# Final Algorithm Taxonomy (7 families, 31 categories) | ID | Family | Category | Includes | |---|---|---|---| | MA-01 | Mathematics | Math Reasoning / Formulas | arithmetic, parity, invariants, inequalities and bounds, closed-form derivation, convexity and duality arguments | | MA-02 | Mathematics | Number Theory | primes/sieve, gcd/lcm, divisors, modular arithmetic/inverse, CRT, multiplicative functions, Dirichlet convolution, min_25 | | MA-03 | Mathematics | Combinatorics & Probability | counting by formula, binomials, inclusion–exclusion, bijections, linearity of expectation, probability | | MA-04 | Mathematics | Game Theory | win/lose states (also when computed by DP or search), Sprague–Grundy, Nim, strategy arguments | | MA-05 | Mathematics | Bitwise / XOR | per-bit reasoning, XOR/AND/OR properties, XOR basis | | MA-06 | Mathematics | Linear Algebra / Polynomials | Gaussian elimination, linear systems, determinants, FFT/NTT, convolution, generating functions | | MA-07 | Mathematics | Geometry | vectors, cross product, convex hull, intersections, areas, computational geometry | | DS-01 | Data Structures | Prefix Sums / Difference Arrays | 1D/2D prefix sums, difference arrays | | DS-02 | Data Structures | Hash / Set / Map | hash table, library (multi)set/map and ordered containers, coordinate compression, XOR/sum hashing of sets | | DS-03 | Data Structures | Stack / Queue / Heap | monotonic stack/deque, priority queue | | DS-04 | Data Structures | Segment Tree / Fenwick / BST | segment tree (lazy propagation, custom merge, beats), sparse table, Fenwick tree, hand-written Treap/Splay, persistent structures | | DS-05 | Data Structures | DSU / Union-Find | union-find, weighted/rollback DSU, small-to-large merging of component data | | DP-01 | Dynamic Programming | Linear / Sequence DP | prefix-state DP, grid DP, LIS/LCS, interval DP, digit DP | | DP-02 | Dynamic Programming | Knapsack / Subset / Bitmask DP | 0/1 and unbounded knapsack, subset-sum, bitmask-state DP, SOS DP | | DP-03 | Dynamic Programming | Tree / DAG DP | DP on trees, rerooting, DP along the topological order of a DAG | | DP-04 | Dynamic Programming | DP Optimization | data-structure-optimized DP, monotone-queue optimization, matrix-power DP, CHT/Li Chao, D&C / Knuth optimization, slope trick, Aliens (WQS) trick, state reduction by observation | | GR-01 | Graph Theory | Traversal / Connectivity | BFS/DFS, components, topological sort, bipartiteness, cycles, functional graphs, Eulerian paths, SCC, 2-SAT, bridges / articulation points | | GR-02 | Graph Theory | Shortest Paths / MST | Dijkstra, 0-1 BFS, Bellman–Ford, Floyd, Kruskal/Prim | | GR-03 | Graph Theory | Flow / Matching | max-flow / min-cut, min-cost flow, bipartite matching, Hall's theorem, König's theorem | | GR-04 | Graph Theory | Tree Algorithms | LCA, binary lifting, Euler tour (tree flattening), HLD, centroid decomposition, small-to-large over subtrees (DSU on tree), diameter | | ST-01 | Strings | Pattern Matching / String Hashing | KMP / prefix function, Z-function, Manacher, rolling hash | | ST-02 | Strings | Trie / Suffix Structures | trie (incl. binary trie), Aho–Corasick, suffix array / automaton, palindromic tree, LCP | | PA-01 | Algorithmic Paradigms | Greedy | exchange argument, sort-then-greedy, regret greedy | | PA-02 | Algorithmic Paradigms | Binary / Ternary Search | binary search on the answer, parallel binary search, ternary / golden-section search | | PA-03 | Algorithmic Paradigms | Two Pointers / Sliding Window / Sweep Line | two pointers, sliding window, sweep line over sorted events, interval merging | | PA-04 | Algorithmic Paradigms | Divide & Conquer / Sqrt Decomposition | D&C on arrays, CDQ, offline D&C over time (segment tree on time), sqrt blocks, Mo's algorithm, sqrt thresholds (heavy/light by size or degree) | | PA-05 | Algorithmic Paradigms | Brute Force / Search | enumeration, backtracking, pruning, meet-in-the-middle, tables precomputed by offline search, randomized algorithms | | AH-01 | Ad-hoc & Implementation | Ad-hoc / Observation / Case Work | problem-specific structural observations that no other category names, case analysis | | AH-02 | Ad-hoc & Implementation | Constructive | build an array / string / graph / order that meets the conditions | | AH-03 | Ad-hoc & Implementation | Implementation / Simulation | direct simulation, heavy implementation, many edge cases to get right | Family sizes: Mathematics 7, Data Structures 5, Dynamic Programming 4, Graph Theory 4, Strings 2, Algorithmic Paradigms 5, Ad-hoc & Implementation 4. GPTClassify/classify.py gives the model the table above and the section below word for word, so human labels and model labels follow the same rules. ## Boundary rules These rules decide between categories when a technique could fit more than one. General 1. Place a technique by the Includes column, not by how close a category's name sounds, and by the technique, not by the data it runs on: KMP on an integer array is ST-01. 2. One idea, one label. A structure that only implements another listed technique the standard way belongs to that technique: the heap in Dijkstra and the DSU in Kruskal are GR-02; the segment tree or deque of an optimized DP is DP-04; the heap of a regret greedy is PA-01. Add the structure's own category only when the structure is itself nonstandard (custom merge, persistence, beats). 3. Not labels by themselves: sorting, a set or map used as a lookup table, lower_bound on a sorted array, a prefix sum taken once, coordinate compression, arithmetic modulo a prime only because the answer is asked modulo it, 64-bit care, fast I/O, a DFS that only computes parents, depths or subtree sizes. Observations, math and DP 4. MA-01 vs AH-01. MA-01 when the key step is a quantitative fact that is derived: a formula or closed form, an equation, an inequality or bound, parity, an invariant, convexity. AH-01 when the key step is a structural fact about the problem's own objects: which choices never matter, the form every optimal answer can be put into, what can be ignored; or a case split. Give both only when each is a separate step of the solution. 5. DP-04 vs AH-01 / MA-01. An observation that shrinks the state or the transitions of a DP that would otherwise be too slow is DP-04 ("state reduction by observation"). An observation that removes the need for DP is AH-01 or MA-01. 6. DP-04 and the base DP. Give the base DP category (DP-01 to DP-03) as well when formulating the DP is itself a main step. 7. Counting with a DP is the DP's category. MA-03 when a count or an expectation comes from a formula, a bijection, inclusion–exclusion or linearity of expectation. 8. Win/lose states computed by DP or search are MA-04; add a DP category only when that DP is substantial on its own, such as a tree DP over positions. Data structures 9. Library std::set/map and other ordered containers are DS-02. DS-04 is for segment trees, Fenwick trees, sparse tables, and hand-written BSTs (treap, splay) that carry augmented data. 10. Hashing a set or multiset (random weights, XOR/sum) is DS-02; hashing strings is ST-01; other randomization (random sampling, shuffling, probabilistic checks) is PA-05. 11. A monotonic stack or deque is DS-03, including sliding-window minimum. PA-03 when moving the ends of a window, or the pointers, is the idea. 12. Small-to-large merging over the subtrees of a tree (DSU on tree) is GR-04. Small-to-large merging of the data attached to DSU components is DS-05. 13. Offline processing over time (a segment tree on time with a rollback DSU, CDQ) is PA-04; the structure inside it (DS-05, DS-04) is labelled on its own. Graphs and trees 14. On a tree, plain DFS/BFS is GR-01 and usually incidental; GR-04 needs one of its listed techniques; a DP whose state is a subtree is DP-03. 15. Topological sort or cycle detection alone is GR-01; a DP computed along the topological order is DP-03. 16. "Euler tour" in GR-04 is flattening a tree into an array. An Eulerian path or circuit through all edges is GR-01. 17. "Heavy/light" in PA-04 is splitting items by a square-root threshold on size or degree. Heavy-light decomposition of a tree is GR-04. 18. GR-03 includes matching theory used only in the argument, such as Hall's or König's theorem, even when the code runs no flow or matching algorithm. 19. When the solution builds a graph (states, residues, layered copies), label the algorithm run on it. The modelling is part of that label, not AH-01. Math 20. Fast exponentiation of a transition matrix (a DP, a linear recurrence, walks in a graph) is DP-04. MA-06 is for other linear algebra and for polynomials. 21. An XOR basis is MA-05, not MA-06. A bitmask used as a DP state is DP-02, not MA-05. A binary trie is ST-02; add MA-05 only when per-bit reasoning is a separate step. 22. MA-07 needs real geometric computation on shapes in the plane or in space: vectors, angles, areas of polygons or other figures, intersections. Not MA-07: integrating a function (area under a curve or a piecewise-linear density), which is MA-01; rotating, reflecting or translating grid cells, which counts toward the method that uses it; axis-parallel rectangles that reduce to 1-D intervals. Paradigms and ad hoc 23. PA-01 needs a greedy choice that makes the solution correct, usually backed by an exchange argument. Sorting before processing is not greedy by itself. 24. PA-02 is binary or ternary search over an answer or a parameter. A lower_bound on a sorted array is not PA-02. 25. PA-05 when exhaustive enumeration or search is the main engine, or when the key is that the search space is small enough, including a table precomputed offline by search. Not for a loop over a few cases inside another method. 26. AH-02 when the output is an object to build (an order, array, graph or string) and building it is a main step. A numeric answer is never AH-02. 27. AH-03 only when implementation or simulation is itself a main difficulty. Long code alone is not AH-03. Case analysis in the reasoning is AH-01; many cases to code correctly is AH-03. 28. Every interactive problem gets AH-04.