Pruning Symbolic Search Spaces using Dimensional Analysis

This direction is split into other directions: dimensional inconsistency. Its last version stays readable and citable here.

Version 1 of 157 papers94 conceptsfrozen 2026-10-04

The "holy grail" of automated science is Symbolic Regression (SR): the ability to look at a table of numbers and discover the underlying mathematical law, like $F = G \frac{m_1 m_2}{r^2}$, that generated them. However, SR faces a daunting "combinatorial explosion." If you have just a few variables and a handful of operators ($+, -, \times, \div, \sin, \exp$), the number of possible equations grows exponentially with the complexity of the formula. Most of these equations are "word salad"—mathematically valid but physically nonsensical.

The research direction across these papers focuses on a powerful solution: Dimensional Analysis. By treating physical units (meters, seconds, kilograms) as a rigorous "type system" for mathematics, researchers can prune the search space of millions of impossible equations, leaving only the handful that are physically plausible.

Units as a Mathematical Type System

In computer science, a type system prevents you from trying to "add" a string to an integer. In physics, the principle of dimensional homogeneity does the same: you cannot add a distance (meters) to a mass (kilograms).

Early systems like COPER, developed by Mieczyslaw M. Kokar, demonstrated that by searching for a "representation base" that respects unit invariance, an algorithm can find exact physical laws without resorting to high-degree polynomial curve fitting [Discovering Functional Formulas through Changing Representation Base]. This approach shifts the problem from "fitting curves" to "finding the right coordinate system" where the law becomes simple.

Brian Falkenhainer expanded this with ABACUS, which used units analysis to ensure that addition and subtraction only occurred between variables with identical units [Proportionality Graphs, Units Analysis, and Domain Constraints]. This simple heuristic acts as a massive filter, discarding physically impossible formulas before they are even evaluated against the data.

Pruning the Combinatorial Explosion

Why does this work so well? Because physical constraints are incredibly restrictive. In a typical symbolic search, the space of possible expressions is effectively infinite. However, as Arjun Guha and colleagues showed in their work on LDIPS, enforcing dimensional consistency can reduce the number of candidate features by nearly two orders of magnitude [Robot Action Selection Learning via Layered Dimension Informed Program Synthesis].

Modern frameworks implement these constraints in several ways:

  1. Grammar-based Constraints: Marc Schoenauer and Michèle Sébag proposed using formal grammars to define the rules of well-formed physical expressions [Using Domain Knowledge in Evolutionary System Identification]. If the grammar only allows adding "length" to "length," the evolutionary algorithm never wastes time exploring "length + mass."
  2. Hard Linear Constraints: In the FIND framework, Tingxiong Xiao treats dimensional consistency as a hard linear constraint on the exponents of the variables [FIND: A Framework for Discovering Formulas in Data]. This reduces the search from millions of iterations to just dozens, making discovery efficient enough for real-time applications.
  3. Dimension Repair: Huan Zhang and others introduced GPDR, which uses Mixed-Integer Linear Programming to "repair" dimensionally inconsistent rules, ensuring that evolved models are always interpretable by human practitioners [Genetic Programming with Model Driven Dimension Repair].

Integrating Physics into Modern AI

The most recent work bridges the gap between classical dimensional analysis and modern deep learning. AI Feynman, developed by Silviu-Marian Udrescu and Max Tegmark, uses neural networks to identify symmetries and separability in data, but relies on dimensional analysis as a primary "simplification" strategy to recover 100% of the equations in the Feynman Lectures [AI Feynman: a Physics-Inspired Method for Symbolic Regression].

Even Large Language Models (LLMs) are being recruited into this effort. Parshin Shojaee introduced LLM-SR, which uses the semantic knowledge of LLMs to understand the meaning of variables (e.g., knowing that "v" stands for velocity) to prune the search space of dimensionally inconsistent hypotheses [LLM-SR: Scientific Equation Discovery via Programming with Large Language Models].

Similarly, the PhyE2E framework by Haowei Lin and colleagues predicts physical units and operators simultaneously, using unit consistency as a strong inductive bias that allows the model to discover laws from as few as five data points [A Neural Symbolic Model for Space Physics].

Why It Works: The Power of Inductive Bias

The success of this research direction stems from a fundamental realization: data alone is often not enough. Purely data-driven symbolic regression, as noted by Michael Kommenda, frequently produces "mathematically accurate but physically nonsensical" equations [Data Mining using Unguided Symbolic Regression].

By embedding the laws of units directly into the search process, researchers provide the algorithm with "physical common sense." This doesn't just make the search faster; it ensures that the discovered formulas generalize better to new scenarios because they respect the fundamental symmetries of the universe. As Miles Cranmer emphasizes with PySR, the goal is to move beyond black-box models toward human-interpretable equations that align with our existing scientific understanding [Interpretable Machine Learning for Science with PySR].

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_pruning_symbolic_search_spaces_using_dim_a49b98d5,
  title = {Pruning Symbolic Search Spaces using Dimensional Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pruning-symbolic-search-spaces-using-dimensional-analysis/rdr_a49b98d51cb44fbda44d75e8e349cf04},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_pruning_symbolic_search_spaces_using_dim_a49b98d5_v1,
  title = {Pruning Symbolic Search Spaces using Dimensional Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pruning-symbolic-search-spaces-using-dimensional-analysis/rdr_a49b98d51cb44fbda44d75e8e349cf04/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}