tropix 0.1.0

A Python library for tropical geometry computation, exploring the connection between ReLU neural networks and tropical rational maps.

tropix

tropix is a Python library for tropical geometry, implementing tropical polynomial arithmetic, Newton polytopes, regular subdivisions, and tropical hypersurfaces, motivated in part by their connection to ReLU neural networks through tropical rational maps (Zhang et al., 2018).

Status

Current version: v0.1.0

• Tropical arithmetic
• Tropical polynomials
• Newton polytopes
• Regular subdivisions
• Analytical duals
• Visualisation
• 3D support, ReLU network interface (planned for v0.2)

Mathematical background

Tropical geometry replaces ordinary polynomial arithmetic with the min-plus semiring (R ∪ {∞}, +, min), where addition becomes "take the minimum" and multiplication becomes ordinary addition. A tropical polynomial is a piecewise-linear function; the set of points where its minimum is achieved by two or more monomials simultaneously is its tropical hypersurface. This hypersurface is combinatorially dual to a regular subdivision of the polynomial's Newton polytope, obtained by lifting each exponent vector by its coefficient and taking the lower convex hull of the resulting point cloud.

This library implements each stage of this pipeline:

arithmetic
    |
polynomials
    |
Newton polytope
    |
subdivision
    |
analytical dual
    |
visualisation

Installation

pip install -e .

Requires Python 3.9+, NumPy, SciPy, and Matplotlib.

Quick start

from tropix import TropicalPolynomial, tropical_edges, plot_tropical_curve_dual

# the standard tropical line: min(x, y, 0)
f = TropicalPolynomial({(1, 0): 0.0, (0, 1): 0.0, (0, 0): 0.0}, nvars=2)

vertices, edges, boundary_edges, cells, index_map = tropical_edges(f)
print(f"{len(vertices)} vertex, {len(boundary_edges)} rays")
# 1 vertex, 3 rays

plot_tropical_curve_dual(f)

Running this example constructs and plots the analytical dual of the standard tropical line.

The tropical line

The tropical line has one vertex at the origin and three unbounded rays:

A non-simplicial tropical conic

This tropical conic contains a non-simplicial (quadrilateral) cell in its regular subdivision, producing three vertices, two bounded edges, and six rays instead of the four-vertex combinatorics of a fully triangulated smooth conic:

f = TropicalPolynomial({
    (2, 0): 0.0, (0, 2): 0.0, (0, 0): 0.0,
    (1, 1): -1.0, (1, 0): -0.5, (0, 1): -0.5,
}, nvars=2)

This is the analytical dual, exact rather than grid-approximated. For comparison, visualisation.py's brute-force grid approximation of the same curve's hypersurface, alongside its Newton polytope:

Together these figures illustrate the complete computational pipeline implemented by tropix, from a tropical polynomial through its Newton polytope and regular subdivision to the corresponding tropical hypersurface and its analytical dual.

Polynomial
    |
    +----------> Grid approximation of T(f)  (independent, numerical)
    |
    v
Newton polytope
    |
    v
Regular subdivision
    |
    v
Analytical dual

What's implemented (v0.1)

• arithmetic.py: TropicalNumber, the min-plus semiring. Validated domain, correct identity elements, consistent equality/hashing.
• polynomial.py: TropicalPolynomial, exponent-to-coefficient maps with full input validation and tropical arithmetic.
• newton.py: Newton polytope vertices and Minkowski sums, correctly handling the 1D case and degenerate inputs.
• subdivision.py: the regular subdivision induced by a polynomial's coefficients. Explicitly detects the degenerate (affine) case via an SVD-based rank test.
• dual.py: the analytical dual, tropical vertices, edges, and rays, computed from each cell's true geometric boundary.
• visualisation.py: brute-force grid approximation of a tropical hypersurface, and Newton polytope plotting.

All five mathematical modules have a corresponding test file; 49 tests in total, covering the semiring axioms, evaluation correctness, the 1D Newton polytope fix, the flat-subdivision fix, and the exact combinatorics of both the tropical line and a genuine tropical conic, verified against known tropical geometry results.

pip install -e ".[dev]"
pytest

Connection to neural networks

Zhang, Naitzat, and Lim (2018) show that a ReLU network computes a tropical rational map: a ratio of tropical polynomials. Under this correspondence, the network's linear regions (the pieces of input space on which it reduces to a single affine map) correspond to cells of a polyhedral complex dual to a Newton polytope subdivision, exactly the structure this library computes directly.

A companion project, nn-from-first-principles, implements a feedforward network from scratch in NumPy and empirically visualises its linear regions. Extracting the actual tropical polynomial a trained network computes, connecting that project's empirical picture to this library's exact machinery, is the natural next step (planned for v0.2).

Notes and limitations (v0.1)

• 2D only: hypersurface visualisation and the analytical dual construction currently require exactly 2 variables.
• The grid-based hypersurface plot (visualisation.py) is a brute-force approximation; the analytical dual (dual.py) is exact.
• No 3D support, no Gröbner bases, no tropical linear algebra yet; these are planned for later versions.

Current roadmap

v0.1 (current): 2D core, arithmetic, polynomials, Newton polytopes, regular subdivisions, the analytical dual, and grid-based visualisation.

v0.2: 3D hypersurfaces, a ReLU network interface for extracting the tropical rational map represented by a trained network, and explicit balancing-condition verification.

Further plans (Gröbner bases, tropical linear algebra, and beyond) are tracked in the repository's issues rather than here.

References

Zhang, L., Naitzat, G., & Lim, L.-H. (2018).
Tropical Geometry of Deep Neural Networks.
Proceedings of the 35th International Conference on Machine Learning (ICML).