MATH-Sigma: A Purely Math-Driven Chess Engine Without Neural Networks or Training Data

MATH-Sigma is a unique chess engine that completely abandons the neural network training and parameter tuning relied upon by traditional engines; all evaluation constants are derived from mathematical principles. Based on three core axioms—average branching factor, effective branching factor, and tempo value—it builds a complete search and evaluation system, achieving a playing strength of approximately 2200 Elo.

• Original Author/Maintainer: theogbob
• Source Platform: GitHub
• Original Title: math-sigma
• Original Link: https://github.com/theogbob/math-sigma
• Release Date: May 26, 2026

---

In the development history of chess engines, we have witnessed the evolution of two mainstream technical routes: traditional hand-tuned evaluation function engines like Fruit and Crafty, and the neural network engines that have become popular in recent years, such as Stockfish's NNUE version. However, an open-source project called MATH-Sigma has taken a completely different path—it attempts to answer a fundamental question: Can a competitive chess engine be built solely based on mathematical principles, without relying on game tuning or neural network training?

MATH-Sigma's answer is yes. This engine, written in Rust with approximately 3200 lines of code, has all evaluation constants derived from mathematical axioms and achieves a playing strength of around 2200 Elo, equivalent to a strong amateur player or national master level.

The core innovation of MATH-Sigma lies in that all its parameters are based on three mathematical axioms about the structure of chess:

b = 35—This is the average branching factor for generating legal moves in chess. In any position, a player usually has about 35 legal moves to choose from. This number is not an empirical estimate but a mathematical property determined by the rules of chess itself.

b_eff ≈ 7.4—The effective branching factor after move ordering optimization. Through transposition tables, killer heuristics, and good capture heuristics, the search tree can be significantly pruned, with an average of only about 7.4 branches that need to be explored in depth.

T = 78cp—Tempo value, derived from the formula P × ln(b)/(1 + ln(b)), where P=100 is the base value of a pawn. This formula quantifies the intrinsic value of having a move ahead in chess.

Based on the three axioms above, MATH-Sigma derives a complete search parameter system: