if (mid + 1 < hi) {
67 int mid1 = (mid + hi + 1) >> 1;
68 long long dist = div_x ? seg_dist(x, minx[mid1], maxx[mid1]) :
69 seg_dist(y, miny[mid1], maxy[mid1]);
70 if (dist*dist < min_dist) nearest_neighbor(mid + 1, hi, !div_x);
71 }
72 } else {
73 nearest_neighbor(mid + 1, hi, !div_x);
74 if (lo < mid) {
75 int mid1 = (lo + mid) >> 1;
76 long long dist = div_x ? seg_dist(x, minx[mid1], maxx[mid1]) :
77 seg_dist(y, miny[mid1], maxy[mid1]);
78 if (dist*dist < min_dist) nearest_neighbor(lo, mid, !div_x);
79 }
80 }
81 }
82
83 static double point_to_segment_squared(int x, int y, const segment & s) {
84 long long dx = s.x2 - s.x1, dy = s.y2 - s.y1;
85 long long px = x - s.x1, py = y - s.y1;
86 long long square_dist = dx*dx + dy*dy;
87 long long dot_product = dx*px + dy*py;
88 if (dot_product <= 0 || square_dist == 0) return px*px + py*py;
89 if (dot_product >= square_dist)

90 return (px - dx)*(px - dx) + (py - dy)*(py - dy);
91 double q = (double)dot_product/square_dist;
92 return (px - q*dx)*(px - q*dx) + (py - q*dy)*(py - q*dy);
93 }
94
95 static inline int seg_dist(int v, int lo, int hi) {
96 return v <= lo ? lo - v : (v >= hi ? v - hi : 0);
97 }
98
99 public:
100 r_tree(int N, segment s[]) {
101 this->s.resize(N);
102 minx.assign(N, INT_MAX);
103 maxx.assign(N, INT_MIN);
104 miny.assign(N, INT_MAX);
105 maxy.assign(N, INT_MIN);
106 build(0, N, true, s);
107 }
108
109 int nearest_neighbor_id(int x, int y) {
110 min_dist = DBL_MAX;
111 this->x = x; this->y = y;
112 nearest_neighbor(0, s.size(), true);
113 return min_dist_id;
114 }
115 };
116
118
121
122 int main() {
123 segment s[4];
124 s[0] = (segment){0, 0, 0, 4};
125 s[1] = (segment){0, 4, 4, 4};
126 s[2] = (segment){4, 4, 4, 0};
127 s[3] = (segment){4, 0, 0, 0};
128 r_tree t(4, s);
129 int id = t.nearest_neighbor_id(-1, 2);
130 cout << s[id].x1 << " " << s[id].y1 << " " <<
131 s[id].x2 << " " << s[id].y2 << "\n"; //0 0 0 4
132 return 0;
133 }
3.4.7 2D Range Tree
1 /*
2
3 Description: A range tree is an ordered tree data structure to
4 hold a list of points. It allows all points within a given range
5 to be reported efficiently. Specifically, for a given query, a
6 range tree will report *all* points that lie in the given range.
7 Note that the initial array passed to construct the tree will be
8 sorted, and all resulting query reports will pertain to the
9 indices of points in the sorted array.
10
11 Time Complexity: A range tree can build() in O(N log^(d-1)(N))
12 and query() in O(log^d(n) + k), where N is the number of points
13 stored in the tree, d is the dimension of each point and k is the
14 number of points reported by a given query. Thus for this 2D case
15 build() is O(N log N) and query() is O(log^2(N) + k).
16
17 Space Complexity: O(N log^(d-1)(N)) for a d-dimensional range tree.
18 Thus for this 2D case, the space complexity is O(N log N).
19
20 */
21
22 #include <algorithm> /* lower_bound(), merge(), sort() */
25
26 class range_tree_2d {
28
29 std::vector<point> P;
30 std::vector<std::vector<point> > seg;
31
32 static inline bool comp1(const point & a, const point & b) {
34 }
35
36 static inline bool comp2(const point & a, int v) {
37 return a.second < v;
38 }
39
41 if (P[lo].first == P[hi].first) {
42 for (int i = lo; i <= hi; i++)
43 seg[n].push_back(point(i, P[i].second));
44 return;
45 }
46 int l = n * 2 + 1, r = n * 2 + 2;
47 build(l, lo, (lo + hi)/2);
48 build(r, (lo + hi)/2 + 1, hi);
49 seg[n].resize(seg[l].size() + seg[r].size());
50 std::merge(seg[l].begin(), seg[l].end(), seg[r].begin(), seg[r].end(),
51 seg[n].begin(), comp1);
52 }
53
54 int xl, xh, yl, yh;
55
57 void query(int n, int lo, int hi, ReportFunction f) {
58 if (P[hi].first < xl || P[lo].first > xh) return;
59 if (xl <= P[lo].first && P[hi].first <= xh) {
60 if (!seg[n].empty() && yh >= yl) {
61 std::vector<point>::iterator it;
62 it = std::lower_bound(seg[n].begin(), seg[n].end(), yl, comp2);
63 for (; it != seg[n].end(); ++it) {
64 if (it->second > yh) break;
65 f(it->first); //or report P[it->first], the actual point
66 }
67 }
68 } else if (lo != hi) {
3.5. Search Trees and Alternatives 127
69 query(n * 2 + 1, lo, (lo + hi) / 2, f);

70 query(n * 2 + 2, (lo + hi) / 2 + 1, hi, f);
71 }
72 }
73
74 public:
75 range_tree_2d(int n, point init[]): seg(4 *n + 1) {
76 std::sort(init, init + n);
77 P = std::vector<point>(init, init + n);
78 build(0, 0, n - 1);
79 }
80
82 void query(int x1, int y1, int x2, int y2, ReportFunction f) {
83 xl = x1; xh = x2;
84 yl = y1; yh = y2;
85 query(0, 0, P.size() - 1, f);
86 }
87 };
88
89 /*** Example Usage (wcipeg.com/problem/boxl) ***/
90
94
95 int N, M; bitset<200005> b;
96 pair<int, int> pts[200005];
97 int x1[200005], y1[200005];
98 int x2[200005], y2[200005];
99
100 void mark(int i) {
101 b[i] = true;
102 }
103
104 int main() {
105 scanf("%d%d", &N, &M);
106 for (int i = 0; i < N; i++)
107 scanf("%d%d%d%d", x1 + i, y1 + i, x2 + i, y2 + i);
108 for (int i = 0; i < M; i++)
109 scanf("%d%d", &pts[i].first, &pts[i].second);
110 range_tree_2d t(M, pts);
111 for (int i = 0; i < N; i++)
112 t.query(x1[i], y1[i], x2[i], y2[i], mark);
113 printf("%d\n", b.count());
114 return 0;
115 }
3.5 Search Trees and Alternatives
3.5.1 Binary Search Tree
1 /*
2
3 Description: A binary search tree (BST) is a node-based binary tree data
4 structure where the left sub-tree of every node has keys less than the
5 nodes key and the right sub-tree of every node has keys greater than the
6 nodes key. A BST may be come degenerate like a linked list resulting in
7 an O(N) running time per operation. A self-balancing binary search tree
8 such as a randomized treap prevents the occurence of this known worst case.
9
10 Note: The following implementation is used similar to an std::map. In order
11 to make it behave like an std::set, modify the code to remove the value
12 associated with each node. In order to make it behave like an std::multiset
13 or std::multimap, make appropriate changes with key comparisons (e.g.
14 change (k < n->key) to (k <= n->key) in search conditions).
15
16 Time Complexity: insert(), erase() and find() are O(log(N)) on average,
17 but O(N) at worst if the tree becomes degenerate. Speed can be improved
18 by randomizing insertion order if it doesnt matter. walk() is O(N).
19
20 Space Complexity: O(N) on the number of nodes.
21
22 */
23
24 template<class key_t, class val_t> class binary_search_tree {
25 struct node_t {
26 key_t key;
27 val_t val;
28 node_t *L, *R;
29
30 node_t(const key_t & k, const val_t & v) {
31 key = k;
32 val = v;
33 L = R = 0;
34 }
35 } *root;
36
37 int num_nodes;
38
39 static bool insert(node_t *& n, const key_t & k, const val_t & v) {
40 if (n == 0) {
41 n = new node_t(k, v);
42 return true;
43 }
44 if (k < n->key) return insert(n->L, k, v);
45 if (n->key < k) return insert(n->R, k, v);
46 return false; //already exists
47 }
48
49 static bool erase(node_t *& n, const key_t & key) {
50 if (n == 0) return false;
51 if (key < n->key) return erase(n->L, key);
52 if (n->key < key) return erase(n->R, key);
53 if (n->L == 0) {
54 node_t *temp = n->R;
55 delete n;
56 n = temp;
57 } else if (n->R == 0) {
58 node_t *temp = n->L;
59 delete n;
60 n = temp;
61 } else {
62 node_t *temp = n->R, *parent = 0;
63 while (temp->L != 0) {
64 parent = temp;
65 temp = temp->L;
66 }
67 n->key = temp->key;
68 n->val = temp->val;
69 if (parent != 0)
70 return erase(parent->L, parent->L->key);
71 return erase(n->R, n->R->key);
72 }
73 return true;
74 }
75
77 static void walk(node_t * n, BinaryFunction f) {
79 walk(n->L, f);
80 f(n->key, n->val);
81 walk(n->R, f);
82 }
83
86 clean_up(n->L);
87 clean_up(n->R);
88 delete n;
89 }
90
91 public:
92 binary_search_tree(): root(0), num_nodes(0) {}
93 ~binary_search_tree() { clean_up(root); }
94 int size() const { return num_nodes; }
96
97 bool insert(const key_t & key, const val_t & val) {
98 if (insert(root, key, val)) {
99 num_nodes++;
100 return true;
101 }
102 return false;
103 }
104
105 bool erase(const key_t & key) {
106 if (erase(root, key)) {
107 num_nodes--;
108 return true;
109 }
110 return false;
111 }
112
113 template<class BinaryFunction> void walk(BinaryFunction f) {
114 walk(root, f);
115 }
116
117 val_t * find(const key_t & key) {
118 for (node_t *n = root; n != 0; ) {
119 if (n->key == key) return &(n->val);
120 n = (key < n->key ? n->L : n->R);
121 }
122 return 0; //key not found
123 }
124 };
125
127
130
131 void printch(int k, char v) { cout << v; }
132
133 int main() {
134 binary_search_tree<int, char> T;
135 T.insert(2, b);
136 T.insert(1, a);
137 T.insert(3, c);
138 T.insert(5, e);
139 T.insert(4, x);
140 *T.find(4) = d;
141 cout << "In-order: ";
142 T.walk(printch); //abcde
143 cout << "\nRemoving node with key 3...";
144 cout << (T.erase(3) ? "Success!" : "Failed");
145 cout << "\n";
146 return 0;
147 }
3.5.2 Treap
1 /*
2
9
10 Treaps use randomly generated priorities to reduce the height of the
11 tree. We assume that the rand() function in <cstdlib> is 16-bits, and
12 call it twice to generate a 32-bit number. For the treap to be
13 effective, the range of the randomly generated numbers should be
14 between 0 and around the number of elements in the treap.
15
18 associated with each node. In order to make it behave like an std::multiset
19 or std::multimap, make appropriate changes with key comparisons (e.g.
20 change (k < n->key) to (k <= n->key) in search conditions).
21
22 Time Complexity: insert(), erase(), and find() are O(log(N)) on average
23 and O(N) in the worst case. Despite the technically O(N) worst case,
24 such cases are still extremely difficult to trigger, making treaps
25 very practice in many programming contest applications. walk() is O(N).
26
27 Space Complexity: O(N) on the number of nodes.
28
29 */
30
31 #include <cstdlib> /* srand(), rand() */
33
34 template<class key_t, class val_t> class treap {
35 struct node_t {
38 }
39
40 key_t key;
41 val_t val;
42 int priority;
43 node_t *L, *R;
44
45 node_t(const key_t & k, const val_t & v): key(k), val(v), L(0), R(0) {
47 }
48 } *root;
49
50 int num_nodes;
51
52 static void rotate_l(node_t *& k2) {
53 node_t *k1 = k2->R;
54 k2->R = k1->L;
55 k1->L = k2;
56 k2 = k1;
57 }
58
59 static void rotate_r(node_t *& k2) {
60 node_t *k1 = k2->L;
61 k2->L = k1->R;
62 k1->R = k2;
63 k2 = k1;
64 }
65
66 static bool insert(node_t *& n, const key_t & k, const val_t & v) {
67 if (n == 0) {
69 return true;
70 }
71 if (k < n->key && insert(n->L, k, v)) {
72 if (n->L->priority < n->priority) rotate_r(n);
73 return true;
74 } else if (n->key < k && insert(n->R, k, v)) {
75 if (n->R->priority < n->priority) rotate_l(n);
76 return true;
77 }
78 return false;
79 }
80
81 static bool erase(node_t *& n, const key_t & k) {
83 if (k < n->key) return erase(n->L, k);
84 if (k > n->key) return erase(n->R, k);
85 if (n->L == 0 || n->R == 0) {
86 node_t *temp = n;
87 n = (n->L != 0) ? n->L : n->R;
88 delete temp;
89 return true;
90 }
91 if (n->L->priority < n->R->priority) {
92 rotate_r(n);
93 return erase(n->R, k);
94 }
95 rotate_l(n);
96 return erase(n->L, k);
97 }
98
101 if (n == 0) return;
102 walk(n->L, f);
104 walk(n->R, f);
105 }
106
108 if (n == 0) return;
109 clean_up(n->L);
110 clean_up(n->R);
111 delete n;
112 }
113
114 public:
115 treap(): root(0), num_nodes(0) { srand(time(0)); }
116 ~treap() { clean_up(root); }
117 int size() const { return num_nodes; }
119
120 bool insert(const key_t & key, const val_t & val) {
121 if (insert(root, key, val)) {
122 num_nodes++;
123 return true;
124 }
125 return false;
126 }
127
128 bool erase(const key_t & key) {
129 if (erase(root, key)) {
130 num_nodes--;
131 return true;
132 }
133 return false;
134 }
135
137 walk(root, f);
138 }
139
141 for (node_t *n = root; n != 0; ) {
143 n = (key < n->key ? n->L : n->R);
144 }
146 }
147 };
148
150
154
156
157 int main() {
158 treap<int, char> T;
159 T.insert(2, b);
160 T.insert(1, a);
161 T.insert(3, c);
162 T.insert(5, e);
163 T.insert(4, x);
164 *T.find(4) = d;
167 cout << "\nRemoving node with key 3...";
168 cout << (T.erase(3) ? "Success!" : "Failed");
169 cout << "\n";
170
171 //stress test - runs in <0.5 seconds
172 //insert keys in an order that would break a normal BST
173 treap<int, int> T2;
174 for (int i = 0; i < 1000000; i++)
175 T2.insert(i, i*1337);
176 for (int i = 0; i < 1000000; i++)
177 assert(*T2.find(i) == i*1337);
178 return 0;
179 }
3.5.3 Size Balanced Tree (Order Statistics)
1 /*
2
9
10 The size balanced tree is a data structure first published in 2007 by
11 Chinese student Chen Qifeng. The tree is rebalanced by examining the sizes
12 of each nodes subtrees. It is popular amongst Chinese OI competitors due
13 to its speed, simplicity to implement, and ability to double up as an
14 ordered statistics tree if necessary.
15 For more info, see: http://wcipeg.com/wiki/Size_Balanced_Tree
16
17 An ordered statistics tree is a BST that supports additional operations:
18 - Select(i): find the i-th smallest element stored in the tree
19 - Rank(x): find the rank of element x in the tree,
20 i.e. its index in the sorted list of elements of the tree
21 For more info, see: http://en.wikipedia.org/wiki/Order_statistic_tree
22

25 associated with each node. Making a size balanced tree behave like an
26 std::multiset or std::multimap is a more complex issue. Refer to the
27 articles above and determine the correct way to preserve the binary search
28 tree property with maintain() if equivalent keys are allowed.
29
30 Time Complexity: insert(), erase(), find(), select() and rank() are
31 O(log N) on the number of elements in the tree. walk() is O(N).
32
33 Space Complexity: O(N) on the number of nodes in the tree.
34
35 */
36
38 #include <utility> /* pair */
39
40 template<class key_t, class val_t> class size_balanced_tree {
41 struct node_t {
42 key_t key;
43 val_t val;
44 int size;
45 node_t * c[2];
46
47 node_t(const key_t & k, const val_t & v) {
48 key = k, val = v;
49 size = 1;
50 c[0] = c[1] = 0;
51 }
52
53 void update() {
54 size = 1;
55 if (c[0]) size += c[0]->size;
56 if (c[1]) size += c[1]->size;
57 }
58 } *root;
59
60 static inline int size(node_t * n) {
61 return n ? n->size : 0;
62 }
63
64 static void rotate(node_t *& n, bool d) {
65 node_t * p = n->c[d];
66 n->c[d] = p->c[!d];
67 p->c[!d] = n;
68 n->update();
69 p->update();
70 n = p;
71 }
72
73 static void maintain(node_t *& n, bool d) {
74 if (n == 0 || n->c[d] == 0) return;
75 node_t *& p = n->c[d];
76 if (size(p->c[d]) > size(n->c[!d])) {
77 rotate(n, d);
78 } else if (size(p->c[!d]) > size(n->c[!d])) {
79 rotate(p, !d);
80 rotate(n, d);
81 } else return;
82 maintain(n->c[0], 0);
83 maintain(n->c[1], 1);
84 maintain(n, 0);
85 maintain(n, 1);
86 }
87
88 static void insert(node_t *& n, const key_t & k, const val_t & v) {
89 if (n == 0) {
91 return;
92 }
93 if (k < n->key) {
94 insert(n->c[0], k, v);
95 maintain(n, 0);
96 } else if (n->key < k) {
97 insert(n->c[1], k, v);
98 maintain(n, 1);
99 } else return;
100 n->update();
101 }
102
103 static void erase(node_t *& n, const key_t & k) {
104 if (n == 0) return;
105 bool d = k < n->key;
106 if (k < n->key) {
107 erase(n->c[0], k);
108 } else if (n->key < k) {
109 erase(n->c[1], k);
110 } else {
111 if (n->c[1] == 0 || n->c[0] == 0) {
112 delete n;
113 n = n->c[1] == 0 ? n->c[0] : n->c[1];
114 return;
115 }
116 node_t * p = n->c[1];
117 while (p->c[0] != 0) p = p->c[0];
118 n->key = p->key;
119 erase(n->c[1], p->key);
120 }
121 maintain(n, d);
122 n->update();
123 }
124
127 if (n == 0) return;
128 walk(n->c[0], f);
130 walk(n->c[1], f);
131 }
132
133 static std::pair<key_t, val_t> select(node_t *& n, int k) {
134 int r = size(n->c[0]);
135 if (k < r) return select(n->c[0], k);
136 if (k > r) return select(n->c[1], k - r - 1);
137 return std::make_pair(n->key, n->val);
138 }
139
140 static int rank(node_t * n, const key_t & k) {
141 if (n == 0)
142 throw std::runtime_error("Cannot rank key not in tree.");
143 int r = size(n->c[0]);
144 if (k < n->key) return rank(n->c[0], k);
145 if (n->key < k) return rank(n->c[1], k) + r + 1;
146 return r;
147 }
148
150 if (n == 0) return;
151 clean_up(n->c[0]);
152 clean_up(n->c[1]);
153 delete n;
154 }
155
156 public:
157 size_balanced_tree() : root(0) {}
158 ~size_balanced_tree() { clean_up(root); }
159 int size() { return size(root); }
161
162 void insert(const key_t & key, const val_t & val) {
163 insert(root, key, val);
164 }
165
166 void erase(const key_t & key) {
167 erase(root, key);
168 }
169
171 walk(root, f);
172 }
173
175 for (node_t *n = root; n != 0; ) {
177 n = (key < n->key ? n->c[0] : n->c[1]);
178 }
180 }
181
182 std::pair<key_t, val_t> select(int k) {
183 if (k >= size(root))
184 throw std::runtime_error("k must be smaller size of tree.");
185 return select(root, k);
186 }
187
188 int rank(const key_t & key) {
189 return rank(root, key);
190 }
191 };
192
194
198
200
201 int main() {
202 size_balanced_tree<int, char> T;
203 T.insert(2, b);
204 T.insert(1, a);
205 T.insert(3, c);
206 T.insert(5, e);
207 T.insert(4, x);
208 *T.find(4) = d;
211 T.erase(3);
212 cout << "\nRank of 2: " << T.rank(2); //1
213 cout << "\nRank of 5: " << T.rank(5); //3
214 cout << "\nValue of 3rd smallest key: ";
215 cout << T.select(2).second; //d
216 cout << "\n";
217
218 //stress test - runs in <1 second
219 //insert keys in an order that would break a normal BST
220 size_balanced_tree<int, int> T2;
221 for (int i = 0; i < 1000000; i++)
222 T2.insert(i, i*1337);
223 for (int i = 0; i < 1000000; i++)
224 assert(*T2.find(i) == i*1337);
225 return 0;
226 }
3.5.4 Hashmap (Chaining)
1 /*
2
3 Description: A hashmap (std::unordered_map in C++11) is an
4 alternative to a binary search tree. Hashmaps use more memory than
5 BSTs, but are usually more efficient. The following implementation
6 uses the chaining method to handle collisions. You can use the
7 hash algorithms provided in the example, or define your own.
8
9 Time Complexity: insert(), remove(), find(), are O(1) amortized.
10 rehash() is O(N).
11
12 Space Complexity: O(N) on the number of entries.
13
14 */
15
16 #include <list>
17
18 template<class key_t, class val_t, class Hash> class hashmap {
19 struct entry_t {
20 key_t key;
21 val_t val;
22 entry_t(const key_t & k, const val_t & v): key(k), val(v) {}
23 };
24
25 std::list<entry_t> * table;
26 int table_size, map_size;
27
28 /**
29 * This doubles the table size, then rehashes every entry.
30 * Rehashing is expensive; it is strongly suggested for the
31 * table to be constructed with a large size to avoid rehashing.
32 */
33 void rehash() {
34 std::list<entry_t> * old = table;
35 int old_size = table_size;
36 table_size = 2*table_size;
37 table = new std::list<entry_t>[table_size];
38 map_size = 0;
39 typename std::list<entry_t>::iterator it;
40 for (int i = 0; i < old_size; i++)
41 for (it = old[i].begin(); it != old[i].end(); ++it)
42 insert(it->key, it->val);
43 delete[] old;
44 }
45
46 public:
47 hashmap(int size = 1024): table_size(size), map_size(0) {
48 table = new std::list<entry_t>[table_size];
49 }
50
51 ~hashmap() { delete[] table; }
52 int size() const { return map_size; }
53
54 void insert(const key_t & key, const val_t & val) {
55 if (find(key) != 0) return;
56 if (map_size >= table_size) rehash();
57 unsigned int i = Hash()(key) % table_size;
58 table[i].push_back(entry_t(key, val));
59 map_size++;
60 }
61
62 void remove(const key_t & key) {
64 typename std::list<entry_t>::iterator it = table[i].begin();
65 while (it != table[i].end() && it->key != key) ++it;
66 if (it == table[i].end()) return;
67 table[i].erase(it);
68 map_size--;
69 }
70
73 typename std::list<entry_t>::iterator it = table[i].begin();
74 while (it != table[i].end() && it->key != key) ++it;
75 if (it == table[i].end()) return 0;
76 return &(it->val);
77 }
78
79 val_t & operator [] (const key_t & key) {
80 val_t * ret = find(key);
81 if (ret != 0) return *ret;
82 insert(key, val_t());
83 return *find(key);
84 }
85 };
86
87 /*** Examples of Hash Algorithm Definitions ***/

88
89 #include <string>
90
91 struct class_hash {
92 unsigned int operator () (int key) {
93 return class_hash()((unsigned int)key);
94 }
95
96 unsigned int operator () (long long key) {
97 return class_hash()((unsigned long long)key);
98 }
99
100 //Knuths multiplicative method (one-to-one)
101 unsigned int operator () (unsigned int key) {
102 return key * 2654435761u; //or just return key
103 }
104
105 //Jenkinss 64-bit hash
106 unsigned int operator () (unsigned long long key) {
107 key += ~(key << 32); key ^= (key >> 22);
108 key += ~(key << 13); key ^= (key >> 8);
109 key += (key << 3); key ^= (key >> 15);
110 key += ~(key << 27); key ^= (key >> 31);
111 return key;
112 }
113
114 //Jenkinss one-at-a-time hash
115 unsigned int operator () (const std::string & key) {
116 unsigned int hash = 0;
117 for (unsigned int i = 0; i < key.size(); i++) {
118 hash += ((hash += key[i]) << 10);
119 hash ^= (hash >> 6);
120 }
121 hash ^= ((hash += (hash << 3)) >> 11);
122 return hash + (hash << 15);
123 }
124 };
125
127
130
131 int main() {
132 hashmap<string, int, class_hash> M;
133 M["foo"] = 1;
134 M.insert("bar", 2);
135 cout << M["foo"] << M["bar"] << endl; //prints 12
136 cout << M["baz"] << M["qux"] << endl; //prints 00
137 M.remove("foo");
138 cout << M.size() << endl; //prints 3
139 cout << M["foo"] << M["bar"] << endl; //prints 02
140 return 0;
141 }
3.5.5 Skip List (Probabilistic)

1 /*
2
3 Description: A skip list is an alternative to binary search trees.
4 Fast search is made possible by maintaining a linked hierarchy of
5 subsequences, each skipping over fewer elements. Searching starts
6 in the sparsest subsequence until two consecutive elements have
7 been found, one smaller and one larger than the element searched for.
8 Skip lists are generally slower than binary search trees, but can
9 be easier to implement. The following version uses randomized levels.
10
11 Time Complexity: insert(), erase(), count() and find() are O(log(N))
12 on average, but O(N) in the worst case. walk() is O(N).
13
14 Space Complexity: O(N) on the number of elements inserted on average,
15 but O(N log N) in the worst case.
16
17 */
18
19 #include <cmath> /* log() */
20 #include <cstdlib> /* rand(), srand() */
21 #include <cstring> /* memset() */
23
24 template<class key_t, class val_t> struct skip_list {
25 static const int MAX_LEVEL = 32; //~ log2(max # of keys)
26
27 static int random_level() { //geometric distribution
28 static const float P = 0.5;
29 int lvl = log((float)rand()/RAND_MAX)/log(1.0 - P);
30 return lvl < MAX_LEVEL ? lvl : MAX_LEVEL;
31 }
32
33 struct node_t {
34 key_t key;
35 val_t val;
36 node_t **next;
37
38 node_t(int level, const key_t & k, const val_t & v) {
39 next = new node_t * [level + 1];
40 memset(next, 0, sizeof(node_t*)*(level + 1));
41 key = k;
42 val = v;
43 }
44
45 ~node_t() { delete[] next; }
46 } *head, *update[MAX_LEVEL + 1];
47
48 int level, num_nodes;
49
50 skip_list() {
51 srand(time(0));
52 head = new node_t(MAX_LEVEL, key_t(), val_t());
53 level = num_nodes = 0;
54 }
55
56 ~skip_list() { delete head; }
57 int size() { return num_nodes; }
58 bool empty() { return num_nodes == 0; }
59 int count(const key_t & k) { return find(k) != 0; }
60
61 void insert(const key_t & k, const val_t & v) {
62 node_t * n = head;
63 memset(update, 0, sizeof(node_t*)*(MAX_LEVEL + 1));
64 for (int i = level; i >= 0; i--) {
65 while (n->next[i] && n->next[i]->key < k) n = n->next[i];
66 update[i] = n;
67 }
68 n = n->next[0];
69 if (!n || n->key != k) {
70 int lvl = random_level();
71 if (lvl > level) {
72 for (int i = level + 1; i <= lvl; i++) update[i] = head;
73 level = lvl;
74 }
75 n = new node_t(lvl, k, v);
76 num_nodes++;
77 for (int i = 0; i <= lvl; i++) {
78 n->next[i] = update[i]->next[i];
79 update[i]->next[i] = n;
80 }
81 } else if (n && n->key == k && n->val != v) {
82 n->val = v;
83 }
84 }
85
86 void erase(const key_t & k) {
88 memset(update, 0, sizeof(node_t*)*(MAX_LEVEL + 1));
89 for (int i = level;i >= 0; i--) {
90 while (n->next[i] && n->next[i]->key < k) n = n->next[i];
91 update[i] = n;
92 }
93 n = n->next[0];
94 if (n->key == k) {
95 for (int i = 0; i <= level; i++) {
96 if (update[i]->next[i] != n) break;
97 update[i]->next[i] = n->next[i];
98 }
99 delete n;
100 num_nodes--;
101 while (level > 0 && !head->next[level]) level--;
102 }
103 }
104
105 val_t * find(const key_t & k) {
107 for (int i = level; i >= 0; i--)
108 while (n->next[i] && n->next[i]->key < k)
109 n = n->next[i];
110 n = n->next[0];
111 if (n && n->key == k) return &(n->val);
112 return 0; //not found
113 }
114
116 node_t *n = head->next[0];
117 while (n) {
119 n = n->next[0];
120 }
121 }
122 };
123
124 /*** Example Usage: Random Tests ***/
125
128 #include <map>
130
131 int main() {
132 map<int, int> m;
133 skip_list<int, int> s;
134 for (int i = 0; i < 50000; i++) {
135 int op = rand() % 3;
136 int val1 = rand(), val2 = rand();
137 if (op == 0) {
138 m[val1] = val2;
139 s.insert(val1, val2);
140 } else if (op == 1) {
141 if (!m.count(val1)) continue;
142 m.erase(val1);
143 s.erase(val1);
144 } else if (op == 2) {
145 assert(s.count(val1) == (int)m.count(val1));
146 if (m.count(val1)) {
147 assert(m[val1] == *s.find(val1));
148 }
149 }
150 }
151 return 0;
152 }
3.6 Tree Data Structures
3.6.1 Heavy-Light Decomposition
1 /*
2
3 Description: Given an undirected, connected graph that is a tree, the
4 heavy-light decomposition (HLD) on the graph is a partitioning of the
5 vertices into disjoint paths to later support dynamic modification and
6 querying of values on paths between pairs of vertices.
7 See: http://wcipeg.com/wiki/Heavy-light_decomposition
8 and: http://blog.anudeep2011.com/heavy-light-decomposition/
9 To support dynamic adding and removal of edges, see link/cut tree.
10
11 Note: The adjacency list tree[] that is passed to the constructor must
12 not be changed afterwards in order for modify() and query() to work.
13
14 Time Complexity: O(N) for the constructor and O(log N) in the worst
15 case for both modify() and query(), where N is the number of vertices.
16
17 Space Complexity: O(N) on the number of vertices in the tree.
3.6. Tree Data Structures 143
18
19 */
20
24
25 template<class T> class heavy_light {
26 //true if you want values on edges, false if you want values on vertices
27 static const bool VALUES_ON_EDGES = true;
28
33 return x + y;
34 }
35
38 }
39
41 if (delta == null_delta()) return null_delta();
46 return delta;
47 }
48
49 static inline T init_value() { return 0; }
52
53 static inline T join_value_with_delta(const T & v, const T & delta) {
54 return delta == null_delta() ? v : modify_op(v, delta);
55 }
56
58 if (delta1 == null_delta()) return delta2;
61 }
62
63 int counter, paths;
64 std::vector<int> *adj;
65 std::vector<std::vector<T> > value, delta;
66 std::vector<std::vector<int> > len;
67 std::vector<int> size, parent, tin, tout;
68 std::vector<int> path, pathlen, pathpos, pathroot;
69
70 void precompute_dfs(int u, int p) {
71 tin[u] = counter++;
72 parent[u] = p;
73 size[u] = 1;
74 for (int j = 0, v; j < (int)adj[u].size(); j++) {
76 precompute_dfs(v, u);
77 size[u] += size[v];
78 }
79 tout[u] = counter++;
80 }
81
82 int new_path(int u) {
83 pathroot[paths] = u;
84 return paths++;
85 }
86
87 void build_paths(int u, int path) {
88 this->path[u] = path;
89 pathpos[u] = pathlen[path]++;
91 if ((v = adj[u][j]) == parent[u]) continue;
92 build_paths(v, 2*size[v] >= size[u] ? path : new_path(v));
93 }
94 }
95
96 inline T join_value_with_delta0(int path, int i) {
97 return join_value_with_delta(value[path][i],
98 delta_on_segment(delta[path][i], len[path][i]));
99 }
100
101 void push_delta(int path, int i) {
102 int d = 0;
103 while ((i >> d) > 0) d++;
104 for (d -= 2; d >= 0; d--) {
105 int x = i >> d;
106 value[path][x >> 1] = join_value_with_delta0(path, x >> 1);
107 delta[path][x] = join_deltas(delta[path][x], delta[path][x >> 1]);
108 delta[path][x ^ 1] = join_deltas(delta[path][x ^ 1], delta[path][x >> 1]);
109 delta[path][x >> 1] = null_delta();
110 }
111 }
112
113 T query(int path, int a, int b) {
114 push_delta(path, a += value[path].size() >> 1);
115 push_delta(path, b += value[path].size() >> 1);
116 T res = null_value();
117 for (; a <= b; a = (a + 1) >> 1, b = (b - 1) >> 1) {
118 if ((a & 1) != 0)
119 res = query_op(res, join_value_with_delta0(path, a));
120 if ((b & 1) == 0)
121 res = query_op(res, join_value_with_delta0(path, b));
122 }
123 return res;
124 }
125
126 void modify(int path, int a, int b, const T & delta) {
127 push_delta(path, a += value[path].size() >> 1);
128 push_delta(path, b += value[path].size() >> 1);
129 int ta = -1, tb = -1;
130 for (; a <= b; a = (a + 1) >> 1, b = (b - 1) >> 1) {
131 if ((a & 1) != 0) {
132 this->delta[path][a] = join_deltas(this->delta[path][a], delta);
133 if (ta == -1) ta = a;
134 }
135 if ((b & 1) == 0) {
136 this->delta[path][b] = join_deltas(this->delta[path][b], delta);

137 if (tb == -1) tb = b;
138 }
139 }
140 for (int i = ta; i > 1; i >>= 1)
141 value[path][i >> 1] = query_op(join_value_with_delta0(path, i),
142 join_value_with_delta0(path, i ^ 1));
143 for (int i = tb; i > 1; i >>= 1)
144 value[path][i >> 1] = query_op(join_value_with_delta0(path, i),
145 join_value_with_delta0(path, i ^ 1));
146 }
147
148 inline bool is_ancestor(int p, int ch) {
149 return tin[p] <= tin[ch] && tout[ch] <= tout[p];
150 }
151
152 public:
153 heavy_light(int N, std::vector<int> tree[]): size(N), parent(N),
154 tin(N), tout(N), path(N), pathlen(N), pathpos(N), pathroot(N) {
155 adj = tree;
156 counter = paths = 0;
157 precompute_dfs(0, -1);
158 build_paths(0, new_path(0));
159 value.resize(paths);
160 delta.resize(paths);
161 len.resize(paths);
162 for (int i = 0; i < paths; i++) {
163 int m = pathlen[i];
164 value[i].assign(2*m, init_value());
165 delta[i].assign(2*m, null_delta());
166 len[i].assign(2*m, 1);
167 for (int j = 2*m - 1; j > 1; j -= 2) {
168 value[i][j >> 1] = query_op(value[i][j], value[i][j ^ 1]);
169 len[i][j >> 1] = len[i][j] + len[i][j ^ 1];
170 }
171 }
172 }
173
175 T res = null_value();
176 for (int root; !is_ancestor(root = pathroot[path[a]], b); a = parent[root])
177 res = query_op(res, query(path[a], 0, pathpos[a]));
178 for (int root; !is_ancestor(root = pathroot[path[b]], a); b = parent[root])
179 res = query_op(res, query(path[b], 0, pathpos[b]));
180 if (VALUES_ON_EDGES && a == b) return res;
181 return query_op(res, query(path[a], std::min(pathpos[a], pathpos[b]) +
182 VALUES_ON_EDGES, std::max(pathpos[a], pathpos[b])));
183 }
184
186 for (int root; !is_ancestor(root = pathroot[path[a]], b); a = parent[root])
187 modify(path[a], 0, pathpos[a], delta);
188 for (int root; !is_ancestor(root = pathroot[path[b]], a); b = parent[root])
189 modify(path[b], 0, pathpos[b], delta);
190 if (VALUES_ON_EDGES && a == b) return;
191 modify(path[a], std::min(pathpos[a], pathpos[b]) + VALUES_ON_EDGES,
192 std::max(pathpos[a], pathpos[b]), delta);
193 }
194 };
195
197
200
203
204 /*
205 w=10 w=20 w=40
206 0---------1---------2---------3
207 \
208 ---------4
209 w=30
210 */
211 int main() {
212 adj[0].push_back(1);
220 heavy_light<int> hld(5, adj);
221 hld.modify(0, 1, 10);
222 hld.modify(1, 2, 20);
223 hld.modify(2, 3, 40);
224 hld.modify(2, 4, 30);
225 cout << hld.query(0, 3) << "\n"; //40
226 cout << hld.query(2, 4) << "\n"; //30
227 hld.modify(3, 4, 50); //w[every edge from 3 to 4] += 50
228 cout << hld.query(1, 4) << "\n"; //80
229 return 0;
230 }
3.6.2 Link-Cut Tree
1 /*
2
3 Description: Given an unweighted forest of trees where each node
4 has an associated value, a link/cut tree can be used to dynamically
5 query and modify values on the path between pairs of nodes a tree.
6 This problem can be solved using heavy-light decomposition, which
7 also supports having values stored on edges rather than the nodes.
8 However in a link/cut tree, nodes in different trees may be
9 dynamically linked, edges between nodes in the same tree may be
10 dynamically split, and connectivity between two nodes (whether they
11 are in the same tree) may be checked.
12
13 Time Complexity: O(log N) amortized for make_root(), link(), cut(),
14 connected(), modify(), and query(), where N is the number of nodes
15 in the forest.
16
17 Space Complexity: O(N) on the number of nodes in the forest.
18
19 */
20
21 #include <algorithm> /* std::max(), std::swap() */
23 #include <map>
25
26 template<class T> class linkcut_forest {
31 return x + y;
32 }
33
36 }
37
39 if (delta == null_delta()) return null_delta();
44 return delta;
45 }
46
49
50 static inline T join_value_with_delta(const T & v, const T & delta) {
51 return delta == null_delta() ? v : modify_op(v, delta);
52 }
53
58 }
59
60 struct node_t {
61 T value, subtree_value, delta;
62 int size;
63 bool rev;
64 node_t *L, *R, *parent;
65
66 node_t(const T & v) {
67 value = subtree_value = v;
69 size = 1;
70 rev = false;
71 L = R = parent = 0;
72 }
73
74 bool is_root() { //is this the root of a splay tree?
75 return parent == 0 || (parent->L != this && parent->R != this);
76 }
77
78 void push() {
79 if (rev) {
80 rev = false;
81 std::swap(L, R);
82 if (L != 0) L->rev = !L->rev;
83 if (R != 0) R->rev = !R->rev;
84 }
85 value = join_value_with_delta(value, delta);
86 subtree_value = join_value_with_delta(subtree_value,
87 delta_on_segment(delta, size));
88 if (L != 0) L->delta = join_deltas(L->delta, delta);
89 if (R != 0) R->delta = join_deltas(R->delta, delta);
91 }
92
93 void update() {
94 subtree_value = query_op(query_op(get_subtree_value(L),
95 join_value_with_delta(value, delta)),
96 get_subtree_value(R));
97 size = 1 + get_size(L) + get_size(R);
98 }
99 };
100
101 static inline int get_size(node_t * n) {
102 return n == 0 ? 0 : n->size;
103 }
104
105 static inline int get_subtree_value(node_t * n) {
106 return n == 0 ? null_value() : join_value_with_delta(n->subtree_value,
107 delta_on_segment(n->delta, n->size));
108 }
109
110 static void connect(node_t * ch, node_t * p, char is_left) {
111 if (ch != 0) ch->parent = p;
112 if (is_left < 0) return;
113 (is_left ? p->L : p->R) = ch;
114 }
115
116 /** rotates edge (n, n.parent)
117 * g g
118 * / /
119 * p n
120 * / \ --> / \
121 * n p.r n.l p
122 * / \ / \
123 * n.l n.r n.r p.r
124 */
125 static void rotate(node_t * n) {
126 node_t *p = n->parent, *g = p->parent;
127 bool is_rootp = p->is_root(), is_left = (n == p->L);
128 connect(is_left ? n->R : n->L, p, is_left);
129 connect(p, n, !is_left);
130 connect(n, g, is_rootp ? -1 : (p == g->L));
131 p->update();
132 }
133
134 /** brings n to the root, balancing tree
135 *
136 * zig-zig case:
137 * g n
138 * / \ p / \
139 * p g.r rot(p) / \ rot(n) n.l p
140 * / \ --> n g --> / \
141 * n p.r / \ / \ n.r g
142 * / \ n.l n.r p.r g.r / \
143 * n.l n.r p.r g.r
144 *
145 * zig-zag case:
146 * g g
147 * / \ / \ n
148 * p g.r rot(n) n g.r rot(n) / \
149 * / \ --> / \ --> p g
150 * p.l n p n.r / \ / \
151 * / \ / \ p.l n.l n.r g.r
152 * n.l n.r p.l n.l
153 */
154 static void splay(node_t * n) {
155 while (!n->is_root()) {
156 node_t *p = n->parent, *g = p->parent;
157 if (!p->is_root()) g->push();
158 p->push();
159 n->push();
160 if (!p->is_root())
161 rotate((n == p->L) == (p == g->L) ? p/*zig-zig*/ : n/*zig-zag*/);
162 rotate(n);
163 }
164 n->push();
165 n->update();
166 }
167
168 //makes node n the root of the virtual tree,
169 //and also n becomes the leftmost node in its splay tree
170 static node_t * expose(node_t * n) {
171 node_t *prev = 0;
172 for (node_t *i = n; i != 0; i = i->parent) {
173 splay(i);
174 i->L = prev;
175 prev = i;
176 }
177 splay(n);
178 return prev;
179 }
180
181 std::map<int, node_t*> nodes; //use array if ID compression not required
182 node_t *u, *v; //temporary
183
184 void get_uv(int a, int b) {
185 static typename std::map<int, node_t*>::iterator it1, it2;
186 it1 = nodes.find(a);
187 it2 = nodes.find(b);
188 if (it1 == nodes.end() || it2 == nodes.end())
189 throw std::runtime_error("Error: a or b does not exist in forest.");
190 u = it1->second;
191 v = it2->second;
192 }
193
194 public:
195 ~linkcut_forest() {
196 static typename std::map<int, node_t*>::iterator it;

197 for (it = nodes.begin(); it != nodes.end(); ++it)
198 delete it->second;
199 }
200
201 void make_root(int id, const T & initv) {
202 if (nodes.find(id) != nodes.end())
203 throw std::runtime_error("Cannot make_root(): ID already exists.");
204 node_t * n = new node_t(initv);
205 expose(n);
206 n->rev = !n->rev;
207 nodes[id] = n;
208 }
209
210 bool connected(int a, int b) {
211 get_uv(a, b);
212 if (a == b) return true;
213 expose(u);
214 expose(v);
215 return u->parent != 0;
216 }
217
218 void link(int a, int b) {
219 if (connected(a, b))
220 throw std::runtime_error("Error: a and b are already connected.");
221 get_uv(a, b);
222 expose(u);
223 u->rev = !u->rev;
224 u->parent = v;
225 }
226
227 void cut(int a, int b) {
228 get_uv(a, b);
229 expose(u);
230 u->rev = !u->rev;
231 expose(v);
232 if (v->R != u || u->L != 0)
233 throw std::runtime_error("Error: edge (a, b) does not exist.");
234 v->R->parent = 0;
235 v->R = 0;
236 }
237
239 if (!connected(a, b))
240 throw std::runtime_error("Error: a and b are not connected.");
241 get_uv(a, b);
242 expose(u);
243 u->rev = !u->rev;
244 expose(v);
245 return get_subtree_value(v);
246 }
247
249 if (!connected(a, b))
250 throw std::runtime_error("Error: a and b are not connected.");
251 get_uv(a, b);
252 expose(u);
253 u->rev = !u->rev;
254 expose(v);
3.7. Lowest Common Ancestor 151
255 v->delta = join_deltas(v->delta, delta);

256 }
257 };
258
260
263
264 int main() {
265 linkcut_forest<int> F;
266 /*
267 v=10 v=40 v=20 v=10
268 0---------1---------2---------3
269 \
270 ---------4
271 v=30
272 */
273 F.make_root(0, 10);
274 F.make_root(1, 40);
275 F.make_root(2, 20);
276 F.make_root(3, 10);
277 F.make_root(4, 30);
278 F.link(0, 1);
279 F.link(1, 2);
280 F.link(2, 3);
281 F.link(2, 4);
282 cout << F.query(1, 4) << "\n"; //40
283 F.modify(1, 1, -10);
284 F.modify(3, 4, -10);
285 /*
286 v=10 v=30 v=10 v=0
287 0---------1---------2---------3
288 \
289 ---------4
290 v=20
291 */
292 cout << F.query(0, 4) << "\n"; //30
293 cout << F.query(3, 4) << "\n"; //20
294 F.cut(1, 2);
295 cout << F.connected(1, 2) << "\n"; //0
298 return 0;
299 }
3.7 Lowest Common Ancestor
3.7.1 Sparse Tables
1 /*
2
3 Description: Given an undirected graph that is a tree, the
4 lowest common ancestor (LCA) of two nodes v and w is the
5 lowest (i.e. deepest) node that has both v and w as descendants,
6 where we define each node to be a descendant of itself (so if
7 v has a direct connection from w, w is the lowest common

8 ancestor). The following program uses sparse tables to solve
9 the problem on an unchanging graph.
10
11 Time Complexity: O(N log N) for build() and O(log N) for lca(),
12 where N is the number of nodes in the tree.
13
14 Space Complexity: O(N log N).
15
16 */
17
19
21 int len, timer, tin[MAXN], tout[MAXN];
22 std::vector<int> adj[MAXN], dp[MAXN];
23
24 void dfs(int u, int p) {
25 tin[u] = timer++;
26 dp[u][0] = p;
27 for (int i = 1; i < len; i++)
28 dp[u][i] = dp[dp[u][i - 1]][i - 1];
29 for (int j = 0, v; j < (int)adj[u].size(); j++)
30 if ((v = adj[u][j]) != p)
31 dfs(v, u);
32 tout[u] = timer++;
33 }
34
35 void build(int nodes, int root) {
36 len = 1;
37 while ((1 << len) <= nodes) len++;
39 dp[i].resize(len);
40 timer = 0;
41 dfs(root, root);
42 }
43
44 inline bool is_parent(int parent, int child) {
45 return tin[parent] <= tin[child] && tout[child] <= tout[parent];
46 }
47
48 int lca(int a, int b) {
49 if (is_parent(a, b)) return a;
50 if (is_parent(b, a)) return b;
51 for (int i = len - 1; i >= 0; i--)
52 if (!is_parent(dp[a][i], b))
53 a = dp[a][i];
54 return dp[a][0];
55 }
56
58
61
62 int main() {
3.7. Lowest Common Ancestor 153
71 build(5, 0);
72 cout << lca(3, 2) << "\n"; //1
73 cout << lca(2, 4) << "\n"; //0
74 return 0;
75 }
3.7.2 Segment Trees
1 /*
2
3 Description: Given a rooted tree, the lowest common ancestor (LCA)
4 of two nodes v and w is the lowest (i.e. deepest) node that has
5 both v and w as descendants, where we define each node to be a
6 descendant of itself (so if v has a direct connection from w, w
7 is the lowest common ancestor). This problem can be reduced to the
8 range minimum query problem using Eulerian tours.
9
10 Time Complexity: O(N log N) for build() and O(log N) for lca(),
11 where N is the number of nodes in the tree.
12
13 Space Complexity: O(N log N).
14
15 */
16
17 #include <algorithm> /* std::fill(), std::min(), std::max() */
19
21 int len, counter;
22 int depth[MAXN], dfs_order[2*MAXN], first[MAXN], minpos[8*MAXN];
23 std::vector<int> adj[MAXN];
24
25 void dfs(int u, int d) {
26 depth[u] = d;
27 dfs_order[counter++] = u;
29 if (depth[v = adj[u][j]] == -1) {
30 dfs(v, d + 1);
31 dfs_order[counter++] = u;
32 }
33 }
34 }
35
36 void build_tree(int n, int l, int h) {
37 if (l == h) {
38 minpos[n] = dfs_order[l];
39 return;
40 }
41 int lchild = 2 * n + 1, rchild = 2 * n + 2;
42 build_tree(lchild, l, (l + h)/2);
43 build_tree(rchild, (l + h) / 2 + 1, h);
44 minpos[n] = depth[minpos[lchild]] < depth[minpos[rchild]] ?
45 minpos[lchild] : minpos[rchild];
46 }
47
48 void build(int nodes, int root) {
49 std::fill(depth, depth + nodes, -1);
50 std::fill(first, first + nodes, -1);
51 len = 2*nodes - 1;
52 counter = 0;
53 dfs(root, 0);
54 build_tree(0, 0, len - 1);
55 for (int i = 0; i < len; i++)
56 if (first[dfs_order[i]] == -1)
57 first[dfs_order[i]] = i;
58 }
59
60 int get_minpos(int a, int b, int n, int l, int h) {
61 if (a == l && b == h) return minpos[n];
62 int mid = (l + h) >> 1;
63 if (a <= mid && b > mid) {
64 int p1 = get_minpos(a, std::min(b, mid), 2 * n + 1, l, mid);
65 int p2 = get_minpos(std::max(a, mid + 1), b, 2 * n + 2, mid + 1, h);
66 return depth[p1] < depth[p2] ? p1 : p2;
67 }
68 if (a <= mid) return get_minpos(a, std::min(b, mid), 2 * n + 1, l, mid);
69 return get_minpos(std::max(a, mid + 1), b, 2 * n + 2, mid + 1, h);
70 }
71
72 int lca(int a, int b) {
73 return get_minpos(std::min(first[a], first[b]),
74 std::max(first[a], first[b]), 0, 0, len - 1);
75 }
76
78
81
82 int main() {
91 build(5, 0);
92 cout << lca(3, 2) << "\n"; //1
93 cout << lca(2, 4) << "\n"; //0
94 return 0;
95 }
Chapter 4
Mathematics
4.1 Mathematics Toolbox
1 /*
2
3 Useful math definitions. Excludes geometry (see next chapter).
4
5 */
6
7 #include <algorithm> /* std::reverse() */
9 #include <cmath> /* a lot of things */
12
13 /* Definitions for Common Floating Point Constants */
14
15 const double PI = acos(-1.0), E = exp(1.0), root2 = sqrt(2.0);
16 const double phi = (1.0 + sqrt(5.0)) / 2.0; //golden ratio
17
18 //Sketchy but working defintions of +infinity, -infinity and quiet NaN
19 //A better way is using functions of std::numeric_limits<T> from <limits>
20 //See main() for identities involving the following special values.
21 const double posinf = 1.0 / 0.0, neginf = -1.0 / 0.0, NaN = -(0.0 / 0.0);
22
23 /*
24
25 Epsilon Comparisons
26
27 The range of values for which X compares EQ() to is [X - eps, X + eps].
28 For values to compare LT() and GT() x, they must fall outside of the range.
29
30 e.g. if eps = 1e-7, then EQ(1e-8, 2e-8) is true and LT(1e-8, 2e-8) is false.
31
32 */
33
34 const double eps = 1e-7;
35 #define EQ(a, b) (fabs((a) - (b)) <= eps) /* equal to */
36 #define NE(a, b) (fabs((a) - (b)) > eps) /* not equal to */
37 #define LT(a, b) ((a) < (b) - eps) /* less than */
155
156 Chapter 4. Mathematics
38 #define GT(a, b) ((a) > (b) + eps) /* greater than */

39 #define LE(a, b) ((a) <= (b) + eps) /* less than or equal to */
40 #define GE(a, b) ((a) >= (b) - eps) /* greater than or equal to */
41
42 /*
43
44 Sign Function:
45
46 Returns: -1 (if x < 0), 0 (if x == 0), or 1 (if x > 0)
47 Doesnt handle the sign of NaN like signbit() or copysign()
48
49 */
50
51 template<class T> int sgn(const T & x) {
52 return (T(0) < x) - (x < T(0));
53 }
54
55 /*
56
57 signbit() and copysign() functions, only in C++11 and later.
58
59 signbit() returns whether the sign bit of the floating point
60 number is set to true. If signbit(x), then x is "negative."
61 Note that signbit(0.0) == 0 but signbit(-0.0) == 1. This
62 also works as expected on NaN, -NaN, posinf, and neginf.
63
64 We implement this by casting the floating point value to an
65 integer type with the same number of bits so we can perform
66 shift operations on it, then we extract the sign bit.
67 Another way is using unions, but this is non-portable
68 depending on endianess of the platform. Unfortunately, we
69 cannot find the signbit of long doubles using the method
70 below because there is no corresponding 96-bit integer type.
71 Note that this will cause complaints with the compiler.
72
73 copysign(x, y) returns a number with the magnitude of x but
74 the sign of y.
75
76 Assumptions: sizeof(float) == sizeof(int) and
77 sizeof(long long) == sizeof(double)
78 CHAR_BITS == 8 (8 bits to a byte)
79
80 */
81
82 inline bool signbit(float x) {
83 return (*(int*)&x) >> (sizeof(float) * 8 - 1);
84 }
85
86 inline bool signbit(double x) {
87 return (*(long long*)&x) >> (sizeof(double) * 8 - 1);
88 }
89
90 template<class Double>
91 inline Double copysign(Double x, Double y) {
92 return signbit(y) ? -fabs(x) : fabs(x);
93 }
94
95 /*
96
4.1. Mathematics Toolbox 157
97 Floating Point Rounding Functions

98
99 floor() in <cmath> asymmetrically rounds down, towards -infinity,
100 while ceil() in <cmath> asymmetrically rounds up, towards +infinity.
101 The following are common alternative ways to round.
102
103 */
104
105 //symmetric round down, bias: towards zero (same as trunc() in C++11)
106 template<class Double> Double floor0(const Double & x) {
107 Double res = floor(fabs(x));
108 return (x < 0.0) ? -res : res;
109 }
110
111 //symmetric round up, bias: away from zero
112 template<class Double> Double ceil0(const Double & x) {
113 Double res = ceil(fabs(x));
114 return (x < 0.0) ? -res : res;
115 }
116
117 //round half up, bias: towards +infinity
118 template<class Double> Double roundhalfup(const Double & x) {
119 return floor(x + 0.5);
120 }
121
122 //round half up, bias: towards -infinity
123 template<class Double> Double roundhalfdown(const Double & x) {
124 return ceil(x - 0.5);
125 }
126
127 //symmetric round half down, bias: towards zero
128 template<class Double> Double roundhalfdown0(const Double & x) {
129 Double res = roundhalfdown(fabs(x));
130 return (x < 0.0) ? -res : res;
131 }
132
133 //symmetric round half up, bias: away from zero
134 template<class Double> Double roundhalfup0(const Double & x) {
135 Double res = roundhalfup(fabs(x));
136 return (x < 0.0) ? -res : res;
137 }
138
139 //round half to even (bankers rounding), bias: none
140 template<class Double>
141 Double roundhalfeven(const Double & x, const Double & eps = 1e-7) {
142 if (x < 0.0) return -roundhalfeven(-x, eps);
143 Double ipart;
144 modf(x, &ipart);
145 if (x - (ipart + 0.5) < eps)
146 return (fmod(ipart, 2.0) < eps) ? ipart : ceil0(ipart + 0.5);
147 return roundhalfup0(x);
148 }
149
150 //round alternating up/down for ties, bias: none for sequential calls
151 template<class Double> Double roundalternate(const Double & x) {
152 static bool up = true;
153 return (up = !up) ? roundhalfup(x) : roundhalfdown(x);
154 }
155
156 //symmetric round alternate, bias: none for sequential calls

157 template<class Double> Double roundalternate0(const Double & x) {
158 static bool up = true;
159 return (up = !up) ? roundhalfup0(x) : roundhalfdown0(x);
160 }
161
162 //round randomly for tie-breaking, bias: generators bias
163 template<class Double> Double roundrandom(const Double & x) {
164 return (rand() % 2 == 0) ? roundhalfup0(x) : roundhalfdown0(x);
165 }
166
167 //round x to N digits after the decimal using the specified round function
168 //e.g. roundplaces(-1.23456, 3, roundhalfdown0<double>) returns -1.235
169 template<class Double, class RoundFunction>
170 double roundplaces(const Double & x, unsigned int N, RoundFunction f) {
171 return f(x * pow(10, N)) / pow(10, N);
172 }
173
174 /*
175
176 Error Function (erf() and erfc() in C++11)
177
178 erf(x) = 2/sqrt(pi) * integral of exp(-t^2) dt from 0 to x
179 erfc(x) = 1 - erf(x)
180 Note that the functions are co-dependent.
181
182 Adapted from: http://www.digitalmars.com/archives/cplusplus/3634.html#N3655
183
184 */
185
186 //calculate 12 significant figs (dont ask for more than 1e-15)
187 static const double rel_error = 1e-12;
188
189 double erf(double x) {
190 if (signbit(x)) return -erf(-x);
191 if (fabs(x) > 2.2) return 1.0 - erfc(x);
192 double sum = x, term = x, xsqr = x * x;
193 int j = 1;
194 do {
195 term *= xsqr / j;
196 sum -= term / (2 * (j++) + 1);
197 term *= xsqr / j;
198 sum += term / (2 * (j++) + 1);
199 } while (fabs(term) / sum > rel_error);
200 return 1.128379167095512574 * sum; //1.128 ~ 2/sqrt(pi)
201 }
202
203 double erfc(double x) {
204 if (fabs(x) < 2.2) return 1.0 - erf(x);
205 if (signbit(x)) return 2.0 - erfc(-x);
206 double a = 1, b = x, c = x, d = x * x + 0.5, q1, q2 = 0, n = 1.0, t;
207 do {
208 t = a * n + b * x; a = b; b = t;
209 t = c * n + d * x; c = d; d = t;
210 n += 0.5;
211 q1 = q2;
212 q2 = b / d;
213 } while (fabs(q1 - q2) / q2 > rel_error);
214 return 0.564189583547756287 * exp(-x * x) * q2; //0.564 ~ 1/sqrt(pi)
215 }
216
217 /*
218
219 Gamma and Log-Gamma Functions (tgamma() and lgamma() in C++11)
220 Warning: unlike the actual standard C++ versions, the following
221 function only works on positive numbers (returns NaN if x <= 0).
222 Adapted from: http://www.johndcook.com/blog/cpp_gamma/
223
224 */
225
226 double lgamma(double x);
227
228 double tgamma(double x) {
229 if (x <= 0.0) return NaN;
230 static const double gamma = 0.577215664901532860606512090;
231 if (x < 1e-3) return 1.0 / (x * (1.0 + gamma * x));
232 if (x < 12.0) {
233 double y = x;
234 int n = 0;
235 bool arg_was_less_than_one = (y < 1.0);
236 if (arg_was_less_than_one) y += 1.0;
237 else y -= (n = static_cast<int>(floor(y)) - 1);
238 static const double p[] = {
239 -1.71618513886549492533811E+0, 2.47656508055759199108314E+1,
240 -3.79804256470945635097577E+2, 6.29331155312818442661052E+2,
241 8.66966202790413211295064E+2,-3.14512729688483675254357E+4,
242 -3.61444134186911729807069E+4, 6.64561438202405440627855E+4
243 };
244 static const double q[] = {
245 -3.08402300119738975254353E+1, 3.15350626979604161529144E+2,
246 -1.01515636749021914166146E+3,-3.10777167157231109440444E+3,
247 2.25381184209801510330112E+4, 4.75584627752788110767815E+3,
248 -1.34659959864969306392456E+5,-1.15132259675553483497211E+5
249 };
250 double num = 0.0, den = 1.0, z = y - 1;
251 for (int i = 0; i < 8; i++) {
252 num = (num + p[i]) * z;
253 den = den * z + q[i];
254 }
255 double result = num / den + 1.0;
256 if (arg_was_less_than_one) result /= (y - 1.0);
257 else for (int i = 0; i < n; i++) result *= y++;
258 return result;
259 }
260 return (x > 171.624) ? DBL_MAX * 2.0 : exp(lgamma(x));
261 }
262
263 double lgamma(double x) {
264 if (x <= 0.0) return NaN;
265 if (x < 12.0) return log(fabs(tgamma(x)));
266 static const double c[8] = {
267 1.0/12.0, -1.0/360.0, 1.0/1260.0, -1.0/1680.0, 1.0/1188.0,
268 -691.0/360360.0, 1.0/156.0, -3617.0/122400.0
269 };
270 double z = 1.0 / (x * x), sum = c[7];
271 for (int i = 6; i >= 0; i--) sum = sum * z + c[i];
272 static const double halflog2pi = 0.91893853320467274178032973640562;
273 return (x - 0.5) * log(x) - x + halflog2pi + sum / x;
274 }
275
276 /*
277
278 Base Conversion - O(N) on the number of digits
279
280 Given the digits of an integer x in base a, returns xs digits in base b.
281 Precondition: the base-10 value of x must be able to fit within an unsigned
282 long long. In other words, the value of x must be between 0 and 2^64 - 1.
283
284 Note: vector[0] stores the most significant digit in all usages below.
285
286 e.g. if x = {1, 2, 3} and a = 5 (i.e. x = 123 in base 5 = 38 in base 10),
287 then convert_base(x, 5, 3) returns {1, 1, 0, 2} (1102 in base 2).
288
289 */
290
291 std::vector<int> convert_base(const std::vector<int> & x, int a, int b) {
292 unsigned long long base10 = 0;
293 for (int i = 0; i < (int)x.size(); i++)
294 base10 += x[i] * pow(a, x.size() - i - 1);
295 int N = ceil(log(base10 + 1) / log(b));
296 std::vector<int> baseb;
297 for (int i = 1; i <= N; i++)
298 baseb.push_back(int(base10 / pow(b, N - i)) % b);
299 return baseb;
300 }
301
302 //returns digits of a number in base b
303 std::vector<int> base_digits(int x, int b = 10) {
304 std::vector<int> baseb;
305 while (x != 0) {
306 baseb.push_back(x % b);
307 x /= b;
308 }
309 std::reverse(baseb.begin(), baseb.end());
310 return baseb;
311 }
312
313 /*
314
315 Integer to Roman Numerals Conversion
316
317 Given an integer x, this function returns the Roman numeral representation
318 of x as a C++ string. More Ms are appended to the front of the resulting
319 string if x is greater than 1000. e.g. to_roman(1234) returns "MCCXXXIV"
320 and to_roman(5678) returns "MMMMMDCLXXVIII".
321
322 */
323
324 std::string to_roman(unsigned int x) {
325 static std::string h[] = {"","C","CC","CCC","CD","D","DC","DCC","DCCC","CM"};
326 static std::string t[] = {"","X","XX","XXX","XL","L","LX","LXX","LXXX","XC"};
327 static std::string o[] = {"","I","II","III","IV","V","VI","VII","VIII","IX"};
328 std::string res(x / 1000, M);
329 x %= 1000;
330 return res + h[x / 100] + t[x / 10 % 10] + o[x % 10];
331 }
332

334
339
340 int main() {
341 cout << "PI: " << PI << "\n";
342 cout << "E: " << E << "\n";
343 cout << "sqrt(2): " << root2 << "\n";
344 cout << "Golden ratio: " << phi << "\n";
345
346 //some properties of posinf, neginf, and NaN:
347 double x = -1234.567890; //any normal value of x will work
348 assert((posinf > x) && (neginf < x) && (posinf == -neginf));
349 assert((posinf + x == posinf) && (posinf - x == posinf));
350 assert((neginf + x == neginf) && (neginf - x == neginf));
351 assert((posinf + posinf == posinf) && (neginf - posinf == neginf));
352 assert((NaN != x) && (NaN != NaN) && (NaN != posinf) && (NaN != neginf));
353 assert(!(NaN < x) && !(NaN > x) && !(NaN <= x) && !(NaN >= x));
354 assert(isnan(0.0*posinf) && isnan(0.0*neginf) && isnan(posinf/neginf));
355 assert(isnan(NaN) && isnan(-NaN) && isnan(NaN*x + x - x/-NaN));
356 assert(isnan(neginf-neginf) && isnan(posinf-posinf) && isnan(posinf+neginf));
357 assert(!signbit(NaN) && signbit(-NaN) && !signbit(posinf) && signbit(neginf));
358
359 assert(copysign(1.0, +2.0) == +1.0 && copysign(posinf, -2.0) == neginf);
360 assert(copysign(1.0, -2.0) == -1.0 && signbit(copysign(NaN, -2.0)));
361 assert(sgn(-1.234) == -1 && sgn(0.0) == 0 && sgn(5678) == 1);
362
363 assert(EQ(floor0(1.5), 1.0) && EQ(floor0(-1.5), -1.0));
364 assert(EQ(ceil0(1.5), 2.0) && EQ(ceil0(-1.5), -2.0));
365 assert(EQ(roundhalfup(1.5), 2.0) && EQ(roundhalfup(-1.5), -1.0));
366 assert(EQ(roundhalfdown(1.5), 1.0) && EQ(roundhalfdown(-1.5), -2.0));
367 assert(EQ(roundhalfup0(1.5), 2.0) && EQ(roundhalfup0(-1.5), -2.0));
368 assert(EQ(roundhalfdown0(1.5), 1.0) && EQ(roundhalfdown0(-1.5), -1.0));
369 assert(EQ(roundhalfeven(1.5), 2.0) && EQ(roundhalfeven(-1.5), -2.0));
370 assert(NE(roundalternate(1.5), roundalternate(1.5)));
371 assert(EQ(roundplaces(-1.23456, 3, roundhalfdown0<double>), -1.235));
372
373 assert(EQ(erf(1.0), 0.8427007929) && EQ(erf(-1.0), -0.8427007929));
374 assert(EQ(tgamma(0.5), 1.7724538509) && EQ(tgamma(1.0), 1.0));
375 assert(EQ(lgamma(0.5), 0.5723649429) && EQ(lgamma(1.0), 0.0));
376
377 int base10digs[] = {1, 2, 3, 4, 5, 6}, a = 20, b = 10;
378 vector<int> basea = base_digits(123456, a);
379 vector<int> baseb = convert_base(basea, a, b);
380 assert(equal(baseb.begin(), baseb.end(), base10digs));
381
382 assert(to_roman(1234) == "MCCXXXIV");
383 assert(to_roman(5678) == "MMMMMDCLXXVIII");
384 return 0;
385 }
4.2 Combinatorics
4.2.1 Combinatorial Calculations
1 /*
2
3 The meanings of the following functions can respectively be
4 found with quick searches online. All of them computes the
5 answer modulo m, since contest problems typically ask us for
6 this due to the actual answer being potentially very large.
7 All functions using tables to generate every answer below
8 n and k can be optimized using recursion and memoization.
9
10 Note: The following are only defined for nonnegative inputs.
11
12 */
13
15
16 typedef std::vector<std::vector<long long> > table;
17
18 //n! mod m in O(n)
19 long long factorial(int n, int m = 1000000007) {
21 for (int i = 2; i <= n; i++) res = (res * i) % m;
22 return res % m;
23 }
24
25 //n! mod p, where p is a prime number, in O(p log n)
26 long long factorialp(long long n, long long p = 1000000007) {
27 long long res = 1, h;
28 while (n > 1) {
29 res = (res * ((n / p) % 2 == 1 ? p - 1 : 1)) % p;
30 h = n % p;
31 for (int i = 2; i <= h; i++) res = (res * i) % p;
32 n /= p;
33 }
34 return res % p;
35 }
36
37 //first n rows of pascals triangle (mod m) in O(n^2)
38 table binomial_table(int n, long long m = 1000000007) {
39 table t(n + 1);
40 for (int i = 0; i <= n; i++)
41 for (int j = 0; j <= i; j++)
42 if (i < 2 || j == 0 || i == j)
43 t[i].push_back(1);
44 else
45 t[i].push_back((t[i - 1][j - 1] + t[i - 1][j]) % m);
46 return t;
47 }
48
49 //if the product of two 64-bit ints (a*a, a*b, or b*b) can
50 //overflow, you must use mulmod (multiplication by adding)
51 long long powmod(long long a, long long b, long long m) {
52 long long x = 1, y = a;
53 for (; b > 0; b >>= 1) {
4.2. Combinatorics 163
54 if (b & 1) x = (x * y) % m;
55 y = (y * y) % m;
56 }
57 return x % m;
58 }
59
60 //n choose k (mod a prime number p) in O(min(k, n - k))
61 //powmod is used to find the mod inverse to get num / den % m
62 long long choose(int n, int k, long long p = 1000000007) {
63 if (n < k) return 0;
64 if (k > n - k) k = n - k;
65 long long num = 1, den = 1;
66 for (int i = 0; i < k; i++)
67 num = (num * (n - i)) % p;
68 for (int i = 1; i <= k; i++)
69 den = (den * i) % p;
70 return num * powmod(den, p - 2, p) % p;
71 }
72
73 //n multichoose k (mod a prime number p) in O(k)
74 long long multichoose(int n, int k, long long p = 1000000007) {
75 return choose(n + k - 1, k, p);
76 }
77
78 //n permute k (mod m) on O(k)
79 long long permute(int n, int k, long long m = 1000000007) {
80 if (n < k) return 0;
82 for (int i = 0; i < k; i++)
83 res = (res * (n - i)) % m;
84 return res % m;
85 }
86
87 //number of partitions of n (mod m) in O(n^2)
88 long long partitions(int n, long long m = 1000000007) {
89 std::vector<long long> p(n + 1, 0);
90 p[0] = 1;
91 for (int i = 1; i <= n; i++)
92 for (int j = i; j <= n; j++)
93 p[j] = (p[j] + p[j - i]) % m;
94 return p[n] % m;
95 }
96
97 //partitions of n into exactly k parts (mod m) in O(n * k)
98 long long partitions(int n, int k, long long m = 1000000007) {
99 table t(n + 1, std::vector<long long>(k + 1, 0));
100 t[0][1] = 1;
101 for (int i = 1; i <= n; i++)
102 for (int j = 1, h = k < i ? k : i; j <= h; j++)
103 t[i][j] = (t[i - 1][j - 1] + t[i - j][j]) % m;
104 return t[n][k] % m;
105 }
106
107 //unsigned Stirling numbers of the 1st kind (mod m) in O(n * k)
108 long long stirling1(int n, int k, long long m = 1000000007) {
110 t[0][0] = 1;
111 for (int i = 1; i <= n; i++)
112 for (int j = 1; j <= k; j++) {
113 t[i][j] = ((i - 1) * t[i - 1][j]) % m;

114 t[i][j] = (t[i][j] + t[i - 1][j - 1]) % m;
115 }
117 }
118
119 //Stirling numbers of the 2nd kind (mod m) in O(n * k)
120 long long stirling2(int n, int k, long long m = 1000000007) {
122 t[0][0] = 1;
123 for (int i = 1; i <= n; i++)
124 for (int j = 1; j <= k; j++) {
125 t[i][j] = (j * t[i - 1][j]) % m;
126 t[i][j] = (t[i][j] + t[i - 1][j - 1]) % m;
127 }
129 }
130
131 //Eulerian numbers of the 1st kind (mod m) in O(n * k)
132 //precondition: n > k
133 long long eulerian1(int n, int k, long long m = 1000000007) {
134 if (k > n - 1 - k) k = n - 1 - k;
136 for (int j = 1; j <= k; j++) t[0][j] = 0;
137 for (int i = 1; i <= n; i++)
138 for (int j = 1; j <= k; j++) {
139 t[i][j] = ((i - j) * t[i - 1][j - 1]) % m;
140 t[i][j] = (t[i][j] + ((j + 1) * t[i - 1][j]) % m) % m;
141 }
143 }
144
145 //Eulerian numbers of the 2nd kind (mod m) in O(n * k)
146 //precondition: n > k
147 long long eulerian2(int n, int k, long long m = 1000000007) {
149 for (int i = 1; i <= n; i++)
150 for (int j = 1; j <= k; j++) {
151 if (i == j) {
152 t[i][j] = 0;
153 } else {
154 t[i][j] = ((j + 1) * t[i - 1][j]) % m;
155 t[i][j] = (((2 * i - 1 - j) * t[i - 1][j - 1]) % m
156 + t[i][j]) % m;
157 }
158 }
160 }
161
162 //nth Catalan number (mod a prime number p) in O(n)
163 long long catalan(int n, long long p = 1000000007) {
164 return choose(2 * n, n, p) * powmod(n + 1, p - 2, p) % p;
165 }
166
168
172
173 int main() {
174 table t = binomial_table(10);
175 for (int i = 0; i < (int)t.size(); i++) {
176 for (int j = 0; j < (int)t[i].size(); j++)
177 cout << t[i][j] << " ";
178 cout << "\n";
179 }
180 assert(factorial(10) == 3628800);
181 assert(factorialp(123456) == 639390503);
182 assert(choose(20, 7) == 77520);
183 assert(multichoose(20, 7) == 657800);
184 assert(permute(10, 4) == 5040);
185 assert(partitions(4) == 5);
186 assert(partitions(100, 5) == 38225);
187 assert(stirling1(4, 2) == 11);
188 assert(stirling2(4, 3) == 6);
189 assert(eulerian1(9, 5) == 88234);
190 assert(eulerian2(8, 3) == 195800);
191 assert(catalan(10) == 16796);
192 return 0;
193 }
4.2.2 Enumerating Arrangements
1 /*
2
3 We shall consider an arrangement to be a permutation of
4 all the integers from 0 to n - 1. For our purposes, the
5 difference between an arrangement and a permutation is
6 simply that a permutation can pertain to a set of any
7 given values, not just distinct integers from 0 to n-1.
8
9 */
10
11 #include <algorithm> /* std::copy(), std::fill() */
13
14 /*
15
16 Changes a[] to the next lexicographically greater
17 permutation of any k distinct integers in range [0, n).
18 The values of a[] thats passed should be k distinct
19 integers, each in range [0, n).
20
21 returns: whether the function could rearrange a[] to
22 a lexicographically greater arrangement.
23
24 examples:
25 next_arrangement(4, 3, {0, 1, 2}) => 1, a[] = {0, 1, 3}
26 next_arrangement(4, 3, {0, 1, 3}) => 1, a[] = {0, 2, 1}
27 next_arrangement(4, 3, {3, 2, 1}) => 0, a[] unchanged
28
29 */
30
31 bool next_arrangement(int n, int k, int a[]) {
32 std::vector<bool> used(n);
33 for (int i = 0; i < k; i++) used[a[i]] = true;

34 for (int i = k - 1; i >= 0; i--) {
35 used[a[i]] = false;
36 for (int j = a[i] + 1; j < n; j++) {
37 if (!used[j]) {
38 a[i++] = j;
39 used[j] = true;
40 for (int x = 0; i < k; x++)
41 if (!used[x]) a[i++] = x;
42 return true;
43 }
44 }
45 }
46 return false;
47 }
48
49 /*
50
51 Computes n permute k using formula: nPk = n!/(n - k)!
52 Complexity: O(k). E.g. n_permute_k(10, 7) = 604800
53
54 */
55
56 long long n_permute_k(int n, int k) {
58 for (int i = 0; i < k; i++) res *= n - i;
59 return res;
60 }
61
62 /*
63
64 Given an integer rank x in range [0, n permute k), returns
65 a vector of integers representing the x-th lexicographically
66 smallest permutation of any k distinct integers in [0, n).
67
68 examples: arrangement_by_rank(4, 3, 0) => {0, 1, 2}
69 arrangement_by_rank(4, 3, 5) => {0, 3, 2}
70
71 */
72
73 std::vector<int> arrangement_by_rank(int n, int k, long long x) {
74 std::vector<int> free(n), res(k);
75 for (int i = 0; i < n; i++) free[i] = i;
76 for (int i = 0; i < k; i++) {
77 long long cnt = n_permute_k(n - 1 - i, k - 1 - i);
78 int pos = (int)(x / cnt);
79 res[i] = free[pos];
80 std::copy(free.begin() + pos + 1, free.end(),
81 free.begin() + pos);
82 x %= cnt;
83 }
84 return res;
85 }
86
87 /*
88
89 Given an array a[] of k integers each in range [0, n), returns
90 the (0-based) lexicographical rank (counting from least to
91 greatest) of the arrangement specified by a[] in all possible
92 permutations of k distinct integers in range [0, n).

93
94 examples: rank_by_arrangement(4, 3, {0, 1, 2}) => 0
95 rank_by_arrangement(4, 3, {0, 3, 2}) => 5
96
97 */
98
99 long long rank_by_arrangement(int n, int k, int a[]) {
101 std::vector<bool> used(n);
102 for (int i = 0; i < k; i++) {
103 int cnt = 0;
104 for (int j = 0; j < a[i]; j++)
105 if (!used[j]) cnt++;
106 res += n_permute_k(n - i - 1, k - i - 1) * cnt;
107 used[a[i]] = true;
108 }
109 return res;
110 }
111
112 /*
113
115 permutation of k (not-necessarily distinct) integers in
116 range [0, n). The values of a[] should be in range [0, n).
117 If a[] was interpreted as a base-n integer that is k digits
118 long, this function would be equivalent to incrementing a.
119 Ergo, there are n^k arrangements if repeats are allowed.
120
121 returns: whether the function could rearrange a[] to a
122 lexicographically greater arrangement with repeats.
123
124 examples:
125 n_a_w_r(4, 3, {0, 0, 0}) => 1, a[] = {0, 0, 1}
126 n_a_w_r(4, 3, {0, 1, 3}) => 1, a[] = {0, 2, 0}
127 n_a_w_r(4, 3, {3, 3, 3}) => 0, a[] unchanged
128
129 */
130
131 bool next_arrangement_with_repeats(int n, int k, int a[]) {
132 for (int i = k - 1; i >= 0; i--) {
133 if (a[i] < n - 1) {
134 a[i]++;
135 std::fill(a + i + 1, a + k, 0);
136 return true;
137 }
138 }
139 return false;
140 }
141
143
147
148 template<class it> void print(it lo, it hi) {
149 for (; lo != hi; ++lo) cout << *lo << " ";
150 cout << "\n";
151 }
152
153 int main() {
154 {
155 int n = 4, k = 3, a[] = {0, 1, 2};
156 cout << n << " permute " << k << " arrangements:\n";
157 int cnt = 0;
158 do {
159 print(a, a + k);
160 vector<int> b = arrangement_by_rank(n, k, cnt);
161 assert(equal(a, a + k, b.begin()));
162 assert(rank_by_arrangement(n, k, a) == cnt);
163 cnt++;
164 } while (next_arrangement(n, k, a));
165 cout << "\n";
166 }
167
168 {
169 int n = 4, k = 2, a[] = {0, 0};
170 cout << n << "^" << k << " arrangements with repeats:\n";
171 do {
173 } while (next_arrangement_with_repeats(n, k, a));
174 }
175 return 0;
176 }
4.2.3 Enumerating Permutations
1 /*
2
3 We shall consider a permutation of n objects to be an
4 ordered list of size n that contains all n elements,
5 where order is important. E.g. 1 1 2 0 and 0 1 2 1
6 are considered two different permutations of 0 1 1 2.
7 Compared to our prior definition of an arrangement, a
8 permutable range of size n may contain repeated values
9 of any type, not just the integers from 0 to n - 1.
10
11 */
12
13 #include <algorithm> /* copy, iter_swap, reverse, swap */
15
16 //identical to std::next_permutation()
17 template<class It> bool _next_permutation(It lo, It hi) {
18 if (lo == hi) return false;
19 It i = lo;
20 if (++i == hi) return false;
21 i = hi; --i;
22 for (;;) {
23 It j = i; --i;
24 if (*i < *j) {
25 It k = hi;
26 while (!(*i < *--k)) /* pass */;
27 std::iter_swap(i, k);
28 std::reverse(j, hi);
29 return true;
30 }
31 if (i == lo) {
32 std::reverse(lo, hi);
33 return false;
34 }
35 }
36 }
37
38 //array version
39 template<class T> bool next_permutation(int n, T a[]) {
40 for (int i = n - 2; i >= 0; i--)
41 if (a[i] < a[i + 1])
42 for (int j = n - 1; ; j--)
43 if (a[i] < a[j]) {
44 std::swap(a[i++], a[j]);
45 for (j = n - 1; i < j; i++, j--)
46 std::swap(a[i], a[j]);
47 return true;
48 }
49 return false;
50 }
51
52 /*
53
54 Calls the custom function f(vector) on all permutations
55 of the integers from 0 to n - 1. This is more efficient
56 than making many consecutive calls to next_permutation(),
57 however, here, the permutations will not be printed in
58 lexicographically increasing order.
59
60 */
61
63 void gen_permutations(int n, ReportFunction report,
64 std::vector<int> & p, int d) {
65 if (d == n) {
66 report(p);
67 return;
68 }
69 for (int i = 0; i < n; i++) {
70 if (p[i] == 0) {
71 p[i] = d;
72 gen_permutations(n, report, p, d + 1);
73 p[i] = 0;
74 }
75 }
76 }
77
79 void gen_permutations(int n, ReportFunction report) {
80 std::vector<int> perms(n, 0);
81 gen_permutations(n, report, perms, 0);
82 }
83
84 /*
85
86 Finds the next lexicographically greater permutation of
87 the binary digits of x. In other words, next_permutation()
88 simply returns the smallest integer greater than x which

89 has the same number of 1 bits (i.e. same popcount) as x.
90
91 examples: next_permutation(10101 base 2) = 10110
92 next_permutation(11100 base 2) = 100011
93
94 This can also be used to generate combinations as follows:
95 If we let k = popcount(x), then we can use this to generate
96 all possible masks to tell us which k items to take out of
97 n total items (represented by the first n bits of x).
98
99 */
100
101 long long next_permutation(long long x) {
102 long long s = x & -x, r = x + s;
103 return r | (((x ^ r) >> 2) / s);
104 }
105
106 /*
107
108 Given an integer rank x in range [0, n!), returns a vector
109 of integers representing the x-th lexicographically smallest
110 permutation of the integers in [0, n).
111
112 examples: permutation_by_rank(4, 0) => {0, 1, 2, 3}
113 permutation_by_rank(4, 5) => {0, 3, 2, 1}
114
115 */
116
117 std::vector<int> permutation_by_rank(int n, long long x) {
118 long long fact[n];
119 fact[0] = 1;
120 for (int i = 1; i < n; i++)
121 fact[i] = i * fact[i - 1];
122 std::vector<int> free(n), res(n);
123 for (int i = 0; i < n; i++) free[i] = i;
124 for (int i = 0; i < n; i++) {
125 int pos = x / fact[n - 1 - i];
126 res[i] = free[pos];
127 std::copy(free.begin() + pos + 1, free.end(),
128 free.begin() + pos);
129 x %= fact[n - 1 - i];
130 }
131 return res;
132 }
133
134 /*
135
136 Given an array a[] of n integers each in range [0, n), returns
138 greatest) of the arrangement specified by a[] in all possible
139 permutations of the integers from 0 to n - 1.
140
141 examples: rank_by_permutation(3, {0, 1, 2}) => 0
142 rank_by_permutation(3, {2, 1, 0}) => 5
143
144 */
145
146 template<class T> long long rank_by_permutation(int n, T a[]) {
147 long long fact[n];

148 fact[0] = 1;
149 for (int i = 1; i < n; i++)
150 fact[i] = i * fact[i - 1];
152 for (int i = 0; i < n; i++) {
153 int v = a[i];
154 for (int j = 0; j < i; j++)
155 if (a[j] < a[i]) v--;
156 res += v * fact[n - 1 - i];
157 }
158 return res;
159 }
160
161 /*
162
163 Given a permutation a[] of the integers from 0 to n - 1,
164 returns a decomposition of the permutation into cycles.
165 A permutation cycle is a subset of a permutation whose
166 elements trade places with one another. For example, the
167 permutation {0, 2, 1, 3} decomposes to {0, 3, 2} and {1}.
168 Here, the notation {0, 3, 2} means that starting from the
169 original ordering {0, 1, 2, 3}, the 0th value is replaced
170 by the 3rd, the 3rd by the 2nd, and the 2nd by the first,
171 See: http://mathworld.wolfram.com/PermutationCycle.html
172
173 */
174
175 typedef std::vector<std::vector<int> > cycles;
176
177 cycles decompose_into_cycles(int n, int a[]) {
178 std::vector<bool> vis(n);
179 cycles res;
180 for (int i = 0; i < n; i++) {
181 if (vis[i]) continue;
182 int j = i;
183 std::vector<int> cur;
184 do {
185 cur.push_back(j);
186 vis[j] = true;
187 j = a[j];
188 } while (j != i);
189 res.push_back(cur);
190 }
191 return res;
192 }
193
195
200
201 void printperm(const vector<int> & perm) {
202 for (int i = 0; i < (int)perm.size(); i++)
203 cout << perm[i] << " ";
204 cout << "\n";
205 }
206
209 cout << "\n";
210 }
211
212 int main() {
213 { //method 1: ordered
214 int n = 4, a[] = {0, 1, 2, 3};
215 int b[n], c[n];
216 for (int i = 0; i < n; i++) b[i] = c[i] = a[i];
217 cout << "Ordered permutations of 0 to " << n-1 << ":\n";
218 int cnt = 0;
219 do {
220 print(a, a + n);
221 assert(equal(b, b + n, a));
222 assert(equal(c, c + n, a));
223 vector<int> d = permutation_by_rank(n, cnt);
224 assert(equal(d.begin(), d.end(), a));
225 assert(rank_by_permutation(n, a) == cnt);
226 cnt++;
227 std::next_permutation(b, b + n);
228 _next_permutation(c, c + n);
229 } while (next_permutation(n, a));
230 cout << "\n";
231 }
232
233 { //method 2: unordered
234 int n = 3;
235 cout << "Unordered permutations of 0 to " << n-1 << ":\n";
236 gen_permutations(n, printperm);
237 cout << "\n";
238 }
239
240 { //permuting binary digits
241 const int n = 5;
242 cout << "Permutations of 2 zeros and 3 ones:\n";
243 long long lo = 7; // 00111 in base 2
244 long long hi = 35; //100011 in base 2
245 do {
246 cout << bitset<n>(lo).to_string() << "\n";
247 } while ((lo = next_permutation(lo)) != hi);
248 cout << "\n";
249 }
250
251 { //permutation cycles
252 int n = 4, a[] = {3, 1, 0, 2};
253 cout << "Decompose 0 2 1 3 into cycles:\n";
254 cycles c = decompose_into_cycles(n, a);
255 for (int i = 0; i < (int)c.size(); i++) {
256 cout << "Cycle " << i + 1 << ":";
257 for (int j = 0; j < (int)c[i].size(); j++)
258 cout << " " << c[i][j];
259 cout << "\n";
260 }
261 }
262 return 0;
263 }
4.2.4 Enumerating Combinations
1 /*
2
3 We shall consider a combination n choose k to be an
4 set of k elements chosen from a total of n elements.
5 Unlike n permute k, the order here doesnt matter.
6 That is, 0 1 2 is considered the same as 0 2 1, so
7 we will consider the sorted representation of each
8 combination for purposes of the functions below.
9
10 */
11
12 #include <algorithm> /* iter_swap, rotate, swap, swap_ranges */
15
16 /*
17
18 Rearranges the values in the range [lo, hi) such that
19 elements in the range [lo, mid) becomes the next
20 lexicographically greater combination of the values from
21 [lo, hi) than it currently is, and returns whether the
22 function could rearrange [lo, hi) to a lexicographically
23 greater combination. If the range [lo, hi) contains n
24 elements and the range [lo, mid) contains k elements,
25 then starting off with a sorted range [lo, hi) and
26 calling next_combination() repeatedly will return true
27 for n choose k iterations before returning false.
28
29 */
30
32 bool next_combination(It lo, It mid, It hi) {
33 if (lo == mid || mid == hi) return false;
34 It l(mid - 1), h(hi - 1);
35 int sz1 = 1, sz2 = 1;
36 while (l != lo && !(*l < *h)) --l, ++sz1;
37 if (l == lo && !(*l < *h)) {
38 std::rotate(lo, mid, hi);
39 return false;
40 }
41 for (; mid < h; ++sz2) if (!(*l < *--h)) { ++h; break; }
42 if (sz1 == 1 || sz2 == 1) {
43 std::iter_swap(l, h);
44 } else if (sz1 == sz2) {
45 std::swap_ranges(l, mid, h);
46 } else {
47 std::iter_swap(l, h);
48 ++l; ++h; --sz1; --sz2;
49 int total = sz1 + sz2, gcd = total;
50 for (int i = sz1; i != 0; ) std::swap(gcd %= i, i);
51 int skip = total / gcd - 1;
52 for (int i = 0; i < gcd; i++) {
53 It curr(i < sz1 ? l + i : h + (i - sz1));
54 int k = i;
55 typename std::iterator_traits<It>::value_type v(*curr);
56 for (int j = 0; j < skip; j++) {
57 k = (k + sz1) % total;
58 It next(k < sz1 ? l + k : h + (k - sz1));
59 *curr = *next;
60 curr = next;
61 }
62 *curr = v;
63 }
64 }
65 return true;
66 }
67
68 /*
69
71 combination of any k distinct integers in range [0, n).
72 The values of a[] thats passed should be k distinct
73 integers, each in range [0, n).
74
75 */
76
77 bool next_combination(int n, int k, int a[]) {
78 for (int i = k - 1; i >= 0; i--) {
79 if (a[i] < n - k + i) {
80 for (++a[i]; ++i < k; ) a[i] = a[i - 1] + 1;
81 return true;
82 }
83 }
84 return false;
85 }
86
87 /*
88
89 Finds the "mask" of the next combination of x. This is
90 equivalent to the next lexicographically greater permutation
91 of the binary digits of x. In other words, the function
92 simply returns the smallest integer greater than x which
93 has the same number of 1 bits (i.e. same popcount) as x.
94
95 examples: next_combination_mask(10101 base 2) = 10110
96 next_combination_mask(11100 base 2) = 100011
97
98 If we arbitrarily number the n items of our collection from
99 0 to n-1, then generating all combinations n choose k can
100 be done as follows: initialize x such that popcount(x) = k
101 and the first (least-significant) k bits are all set to 1
102 (e.g. to do 5 choose 3, start at x = 00111 (base 2) = 7).
103 Then, we repeatedly call x = next_combination_mask(x) until
104 we reach 11100 (the lexicographically greatest mask for 5
105 choose 3), after which we stop. At any point in the process,
106 we can say that the i-th item is being "taken" (0 <= i < n)
107 iff the i-th bit of x is set.
108
109 Note: this does not produce combinations in the same order
110 as next_combination, nor does it work if your n items have
111 repeated values (in that case, repeated combos will be
112 generated).
113
114 */
115
116 long long next_combination_mask(long long x) {

117 long long s = x & -x, r = x + s;
118 return r | (((x ^ r) >> 2) / s);
119 }
120
121 //n choose k in O(min(k, n - k))
122 long long n_choose_k(long long n, long long k) {
123 if (k > n - k) k = n - k;
125 for (int i = 0; i < k; i++)
126 res = res * (n - i) / (i + 1);
127 return res;
128 }
129
130 /*
131
132 Given an integer rank x in range [0, n choose k), returns
133 a vector of integers representing the x-th lexicographically
134 smallest combination k distinct integers in [0, n).
135
136 examples: combination_by_rank(4, 3, 0) => {0, 1, 2}
137 combination_by_rank(4, 3, 2) => {0, 2, 3}
138
139 */
140
141 std::vector<int> combination_by_rank(int n, int k, long long x) {
142 std::vector<int> res(k);
143 int cnt = n;
144 for (int i = 0; i < k; i++) {
145 int j = 1;
146 for (;; j++) {
147 long long am = n_choose_k(cnt - j, k - 1 - i);
148 if (x < am) break;
149 x -= am;
150 }
151 res[i] = i > 0 ? (res[i - 1] + j) : (j - 1);
152 cnt -= j;
153 }
154 return res;
155 }
156
157 /*
158
159 Given an array a[] of k integers each in range [0, n), returns
161 greatest) of the combination specified by a[] in all possible
162 combination of k distinct integers in range [0, n).
163
164 examples: rank_by_combination(4, 3, {0, 1, 2}) => 0
165 rank_by_combination(4, 3, {0, 2, 3}) => 2
166
167 */
168
169 long long rank_by_combination(int n, int k, int a[]) {
171 int prev = -1;
172 for (int i = 0; i < k; i++) {
173 for (int j = prev + 1; j < a[i]; j++)
174 res += n_choose_k(n - 1 - j, k - 1 - i);
175 prev = a[i];

176 }
177 return res;
178 }
179
180 /*
181
183 combination of any k (not necessarily distinct) integers
184 in range [0, n). The values of a[] thats passed should
185 be k integers, each in range [0, n). Note that there are
186 a total of n multichoose k combinations with repetition,
187 where n multichoose k = (n + k - 1) choose k
188
189 */
190
191 bool next_combination_with_repeats(int n, int k, int a[]) {
192 for (int i = k - 1; i >= 0; i--) {
193 if (a[i] < n - 1) {
194 for (++a[i]; ++i < k; ) a[i] = a[i - 1];
195 return true;
196 }
197 }
198 return false;
199 }
200
202
206
209 cout << "\n";
210 }
211
212 int main() {
213 { //like std::next_permutation(), repeats in the range allowed
214 int k = 3;
215 string s = "11234";
216 cout << s << " choose " << k << ":\n";
217 do {
218 cout << s.substr(0, k) << "\n";
219 } while (next_combination(s.begin(), s.begin() + k, s.end()));
220 cout << "\n";
221 }
222
223 { //unordered combinations with masks
224 int n = 5, k = 3;
225 string s = "abcde"; //must be distinct values
226 cout << s << " choose " << k << " with masks:\n";
227 long long mask = 0, dest = 0;
228 for (int i = 0; i < k; i++) mask |= 1 << i;
229 for (int i = k - 1; i < n; i++) dest |= 1 << i;
230 do {
231 for (int i = 0; i < n; i++)
232 if ((mask >> i) & 1) cout << s[i];
233 cout << "\n";
234 mask = next_combination_mask(mask);

235 } while (mask != dest);
236 cout << "\n";
237 }
238
239 { //only combinations of distinct integers from 0 to n - 1
240 int n = 5, k = 3, a[] = {0, 1, 2};
241 cout << n << " choose " << k << ":\n";
242 int cnt = 0;
243 do {
245 vector<int> b = combination_by_rank(n, k, cnt);
246 assert(equal(a, a + k, b.begin()));
247 assert(rank_by_combination(n, k, a) == cnt);
248 cnt++;
249 } while (next_combination(n, k, a));
250 cout << "\n";
251 }
252
253 { //combinations with repetition
254 int n = 3, k = 2, a[] = {0, 0};
255 cout << n << " multichoose " << k << ":\n";
256 do {
258 } while (next_combination_with_repeats(n, k, a));
259 }
260 return 0;
261 }
4.2.5 Enumerating Partitions
1 /*
2
3 We shall consider a partition of an integer n to be an
4 unordered multiset of positive integers that has a total
5 sum equal to n. Since both 2 1 1 and 1 2 1 represent the
6 same partition of 4, we shall consider only descending
7 sorted lists as "valid" partitions for functions below.
8
9 */
10
12
13 /*
14
15 Given a vector representing a partition of some
16 integer n (the sum of all values in the vector),
17 changes p to the next lexicographically greater
18 partition of n and returns whether the change was
19 successful (whether a lexicographically greater
20 partition existed). Note that the "initial" value
21 of p must be a vector of size n, all initialized 1.
22
23 e.g. next_partition({2, 1, 1}) => 1, p becomes {2, 2}
24 next_partition({2, 2}) => 1, p becomes {3, 1}
25 next_partition({4}) => 0, p is unchanged
26
27 */
28
29 bool next_partition(std::vector<int> & p) {
30 int n = p.size();
31 if (n <= 1) return false;
32 int s = p[n - 1] - 1, i = n - 2;
33 p.pop_back();
34 for (; i > 0 && p[i] == p[i - 1]; i--) {
35 s += p[i];
36 p.pop_back();
37 }
38 for (p[i]++; s-- > 0; ) p.push_back(1);
39 return true;
40 }
41
42 /* Returns the number of partitions of n. */
43
44 long long count_partitions(int n) {
45 std::vector<long long> p(n + 1, 0);
46 p[0] = 1;
47 for (int i = 1; i <= n; i++)
48 for (int j = i; j <= n; j++)
49 p[j] += p[j - i];
50 return p[n];
51 }
52
53 /* Helper function for partitioning by rank */
54
55 std::vector< std::vector<long long> >
56 p(1, std::vector<long long>(1, 1)); //memoization
57
58 long long partition_function(int a, int b) {
59 if (a >= (int)p.size()) {
60 int old = p.size();
61 p.resize(a + 1);
62 p[0].resize(a + 1);
63 for (int i = 1; i <= a; i++) {
64 p[i].resize(a + 1);
65 for (int j = old; j <= i; j++)
66 p[i][j] = p[i - 1][j - 1] + p[i - j][j];
67 }
68 }
69 return p[a][b];
70 }
71
72 /*
73
74 Given an integer n to partition and a 0-based rank x,
75 returns a vector of integers representing the x-th
76 lexicographically smallest partition of n (if values
77 in each partition were sorted in decreasing order).
78
79 examples: partition_by_rank(4, 0) => {1, 1, 1, 1}
80 partition_by_rank(4, 3) => {3, 1}
81
82 */
83
84 std::vector<int> partition_by_rank(int n, long long x) {
85 std::vector<int> res;
86 for (int i = n; i > 0; ) {

87 int j = 1;
88 for (;; j++) {
89 long long cnt = partition_function(i, j);
90 if (x < cnt) break;
91 x -= cnt;
92 }
93 res.push_back(j);
94 i -= j;
95 }
96 return res;
97 }
98
99 /*
100
101 Given a partition of an integer n (sum of all values
102 in vector p), returns a 0-based rank x of the partition
103 represented by p, considering partitions from least to
104 greatest in lexicographical order (if each partition
105 had values sorted in descending order).
106
107 examples: rank_by_partition({1, 1, 1, 1}) => 0
108 rank_by_partition({3, 1}) => 3
109
110 */
111
112 long long rank_by_partition(const std::vector<int> & p) {
114 int sum = 0;
115 for (int i = 0; i < (int)p.size(); i++) sum += p[i];
116 for (int i = 0; i < (int)p.size(); i++) {
117 for (int j = 0; j < p[i]; j++)
118 res += partition_function(sum, j);
119 sum -= p[i];
120 }
121 return res;
122 }
123
124 /*
125
126 Calls the custom function f(vector) on all partitions
127 which consist of strictly *increasing* integers.
128 This will exclude partitions such as {1, 1, 1, 1}.
129
130 */
131
133 void gen_increasing_partitons(int left, int prev, int i,
134 ReportFunction f, std::vector<int> & p) {
135 if (left == 0) {
136 //warning: slow constructor - modify accordingly
137 f(std::vector<int>(p.begin(), p.begin() + i));
138 return;
139 }
140 for (p[i] = prev + 1; p[i] <= left; p[i]++)
141 gen_increasing_partitons(left - p[i], p[i], i + 1, f, p);
142 }
143
145 void gen_increasing_partitons(int n, ReportFunction f) {

146 std::vector<int> partitions(n, 0);
147 gen_increasing_partitons(n, 0, 0, f, partitions);
148 }
149
151
155
156 void print(const vector<int> & v) {
158 cout << v[i] << " ";
159 cout << "\n";
160 }
161
162 int main() {
163 assert(count_partitions(5) == 7);
168
169 {
170 int n = 4;
171 vector<int> a(n, 1);
172 cout << "Partitions of " << n << ":\n";
173 int cnt = 0;
174 do {
175 print(a);
176 vector<int> b = partition_by_rank(n, cnt);
177 assert(equal(a.begin(), a.end(), b.begin()));
178 assert(rank_by_partition(a) == cnt);
179 cnt++;
180 } while (next_partition(a));
181 cout << "\n";
182 }
183
184 {
185 int n = 8;
186 cout << "Increasing partitions of " << n << ":\n";
187 gen_increasing_partitons(n, print);
188 }
189 return 0;
190 }
4.2.6 Enumerating Generic Combinatorial Sequences
1 /*
2
3 The follow provides a universal method for enumerating
4 abstract combinatorial sequences in O(n^2) time.
5
6 */
7
8 #include <vector>
9
10 class abstract_enumeration {
11 protected:
12 int range, length;
13
14 abstract_enumeration(int r, int l): range(r), length(l) {}
15
16 virtual long long count(const std::vector<int> & pre) {
17 return 0;
18 }
19
20 std::vector<int> next(std::vector<int> & seq) {
21 return from_number(to_number(seq) + 1);
22 }
23
24 long long total_count() {
25 return count(std::vector<int>(0));
26 }
27
28 public:
29 long long to_number(const std::vector<int> & seq) {
31 for (int i = 0; i < (int)seq.size(); i++) {
32 std::vector<int> pre(seq.begin(), seq.end());
33 pre.resize(i + 1);
34 for (pre[i] = 0; pre[i] < seq[i]; ++pre[i])
35 res += count(pre);
36 }
37 return res;
38 }
39
40 std::vector<int> from_number(long long x) {
41 std::vector<int> seq(length);
42 for (int i = 0; i < (int)seq.size(); i++) {
43 std::vector<int> pre(seq.begin(), seq.end());
44 pre.resize(i + 1);
45 for (pre[i] = 0; pre[i] < range; ++pre[i]) {
46 long long cur = count(pre);
47 if (x < cur) break;
48 x -= cur;
49 }
50 seq[i] = pre[i];
51 }
52 return seq;
53 }
54
56 void enumerate(ReportFunction report) {
57 long long total = total_count();
58 for (long long i = 0; i < total; i++) {
59 //assert(i == to_number(from_number(i));
60 report(from_number(i));
61 }
62 }
63 };
64
65 class arrangements: public abstract_enumeration {
66 public:
67 arrangements(int n, int k) : abstract_enumeration(n, k) {}
68
69 long long count(const std::vector<int> & pre) {
70 int sz = pre.size();
71 for (int i = 0; i < sz - 1; i++)
72 if (pre[i] == pre[sz - 1]) return 0;
74 for (int i = 0; i < length - sz; i++)
75 res *= range - sz - i;
76 return res;
77 }
78 };
79
80 class permutations: public arrangements {
81 public:
82 permutations(int n) : arrangements(n, n) {}
83 };
84
85 class combinations: public abstract_enumeration {
86 std::vector<std::vector<long long> > binomial;
87
88 public:
89 combinations(int n, int k) : abstract_enumeration(n, k),
90 binomial(n + 1, std::vector<long long>(n + 1, 0)) {
91 for (int i = 0; i <= n; i++)
92 for (int j = 0; j <= i; j++)
93 binomial[i][j] = (j == 0) ? 1 :
94 binomial[i - 1][j - 1] + binomial[i - 1][j];
95 }
96
98 int sz = pre.size();
99 if (sz >= 2 && pre[sz - 1] <= pre[sz - 2]) return 0;
100 int last = sz > 0 ? pre[sz - 1] : -1;
101 return binomial[range - 1 - last][length - sz];
102 }
103 };
104
105 class partitions: public abstract_enumeration {
106 std::vector<std::vector<long long> > p;
107
108 public:
109 partitions(int n) : abstract_enumeration(n + 1, n),
110 p(n + 1, std::vector<long long>(n + 1, 0)) {
111 std::vector<std::vector<long long> > pp(p);
112 pp[0][0] = 1;
113 for (int i = 1; i <= n; i++)
114 for (int j = 1; j <= i; j++)
115 pp[i][j] = pp[i - 1][j - 1] + pp[i - j][j];
116 for (int i = 1; i <= n; i++)
117 for (int j = 1; j <= n; j++)
118 p[i][j] = pp[i][j] + p[i][j - 1];
119 }
120
122 int size = pre.size(), sum = 0;
123 for (int i = 0; i < (int)pre.size(); i++) sum += pre[i];
124 if (sum == range - 1) return 1;
125 if (sum > range - 1 || (size > 0 && pre[size - 1] == 0) ||
126 (size >= 2 && pre[size - 1] > pre[size - 2])) return 0;
4.3. Number Theory 183
127 int last = size > 0 ? pre[size - 1] : range - 1;

128 return p[range - 1 - sum][last];
129 }
130 };
131
133
136
137 void print(const std::vector<int> & v) {
139 cout << v[i] << " ";
140 cout << "\n";
141 }
142
143 int main() {
144 cout << "Arrangement(3, 2):\n";
145 arrangements arrg(3, 2);
146 arrg.enumerate(print);
147
148 cout << "Permutation(3):\n";
149 permutations perm(3);
150 perm.enumerate(print);
151
152 cout << "Combination(4, 3):\n";
153 combinations comb(4, 3);
154 comb.enumerate(print);
155
156 cout << "Partition(4):\n";
157 partitions part(4);
158 part.enumerate(print);
159 return 0;
160 }
4.3 Number Theory
4.3.1 GCD, LCM, Mod Inverse, Chinese Remainder
1 /*
2
3 GCD, LCM, Modular Inverse, Chinese Remainder Theorem
4
5 */
6
8 #include <vector>
9
10 //C++98 does not have abs() declared for long long
11 template<class T> inline T _abs(const T & x) {
12 return x < 0 ? -x : x;
13 }
14
15 //GCD using Euclids algorithm - O(log(a + b))
16 template<class Int> Int gcd(Int a, Int b) {
17 return b == 0 ? _abs(a) : gcd(b, a % b);
18 }
19
20 //non-recursive version
21 template<class Int> Int gcd2(Int a, Int b) {
22 while (b != 0) {
23 Int t = b;
24 b = a % b;
25 a = t;
26 }
27 return _abs(a);
28 }
29
30 template<class Int> Int lcm(Int a, Int b) {
31 return _abs(a / gcd(a, b) * b);
32 }
33
34 //returns <gcd(a, b), <x, y>> such that gcd(a, b) = ax + by
35 template<class Int>
36 std::pair<Int, std::pair<Int, Int> > euclid(Int a, Int b) {
37 Int x = 1, y = 0, x1 = 0, y1 = 1;
38 //invariant: a = a * x + b * y, b = a * x1 + b * y1
39 while (b != 0) {
40 Int q = a / b, _x1 = x1, _y1 = y1, _b = b;
41 x1 = x - q * x1;
42 y1 = y - q * y1;
43 b = a - q * b;
44 x = _x1;
45 y = _y1;
46 a = _b;
47 }
48 return a > 0 ? std::make_pair(a, std::make_pair(x, y)) :
49 std::make_pair(-a, std::make_pair(-x, -y));
50 }
51
52 //recursive version
54 std::pair<Int, std::pair<Int, Int> > euclid2(Int a, Int b) {
55 if (b == 0) {
56 return a > 0 ? std::make_pair(a, std::make_pair(1, 0)) :
57 std::make_pair(-a, std::make_pair(-1, 0));
58 }
59 std::pair<Int, std::pair<Int, Int> > r = euclid2(b, a % b);
60 return std::make_pair(r.first, std::make_pair(r.second.second,
61 r.second.first - a / b * r.second.second));
62 }
63
64 /*
65
66 Modulo Operation - Euclidean Definition
67
68 The % operator in C/C++ returns the remainder of division (which
69 may be positive or negative) The true Euclidean definition of
70 modulo, however, defines the remainder to be always nonnegative.
71 For positive operators, % and mod are the same. But for negative
72 operands, they differ. The result here is consistent with the
73 Euclidean division algorithm.
74
75 e.g. -21 % 4 == -1 since -21 / 4 == -5 and 4 * -5 + (-1) == -21
76 however, -21 mod 4 is equal to 3 because -21 + 4 * 6 is 3.
77
78 */
79
80 template<class Int> Int mod(Int a, Int m) {
81 Int r = (Int)(a % m);
82 return r >= 0 ? r : r + m;
83 }
84
85 //returns x such that a * x = 1 (mod m)
86 //precondition: m > 0 && gcd(a, m) = 1
87 template<class Int> Int mod_inverse(Int a, Int m) {
88 a = mod(a, m);
89 return a == 0 ? 0 : mod((1 - m * mod_inverse(m % a, a)) / a, m);
90 }
91
92 //precondition: m > 0 && gcd(a, m) = 1
93 template<class Int> Int mod_inverse2(Int a, Int m) {
94 return mod(euclid(a, m).second.first, m);
95 }
96
97 //returns a vector where i*v[i] = 1 (mod p) in O(p) time
98 //precondition: p is prime
99 std::vector<int> generate_inverses(int p) {
100 std::vector<int> res(p);
101 res[1] = 1;
102 for (int i = 2; i < p; i++)
103 res[i] = (p - (p / i) * res[p % i] % p) % p;
104 return res;
105 }
106
107 /*
108
109 Chinese Remainder Theorem
110
111 Let r and s be positive integers which are relatively prime and
112 let a and b be any two integers. Then there exists an integer N
113 such that N = a (mod r) and N = b (mod s). Moreover, N is
114 uniquely determined modulo rs.
115
116 More generally, given a set of simultaneous congruences for
117 which all values in p[] are pairwise relative prime:
118
119 x = a[i] (mod p[i]), for i = 1..n
120
121 the solution of the set of congruences is:
122
123 x = a[1] * b[1] * (M/p[1]) + ... + a[n] * b[n] * (M/p[n]) (mod M)
124
125 where M = p[1] * p[2] ... * p[n] and the b[i] are determined for
126
127 b[i] * (M/p[i]) = 1 (mod p[i]).
128
129 The following functions solves for this value of x, with the
130 first function computed using the method above while the
131 second function using a special case of Garners algorithm.
132
133 http://e-maxx-eng.github.io/algebra/chinese-remainder-theorem.html
134
135 */
136
137 long long simple_restore(int n, int a[], int p[]) {
138 long long res = 0, m = 1;
139 for (int i = 0; i < n; i++) {
140 while (res % p[i] != a[i]) res += m;
141 m *= p[i];
142 }
143 return res;
144 }
145
146 long long garner_restore(int n, int a[], int p[]) {
147 int x[n];
148 for (int i = 0; i < n; i++) x[i] = a[i];
149 for (int i = 0; i < n; i++) {
150 for (int j = 0; j < i; j++)
151 x[i] = mod_inverse((long long)p[j], (long long)p[i]) *
152 (long long)(x[i] - x[j]);
153 x[i] = (x[i] % p[i] + p[i]) % p[i];
154 }
155 long long res = x[0], m = 1;
156 for (int i = 1; i < n; i++) {
157 m *= p[i - 1];
158 res += x[i] * m;
159 }
160 return res;
161 }
162
164
170
171 int main() {
172 {
173 srand(time(0));
174 for (int steps = 0; steps < 10000; steps++) {
175 int a = rand() % 200 - 10;
176 int b = rand() % 200 - 10;
177 int g1 = gcd(a, b), g2 = gcd2(a, b);
178 assert(g1 == g2);
179 if (g1 == 1 && b > 1) {
180 int inv1 = mod_inverse(a, b);
181 int inv2 = mod_inverse2(a, b);
182 assert(inv1 == inv2 && mod(a * inv1, b) == 1);
183 }
184 pair<int, pair<int, int> > euc1 = euclid(a, b);
185 pair<int, pair<int, int> > euc2 = euclid2(a, b);
186 assert(euc1.first == g1 && euc1 == euc2);
187 int x = euc1.second.first;
188 int y = euc1.second.second;
189 assert(g1 == a * x + b * y);
190 }
191 }
192
193 {
194 long long a = 6, b = 9;
195 pair<int, pair<int, int> > r = euclid(6, 9);

196 cout << r.second.first << " * (" << a << ")" << " + ";
197 cout << r.second.second << " * (" << b << ") = gcd(";
198 cout << a << "," << b << ") = " << r.first << "\n";
199 }
200
201 {
202 int prime = 17;
203 std::vector<int> res = generate_inverses(prime);
204 for (int i = 0; i < prime; i++) {
205 if (i > 0) assert(mod(i * res[i], prime) == 1);
206 cout << res[i] << " ";
207 }
208 cout << "\n";
209 }
210
211 {
212 int n = 3, a[] = {2, 3, 1}, m[] = {3, 4, 5};
213 //solves for x in the simultaneous congruences:
214 //x = 2 (mod 3)
215 //x = 3 (mod 4)
216 //x = 1 (mod 5)
217 int x1 = simple_restore(n, a, m);
218 int x2 = garner_restore(n, a, m);
219 assert(x1 == x2);
220 for (int i = 0; i < n; i++)
221 assert(mod(x1, m[i]) == a[i]);
222 cout << "Solution: " << x1 << "\n"; //11
223 }
224
225 return 0;
226 }
4.3.2 Generating Primes
1 /*
2
3 The following are three methods to generate primes.
4 Although the latter two functions are theoretically
5 linear, the former function with the sieve of
6 Eratosthenes is still significantly the fastest even
7 for n under 1 billion, since its constant factor is
8 so much better because of its minimal arithmetic
9 operations. For this reason, it should be favored
10 over the other two algorithms in most contest
11 applications. For the computation of larger primes,
12 you should replace int with long long or an arbitrary
13 precision class.
14
15 */
16
17 #include <cmath> /* ceil(), sqrt() */
19
20 //Sieve of Eratosthenes in ~ O(n log log n)
21 //returns: a vector of all primes under n
22 std::vector<int> gen_primes(int n) {
23 std::vector<bool> prime(n + 1, true);

24 int sqrtn = (int)ceil(sqrt(n));
25 for (int i = 2; i <= sqrtn; i++) {
26 if (prime[i])
27 for (int j = i * i; j <= n; j += i)
28 prime[j] = false;
29 }
31 for (int i = 2; i <= n; i++)
32 if (prime[i]) res.push_back(i);
33 return res;
34 }
35
36 //Technically O(n), but on -O2, this is about
37 //as fast as the above sieve for n = 100 million
38 std::vector<int> gen_primes_linear(int n) {
39 std::vector<int> lp(n + 1), res;
40 for (int i = 2; i <= n; i++) {
41 if (lp[i] == 0) {
42 lp[i] = i;
43 res.push_back(i);
44 }
45 for (int j = 0; j < (int)res.size(); j++) {
46 if (res[j] > lp[i] || i * res[j] > n)
47 break;
48 lp[i * res[j]] = res[j];
49 }
50 }
51 return res;
52 }
53
54 //Sieve of Atkins in O(n), somewhat slow due to
55 //its heavier arithmetic compared to the above
56 std::vector<int> gen_primes_atkins(int n) {
57 std::vector<bool> prime(n + 1, false);
59 prime[2] = true;
60 prime[3] = true;
61 int num, lim = ceil(sqrt(n));
62 for (int x = 1; x <= lim; x++) {
63 for (int y = 1; y <= lim; y++) {
64 num = 4 * x * x + y * y;
65 if (num <= n && (num % 12 == 1 || num % 12 == 5))
66 prime[num] = true;
67 num = 3 * x * x + y * y;
68 if (num <= n && (num % 12 == 7))
70 if (x > y) {
71 num = (3 * x * x - y * y);
72 if (num <= n && num % 12 == 11)
74 }
75 }
76 }
77 for (int i = 5; i <= lim; i++) {
78 if (prime[i])
79 for (int j = i * i; j <= n; j += i)
80 prime[j] = false;
81 }
82 for (int i = 2; i <= n; i++)

83 if (prime[i]) res.push_back(i);
84 return res;
85 }
86
87 //Double sieve to find primes in [l, h]
88 //Approximately O(sqrt(h) * log log(h - l))
89 std::vector<int> gen_primes(int l, int h) {
90 int sqrth = (int)ceil(sqrt(h));
91 int sqrtsqrth = (int)ceil(sqrt(sqrth));
92 std::vector<bool> prime1(sqrth + 1, true);
93 std::vector<bool> prime2(h - l + 1, true);
94 for (int i = 2; i <= sqrtsqrth; i++) {
95 if (prime1[i])
96 for (int j = i * i; j <= sqrth; j += i)
97 prime1[j] = false;
98 }
99 for (int i = 2, n = h - l; i <= sqrth; i++) {
100 if (prime1[i])
101 for (int j = l / i * i - l; j <= n; j += i)
102 if (j >= 0 && j + l != i)
103 prime2[j] = false;
104 }
106 for (int i = l > 1 ? l : 2; i <= h; i++)
107 if (prime2[i - l]) res.push_back(i);
108 return res;
109 }
110
112
117
118 template<class It> void print(It lo, It hi) {
119 while (lo != hi) cout << *(lo++) << " ";
120 cout << "\n";
121 }
122
123 int main() {
124 int pmax = 10000000;
125 vector<int> p;
126 time_t start;
127 double delta;
128
129 cout << "Generating primes up to " << pmax << "...\n";
130 start = clock();
131 p = gen_primes(pmax);
132 delta = (double)(clock() - start)/CLOCKS_PER_SEC;
133 cout << "gen_primes() took " << delta << "s.\n";
134
136 p = gen_primes_linear(pmax);
138 cout << "gen_primes_linear() took " << delta << "s.\n";
139
141 p = gen_primes_atkins(pmax);
143 cout << "gen_primes_atkins() took " << delta << "s.\n";
144
145 cout << "Generated " << p.size() << " primes.\n";
146 //print(p.begin(), p.end());
147
148 for (int i = 0; i <= 1000; i++) {
149 assert(gen_primes(i) == gen_primes_linear(i));
150 assert(gen_primes(i) == gen_primes_atkins(i));
151 }
152
153 int l = 1000000000, h = 1000000500;
154 cout << "Generating primes in [" << l << ", " << h << "]...\n";
156 p = gen_primes(l, h);
158 cout << "Generated " << p.size() << " primes in " << delta << "s.\n";
159 print(p.begin(), p.end());
160 return 0;
161 }
4.3.3 Primality Testing
1 /*
2
3 Primality Testing
4
5 */
6
9 #include <stdint.h> /* uint64_t */
10
11 /*
12
13 Trial division in O(sqrt(n)) to return whether n is prime
14 Applies an optimization based on the fact that all
15 primes greater than 3 take the form 6n + 1 or 6n - 1.
16
17 */
18
19 template<class Int> bool is_prime(Int n) {
20 if (n == 2 || n == 3) return true;
21 if (n < 2 || !(n % 2) || !(n % 3)) return false;
22 for (Int i = 5, w = 4; i * i <= n; i += (w = 6 - w))
23 if (n % i == 0) return false;
24 return true;
25 }
26
27 /*
28
29 Miller-Rabin Primality Test (Probabilistic)
30
31 Checks whether a number n is probably prime. If n is prime,
32 the function is guaranteed to return 1. If n is composite,
33 the function returns 1 with a probability of (1/4)^k,
34 where k is the number of iterations. With k = 1, the

35 probability of a composite being falsely predicted to be a
36 prime is 25%. If k = 5, the probability for this error is
37 just less than 0.1%. Thus, k = 18 to 20 is accurate enough
38 for most applications. All values of n < 2^63 is supported.
39
40 Complexity: O(k log^3(n)). In comparison to trial division,
41 the Miller-Rabin algorithm on 32-bit ints take ~45
42 operations for k = 10 iterations (~0.0001% error), while the
43 former takes ~10,000.
44
45 Warning: Due to the overflow of modular exponentiation,
46 this will only work on inputs less than 2^63.
47
48 */
49
50 uint64_t mulmod(uint64_t a, uint64_t b, uint64_t m) {
51 uint64_t x = 0, y = a % m;
52 for (; b > 0; b >>= 1) {
53 if (b & 1) x = (x + y) % m;
54 y = (y << 1) % m;
55 }
56 return x % m;
57 }
58
59 uint64_t powmod(uint64_t a, uint64_t b, uint64_t m) {
60 uint64_t x = 1, y = a;
61 for (; b > 0; b >>= 1) {
62 if (b & 1) x = mulmod(x, y, m);
63 y = mulmod(y, y, m);
64 }
65 return x % m;
66 }
67
68 //5 calls to rand() is unnecessary if RAND_MAX is 2^31-1
69 uint64_t rand64u() {
70 return ((uint64_t)(rand() & 0xf) << 60) |
71 ((uint64_t)(rand() & 0x7fff) << 45) |
72 ((uint64_t)(rand() & 0x7fff) << 30) |
73 ((uint64_t)(rand() & 0x7fff) << 15) |
74 ((uint64_t)(rand() & 0x7fff));
75 }
76
77 bool is_probable_prime(long long n, int k = 20) {
78 if (n < 2 || (n != 2 && !(n & 1))) return false;
79 uint64_t s = n - 1, p = n - 1, x, r;
80 while (!(s & 1)) s >>= 1;
81 for (int i = 0; i < k; i++) {
82 r = powmod(rand64u() % p + 1, s, n);
83 for (x = s; x != p && r != 1 && r != p; x <<= 1)
84 r = mulmod(r, r, n);
85 if (r != p && !(x & 1)) return false;
86 }
87 return true;
88 }
89
90 /*
91
92 Miller-Rabin - Deterministic for all unsigned long long
93
94 Although Miller-Rabin is generally probabilistic, the seven
95 bases 2, 325, 9375, 28178, 450775, 9780504, 1795265022 have
96 been proven to deterministically test the primality of all
97 numbers under 2^64. See: http://miller-rabin.appspot.com/
98
99 Complexity: O(log^3(n)).
100 Warning: Due to the overflow of modular exponentiation,
101 this will only work on inputs less than 2^63.
102
103 */
104
105 bool is_prime_fast(long long n) {
106 static const uint64_t witnesses[] =
107 {2, 325, 9375, 28178, 450775, 9780504, 1795265022};
108 if (n <= 1) return false;
109 if (n <= 3) return true;
110 if ((n & 1) == 0) return false;
111 uint64_t d = n - 1;
112 int s = 0;
113 for (; ~d & 1; s++) d >>= 1;
114 for (int i = 0; i < 7; i++) {
115 if (witnesses[i] > (uint64_t)n - 2) break;
116 uint64_t x = powmod(witnesses[i], d, n);
117 if (x == 1 || x == (uint64_t)n - 1) continue;
118 bool flag = false;
119 for (int j = 0; j < s; j++) {
120 x = powmod(x, 2, n);
121 if (x == 1) return false;
122 if (x == (uint64_t)n - 1) {
123 flag = true;
124 break;
125 }
126 }
127 if (!flag) return false;
128 }
129 return true;
130 }
131
133
135
136 int main() {
137 int len = 20;
138 unsigned long long v[] = {
139 0, 1, 2, 3, 4, 5, 11,
140 1000000ull,
141 772023803ull,
142 792904103ull,
143 813815117ull,
144 834753187ull,
145 855718739ull,
146 876717799ull,
147 897746119ull,
148 2147483647ull,
149 5705234089ull,
150 5914686649ull,
151 6114145249ull,
152 6339503641ull,
153 6548531929ull
154 };
155 for (int i = 0; i < len; i++) {
156 bool p = is_prime(v[i]);
157 assert(p == is_prime_fast(v[i]));
158 assert(p == is_probable_prime(v[i]));
159 }
160 return 0;
161 }
4.3.4 Integer Factorization
1 /*
2
3 Integer Factorization
4
5 */
6
10 #include <stdint.h> /* uint64_t */
12
13 /*
14
15 Trial division in O(sqrt(n))
16
17 Returns a vector of pair<prime divisor, exponent>
18 e.g. prime_factorize(15435) => {(3,2),(5,1),(7,3)}
19 because 3^2 * 5^1 * 7^3 = 15435
20
21 */
22
24 std::vector<std::pair<Int, int> > prime_factorize(Int n) {
25 std::vector<std::pair<Int, int> > res;
26 for (Int d = 2; ; d++) {
27 int power = 0, quot = n / d, rem = n - quot * d;
28 if (d > quot || (d == quot && rem > 0)) break;
29 for (; rem == 0; rem = n - quot * d) {
30 power++;
31 n = quot;
32 quot = n / d;
33 }
34 if (power > 0) res.push_back(std::make_pair(d, power));
35 }
36 if (n > 1) res.push_back(std::make_pair(n, 1));
37 return res;
38 }
39
40 /*
41
42 Trial division in O(sqrt(n))
43
44 Returns a sorted vector of all divisors of n.
45 e.g. get_all_divisors(28) => {1, 2, 4, 7, 14, 28}

46
47 */
48
50 std::vector<Int> get_all_divisors(Int n) {
51 std::vector<Int> res;
52 for (int d = 1; d * d <= n; d++) {
53 if (n % d == 0) {
54 res.push_back(d);
55 if (d * d != n)
56 res.push_back(n / d);
57 }
58 }
59 std::sort(res.begin(), res.end());
60 return res;
61 }
62
63 /*
64
65 Fermats Method ~ O(sqrt(N))
66
67 Given a number n, returns one factor of n that is
68 not necessary prime. Fermats algorithm is pretty
69 good when the number you wish to factor has two
70 factors very near to sqrt(n). Otherwise, it is just
71 as slow as the basic trial division algorithm.
72
73 e.g. 14917627 => 1 (its a prime), or
74 1234567 => 127 (because 127*9721 = 1234567)
75
76 */
77
78 long long fermat(long long n) {
79 if (n % 2 == 0) return 2;
80 long long x = sqrt(n), y = 0;
81 long long r = x * x - y * y - n;
82 while (r != 0) {
83 if (r < 0) {
84 r += x + x + 1;
85 x++;
86 } else {
87 r -= y + y + 1;
88 y++;
89 }
90 }
91 return x != y ? x - y : x + y;
92 }
93
94 /*
95
96 Pollards rho Algorithm with Brents Optimization
97
98 Brents algorithm is a much faster variant of Pollards
99 rho algorithm using Brents cycle-finding method. The
100 following function returns a (not necessarily prime) factor
101 of n, or n if n is prime. Note that this is not necessarily
102 guaranteed to always work perfectly. brent(9) may return 9
103 instead of 3. However, it works well when coupled with trial
104 division in the function prime_factorize_big() below.

105
106 */
107
108 uint64_t mulmod(uint64_t a, uint64_t b, uint64_t m) {
109 uint64_t x = 0, y = a % m;
110 for (; b > 0; b >>= 1) {
111 if (b & 1) x = (x + y) % m;
112 y = (y << 1) % m;
113 }
114 return x % m;
115 }
116
117 //5 calls to rand() is unnecessary if RAND_MAX is 2^31-1
118 uint64_t rand64u() {
119 return ((uint64_t)(rand() & 0xf) << 60) |
120 ((uint64_t)(rand() & 0x7fff) << 45) |
121 ((uint64_t)(rand() & 0x7fff) << 30) |
122 ((uint64_t)(rand() & 0x7fff) << 15) |
123 ((uint64_t)(rand() & 0x7fff));
124 }
125
126 uint64_t gcd(uint64_t a, uint64_t b) {
127 return b == 0 ? a : gcd(b, a % b);
128 }
129
130 long long brent(long long n) {
131 if (n % 2 == 0) return 2;
132 long long y = rand64u() % (n - 1) + 1;
133 long long c = rand64u() % (n - 1) + 1;
134 long long m = rand64u() % (n - 1) + 1;
135 long long g = 1, r = 1, q = 1, ys = 0, hi = 0, x = 0;
136 while (g == 1) {
137 x = y;
138 for (int i = 0; i < r; i++)
139 y = (mulmod(y, y, n) + c) % n;
140 for (long long k = 0; k < r && g == 1; k += m) {
141 ys = y;
142 hi = std::min(m, r - k);
143 for (int j = 0; j < hi; j++) {
144 y = (mulmod(y, y, n) + c) % n;
145 q = mulmod(q, x > y ? x - y : y - x, n);
146 }
147 g = gcd(q, n);
148 }
149 r *= 2;
150 }
151 if (g == n) do {
152 ys = (mulmod(ys, ys, n) + c) % n;
153 g = gcd(x > ys ? x - ys : ys - x, n);
154 } while (g <= 1);
155 return g;
156 }
157
158 /*
159
160 Combines Brents method with trial division to efficiently
161 generate the prime factorization of large integers.
162
163 Returns a vector of prime divisors that multiply to n.

164 e.g. prime_factorize(15435) => {3, 3, 5, 7, 7, 7}
165 because 3^2 * 5^1 * 7^3 = 15435
166
167 */
168
169 std::vector<long long> prime_factorize_big(long long n) {
170 if (n <= 0) return std::vector<long long>(0);
171 if (n == 1) return std::vector<long long>(1, 1);
172 std::vector<long long> res;
173 for (; n % 2 == 0; n /= 2) res.push_back(2);
174 for (; n % 3 == 0; n /= 3) res.push_back(3);
175 int mx = 1000000; //trial division for factors <= 1M
176 for (int i = 5, w = 2; i <= mx; i += w, w = 6 - w) {
177 for (; n % i == 0; n /= i) res.push_back(i);
178 }
179 for (long long p = 0, p1; n > mx; n /= p1) { //brent
180 for (p1 = n; p1 != p; p1 = brent(p)) p = p1;
181 res.push_back(p1);
182 }
183 if (n != 1) res.push_back(n);
184 sort(res.begin(), res.end());
185 return res;
186 }
187
189
194
195 template<class It> void print(It lo, It hi) {
196 while (lo != hi) cout << *(lo++) << " ";
197 cout << "\n";
198 }
199
200 template<class It> void printp(It lo, It hi) {
201 for (; lo != hi; ++lo)
202 cout << "(" << lo->first << "," << lo->second << ") ";
203 cout << "\n";
204 }
205
206 int main() {
207 srand(time(0));
208
209 vector< pair<int, int> > v1 = prime_factorize(15435);
210 printp(v1.begin(), v1.end());
211
212 vector<int> v2 = get_all_divisors(28);
213 print(v2.begin(), v2.end());
214
215 long long n = 1000003ll*100000037ll;
216 assert(fermat(n) == 1000003ll);
217
218 vector<long long> v3 = prime_factorize_big(n);
219 print(v3.begin(), v3.end());
220
221 return 0;
222 }
4.3.5 Eulers Totient Function
1 /*
2
3 Eulers totient function (or Eulers phi function) counts
4 the positive integers less than or equal to n that are
5 relatively prime to n. (These integers are sometimes
6 referred to as totatives of n.) Thus, phi(n) is the number
7 of integers k in the range [1, n] for which gcd(n, k) = 1.
8
9 E.g. if n = 9. Then gcd(9, 3) = gcd(9, 6) = 3 and gcd(9, 9)
10 = 9. The other six numbers in the range [1, 9], i.e. 1, 2,
11 4, 5, 7 and 8 are relatively prime to 9. Thus, phi(9) = 6.
12
13 */
14
16
17 int phi(int n) {
18 int res = n;
19 for (int i = 2; i * i <= n; i++)
20 if (n % i == 0) {
21 while (n % i == 0) n /= i;
22 res -= res / i;
23 }
24 if (n > 1) res -= res / n;
25 return res;
26 }
27
28 std::vector<int> phi_table(int n) {
29 std::vector<int> res(n + 1);
30 for (int i = 1; i <= n; i++)
31 res[i] = i;
32 for (int i = 1; i <= n; i++)
33 for (int j = i + i; j <= n; j += i)
34 res[j] -= res[i];
35 return res;
36 }
37
39
43
44 int main() {
45 cout << phi(1) << "\n"; //1
46 cout << phi(9) << "\n"; //6
47 cout << phi(1234567) << "\n"; //1224720
48
49 int n = 1000;
50 vector<int> v = phi_table(n);
51 for (int i = 0; i <= n; i++)
52 assert(v[i] == phi(i));
53 return 0;
54 }
4.4 Arbitrary Precision Arithmetic
4.4.1 Big Integers (Simple)
1 /*
2
3 Description: Integer arbitrary precision functions.
4 To use, pass bigints to the functions by addresses.
5 e.g. add(&a, &b, &c) stores the sum of a and b into c.
6
7 Complexity: comp(), to_string(), digit_shift(), add(),
8 and sub() are O(N) on the number of digits. mul() and
9 div() are O(N^2). zero_justify() is amortized constant.
10
11 */
12
14
15 struct bigint {
16 static const int maxdigits = 1000;
17
18 char dig[maxdigits], sign;
19 int last;
20
21 bigint(long long x = 0): sign(x < 0 ? -1 : 1) {
22 for (int i = 0; i < maxdigits; i++) dig[i] = 0;
23 if (x == 0) { last = 0; return; }
24 if (x < 0) x = -x;
25 for (last = -1; x > 0; x /= 10) dig[++last] = x % 10;
26 }
27
28 bigint(const std::string & s): sign(s[0] == - ? -1 : 1) {
29 for (int i = 0; i < maxdigits; i++) dig[i] = 0;
30 last = -1;
31 for (int i = s.size() - 1; i >= 0; i--)
32 dig[++last] = (s[i] - 0);
33 if (dig[last] + 0 == -) dig[last--] = 0;
34 }
35 };
36
37 void zero_justify(bigint * x) {
38 while (x->last > 0 && !x->dig[x->last]) x->last--;
39 if (x->last == 0 && x->dig[0] == 0) x->sign = 1;
40 }
41
42 void add(bigint * a, bigint * b, bigint * c);
43 void sub(bigint * a, bigint * b, bigint * c);
44
45 //returns: -1 if a < b, 0 if a == b, or 1 if a > b
46 int comp(bigint * a, bigint * b) {
47 if (a->sign != b->sign) return b->sign;
48 if (b->last > a->last) return a->sign;
49 if (a->last > b->last) return -a->sign;
50 for (int i = a->last; i >= 0; i--) {
4.4. Arbitrary Precision Arithmetic 199
51 if (a->dig[i] > b->dig[i]) return -a->sign;

52 if (b->dig[i] > a->dig[i]) return a->sign;
53 }
54 return 0;
55 }
56
57 void add(bigint * a, bigint * b, bigint * c) {
58 if (a->sign != b->sign) {
59 if (a->sign == -1)
60 a->sign = 1, sub(b, a, c), a->sign = -1;
61 else
62 b->sign = 1, sub(a, b, c), b->sign = -1;
63 return;
64 }
65 c->sign = a->sign;
66 c->last = (a->last > b->last ? a->last : b->last) + 1;
67 for (int i = 0, carry = 0; i <= c->last; i++) {
68 c->dig[i] = (carry + a->dig[i] + b->dig[i]) % 10;
69 carry = (carry + a->dig[i] + b->dig[i]) / 10;
70 }
71 zero_justify(c);
72 }
73
74 void sub(bigint * a, bigint * b, bigint * c) {
75 if (a->sign == -1 || b->sign == -1) {
76 b->sign *= -1, add(a, b, c), b->sign *= -1;
77 return;
78 }
79 if (comp(a, b) == 1) {
80 sub(b, a, c), c->sign = -1;
81 return;
82 }
83 c->last = (a->last > b->last) ? a->last : b->last;
84 for (int i = 0, borrow = 0, v; i <= c->last; i++) {
85 v = a->dig[i] - borrow;
86 if (i <= b->last) v -= b->dig[i];
87 if (a->dig[i] > 0) borrow = 0;
88 if (v < 0) v += 10, borrow = 1;
89 c->dig[i] = v % 10;
90 }
91 zero_justify(c);
92 }
93
94 void digit_shift(bigint * x, int n) {
95 if (!x->last && !x->dig[0]) return;
96 for (int i = x->last; i >= 0; i--)
97 x->dig[i + n] = x->dig[i];
98 for (int i = 0; i < n; i++) x->dig[i] = 0;
99 x->last += n;
100 }
101
102 void mul(bigint * a, bigint * b, bigint * c) {
103 bigint row = *a, tmp;
104 for (int i = 0; i <= b->last; i++) {
105 for (int j = 1; j <= b->dig[i]; j++) {
106 add(c, &row, &tmp);
107 *c = tmp;
108 }
109 digit_shift(&row, 1);
110 }
111 c->sign = a->sign * b->sign;
112 zero_justify(c);
113 }
114
115 void div(bigint * a, bigint * b, bigint * c) {
116 bigint row, tmp;
117 int asign = a->sign, bsign = b->sign;
118 a->sign = b->sign = 1;
119 c->last = a->last;
120 for (int i = a->last; i >= 0; i--) {
121 digit_shift(&row, 1);
122 row.dig[0] = a->dig[i];
123 c->dig[i] = 0;
124 for (; comp(&row, b) != 1; row = tmp) {
125 c->dig[i]++;
126 sub(&row, b, &tmp);
127 }
128 }
129 c->sign = (a->sign = asign) * (b->sign = bsign);
130 zero_justify(c);
131 }
132
133 std::string to_string(bigint * x) {
134 std::string s(x->sign == -1 ? "-" : "");
135 for (int i = x->last; i >= 0; i--)
136 s += (char)(0 + x->dig[i]);
137 return s;
138 }
139
141
143
144 int main() {
145 bigint a("-9899819294989142124"), b("12398124981294214");
146 bigint sum; add(&a, &b, &sum);
147 bigint dif; sub(&a, &b, &dif);
148 bigint prd; mul(&a, &b, &prd);
149 bigint quo; div(&a, &b, &quo);
150 assert(to_string(&sum) == "-9887421170007847910");
151 assert(to_string(&dif) == "-9912217419970436338");
152 assert(to_string(&prd) == "-122739196911503356525379735104870536");
153 assert(to_string(&quo) == "-798");
154 return 0;
155 }
4.4.2 Big Integer and Rational Class
1 /*
2
3 The following bigint class is implemented by storing "chunks"
4 of the big integer in a large base that is a power of 10 so
5 it can be efficiently stored, operated on, and printed.
6
7 It has extensive features including karatsuba multiplication,
8 exponentiation by squaring, and n-th root using binary search.
9 The class is thoroughly templatized, so you can use it as

10 easily as you do for normal ints. For example, you may use
11 operators with a bigint and a string (e.g. bigint(1234)+"-567"
12 and the result will be correctly promoted to a bigint that has
13 a value of 667). I/O is done using <iostream>. For example:
14 bigint a, b; cin >> a >> b; cout << a + b << "\n";
15 adds two integers together and prints the result, just as you
16 would expect for a normal int, except with arbitrary precision.
17 The class also supports other streams such as fstream.
18
19 After the bigint class, a class for rational numbers is
20 implemented, using two bigints to store its numerators and
21 denominators. It is useful for when exact results of division
22 operations are needed.
23
24 */
25
26 #include <algorithm> /* std::max(), std::swap() */
28 #include <cstdlib> /* rand() */
29 #include <iomanip> /* std::setw(), std::setfill() */
30 #include <istream>
31 #include <ostream>
32 #include <sstream>
37
38 struct bigint {
39 //base should be a power of 10 for I/O to work
40 //base and base_digits should be consistent
41 static const int base = 1000000000, base_digits = 9;
42
43 typedef std::vector<int> vint;
44 typedef std::vector<long long> vll;
45
46 vint a; //a[0] stores right-most (least significant) base-digit
47 int sign;
48
49 bigint() : sign(1) {}
50 bigint(int v) { *this = (long long)v; }
51 bigint(long long v) { *this = v; }
52 bigint(const std::string & s) { read(s); }
53 bigint(const char * s) { read(std::string(s)); }
54
55 void trim() {
56 while (!a.empty() && a.back() == 0) a.pop_back();
57 if (a.empty()) sign = 1;
58 }
59
60 void read(const std::string & s) {
61 sign = 1;
62 a.clear();
63 int pos = 0;
64 while (pos < (int)s.size() && (s[pos] == - || s[pos] == +)) {
65 if (s[pos] == -) sign = -sign;
66 pos++;
67 }
68 for (int i = s.size() - 1; i >= pos; i -= base_digits) {

69 int x = 0;
70 for (int j = std::max(pos, i - base_digits + 1); j <= i; j++)
71 x = x * 10 + s[j] - 0;
72 a.push_back(x);
73 }
74 trim();
75 }
76
77 void operator = (const bigint & v) {
78 sign = v.sign;
79 a = v.a;
80 }
81
82 void operator = (long long v) {
83 sign = 1;
84 if (v < 0) sign = -1, v = -v;
85 a.clear();
86 for (; v > 0; v /= base) a.push_back(v % base);
87 }
88
89 bigint operator + (const bigint & v) const {
90 if (sign == v.sign) {
91 bigint res = v;
92 int carry = 0;
93 for (int i = 0; i < (int)std::max(a.size(), v.a.size()) || carry; i++) {
94 if (i == (int)res.a.size()) res.a.push_back(0);
95 res.a[i] += carry + (i < (int)a.size() ? a[i] : 0);
96 carry = res.a[i] >= base;
97 if (carry) res.a[i] -= base;
98 }
99 return res;
100 }
101 return *this - (-v);
102 }
103
104 bigint operator - (const bigint & v) const {
105 if (sign == v.sign) {
106 if (abs() >= v.abs()) {
107 bigint res(*this);
108 for (int i = 0, carry = 0; i < (int)v.a.size() || carry; i++) {
109 res.a[i] -= carry + (i < (int)v.a.size() ? v.a[i] : 0);
110 carry = res.a[i] < 0;
111 if (carry) res.a[i] += base;
112 }
113 res.trim();
114 return res;
115 }
116 return -(v - *this);
117 }
118 return *this + (-v);
119 }
120
121 void operator *= (int v) {
122 if (v < 0) sign = -sign, v = -v;
123 for (int i = 0, carry = 0; i < (int)a.size() || carry; i++) {
124 if (i == (int)a.size()) a.push_back(0);
125 long long cur = a[i] * (long long)v + carry;
126 carry = (int)(cur / base);
127 a[i] = (int)(cur % base);

128 //asm("divl %%ecx" : "=a"(carry), "=d"(a[i]) : "A"(cur), "c"(base));
129 }
130 trim();
131 }
132
133 bigint operator * (int v) const {
135 res *= v;
136 return res;
137 }
138
139 static vint convert_base(const vint & a, int l1, int l2) {
140 vll p(std::max(l1, l2) + 1);
141 p[0] = 1;
142 for (int i = 1; i < (int)p.size(); i++) p[i] = p[i - 1] * 10;
143 vint res;
144 long long cur = 0;
145 for (int i = 0, cur_digits = 0; i < (int)a.size(); i++) {
146 cur += a[i] * p[cur_digits];
147 cur_digits += l1;
148 while (cur_digits >= l2) {
149 res.push_back((int)(cur % p[l2]));
150 cur /= p[l2];
151 cur_digits -= l2;
152 }
153 }
154 res.push_back((int)cur);
155 while (!res.empty() && res.back() == 0) res.pop_back();
156 return res;
157 }
158
159 //complexity: O(3N^log2(3)) ~ O(3N^1.585)
160 static vll karatsuba_multiply(const vll & a, const vll & b) {
161 int n = a.size();
162 vll res(n + n);
163 if (n <= 32) {
164 for (int i = 0; i < n; i++)
165 for (int j = 0; j < n; j++)
166 res[i + j] += a[i] * b[j];
167 return res;
168 }
169 int k = n >> 1;
170 vll a1(a.begin(), a.begin() + k), a2(a.begin() + k, a.end());
171 vll b1(b.begin(), b.begin() + k), b2(b.begin() + k, b.end());
172 vll a1b1 = karatsuba_multiply(a1, b1);
173 vll a2b2 = karatsuba_multiply(a2, b2);
174 for (int i = 0; i < k; i++) a2[i] += a1[i];
175 for (int i = 0; i < k; i++) b2[i] += b1[i];
176 vll r = karatsuba_multiply(a2, b2);
177 for (int i = 0; i < (int)a1b1.size(); i++) r[i] -= a1b1[i];
178 for (int i = 0; i < (int)a2b2.size(); i++) r[i] -= a2b2[i];
179 for (int i = 0; i < (int)r.size(); i++) res[i + k] += r[i];
180 for (int i = 0; i < (int)a1b1.size(); i++) res[i] += a1b1[i];
181 for (int i = 0; i < (int)a2b2.size(); i++) res[i + n] += a2b2[i];
182 return res;
183 }
184
185 bigint operator * (const bigint & v) const {
186 //if really big values cause overflow, use smaller _base
187 static const int _base = 10000, _base_digits = 4;
188 vint _a = convert_base(this->a, base_digits, _base_digits);
189 vint _b = convert_base(v.a, base_digits, _base_digits);
190 vll a(_a.begin(), _a.end());
191 vll b(_b.begin(), _b.end());
192 while (a.size() < b.size()) a.push_back(0);
193 while (b.size() < a.size()) b.push_back(0);
194 while (a.size() & (a.size() - 1)) {
195 a.push_back(0);
196 b.push_back(0);
197 }
198 vll c = karatsuba_multiply(a, b);
199 bigint res;
200 res.sign = sign * v.sign;
201 for (int i = 0, carry = 0; i < (int)c.size(); i++) {
202 long long cur = c[i] + carry;
203 res.a.push_back((int)(cur % _base));
204 carry = (int)(cur / _base);
205 }
206 res.a = convert_base(res.a, _base_digits, base_digits);
207 res.trim();
208 return res;
209 }
210
211 bigint operator ^ (const bigint & v) const {
212 if (v.sign == -1) return bigint(0);
213 bigint x(*this), n(v), res(1);
214 while (!n.is_zero()) {
215 if (n.a[0] % 2 == 1) res *= x;
216 x *= x;
217 n /= 2;
218 }
219 return res;
220 }
221
222 friend std::pair<bigint, bigint> divmod(const bigint & a1, const bigint & b1) {
223 int norm = base / (b1.a.back() + 1);
224 bigint a = a1.abs() * norm;
225 bigint b = b1.abs() * norm;
226 bigint q, r;
227 q.a.resize(a.a.size());
228 for (int i = a.a.size() - 1; i >= 0; i--) {
229 r *= base;
230 r += a.a[i];
231 int s1 = r.a.size() <= b.a.size() ? 0 : r.a[b.a.size()];
232 int s2 = r.a.size() <= b.a.size() - 1 ? 0 : r.a[b.a.size() - 1];
233 int d = ((long long)base * s1 + s2) / b.a.back();
234 for (r -= b * d; r < 0; r += b) d--;
235 q.a[i] = d;
236 }
237 q.sign = a1.sign * b1.sign;
238 r.sign = a1.sign;
239 q.trim();
240 r.trim();
241 return std::make_pair(q, r / norm);
242 }
243
244 bigint operator / (const bigint & v) const { return divmod(*this, v).first; }
245 bigint operator % (const bigint & v) const { return divmod(*this, v).second; }
246
247 bigint & operator /= (int v) {
248 if (v < 0) sign = -sign, v = -v;
249 for (int i = a.size() - 1, rem = 0; i >= 0; i--) {
250 long long cur = a[i] + rem * (long long)base;
251 a[i] = (int)(cur / v);
252 rem = (int)(cur % v);
253 }
254 trim();
255 return *this;
256 }
257
258 bigint operator / (int v) const {
260 res /= v;
261 return res;
262 }
263
264 int operator % (int v) const {
265 if (v < 0) v = -v;
266 int m = 0;
267 for (int i = a.size() - 1; i >= 0; i--)
268 m = (a[i] + m * (long long)base) % v;
269 return m * sign;
270 }
271
272 bigint operator ++(int) { bigint t(*this); operator++(); return t; }
273 bigint operator --(int) { bigint t(*this); operator--(); return t; }
274 bigint & operator ++() { *this = *this + bigint(1); return *this; }
275 bigint & operator --() { *this = *this - bigint(1); return *this; }
276 bigint & operator += (const bigint & v) { *this = *this + v; return *this; }
277 bigint & operator -= (const bigint & v) { *this = *this - v; return *this; }
278 bigint & operator *= (const bigint & v) { *this = *this * v; return *this; }
279 bigint & operator /= (const bigint & v) { *this = *this / v; return *this; }
280 bigint & operator %= (const bigint & v) { *this = *this % v; return *this; }
281 bigint & operator ^= (const bigint & v) { *this = *this ^ v; return *this; }
282
283 bool operator < (const bigint & v) const {
284 if (sign != v.sign) return sign < v.sign;
285 if (a.size() != v.a.size())
286 return a.size() * sign < v.a.size() * v.sign;
287 for (int i = a.size() - 1; i >= 0; i--)
288 if (a[i] != v.a[i])
289 return a[i] * sign < v.a[i] * sign;
290 return false;
291 }
292
293 bool operator > (const bigint & v) const { return v < *this; }
294 bool operator <= (const bigint & v) const { return !(v < *this); }
295 bool operator >= (const bigint & v) const { return !(*this < v); }
296 bool operator == (const bigint & v) const { return !(*this < v) && !(v < *this); }
297 bool operator != (const bigint & v) const { return *this < v || v < *this; }
298
299 int size() const {
300 if (a.empty()) return 1;
301 std::ostringstream oss;
302 oss << a.back();
303 return oss.str().length() + base_digits*(a.size() - 1);
304 }
305
306 bool is_zero() const {
307 return a.empty() || (a.size() == 1 && !a[0]);
308 }
309
310 bigint operator - () const {
312 res.sign = -sign;
313 return res;
314 }
315
316 bigint abs() const {
318 res.sign *= res.sign;
319 return res;
320 }
321
322 friend bigint abs(const bigint & a) {
323 return a.abs();
324 }
325
326 friend bigint gcd(const bigint & a, const bigint & b) {
327 return b.is_zero() ? a : gcd(b, a % b);
328 }
329
330 friend bigint lcm(const bigint & a, const bigint & b) {
331 return a / gcd(a, b) * b;
332 }
333
334 friend bigint sqrt(const bigint & x) {
335 bigint a = x;
336 while (a.a.empty() || a.a.size() % 2 == 1) a.a.push_back(0);
337 int n = a.a.size();
338 int firstdig = sqrt((double)a.a[n - 1] * base + a.a[n - 2]);
339 int norm = base / (firstdig + 1);
340 a *= norm;
341 a *= norm;
342 while (a.a.empty() || a.a.size() % 2 == 1) a.a.push_back(0);
343 bigint r = (long long)a.a[n - 1] * base + a.a[n - 2];
344 firstdig = sqrt((double)a.a[n - 1] * base + a.a[n - 2]);
345 int q = firstdig;
346 bigint res;
347 for (int j = n / 2 - 1; j >= 0; j--) {
348 for (;; q--) {
349 bigint r1 = (r - (res * 2 * base + q) * q) * base * base + (j > 0 ?
350 (long long)a.a[2 * j - 1] * base + a.a[2 * j - 2] : 0);
351 if (r1 >= 0) {
352 r = r1;
353 break;
354 }
355 }
356 res = (res * base) + q;
357 if (j > 0) {
358 int d1 = res.a.size() + 2 < r.a.size() ? r.a[res.a.size() + 2] : 0;
359 int d2 = res.a.size() + 1 < r.a.size() ? r.a[res.a.size() + 1] : 0;
360 int d3 = res.a.size() < r.a.size() ? r.a[res.a.size()] : 0;
361 q = ((long long)d1*base*base + (long long)d2*base + d3)/(firstdig * 2);
362 }
363 }
364 res.trim();
365 return res / norm;
366 }
367
368 friend bigint nthroot(const bigint & x, const bigint & n) {
369 bigint hi = 1;
370 while ((hi ^ n) <= x) hi *= 2;
371 bigint lo = hi / 2, mid, midn;
373 mid = (lo + hi) / 2;
374 midn = mid ^ n;
375 if (lo < mid && midn < x) {
376 lo = mid;
377 } else if (mid < hi && x < midn) {
378 hi = mid;
379 } else {
380 return mid;
381 }
382 }
383 return mid + 1;
384 }
385
386 friend std::istream & operator >> (std::istream & in, bigint & v) {
387 std::string s;
388 in >> s;
389 v.read(s);
390 return in;
391 }
392
393 friend std::ostream & operator << (std::ostream & out, const bigint & v) {
394 if (v.sign == -1) out << -;
395 out << (v.a.empty() ? 0 : v.a.back());
396 for (int i = v.a.size() - 2; i >= 0; i--)
397 out << std::setw(base_digits) << std::setfill(0) << v.a[i];
398 return out;
399 }
400
401 std::string to_string() const {
403 if (sign == -1) oss << -;
404 oss << (a.empty() ? 0 : a.back());
405 for (int i = a.size() - 2; i >= 0; i--)
406 oss << std::setw(base_digits) << std::setfill(0) << a[i];
407 return oss.str();
408 }
409
410 long long to_llong() const {
412 for (int i = a.size() - 1; i >= 0; i--)
413 res = res * base + a[i];
414 return res * sign;
415 }
416
417 double to_double() const {
418 std::stringstream ss(to_string());
419 double res;
420 ss >> res;
421 return res;
422 }
423
424 long double to_ldouble() const {
425 std::stringstream ss(to_string());
426 long double res;
427 ss >> res;
428 return res;
429 }
430
431 static bigint rand(int len) {
432 if (len == 0) return bigint(0);
433 std::string s(1, 1 + (::rand() % 9));
434 for (int i = 1; i < len; i++) s += 0 + (::rand() % 10);
435 return bigint(s);
436 }
437 };
438
439 template<class T> bool operator > (const T & a, const bigint & b) { return bigint(a) > b; }
440 template<class T> bool operator < (const T & a, const bigint & b) { return bigint(a) < b; }
441 template<class T> bool operator >= (const T & a, const bigint & b) { return bigint(a) >= b; }
442 template<class T> bool operator <= (const T & a, const bigint & b) { return bigint(a) <= b; }
443 template<class T> bool operator == (const T & a, const bigint & b) { return bigint(a) == b; }
444 template<class T> bool operator != (const T & a, const bigint & b) { return bigint(a) != b; }
445 template<class T> bigint operator + (const T & a, const bigint & b) { return bigint(a) + b; }
446 template<class T> bigint operator - (const T & a, const bigint & b) { return bigint(a) - b; }
447 template<class T> bigint operator ^ (const T & a, const bigint & b) { return bigint(a) ^ b; }
448
449 /*
450
451 Exclude *, /, and % to force a user decision between int and bigint algorithms
452
453 bigint operator * (bigint a, bigint b) vs. bigint operator * (bigint a, int b)
454 bigint operator / (bigint a, bigint b) vs. bigint operator / (bigint a, int b)
455 bigint operator % (bigint a, bigint b) vs. int operator % (bigint a, int b)
456
457 */
458
459 struct rational {
460 bigint num, den;
461
462 rational(): num(0), den(1) {}
463 rational(long long n): num(n), den(1) {}
464 rational(const bigint & n) : num(n), den(1) {}
465
466 template<class T1, class T2>
467 rational(const T1 & n, const T2 & d): num(n), den(d) {
468 if (den == 0)
469 throw std::runtime_error("Rational division by zero.");
470 if (den < 0) {
471 num = -num;
472 den = -den;
473 }
474 bigint a(num < 0 ? -num : num), b(den), tmp;
475 while (a != 0 && b != 0) {
476 tmp = a % b;
477 a = b;
478 b = tmp;
479 }
480 bigint gcd = (b == 0) ? a : b;
481 num /= gcd;

482 den /= gcd;
483 }
484
485 bool operator < (const rational & r) const {
486 return num * r.den < r.num * den;
487 }
488
489 bool operator > (const rational & r) const {
490 return r.num * den < num * r.den;
491 }
492
493 bool operator <= (const rational & r) const {
494 return !(r < *this);
495 }
496
497 bool operator >= (const rational & r) const {
498 return !(*this < r);
499 }
500
501 bool operator == (const rational & r) const {
502 return num == r.num && den == r.den;
503 }
504
505 bool operator != (const rational & r) const {
506 return num != r.num || den != r.den;
507 }
508
509 rational operator + (const rational & r) const {
510 return rational(num * r.den + r.num * den, den * r.den);
511 }
512
513 rational operator - (const rational & r) const {
514 return rational(num * r.den - r.num * den, r.den * den);
515 }
516
517 rational operator * (const rational & r) const {
518 return rational(num * r.num, r.den * den);
519 }
520
521 rational operator / (const rational & r) const {
522 return rational(num * r.den, den * r.num);
523 }
524
525 rational operator % (const rational & r) const {
526 return *this - r * rational(num * r.den / (r.num * den), 1);
527 }
528
529 rational operator ^ (const bigint & p) const {
530 return rational(num ^ p, den ^ p);
531 }
532
533 rational operator ++(int) { rational t(*this); operator++(); return t; }
534 rational operator --(int) { rational t(*this); operator--(); return t; }
535 rational & operator ++() { *this = *this + 1; return *this; }
536 rational & operator --() { *this = *this - 1; return *this; }
537 rational & operator += (const rational & r) { *this = *this + r; return *this; }
538 rational & operator -= (const rational & r) { *this = *this - r; return *this; }
539 rational & operator *= (const rational & r) { *this = *this * r; return *this; }
540 rational & operator /= (const rational & r) { *this = *this / r; return *this; }
541 rational & operator %= (const rational & r) { *this = *this % r; return *this; }
542 rational & operator ^= (const bigint & r) { *this = *this ^ r; return *this; }
543
544 rational operator - () const {
545 return rational(-num, den);
546 }
547
548 rational abs() const {
549 return rational(num.abs(), den);
550 }
551
552 long long to_llong() const {
553 return num.to_llong() / den.to_llong();
554 }
555
556 double to_double() const {
557 return num.to_double() / den.to_double();
558 }
559
560 friend rational abs(const rational & r) {
561 return rational(r.num.abs(), r.den);
562 }
563
564 friend std::istream & operator >> (std::istream & in, rational & r) {
565 std::string s;
566 in >> r.num;
567 r.den = 1;
568 return in;
569 }
570
571 friend std::ostream & operator << (std::ostream & out, const rational & r) {
572 out << r.num << "/" << r.den;
573 return out;
574 }
575
576 //rational in range [0, 1] with precision no greater than prec
577 static rational rand(int prec) {
578 rational r(bigint::rand(prec), bigint::rand(prec));
579 if (r.num > r.den) std::swap(r.num, r.den);
580 return r;
581 }
582 };
583
584 template<class T> bool operator > (const T & a, const rational & b) { return rational(a) > b; }
585 template<class T> bool operator < (const T & a, const rational & b) { return rational(a) < b; }
586 template<class T> bool operator >= (const T & a, const rational & b) { return rational(a) >= b; }
587 template<class T> bool operator <= (const T & a, const rational & b) { return rational(a) <= b; }
588 template<class T> bool operator == (const T & a, const rational & b) { return rational(a) == b; }
589 template<class T> bool operator != (const T & a, const rational & b) { return rational(a) != b; }
590 template<class T> rational operator + (const T & a, const rational & b) { return rational(a) + b; }
591 template<class T> rational operator - (const T & a, const rational & b) { return rational(a) - b; }
592 template<class T> rational operator * (const T & a, const rational & b) { return rational(a) * b; }
593 template<class T> rational operator / (const T & a, const rational & b) { return rational(a) / b; }
594 template<class T> rational operator % (const T & a, const rational & b) { return rational(a) % b; }
595 template<class T> rational operator ^ (const T & a, const rational & b) { return rational(a) ^ b; }
596
598

604
605 int main() {
606 for (int i = 0; i < 20; i++) {
607 int n = rand() % 100 + 1;
608 bigint a = bigint::rand(n);
609 bigint res = sqrt(a);
610 bigint xx(res * res);
611 bigint yy(res + 1);
612 yy *= yy;
613 assert(xx <= a && yy > a);
614 int m = rand() % n + 1;
615 bigint b = bigint::rand(m) + 1;
616 res = a / b;
617 xx = res * b;
618 yy = b * (res + 1);
619 assert(a >= xx && a < yy);
620 }
621
622 assert("995291497" ==
623 nthroot(bigint("981298591892498189249182998429898124"), 4));
624
625 bigint x(5);
626 x = -6;
627 assert(x.to_llong() == -6ll);
628 assert(x.to_string() == "-6");
629
630 clock_t start;
631
633 bigint c = bigint::rand(10000) / bigint::rand(2000);
634 cout << "Div took " << (float)(clock() - start)/CLOCKS_PER_SEC << "s\n";
635
637 assert((20^bigint(12345)).size() == 16062);
638 cout << "Pow took " << (float)(clock() - start)/CLOCKS_PER_SEC << "s\n";
639
640 int nn = -21, dd = 2;
641 rational n(nn, 1), d(dd);
642 cout << (nn % dd) << "\n";
643 cout << (n % d) << "\n";
644 cout << fmod(-5.3, -1.7) << "\n";
645 cout << rational(-53, 10) % rational(-17, 10) << "\n";
646 cout << rational(-53, 10).abs() << "\n";
647 cout << (rational(-53, 10) ^ 20) << "\n";
648 cout << rational::rand(20) << "\n";
649 return 0;
650 }
4.4.3 FFT and Multiplication
1 /*
2
3 A discrete Fourier transform (DFT) converts a list of equally

4 spaced samples of a function into the list of coefficients of
5 a finite combination of complex sinusoids, ordered by their
6 frequencies, that has those same sample values. A Fast Fourier
7 Transform (FFT) rapidly computes the DFT by factorizing the
8 DFT matrix into a product of sparse (mostly zero) factors.
9 The FFT can be used to solve problems such as efficiently
10 multiplying big integers or polynomials
11
12 The fft() function below is a generic function that will
13 work well in many applications beyond just multiplying
14 big integers. While Karatsuba multiplication is ~ O(n^1.58),
15 the complexity of the fft multiplication is only O(n log n).
16
17 Note that mul(string, string) in the following implementation
18 only works for strings of strictly digits from 0 to 9.
19 It is also easy to adapt this for the bigint class in the
20 previous section. Simply replace the old bigint operator *
21 definition with the following modified version of mul():
22
23 bigint operator * (const bigint & v) const {
24 static const int _base = 10000, _base_digits = 4;
25 vint _a = convert_base(this->a, base_digits, _base_digits);
26 vint _b = convert_base(v.a, base_digits, _base_digits);
27 int len = 32 - __builtin_clz(std::max(_a.size(), _b.size()) - 1);
28 len = 1 << (len + 1);
29 vcd a(len), b(len);
30 for (int i = 0; i < _a.size(); i++) a[i] = cd(_a[i], 0);
31 for (int i = 0; i < _b.size(); i++) b[i] = cd(_b[i], 0);
32 a = fft(a);
33 b = fft(b);
34 for (int i = 0; i < len; i++) {
35 double real = a[i].real() * b[i].real() - a[i].imag() * b[i].imag();
36 a[i].imag() = a[i].imag() * b[i].real() + b[i].imag() * a[i].real();
37 a[i].real() = real;
38 }
39 a = fft(a, true);
40 vll c(len);
41 for (int i = 0; i < len; i++) c[i] = (long long)(a[i].real() + 0.5);
42 bigint res;
43 res.sign = sign * v.sign;
44 for (int i = 0, carry = 0; i < c.size(); i++) {
45 long long cur = c[i] + carry;
46 res.a.push_back((int)(cur % _base));
47 carry = (int)(cur / _base);
48 }
49 res.a = convert_base(res.a, _base_digits, base_digits);
50 res.trim();
51 return res;
52 }
53
54 */
55
56 #include <algorithm> /* std::max(), std::reverse() */
57 #include <cmath> /* M_PI, cos(), sin() */
58 #include <complex>
59 #include <iomanip> /* std::setw(), std::setfill() */
63
64 typedef std::complex<double> cd;
65 typedef std::vector<cd> vcd;
66
67 vcd fft(const vcd & v, bool inverse = false) {
68 static const double PI = acos(-1.0);
69 int n = v.size(), k = 0, high1 = -1;
70 while ((1 << k) < n) k++;
71 std::vector<int> rev(n);
72 rev[0] = 0;
73 for (int i = 1; i < n; i++) {
74 if ((i & (i - 1)) == 0) high1++;
75 rev[i] = rev[i ^ (1 << high1)];
76 rev[i] |= (1 << (k - high1 - 1));
77 }
78 vcd roots(n), res(n);
79 for (int i = 0; i < n; i++) {
80 double alpha = 2 * PI * i / n;
81 roots[i] = cd(cos(alpha), sin(alpha));
82 }
83 for (int i = 0; i < n; i++) res[i] = v[rev[i]];
84 for (int len = 1; len < n; len <<= 1) {
85 vcd tmp(n);
86 int rstep = roots.size() / (len * 2);
87 for (int pdest = 0; pdest < n;) {
88 int p1 = pdest;
89 for (int i = 0; i < len; i++) {
90 cd val = roots[i * rstep] * res[p1 + len];
91 tmp[pdest] = res[p1] + val;
92 tmp[pdest + len] = res[p1] - val;
93 pdest++, p1++;
94 }
95 pdest += len;
96 }
97 res.swap(tmp);
98 }
99 if (inverse) {
100 for (int i = 0; i < (int)res.size(); i++) res[i] /= v.size();
101 std::reverse(res.begin() + 1, res.end());
102 }
103 return res;
104 }
105
106 typedef std::vector<long long> vll;
107
108 vll mul(const vll & va, const vll & vb) {
109 int len = 32 - __builtin_clz(std::max(va.size(), vb.size()) - 1);
110 len = 1 << (len + 1);
111 vcd a(len), b(len);
112 for (int i = 0; i < (int)va.size(); i++) a[i] = cd(va[i], 0);
113 for (int i = 0; i < (int)vb.size(); i++) b[i] = cd(vb[i], 0);
114 a = fft(a);
115 b = fft(b);
116 for (int i = 0; i < len; i++) {
117 double real = a[i].real() * b[i].real() - a[i].imag() * b[i].imag();
118 a[i].imag() = a[i].imag() * b[i].real() + b[i].imag() * a[i].real();
119 a[i].real() = real;
120 }
121 a = fft(a, true);

122 vll res(len);
123 for (int i = 0; i < len; i++) res[i] = (long long)(a[i].real() + 0.5);
124 return res;
125 }
126
127 const int base = 10000, base_digits = 4;
128
129 std::string mul(const std::string & as, const std::string & bs) {
130 vll a, b;
131 for (int i = as.size() - 1; i >= 0; i -= base_digits) {
132 int x = 0;
133 for (int j = std::max(0, i - base_digits + 1); j <= i; j++)
134 x = x * 10 + as[j] - 0;
135 a.push_back(x);
136 }
137 for (int i = bs.size() - 1; i >= 0; i -= base_digits) {
138 int x = 0;
139 for (int j = std::max(0, i - base_digits + 1); j <= i; j++)
140 x = x * 10 + bs[j] - 0;
141 b.push_back(x);
142 }
143 vll c = mul(a, b);
144 long long carry = 0;
145 for (int i = 0; i < (int)c.size(); i++) {
146 c[i] += carry;
147 carry = c[i] / base;
148 c[i] %= base;
149 }
150 while (c.back() == 0) c.pop_back();
151 if (c.empty()) c.push_back(0);
153 oss << (c.empty() ? 0 : c.back());
154 for (int i = c.size() - 2; i >= 0; i--)
155 oss << std::setw(base_digits) << std::setfill(0) << c[i];
157 }
158
160
162
163 int main() {
164 assert(mul("98904189", "244212") == "24153589804068");
165 return 0;
166 }
4.5 Linear Algebra
4.5.1 Matrix Class
1 /*
2
3 Basic matrix class with support for arithmetic operations
4 as well as matrix multiplication and exponentiation. You
5 can access/modify indices using m(r, c) or m[r][c]. You
4.5. Linear Algebra 215
6 can also treat it as a 2d vector, since the cast operator

7 to a reference to its internal 2d vector is defined. This
8 makes it compatible with the 2d vector functions such as
9 det() and lu_decompose() in later sections.
10
11 */
12
16
17 template<class val_t> class matrix {
18 int r, c;
19 std::vector<std::vector<val_t> > mat;
20
21 public:
22 matrix(int rows, int cols, val_t init = val_t()) {
23 r = rows;
24 c = cols;
25 mat.resize(r, std::vector<val_t>(c, init));
26 }
27
28 matrix(const std::vector<std::vector<val_t> > & m) {
29 r = m.size();
30 c = m[0].size();
31 mat = m;
32 mat.resize(r, std::vector<val_t>(c));
33 }
34
35 template<size_t rows, size_t cols>
36 matrix(val_t (&init)[rows][cols]) {
37 r = rows;
38 c = cols;
39 mat.resize(r, std::vector<val_t>(c));
40 for (int i = 0; i < r; i++)
41 for (int j = 0; j < c; j++)
42 mat[i][j] = init[i][j];
43 }
44
45 operator std::vector<std::vector<val_t> > &() { return mat; }
46 val_t & operator() (int r, int c) { return mat[r][c]; }
47 std::vector<val_t> & operator[] (int r) { return mat[r]; }
48 val_t at(int r, int c) const { return mat[r][c]; }
49 int rows() const { return r; }
50 int cols() const { return c; }
51
52 friend bool operator < (const matrix & a, const matrix & b) { return a.mat < b.mat; }
53 friend bool operator > (const matrix & a, const matrix & b) { return a.mat > b.mat; }
54 friend bool operator <= (const matrix & a, const matrix & b) { return a.mat <= b.mat; }
55 friend bool operator >= (const matrix & a, const matrix & b) { return a.mat >= b.mat; }
56 friend bool operator == (const matrix & a, const matrix & b) { return a.mat == b.mat; }
57 friend bool operator != (const matrix & a, const matrix & b) { return a.mat != b.mat; }
58
59 friend matrix operator + (const matrix & a, const matrix & b) {
60 if (a.r != b.r || a.c != b.c)
61 throw std::runtime_error("Matrix dimensions dont match.");
62 matrix res(a);
63 for (int i = 0; i < res.r; i++)
64 for (int j = 0; j < res.c; j++)
65 res.mat[i][j] += b.mat[i][j];
66 return res;
67 }
68
69 friend matrix operator - (const matrix & a, const matrix & b) {
70 if (a.r != b.r || a.c != b.c)
71 throw std::runtime_error("Matrix dimensions dont match.");
72 matrix res(a);
73 for (int i = 0; i < a.r; i++)
74 for (int j = 0; j < a.c; j++)
75 res.mat[i][j] -= b.mat[i][j];
76 return res;
77 }
78
79 friend matrix operator * (const matrix & a, const matrix & b) {
80 if (a.c != b.r)
81 throw std::runtime_error("# of a cols must equal # of b rows.");
82 matrix res(a.r, b.c, 0);
83 for (int i = 0; i < a.r; i++)
84 for (int j = 0; j < b.c; j++)
85 for (int k = 0; k < a.c; k++)
86 res.mat[i][j] += a.mat[i][k] * b.mat[k][j];
87 return res;
88 }
89
90 friend matrix operator + (const matrix & a, const val_t & v) {
91 matrix res(a);
92 for (int i = 0; i < a.r; i++)
93 for (int j = 0; j < a.c; j++) res.mat[i][j] += v;
94 return res;
95 }
96
97 friend matrix operator - (const matrix & a, const val_t & v) {
98 matrix res(a);
99 for (int i = 0; i < a.r; i++)
100 for (int j = 0; j < a.c; j++) res.mat[i][j] -= v;
101 return res;
102 }
103
104 friend matrix operator * (const matrix & a, const val_t & v) {
105 matrix res(a);
106 for (int i = 0; i < a.r; i++)
107 for (int j = 0; j < a.c; j++) res.mat[i][j] *= v;
108 return res;
109 }
110
111 friend matrix operator / (const matrix & a, const val_t & v) {
112 matrix res(a);
113 for (int i = 0; i < a.r; i++)
114 for (int j = 0; j < a.c; j++)
115 res.mat[i][j] /= v;
116 return res;
117 }
118
119 //raise matrix to the n-th power. precondition: a must be a square matrix
120 friend matrix operator ^ (const matrix & a, unsigned int n) {
121 if (a.r != a.c)
122 throw std::runtime_error("Matrix must be square for exponentiation.");
123 if (n == 0) return identity_matrix(a.r);
124 if (n % 2 == 0) return (a * a) ^ (n / 2);

125 return a * (a ^ (n - 1));
126 }
127
128 //returns a^1 + a^2 + ... + a^n
129 friend matrix powsum(const matrix & a, unsigned int n) {
130 if (n == 0) return matrix(a.r, a.r);
131 if (n % 2 == 0)
132 return powsum(a, n / 2) * (identity_matrix(a.r) + (a ^ (n / 2)));
133 return a + a * powsum(a, n - 1);
134 }
135
136 matrix & operator += (const matrix & m) { *this = *this + m; return *this; }
137 matrix & operator -= (const matrix & m) { *this = *this - m; return *this; }
138 matrix & operator *= (const matrix & m) { *this = *this * m; return *this; }
139 matrix & operator += (const val_t & v) { *this = *this + v; return *this; }
140 matrix & operator -= (const val_t & v) { *this = *this - v; return *this; }
141 matrix & operator *= (const val_t & v) { *this = *this * v; return *this; }
142 matrix & operator /= (const val_t & v) { *this = *this / v; return *this; }
143 matrix & operator ^= (unsigned int n) { *this = *this ^ n; return *this; }
144
145 static matrix identity_matrix(int n) {
146 matrix res(n, n);
147 for (int i = 0; i < n; i++) res[i][i] = 1;
148 return res;
149 }
150
151 friend std::ostream & operator << (std::ostream & out, const matrix & m) {
152 out << "[";
153 for (int i = 0; i < m.r; i++) {
154 out << (i > 0 ? ",[" : "[");
155 for (int j = 0; j < m.c; j++)
156 out << (j > 0 ? "," : "") << m.mat[i][j];
157 out << "]";
158 }
159 out << "]";
160 return out;
161 }
162 };
163
165
169
170 int main() {
171 int a[2][2] = {{1,8}, {5,9}};
172 matrix<int> m(5, 5, 10), m2(a);
173 m += 10;
174 m[0][0] += 10;
175 assert(m[0][0] == 30 && m[1][1] == 20);
176 assert(powsum(m2, 3) == m2 + m2*m2 + (m2^3));
177 return 0;
178 }
4.5.2 Determinant (Gauss)

1 /*
2
3 The following are ways to compute the determinant of a
4 matrix directly using Gaussian elimination. See the
5 following section for a generalized solution using LU
6 decompositions. Since the determinant can get very large,
7 look out for overflows and floating-point inaccuracies.
8 Bignums are recommended for maximal correctness.
9
10 Complexity: O(N^3), except for the adjustment for
11 overflow in the integer det() function.
12
13 Precondition: All input matrices must be square.
14
15 */
16
19 #include <cmath> /* fabs() */
20 #include <map>
22
23 static const double eps = 1e-10;
24 typedef std::vector<std::vector<int> > vvi;
25 typedef std::vector<std::vector<double> > vvd;
26
27 double det(vvd a) {
29 assert(!a.empty() && n == (int)a[0].size());
30 double res = 1;
31 std::vector<bool> used(n, false);
32 for (int i = 0; i < n; i++) {
33 int p;
34 for (p = 0; p < n; p++)
35 if (!used[p] && fabs(a[p][i]) > eps)
36 break;
37 if (p >= n) return 0;
38 res *= a[p][i];
39 used[p] = true;
40 double z = 1 / a[p][i];
41 for (int j = 0; j < n; j++) a[p][j] *= z;
42 for (int j = 0; j < n; j++) {
43 if (j == p) continue;
44 z = a[j][i];
45 for (int k = 0; k < n; k++)
46 a[j][k] -= z * a[p][k];
47 }
48 }
49 return res;
50 }
51
52 /*
53
54 Determinant of Integer Matrix
55
56 This is prone to overflow, so it is recommended you use your
57 own bigint class instead of long long. At the end of this
58 function, the final answer is found as a product of powers.
59 You have two choices: change the "#if 0" to "#if 1" and use
60 the naive method to compute this product and risk overflow,

61 or keep it as "#if 0" and try to make the situation better
62 through prime factorization (less efficient). Note that
63 even in the prime factorization method, overflow may happen
64 if the final answer is too big for a long long.
65
66 */
67
68 //C++98 doesnt have an abs() for long long
69 template<class T> inline T _abs(const T & x) {
70 return x < 0 ? -x : x;
71 }
72
73 long long det(const vvi & a) {
76 long long b[n][n], det = 1;
77 for (int i = 0; i < n; i++)
78 for (int j = 0; j < n; j++) b[i][j] = a[i][j];
79 int sign = 1, exponent[n];
80 for (int i = 0; i < n; i++) {
81 exponent[i] = 0;
82 int k = i;
83 for (int j = i + 1; j < n; j++) {
84 if (b[k][i] == 0 || (b[j][i] != 0 && _abs(b[k][i]) > _abs(b[j][i])))
85 k = j;
86 }
87 if (b[k][i] == 0) return 0;
88 if (i != k) {
89 sign = -sign;
90 for (int j = 0; j < n; j++)
91 std::swap(b[i][j], b[k][j]);
92 }
93 exponent[i]++;
94 for (int j = i + 1; j < n; j++)
95 if (b[j][i] != 0) {
96 for (int p = i + 1; p < n; ++p)
97 b[j][p] = b[j][p] * b[i][i] - b[i][p] * b[j][i];
98 exponent[i]--;
99 }
100 }
101
102 #if 0
103 for (int i = 0; i < n; i++)
104 for (; exponent[i] > 0; exponent[i]--)
105 det *= b[i][i];
106 for (int i = 0; i < n; i++)
107 for (; exponent[i] < 0; exponent[i]++)
108 det /= b[i][i];
109 #else
110 std::map<long long, int> m;
111 for (int i = 0; i < n; i++) {
112 long long x = b[i][i];
113 for (long long d = 2; ; d++) {
114 long long power = 0, quo = x / d, rem = x - quo * d;
115 if (d > quo || (d == quo && rem > 0)) break;
116 for (; rem == 0; rem = x - quo * d) {
117 power++;
118 x = quo;
119 quo = x / d;
120 }
121 if (power > 0) m[d] += power * exponent[i];
122 }
123 if (x > 1) m[x] += exponent[i];
124 }
125 std::map<long long, int>::iterator it;
126 for (it = m.begin(); it != m.end(); ++it)
127 for (int i = 0; i < it->second; i++)
128 det *= it->first;
129 #endif
130
131 return sign < 0 ? -det : det;
132 }
133
135
138
139 int main() {
141 int a[n][n] = {{6,1,1},{4,-2,5},{2,8,7}};
142 vvi v1(n);
143 vvd v2(n);
144 for (int i = 0; i < n; i++) {
145 v1[i] = vector<int>(a[i], a[i] + n);
146 v2[i] = vector<double>(a[i], a[i] + n);
147 }
148 int d1 = det(v1);
149 int d2 = (int)det(v2);
150 assert(d1 == d2 && d2 == -306);
151 return 0;
152 }
4.5.3 Gaussian Elimination
1 /*
2
3 Given a system of m linear equations with n unknowns:
4
5 A(1,1)*x(1) + A(1,2)*x(2) + ... + A(1,n)*x(n) = B(1)
6 A(2,1)*x(1) + A(2,2)*x(2) + ... + A(2,n)*x(n) = B(2)
7 ...
8 A(m,1)*x(1) + A(m,2)*x(2) + ... + A(m,n)*x(n) = B(m)
9
10 For any system of linear equations, there will either
11 be no solution (in 2d, lines are parallel), a single
12 solution (in 2d, the lines intersect at a point), or
13 or infinite solutions (in 2d, lines are the same).
14
15 Using Gaussian elimination in O(n^3), this program
16 solves for the values of x(1) ... x(n) or determines
17 that no unique solution of x() exists. Note that
18 the implementation below uses 0-based indices.
19
20 */
21
25
27 typedef std::vector<double> vd;
28 typedef std::vector<vd> vvd;
29
30 //note: A[i][n] stores B[i]
31 //if no unique solution found, returns empty vector
32 vd solve_system(vvd A) {
33 int m = A.size(), n = A[0].size() - 1;
34 vd x(n);
35 if (n > m) goto fail;
36 for (int k = 0; k < n; k++) {
37 double mv = 0;
38 int mi = -1;
39 for (int i = k; i < m; i++)
40 if (mv < fabs(A[i][k])) {
41 mv = fabs(A[i][k]);
42 mi = i;
43 }
44 if (mv < eps) goto fail;
45 for (int i = 0; i <= n; i++)
46 std::swap(A[mi][i], A[k][i]);
47 for (int i = k + 1; i < m; i++) {
48 double v = A[i][k] / A[k][k];
49 for (int j = k; j <= n; j++)
50 A[i][j] -= v * A[k][j];
51 A[i][k] = 0;
52 }
53 }
54 for (int i = n; i < m; i++)
55 if (fabs(A[i][n]) > eps) goto fail;
56 for (int i = n - 1; i >= 0; i--) {
57 if (fabs(A[i][i]) < eps) goto fail;
58 double v = 0;
59 for (int j = i + 1; j < n; j++)
60 v += A[i][j] * x[j];
61 v = A[i][n] - v;
62 x[i] = v / A[i][i];
63 }
64 return x;
65 fail:
66 return vd();
67 }
68
69 /*** Example Usage (wcipeg.com/problem/syssolve) ***/
70
73
74 int main() {
75 int n, m;
76 cin >> n >> m;
77 vvd a(m, vd(n + 1));
78 for (int i = 0; i < m; i++)
79 for (int j = 0; j <= n; j++)
80 cin >> a[i][j];

81 vd x = solve_system(a);
82 if (x.empty()) {
83 cout << "NO UNIQUE SOLUTION\n";
84 } else {
86 for (int i = 0; i < n; i++)
87 cout << fixed << x[i] << "\n";
88 }
89 return 0;
90 }
4.5.4 LU Decomposition
1 /*
2
3 The LU (lower upper) decomposition of a matrix is a factorization
4 of a matrix as the product of a lower triangular matrix and an
5 upper triangular matrix. With the LU decomposition, we can solve
6 many problems, including the determinant of the matrix, a systems
7 of linear equations, and the inverse of a matrix.
8
9 Note: in the following implementation, each call to det(),
10 solve_system(), and inverse() recomputes the lu decomposition.
11 For the same matrix, you should precompute the lu decomposition
12 and reuse it for several of these operations afterwards.
13
14 Complexity: O(n^3) for lu_decompose(). det() uses the running time
15 of lu_decompose(), plus an addition O(n) term. solve_system() and
16 inverse() both have the running time of lu_decompose(), plus an
17 additional O(n^3) term.
18
19 */
20
25
29
30 /*
31
32 LU decomposition with Gauss-Jordan elimination. This is generalized
33 for rectangular matrices. Since the resulting L and U matrices have
34 all mutually exclusive 0s (except when i == j), we can merge them
35 into a single LU matrix to save memory. Note: l[i][i] = 1 for all i.
36
37 Optionally determine the permutation vector p. If an array p is
38 passed, p[i] will be populated such that p[i] is the only column of
39 the i-th row of the permutation matrix that is equal to 1.
40
41 Returns: a matrix m, the merged lower/upper triangular matrix:
42 m[i][j] = l[i][j] (for i > j) or u[i][j] (for i <= j)
43
44 */
45
46 vvd lu_decompose(vvd a, int * detsign = 0, int * p = 0) {
47 int n = a.size(), m = a[0].size();
48 int sign = 1;
49 if (p != 0)
50 for (int i = 0; i < n; i++) p[i] = i;
51 for (int r = 0, c = 0; r < n && c < m; r++, c++) {
52 int pr = r;
53 for (int i = r + 1; i < n; i++)
54 if (fabs(a[i][c]) > fabs(a[pr][c]))
55 pr = i;
56 if (fabs(a[pr][c]) <= eps) {
57 r--;
58 continue;
59 }
60 if (pr != r) {
61 if (p != 0) std::swap(p[r], p[pr]);
62 sign = -sign;
63 for (int i = 0; i < m; i++)
64 std::swap(a[r][i], a[pr][i]);
65 }
66 for (int s = r + 1; s < n; s++) {
67 a[s][c] /= a[r][c];
68 for (int d = c + 1; d < m; d++)
69 a[s][d] -= a[s][c] * a[r][d];
70 }
71 }
72 if (detsign != 0) *detsign = sign;
73 return a;
74 }
75
76 double getl(const vvd & lu, int i, int j) {
77 if (i > j) return lu[i][j];
78 return i < j ? 0.0 : 1.0;
79 }
80
81 double getu(const vvd & lu, int i, int j) {
82 return i <= j ? lu[i][j] : 0.0;
83 }
84
85 //Precondition: A is square matrix.
86 double det(const vvd & a) {
87 int n = a.size(), detsign;
89 vvd lu = lu_decompose(a, &detsign);
90 double det = 1;
91 for (int i = 0; i < n; i++)
92 det *= lu[i][i];
93 return detsign < 0 ? -det : det;
94 }
95
96 /*
97
98 Solves system of linear equations with forward/backwards
99 substitution. Precondition: A must be n*n and B must be n*m.
100 Returns: an n by m matrix X such that A*X = B.
101
102 */
103
104 vvd solve_system(const vvd & a, const vvd & b) {
105 int n = b.size(), m = b[0].size();
106 assert(!a.empty() && n == (int)a.size() && n == (int)a[0].size());
107 int detsign, p[a.size()];
108 vvd lu = lu_decompose(a, &detsign, p);
109 //forward substitute for Y in L*Y = B
110 vvd y(n, vd(m));
111 for (int j = 0; j < m; j++) {
112 y[0][j] = b[p[0]][j] / getl(lu, 0, 0);
113 for (int i = 1; i < n; i++) {
114 double s = 0;
115 for (int k = 0; k < i; k++)
116 s += getl(lu, i, k) * y[k][j];
117 y[i][j] = (b[p[i]][j] - s) / getl(lu, i, i);
118 }
119 }
120 //backward substitute for X in U*X = Y
121 vvd x(n, vd(m));
122 for (int j = 0; j < m; j++) {
123 x[n - 1][j] = y[n - 1][j] / getu(lu, n-1, n-1);
124 for (int i = n - 2; i >= 0; i--) {
125 double s = 0;
126 for (int k = i + 1; k < n; k++)
127 s += getu(lu, i, k) * x[k][j];
128 x[i][j] = (y[i][j] - s) / getu(lu, i, i);
129 }
130 }
131 return x;
132 }
133
134 /*
135
136 Find the inverse A^-1 of a matrix A. The inverse of a matrix
137 satisfies A * A^-1 = I, where I is the identity matrix (for
138 all pairs (i, j), I[i][j] = 1 iff i = j, else I[i][j] = 0).
139 The inverse of a matrix exists if and only if det(a) is not 0.
140 Were lazy, so we just generate I and call solve_system().
141
142 Precondition: A is a square and det(A) != 0.
143
144 */
145
146 vvd inverse(const vvd & a) {
149 vvd I(n, vd(n));
150 for (int i = 0; i < n; i++) I[i][i] = 1;
151 return solve_system(a, I);
152 }
153
155
159
160 void print(const vvd & m) {
161 cout << "[";
162 for (int i = 0; i < (int)m.size(); i++) {

163 cout << (i > 0 ? ",[" : "[");
164 for (int j = 0; j < (int)m[0].size(); j++)
165 cout << (j > 0 ? "," : "") << m[i][j];
166 cout << "]";
167 }
168 cout << "]\n";
169 }
170
171 void printlu(const vvd & lu) {
172 printf("L:\n");
173 for (int i = 0; i < (int)lu.size(); i++) {
174 for (int j = 0; j < (int)lu[0].size(); j++)
175 printf("%10.5f ", getl(lu, i, j));
176 printf("\n");
177 }
178 printf("U:\n");
179 for (int i = 0; i < (int)lu.size(); i++) {
180 for (int j = 0; j < (int)lu[0].size(); j++)
181 printf("%10.5f ", getu(lu, i, j));
182 printf("\n");
183 }
184 }
185
186 int main() {
187 { //determinant of 3x3
189 double a[n][n] = {{1,3,5},{2,4,7},{1,1,0}};
190 vvd v(n);
191 for (int i = 0; i < n; i++)
192 v[i] = vector<double>(a[i], a[i] + n);
193 printlu(lu_decompose(v));
194 cout << "determinant: " << det(v) << "\n"; //4
195 }
196
197 { //determinant of 4x4
199 double a[n][n] = {{11,9,24,2},{1,5,2,6},{3,17,18,1},{2,5,7,1}};
200 vvd v(n);
201 for (int i = 0; i < n; i++)
203 printlu(lu_decompose(v));
204 cout << "determinant: " << det(v) << "\n"; //284
205 }
206
207 { //solve for [x, y] in x + 3y = 4 && 2x + 3y = 6
209 double a[n][n] = {{1,3},{2,3}};
210 double b[n] = {4, 6};
211 vvd va(n), vb(n);
212 for (int i = 0; i < n; i++) {
213 va[i] = vector<double>(a[i], a[i] + n);
214 vb[i] = vector<double>(1, b[i]);
215 }
216 vvd x = solve_system(va, vb);
217 for (int i = 0; i < n; i++) {
218 assert(fabs(a[i][0]*x[0][0] + a[i][1]*x[1][0] - b[i]) < eps);
219 }
220 }
221
222 { //find inverse by solving a system
224 double a[n][n] = {{2,3},{1,2}};
225 vvd v(n);
226 for (int i = 0; i < n; i++)
228 print(inverse(v)); //[[2,-3],[-1,2]]
229 }
230 return 0;
231 }
4.5.5 Simplex Algorithm
1 /*
2
3 Description: The canonical form of a linear programming
4 problem is to maximize c^T*x, subject to Ax <= b, and x >= 0.
5 where x is the vector of variables (to be solved), c and b
6 are vectors of (known) coefficients, A is a (known) matrix of
7 coefficients, and (.)^T is the matrix transpose. The following
8 implementation solves n variables in a system of m constraints.
9
10 Precondition: ab has dimensions m by n+1 and c has length n+1.
11
12 Complexity: The simplex method is remarkably efficient in
13 practice, usually taking 2m or 3m iterations, converging in
14 expected polynomial time for certain distributions of random
15 inputs. However, its worst-case complexity is exponential,
16 and can be demonstrated with carefully constructed examples.
17
18 */
19
24
27
28 //ab[i][0..n-1] stores A and ab[i][n] stores B
29 vd simplex(const vvd & ab, const vd & c, bool max = true) {
31 int n = c.size() - 1, m = ab.size();
32 vvd ts(m + 2, vd(n + 2));
33 ts[1][1] = max ? c[n] : -c[n];
34 for (int j = 1; j <= n; j++)
35 ts[1][j + 1] = max ? c[j - 1] : -c[j - 1];
36 for (int i = 1; i <= m; i++) {
37 for (int j = 1; j <= n; j++)
38 ts[i + 1][j + 1] = -ab[i - 1][j - 1];
39 ts[i + 1][1] = ab[i - 1][n];
40 }
41 for (int j = 1; j <= n; j++)
42 ts[0][j + 1] = j;
43 for (int i = n + 1; i <= n + m; i++)
44 ts[i - n + 1][0] = i;
45 double p1 = 0.0, p2 = 0.0;
46 bool done = true;
47 do {
48 double mn = DBL_MAX, xmax = 0.0, v;
49 for (int j = 2; j <= n + 1; j++)
50 if (ts[1][j] > 0.0 && ts[1][j] > xmax) {
51 p2 = j;
52 xmax = ts[1][j];
53 }
54 for (int i = 2; i <= m + 1; i++) {
55 v = fabs(ts[i][1] / ts[i][p2]);
56 if (ts[i][p2] < 0.0 && mn > v) {
57 mn = v;
58 p1 = i;
59 }
60 }
61 std::swap(ts[p1][0], ts[0][p2]);
62 for (int i = 1; i <= m + 1; i++) {
63 if (i == p1) continue;
64 for (int j = 1; j <= n + 1; j++)
65 if (j != p2)
66 ts[i][j] -= ts[p1][j] * ts[i][p2] / ts[p1][p2];
67 }
68 ts[p1][p2] = 1.0 / ts[p1][p2];
69 for (int j = 1; j <= n + 1; j++) {
70 if (j != p2)
71 ts[p1][j] *= fabs(ts[p1][p2]);
72 }
73 for (int i = 1; i <= m + 1; i++) {
74 if (i != p1)
75 ts[i][p2] *= ts[p1][p2];
76 }
77 for (int i = 2; i <= m + 1; i++)
78 if (ts[i][1] < 0.0) return vd(); //no solution
79 done = true;
80 for (int j = 2; j <= n + 1; j++)
81 if (ts[1][j] > 0) done = false;
82 } while (!done);
83 vd res;
84 for (int i = 1; i <= n; i++)
85 for (int j = 2; j <= m + 1; j++)
86 if (fabs(ts[j][0] - i) <= eps)
87 res.push_back(ts[j][1]);
88 //the solution is stored in ts[1][1]
89 return res;
90 }
91
93
96
97 /*
98 Maximize 3x + 4y + 5, subject to x, y >= 0 and:
99 -2x + 1y <= 0
100 1x + 0.85y <= 9
101 1x + 2y <= 14
102
103 Note: The solution is 38.3043 at (5.30435, 4.34783).

104 */
105
106 int main() {
107 const int n = 2, m = 3;
108 double ab[m][n + 1] = {{-2, 1, 0}, {1, 0.85, 9}, {1, 2, 14}};
109 double c[n + 1] = {3, 4, 5};
110 vvd vab(m, vd(n + 1));
111 vd vc(c, c + n + 1);
112 for (int i = 0; i < m; i++) {
113 for (int j = 0; j <= n; j++)
114 vab[i][j] = ab[i][j];
115 }
116 vd x = simplex(vab, vc);
117 if (x.empty()) {
118 cout << "No solution.\n";
119 } else {
120 double solval = c[n];
122 solval += c[i] * x[i];
123 cout << "Solution = " << solval;
124 cout << " at (" << x[0];
126 cout << ", " << x[i];
127 cout << ").\n";
128 }
129 return 0;
130 }
4.6 Root-Finding
4.6.1 Real Root Finding (Differentiation)
1 /*
2
3 Real roots can be found via binary searching, a.k.a the bisection
4 method. If two x-coordinates evaluate to y-coordinates that have
5 opposite signs, a root must exist between them. For a polynomial
6 function, at most 1 root lies between adjacent local extrema.
7 Since local extrema exist where the derivative equals 0, we can
8 break root-finding into the subproblem of finding the roots of
9 the derivative. Recursively solve for local extrema until we get
10 to a base case of degree 0. For each set of local extrema found,
11 binary search between pairs of extrema for a root. This method is
12 easy, robust, and allows us to find the root to an arbitrary level
13 of accuracy. Were limited only by the precision of the arithmetic.
14
15 Complexity: For a degree N polynomial, repeatedly differentiating
16 it will take N + (N-1) + ... + 1 = O(N^2) operations. At each step
17 we binary search the number of times equal to the current degree.
18 If we want to make roots precise to eps=10^-P, each binary search
19 will take O(log P). Thus the overall complexity is O(N^2 log P).
20
21 */
22
23 #include <cmath> /* fabsl(), powl() */
4.6. Root-Finding 229
24 #include <limits> /* std::numeric_limits<>::quiet_NaN() */

25 #include <utility> /* std::pair<> */
27
28 typedef long double Double;
29 typedef std::vector<std::pair<Double, int> > poly;
30
31 const Double epsa = 1e-11; //required precision of roots in absolute error
32 const Double epsr = 1e-15; //required precision of roots in relative error
33 const Double eps0 = 1e-17; //x is considered a root if fabs(eval(x))<=eps0
34 const Double inf = 1e20; //[-inf, inf] is the range of roots to consider
35 const Double NaN = std::numeric_limits<Double>::quiet_NaN();
36
37 Double eval(const poly & p, Double x) {
38 Double res = 0;
39 for (int i = 0; i < (int)p.size(); i++)
40 res += p[i].first * powl(x, p[i].second);
41 return res;
42 }
43
44 Double find_root(const poly & p, Double x1, Double x2) {
45 Double y1 = eval(p, x1), y2 = eval(p, x2);
46 if (fabsl(y1) <= eps0) return x1;
47 bool neg1 = (y1 < 0), neg2 = (y2 < 0);
48 if (fabsl(y2) <= eps0 || neg1 == neg2) return NaN;
49 while (x2 - x1 > epsa && x1 * (1 + epsr) < x2 && x2 * (1 + epsr) > x1) {
50 Double x = (x1 + x2) / 2;
51 ((eval(p, x) < 0) == neg1 ? x1 : x2) = x;
52 }
53 return x1;
54 }
55
56 std::vector<Double> find_all_roots(const poly & p) {
57 poly dif;
58 for (int i = 0; i < (int)p.size(); i++)
59 if (p[i].second > 0)
60 dif.push_back(std::make_pair(p[i].first * p[i].second, p[i].second - 1));
61 if (dif.empty()) return std::vector<Double>();
62 std::vector<Double> res, r = find_all_roots(dif);
63 r.insert(r.begin(), -inf);
64 r.push_back(inf);
65 for (int i = 0; i < (int)r.size() - 1; i++) {
66 Double root = find_root(p, r[i], r[i + 1]);
67 if (root != root) continue; //NaN, not found
68 if (res.empty() || root != res.back())
69 res.push_back(root);
70 }
71 return res;
72 }
73
74 /*** Example Usage (http://wcipeg.com/problem/rootsolve) ***/
75
78
79 int main() {
80 int n, d;
81 Double c;
82 poly p;
83 cin >> n;
84 for (int i = 0; i < n; i++) {
85 cin >> c >> d;
86 p.push_back(make_pair(c, d));
87 }
88 vector<Double> sol = find_all_roots(p);
89 if (sol.empty()) {
90 cout << "NO REAL ROOTS\n";
91 } else {
93 for (int i = 0; i < (int)sol.size(); i++)
94 cout << fixed << sol[i] << "\n";
95 }
96 return 0;
97 }
4.6.2 Complex Root Finding (Laguerres)
1 /*
2
3 Laguerres method can be used to not only find complex roots of
4 a polynomial, the polynomial may also have complex coefficients.
5 From extensive empirical study, Laguerres method is observed to
6 be very close to being a "sure-fire" method, as it is almost
7 guaranteed to always converge to some root of the polynomial
8 regardless of what initial guess is chosen.
9
10 */
11
13 #include <cstdlib> /* rand(), RAND_MAX */
15
16 typedef long double Double;
17 typedef std::complex<Double> cdouble;
18 typedef std::vector<cdouble> poly;
19
20 const Double eps = 1e-12;
21
22 std::pair<poly, cdouble> horner(const poly & a, const cdouble & x) {
24 poly b = poly(std::max(1, n - 1));
25 for (int i = n - 1; i > 0; i--)
26 b[i - 1] = a[i] + (i < n - 1 ? b[i] * x : 0);
27 return std::make_pair(b, a[0] + b[0] * x);
28 }
29
30 cdouble eval(const poly & p, const cdouble & x) {
31 return horner(p, x).second;
32 }
33
34 poly derivative(const poly & p) {
35 int n = p.size();
36 poly r(std::max(1, n - 1));
37 for(int i = 1; i < n; i++)
38 r[i - 1] = p[i] * cdouble(i);
39 return r;
40 }
41
42 int comp(const cdouble & x, const cdouble & y) {
43 Double diff = std::abs(x) - std::abs(y);
44 return diff < -eps ? -1 : (diff > eps ? 1 : 0);
45 }
46
47 cdouble find_one_root(const poly & p, cdouble x) {
48 int n = p.size() - 1;
49 poly p1 = derivative(p), p2 = derivative(p1);
50 for (int step = 0; step < 10000; step++) {
51 cdouble y0 = eval(p, x);
52 if (comp(y0, 0) == 0) break;
53 cdouble G = eval(p1, x) / y0;
54 cdouble H = G * G - eval(p2, x) / y0;
55 cdouble R = std::sqrt(cdouble(n - 1) * (H * cdouble(n) - G * G));
56 cdouble D1 = G + R, D2 = G - R;
57 cdouble a = cdouble(n) / (comp(D1, D2) > 0 ? D1 : D2);
58 x -= a;
59 if (comp(a, 0) == 0) break;
60 }
61 return x;
62 }
63
64 std::vector<cdouble> find_all_roots(const poly & p) {
65 std::vector<cdouble> res;
66 poly q = p;
67 while (q.size() > 2) {
68 cdouble z(rand()/Double(RAND_MAX), rand()/Double(RAND_MAX));
69 z = find_one_root(q, z);
70 z = find_one_root(p, z);
71 q = horner(q, z).first;
72 res.push_back(z);
73 }
74 res.push_back(-q[0] / q[1]);
75 return res;
76 }
77
79
83
84 void print_roots(vector<cdouble> roots) {
85 for (int i = 0; i < (int)roots.size(); i++) {
86 printf("(%9.5f, ", (double)roots[i].real());
87 printf("%9.5f)\n", (double)roots[i].imag());
88 }
89 }
90
91 int main() {
92 { // x^3 - 8x^2 - 13x + 140 = (x + 4)(x - 5)(x - 7)
93 printf("Roots of x^3 - 8x^2 - 13x + 140:\n");
94 poly p;
95 p.push_back(140);
96 p.push_back(-13);
97 p.push_back(-8);
98 p.push_back(1);
99 vector<cdouble> roots = find_all_roots(p);

100 print_roots(roots);
101 }
102
103 { //(-6+4i)x^4 + (-26+12i)x^3 + (-30+40i)x^2 + (-26+12i)x + (-24+36i)
104 // = ((2+3i)x + 6)*(x + i)*(2x + (6+4i))*(x*i + 1)
105 printf("Roots of ((2+3i)x + 6)(x + i)(2x + (6+4i))(x*i + 1):\n");
106 poly p;
107 p.push_back(cdouble(-24, 36));
112 vector<cdouble> roots = find_all_roots(p);
113 print_roots(roots);
114 }
115 return 0;
116 }
4.6.3 Complex Root Finding (RPOLY)
1 /*
2
3 Determine the complex roots of a polynomial with real coefficients.
4 This is the variant of the Jenkins-Traub algorithm for polynomials
5 with real coefficient, known as RPOLY. RPOLY follows follows the
6 same pattern as the CPOLY algorithm, but computes two roots at a
7 time, either two real roots or a pair of conjugate complex roots.
8 See: https://en.wikipedia.org/wiki/Jenkins%E2%80%93Traub_algorithm
9
10 The following is a translation of TOMS493 (www.netlib.org/toms/)
11 from FORTRAN to C++, with a simple wrapper at the end for the C++
12 <complex> class. Although the code is not meant to be read, it is
13 extremely efficient and robust, capable of achieving an accuracy
14 of at least 5 decimal places for even the most strenuous inputs.
15
16 */
17
18 #include <cfloat> /* LDBL_EPSILON, LDBL_MAX, LDBL_MIN */
19 #include <cmath> /* cosl, expl, fabsl, logl, powl, sinl, sqrtl */
20
21 typedef long double LD;
22
23 void divide_quadratic(int n, LD u, LD v, LD p[], LD q[], LD * a, LD * b) {
24 q[0] = *b = p[0];
25 q[1] = *a = -((*b) * u) + p[1];
26 for (int i = 2; i < n; i++) {
27 q[i] = -((*a) * u + (*b) * v) + p[i];
28 *b = *a;
29 *a = q[i];
30 }
31 }
32
33 int get_flag(int n, LD a, LD b, LD * a1, LD * a3, LD * a7,
34 LD * c, LD * d, LD * e, LD * f, LD * g, LD * h,
35 LD k[], LD u, LD v, LD qk[]) {
36 divide_quadratic(n, u, v, k, qk, c, d);
37 if (fabsl(*c) <= 100.0 * LDBL_EPSILON * fabsl(k[n - 1]) &&

38 fabsl(*d) <= 100.0 * LDBL_EPSILON * fabsl(k[n - 2])) return 3;
39 *h = v * b;
40 if (fabsl(*d) >= fabsl(*c)) {
41 *e = a / (*d);
42 *f = (*c) / (*d);
43 *g = u * b;
44 *a1 = (*f) * b - a;
45 *a3 = (*e) * ((*g) + a) + (*h) * (b / (*d));
46 *a7 = (*h) + ((*f) + u) * a;
47 return 2;
48 }
49 *e = a / (*c);
50 *f = (*d) / (*c);
51 *g = (*e) * u;
52 *a1 = -(a * ((*d) / (*c))) + b;
53 *a3 = (*e) * a + ((*g) + (*h) / (*c)) * b;
54 *a7 = (*g) * (*d) + (*h) * (*f) + a;
55 return 1;
56 }
57
58 void find_polynomials(int n, int flag, LD a, LD b, LD a1, LD * a3,
59 LD * a7, LD k[], LD qk[], LD qp[]) {
60 if (flag == 3) {
61 k[1] = k[0] = 0.0;
62 for (int i = 2; i < n; i++) k[i] = qk[i - 2];
63 return;
64 }
65 if (fabsl(a1) > 10.0 * LDBL_EPSILON * fabsl(flag == 1 ? b : a)) {
66 *a7 /= a1;
67 *a3 /= a1;
68 k[0] = qp[0];
69 k[1] = qp[1] - (*a7) * qp[0];
70 for (int i = 2; i < n; i++)
71 k[i] = qp[i] - ((*a7) * qp[i - 1]) + (*a3) * qk[i - 2];
72 } else {
73 k[0] = 0.0;
74 k[1] = -(*a7) * qp[0];
75 for (int i = 2; i < n; i++)
76 k[i] = (*a3) * qk[i - 2] - (*a7) * qp[i - 1];
77 }
78 }
79
80 void estimate_coeff(int flag, LD * uu, LD * vv, LD a, LD a1, LD a3, LD a7,
81 LD b, LD c, LD d, LD f, LD g, LD h, LD u, LD v, LD k[],
82 int n, LD p[]) {
83 LD a4, a5, b1, b2, c1, c2, c3, c4, temp;
84 *vv = *uu = 0.0;
85 if (flag == 3) return;
86 if (flag != 2) {
87 a4 = a + u * b + h * f;
88 a5 = c + (u + v * f) * d;
89 } else {
90 a4 = (a + g) * f + h;
91 a5 = (f + u) * c + v * d;
92 }
93 b1 = -k[n - 1] / p[n];
94 b2 = -(k[n - 2] + b1 * p[n - 1]) / p[n];
95 c1 = v * b2 * a1;
96 c2 = b1 * a7;
97 c3 = b1 * b1 * a3;
98 c4 = c1 - c2 - c3;
99 temp = b1 * a4 - c4 + a5;
100 if (temp != 0.0) {
101 *uu= u - (u * (c3 + c2) + v * (b1 * a1 + b2 * a7)) / temp;
102 *vv = v * (1.0 + c4 / temp);
103 }
104 }
105
106 void solve_quadratic(LD a, LD b1, LD c, LD * sr, LD * si, LD * lr, LD * li) {
107 LD b, d, e;
108 *sr = *si = *lr = *li = 0.0;
109 if (a == 0) {
110 *sr = (b1 != 0) ? -c / b1 : *sr;
111 return;
112 }
113 if (c == 0) {
114 *lr = -b1 / a;
115 return;
116 }
117 b = b1 / 2.0;
118 if (fabsl(b) < fabsl(c)) {
119 e = (c >= 0) ? a : -a;
120 e = b * (b / fabsl(c)) - e;
121 d = sqrtl(fabsl(e)) * sqrtl(fabsl(c));
122 } else {
123 e = 1.0 - (a / b) * (c / b);
124 d = sqrtl(fabsl(e)) * fabsl(b);
125 }
126 if (e >= 0) {
127 d = (b >= 0) ? -d : d;
128 *lr = (d - b) / a;
129 *sr = (*lr != 0) ? (c / *lr / a) : *sr;
130 } else {
131 *lr = *sr = -b / a;
132 *si = fabsl(d / a);
133 *li = -(*si);
134 }
135 }
136
137 void quadratic_iterate(int N, int * NZ, LD uu, LD vv,
138 LD * szr, LD * szi, LD * lzr, LD * lzi, LD qp[],
139 int n, LD * a, LD * b, LD p[], LD qk[],
140 LD * a1, LD * a3, LD * a7, LD * c, LD * d, LD * e,
141 LD * f, LD * g, LD * h, LD k[]) {
142 int steps = 0, flag, tried_flag = 0;
143 LD ee, mp, omp = 0.0, relstp = 0.0, t, u, ui, v, vi, zm;
144 *NZ = 0;
145 u = uu;
146 v = vv;
147 do {
148 solve_quadratic(1.0, u, v, szr, szi, lzr, lzi);
149 if (fabsl(fabsl(*szr) - fabsl(*lzr)) > 0.01 * fabsl(*lzr)) break;
150 divide_quadratic(n, u, v, p, qp, a, b);
151 mp = fabsl(-((*szr) * (*b)) + *a) + fabsl((*szi) * (*b));
152 zm = sqrtl(fabsl(v));
153 ee = 2.0 * fabsl(qp[0]);
154 t = -(*szr) * (*b);
155 for (int i = 1; i < N; i++) ee = ee * zm + fabsl(qp[i]);

156 ee = ee * zm + fabsl(*a + t);
157 ee = ee * 9.0 + 2.0 * fabsl(t) - 7.0 * (fabsl(*a + t) + zm * fabsl(*b));
158 ee *= LDBL_EPSILON;
159 if (mp <= 20.0 * ee) {
160 *NZ = 2;
161 break;
162 }
163 if (++steps > 20) break;
164 if (steps >= 2 && relstp <= 0.01 && mp >= omp && !tried_flag) {
165 relstp = (relstp < LDBL_EPSILON) ? sqrtl(LDBL_EPSILON) : sqrtl(relstp);
166 u -= u * relstp;
167 v += v * relstp;
168 divide_quadratic(n, u, v, p, qp, a, b);
169 for (int i = 0; i < 5; i++) {
170 flag = get_flag(N, *a, *b, a1, a3, a7, c, d, e, f, g, h, k, u, v, qk);
171 find_polynomials(N, flag, *a, *b, *a1, a3, a7, k, qk, qp);
172 }
173 tried_flag = 1;
174 steps = 0;
175 }
176 omp = mp;
178 find_polynomials(N, flag, *a, *b, *a1, a3, a7, k, qk, qp);
180 estimate_coeff(flag, &ui, &vi, *a, *a1, *a3, *a7, *b, *c, *d, *f, *g, *h,
181 u, v, k, N, p);
182 if (vi != 0) {
183 relstp = fabsl((-v + vi) / vi);
184 u = ui;
185 v = vi;
186 }
187 } while (vi != 0);
188 }
189
190 void real_iterate(int * flag, int * nz, LD * sss, int n, LD p[],
191 int nn, LD qp[], LD * szr, LD * szi, LD k[], LD qk[]) {
192 int steps = 0;
193 LD ee, kv, mp, ms, omp = 0.0, pv, s, t = 0.0;
194 *flag = *nz = 0;
195 for (s = *sss; ; s += t) {
196 pv = p[0];
197 qp[0] = pv;
198 for (int i = 1; i < nn; i++) qp[i] = pv = pv * s + p[i];
199 mp = fabsl(pv);
200 ms = fabsl(s);
201 ee = 0.5 * fabsl(qp[0]);
202 for (int i = 1; i < nn; i++) ee = ee * ms + fabsl(qp[i]);
203 if (mp <= 20.0 * LDBL_EPSILON * (2.0 * ee - mp)) {
204 *nz = 1;
205 *szr = s;
206 *szi = 0.0;
207 break;
208 }
209 if (++steps > 10) break;
210 if (steps >= 2 && fabsl(t) <= 0.001 * fabsl(s - t) && mp > omp) {
211 *flag = 1;
212 *sss = s;
213 break;
214 }
215 omp = mp;
216 qk[0] = kv = k[0];
217 for (int i = 1; i < n; i++) qk[i] = kv = kv * s + k[i];
218 if (fabsl(kv) > fabsl(k[n - 1]) * 10.0 * LDBL_EPSILON) {
219 t = -pv / kv;
220 k[0] = qp[0];
221 for (int i = 1; i < n; i++)
222 k[i] = t * qk[i - 1] + qp[i];
223 } else {
224 k[0] = 0.0;
225 for (int i = 1; i < n; i++)
226 k[i] = qk[i - 1];
227 }
228 kv = k[0];
229 for (int i = 1; i < n; i++) kv = kv * s + k[i];
230 t = fabsl(kv) > (fabsl(k[n - 1]) * 10.0 * LDBL_EPSILON) ? -pv / kv : 0.0;
231 }
232 }
233
234 void solve_fixedshift(int l2, int * nz, LD sr, LD v, LD k[], int n,
235 LD p[], int nn, LD qp[], LD u, LD qk[], LD svk[],
236 LD * lzi, LD * lzr, LD * szi, LD * szr) {
237 int flag, _flag, __flag = 1, spass, stry, vpass, vtry;
238 LD a, a1, a3, a7, b, betas, betav, c, d, e, f, g, h;
239 LD oss, ots = 0.0, otv = 0.0, ovv, s, ss, ts, tss, tv, tvv, ui, vi, vv;
240 *nz = 0;
241 betav = betas = 0.25;
242 oss = sr;
243 ovv = v;
244 divide_quadratic(nn, u, v, p, qp, &a, &b);
245 flag = get_flag(n, a, b, &a1, &a3, &a7, &c, &d, &e, &f, &g, &h,
246 k, u, v, qk);
247 for (int j = 0; j < l2; j++) {
248 _flag = 1;
249 find_polynomials(n, flag, a, b, a1, &a3, &a7, k, qk, qp);
251 k, u, v, qk);
252 estimate_coeff(flag, &ui, &vi, a, a1, a3, a7, b, c, d, f, g, h,
253 u, v, k, n, p);
254 vv = vi;
255 ss = k[n - 1] != 0.0 ? -p[n] / k[n - 1] : 0.0;
256 ts = tv = 1.0;
257 if (j != 0 && flag != 3) {
258 tv = (vv != 0.0) ? fabsl((vv - ovv) / vv) : tv;
259 ts = (ss != 0.0) ? fabsl((ss - oss) / ss) : ts;
260 tvv = (tv < otv) ? tv * otv : 1.0;
261 tss = (ts < ots) ? ts * ots : 1.0;
262 vpass = (tvv < betav) ? 1 : 0;
263 spass = (tss < betas) ? 1 : 0;
264 if (spass || vpass) {
265 for (int i = 0; i < n; i++) svk[i] = k[i];
266 s = ss; stry = vtry = 0;
267 for (;;) {
268 if (!(_flag && spass && (!vpass || tss < tvv))) {
269 quadratic_iterate(n, nz, ui, vi, szr, szi, lzr, lzi, qp, nn,
270 &a, &b, p, qk, &a1, &a3, &a7, &c, &d, &e, &f, &g, &h, k);
271 if (*nz > 0) return;
272 __flag = vtry = 1;
273 betav *= 0.25;

274 if (stry || !spass) {
275 __flag = 0;
276 } else {
277 for (int i = 0; i < n; i++) k[i] = svk[i];
278 }
279 }
280 _flag = 0;
281 if (__flag != 0) {
282 real_iterate(&__flag, nz, &s, n, p, nn, qp, szr, szi, k, qk);
283 if (*nz > 0) return;
284 stry = 1;
285 betas *= 0.25;
286 if (__flag != 0) {
287 ui = -(s + s);
288 vi = s * s;
289 continue;
290 }
291 }
292 for (int i = 0; i < n; i++) k[i] = svk[i];
293 if (!vpass || vtry) break;
294 }
295 divide_quadratic(nn, u, v, p, qp, &a, &b);
297 k, u, v, qk);
298 }
299 }
300 ovv = vv;
301 oss = ss;
302 otv = tv;
303 ots = ts;
304 }
305 }
306
307 void find_roots(int degree, LD co[], LD re[], LD im[]) {
308 int j, jj, n, nm1, nn, nz, zero, SZ = degree + 1;
309 LD k[SZ], p[SZ], pt[SZ], qp[SZ], temp[SZ], qk[SZ], svk[SZ];
310 LD bnd, df, dx, factor, ff, moduli_max, moduli_min, sc, x, xm;
311 LD aa, bb, cc, lzi, lzr, sr, szi, szr, t, u, xx, xxx, yy;
312 n = degree;
313 xx = sqrtl(0.5);
314 yy = -xx;
315 for (j = 0; co[n] == 0; n--, j++) re[j] = im[j] = 0.0;
316 nn = n + 1;
317 for (int i = 0; i < nn; i++) p[i] = co[i];
318 while (n >= 1) {
319 if (n <= 2) {
320 if (n < 2) {
321 re[degree - 1] = -p[1] / p[0];
322 im[degree - 1] = 0.0;
323 } else {
324 solve_quadratic(p[0], p[1], p[2], &re[degree - 2], &im[degree - 2],
325 &re[degree - 1], &im[degree - 1]);
326 }
327 break;
328 }
329 moduli_max = 0.0;
330 moduli_min = LDBL_MAX;
331 for (int i = 0; i < nn; i++) {
332 x = fabsl(p[i]);
333 if (x > moduli_max) moduli_max = x;
334 if (x != 0 && x < moduli_min) moduli_min = x;
335 }
336 sc = LDBL_MIN / LDBL_EPSILON / moduli_min;
337 if ((sc <= 1.0 && moduli_max >= 10) ||
338 (sc > 1.0 && LDBL_MAX / sc >= moduli_max)) {
339 sc = (sc == 0) ? LDBL_MIN : sc;
340 factor = powl(2.0, logl(sc) / logl(2.0));
341 if (factor != 1.0)
342 for (int i = 0; i < nn; i++) p[i] *= factor;
343 }
344 for (int i = 0; i < nn; i++) pt[i] = fabsl(p[i]);
345 pt[n] = -pt[n];
346 nm1 = n - 1;
347 x = expl((logl(-pt[n]) - logl(pt[0])) / (LD)n);
348 if (pt[nm1] != 0) {
349 xm = -pt[n] / pt[nm1];
350 if (xm < x) x = xm;
351 }
352 xm = x;
353 do {
354 x = xm;
355 xm = 0.1 * x;
356 ff = pt[0];
357 for (int i = 1; i < nn; i++) ff = ff * xm + pt[i];
358 } while (ff > 0);
359 dx = x;
360 do {
361 df = ff = pt[0];
362 for (int i = 1; i < n; i++) {
363 ff = x * ff + pt[i];
364 df = x * df + ff;
365 }
366 ff = x * ff + pt[n];
367 dx = ff / df;
368 x -= dx;
369 } while (fabsl(dx / x) > 0.005);
370 bnd = x;
371 for (int i = 1; i < n; i++)
372 k[i] = (LD)(n - i) * p[i] / (LD)n;
373 k[0] = p[0];
374 aa = p[n];
375 bb = p[nm1];
376 zero = (k[nm1] == 0) ? 1 : 0;
377 for (jj = 0; jj < 5; jj++) {
378 cc = k[nm1];
379 if (zero) {
380 for (int i = 0; i < nm1; i++) {
381 j = nm1 - i;
382 k[j] = k[j - 1];
383 }
384 k[0] = 0;
385 zero = (k[nm1] == 0) ? 1 : 0;
386 } else {
387 t = -aa / cc;
388 for (int i = 0; i < nm1; i++) {
389 j = nm1 - i;
390 k[j] = t * k[j - 1] + p[j];
391 }
392 k[0] = p[0];
393 zero = (fabsl(k[nm1]) <= fabsl(bb) * LDBL_EPSILON * 10.0) ? 1 : 0;
394 }
395 }
396 for (int i = 0; i < n; i++) temp[i] = k[i];
397 static const LD DEG = 0.01745329251994329576923690768489L;
398 for (jj = 1; jj <= 20; jj++) {
399 xxx = -sinl(94.0 * DEG) * yy + cosl(94.0 * DEG) * xx;
400 yy = sinl(94.0 * DEG) * xx + cosl(94.0 * DEG) * yy;
401 xx = xxx;
402 sr = bnd * xx;
403 u = -2.0 * sr;
404 for (int i = 0; i < nn; i++) qk[i] = svk[i] = 0.0;
405 solve_fixedshift(20 * jj, &nz, sr, bnd, k, n, p, nn, qp, u,
406 qk, svk, &lzi, &lzr, &szi, &szr);
407 if (nz != 0) {
408 j = degree - n;
409 re[j] = szr;
410 im[j] = szi;
411 nn = nn - nz;
412 n = nn - 1;
413 for (int i = 0; i < nn; i++) p[i] = qp[i];
414 if (nz != 1) {
415 re[j + 1] = lzr;
416 im[j + 1] = lzi;
417 }
418 break;
419 } else {
420 for (int i = 0; i < n; i++) k[i] = temp[i];
421 }
422 }
423 if (jj > 20) break;
424 }
425 }
426
427 /*** Wrapper ***/
428
429 #include <algorithm> /* std::reverse(), std::sort() */
432
433 typedef std::complex<LD> root;
434
435 bool comp(const root & a, const root & b) {
436 if (real(a) != real(b)) return real(a) < real(b);
437 return imag(a) < imag(b);
438 }
439
440 std::vector<root> find_roots(int degree, LD coefficients[]) {
441 std::reverse(coefficients, coefficients + degree + 1);
442 LD re[degree], im[degree];
443 find_roots(degree, coefficients, re, im);
444 std::vector<root> res;
445 for (int i = 0; i < degree; i++)
446 res.push_back(root(re[i], im[i]));
447 std::sort(res.begin(), res.end(), comp);
448 return res;
449 }
450
451 /*** Example Usage (http://wcipeg.com/problem/rootsolve) ***/
452
455
456 int T, degree, p;
457 LD c, coeff[101];
458
459 int main() {
460 degree = 0;
461 cin >> T;
462 for (int i = 0; i < T; i++) {
463 cin >> c >> p;
464 if (p > degree) degree = p;
465 coeff[p] = c;
466 }
467 std::vector<root> roots = find_roots(degree, coeff);
468 bool printed = false;
470 for (int i = 0; i < (int)roots.size(); i++) {
471 if (fabsl(roots[i].imag()) < LDBL_EPSILON) {
472 cout << fixed << roots[i].real() << "\n";
473 printed = true;
474 }
475 }
476 if (!printed) cout << "NO REAL ROOTS\n";
477 return 0;
478 }
4.7 Integration
4.7.1 Simpsons Rule
1 /*
2
3 Simpsons rule is a method for numerical integration, the
4 numerical approximation of definite integrals. The rule is:
5
6 Integral of f(x) dx from a to b ~=
7 [f(a) + 4*f((a + b)/2) + f(b)] * (b - a)/6
8
9 */
10
12
13 template<class DoubleFunction>
14 double simpsons(DoubleFunction f, double a, double b) {
15 return (f(a) + 4 * f((a + b)/2) + f(b)) * (b - a)/6;
16 }
17
18 template<class DoubleFunction>
19 double integrate(DoubleFunction f, double a, double b) {
21 double m = (a + b) / 2;
22 double am = simpsons(f, a, m);
4.7. Integration 241
23 double mb = simpsons(f, m, b);

24 double ab = simpsons(f, a, b);
25 if (fabs(am + mb - ab) < eps) return ab;
26 return integrate(f, a, m) + integrate(f, m, b);
27 }
28
30
33
34 double f(double x) { return sin(x); }
35
36 int main () {
37 double PI = acos(-1.0);
38 cout << integrate(f, 0.0, PI/2) << "\n"; //1
39 return 0;
40 }
Chapter 5
Geometry
5.1 Geometric Classes
5.1.1 Point
1 /*
2
3 This class is very similar to std::complex, except it uses epsilon
4 comparisons and also supports other operations such as reflection
5 and rotation. In addition, this class supports many arithmetic
6 operations (e.g. overloaded operators for vector addition, subtraction,
7 multiplication, and division; dot/cross products, etc.) pertaining to
8 2D cartesian vectors.
9
10 All operations are O(1) in time and space.
11
12 */
13
14 #include <cmath> /* atan(), fabs(), sqrt() */
17
19
22
23 struct point {
24
25 double x, y;
26
27 point() : x(0), y(0) {}
28 point(const point & p) : x(p.x), y(p.y) {}
29 point(const std::pair<double, double> & p) : x(p.first), y(p.second) {}
30 point(const double & a, const double & b) : x(a), y(b) {}
31
32 bool operator < (const point & p) const {
33 return EQ(x, p.x) ? LT(y, p.y) : LT(x, p.x);
34 }
35
242
5.1. Geometric Classes 243
36 bool operator > (const point & p) const {

37 return EQ(x, p.x) ? LT(p.y, y) : LT(p.x, x);
38 }
39
40 bool operator == (const point & p) const { return EQ(x, p.x) && EQ(y, p.y); }
41 bool operator != (const point & p) const { return !(*this == p); }
42 bool operator <= (const point & p) const { return !(*this > p); }
43 bool operator >= (const point & p) const { return !(*this < p); }
44 point operator + (const point & p) const { return point(x + p.x, y + p.y); }
45 point operator - (const point & p) const { return point(x - p.x, y - p.y); }
46 point operator + (const double & v) const { return point(x + v, y + v); }
47 point operator - (const double & v) const { return point(x - v, y - v); }
48 point operator * (const double & v) const { return point(x * v, y * v); }
49 point operator / (const double & v) const { return point(x / v, y / v); }
50 point & operator += (const point & p) { x += p.x; y += p.y; return *this; }
51 point & operator -= (const point & p) { x -= p.x; y -= p.y; return *this; }
52 point & operator += (const double & v) { x += v; y += v; return *this; }
53 point & operator -= (const double & v) { x -= v; y -= v; return *this; }
54 point & operator *= (const double & v) { x *= v; y *= v; return *this; }
55 point & operator /= (const double & v) { x /= v; y /= v; return *this; }
56 friend point operator + (const double & v, const point & p) { return p + v; }
57 friend point operator * (const double & v, const point & p) { return p * v; }
58
59 double norm() const { return x * x + y * y; }
60 double abs() const { return sqrt(x * x + y * y); }
61 double arg() const { return atan2(y, x); }
62 double dot(const point & p) const { return x * p.x + y * p.y; }
63 double cross(const point & p) const { return x * p.y - y * p.x; }
64 double proj(const point & p) const { return dot(p) / p.abs(); } //onto p
65 point rot90() const { return point(-y, x); }
66
67 //proportional unit vector of (x, y) such that x^2 + y^2 = 1
68 point normalize() const {
69 return (EQ(x, 0) && EQ(y, 0)) ? point(0, 0) : (point(x, y) / abs());
70 }
71
72 //rotate t radians CW about origin
73 point rotateCW(const double & t) const {
74 return point(x * cos(t) + y * sin(t), y * cos(t) - x * sin(t));
75 }
76
77 //rotate t radians CCW about origin
78 point rotateCCW(const double & t) const {
79 return point(x * cos(t) - y * sin(t), x * sin(t) + y * cos(t));
80 }
81
82 //rotate t radians CW about point p
83 point rotateCW(const point & p, const double & t) const {
84 return (*this - p).rotateCW(t) + p;
85 }
86
87 //rotate t radians CCW about point p
88 point rotateCCW(const point & p, const double & t) const {
89 return (*this - p).rotateCCW(t) + p;
90 }
91
92 //reflect across point p
93 point reflect(const point & p) const {
94 return point(2 * p.x - x, 2 * p.y - y);
244 Chapter 5. Geometry
95 }
96
97 //reflect across the line containing points p and q
98 point reflect(const point & p, const point & q) const {
99 if (p == q) return reflect(p);
100 point r(*this - p), s = q - p;
101 r = point(r.x * s.x + r.y * s.y, r.x * s.y - r.y * s.x) / s.norm();
102 r = point(r.x * s.x - r.y * s.y, r.x * s.y + r.y * s.x) + p;
103 return r;
104 }
105
106 friend double norm(const point & p) { return p.norm(); }
107 friend double abs(const point & p) { return p.abs(); }
108 friend double arg(const point & p) { return p.arg(); }
109 friend double dot(const point & p, const point & q) { return p.dot(q); }
110 friend double cross(const point & p, const point & q) { return p.cross(q); }
111 friend double proj(const point & p, const point & q) { return p.proj(q); }
112 friend point rot90(const point & p) { return p.rot90(); }
113 friend point normalize(const point & p) { return p.normalize(); }
114 friend point rotateCW(const point & p, const double & t) { return p.rotateCW(t); }
115 friend point rotateCCW(const point & p, const double & t) { return p.rotateCCW(t); }
116 friend point rotateCW(const point & p, const point & q, const double & t) { return p.rotateCW(q, t); }
117 friend point rotateCCW(const point & p, const point & q, const double & t) { return p.rotateCCW(q, t);
}
118 friend point reflect(const point & p, const point & q) { return p.reflect(q); }
119 friend point reflect(const point & p, const point & a, const point & b) { return p.reflect(a, b); }
120
121 friend std::ostream & operator << (std::ostream & out, const point & p) {
122 out << "(";
123 out << (fabs(p.x) < eps ? 0 : p.x) << ",";
124 out << (fabs(p.y) < eps ? 0 : p.y) << ")";
125 return out;
126 }
127 };
128
130
132 #define pt point
133
134 const double PI = acos(-1.0);
135
136 int main() {
137 pt p(-10, 3);
138 assert(pt(-18, 29) == p + pt(-3, 9) * 6 / 2 - pt(-1, 1));
139 assert(EQ(109, p.norm()));
140 assert(EQ(10.44030650891, p.abs()));
141 assert(EQ(2.850135859112, p.arg()));
142 assert(EQ(0, p.dot(pt(3, 10))));
143 assert(EQ(0, p.cross(pt(10, -3))));
144 assert(EQ(10, p.proj(pt(-10, 0))));
145 assert(EQ(1, p.normalize().abs()));
146 assert(pt(-3, -10) == p.rot90());
147 assert(pt(3, 12) == p.rotateCW(pt(1, 1), PI / 2));
148 assert(pt(1, -10) == p.rotateCCW(pt(2, 2), PI / 2));
149 assert(pt(10, -3) == p.reflect(pt(0, 0)));
150 assert(pt(-10, -3) == p.reflect(pt(-2, 0), pt(5, 0)));
151 return 0;
152 }
5.1.2 Line
1 /*
2
3 A 2D line is expressed in the form Ax + By + C = 0. All lines can be
4 "normalized" to a canonical form by insisting that the y-coefficient
5 equal 1 if it is non-zero. Otherwise, we set the x-coefficient to 1.
6 If B is non-zero, then we have the common case where the slope = -A
7 after normalization.
8
10
11 */
12
17
18 const double eps = 1e-9, NaN = std::numeric_limits<double>::quiet_NaN();
19
22
23 typedef std::pair<double, double> point;
24 #define x first
25 #define y second
26
27 struct line {
28
29 double a, b, c;
30
31 line(): a(0), b(0), c(0) {} //invalid or uninitialized line
32
33 line(const double & A, const double & B, const double & C) {
34 a = A;
35 b = B;
36 c = C;
37 if (!EQ(b, 0)) {
38 a /= b; c /= b; b = 1;
39 } else {
40 c /= a; a = 1; b = 0;
41 }
42 }
43
44 line(const double & slope, const point & p) {
45 a = -slope;
46 b = 1;
47 c = slope * p.x - p.y;
48 }
49
50 line(const point & p, const point & q): a(0), b(0), c(0) {
51 if (EQ(p.x, q.x)) {
52 if (EQ(p.y, q.y)) return; //invalid line
53 //vertical line
54 a = 1;
55 b = 0;
56 c = -p.x;
57 return;
58 }
59 a = -(p.y - q.y) / (p.x - q.x);
60 b = 1;
61 c = -(a * p.x) - (b * p.y);
62 }
63
64 bool operator == (const line & l) const {
65 return EQ(a, l.a) && EQ(b, l.b) && EQ(c, l.c);
66 }
67
68 bool operator != (const line & l) const {
69 return !(*this == l);
70 }
71
72 //whether the line is initialized and normalized
73 bool valid() const {
74 if (EQ(a, 0)) return !EQ(b, 0);
75 return EQ(b, 1) || (EQ(b, 0) && EQ(a, 1));
76 }
77
78 bool horizontal() const { return valid() && EQ(a, 0); }
79 bool vertical() const { return valid() && EQ(b, 0); }
80
81 double slope() const {
82 if (!valid() || EQ(b, 0)) return NaN; //vertical
83 return -a;
84 }
85
86 //solve for x, given y
87 //for horizontal lines, either +inf, -inf, or nan is returned
88 double x(const double & y) const {
89 if (!valid() || EQ(a, 0)) return NaN; //invalid or horizontal
90 return (-c - b * y) / a;
91 }
92
93 //solve for y, given x
94 //for vertical lines, either +inf, -inf, or nan is returned
95 double y(const double & x) const {
96 if (!valid() || EQ(b, 0)) return NaN; //invalid or vertical
97 return (-c - a * x) / b;
98 }
99
100 //returns whether p exists on the line
101 bool contains(const point & p) const {
102 return EQ(a * p.x + b * p.y + c, 0);
103 }
104
105 //returns whether the line is parallel to l
106 bool parallel(const line & l) const {
107 return EQ(a, l.a) && EQ(b, l.b);
108 }
109
110 //returns whether the line is perpendicular to l
111 bool perpendicular(const line & l) const {
112 return EQ(-a * l.a, b * l.b);
113 }
114
115 //return the parallel line passing through point p
116 line parallel(const point & p) const {

117 return line(a, b, -a * p.x - b * p.y);
118 }
119
120 //return the perpendicular line passing through point p
121 line perpendicular(const point & p) const {
122 return line(-b, a, b * p.x - a * p.y);
123 }
124
125 friend std::ostream & operator << (std::ostream & out, const line & l) {
126 out << (fabs(l.a) < eps ? 0 : l.a) << "x" << std::showpos;
127 out << (fabs(l.b) < eps ? 0 : l.b) << "y";
128 out << (fabs(l.c) < eps ? 0 : l.c) << "=0" << std::noshowpos;
129 return out;
130 }
131 };
132
134
136
137 int main() {
138 line l(2, -5, -8);
139 line para = line(2, -5, -8).parallel(point(-6, -2));
140 line perp = line(2, -5, -8).perpendicular(point(-6, -2));
141 assert(l.parallel(para) && l.perpendicular(perp));
142 assert(l.slope() == 0.4);
143 assert(para == line(-0.4, 1, -0.4)); //-0.4x+1y-0.4=0
144 assert(perp == line(2.5, 1, 17)); //2.5x+1y+17=0
145 return 0;
146 }
5.1.3 Circle
1 /*
2
3 A 2D circle with center at (h, k) and a radius of r can be expressed by
4 the relation (x - h)^2 + (y - k)^2 = r^2. In the following definition,
5 the radius used to construct it is forced to be a positive number.
6
8
9 */
10
11 #include <cmath> /* fabs(), sqrt() */
15
17
21
23 #define x first
24 #define y second
25
26 double norm(const point & a) { return a.x * a.x + a.y * a.y; }
27 double abs(const point & a) { return sqrt(norm(a)); }
28
29 struct circle {
30
31 double h, k, r;
32
33 circle(): h(0), k(0), r(0) {}
34 circle(const double & R): h(0), k(0), r(fabs(R)) {}
35 circle(const point & o, const double & R): h(o.x), k(o.y), r(fabs(R)) {}
36 circle(const double & H, const double & K, const double & R):
37 h(H), k(K), r(fabs(R)) {}
38
39 //circumcircle with the diameter equal to the distance from a to b
40 circle(const point & a, const point & b) {
41 h = (a.x + b.x) / 2.0;
42 k = (a.y + b.y) / 2.0;
43 r = abs(point(a.x - h, a.y - k));
44 }
45
46 //circumcircle of 3 points - throws exception if abc are collinear/equal
47 circle(const point & a, const point & b, const point & c) {
48 double an = norm(point(b.x - c.x, b.y - c.y));
49 double bn = norm(point(a.x - c.x, a.y - c.y));
50 double cn = norm(point(a.x - b.x, a.y - b.y));
51 double wa = an * (bn + cn - an);
52 double wb = bn * (an + cn - bn);
53 double wc = cn * (an + bn - cn);
54 double w = wa + wb + wc;
55 if (fabs(w) < eps)
56 throw std::runtime_error("No circle from collinear points.");
57 h = (wa * a.x + wb * b.x + wc * c.x) / w;
58 k = (wa * a.y + wb * b.y + wc * c.y) / w;
60 }
61
62 //circle from 2 points and a radius - many possible edge cases!
63 //in the "normal" case, there will be 2 possible circles, one
64 //centered at (h1, k1) and the other (h2, k2). Only one is used.
65 //note that (h1, k1) equals (h2, k2) if dist(a, b) = 2 * r = d
66 circle(const point & a, const point & b, const double & R) {
67 r = fabs(R);
68 if (LE(r, 0) && a == b) { //circle is a point
69 h = a.x;
70 k = a.y;
71 return;
72 }
73 double d = abs(point(b.x - a.x, b.y - a.y));
74 if (EQ(d, 0))
75 throw std::runtime_error("Identical points, infinite circles.");
76 if (GT(d, r * 2.0))
77 throw std::runtime_error("Points too far away to make circle.");
78 double v = sqrt(r * r - d * d / 4.0) / d;
79 point m((a.x + b.x) / 2.0, (a.y + b.y) / 2.0);
80 h = m.x + (a.y - b.y) * v;
81 k = m.y + (b.x - a.x) * v;
82 //other answer is (h, k) = (m.x-(a.y-b.y)*v, m.y-(b.x-a.x)*v)
83 }
84
85 bool operator == (const circle & c) const {
86 return EQ(h, c.h) && EQ(k, c.k) && EQ(r, c.r);
87 }
88
89 bool operator != (const circle & c) const {
90 return !(*this == c);
91 }
92
94 return LE(norm(point(p.x - h, p.y - k)), r * r);
95 }
96
97 bool on_edge(const point & p) const {
98 return EQ(norm(point(p.x - h, p.y - k)), r * r);
99 }
100
101 point center() const {
102 return point(h, k);
103 }
104
105 friend std::ostream & operator << (std::ostream & out, const circle & c) {
106 out << std::showpos;
107 out << "(x" << -(fabs(c.h) < eps ? 0 : c.h) << ")^2+";
108 out << "(y" << -(fabs(c.k) < eps ? 0 : c.k) << ")^2";
109 out << std::noshowpos;
110 out << "=" << (fabs(c.r) < eps ? 0 : c.r * c.r);
111 return out;
112 }
113 };
114
115 //circle inscribed within points a, b, and c
116 circle incircle(const point & a, const point & b, const point & c) {
117 double al = abs(point(b.x - c.x, b.y - c.y));
118 double bl = abs(point(a.x - c.x, a.y - c.y));
119 double cl = abs(point(a.x - b.x, a.y - b.y));
120 double p = al + bl + cl;
121 if (EQ(p, 0)) return circle(a.x, a.y, 0);
122 circle res;
123 res.h = (al * a.x + bl * b.x + cl * c.x) / p;
124 res.k = (al * a.y + bl * b.y + cl * c.y) / p;
125 res.r = fabs((a.x - c.x) * (b.y - c.y) - (a.y - c.y) * (b.x - c.x)) / p;
126 return res;
127 }
128
130
132
133 int main() {
134 circle c(-2, 5, sqrt(10)); //(x+2)^2+(y-5)^2=10
135 assert(c == circle(point(-2, 5), sqrt(10)));
136 assert(c == circle(point(1, 6), point(-5, 4)));
137 assert(c == circle(point(-3, 2), point(-3, 8), point(-1, 8)));
138 assert(c == incircle(point(-12, 5), point(3, 0), point(0, 9)));
139 assert(c.contains(point(-2, 8)) && !c.contains(point(-2, 9)));
140 assert(c.on_edge(point(-1, 2)) && !c.on_edge(point(-1.01, 2)));
141 return 0;
142 }
5.2 Geometric Calculations
5.2.1 Angles
1 /*
2
3 Angle calculations in 2 dimensions. All returned angles are in radians,
4 except for reduce_deg(). If x is an angle in radians, then you may use
5 x * DEG to convert x to degrees, and vice versa to radians with x * RAD.
6
8
9 */
10
11 #include <cmath> /* acos(), fabs(), sqrt(), atan2() */
13
15 #define x first
16 #define y second
17
18 const double PI = acos(-1.0), RAD = 180 / PI, DEG = PI / 180;
19
20 double abs(const point & a) { return sqrt(a.x * a.x + a.y * a.y); }
21
22 //reduce angles to the range [0, 360) degrees. e.g. reduce_deg(-630) = 90
23 double reduce_deg(const double & t) {
24 if (t < -360) return reduce_deg(fmod(t, 360));
25 if (t < 0) return t + 360;
26 return t >= 360 ? fmod(t, 360) : t;
27 }
28
29 //reduce angles to the range [0, 2*pi) radians. e.g. reduce_rad(720.5) = 0.5
30 double reduce_rad(const double & t) {
31 if (t < -2 * PI) return reduce_rad(fmod(t, 2 * PI));
32 if (t < 0) return t + 2 * PI;
33 return t >= 2 * PI ? fmod(t, 2 * PI) : t;
34 }
35
36 //like std::polar(), but returns a point instead of an std::complex
37 point polar_point(const double & r, const double & theta) {
38 return point(r * cos(theta), r * sin(theta));
39 }
40
41 //angle of segment (0, 0) to p, relative (CCW) to the +ve x-axis in radians
42 double polar_angle(const point & p) {
43 double t = atan2(p.y, p.x);
44 return t < 0 ? t + 2 * PI : t;
45 }
46
47 //smallest angle formed by points aob (angle is at point o) in radians
48 double angle(const point & a, const point & o, const point & b) {
49 point u(o.x - a.x, o.y - a.y), v(o.x - b.x, o.y - b.y);
50 return acos((u.x * v.x + u.y * v.y) / (abs(u) * abs(v)));
5.2. Geometric Calculations 251
51 }
52
53 //angle of line segment ab relative (CCW) to the +ve x-axis in radians
54 double angle_between(const point & a, const point & b) {
55 double t = atan2(a.x * b.y - a.y * b.x, a.x * b.x + a.y * b.y);
56 return t < 0 ? t + 2 * PI : t;
57 }
58
59 //Given the A, B values of two lines in Ax + By + C = 0 form, finds the
60 //minimum angle in radians between the two lines in the range [0, PI/2]
61 double angle_between(const double & a1, const double & b1,
62 const double & a2, const double & b2) {
63 double t = atan2(a1 * b2 - a2 * b1, a1 * a2 + b1 * b2);
64 if (t < 0) t += PI; //force angle to be positive
65 if (t > PI / 2) t = PI - t; //force angle to be <= 90 degrees
66 return t;
67 }
68
69 //magnitude of the 3D cross product with Z component implicitly equal to 0
70 //the answer assumes the origin (0, 0) is instead shifted to point o.
71 //this is equal to 2x the signed area of the triangle from these 3 points.
72 double cross(const point & o, const point & a, const point & b) {
73 return (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x);
74 }
75
76 //does the path a->o->b form:
77 // -1 ==> a left turn on the plane?
78 // 0 ==> a single straight line segment? (i.e. are a,o,b collinear?) or
79 // +1 ==> a right turn on the plane?
80 //warning: the order of parameters is a,o,b, and NOT o,a,b as in cross()
81 int turn(const point & a, const point & o, const point & b) {
82 double c = cross(o, a, b);
83 return c < 0 ? -1 : (c > 0 ? 1 : 0);
84 }
85
87
89 #define pt point
90 #define EQ(a, b) (fabs((a) - (b)) <= 1e-9)
91
92 int main() {
93 assert(EQ(123, reduce_deg(-(8 * 360) + 123)));
94 assert(EQ(1.2345, reduce_rad(2 * PI * 8 + 1.2345)));
95 point p = polar_point(4, PI), q = polar_point(4, -PI / 2);
96 assert(EQ(p.x, -4) && EQ(p.y, 0));
97 assert(EQ(q.x, 0) && EQ(q.y, -4));
98 assert(EQ(45, polar_angle(pt(5, 5)) * RAD));
99 assert(EQ(135, polar_angle(pt(-4, 4)) * RAD));
100 assert(EQ(90, angle(pt(5, 0), pt(0, 5), pt(-5, 0)) * RAD));
101 assert(EQ(225, angle_between(pt(0, 5), pt(5, -5)) * RAD));
102 assert(EQ(90, angle_between(-1, 1, -1, -1) * RAD)); //y=x and y=-x
103 assert(-1 == cross(pt(0, 0), pt(0, 1), pt(1, 0)));
104 assert(+1 == turn(pt(0, 1), pt(0, 0), pt(-5, -5)));
105 return 0;
106 }
5.2.2 Distances
1 /*
2
3 Distance calculations in 2 dimensions between points, lines, and segments.
5
6 */
7
11
13 #define x first
14 #define y second
15
17
21
24
25 //distance from point a to point b
26 double dist(const point & a, const point & b) {
27 return abs(point(b.x - a.x, b.y - a.y));
28 }
29
30 //squared distance from point a to point b
31 double dist2(const point & a, const point & b) {
32 return norm(point(b.x - a.x, b.y - a.y));
33 }
34
35 //minimum distance from point p to line l denoted by ax + by + c = 0
36 //if a = b = 0, then -inf, nan, or +inf is returned depending on sgn(c)
37 double dist_line(const point & p,
38 const double & a, const double & b, const double & c) {
39 return fabs(a * p.x + b * p.y + c) / sqrt(a * a + b * b);
40 }
41
42 //minimum distance from point p to the infinite line containing a and b
43 //if a = b, then the point distance from p to the single point is returned
44 double dist_line(const point & p, const point & a, const point & b) {
45 double ab2 = dist2(a, b);
46 if (EQ(ab2, 0)) return dist(p, a);
47 double u = ((p.x - a.x) * (b.x - a.x) + (p.y - a.y) * (b.y - a.y)) / ab2;
48 return abs(point(a.x + u * (b.x - a.x) - p.x, a.y + u * (b.y - a.y) - p.y));
49 }
50
51 //distance between two lines each denoted by the form ax + by + c = 0
52 //if the lines are nonparallel, then the distance is 0, otherwise
53 //it is the perpendicular distance from a point on one line to the other
54 double dist_lines(const double & a1, const double & b1, const double & c1,
55 const double & a2, const double & b2, const double & c2) {
56 if (EQ(a1 * b2, a2 * b1)) {
57 double factor = EQ(b1, 0) ? (a1 / a2) : (b1 / b2);

58 if (EQ(c1, c2 * factor)) return 0;
59 return fabs(c2 * factor - c1) / sqrt(a1 * a1 + b1 * b1);
60 }
61 return 0;
62 }
63
64 //distance between two infinite lines respectively containing ab and cd
65 //same results as above, except we solve for the lines here first.
66 double dist_lines(const point & a, const point & b,
67 const point & c, const point & d) {
68 double A1 = a.y - b.y, B1 = b.x - a.x;
69 double A2 = c.y - d.y, B2 = d.x - c.x;
70 double C1 = -A1 * a.x - B1 * a.y, C2 = -A2 * c.x - B2 * c.y;
71 return dist_lines(A1, B1, C1, A2, B2, C2);
72 }
73
74 //minimum distance from point p to any point on segment ab
75 double dist_seg(const point & p, const point & a, const point & b) {
76 if (a == b) return dist(p, a);
77 point ab(b.x - a.x, b.y - a.y), ap(p.x - a.x, p.y - a.y);
78 double n = norm(ab), d = ab.x * ap.x + ab.y * ap.y;
79 if (LE(d, 0) || EQ(n, 0)) return abs(ap);
80 if (GE(d, n)) return abs(point(ap.x - ab.x, ap.y - ab.y));
81 return abs(point(ap.x - ab.x * (d / n), ap.y - ab.y * (d / n)));
82 }
83
84 double dot(const point & a, const point & b) { return a.x * b.x + a.y * b.y; }
85 double cross(const point & a, const point & b) { return a.x * b.y - a.y * b.x; }
86
87 //minimum distance from any point on segment ab to any point on segment cd
88 double dist_segs(const point & a, const point & b,
90 //check if segments are touching or intersecting - if so, distance is 0
91 point ab(b.x - a.x, b.y - a.y);
92 point ac(c.x - a.x, c.y - a.y);
93 point cd(d.x - c.x, d.y - c.y);
94 double c1 = cross(ab, cd), c2 = cross(ac, ab);
95 if (EQ(c1, 0) && EQ(c2, 0)) {
96 double t0 = dot(ac, ab) / norm(ab);
97 double t1 = t0 + dot(cd, ab) / norm(ab);
98 if (LE(std::min(t0, t1), 1) && LE(0, std::max(t0, t1)))
99 return 0;
100 } else {
101 double t = cross(ac, cd) / c1, u = c2 / c1;
102 if (!EQ(c1, 0) && LE(0, t) && LE(t, 1) && LE(0, u) && LE(u, 1))
103 return 0;
104 }
105 //find min distances across each endpoint to opposing segment
106 return std::min(std::min(dist_seg(a, c, d), dist_seg(b, c, d)),
107 std::min(dist_seg(c, a, b), dist_seg(d, a, b)));
108 }
109
111
114
115 int main() {
116 assert(EQ(5, dist(pt(-1, -1), pt(2, 3))));

117 assert(EQ(25, dist2(pt(-1, -1), pt(2, 3))));
118 assert(EQ(1.2, dist_line(pt(2, 1), -4, 3, -1)));
119 assert(EQ(0.8, dist_line(pt(3, 3), pt(-1, -1), pt(2, 3))));
120 assert(EQ(1.2, dist_line(pt(2, 1), pt(-1, -1), pt(2, 3))));
121 assert(EQ(0.0, dist_lines(-4, 3, -1, 8, 6, 2)));
122 assert(EQ(0.8, dist_lines(-4, 3, -1, -8, 6, -10)));
123 assert(EQ(1.0, dist_seg(pt(3, 3), pt(-1, -1), pt(2, 3))));
124 assert(EQ(1.2, dist_seg(pt(2, 1), pt(-1, -1), pt(2, 3))));
125 assert(EQ(0.0, dist_segs(pt(0, 2), pt(3, 3), pt(-1, -1), pt(2, 3))));
126 assert(EQ(0.6, dist_segs(pt(-1, 0), pt(-2, 2), pt(-1, -1), pt(2, 3))));
127 return 0;
128 }
5.2.3 Line Intersections
1 /*
2
3 Intersections between straight lines, as well as between line segments
4 in 2 dimensions. Also included are functions to determine the closest
5 point to a line, which is done by finding the intersection through the
6 perpendicular. Note that you should modify the TOUCH_IS_INTERSECT flag
7 used for line segment intersection, depending on whether you wish for
8 the algorithm to consider barely touching segments to intersect.
9
11
12 */
13
14 #include <algorithm> /* std::min(), std::max() */
17
19 #define x first
20 #define y second
21
23
27
28 //intersection of line l1 and line l2, each in ax + by + c = 0 form
29 //returns: -1, if lines do not intersect,
30 // 0, if there is exactly one intersection point, or
31 // +1, if there are infinite intersection
32 //in the 2nd case, the intersection point is optionally stored into p
33 int line_intersection(const double & a1, const double & b1, const double & c1,
34 const double & a2, const double & b2, const double & c2,
35 point * p = 0) {
36 if (EQ(a1 * b2, a2 * b1))
37 return (EQ(a1 * c2, a2 * c1) || EQ(b1 * c2, b2 * c1)) ? 1 : -1;
38 if (p != 0) {
39 p->x = (b1 * c1 - b1 * c2) / (a2 * b1 - a1 * b2);
40 if (!EQ(b1, 0)) p->y = -(a1 * p->x + c1) / b1;
41 else p->y = -(a2 * p->x + c2) / b2;
42 }
43 return 0;
44 }
45
46 //intersection of line through p1, p2, and line through p2, p3
47 //returns: -1, if lines do not intersect,
48 // 0, if there is exactly one intersection point, or
49 // +1, if there are infinite intersections
50 //in the 2nd case, the intersection point is optionally stored into p
51 int line_intersection(const point & p1, const point & p2,
52 const point & p3, const point & p4, point * p = 0) {
53 double a1 = p2.y - p1.y, b1 = p1.x - p2.x;
54 double c1 = -(p1.x * p2.y - p2.x * p1.y);
57 double x = -(c1 * b2 - c2 * b1), y = -(a1 * c2 - a2 * c1);
58 double det = a1 * b2 - a2 * b1;
59 if (EQ(det, 0))
60 return (EQ(x, 0) && EQ(y, 0)) ? 1 : -1;
61 if (p != 0) *p = point(x / det, y / det);
62 return 0;
63 }
64
65 //Line Segment Intersection (http://stackoverflow.com/a/565282)
66
71
72 //should we consider barely touching segments an intersection?
73 const bool TOUCH_IS_INTERSECT = true;
74
75 //does [l, h] contain m?
76 //precondition: l <= h
77 bool contain(const double & l, const double & m, const double & h) {
78 if (TOUCH_IS_INTERSECT) return LE(l, m) && LE(m, h);
79 return LT(l, m) && LT(m, h);
80 }
81
82 //does [l1, h1] overlap with [l2, h2]?
83 //precondition: l1 <= h1 and l2 <= h2
84 bool overlap(const double & l1, const double & h1,
85 const double & l2, const double & h2) {
86 if (TOUCH_IS_INTERSECT) return LE(l1, h2) && LE(l2, h1);
87 return LT(l1, h2) && LT(l2, h1);
88 }
89
90 //intersection of line segment ab with line segment cd
91 //returns: -1, if segments do not intersect,
92 // 0, if there is exactly one intersection point
93 // +1, if the intersection is another line segment
94 //In case 2, the intersection point is stored into p
95 //In case 3, the intersection segment is stored into p and q
96 int seg_intersection(const point & a, const point & b,
97 const point & c, const point & d,
98 point * p = 0, point * q = 0) {

103 if (EQ(c1, 0) && EQ(c2, 0)) { //collinear
106 if (overlap(std::min(t0, t1), std::max(t0, t1), 0, 1)) {
107 point res1 = std::max(std::min(a, b), std::min(c, d));
108 point res2 = std::min(std::max(a, b), std::max(c, d));
109 if (res1 == res2) {
110 if (p != 0) *p = res1;
111 return 0; //collinear, meeting at an endpoint
112 }
113 if (p != 0 && q != 0) *p = res1, *q = res2;
114 return 1; //collinear and overlapping
115 } else {
116 return -1; //collinear and disjoint
117 }
118 }
119 if (EQ(c1, 0)) return -1; //parallel and disjoint
121 if (contain(0, t, 1) && contain(0, u, 1)) {
122 if (p != 0) *p = point(a.x + t * ab.x, a.y + t * ab.y);
123 return 0; //non-parallel with one intersection
124 }
125 return -1; //non-parallel with no intersections
126 }
127
128 //determines the point on line ax + by + c = 0 that is closest to point p
129 //this always lies on the line through p perpendicular to l.
130 point closest_point(const double & a, const double & b, const double & c,
131 const point & p) {
132 if (EQ(a, 0)) return point(p.x, -c); //horizontal line
133 if (EQ(b, 0)) return point(-c, p.y); //vertical line
134 point res;
135 line_intersection(a, b, c, -b, a, b * p.x - a * p.y, &res);
136 return res;
137 }
138
139 //determines the point on segment ab closest to point p
140 point closest_point(const point & a, const point & b, const point & p) {
141 if (a == b) return a;
142 point ap(p.x - a.x, p.y - a.y), ab(b.x - a.x, b.y - a.y);
143 double t = dot(ap, ab) / norm(ab);
144 if (t <= 0) return a;
145 if (t >= 1) return b;
146 return point(a.x + t * ab.x, a.y + t * ab.y);
147 }
148
150
153
154 int main() {
155 point p;
156 assert(line_intersection(-1, 1, 0, 1, 1, -3, &p) == 0);
157 assert(p == pt(1.5, 1.5));
158 assert(line_intersection(pt(0, 0), pt(1, 1), pt(0, 4), pt(4, 0), &p) == 0);
159 assert(p == pt(2, 2));
160
161 //tests for segment intersection (examples in order from link below)
162 //http://martin-thoma.com/how-to-check-if-two-line-segments-intersect/
163 {
164 #define test(a,b,c,d,e,f,g,h) seg_intersection(pt(a,b),pt(c,d),pt(e,f),pt(g,h),&p,&q)
165 pt p, q;
166 //intersection is a point
167 assert(0 == test(-4, 0, 4, 0, 0, -4, 0, 4)); assert(p == pt(0, 0));
168 assert(0 == test(0, 0, 10, 10, 2, 2, 16, 4)); assert(p == pt(2, 2));
169 assert(0 == test(-2, 2, -2, -2, -2, 0, 0, 0)); assert(p == pt(-2, 0));
170 assert(0 == test(0, 4, 4, 4, 4, 0, 4, 8)); assert(p == pt(4, 4));
171
172 //intersection is a segment
173 assert(1 == test(10, 10, 0, 0, 2, 2, 6, 6));
174 assert(p == pt(2, 2) && q == pt(6, 6));
175 assert(1 == test(6, 8, 14, -2, 14, -2, 6, 8));
176 assert(p == pt(6, 8) && q == pt(14, -2));
177
178 //no intersection
179 assert(-1 == test(6, 8, 8, 10, 12, 12, 4, 4));
180 assert(-1 == test(-4, 2, -8, 8, 0, 0, -4, 6));
181 assert(-1 == test(4, 4, 4, 6, 0, 2, 0, 0));
182 assert(-1 == test(4, 4, 6, 4, 0, 2, 0, 0));
183 assert(-1 == test(-2, -2, 4, 4, 10, 10, 6, 6));
184 assert(-1 == test(0, 0, 2, 2, 4, 0, 1, 4));
185 assert(-1 == test(2, 2, 2, 8, 4, 4, 6, 4));
186 assert(-1 == test(4, 2, 4, 4, 0, 8, 10, 0));
187 }
188 assert(pt(2.5, 2.5) == closest_point(-1, -1, 5, pt(0, 0)));
189 assert(pt(3, 0) == closest_point(1, 0, -3, pt(0, 0)));
190 assert(pt(0, 3) == closest_point(0, 1, -3, pt(0, 0)));
191
192 assert(pt(3, 0) == closest_point(pt(3, 0), pt(3, 3), pt(0, 0)));
193 assert(pt(2, -1) == closest_point(pt(2, -1), pt(4, -1), pt(0, 0)));
194 assert(pt(4, -1) == closest_point(pt(2, -1), pt(4, -1), pt(5, 0)));
195 return 0;
196 }
5.2.4 Circle Intersections
1 /*
2
3 Tangent lines to circles, circle-line intersections, and circle-circle
4 intersections (intersection point(s) as well as area) in 2 dimensions.
5
7
8 */
9
11 #include <cmath> /* acos(), fabs(), sqrt() */
13
15 #define x first
16 #define y second
17

19
21 #define NE(a, b) (fabs((a) - (b)) > eps) /* not equal to */
26
27 struct circle {
28 double h, k, r;
29
30 circle(const double & h, const double & k, const double & r) {
31 this->h = h;
32 this->k = k;
33 this->r = r;
34 }
35 };
36
37 //note: this is a simplified version of line that is not canonicalized.
38 // e.g. comparing lines with == signs will not work as intended. For a
39 // fully featured line class, see the whole geometry library.
40 struct line {
41 double a, b, c;
42
43 line() { a = b = c = 0; }
44
45 line(const double & a, const double & b, const double & c) {
46 this->a = a;
47 this->b = b;
48 this->c = c;
49 }
50
51 line(const point & p, const point & q) {
52 a = p.y - q.y,
53 b = q.x - p.x;
54 c = -a * p.x - b * p.y;
55 }
56 };
57
61
62 //tangent line(s) to circle c passing through p. there are 3 cases:
63 //returns: 0, if there are no lines (p is strictly inside c)
64 // 1, if there is 1 tangent line (p is on the edge)
65 // 2, if there are 2 tangent lines (p is strictly outside)
66 //If there is only 1 tangent, then the line will be stored in l1.
67 //If there are 2, then they will be stored in l1 and l2 respectively.
68 int tangents(const circle & c, const point & p, line * l1 = 0, line * l2 = 0) {
69 point vop(p.x - c.h, p.y - c.k);
70 if (EQ(norm(vop), c.r * c.r)) { //on an edge, get perpendicular through p
71 if (l1 != 0) {
72 *l1 = line(point(c.h, c.k), p);
73 *l1 = line(-l1->b, l1->a, l1->b * p.x - l1->a * p.y);
74 }
75 return 1;
76 }
77 if (LE(norm(vop), c.r * c.r)) return 0; //inside circle

78 point q(vop.x / c.r, vop.y / c.r);
79 double n = norm(q), d = q.y * sqrt(norm(q) - 1.0);
80 point t1((q.x - d) / n, c.k), t2((q.x + d) / n, c.k);
81 if (NE(q.y, 0)) { //common case
82 t1.y += c.r * (1.0 - t1.x * q.x) / q.y;
83 t2.y += c.r * (1.0 - t2.x * q.x) / q.y;
84 } else { //point at center horizontal, y = 0
85 d = c.r * sqrt(1.0 - t1.x * t1.x);
86 t1.y += d;
87 t2.y -= d;
88 }
89 t1.x = t1.x * c.r + c.h;
90 t2.x = t2.x * c.r + c.h;
91 //note: here, t1 and t2 are the two points of tangencies
92 if (l1 != 0) *l1 = line(p, t1);
93 if (l2 != 0) *l2 = line(p, t2);
94 return 2;
95 }
96
97 //determines the intersection(s) between a circle c and line l
98 //returns: 0, if the line does not intersect with the circle
99 // 1, if the line is tangent (one intersection)
100 // 2, if the line crosses through the circle
101 //If there is 1 intersection point, it will be stored in p
102 //If there are 2, they will be stored in p and q respectively
103 int intersection(const circle & c, const line & l,
104 point * p = 0, point * q = 0) {
105 double v = c.h * l.a + c.k * l.b + l.c;
106 double aabb = l.a * l.a + l.b * l.b;
107 double disc = v * v / aabb - c.r * c.r;
108 if (disc > eps) return 0;
109 double x0 = -l.a * l.c / aabb, y0 = -l.b * v / aabb;
110 if (disc > -eps) {
111 if (p != 0) *p = point(x0 + c.h, y0 + c.k);
112 return 1;
113 }
114 double k = sqrt((disc /= -aabb) < 0 ? 0 : disc);
115 if (p != 0) *p = point(x0 + k * l.b + c.h, y0 - k * l.a + c.k);
116 if (q != 0) *q = point(x0 - k * l.b + c.h, y0 + k * l.a + c.k);
117 return 2;
118 }
119
120 //determines the intersection points between two circles c1 and c2
121 //returns: -2, if circle c2 completely encloses circle c1
122 // -1, if circle c1 completely encloses circle c2
123 // 0, if the circles are completely disjoint
124 // 1, if the circles are tangent (one intersection point)
125 // 2, if the circles intersect at two points
126 // 3, if the circles intersect at infinite points (c1 = c2)
127 //If one intersection, the intersection point is stored in p
128 //If two, the intersection points are stored in p and q respectively
129 int intersection(const circle & c1, const circle & c2,
130 point * p = 0, point * q = 0) {
131 if (EQ(c1.h, c2.h) && EQ(c1.k, c2.k))
132 return EQ(c1.r, c2.r) ? 3 : (c1.r > c2.r ? -1 : -2);
133 point d12(point(c2.h - c1.h, c2.k - c1.k));
134 double d = abs(d12);
135 if (GT(d, c1.r + c2.r)) return 0;
136 if (LT(d, fabs(c1.r - c2.r))) return c1.r > c2.r ? -1 : -2;

137 double a = (c1.r * c1.r - c2.r * c2.r + d * d) / (2 * d);
138 double x0 = c1.h + (d12.x * a / d);
139 double y0 = c1.k + (d12.y * a / d);
140 double s = sqrt(c1.r * c1.r - a * a);
141 double rx = -d12.y * s / d, ry = d12.x * s / d;
142 if (EQ(rx, 0) && EQ(ry, 0)) {
143 if (p != 0) *p = point(x0, y0);
144 return 1;
145 }
146 if (p != 0) *p = point(x0 - rx, y0 - ry);
147 if (q != 0) *q = point(x0 + rx, y0 + ry);
148 return 2;
149 }
150
151 const double PI = acos(-1.0);
152
153 //intersection area of circles c1 and c2
154 double intersection_area(const circle & c1, const circle c2) {
155 double r = std::min(c1.r, c2.r), R = std::max(c1.r, c2.r);
156 double d = abs(point(c2.h - c1.h, c2.k - c1.k));
157 if (LE(d, R - r)) return PI * r * r;
158 if (GE(d, R + r)) return 0;
159 return r * r * acos((d * d + r * r - R * R) / 2 / d / r) +
160 R * R * acos((d * d + R * R - r * r) / 2 / d / R) -
161 0.5 * sqrt((-d + r + R) * (d + r - R) * (d - r + R) * (d + r + R));
162 }
163
165
170
171 int main() {
172 line l1, l2;
173 assert(0 == tangents(circle(0, 0, 4), pt(1, 1), &l1, &l2));
174 assert(1 == tangents(circle(0, 0, sqrt(2)), pt(1, 1), &l1, &l2));
175 cout << l1.a << " " << l1.b << " " << l1.c << "\n"; // -x - y + 2 = 0
176 assert(2 == tangents(circle(0, 0, 2), pt(2, 2), &l1, &l2));
177 cout << l1.a << " " << l1.b << " " << l1.c << "\n"; // -2y + 4 = 0
178 cout << l2.a << " " << l2.b << " " << l2.c << "\n"; // 2x - 4 = 0
179
180 pt p, q;
181 assert(0 == intersection(circle(1, 1, 3), line(5, 3, -30), &p, &q));
183 assert(p == pt(1, 4));
185 assert(p == pt(4, 1));
186 assert(q == pt(-2, 1));
187
188 assert(-2 == intersection(circle(1, 1, 1), circle(0, 0, 3), &p, &q));
189 assert(-1 == intersection(circle(0, 0, 3), circle(1, 1, 1), &p, &q));
190 assert(0 == intersection(circle(5, 0, 4), circle(-5, 0, 4), &p, &q));
191 assert(1 == intersection(circle(-5, 0, 5), circle(5, 0, 5), &p, &q));
192 assert(p == pt(0, 0));
193 assert(2 == intersection(circle(-0.5, 0, 1), circle(0.5, 0, 1), &p, &q));
194 assert(p == pt(0, -sqrt(3) / 2));
5.3. Common Geometric Computations 261
195 assert(q == pt(0, sqrt(3) / 2));

196
197 //example where each circle passes through the other circles center
198 //http://math.stackexchange.com/a/402891
199 double r = 3;
200 double a = intersection_area(circle(-r / 2, 0, r), circle(r / 2, 0, r));
201 assert(EQ(a, r * r * (2 * PI / 3 - sqrt(3) / 2)));
202 return 0;
203 }
5.3 Common Geometric Computations
5.3.1 Polygon Sorting and Area
1 /*
2
3 centroid() - Simply returns the geometric average point of all the
4 points given. This could be used to find the reference center point
5 for the following function. An empty range will result in (0, 0).
6 Complexity: O(n) on the number of points in the given range.
7
8 cw_comp() - Given a set of points, these points could possibly form
9 many different polygons. The following sorting comparators, when
10 used in conjunction with std::sort, will produce one such ordering
11 of points which is sorted in clockwise order relative to a custom-
12 defined center point that must be set beforehand. This could very
13 well be the result of mean_point(). ccw_comp() is the opposite
14 function, which produces the points in counterclockwise order.
15 Complexity: O(1) per call.
16
17 polygon_area() - A given range of points is interpreted as a polygon
18 based on the ordering theyre given in. The shoelace formula is used
19 to determine its area. The polygon does not necessarily have to be
20 sorted using one of the functions above, but may be any ordering that
21 produces a valid polygon. You may optionally pass the last point in
22 the range equal to the first point and still expect the correct result.
23 Complexity: O(n) on the number of points in the range, assuming that
24 the points are already sorted in the order that specifies the polygon.
25
26 */
27
31
33 #define x first
34 #define y second
35
37
41

47 }
48
49 point ctr;
50
51 template<class It> point centroid(It lo, It hi) {
52 if (lo == hi) return point(0, 0);
53 double xtot = 0, ytot = 0, points = hi - lo;
54 for (; lo != hi; ++lo) {
55 xtot += lo->x;
56 ytot += lo->y;
57 }
58 return point(xtot / points, ytot / points);
59 }
60
61 //ctr must be defined beforehand
62 bool cw_comp(const point & a, const point & b) {
63 if (GE(a.x - ctr.x, 0) && LT(b.x - ctr.x, 0)) return true;
64 if (LT(a.x - ctr.x, 0) && GE(b.x - ctr.x, 0)) return false;
65 if (EQ(a.x - ctr.x, 0) && EQ(b.x - ctr.x, 0)) {
66 if (GE(a.y - ctr.y, 0) || GE(b.y - ctr.y, 0))
67 return a.y > b.y;
68 return b.y > a.y;
69 }
70 double det = cross(ctr, a, b);
71 if (EQ(det, 0))
72 return (a.x - ctr.x) * (a.x - ctr.x) + (a.y - ctr.y) * (a.y - ctr.y) >
73 (b.x - ctr.x) * (b.x - ctr.x) + (b.y - ctr.y) * (b.y - ctr.y);
74 return det < 0;
75 }
76
77 bool ccw_comp(const point & a, const point & b) {
78 return cw_comp(b, a);
79 }
80
81 //area of a polygon specified by range [lo, hi) - shoelace formula in O(n)
82 //[lo, hi) must point to the polygon vertices, sorted in CW or CCW order
83 template<class It> double polygon_area(It lo, It hi) {
84 if (lo == hi) return 0;
85 double area = 0;
86 if (*lo != *--hi)
87 area += (lo->x - hi->x) * (lo->y + hi->y);
88 for (It i = hi, j = hi - 1; i != lo; --i, --j)
89 area += (i->x - j->x) * (i->y + j->y);
90 return fabs(area / 2.0);
91 }
92
94
98 #define pt point
99
100 int main() {
101 //irregular pentagon with only (1, 2) not on the convex hull
102 //the ordering here is already sorted in ccw order around their centroid
103 //we will scramble them and see if our comparator works
104 pt pts[] = {pt(1, 3), pt(1, 2), pt(2, 1), pt(0, 0), pt(-1, 3)};
105 vector<pt> v(pts, pts + 5);
106 std::random_shuffle(v.begin(), v.end());
107 ctr = centroid(v.begin(), v.end()); //note: ctr is a global variable
108 assert(EQ(ctr.x, 0.6) && EQ(ctr.y, 1.8));
109 sort(v.begin(), v.end(), cw_comp);
110 for (int i = 0; i < (int)v.size(); i++) assert(v[i] == pts[i]);
111 assert(EQ(polygon_area(v.begin(), v.end()), 5));
112 return 0;
113 }
5.3.2 Point in Polygon Query
1 /*
2
3 Given a single point p and another range of points specifying a
4 polygon, determine whether p lies within the polygon. Note that
5 you should modify the EDGE_IS_INSIDE flag, depending on whether
6 you wish for the algorithm to consider points lying on an edge of
7 the polygon to be inside it.
8
9 Complexity: O(n) on the number of vertices in the polygon.
10
11 */
12
16
18 #define x first
19 #define y second
20
22
26
27 //should we consider points lying on an edge to be inside the polygon?
28 const bool EDGE_IS_INSIDE = true;
29
35 }
36
37 //return whether point p is in polygon specified by range [lo, hi) in O(n)
38 //[lo, hi) must point to the polygon vertices, sorted in CW or CCW order
39 template<class It> bool point_in_polygon(const point & p, It lo, It hi) {
40 int cnt = 0;
41 for (It i = lo, j = hi - 1; i != hi; j = i++) {
42 if (EQ(i->y, p.y) && (EQ(i->x, p.x) ||
43 (EQ(j->y, p.y) && (LE(i->x, p.x) || LE(j->x, p.x)))))

44 return EDGE_IS_INSIDE; //on an edge
45 if (GT(i->y, p.y) != GT(j->y, p.y)) {
46 double det = cross(p, *i, *j);
47 if (EQ(det, 0)) return EDGE_IS_INSIDE; //on an edge
48 if (GT(det, 0) != GT(j->y, i->y)) cnt++;
49 }
50 }
51 return cnt % 2 == 1;
52 }
53
55
58 #define pt point
59
60 int main() {
61 //irregular trapezoid
62 pt p[] = {pt(-1, 3), pt(1, 3), pt(2, 1), pt(0, 0)};
63 assert(point_in_polygon(pt(1, 2), p, p + 4));
64 assert(point_in_polygon(pt(0, 3), p, p + 4));
65 assert(!point_in_polygon(pt(0, 3.01), p, p + 4));
66 assert(!point_in_polygon(pt(2, 2), p, p + 4));
67 return 0;
68 }
5.3.3 Convex Hull
1 /*
2
3 Determines the convex hull from a range of points, that is, the
4 smallest convex polygon (a polygon such that every line which
5 crosses through it will only cross through it once) that contains
6 all of the points. This function uses the monotone chain algorithm
7 to compute the upper and lower hulls separately.
8
9 Returns: a vector of the convex hull points in clockwise order.
10 Complexity: O(n log n) on the number of points given
11
12 Notes: To yield the hull points in counterclockwise order,
13 replace every usage of GE() in the function with LE().
14 To have the first point on the hull repeated as the last,
15 replace the last line of the function to res.resize(k);
16
17 */
18
23
25 #define x first
26 #define y second
27
28 //change < 0 comparisons to > 0 to produce hull points in CCW order
29 double cw(const point & o, const point & a, const point & b) {
30 return (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x) < 0;
31 }
32
33 //convex hull from a range [lo, hi) of points
34 //monotone chain in O(n log n) to find hull points in CW order
35 //notes: the range of input points will be sorted lexicographically
36 template<class It> std::vector<point> convex_hull(It lo, It hi) {
37 int k = 0;
38 if (hi - lo <= 1) return std::vector<point>(lo, hi);
39 std::vector<point> res(2 * (int)(hi - lo));
40 std::sort(lo, hi); //compare by x, then by y if x-values are equal
42 while (k >= 2 && !cw(res[k - 2], res[k - 1], *it)) k--;
43 res[k++] = *it;
44 }
45 int t = k + 1;
46 for (It it = hi - 2; it != lo - 1; --it) {
47 while (k >= t && !cw(res[k - 2], res[k - 1], *it)) k--;
48 res[k++] = *it;
49 }
50 res.resize(k - 1);
51 return res;
52 }
53
55
58
59 int main() {
61 vector<point> v;
62 v.push_back(point(1, 3));
66 v.push_back(point(-1, 3));
67 std::random_shuffle(v.begin(), v.end());
68 vector<point> h = convex_hull(v.begin(), v.end());
69 cout << "hull points:";
70 for (int i = 0; i < (int)h.size(); i++)
71 cout << " (" << h[i].x << "," << h[i].y << ")";
72 cout << "\n";
73 return 0;
74 }
5.3.4 Minimum Enclosing Circle
1 /*
2
3 Given a range of points on the 2D cartesian plane, determine
4 the equation of the circle with smallest possible area which
5 encloses all of the points. Note: in an attempt to avoid the
6 worst case, the circles are randomly shuffled before the
7 algorithm is performed. This is not necessary to obtain the
8 correct answer, and may be removed if the input order must
9 be preserved.
10
11 Time Complexity: O(n) average on the number of points given.
12
13 */
14
16 #include <cmath>
19
21
23
25 #define x first
26 #define y second
27
30
31 struct circle {
32
33 double h, k, r;
34
35 circle(): h(0), k(0), r(0) {}
36 circle(const double & H, const double & K, const double & R):
37 h(H), k(K), r(fabs(R)) {}
38
39 //circumcircle with the diameter equal to the distance from a to b
40 circle(const point & a, const point & b) {
41 h = (a.x + b.x) / 2.0;
42 k = (a.y + b.y) / 2.0;
44 }
45
46 //circumcircle of 3 points - throws exception if abc are collinear/equal
47 circle(const point & a, const point & b, const point & c) {
48 double an = norm(point(b.x - c.x, b.y - c.y));
49 double bn = norm(point(a.x - c.x, a.y - c.y));
50 double cn = norm(point(a.x - b.x, a.y - b.y));
51 double wa = an * (bn + cn - an);
52 double wb = bn * (an + cn - bn);
53 double wc = cn * (an + bn - cn);
54 double w = wa + wb + wc;
55 if (fabs(w) < eps)
56 throw std::runtime_error("No circle from collinear points.");
57 h = (wa * a.x + wb * b.x + wc * c.x) / w;
58 k = (wa * a.y + wb * b.y + wc * c.y) / w;
60 }
61
63 return LE(norm(point(p.x - h, p.y - k)), r * r);
64 }
65
66 };
67
68 template<class It> circle smallest_circle(It lo, It hi) {

69 if (lo == hi) return circle(0, 0, 0);
70 if (lo + 1 == hi) return circle(lo->x, lo->y, 0);
71 std::random_shuffle(lo, hi);
72 circle res(*lo, *(lo + 1));
73 for (It i = lo + 2; i != hi; ++i) {
74 if (res.contains(*i)) continue;
75 res = circle(*lo, *i);
76 for (It j = lo + 1; j != i; ++j) {
77 if (res.contains(*j)) continue;
78 res = circle(*i, *j);
79 for (It k = lo; k != j; ++k)
80 if (!res.contains(*k)) res = circle(*i, *j, *k);
81 }
82 }
83 return res;
84 }
85
87
91
92 int main() {
93 vector<point> v;
98 circle res = smallest_circle(v.begin(), v.end());
99 cout << "center: (" << res.h << ", " << res.k << ")\n";
100 cout << "radius: " << res.r << "\n";
101 return 0;
102 }
5.3.5 Diameter of Point Set
1 /*
2
3 Determines the diametral pair of a range of points. The diamter
4 of a set of points is the largest distance between any two
5 points in the set. A diametral pair is a pair of points in the
6 set whose distance is equal to the sets diameter. The following
7 program uses rotating calipers method to find a solution.
8
9 Time Complexity: O(n log n) on the number of points in the set.
10
11 */
12
17
19 #define x first
20 #define y second
21
22 double sqdist(const point & a, const point & b) {
23 double dx = a.x - b.x, dy = a.y - b.y;
24 return sqrt(dx * dx + dy * dy);
25 }
26
29 }
30
31 bool cw(const point & o, const point & a, const point & b) {
32 return cross(o, a, b) < 0;
33 }
34
35 double area(const point & o, const point & a, const point & b) {
36 return fabs(cross(o, a, b));
37 }
38
39 template<class It> std::vector<point> convex_hull(It lo, It hi) {
40 int k = 0;
41 if (hi - lo <= 1) return std::vector<point>(lo, hi);
42 std::vector<point> res(2 * (int)(hi - lo));
43 std::sort(lo, hi); //compare by x, then by y if x-values are equal
45 while (k >= 2 && !cw(res[k - 2], res[k - 1], *it)) k--;
46 res[k++] = *it;
47 }
48 int t = k + 1;
49 for (It it = hi - 2; it != lo - 1; --it) {
50 while (k >= t && !cw(res[k - 2], res[k - 1], *it)) k--;
51 res[k++] = *it;
52 }
53 res.resize(k - 1);
54 return res;
55 }
56
57 template<class It> std::pair<point, point> diametral_pair(It lo, It hi) {
58 std::vector<point> h = convex_hull(lo, hi);
59 int m = h.size();
60 if (m == 1) return std::make_pair(h[0], h[0]);
61 if (m == 2) return std::make_pair(h[0], h[1]);
62 int k = 1;
63 while (area(h[m - 1], h[0], h[(k + 1) % m]) > area(h[m - 1], h[0], h[k]))
64 k++;
65 double maxdist = 0, d;
66 std::pair<point, point> res;
67 for (int i = 0, j = k; i <= k && j < m; i++) {
68 d = sqdist(h[i], h[j]);
69 if (d > maxdist) {
70 maxdist = d;
71 res = std::make_pair(h[i], h[j]);
72 }
73 while (j < m && area(h[i], h[(i + 1) % m], h[(j + 1) % m]) >
74 area(h[i], h[(i + 1) % m], h[j])) {
75 d = sqdist(h[i], h[(j + 1) % m]);
76 if (d > maxdist) {
77 maxdist = d;
78 res = std::make_pair(h[i], h[(j + 1) % m]);
79 }
80 j++;
81 }
82 }
83 return res;
84 }
85
87
90
91 int main() {
92 vector<point> v;
98 pair<point, point> res = diametral_pair(v.begin(), v.end());
99 cout << "diametral pair: (" << res.first.x << "," << res.first.y << ") ";
100 cout << "(" << res.second.x << "," << res.second.y << ")\n";
101 cout << "diameter: " << sqrt(sqdist(res.first, res.second)) << "\n";
102 return 0;
103 }
5.3.6 Closest Point Pair
1 /*
2
3 Given a range containing distinct points on the Cartesian plane,
4 determine two points which have the closest possible distance.
5 A divide and conquer algorithm is used. Note that the ordering
6 of points in the input range may be changed by the function.
7
8 Time Complexity: O(n log^2 n) where n is the number of points.
9
10 */
11
12 #include <algorithm> /* std::min, std::sort */
14 #include <cmath> /* fabs */
16
18 #define x first
19 #define y second
20
21 double sqdist(const point & a, const point & b) {
22 double dx = a.x - b.x, dy = a.y - b.y;
23 return dx * dx + dy * dy;
24 }
25
26 bool cmp_x(const point & a, const point & b) { return a.x < b.x; }
27 bool cmp_y(const point & a, const point & b) { return a.y < b.y; }
28
30 double rec(It lo, It hi, std::pair<point, point> & res, double mindist) {
31 if (lo == hi) return DBL_MAX;
32 It mid = lo + (hi - lo) / 2;
33 double midx = mid->x;
34 double d1 = rec(lo, mid, res, mindist);
35 mindist = std::min(mindist, d1);
36 double d2 = rec(mid + 1, hi, res, mindist);
37 mindist = std::min(mindist, d2);
38 std::sort(lo, hi, cmp_y);
39 int size = 0;
40 It t[hi - lo];
42 if (fabs(it->x - midx) < mindist)
43 t[size++] = it;
44 for (int i = 0; i < size; i++) {
45 for (int j = i + 1; j < size; j++) {
46 point a = *t[i], b = *t[j];
47 if (b.y - a.y >= mindist) break;
48 double dist = sqdist(a, b);
49 if (mindist > dist) {
50 mindist = dist;
51 res = std::make_pair(a, b);
52 }
53 }
54 }
55 return mindist;
56 }
57
58 template<class It> std::pair<point, point> closest_pair(It lo, It hi) {
59 std::pair<point, point> res;
60 std::sort(lo, hi, cmp_x);
61 rec(lo, hi, res, DBL_MAX);
62 return res;
63 }
64
66
70
71 int main() {
72 vector<point> v;
79 pair<point, point> res = closest_pair(v.begin(), v.end());
80 cout << "closest pair: (" << res.first.x << "," << res.first.y << ") ";
81 cout << "(" << res.second.x << "," << res.second.y << ")\n";
82 cout << "dist: " << sqrt(sqdist(res.first, res.second)) << "\n"; //1.41421
83 return 0;
84 }
5.3.7 Segment Intersection Finding

1 /*
2
3 Given a range of segments on the Cartesian plane, identify one
4 pair of segments which intersect each other. This is done using
5 a sweep line algorithm.
6
7 Time Complexity: O(n log n) where n is the number of segments.
8
9 */
10
11 #include <algorithm> /* std::min(), std::max(), std::sort() */
13 #include <set>
15
17 #define x first
18 #define y second
19
21
25
31 }
32
34
38 }
39
44 }
45
52 if (EQ(c1, 0) && EQ(c2, 0)) {
58 return (res1 == res2) ? 0 : 1;
59 }
60 return -1;
61 }
62 if (EQ(c1, 0)) return -1;
64 if (contain(0, t, 1) && contain(0, u, 1)) return 0;
65 return -1;
66 }
67
68 struct segment {
69 point p, q;
70
71 segment() {}
72 segment(const point & p, const point & q) {
73 if (p < q) {
74 this->p = p;
75 this->q = q;
76 } else {
77 this->p = q;
78 this->q = p;
79 }
80 }
81
82 bool operator < (const segment & rhs) const {
83 if (p.x < rhs.p.x) {
84 double c = cross(p, q, rhs.p);
85 if (c != 0) return c > 0;
86 } else if (p.x > rhs.p.x) {
87 double c = cross(rhs.p, rhs.q, q);
88 if (c != 0) return c < 0;
89 }
90 return p.y < rhs.p.y;
91 }
92 };
93
94 template<class SegIt> struct event {
95 point p;
96 int type;
97 SegIt seg;
98
99 event() {}
100 event(const point & p, const int type, SegIt seg) {
101 this->p = p;
102 this->type = type;
103 this->seg = seg;
104 }
105
106 bool operator < (const event & rhs) const {
107 if (p.x != rhs.p.x) return p.x < rhs.p.x;
108 if (type != rhs.type) return type > rhs.type;
109 return p.y < rhs.p.y;
110 }
111 };
112
113 bool intersect(const segment & s1, const segment & s2) {
114 return seg_intersection(s1.p, s1.q, s2.p, s2.q) >= 0;
115 }
116
117 //returns whether any pair of segments in the range [lo, hi) intersect
118 //if the result is true, one such intersection pair will be stored
119 //into values pointed to by res1 and res2.

121 bool find_intersection(It lo, It hi, segment * res1, segment * res2) {
122 int cnt = 0;
123 event<It> e[2 * (hi - lo)];
125 if (it->p > it->q) std::swap(it->p, it->q);
126 e[cnt++] = event<It>(it->p, 1, it);
127 e[cnt++] = event<It>(it->q, -1, it);
128 }
129 std::sort(e, e + cnt);
130 std::set<segment> s;
131 std::set<segment>::iterator it, next, prev;
132 for (int i = 0; i < cnt; i++) {
133 It seg = e[i].seg;
134 if (e[i].type == 1) {
135 it = s.lower_bound(*seg);
136 if (it != s.end() && intersect(*it, *seg)) {
137 *res1 = *it; *res2 = *seg;
138 return true;
139 }
140 if (it != s.begin() && intersect(*--it, *seg)) {
141 *res1 = *it; *res2 = *seg;
142 return true;
143 }
144 s.insert(*seg);
145 } else {
146 it = s.lower_bound(*seg);
147 next = prev = it;
148 prev = it;
149 if (it != s.begin() && it != --s.end()) {
150 ++next;
151 --prev;
152 if (intersect(*next, *prev)) {
153 *res1 = *next; *res2 = *prev;
154 return true;
155 }
156 }
157 s.erase(it);
158 }
159 }
160 return false;
161 }
162
164
168
169 void print(const segment & s) {
170 cout << "(" << s.p.x << "," << s.p.y << ")<->";
171 cout << "(" << s.q.x << "," << s.q.y << ")\n";
172 }
173
174 int main() {
175 vector<segment> v;
176 v.push_back(segment(point(0, 0), point(2, 2)));
177 v.push_back(segment(point(3, 0), point(0, -1)));
178 v.push_back(segment(point(0, 2), point(2, -2)));

179 v.push_back(segment(point(0, 3), point(9, 0)));
180 segment res1, res2;
181 bool res = find_intersection(v.begin(), v.end(), &res1, &res2);
182 if (res) {
183 print(res1);
184 print(res2);
185 } else {
186 cout << "No intersections.\n";
187 }
188 return 0;
189 }
5.4 Advanced Geometric Computations
5.4.1 Convex Polygon Cut
1 /*
2
3 Given a range of points specifying a polygon on the Cartesian
4 plane, as well as two points specifying an infinite line, "cut"
5 off the right part of the polygon with the line and return the
6 resulting polygon that is the left part.
7
8 Time Complexity: O(n) on the number of points in the poylgon.
9
10 */
11
15
17 #define x first
18 #define y second
19
21
25
28 }
29
30 int orientation(const point & o, const point & a, const point & b) {
31 double c = cross(o, a, b);
32 return LT(c, 0) ? -1 : (GT(c, 0) ? 1 : 0);
33 }
34
5.4. Advanced Geometric Computations 275

42 double det = a1 * b2 - a2 * b1;
43 if (EQ(det, 0))
44 return (EQ(x, 0) && EQ(y, 0)) ? 1 : -1;
46 return 0;
47 }
48
50 std::vector<point> convex_cut(It lo, It hi, const point & p, const point & q) {
51 std::vector<point> res;
52 for (It i = lo, j = hi - 1; i != hi; j = i++) {
53 int d1 = orientation(p, q, *j), d2 = orientation(p, q, *i);
54 if (d1 >= 0) res.push_back(*j);
55 if (d1 * d2 < 0) {
56 point r;
57 line_intersection(p, q, *j, *i, &r);
58 res.push_back(r);
59 }
60 }
61 return res;
62 }
63
65
68
69 int main() {
71 vector<point> v;
77 //cut using the vertical line through (0, 0)
78 vector<point> res = convex_cut(v.begin(), v.end(), point(0, 0), point(0, 1));
79 cout << "left cut:\n";
80 for (int i = 0; i < (int)res.size(); i++)
81 cout << "(" << res[i].x << "," << res[i].y << ")\n";
82 return 0;
83 }
5.4.2 Polygon Union and Intersection
1 /*
2
3 Given two ranges of points respectively denoting the vertices of
4 two polygons, determine the intersection area of those polygons.
5 Using this, we can easily calculate their union with the forumla:
6 union_area(A, B) = area(A) + area(B) - intersection_area(A, B)
7
8 Time Complexity: O(n^2 log n), where n is the total number of vertices.
9
10 */
11
14 #include <set>
17
19
23
25 #define x first
26 #define y second
27
28 inline int sgn(const double & x) {
29 return (0.0 < x) - (x < 0.0);
30 }
31
32 //Line and line segment intersection (see their own sections)
33
41 double det = a1 * b2 - a2 * b1;
42 if (EQ(det, 0))
43 return (EQ(x, 0) && EQ(y, 0)) ? 1 : -1;
45 return 0;
46 }
47
51
53
57 }
58
63 }
64
66 const point & c, const point & d,
67 point * p = 0, point * q = 0) {

72 if (EQ(c1, 0) && EQ(c2, 0)) { //collinear
78 if (res1 == res2) {
79 if (p != 0) *p = res1;
80 return 0; //collinear, meeting at an endpoint
81 }
82 if (p != 0 && q != 0) *p = res1, *q = res2;
83 return 1; //collinear and overlapping
84 } else {
85 return -1; //collinear and disjoint
86 }
87 }
88 if (EQ(c1, 0)) return -1; //parallel and disjoint
90 if (contain(0, t, 1) && contain(0, u, 1)) {
91 if (p != 0) *p = point(a.x + t * ab.x, a.y + t * ab.y);
92 return 0; //non-parallel with one intersection
93 }
94 return -1; //non-parallel with no intersections
95 }
96
97 struct event {
98 double y;
99 int mask_delta;
100
101 event(double y = 0, int mask_delta = 0) {
102 this->y = y;
103 this->mask_delta = mask_delta;
104 }
105
106 bool operator < (const event & e) const {
107 if (y != e.y) return y < e.y;
108 return mask_delta < e.mask_delta;
109 }
110 };
111
113 double intersection_area(It lo1, It hi1, It lo2, It hi2) {
114 It plo[2] = {lo1, lo2}, phi[] = {hi1, hi2};
115 std::set<double> xs;
116 for (It i1 = lo1; i1 != hi1; ++i1) xs.insert(i1->x);
117 for (It i2 = lo2; i2 != hi2; ++i2) xs.insert(i2->x);
118 for (It i1 = lo1, j1 = hi1 - 1; i1 != hi1; j1 = i1++) {
119 for (It i2 = lo2, j2 = hi2 - 1; i2 != hi2; j2 = i2++) {
120 point p;
121 if (seg_intersection(*i1, *j1, *i2, *j2, &p) == 0)
122 xs.insert(p.x);
123 }
124 }
125 std::vector<double> xsa(xs.begin(), xs.end());
126 double res = 0;
127 for (int k = 0; k < (int)xsa.size() - 1; k++) {
128 double x = (xsa[k] + xsa[k + 1]) / 2;
129 point sweep0(x, 0), sweep1(x, 1);

130 std::vector<event> events;
131 for (int poly = 0; poly < 2; poly++) {
132 It lo = plo[poly], hi = phi[poly];
133 double area = 0;
134 for (It i = lo, j = hi - 1; i != hi; j = i++)
135 area += (j->x - i->x) * (j->y + i->y);
136 for (It j = lo, i = hi - 1; j != hi; i = j++) {
137 point p;
138 if (line_intersection(*j, *i, sweep0, sweep1, &p) == 0) {
139 double y = p.y, x0 = i->x, x1 = j->x;
140 if (x0 < x && x1 > x) {
141 events.push_back(event(y, sgn(area) * (1 << poly)));
142 } else if (x0 > x && x1 < x) {
143 events.push_back(event(y, -sgn(area) * (1 << poly)));
144 }
145 }
146 }
147 }
148 std::sort(events.begin(), events.end());
149 double a = 0.0;
150 int mask = 0;
151 for (int j = 0; j < (int)events.size(); j++) {
152 if (mask == 3)
153 a += events[j].y - events[j - 1].y;
154 mask += events[j].mask_delta;
155 }
156 res += a * (xsa[k + 1] - xsa[k]);
157 }
158 return res;
159 }
160
161 template<class It> double polygon_area(It lo, It hi) {
162 if (lo == hi) return 0;
163 double area = 0;
164 if (*lo != *--hi)
165 area += (lo->x - hi->x) * (lo->y + hi->y);
166 for (It i = hi, j = hi - 1; i != lo; --i, --j)
167 area += (i->x - j->x) * (i->y + j->y);
168 return fabs(area / 2.0);
169 }
170
172 double union_area(It lo1, It hi1, It lo2, It hi2) {
173 return polygon_area(lo1, hi1) + polygon_area(lo2, hi2) -
174 intersection_area(lo1, hi1, lo2, hi2);
175 }
176
178
181
182 int main() {
183 vector<point> p1, p2;
184
185 //irregular pentagon with area 1.5 triangle in quadrant 2
186 p1.push_back(point(1, 3));

190 p1.push_back(point(-1, 3));
191 //a big square in quadrant 2
196
197 assert(EQ(1.5, intersection_area(p1.begin(), p1.end(),
198 p2.begin(), p2.end())));
199 assert(EQ(12.5, union_area(p1.begin(), p1.end(),
200 p2.begin(), p2.end())));
201 return 0;
202 }
5.4.3 Delaunay Triangulation (Simple)
1 /*
2
3 Given a range of points P on the Cartesian plane, the Delaunay
4 Triangulation of said points is a set of non-overlapping triangles
5 covering the entire convex hull of P, such that no point in P lies
6 within the circumcircle of any of the resulting triangles. The
7 triangulation maximizes the minimum angle of all the angles of the
8 triangles in the triangulation. In addition, for any point p in the
9 convex hull (not necessarily in P), the nearest point is guaranteed
10 to be a vertex of the enclosing triangle from the triangulation.
11 See: https://en.wikipedia.org/wiki/Delaunay_triangulation
12
13 The triangulation may not exist (e.g. for a set of collinear points)
14 or it may not be unique (multiple possible triangulations may exist).
17 The following program assumes that a triangulation exists, and
18 produces one such valid result using one of the simplest algorithms
19 to solve this problem. It involves encasing the simplex in a circle
20 and rejecting the simplex if another point in the tessellation is
21 within the generalized circle.
22
23 Time Complexity: O(n^4) on the number of input points.
24
25 */
26
31
33
39

41 #define x first
42 #define y second
43
47
48 const bool TOUCH_IS_INTERSECT = false;
49
53 }
54
59 }
60
67 if (EQ(c1, 0) && EQ(c2, 0)) {
73 return (res1 == res2) ? 0 : 1;
74 }
75 return -1;
76 }
77 if (EQ(c1, 0)) return -1;
79 if (contain(0, t, 1) && contain(0, u, 1)) return 0;
80 return -1;
81 }
82
83 struct triangle { point a, b, c; };
84
86 std::vector<triangle> delaunay_triangulation(It lo, It hi) {
87 int n = hi - lo;
88 std::vector<double> x, y, z;
90 x.push_back(it->x);
91 y.push_back(it->y);
92 z.push_back((it->x) * (it->x) + (it->y) * (it->y));
93 }
94 std::vector<triangle> res;
95 for (int i = 0; i < n - 2; i++) {
96 for (int j = i + 1; j < n; j++) {
97 for (int k = i + 1; k < n; k++) {
98 if (j == k) continue;
99 double nx = (y[j] - y[i]) * (z[k] - z[i]) - (y[k] - y[i]) * (z[j] - z[i]);

100 double ny = (x[k] - x[i]) * (z[j] - z[i]) - (x[j] - x[i]) * (z[k] - z[i]);
101 double nz = (x[j] - x[i]) * (y[k] - y[i]) - (x[k] - x[i]) * (y[j] - y[i]);
102 if (GE(nz, 0)) continue;
103 bool done = false;
104 for (int m = 0; m < n; m++)
105 if (x[m] - x[i]) * nx + (y[m] - y[i]) * ny + (z[m] - z[i]) * nz > 0) {
106 done = true;
107 break;
108 }
109 if (!done) { //handle 4 points on a circle
110 point s1[] = { *(lo + i), *(lo + j), *(lo + k), *(lo + i) };
111 for (int t = 0; t < (int)res.size(); t++) {
112 point s2[] = { res[t].a, res[t].b, res[t].c, res[t].a };
113 for (int u = 0; u < 3; u++)
114 for (int v = 0; v < 3; v++)
115 if (seg_intersection(s1[u], s1[u + 1], s2[v], s2[v + 1]) == 0)
116 goto skip;
117 }
118 res.push_back((triangle){*(lo + i), *(lo + j), *(lo + k)});
119 }
120 skip:;
121 }
122 }
123 }
124 return res;
125 }
126
128
131
132 int main() {
133 vector<point> v;
139 vector<triangle> dt = delaunay_triangulation(v.begin(), v.end());
140 for (int i = 0; i < (int)dt.size(); i++) {
141 cout << "Triangle: ";
142 cout << "(" << dt[i].a.x << "," << dt[i].a.y << ") ";
143 cout << "(" << dt[i].b.x << "," << dt[i].b.y << ") ";
144 cout << "(" << dt[i].c.x << "," << dt[i].c.y << ")\n";
145 }
146 return 0;
147 }
5.4.4 Delaunay Triangulation (Fast)
1 /*
2
3 Given a range of points P on the Cartesian plane, the Delaunay
4 Triangulation of said points is a set of non-overlapping triangles
5 covering the entire convex hull of P, such that no point in P lies
6 within the circumcircle of any of the resulting triangles. The

7 triangulation maximizes the minimum angle of all the angles of the
8 triangles in the triangulation. In addition, for any point p in the
9 convex hull (not necessarily in P), the nearest point is guaranteed
10 to be a vertex of the enclosing triangle from the triangulation.
11 See: https://en.wikipedia.org/wiki/Delaunay_triangulation
12
15 The following program assumes that a triangulation exists, and
16 produces one such valid result. The following is a C++ adaptation of
17 a FORTRAN90 program, which applies a divide and conquer algorithm
18 with complex linear-time merging. The original program can be found
19 via the following link. It contains more thorough documentation,
20 comments, and debugging messages associated with the current asserts().
21 http://people.sc.fsu.edu/~burkardt/f_src/table_delaunay/table_delaunay.html
22
23 Time Complexity: O(n log n) on the number of input points.
24
25 */
26
32
33 int wrap(int ival, int ilo, int ihi) {
34 int jlo = std::min(ilo, ihi), jhi = std::max(ilo, ihi);
35 int wide = jhi + 1 - jlo, res = jlo;
36 if (wide != 1) {
37 assert(wide != 0);
38 int tmp = (ival - jlo) % wide;
39 if (tmp < 0) res += abs(wide);
40 res += tmp;
41 }
42 return res;
43 }
44
45 double epsilon() {
46 double r = 1;
47 while (1 < (double)(r + 1)) r /= 2;
48 return 2 * r;
49 }
50
51 void permute(int n, double a[][2], int p[]) {
52 for (int istart = 1; istart <= n; istart++) {
53 if (p[istart - 1] < 0) continue;
54 if (p[istart - 1] == istart) {
55 p[istart - 1] = -p[istart - 1];
56 continue;
57 }
58 double tmp0 = a[istart - 1][0];
59 double tmp1 = a[istart - 1][1];
60 int iget = istart;
61 for (;;) {
62 int iput = iget;
63 iget = p[iget - 1];
64 p[iput - 1] = -p[iput - 1];
65 assert(!(iget < 1 || n < iget));

66 if (iget == istart) {
67 a[iput - 1][0] = tmp0;
68 a[iput - 1][1] = tmp1;
69 break;
70 }
71 a[iput - 1][0] = a[iget - 1][0];
72 a[iput - 1][1] = a[iget - 1][1];
73 }
74 }
75 for (int i = 0; i < n; i++) p[i] = -p[i];
76 return;
77 }
78
79 int * sort_heap(int n, double a[][2]) {
80 double aval[2];
81 int i, ir, j, l, idxt;
82 int *idx;
83 if (n < 1) return NULL;
84 if (n == 1) {
85 idx = new int[1];
86 idx[0] = 1;
87 return idx;
88 }
89 idx = new int[n];
90 for (int i = 0; i < n; i++) idx[i] = i + 1;
91 l = n / 2 + 1;
92 ir = n;
93 for (;;) {
94 if (1 < l) {
95 l--;
96 idxt = idx[l - 1];
97 aval[0] = a[idxt - 1][0];
98 aval[1] = a[idxt - 1][1];
99 } else {
100 idxt = idx[ir - 1];
101 aval[0] = a[idxt - 1][0];
102 aval[1] = a[idxt - 1][1];
103 idx[ir - 1] = idx[0];
104 if (--ir == 1) {
105 idx[0] = idxt;
106 break;
107 }
108 }
109 i = l;
110 j = 2 * l;
111 while (j <= ir) {
112 if (j < ir && (a[idx[j - 1] - 1][0] < a[idx[j] - 1][0] ||
113 (a[idx[j - 1] - 1][0] == a[idx[j] - 1][0] &&
114 a[idx[j - 1] - 1][1] < a[idx[j] - 1][1]))) {
115 j++;
116 }
117 if ( aval[0] < a[idx[j - 1] - 1][0] ||
118 (aval[0] == a[idx[j - 1] - 1][0] &&
119 aval[1] < a[idx[j - 1] - 1][1])) {
120 idx[i - 1] = idx[j - 1];
121 i = j;
122 j *= 2;
123 } else {
124 j = ir + 1;
125 }
126 }
127 idx[i - 1] = idxt;
128 }
129 return idx;
130 }
131
132 int lrline(double xu, double yu, double xv1, double yv1,
133 double xv2, double yv2, double dv) {
134 double tol = 1e-7;
135 double dx = xv2 - xv1, dy = yv2 - yv1;
136 double dxu = xu - xv1, dyu = yu - yv1;
137 double t = dy * dxu - dx * dyu + dv * sqrt(dx * dx + dy * dy);
138 double tolabs = tol * std::max(std::max(fabs(dx), fabs(dy)),
139 std::max(fabs(dxu), std::max(fabs(dyu), fabs(dv))));
140 if (tolabs < t) return 1;
141 if (-tolabs <= t) return 0;
142 return -1;
143 }
144
145 void vbedg(double x, double y, int point_num, double point_xy[][2],
146 int tri_num, int tri_nodes[][3], int tri_neigh[][3],
147 int *ltri, int *ledg, int *rtri, int *redg) {
148 int a, b;
149 double ax, ay, bx, by;
150 bool done;
151 int e, l, t;
152 if (*ltri == 0) {
153 done = false;
154 *ltri = *rtri;
155 *ledg = *redg;
156 } else {
157 done = true;
158 }
159 for (;;) {
160 l = -tri_neigh[(*rtri) - 1][(*redg) - 1];
161 t = l / 3;
162 e = l % 3 + 1;
163 a = tri_nodes[t - 1][e - 1];
164 if (e <= 2) {
165 b = tri_nodes[t - 1][e];
166 } else {
167 b = tri_nodes[t - 1][0];
168 }
169 ax = point_xy[a - 1][0];
170 ay = point_xy[a - 1][1];
171 bx = point_xy[b - 1][0];
172 by = point_xy[b - 1][1];
173 if (lrline(x, y, ax, ay, bx, by, 0.0) <= 0) break;
174 *rtri = t;
175 *redg = e;
176 }
177 if (done) return;
178 t = *ltri;
179 e = *ledg;
180 for (;;) {
181 b = tri_nodes[t - 1][e - 1];
182 e = wrap(e - 1, 1, 3);
183 while (0 < tri_neigh[t - 1][e - 1]) {

184 t = tri_neigh[t - 1][e - 1];
185 if (tri_nodes[t - 1][0] == b) {
186 e = 3;
187 } else if (tri_nodes[t - 1][1] == b) {
188 e = 1;
189 } else {
190 e = 2;
191 }
192 }
193 a = tri_nodes[t - 1][e - 1];
194 ax = point_xy[a - 1][0];
195 ay = point_xy[a - 1][1];
196 bx = point_xy[b - 1][0];
197 by = point_xy[b - 1][1];
198 if (lrline(x, y, ax, ay, bx, by, 0.0) <= 0) break;
199 }
200 *ltri = t;
201 *ledg = e;
202 return;
203 }
204
205 int diaedg(double x0, double y0, double x1, double y1,
206 double x2, double y2, double x3, double y3) {
207 double ca, cb, s, tol, tola, tolb;
208 int value;
209 tol = 100.0 * epsilon();
210 double dx10 = x1 - x0, dy10 = y1 - y0;
211 double dx12 = x1 - x2, dy12 = y1 - y2;
212 double dx30 = x3 - x0, dy30 = y3 - y0;
213 double dx32 = x3 - x2, dy32 = y3 - y2;
214 tola = tol * std::max(std::max(fabs(dx10), fabs(dy10)),
215 std::max(fabs(dx30), fabs(dy30)));
216 tolb = tol * std::max(std::max(fabs(dx12), fabs(dy12)),
217 std::max(fabs(dx32), fabs(dy32)));
218 ca = dx10 * dx30 + dy10 * dy30;
219 cb = dx12 * dx32 + dy12 * dy32;
220 if (tola < ca && tolb < cb) {
221 value = -1;
222 } else if (ca < -tola && cb < -tolb) {
223 value = 1;
224 } else {
225 tola = std::max(tola, tolb);
226 s = (dx10 * dy30 - dx30 * dy10) * cb + (dx32 * dy12 - dx12 * dy32) * ca;
227 if (tola < s) {
228 value = -1;
229 } else if (s < -tola) {
230 value = 1;
231 } else {
232 value = 0;
233 }
234 }
235 return value;
236 }
237
238 int swapec(int i, int *top, int *btri, int *bedg,
239 int point_num, double point_xy[][2],
240 int tri_num, int tri_nodes[][3], int tri_neigh[][3], int stack[]) {
241 int a, b, c, e, ee, em1, ep1, f, fm1, fp1, l, r, s, swap, t, tt, u;
242 double x = point_xy[i - 1][0];

243 double y = point_xy[i - 1][1];
244 for (;;) {
245 if (*top <= 0) break;
246 t = stack[*top - 1];
247 *top = *top - 1;
248 if (tri_nodes[t - 1][0] == i) {
249 e = 2;
250 b = tri_nodes[t - 1][2];
251 } else if (tri_nodes[t - 1][1] == i) {
252 e = 3;
253 b = tri_nodes[t - 1][0];
254 } else {
255 e = 1;
256 b = tri_nodes[t - 1][1];
257 }
258 a = tri_nodes[t - 1][e - 1];
259 u = tri_neigh[t - 1][e - 1];
260 if (tri_neigh[u - 1][0] == t) {
261 f = 1;
262 c = tri_nodes[u - 1][2];
263 } else if (tri_neigh[u - 1][1] == t) {
264 f = 2;
265 c = tri_nodes[u - 1][0];
266 } else {
267 f = 3;
268 c = tri_nodes[u - 1][1];
269 }
270 swap = diaedg(x, y, point_xy[a - 1][0], point_xy[a - 1][1],
271 point_xy[c - 1][0], point_xy[c - 1][1],
272 point_xy[b - 1][0], point_xy[b - 1][1]);
273 if (swap == 1) {
274 em1 = wrap(e - 1, 1, 3);
275 ep1 = wrap(e + 1, 1, 3);
276 fm1 = wrap(f - 1, 1, 3);
277 fp1 = wrap(f + 1, 1, 3);
278 tri_nodes[t - 1][ep1 - 1] = c;
279 tri_nodes[u - 1][fp1 - 1] = i;
280 r = tri_neigh[t - 1][ep1 - 1];
281 s = tri_neigh[u - 1][fp1 - 1];
282 tri_neigh[t - 1][ep1 - 1] = u;
283 tri_neigh[u - 1][fp1 - 1] = t;
284 tri_neigh[t - 1][e - 1] = s;
285 tri_neigh[u - 1][f - 1] = r;
286 if (0 < tri_neigh[u - 1][fm1 - 1]) {
287 *top = *top + 1;
288 stack[*top - 1] = u;
289 }
290 if (0 < s) {
291 if (tri_neigh[s - 1][0] == u) {
292 tri_neigh[s - 1][0] = t;
293 } else if (tri_neigh[s - 1][1] == u) {
294 tri_neigh[s - 1][1] = t;
295 } else {
296 tri_neigh[s - 1][2] = t;
297 }
298 *top = *top + 1;
299 if (point_num < *top) return 8;
300 stack[*top - 1] = t;
301 } else {
302 if (u == *btri && fp1 == *bedg) {
303 *btri = t;
304 *bedg = e;
305 }
306 l = - (3 * t + e - 1);
307 tt = t;
308 ee = em1;
309 while (0 < tri_neigh[tt - 1][ee - 1]) {
310 tt = tri_neigh[tt - 1][ee - 1];
311 if (tri_nodes[tt - 1][0] == a) {
312 ee = 3;
313 } else if (tri_nodes[tt - 1][1] == a) {
314 ee = 1;
315 } else {
316 ee = 2;
317 }
318 }
319 tri_neigh[tt - 1][ee - 1] = l;
320 }
321 if (0 < r) {
322 if (tri_neigh[r - 1][0] == t) {
323 tri_neigh[r - 1][0] = u;
324 } else if (tri_neigh[r - 1][1] == t) {
325 tri_neigh[r - 1][1] = u;
326 } else {
327 tri_neigh[r - 1][2] = u;
328 }
329 } else {
330 if (t == *btri && ep1 == *bedg) {
331 *btri = u;
332 *bedg = f;
333 }
334 l = -(3 * u + f - 1);
335 tt = u;
336 ee = fm1;
337 while (0 < tri_neigh[tt - 1][ee - 1]) {
338 tt = tri_neigh[tt - 1][ee - 1];
339 if (tri_nodes[tt - 1][0] == b) {
340 ee = 3;
341 } else if (tri_nodes[tt - 1][1] == b) {
342 ee = 1;
343 } else {
344 ee = 2;
345 }
346 }
347 tri_neigh[tt - 1][ee - 1] = l;
348 }
349 }
350 }
351 return 0;
352 }
353
354 void perm_inv(int n, int p[]) {
355 int i, i0, i1, i2;
356 assert(n > 0);
357 for (i = 1; i <= n; i++) {
358 i1 = p[i - 1];
359 while (i < i1) {
360 i2 = p[i1 - 1];

361 p[i1 - 1] = -i2;
362 i1 = i2;
363 }
364 p[i - 1] = -p[i - 1];
365 }
366 for (i = 1; i <= n; i++) {
367 i1 = -p[i - 1];
368 if (0 <= i1) {
369 i0 = i;
370 for (;;) {
371 i2 = p[i1 - 1];
372 p[i1 - 1] = i0;
373 if (i2 < 0) break;
374 i0 = i1;
375 i1 = i2;
376 }
377 }
378 }
379 return;
380 }
381
382 int dtris2(int point_num, double point_xy[][2],
383 int tri_nodes[][3], int tri_neigh[][3]) {
384 double cmax;
385 int e, error;
386 int i, j, k, l, m, m1, m2, n;
387 int ledg, lr, ltri, redg, rtri, t, top;
388 double tol;
389 int *stack = new int[point_num];
390 tol = 100.0 * epsilon();
391 int *idx = sort_heap(point_num, point_xy);
392 permute(point_num, point_xy, idx);
393 m1 = 0;
394 for (i = 1; i < point_num; i++) {
395 m = m1;
396 m1 = i;
397 k = -1;
398 for (j = 0; j <= 1; j++) {
399 cmax = std::max(fabs(point_xy[m][j]), fabs(point_xy[m1][j]));
400 if (tol * (cmax + 1.0) < fabs(point_xy[m][j] - point_xy[m1][j])) {
401 k = j;
402 break;
403 }
404 }
405 assert(k != -1);
406 }
407 m1 = 1;
408 m2 = 2;
409 j = 3;
410 for (;;) {
411 assert(point_num >= j);
412 m = j;
413 lr = lrline(point_xy[m - 1][0], point_xy[m - 1][1],
414 point_xy[m1 - 1][0], point_xy[m1 - 1][1],
415 point_xy[m2 - 1][0], point_xy[m2 - 1][1], 0.0);
416 if (lr != 0) break;
417 j++;
418 }
419 int tri_num = j - 2;

420 if (lr == -1) {
421 tri_nodes[0][0] = m1;
422 tri_nodes[0][1] = m2;
423 tri_nodes[0][2] = m;
424 tri_neigh[0][2] = -3;
425 for (i = 2; i <= tri_num; i++) {
426 m1 = m2;
427 m2 = i + 1;
428 tri_nodes[i - 1][0] = m1;
429 tri_nodes[i - 1][1] = m2;
430 tri_nodes[i - 1][2] = m;
431 tri_neigh[i - 1][0] = -3 * i;
432 tri_neigh[i - 1][1] = i;
433 tri_neigh[i - 1][2] = i - 1;
434 }
435 tri_neigh[tri_num - 1][0] = -3 * tri_num - 1;
436 tri_neigh[tri_num - 1][1] = -5;
437 ledg = 2;
438 ltri = tri_num;
439 } else {
440 tri_nodes[0][0] = m2;
441 tri_nodes[0][1] = m1;
442 tri_nodes[0][2] = m;
443 tri_neigh[0][0] = -4;
444 for (i = 2; i <= tri_num; i++) {
445 m1 = m2;
446 m2 = i+1;
447 tri_nodes[i - 1][0] = m2;
448 tri_nodes[i - 1][1] = m1;
449 tri_nodes[i - 1][2] = m;
450 tri_neigh[i - 2][2] = i;
451 tri_neigh[i - 1][0] = -3 * i - 3;
452 tri_neigh[i - 1][1] = i - 1;
453 }
454 tri_neigh[tri_num - 1][2] = -3 * (tri_num);
455 tri_neigh[0][1] = -3 * (tri_num) - 2;
456 ledg = 2;
457 ltri = 1;
458 }
459 top = 0;
460 for (i = j + 1; i <= point_num; i++) {
461 m = i;
462 m1 = tri_nodes[ltri - 1][ledg - 1];
463 if (ledg <= 2) {
464 m2 = tri_nodes[ltri - 1][ledg];
465 } else {
466 m2 = tri_nodes[ltri - 1][0];
467 }
468 lr = lrline(point_xy[m - 1][0], point_xy[m - 1][1],
469 point_xy[m1 - 1][0], point_xy[m1 - 1][1],
470 point_xy[m2 - 1][0], point_xy[m2 - 1][1], 0.0);
471 if (0 < lr) {
472 rtri = ltri;
473 redg = ledg;
474 ltri = 0;
475 } else {
476 l = -tri_neigh[ltri - 1][ledg - 1];
477 rtri = l / 3;
478 redg = (l % 3) + 1;
479 }
480 vbedg(point_xy[m - 1][0], point_xy[m - 1][1],
481 point_num, point_xy, tri_num, tri_nodes, tri_neigh,
482 &ltri, &ledg, &rtri, &redg);
483 n = tri_num + 1;
484 l = -tri_neigh[ltri - 1][ledg - 1];
485 for (;;) {
486 t = l / 3;
487 e = (l % 3) + 1;
488 l = -tri_neigh[t - 1][e - 1];
489 m2 = tri_nodes[t - 1][e - 1];
490 if (e <= 2) {
491 m1 = tri_nodes[t - 1][e];
492 } else {
493 m1 = tri_nodes[t - 1][0];
494 }
495 tri_num++;
496 tri_neigh[t - 1][e - 1] = tri_num;
497 tri_nodes[tri_num - 1][0] = m1;
498 tri_nodes[tri_num - 1][1] = m2;
499 tri_nodes[tri_num - 1][2] = m;
500 tri_neigh[tri_num - 1][0] = t;
501 tri_neigh[tri_num - 1][1] = tri_num - 1;
502 tri_neigh[tri_num - 1][2] = tri_num + 1;
503 top++;
504 assert(point_num >= top);
505 stack[top - 1] = tri_num;
506 if (t == rtri && e == redg) break;
507 }
508 tri_neigh[ltri - 1][ledg - 1] = -3 * n - 1;
509 tri_neigh[n - 1][1] = -3 * tri_num - 2;
510 tri_neigh[tri_num - 1][2] = -l;
511 ltri = n;
512 ledg = 2;
513 error = swapec(m, &top, &ltri, &ledg, point_num, point_xy,
514 tri_num, tri_nodes, tri_neigh, stack);
515 assert(error == 0);
516 }
517 for (i = 0; i < 3; i++)
518 for (j = 0; j < tri_num; j++)
519 tri_nodes[j][i] = idx[tri_nodes[j][i] - 1];
520 perm_inv(point_num, idx);
521 permute(point_num, point_xy, idx);
522 delete[] idx;
523 delete[] stack;
524 return tri_num;
525 }
526
527 /*** C++ Wrapper ***/
528
530 #define x first
531 #define y second
532
533 struct triangle { point a, b, c; };
534
536 std::vector<triangle> delaunay_triangulation(It lo, It hi) {
537 int n = hi - lo;

538 double points[n][2];
539 int tri_nodes[3 * n][3], tri_neigh[3 * n][3];
540 int curr = 0;
541 for (It it = lo; it != hi; ++curr, ++it) {
542 points[curr][0] = it->x;
543 points[curr][1] = it->y;
544 }
545 int m = dtris2(n, points, tri_nodes, tri_neigh);
546 std::vector<triangle> res;
547 for (int i = 0; i < m; i++)
548 res.push_back((triangle){*(lo + (tri_nodes[i][0] - 1)),
549 *(lo + (tri_nodes[i][1] - 1)),
550 *(lo + (tri_nodes[i][2] - 1))});
551 return res;
552 }
553
555
558
559 int main() {
560 vector<point> v;
566 vector<triangle> dt = delaunay_triangulation(v.begin(), v.end());
567 for (int i = 0; i < (int)dt.size(); i++) {
568 cout << "Triangle: ";
569 cout << "(" << dt[i].a.x << "," << dt[i].a.y << ") ";
570 cout << "(" << dt[i].b.x << "," << dt[i].b.y << ") ";
571 cout << "(" << dt[i].c.x << "," << dt[i].c.y << ")\n";
572 }
573 return 0;
574 }
Chapter 6
Strings
6.1 Strings Toolbox
1 /*
2
3 Useful or trivial string operations. These functions are not particularly
4 algorithmic. They are typically naive implementations using C++ features.
5 They depend on many features of the C++ <string> library, which tend to
6 have an unspecified complexity. They may not be optimally efficient.
7
8 */
9
14
15 //integer to string conversion and vice versa using C++ features
16
17 //note that a similar std::to_string is introduced in C++0x
19 std::string to_string(const Int & i) {
21 oss << i;
23 }
24
25 //like atoi, except during special cases like overflows
26 int to_int(const std::string & s) {
27 std::istringstream iss(s);
28 int res;
29 if (!(iss >> res)) /* complain */;
30 return res;
31 }
32
33 /*
34
35 itoa implementation (fast)
36 documentation: http://www.cplusplus.com/reference/cstdlib/itoa/
37 taken from: http://www.jb.man.ac.uk/~slowe/cpp/itoa.html
292
6.1. Strings Toolbox 293
38
39 */
40
41 char* itoa(int value, char * str, int base = 10) {
42 if (base < 2 || base > 36) {
43 *str = \0;
44 return str;
45 }
46 char *ptr = str, *ptr1 = str, tmp_c;
47 int tmp_v;
48 do {
49 tmp_v = value;
50 value /= base;
51 *ptr++ = "zyxwvutsrqponmlkjihgfedcba9876543210123456789"
52 "abcdefghijklmnopqrstuvwxyz"[35 + (tmp_v - value * base)];
53 } while (value);
54 if (tmp_v < 0) *ptr++ = -;
55 for (*ptr-- = \0; ptr1 < ptr; *ptr1++ = tmp_c) {
56 tmp_c = *ptr;
57 *ptr-- = *ptr1;
58 }
59 return str;
60 }
61
62 /*
63
64 Trimming functions (in place). Given a string and optionally a series
65 of characters to be considered for trimming, trims the strings ends
66 (left, right, or both) and returns the string. Note that the ORIGINAL
67 string is trimmed as its passed by reference, despite the original
68 reference being returned for convenience.
69
70 */
71
72 std::string& ltrim(std::string & s, const std::string & delim = " \n\t\v\f\r") {
73 unsigned int pos = s.find_first_not_of(delim);
74 if (pos != std::string::npos) s.erase(0, pos);
75 return s;
76 }
77
78 std::string& rtrim(std::string & s, const std::string & delim = " \n\t\v\f\r") {
79 unsigned int pos = s.find_last_not_of(delim);
80 if (pos != std::string::npos) s.erase(pos);
81 return s;
82 }
83
84 std::string& trim(std::string & s, const std::string & delim = " \n\t\v\f\r") {
85 return ltrim(rtrim(s));
86 }
87
88 /*
89
90 Returns a copy of the string s with all occurrences of the given
91 string search replaced with the given string replace.
92
93 Time Complexity: Unspecified, but proportional to the number of times
94 the search string occurs and the complexity of std::string::replace,
95 which is unspecified.
96
294 Chapter 6. Strings
97 */
98
99 std::string replace(std::string s,
100 const std::string & search,
101 const std::string & replace) {
102 if (search.empty()) return s;
103 unsigned int pos = 0;
104 while ((pos = s.find(search, pos)) != std::string::npos) {
105 s.replace(pos, search.length(), replace);
106 pos += replace.length();
107 }
108 return s;
109 }
110
111 /*
112
113 Tokenizes the string s based on single character delimiters.
114
115 Version 1: Simpler. Only one delimiter character allowed, and this will
116 not skip empty tokens.
117 e.g. split("a::b", ":") yields {"a", "b"}, not {"a", "", "b"}.
118
119 Version 2: All of the characters in the delim parameter that also exists
120 in s will be removed from s, and the token(s) of s that are left over will
121 be added sequentially to a vector and returned. Empty tokens are skipped.
122 e.g. split("a::b", ":") yields {"a", "b"}, not {"a", "", "b"}.
123
124 Time Complexity: O(s.length() * delim.length())
125
126 */
127
128 std::vector<std::string> split(const std::string & s, char delim) {
129 std::vector<std::string> res;
130 std::stringstream ss(s);
131 std::string curr;
132 while (std::getline(ss, curr, delim))
133 res.push_back(curr);
134 return res;
135 }
136
137 std::vector<std::string> split(const std::string & s,
138 const std::string & delim = " \n\t\v\f\r") {
140 std::string curr;
142 if (delim.find(s[i]) == std::string::npos) {
143 curr += s[i];
144 } else if (!curr.empty()) {
145 res.push_back(curr);
146 curr = "";
147 }
148 }
149 if (!curr.empty()) res.push_back(curr);
150 return res;
151 }
152
153 /*
154
155 Like the explode() function in PHP, the string s is tokenized based
6.1. Strings Toolbox 295
156 on delim, which is considered as a whole boundary string, not just a

157 sequence of possible boundary characters like the split() function above.
158 This will not skip empty tokens.
159 e.g. explode("a::b", ":") yields {"a", "", "b"}, not {"a", "b"}.
160
161 Time Complexity: O(s.length() * delim.length())
162
163 */
164
165 std::vector<std::string> explode(const std::string & s,
166 const std::string & delim) {
168 unsigned int last = 0, next = 0;
169 while ((next = s.find(delim, last)) != std::string::npos) {
170 res.push_back(s.substr(last, next - last));
171 last = next + delim.size();
172 }
173 res.push_back(s.substr(last));
174 return res;
175 }
176
178
183
184 void print(const vector<string> & v) {
185 cout << "[";
187 cout << (i ? "\", \"" : "\"") << v[i];
188 cout << "\"]\n";
189 }
190
191 int main() {
192 assert(to_string(123) + "4" == "1234");
193 assert(to_int("1234") == 1234);
194 char buffer[50];
195 assert(string(itoa(1750, buffer, 10)) == "1750");
196 assert(string(itoa(1750, buffer, 16)) == "6d6");
197 assert(string(itoa(1750, buffer, 2)) == "11011010110");
198
199 string s(" abc \n");
200 string t = s;
201 assert(ltrim(s) == "abc \n");
202 assert(rtrim(s) == trim(t));
203 assert(replace("abcdabba", "ab", "00") == "00cd00ba");
204
205 vector<string> tokens;
206
207 tokens = split("a\nb\ncde\nf", \n);
208 cout << "split v1: ";
209 print(tokens); //["a", "b", "cde", "f"]
210
211 tokens = split("a::b,cde:,f", ":,");
212 cout << "split v2: ";
213 print(tokens); //["a", "b", "cde", "f"]
214
215 tokens = explode("a..b.cde....f", "..");

216 cout << "explode: ";
217 print(tokens); //["a", ".b.cde", "", ".f"]
218 return 0;
219 }
6.2 Expression Parsing
6.2.1 Recursive Descent
1 /*
2
3 Evaluate a mathematical expression in accordance to the order
4 of operations (parentheses, exponents, multiplication, division,
5 addition, subtraction). Does not handle unary operators like -.
6
7 */
8
10
11 #include <cctype>
12 #include <cmath>
16
17 class parser {
18 int pos;
19 double tokval;
20 std::string s;
21
22 bool is_dig_or_dot(char c) {
23 return isdigit(c) || c == .;
24 }
25
26 double to_double(const std::string & s) {
27 std::stringstream ss(s);
28 double res;
29 ss >> res;
30 return res;
31 }
32
33 public:
34 char token;
35
36 parser(const std::string & s) {
37 this->s = s;
38 pos = 0;
39 }
40
41 int next() {
42 for (;;) {
43 if (pos == (int)s.size())
44 return token = -1;
45 char c = s[pos++];
46 if (std::string("+-*/^()\n").find(c) != std::string::npos)
6.2. Expression Parsing 297
47 return token = c;
48 if (isspace(c)) continue;
49 if (isdigit(c) || c == .) {
50 std::string operand(1, c);
51 while (pos < (int)s.size() && is_dig_or_dot(s[pos]))
52 operand += (c = s[pos++]);
53 tokval = to_double(operand);
54 return token = n;
55 }
56 throw std::runtime_error(std::string("Bad character: ") + c);
57 }
58 }
59
60 void skip(int ch) {
61 if (token != ch)
62 throw std::runtime_error(std::string("Bad character: ") + token + std::string(", expected: ") +
(char)ch);
63 next();
64 }
65
66 double number() {
67 if (token == n) {
68 double v = tokval;
69 skip(n);
70 return v;
71 }
72 skip(();
73 double v = expression();
74 skip());
75 return v;
76 }
77
78 // factor ::= number | number ^ factor
79 double factor() {
80 double v = number();
81 if (token == ^) {
82 skip(^);
83 v = pow(v, factor());
84 }
85 return v;
86 }
87
88 // term ::= factor | term * factor | term / factor
89 double term() {
90 double v = factor();
91 for (;;) {
92 if (token == *) {
93 skip(*);
94 v *= factor();
95 } else if (token == /) {
96 skip(/);
97 v /= factor();
98 } else {
99 return v;
100 }
101 }
102 }
103
104 // expression ::= term | expression + term | expression - term
105 double expression() {

106 double v = term();
107 for (;;) {
108 if (token == +) {
109 skip(+);
110 v += term();
111 } else if (token == -) {
112 skip(-);
113 v -= term();
114 } else {
115 return v;
116 }
117 }
118 }
119 };
120
123
124 int main() {
125 parser p("1+2*3*4+3*(2+2)-100\n");
126 p.next();
127 while (p.token != -1) {
128 if (p.token == \n) {
129 p.skip(\n);
130 continue;
131 }
132 cout << p.expression() << "\n";
133 }
134 return 0;
135 }
6.2.2 Recursive Descent (Simple)
1 /*
2
3 Evaluate a mathematica expression in accordance to the order
5 addition, subtraction). This handles unary operators like -.
6
7 */
8
9 #include <string>
10
11 template<class It> int eval(It & it, int prec) {
12 if (prec == 0) {
13 int sign = 1, ret = 0;
14 for (; *it == -; it++) sign *= -1;
15 if (*it == () {
16 ret = eval(++it, 2);
17 it++;
18 } else while (*it >= 0 && *it <= 9) {
19 ret = 10 * ret + (*(it++) - 0);
20 }
21 return sign * ret;
22 }
23 int num = eval(it, prec - 1);
6.2. Expression Parsing 299
24 while (!((prec == 2 && *it != + && *it != -) ||

25 (prec == 1 && *it != * && *it != /))) {
26 switch (*(it++)) {
27 case +: num += eval(it, prec - 1); break;
28 case -: num -= eval(it, prec - 1); break;
29 case *: num *= eval(it, prec - 1); break;
30 case /: num /= eval(it, prec - 1); break;
31 }
32 }
33 return num;
34 }
35
36 /*** Wrapper Function ***/
37
38 int eval(const std::string & s) {
39 std::string::iterator it = std::string(s).begin();
40 return eval(it, 2);
41 }
42
44
47
48 int main() {
49 cout << eval("1+2*3*4+3*(2+2)-100") << "\n";
50 return 0;
51 }
6.2.3 Shunting Yard Algorithm
1 /*
2
3 Evaluate a mathematica expression in accordance to the order
5 addition, subtraction). This also handles unary operators like -.
6 We use strings for operators so we can even define things like "sqrt"
7 and "mod" as unary operators by changing prec() and split_expr()
8 accordingly.
9
10 Time Complexity: O(n) on the total number of operators and operands.
11
12 */
13
14 #include <cstdlib> /* strtol() */
15 #include <stack>
19
20 // Classify the precedences of operators here.
21 inline int prec(const std::string & op, bool unary) {
22 if (unary) {
23 if (op == "+" || op == "-") return 3;
24 return 0; // not a unary operator
25 }
26 if (op == "*" || op == "/") return 2;
27 if (op == "+" || op == "-") return 1;

28 return 0; // not a binary operator
29 }
30
31 inline int calc1(const std::string & op, int val) {
32 if (op == "+") return +val;
33 if (op == "-") return -val;
34 throw std::runtime_error("Invalid unary operator: " + op);
35 }
36
37 inline int calc2(const std::string & op, int L, int R) {
38 if (op == "+") return L + R;
39 if (op == "-") return L - R;
40 if (op == "*") return L * R;
41 if (op == "/") return L / R;
42 throw std::runtime_error("Invalid binary operator: " + op);
43 }
44
45 inline bool is_operand(const std::string & s) {
46 return s != "(" && s != ")" && !prec(s, 0) && !prec(s, 1);
47 }
48
49 int eval(std::vector<std::string> E) { // E stores the tokens
50 E.insert(E.begin(), "(");
51 E.push_back(")");
52 std::stack<std::pair<std::string, bool> > ops;
53 std::stack<int> vals;
54 for (int i = 0; i < (int)E.size(); i++) {
55 if (is_operand(E[i])) {
56 vals.push(strtol(E[i].c_str(), 0, 10)); // convert to int
57 continue;
58 }
59 if (E[i] == "(") {
60 ops.push(std::make_pair("(", 0));
61 continue;
62 }
63 if (prec(E[i], 1) && (i == 0 || E[i - 1] == "(" || prec(E[i - 1], 0))) {
64 ops.push(std::make_pair(E[i], 1));
65 continue;
66 }
67 while(prec(ops.top().first, ops.top().second) >= prec(E[i], 0)) {
68 std::string op = ops.top().first;
69 bool is_unary = ops.top().second;
70 ops.pop();
71 if (op == "(") break;
72 int y = vals.top(); vals.pop();
73 if (is_unary) {
74 vals.push(calc1(op, y));
75 } else {
76 int x = vals.top(); vals.pop();
77 vals.push(calc2(op, x, y));
78 }
79 }
80 if (E[i] != ")") ops.push(std::make_pair(E[i], 0));
81 }
82 return vals.top();
83 }
84
85 /*
6.3. String Searching 301
86
87 Split a string expression to tokens, ignoring whitespace delimiters.
88 A vector of tokens is a more flexible format since you can decide to
89 parse the expression however you wish just by modifying this function.
90 e.g. "1+(51 * -100)" converts to {"1","+","(","51","*","-","100",")"}
91
92 */
93
94 std::vector<std::string> split_expr(const std::string &s,
95 const std::string &delim = " \n\t\v\f\r") {
96 std::vector<std::string> ret;
97 std::string acc = "";
98 for (int i = 0; i < (int)s.size(); i++)
99 if (s[i] >= 0 && s[i] <= 9) {
100 acc += s[i];
101 } else {
102 if (i > 0 && s[i - 1] >= 0 && s[i - 1] <= 9)
103 ret.push_back(acc);
104 acc = "";
105 if (delim.find(s[i]) != std::string::npos) continue;
106 ret.push_back(std::string("") + s[i]);
107 }
108 if (s[s.size() - 1] >= 0 && s[s.size() - 1] <= 9)
109 ret.push_back(acc);
110 return ret;
111 }
112
113 int eval(const std::string & s) {
114 return eval(split_expr(s));
115 }
116
118
121
122 int main() {
123 cout << eval("1+2*3*4+3*(2+2)-100") << endl;
124 return 0;
125 }
6.3 String Searching
6.3.1 Longest Common Substring
1 /*
2
3 Given an text and a pattern to be searched for within the text,
4 determine the first position in which the pattern occurs in
5 the text. The KMP algorithm is much faster than the naive,
6 quadratic time, string searching algorithm that is found in
7 string.find() in the C++ standard library.
8
9 KMP generates a table using a prefix function of the pattern.
10 Then, the precomputed table of the pattern can be used indefinitely
11 for any number of texts.
12
13 Time Complexity: O(n + m) where n is the length of the text
14 and m is the length of the pattern.
15
16 Space Complexity: O(m) auxiliary on the length of the pattern.
17
18 */
19
22
23 int find(const std::string & text, const std::string & pattern) {
24 if (pattern.empty()) return 0;
25 //generate table using pattern
26 std::vector<int> p(pattern.size());
27 for (int i = 0, j = p[0] = -1; i < (int)pattern.size(); ) {
28 while (j >= 0 && pattern[i] != pattern[j])
29 j = p[j];
30 i++;
31 j++;
32 p[i] = (pattern[i] == pattern[j]) ? p[j] : j;
33 }
34 //use the precomputed table to search within text
35 //the following can be repeated on many different texts
36 for (int i = 0, j = 0; j < (int)text.size(); ) {
37 while (i >= 0 && pattern[i] != text[j])
38 i = p[i];
39 i++;
40 j++;
41 if (i >= (int)pattern.size())
42 return j - i;
43 }
44 return std::string::npos;
45 }
46
48
50
51 int main() {
52 assert(15 == find("ABC ABCDAB ABCDABCDABDE", "ABCDABD"));
53 return 0;
54 }
6.3.2 Longest Common Subsequence
1 /*
2
3 Given a text and multiple patterns to be searched for within the
4 text, simultaneously determine the position of all matches.
5 All of the patterns will be first required for precomputing
6 the automata, after which any input text may be given without
7 having to recompute the automata for the pattern.
8
9 Time Complexity: O(n) for build_automata(), where n is the sum of
10 all pattern lengths, and O(1) amortized for next_state(). However,
11 since it must be called m times for an input text of length m, and
6.3. String Searching 303
12 if there are z matches throughout the entire text, then the entire
13 algorithm will have a running time of O(n + m + z).
14
15 Note that in this implementation, a bitset is used to speed up
16 build_automata() at the cost of making the later text search cost
17 O(n * m). To truly make the algorithm O(n + m + z), bitset must be
18 substituted for an unordered_set, which will not encounter any
19 blank spaces during iteration of the bitset. However, for simply
20 counting the number of matches, bitsets are clearly advantages.
21
22 Space Complexity: O(l * c), where l is the sum of all pattern
23 lengths and c is the size of the alphabet.
24
25 */
26
28 #include <cstring>
29 #include <queue>
32
33 const int MAXP = 1000; //maximum number of patterns
34 const int MAXL = 10000; //max possible sum of all pattern lengths
35 const int MAXC = 26; //size of the alphabet (e.g. a..z)
36
37 //This function should be customized to return a mapping from
38 //the input alphabet (e.g. a..z) to the integers 0..MAXC-1
39 inline int map_alphabet(char c) {
40 return (int)(c - a);
41 }
42
43 std::bitset<MAXP> out[MAXL]; //std::unordered_set<int> out[MAXL]
44 int fail[MAXL], g[MAXL][MAXC + 1];
45
46 int build_automata(const std::vector<std::string> & patterns) {
47 memset(fail, -1, sizeof fail);
48 memset(g, -1, sizeof g);
49 for (int i = 0; i < MAXL; i++)
50 out[i].reset(); //out[i].clear();
51 int states = 1;
52 for (int i = 0; i < (int)patterns.size(); i++) {
53 const std::string & pattern = patterns[i];
54 int curr = 0;
55 for (int j = 0; j < (int)pattern.size(); j++) {
56 int c = map_alphabet(pattern[j]);
57 if (g[curr][c] == -1)
58 g[curr][c] = states++;
59 curr = g[curr][c];
60 }
61 out[curr][i] = out[curr][i] | 1; //out[curr].insert(i);
62 }
63 for (int c = 0; c < MAXC; c++)
64 if (g[0][c] == -1) g[0][c] = 0;
65 std::queue<int> q;
66 for (int c = 0; c <= MAXC; c++) {
67 if (g[0][c] != -1 && g[0][c] != 0) {
68 fail[g[0][c]] = 0;
69 q.push(g[0][c]);
70 }
71 }
73 int s = q.front(), t;
74 q.pop();
75 for (int c = 0; c <= MAXC; c++) {
76 t = g[s][c];
77 if (t != -1) {
78 int f = fail[s];
79 while (g[f][c] == -1)
80 f = fail[f];
81 f = g[f][c];
82 fail[t] = f;
83 out[t] |= out[f]; //out[t].insert(out[f].begin(), out[f].end());
84 q.push(t);
85 }
86 }
87 }
88 return states;
89 }
90
91 int next_state(int curr, char ch) {
92 int next = curr, c = map_alphabet(ch);
93 while (g[next][c] == -1)
94 next = fail[next];
95 return g[next][c];
96 }
97
98 /*** Example Usage (en.wikipedia.org/wiki/AhoCorasick_algorithm) ***/
99
102
103 int main() {
104 vector<string> patterns;
105 patterns.push_back("a");
106 patterns.push_back("ab");
107 patterns.push_back("bab");
108 patterns.push_back("bc");
109 patterns.push_back("bca");
110 patterns.push_back("c");
111 patterns.push_back("caa");
112 build_automata(patterns);
113
114 string text("abccab");
115 int state = 0;
116 for (int i = 0; i < (int)text.size(); i++) {
117 state = next_state(state, text[i]);
118 cout << "Matches ending at position " << i << ":" << endl;
119 if (out[state].any())
120 for (int j = 0; j < (int)out[state].size(); j++)
121 if (out[state][j])
122 cout << "" << patterns[j] << "" << endl;
123 }
124 return 0;
125 }
6.3.3 Edit Distance

6.4. Dynamic Programming 305
1 /*
2
3 Given an text and a pattern to be searched for within the text,
4 determine the positions of all patterns within the text. This
5 is as efficient as KMP, but does so through computing the
6 "Z function." For a string S, Z[i] stores the length of the longest
7 substring starting from S[i] which is also a prefix of S, i.e. the
8 maximum k such that S[j] = S [ i + j ] for all 0 <= j < k .
9
10 Time Complexity: O(n + m) where n is the length of the text
11 and m is the length of the pattern.
12
13 Space Complexity: O(m) auxiliary on the length of the pattern.
14
15 */
16
20
21 std::vector<int> z_function(const std::string & s) {
22 std::vector<int> z(s.size());
23 for (int i = 1, l = 0, r = 0; i < (int)z.size(); i++) {
24 if (i <= r)
25 z[i] = std::min(r - i + 1, z[i - l]);
26 while (i + z[i] < (int)z.size() && s[z[i]] == s[i + z[i]])
27 z[i]++;
28 if (r < i + z[i] - 1) {
29 l = i;
30 r = i + z[i] - 1;
31 }
32 }
33 return z;
34 }
35
37
40
41 int main() {
42 string text = "abcabaaaababab";
43 string pattern = "aba";
44 vector<int> z = z_function(pattern + "$" + text);
45 for (int i = (int)pattern.size() + 1; i < (int)z.size(); i++) {
46 if (z[i] == (int)pattern.size())
47 cout << "Pattern found starting at index "
48 << (i - (int)pattern.size() - 1) << "." << endl;
49 }
50 return 0;
51 }
6.4 Dynamic Programming
6.4.1 Longest Common Substring

1 /*
2
3 A substring is a consecutive part of a longer string (e.g. "ABC" is
4 a substring of "ABCDE" but "ABD" is not). Using dynamic programming,
5 determine the longest string which is a substring common to any two
6 input strings.
7
8 Time Complexity: O(n * m) where n and m are the lengths of the two
9 input strings, respectively.
10
11 Space Complexity: O(min(n, m)) auxiliary.
12
13 */
14
16
17 std::string longest_common_substring
18 (const std::string & s1, const std::string & s2) {
19 if (s1.empty() || s2.empty()) return "";
20 if (s1.size() < s2.size())
21 return longest_common_substring(s2, s1);
22 int * A = new int[s2.size()];
23 int * B = new int[s2.size()];
24 int startpos = 0, maxlen = 0;
25 for (int i = 0; i < (int)s1.size(); i++) {
26 for (int j = 0; j < (int)s2.size(); j++) {
27 if (s1[i] == s2[j]) {
28 A[j] = (i > 0 && j > 0) ? 1 + B[j - 1] : 1;
29 if (maxlen < A[j]) {
30 maxlen = A[j];
31 startpos = i - A[j] + 1;
32 }
33 } else {
34 A[j] = 0;
35 }
36 }
37 int * temp = A;
38 A = B;
39 B = temp;
40 }
41 delete[] A;
42 delete[] B;
43 return s1.substr(startpos, maxlen);
44 }
45
47
49
50 int main() {
51 assert(longest_common_substring("bbbabca", "aababcd") == "babc");
52 return 0;
53 }
6.4.2 Longest Common Subsequence
1 /*
6.4. Dynamic Programming 307
2
3 A subsequence is a sequence that can be derived from another sequence
4 by deleting some elements without changing the order of the remaining
5 elements (e.g. "ACE" is a subsequence of "ABCDE", but "BAE" is not).
6 Using dynamic programming, determine the longest string which
7 is a subsequence common to any two input strings.
8
9 In addition, the shortest common supersequence between two strings is
10 a closely related problem, which involves finding the shortest string
11 which has both input strings as subsequences (e.g. "ABBC" and "BCB" has
12 the shortest common supersequence of "ABBCB"). The answer is simply:
13 (sum of lengths of s1 and s2) - (length of LCS of s1 and s2)
14
17
18 Space Complexity: O(n * m) auxiliary.
19
20 */
21
24
25 std::string longest_common_subsequence
27 int n = s1.size(), m = s2.size();
28 std::vector< std::vector<int> > dp;
29 dp.resize(n + 1, std::vector<int>(m + 1, 0));
30 for (int i = 0; i < n; i++) {
31 for (int j = 0; j < m; j++) {
32 if (s1[i] == s2[j]) {
33 dp[i + 1][j + 1] = dp[i][j] + 1;
34 } else if (dp[i + 1][j] > dp[i][j + 1]) {
35 dp[i + 1][j + 1] = dp[i + 1][j];
36 } else {
37 dp[i + 1][j + 1] = dp[i][j + 1];
38 }
39 }
40 }
41 std::string ret;
42 for (int i = n, j = m; i > 0 && j > 0; ) {
43 if (s1[i - 1] == s2[j - 1]) {
44 ret = s1[i - 1] + ret;
45 i--;
46 j--;
47 } else if (dp[i - 1][j] < dp[i][j - 1]) {
48 j--;
49 } else {
50 i--;
51 }
52 }
53 return ret;
54 }
55
57
59
60 int main() {
61 assert(longest_common_subsequence("xmjyauz", "mzjawxu") == "mjau");

62 return 0;
63 }
6.4.3 Edit Distance
1 /*
2
3 Given two strings s1 and s2, the edit distance between them is the
4 minimum number of operations required to transform s1 into s2,
5 where each operation can be any one of the following:
6 - insert a letter anywhere into the current string
7 - delete any letter from the current string
8 - replace any letter of the current string with any other letter
9
12
13 Space Complexity: O(n * m) auxiliary.
14
15 */
16
20
21 int edit_distance(const std::string & s1, const std::string & s2) {
22 int n = s1.size(), m = s2.size();
23 std::vector< std::vector<int> > dp;
24 dp.resize(n + 1, std::vector<int>(m + 1, 0));
25 for (int i = 0; i <= n; i++) dp[i][0] = i;
26 for (int j = 0; j <= m; j++) dp[0][j] = j;
27 for (int i = 0; i < n; i++) {
28 for (int j = 0; j < m; j++) {
29 if (s1[i] == s2[j]) {
30 dp[i + 1][j + 1] = dp[i][j];
31 } else {
32 dp[i + 1][j + 1] = 1 + std::min(dp[i][j], //replace
33 std::min(dp[i + 1][j], //insert
34 dp[i][j + 1])); //delete
35 }
36 }
37 }
38 return dp[n][m];
39 }
40
42
44
45 int main() {
46 assert(edit_distance("abxdef", "abcdefg") == 2);
47 return 0;
48 }
6.5. Suffix Array and LCP 309
6.5 Suffix Array and LCP
6.5.1 O(N log 2 N ) Construction
1 /*
2
3 A suffix array SA of a string S[1..n] is a sorted array of indices of
4 all the suffixes of S ("abc" has suffixes "abc", "bc", and "c").
5 SA[i] contains the starting position of the i-th smallest suffix in S,
6 ensuring that for all 1 < i <= n, S[SA[i - 1], n] < S[A[i], n] holds.
7 It is a simple, space efficient alternative to suffix trees.
8 By binary searching on a suffix array, one can determine whether a
9 substring exists in a string in O(log n) time per query.
10
11 The longest common prefix array (LCP array) stores the lengths of the
12 longest common prefixes between all pairs of consecutive suffixes in
13 a sorted suffix array and can be found in O(n) given the suffix array.
14
15 The following algorithm uses a "gap" partitioning algorithm
16 explained here: http://stackoverflow.com/a/17763563
17
18 Time Complexity: O(n log^2 n) for suffix_array() and O(n) for
19 lcp_array(), where n is the length of the input string.
20
22
23 */
24
28
29 std::vector<long long> rank2;
30
31 bool comp(const int & a, const int & b) {
32 return rank2[a] < rank2[b];
33 }
34
35 std::vector<int> suffix_array(const std::string & s) {
36 int n = s.size();
37 std::vector<int> sa(n), rank(n);
38 for (int i = 0; i < n; i++) {
39 sa[i] = i;
40 rank[i] = (int)s[i];
41 }
42 rank2.resize(n);
43 for (int len = 1; len < n; len *= 2) {
44 for (int i = 0; i < n; i++)
45 rank2[i] = ((long long)rank[i] << 32) +
46 (i + len < n ? rank[i + len] + 1 : 0);
47 std::sort(sa.begin(), sa.end(), comp);
48 for (int i = 0; i < n; i++)
49 rank[sa[i]] = (i > 0 && rank2[sa[i - 1]] == rank2[sa[i]]) ?
50 rank[sa[i - 1]] : i;
51 }
52 return sa;
53 }
54
55 std::vector<int> lcp_array(const std::string & s,
56 const std::vector<int> & sa) {
57 int n = sa.size();
58 std::vector<int> rank(n), lcp(n - 1);
59 for (int i = 0; i < n; i++)
60 rank[sa[i]] = i;
61 for (int i = 0, h = 0; i < n; i++) {
62 if (rank[i] < n - 1) {
63 int j = sa[rank[i] + 1];
64 while (std::max(i, j) + h < n && s[i + h] == s[j + h])
65 h++;
66 lcp[rank[i]] = h;
67 if (h > 0) h--;
68 }
69 }
70 return lcp;
71 }
72
74
77
78 int main() {
79 string s("banana");
80 vector<int> sa = suffix_array(s);
81 vector<int> lcp = lcp_array(s, sa);
82 int sa_ans[] = {5, 3, 1, 0, 4, 2};
83 int lcp_ans[] = {1, 3, 0, 0, 2};
84 assert(equal(sa.begin(), sa.end(), sa_ans));
85 assert(equal(lcp.begin(), lcp.end(), lcp_ans));
86 return 0;
87 }
6.5.2 O(N log N ) Construction
1 /*
2
3 A suffix array SA of a string S[1..n] is a sorted array of indices of
10
14
15 The following algorithm uses a "gap" partitioning algorithm
16 explained here: http://stackoverflow.com/a/17763563, except that the
17 O(n log n) comparison-based sort is substituted for an O(n) counting
18 sort to reduce the running time by an order of log n.
19
20 Time Complexity: O(n log n) for suffix_array() and O(n) for
21 lcp_array(), where n is the length of the input string.

22
24
25 */
26
30
31 const std::string * str;
32
33 bool comp(const int & a, const int & b) {
34 return (*str)[a] < (*str)[b];
35 }
36
39 std::vector<int> sa(n), order(n), rank(n);
40 for (int i = 0; i < n; i++)
41 order[i] = n - 1 - i;
42 str = &s;
43 std::stable_sort(order.begin(), order.end(), comp);
44 for (int i = 0; i < n; i++) {
45 sa[i] = order[i];
46 rank[i] = (int)s[i];
47 }
48 std::vector<int> r(n), cnt(n), _sa(n);
49 for (int len = 1; len < n; len *= 2) {
50 r = rank;
51 _sa = sa;
52 for (int i = 0; i < n; i++)
53 cnt[i] = i;
54 for (int i = 0; i < n; i++) {
55 if (i > 0 && r[sa[i - 1]] == r[sa[i]] && sa[i - 1] + len < n &&
56 r[sa[i - 1] + len / 2] == r[sa[i] + len / 2]) {
57 rank[sa[i]] = rank[sa[i - 1]];
58 } else {
59 rank[sa[i]] = i;
60 }
61 }
62 for (int i = 0; i < n; i++) {
63 int s1 = _sa[i] - len;
64 if (s1 >= 0)
65 sa[cnt[rank[s1]]++] = s1;
66 }
67 }
68 return sa;
69 }
70
75 for (int i = 0; i < n; i++)
76 rank[sa[i]] = i;
77 for (int i = 0, h = 0; i < n; i++) {
78 if (rank[i] < n - 1) {
79 int j = sa[rank[i] + 1];

81 h++;
83 if (h > 0) h--;
84 }
85 }
86 return lcp;
87 }
88
90
93
94 int main() {
98 int sa_ans[] = {5, 3, 1, 0, 4, 2};
99 int lcp_ans[] = {1, 3, 0, 0, 2};
102 return 0;
103 }
6.5.3 O(N log N ) Construction (DC3/Skew)
1 /*
2
3 A suffix array SA of a string S[1, n] is a sorted array of indices of
10
14
15 The following implementation uses the sophisticated DC3/skew algorithm
16 by Karkkainen & Sanders (2003), using radix sort on integer alphabets
17 for linear construction. The function suffix_array(s, SA, n, K) takes
18 in s, an array [0, n - 1] of ints with n values in the range [1, K].
19 It stores the indices defining the suffix array into SA. The last value
20 of the input array s[n - 1] must be equal 0, the sentinel character. A
21 C++ wrapper function suffix_array(std::string) is implemented below it.
22
23 Time Complexity: O(n) for suffix_array() and lcp_array(), where n is
24 the length of the input string.
25
27
28 */
29
30 inline bool leq(int a1, int a2, int b1, int b2) {
31 return a1 < b1 || (a1 == b1 && a2 <= b2);

32 }
33
34 inline bool leq(int a1, int a2, int a3, int b1, int b2, int b3) {
35 return a1 < b1 || (a1 == b1 && leq(a2, a3, b2, b3));
36 }
37
38 static void radix_pass(int * a, int * b, int * r, int n, int K) {
39 int *c = new int[K + 1];
40 for (int i = 0; i <= K; i++)
41 c[i] = 0;
42 for (int i = 0; i < n; i++)
43 c[r[a[i]]]++;
44 for (int i = 0, sum = 0; i <= K; i++) {
45 int tmp = c[i];
46 c[i] = sum;
47 sum += tmp;
48 }
49 for (int i = 0; i < n; i++)
50 b[c[r[a[i]]]++] = a[i];
51 delete[] c;
52 }
53
54 void suffix_array(int * s, int * sa, int n, int K) {
55 int n0 = (n + 2) / 3, n1 = (n + 1) / 3, n2 = n / 3, n02 = n0 + n2;
56 int *s12 = new int[n02 + 3], *SA12 = new int[n02 + 3];
57 s12[n02] = s12[n02 + 1] = s12[n02 + 2] = 0;
58 SA12[n02] = SA12[n02 + 1] = SA12[n02 + 2] = 0;
59 int *s0 = new int[n0], *SA0 = new int[n0];
60 for (int i = 0, j = 0; i < n + n0 - n1; i++)
61 if (i % 3 != 0) s12[j++] = i;
62 radix_pass(s12 , SA12, s + 2, n02, K);
63 radix_pass(SA12, s12 , s + 1, n02, K);
64 radix_pass(s12 , SA12, s , n02, K);
65 int name = 0, c0 = -1, c1 = -1, c2 = -1;
66 for (int i = 0; i < n02; i++) {
67 if (s[SA12[i]] != c0 || s[SA12[i] + 1] != c1 || s[SA12[i] + 2] != c2) {
68 name++;
69 c0 = s[SA12[i]];
70 c1 = s[SA12[i] + 1];
71 c2 = s[SA12[i] + 2];
72 }
73 if (SA12[i] % 3 == 1)
74 s12[SA12[i] / 3] = name;
75 else
76 s12[SA12[i] / 3 + n0] = name;
77 }
78 if (name < n02) {
79 suffix_array(s12, SA12, n02, name);
80 for (int i = 0; i < n02; i++)
81 s12[SA12[i]] = i + 1;
82 } else {
83 for (int i = 0; i < n02; i++)
84 SA12[s12[i] - 1] = i;
85 }
86 for (int i = 0, j = 0; i < n02; i++)
87 if (SA12[i] < n0)
88 s0[j++] = 3 * SA12[i];
89 radix_pass(s0, SA0, s, n0, K);
90 #define GetI() (SA12[t] < n0 ? SA12[t] * 3 + 1 : (SA12[t] - n0) * 3 + 2)

91 for (int p = 0, t = n0 - n1, k = 0; k < n; k++) {
92 int i = GetI(), j = SA0[p];
93 if (SA12[t] < n0 ? leq(s[i], s12[SA12[t] + n0],s[j], s12[j/3]) :
94 leq(s[i], s[i + 1], s12[SA12[t] - n0 + 1], s[j], s[j + 1], s12[j / 3 + n0])) {
95 sa[k] = i;
96 if (++t == n02)
97 for (k++; p < n0; p++, k++)
98 sa[k] = SA0[p];
99 } else {
100 sa[k] = j;
101 if (++p == n0)
102 for (k++; t < n02; t++, k++)
103 sa[k] = GetI();
104 }
105 }
106 #undef GetI
107 delete[] s12;
108 delete[] SA12;
109 delete[] SA0;
110 delete[] s0;
111 }
112
115
116 // C++ wrapper function
119 int *str = new int[n + 5], *sa = new int[n + 1];
120 for (int i = 0; i < n + 5; i++) str[i] = 0;
121 for (int i = 0; i < n; i++) str[i] = (int)s[i];
122 suffix_array(str, sa, n + 1, 256);
123 return std::vector<int>(sa + 1, sa + n + 1);
124 }
125
130 for (int i = 0; i < n; i++)
131 rank[sa[i]] = i;
132 for (int i = 0, h = 0; i < n; i++) {
133 if (rank[i] < n - 1) {
134 int j = sa[rank[i] + 1];
136 h++;
138 if (h > 0) h--;
139 }
140 }
141 return lcp;
142 }
143
145
148
6.6. String Data Structures 315
149 int main() {

153 int sa_ans[] = {5, 3, 1, 0, 4, 2};
154 int lcp_ans[] = {1, 3, 0, 0, 2};
157 return 0;
158 }
6.6 String Data Structures
6.5.1 Simple Trie
1 /*
2
3 A trie, digital tree, or prefix tree, is an ordered tree data
4 structure that is used to store a dynamic set or associative array
5 where the keys are strings. Each leaf node represents a string that
6 has been inserted into the trie. This makes tries easier to implement
7 than balanced binary search trees, and also potentially faster.
8
9 Time Complexity: O(n) for insert(), contains(), and erase(), where
10 n is the length of the string being inserted, searched, or erased.
11
12 Space Complexity: At worst O(l * ALPHABET_SIZE), where l is the
13 sum of all lengths of strings that have been inserted so far.
14
15 */
16
18
19 class trie {
20 static const int ALPHABET_SIZE = 26;
21
22 static int map_alphabet(char c) {
24 }
25
26 struct node_t {
27 bool leaf;
28
29 node_t * children[ALPHABET_SIZE];
30
31 node_t(): leaf(false) {
32 for (int i = 0; i < ALPHABET_SIZE; i++)
33 children[i] = 0;
34 }
35
36 bool is_free() {
38 if (this->children[i] != 0) return true;
39 return false;
40 }
41 } *root;
42
43 bool erase(const std::string & s, node_t * n, int depth) {
45 if (depth == (int)s.size()) {
46 if (n->leaf) {
47 n->leaf = false;
48 return n->is_free();
49 }
50 } else {
51 int idx = map_alphabet(s[depth]);
52 if (erase(s, n->children[idx], depth + 1)) {
53 delete n->children[idx];
54 return !n->leaf && n->is_free();
55 }
56 }
57 return false;
58 }
59
61 if (n == 0 || n->leaf) return;
63 clean_up(n->children[i]);
64 delete n;
65 }
66
67 public:
68 trie() { root = new node_t(); }
69 ~trie() { clean_up(root); }
70
71 void insert(const std::string & s) {
72 node_t * n = root;
74 int c = map_alphabet(s[i]);
75 if (n->children[c] == 0)
76 n->children[c] = new node_t();
77 n = n->children[c];
78 }
79 n->leaf = true;
80 }
81
82 bool contains(const std::string & s) {
83 node_t *n = root;
86 if (n->children[c] == 0)
87 return false;
88 n = n->children[c];
89 }
90 return n != 0 && n->leaf;
91 }
92
93 bool erase(const std::string & s) {
94 return erase(s, root, 0);
95 }
96 };
97
99

102
103 int main() {
104 string s[8] = {"a", "to", "tea", "ted", "ten", "i", "in", "inn"};
105 trie t;
106 for (int i = 0; i < 8; i++)
107 t.insert(s[i]);
108 assert(t.contains("ten"));
109 t.erase("tea");
110 assert(!t.contains("tea"));
111 return 0;
112 }
6.5.2 Radix Trie
1 /*
2
3 A radix tree, radix trie, patricia trie, or compressed trie is a
4 data structure that is used to store a dynamic set or associative
5 array where the keys are strings. Each leaf node represents a string
6 that has been inserted into the trie. Unlike simple tries, radix
7 tries are space-optimized by merging each node that is an only child
8 with its parent.
9
10 Time Complexity: O(n) for insert(), contains(), and erase(), where
11 n is the length of the string being inserted, searched, or erased.
12
13 Space Complexity: At worst O(l), where l is the sum of all lengths
14 of strings that have been inserted so far.
15
16 */
17
20
21 class radix_trie {
22 struct node_t {
23 std::string label;
24 std::vector<node_t*> children;
25
26 node_t(const std::string & s = "") {
27 label = s;
28 }
29 } *root;
30
31 unsigned int lcplen(const std::string & s, const std::string & t) {
32 int minsize = (t.size() < s.size()) ? t.size() : s.size();
33 if (minsize == 0) return 0;
34 unsigned int res = 0;
35 for (int i = 0; i < minsize && s[i] == t[i]; i++)
36 res++;
37 return res;
38 }
39
40 void insert(const std::string & s, node_t * n) {
41 unsigned int lcp = lcplen(s, n->label);
42 if (lcp == 0 || n == root ||
43 (lcp > 0 && lcp < s.size() && lcp >= n->label.size())) {
44 bool inserted = false;
45 std::string newstr = s.substr(lcp, s.size() - lcp);
46 for (int i = 0; i < (int)n->children.size(); i++) {
47 if (n->children[i]->label[0] == newstr[0]) {
48 inserted = true;
49 insert(newstr, n->children[i]);
50 }
51 }
52 if (!inserted)
53 n->children.push_back(new node_t(newstr));
54 } else if (lcp < s.size()) {
55 node_t * t = new node_t();
56 t->label = n->label.substr(lcp, n->label.size() - lcp);
57 t->children.assign(n->children.begin(), n->children.end());
58 n->label = s.substr(0, lcp);
59 n->children.assign(1, t);
60 n->children.push_back(new node_t(s.substr(lcp, s.size() - lcp)));
61 }
62 }
63
64 void erase(const std::string & s, node_t * n) {
66 if (lcp == 0 || n == root ||
69 for (int i = 0; i < (int)n->children.size(); i++) {
70 if (n->children[i]->label[0] == newstr[0]) {
71 if (newstr == n->children[i]->label &&
72 n->children[i]->children.empty()) {
73 n->children.erase(n->children.begin() + i);
74 return;
75 }
76 erase(newstr, n->children[i]);
77 }
78 }
79 }
80 }
81
82 bool contains(const std::string & s, node_t * n) {
84 if (lcp == 0 || n == root ||
87 for (int i = 0; i < (int)n->children.size(); i++)
88 if (n->children[i]->label[0] == newstr[0])
89 return contains(newstr, n->children[i]);
90 return false;
91 }
92 return lcp == n->label.size();
93 }
94
98 clean_up(n->children[i]);
99 delete n;
100 }
101
102 public:
103 template <class UnaryFunction>
104 void walk(node_t * n, UnaryFunction f) {
105 if (n == 0) return;
106 if (n != root) f(n->label);
108 walk(n->children[i], f);
109 }
110
111 radix_trie() { root = new node_t(); }
112 ~radix_trie() { clean_up(root); }
113
114 void insert(const std::string & s) { insert(s, root); }
115 void erase(const std::string & s) { erase(s, root); }
116 bool contains(const std::string & s) { return contains(s, root); }
117
118 template <class UnaryFunction> void walk(UnaryFunction f) {
119 walk(root, f);
120 }
121 };
122
124
127
128 string preorder;
129
130 void concat(const string & s) {
131 preorder += (s + " ");
132 }
133
134 int main() {
135 {
136 string s[8] = {"a", "to", "tea", "ted", "ten", "i", "in", "inn"};
137 radix_trie t;
138 for (int i = 0; i < 8; i++)
139 t.insert(s[i]);
140 assert(t.contains("ten"));
141 t.erase("tea");
142 assert(!t.contains("tea"));
143 }
144 {
145 radix_trie t;
146 t.insert("test");
147 t.insert("toaster");
148 t.insert("toasting");
149 t.insert("slow");
150 t.insert("slowly");
151 preorder = "";
152 t.walk(concat);
153 assert(preorder == "t est oast er ing slow ly ");
154 }
155 return 0;
156 }
6.5.3 Suffix Trie

1 /*
2
3 A suffix tree of a string S is a compressed trie of all the suffixes
4 of S. While it can be constructed in O(n^2) time on the length of S
5 by simply inserting the suffixes into a radix tree, Ukkonen (1995)
6 provided an algorithm to construct one in O(n * ALPHABET_SIZE).
7
8 Suffix trees can be used for string searching, pattern matching, and
9 solving the longest common substring problem. The implementation
10 below is optimized for solving the latter.
11
12 Time Complexity: O(n) for construction of suffix_tree() and
13 per call to longest_common_substring(), respectively.
14
16
17 */
18
21
22 struct suffix_tree {
23
25
27 static const std::string ALPHABET(
28 "abcdefghijklmnopqrstuvwxyz0123456789\01\02"
29 );
30 return ALPHABET.find(c);
31 }
32
33 struct node_t {
34 int begin, end, depth;
35 node_t *parent, *suffix_link;
36 node_t *children[ALPHABET_SIZE];
37
38 node_t(int begin, int end, int depth, node_t * parent) {
39 this->begin = begin;
40 this->end = end;
41 this->depth = depth;
42 this->parent = parent;
44 children[i] = 0;
45 }
46 } *root;
47
48 suffix_tree(const std::string & s) {
50 int * c = new int[n];
51 for (int i = 0; i < n; i++) c[i] = map_alphabet(s[i]);
52 root = new node_t(0, 0, 0, 0);
53 node_t *node = root;
54 for (int i = 0, tail = 0; i < n; i++, tail++) {
55 node_t *last = 0;
56 while (tail >= 0) {
57 node_t *ch = node->children[c[i - tail]];
58 while (ch != 0 && tail >= ch->end - ch->begin) {
59 tail -= ch->end - ch->begin;
60 node = ch;
61 ch = ch->children[c[i - tail]];
62 }
63 if (ch == 0) {
64 node->children[c[i]] = new node_t(i, n,
65 node->depth + node->end - node->begin, node);
66 if (last != 0) last->suffix_link = node;
67 last = 0;
68 } else {
69 int aftertail = c[ch->begin + tail];
70 if (aftertail == c[i]) {
71 if (last != 0) last->suffix_link = node;
72 break;
73 } else {
74 node_t *split = new node_t(ch->begin, ch->begin + tail,
75 node->depth + node->end - node->begin, node);
76 split->children[c[i]] = new node_t(i, n, ch->depth + tail, split);
77 split->children[aftertail] = ch;
78 ch->begin += tail;
79 ch->depth += tail;
80 ch->parent = split;
81 node->children[c[i - tail]] = split;
82 if (last != 0)
83 last->suffix_link = split;
84 last = split;
85 }
86 }
87 if (node == root) {
88 tail--;
89 } else {
90 node = node->suffix_link;
91 }
92 }
93 }
94 }
95 };
96
97 int lcs_begin, lcs_len;
98
99 int lcs_rec(suffix_tree::node_t * n, int i1, int i2) {
100 if (n->begin <= i1 && i1 < n->end) return 1;
101 if (n->begin <= i2 && i2 < n->end) return 2;
102 int mask = 0;
103 for (int i = 0; i < suffix_tree::ALPHABET_SIZE; i++) {
104 if (n->children[i] != 0)
105 mask |= lcs_rec(n->children[i], i1, i2);
106 }
107 if (mask == 3) {
108 int curr_len = n->depth + n->end - n->begin;
109 if (lcs_len < curr_len) {
110 lcs_len = curr_len;
111 lcs_begin = n->begin;
112 }
113 }
114 return mask;
115 }
116
117 std::string longest_common_substring
119 std::string s(s1 + \01 + s2 + \02);

120 suffix_tree tree(s);
121 lcs_begin = lcs_len = 0;
122 lcs_rec(tree.root, s1.size(), s1.size() + s2.size() + 1);
123 return s.substr(lcs_begin - 1, lcs_len);
124 }
125
127
129
130 int main() {
131 assert(longest_common_substring("bbbabca", "aababcd") == "babc");
132 return 0;
133 }
6.5.4 Suffix Automaton
1 /*
2
3 A suffix automaton is a data structure to efficiently represent the
4 suffixes of a string. It can be considered a compressed version of
5 a suffix tree. The data structure supports querying for substrings
6 within the text from with the automaton is constructed in linear
7 time. It also supports computation of the longest common substring
8 in linear time.
9
10 Time Complexity: O(n * ALPHABET_SIZE) for construction, and O(n)
11 for find_all(), as well as longest_common_substring().
12
13 Space Complexity: O(n * ALPHABET_SIZE) auxiliary.
14
15 */
16
18 #include <queue>
21
22 struct suffix_automaton {
23
25
28 }
29
30 struct state_t {
31 int length, suffix_link;
32 int firstpos, next[ALPHABET_SIZE];
33 std::vector<int> invlinks;
34
35 state_t() {
36 length = 0;
37 suffix_link = 0;
38 firstpos = -1;
40 next[i] = -1;
41 }
42 };
43
44 std::vector<state_t> states;
45
46 suffix_automaton(const std::string & s) {
48 states.resize(std::max(2, 2 * n - 1));
49 states[0].suffix_link = -1;
50 int last = 0;
51 int size = 1;
52 for (int i = 0; i < n; i++) {
54 int curr = size++;
55 states[curr].length = i + 1;
56 states[curr].firstpos = i;
57 int p = last;
58 while (p != -1 && states[p].next[c] == -1) {
59 states[p].next[c] = curr;
60 p = states[p].suffix_link;
61 }
62 if (p == -1) {
63 states[curr].suffix_link = 0;
64 } else {
65 int q = states[p].next[c];
66 if (states[p].length + 1 == states[q].length) {
67 states[curr].suffix_link = q;
68 } else {
69 int clone = size++;
70 states[clone].length = states[p].length + 1;
72 states[clone].next[i] = states[q].next[i];
73 states[clone].suffix_link = states[q].suffix_link;
74 while (p != -1 && states[p].next[c] == q) {
75 states[p].next[c] = clone;
76 p = states[p].suffix_link;
77 }
78 states[q].suffix_link = clone;
79 states[curr].suffix_link = clone;
80 }
81 }
82 last = curr;
83 }
84 for (int i = 1; i < size; i++)
85 states[states[i].suffix_link].invlinks.push_back(i);
86 states.resize(size);
87 }
88
89 std::vector<int> find_all(const std::string & s) {
91 int node = 0;
93 int next = states[node].next[map_alphabet(s[i])];
94 if (next == -1) return res;
95 node = next;
96 }
97 std::queue<int> q;
98 q.push(node);
100 int curr = q.front();
101 q.pop();
102 if (states[curr].firstpos != -1)
103 res.push_back(states[curr].firstpos - (int)s.size() + 1);
104 for (int j = 0; j < (int)states[curr].invlinks.size(); j++)
105 q.push(states[curr].invlinks[j]);
106 }
107 return res;
108 }
109
110 std::string longest_common_substring(const std::string & s) {
111 int len = 0, bestlen = 0, bestpos = -1;
112 for (int i = 0, cur = 0; i < (int)s.size(); i++) {
114 if (states[cur].next[c] == -1) {
115 while (cur != -1 && states[cur].next[c] == -1)
116 cur = states[cur].suffix_link;
117 if (cur == -1) {
118 cur = len = 0;
119 continue;
120 }
121 len = states[cur].length;
122 }
123 len++;
124 cur = states[cur].next[c];
125 if (bestlen < len) {
126 bestlen = len;
127 bestpos = i;
128 }
129 }
130 return s.substr(bestpos - bestlen + 1, bestlen);
131 }
132 };
133
135
139
140 int main() {
141 {
142 suffix_automaton sa("bananas");
143 vector<int> pos_a, pos_an, pos_ana;
144 int ans_a[] = {1, 3, 5};
145 int ans_an[] = {1, 3};
146 int ans_ana[] = {1, 3};
147 pos_a = sa.find_all("a");
148 pos_an = sa.find_all("an");
149 pos_ana = sa.find_all("ana");
150 assert(equal(pos_a.begin(), pos_a.end(), ans_a));
151 assert(equal(pos_an.begin(), pos_an.end(), ans_an));
152 assert(equal(pos_ana.begin(), pos_ana.end(), ans_ana));
153 }
154 {
155 suffix_automaton sa("bbbabca");
156 assert(sa.longest_common_substring("aababcd") == "babc");
157 }
158 return 0;
159 }

A3c5 PDF

Uploaded by

Copyright:

Available Formats

A3c5 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A3c5 PDF

Uploaded by

Copyright:

Available Formats

Alexs Anthology of Algorithms

Common Code for Contests in Concise C++

Compiled, Edited, and Written by

Note: Visit http://github.com/Alextrovert/Algorithm-Anthology for the most up-to-date digital

0.3 Usage Notes

1.4.2 Brents Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.8.2 Shortest Hamiltonian Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

3.5.4 Hashmap (Chaining) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

4.6.1 Real Root Finding (Differentiation) . . . . . . . . . . . . . . . . . . . . . . . . . . 228

6.2.3 Shunting Yard Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299

1.1 Array Transformations

1.1.1 Sorting Algorithms

95 while (a <= mid && c < hi)

154 if (child + 1 < j && comp(*child, *(child + 1)))

272 Sorting five million integers...

1.1.2 Array Rotation

1.1.3 Counting Inversions

1.1.4 Coordinate Compression

1.1.5 Selection (Quickselect)

23 #include <iterator> /* std::iterator_traits */

1.2 Array Queries

1.2.1 Longest Increasing Subsequence

1.2.2 Maximal Subarray Sum (Kadanes)

99 sum = max_subarray_sum(sums.begin(), sums.end(), &begin, &end);

1.2.3 Majority Element (Boyer-Moore)

1.2.4 Subset Sum (Meet-in-the-Middle)

50 while (l < llen && r >= 0) {

1.2.5 Maximal Zero Submatrix

1.3.1 Discrete Binary Search

17 predicate pred(k) tests true. This function is correct if and only if

1.3.2 Ternary Search

6 ternary_search_min() takes the domain [lo, hi] of a continuous function

1.3.3 Hill Climbing

37 double x = x0, y = y0, res = f(x0, y0);

1.3.4 Convex Hull Trick (Semi-Dynamic)

4 process a set of x-coordinate queries each asking to find the minimum

1.3.5 Convex Hull Trick (Fully Dynamic)

1.4 Cycle Detection

1.4.1 Floyds Algorithm

60 using namespace std;

1.4.2 Brents Algorithm

23 sequence on which a cycle starts, and lambda is the cycles length.

1.5 Binary Exponentiation

2.1 Depth-First Search

2.1.1 Graph Class and Depth-First Search

36 #include <algorithm> /* std::max */

95 dfs(start, vis, act);

154 g.add_edge(2, 3);

2.1.2 Topological Sorting

31 #include <algorithm> /* std::fill(), std::reverse() */

2.1.3 Eulerian Cycles

10 degree belong to a single connected component.

2.1.4 Unweighted Tree Centers

80 pair<int, int> res = make_pair(depth, u);

2.2 Shortest Paths

2.2.1 Breadth First Search

24 Take the path: 0->1->3.

2.2.2 Dijkstras Algorithm

57 vector<bool> vis(nodes, false);

154 if (child + 1 < j && comp(child, (child + 1)))