eml-graph 0.1.0

Equivalence-class graph over corpora of SymPy expressions.

eml-graph

Equivalence-class graph over corpora of SymPy expressions. Cluster a codebase's mathematical content by Pfaffian cost class; optionally add cost-monotone rewrite edges; render to Graphviz DOT for visualization.

Quick start

pip install eml-graph

from eml_graph import build_graph, to_dot
import sympy as sp

x, y = sp.symbols("x y")
corpus = [
    sp.sin(x),
    sp.sin(y),                    # collapses with sin(x) by axes
    sp.exp(x),
    sp.cos(y),                    # collapses with sin(*) by axes
    sp.exp(x) / (1 + sp.exp(x)),  # textbook sigmoid
    1 / (1 + sp.exp(-x)),         # canonical sigmoid
]

g = build_graph(corpus, label_with_discover=True)
print(g.num_nodes(), "nodes across", g.num_classes(), "Pfaffian-cost classes")

# Save as DOT, then render with graphviz:
with open("corpus.dot", "w") as f:
    f.write(to_dot(g, include_edges=True))
# $ dot -Tsvg corpus.dot -o corpus.svg

What does this do that nothing else does?

Software architects have call graphs. Mathematicians have equivalence classes. eml-graph is the first tool to give you both at once for a real corpus of code.

Open numpy/, scipy/, torch/, or your own pipeline; scrape every SymPy expression (or build them from your symbolic layer); feed the list to build_graph(). You get back:

• Cost-class clusters. All expressions sharing the same Pfaffian profile (pfaffian_r, max_path_r, eml_depth, correction sum) collapse into one cluster. You see immediately which mathematical shapes the codebase actually uses.
• Rewrite edges (optional). Within a cluster, two expressions are connected if the eml-rewrite library can walk one into the other under monotone-decreasing cost. Edge weights are the rewrite step counts.
• Named labels (optional). When eml-discover is installed and label_with_discover=True, nodes that match a registry formula (sigmoid, GELU, Pythagorean identity, ...) get the canonical name.

Output

to_dot(graph) returns a Graphviz DOT-format string. No binary deps; pipe through dot, neato, sfdp, or paste into a DOT web viewer. For SVG: dot -Tsvg input.dot -o output.svg.

Status

Beta. v0.1.0 is the MVP — clustering + DOT rendering + lazy path-based edges. Roadmap:

• v0.2: matplotlib renderer (no graphviz binary needed)
• v0.3: streaming graph for corpora that don't fit in memory
• v0.4: per-class summary statistics (entropy, modularity, ...)
• v1.0: stable API + first public release with paper

License

PROPRIETARY-PRE-RELEASE. Patent pending. Do not redistribute.