OpenMander

OpenMander is a fast, memory-efficient, research-oriented toolkit for computational redistricting, implemented in Rust with Python bindings. It supports large-scale, multi-objective redistricting experiments at census-block resolution, with a focus on incremental evaluation, contiguity-aware moves, and heuristic search.

Paper

OpenMander: A Multi-Objective Optimization Framework for Computational Redistricting
Under submission.
Preprint: TBD

Quickstart (Python, Rust)

Python

Install from PyPi:

python -m pip install openmander

Or build locally from source (requires [maturin]):

python -m pip install -U maturin
cd bindings/python
maturin develop -r

Example usage:

import openmander as om

# Download and prepare a state-level data pack (Illinois, 2020 census)
pack_path = om.download("IL")

# Load the map and initialize a 17-district plan
mp = om.Map(pack_path)
plan = om.Plan(mp, num_districts=17)

# Generate a random initial plan
plan.randomize()

# Balance population across districts
plan.equalize(
    "T_20_CENS_Total",
    tolerance=0.002,
    max_iter=1000,
)

# Export block-level assignments
plan.to_csv("block_assignments.csv")

The Python interface exposes high-level abstractions (Map, Plan, and optimization helpers) suitable for exploratory analysis, batch experiments, and integration with scientific Python tooling.

Rust

Add OpenMander as a dependency:

# Cargo.toml
[dependencies]
openmander = { git = "https://github.com/Ben1152000/openmander-core" }
anyhow = "1"

Minimal example:

use std::sync::Arc;
use anyhow::Result;
use openmander::{Map, Plan};

fn main() -> Result<()> {
    // Load a previously prepared data pack
    let map = Arc::new(Map::read_from_pack("IA_2020_pack")?);

    // Create a 4-district plan
    let mut plan = Plan::new(map, 4);

    // Randomize and export assignments
    plan.randomize()?;
    plan.to_csv("plan.csv")?;

    Ok(())
}

Components

Map

A Map represents a multi-layer geographic dataset with explicit graph structure. Each map consists of ordered layers (e.g., state → county → tract → block), where each layer contains:

• Dense node indices and stable geographic identifiers
• Tabular attributes (e.g., census variables)
• Adjacency graphs with optional shared-perimeter weights (CSR format)
• Optional geometries (stored as FlatGeobuf)

Maps are immutable after construction and shared across plans.

• Plan (a.k.a. GraphPartition) A partition of the node graph into parts (districts):

  • assignments[u] → part
  • boundary[u] + per-part frontier sets
  • per-part size/weight totals (WeightMatrix)
  • contiguity checks and articulation-aware moves
  • simulated annealing helpers (balance, optimize)

Plan

A Plan (internally a graph partition) assigns each node in a map layer to a district, and contains:

• Incremental boundary tracking and per-district frontier sets
• Per-district aggregate weights (WeightMatrix)
• Fast contiguity checks and articulation-aware moves
• Built-in local search helpers (randomization, balancing, optimization)

Plans are designed to support heuristic search algorithms such as simulated annealing, tabu search, and beam search.

Data Packs

OpenMander operates on preprocessed data packs, which bundle all required data for a state and census decade.

Example layout:

<STATE>_2020_pack/
  data/             # per-level attribute tables (parquet)
  adj/              # CSR graphs per level (*.csr.bin)
  geom/             # per-level FlatGeobuf (*.fgb)
  hull/             # per-level convex hulls (*.fgb)
  manifest.json     # schema & provenance

License

License: TBD