Dimensional Constraints for Symbolic Equation Discovery

67 concepts40 papers

Symbolic regression aims to uncover the underlying mathematical formulas that govern empirical data. Unlike black-box neural networks, it provides interpretable, "white-box" equations. However, searching the infinite space of mathematical expressions introduces a severe combinatorial explosion. If an algorithm blindly combines operators ($+, -, \times, \div$) and variables, the number of candidate equations grows exponentially with the complexity of the formula.

Worse, standard data-driven optimization algorithms are "physics-blind." As highlighted by Reissmann et al. and Shojaee et al., unconstrained symbolic regression frequently produces formulas that fit training data points with high numerical accuracy but are physically nonsensical—such as attempting to add meters to seconds.

To solve this, a powerful paradigm has emerged: using dimensional consistency as a hard constraint to prune the mathematical search space. By treating the laws of physics as a strict type system, researchers can eliminate physically impossible equations before they are ever evaluated.


Physics as a Type System

In computer science, a strongly typed language prevents you from adding a string to an integer. In the physical sciences, the principle of dimensional homogeneity acts exactly the same way. Every physical variable possesses a dimension that can be represented as a vector of base unit exponents (such as Mass, Length, Time, and Charge).

As Durasevic et al. demonstrate, representing physical quantities as vectors of base unit exponents allows automated discovery systems to formally verify dimensional consistency. This creates a strict grammar for mathematical operations:

  • Multiplication and Division: Allowed between any dimensions. The resulting dimension vector is simply the sum or difference of the operand vectors (e.g., Length divided by Time yields Velocity, $[1, 0, 0] - [0, 1, 0] = [1, -1, 0]$).
  • Addition and Subtraction: Strictly forbidden unless the dimensional signatures of both operands are identical. You can add a velocity to a velocity, but never a velocity to a distance.
  • Transcendental Functions: Functions like $\sin(x)$, $\exp(x)$, or $\log(x)$ require their arguments to be completely dimensionless ($[0, 0, 0]$).

By enforcing these rules, the search space undergoes a massive reduction. Guha et al. show that this physical type-checking can reduce the number of candidate features by nearly two orders of magnitude. In fact, Durasevic et al. point out that for certain dimensional signatures, the constraints are so restrictive that only a single mathematically valid expression can exist, collapsing a complex optimization problem into an instantaneous discovery.


Mechanisms for Enforcing Consistency

Researchers have developed several distinct strategies to embed these dimensional guardrails into search algorithms, moving from early heuristic filters to modern generative architectures.

1. Grammars and Algebraic Constraints

Instead of generating arbitrary equations and filtering out the bad ones, modern frameworks use dimensional analysis to restrict the generation process itself. Srivastava et al. utilize a context-free grammar built directly on dimensional analysis, ensuring that the search engine only constructs physically meaningful combinations of units. Similarly, Xiao et al. derive linear constraints from the physical units of variables to restrict the search space of exponents, ensuring that all generated mathematical operations naturally obey unit laws.

2. Knowledge-Guided Deep Learning

Rather than relying purely on evolutionary algorithms, dimensional constraints can be embedded directly into neural networks. For instance, in earth science applications, Li et al. integrate dimensional consistency and hierarchical sub-libraries into deep symbolic regression. This structural guardrail prevents the network from exploring physically impossible combinations, enabling the recovery of complex, multi-variable scientific laws from noisy data.

3. Continuous Generative Spaces

To avoid navigating discrete, chaotic equation trees, recent frameworks map equations into smooth, continuous representations. The GenSR framework, developed by Li et al., integrates dimensional analysis directly into a generative latent space. This ensures that when the model samples or evolves equations within this continuous space, the resulting expressions inherently respect physical units and laws.


Beyond Units: Symmetries and Invariants

Dimensional consistency is the most fundamental layer of a broader philosophy: constraining data-driven models using the invariant properties of the physical universe. When units alone are not enough, researchers leverage other physical priors to prune search spaces:

  • Coordinate and Scale Invariance: Early foundational systems like COPER, developed by Kokar, focused on finding a "representation base" that remains invariant under unit transformations. This prevents algorithms from falling back on high-degree polynomial curve fitting. Similarly, Washio et al. utilize constraints derived from measurement scale types (such as ratio and interval scales) to eliminate inadmissible differential equations.
  • Geometric Symmetries: Physical systems often obey rotational, translational, or permutational symmetries. Yang et al. demonstrate that constraining symbolic regression to equations that respect these fundamental symmetries dramatically improves accuracy and provides resilience against experimental noise.
  • Conservation Laws: Systems can discover hidden physical properties by enforcing conservation principles. DeGennaro et al. show that enforcing energy conservation laws drastically reduces the dimensionality of a model's parameter space. To find these laws automatically, Kasim et al. evaluate models across varying dimensions of invariant outputs; identifying the "elbow" in the loss curve reveals the exact number of independent conservation laws operating within a hidden system.

By shifting the role of AI from a blind curve-fitter to a constrained explorer of physically permissible math, dimensional and structural consistency constraints turn the combinatorial explosion of symbolic regression into a tractable, robust tool for true scientific discovery.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dimensional_inconsistency_7b8fa55f,
  title = {dimensional inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dimensional-inconsistency/rdr_7b8fa55f9d5e4e9685cae1bdd59924f3},
  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_dimensional_inconsistency_7b8fa55f_v1,
  title = {dimensional inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dimensional-inconsistency/rdr_7b8fa55f9d5e4e9685cae1bdd59924f3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake