AIP Engine

Algebraic Independence Processor — auto-detection of matrix structure + memory-efficient computation for ultra-large sparse systems.

pip install aip-engine

What it does

AIP Engine solves a real problem: building and solving sparse linear systems that are too large for conventional tools. It does three things:

1. Detects matrix structure automatically (sparse/dense, square/rectangular)
2. Routes to the optimal solver (LSQR, spsolve, LAPACK)
3. Accordion Memory: builds and solves ultra-large systems without running out of RAM

Quick start

import aip

# Auto-detect structure and solve
report = aip.detect_matrix(A)
x = aip.solve(A, b)

Accordion Memory

For systems with millions or billions of entries that don't fit in RAM:

from aip.accordion import PascalIndex, AccordionBuilder, solve_chunks

# 1. Mathematical indexing (replaces dictionary, saves GB of RAM)
index = PascalIndex(num_vars=20, max_degree=6)
idx = index.combo_to_index((3, 7, 12))  # O(k) time, 0 extra memory

# 2. Batch construction (never all in RAM at once)
builder = AccordionBuilder(num_rows=1_000_000)
# ... add entries in batches ...
builder.flush()  # converts to CSR chunk, frees raw data
chunks = builder.finalize()

# 3. Streaming solve (never assembles full matrix)
result = solve_chunks(chunks, b, max_iter=10000)
print(result['residual'], result['size_l2'])

Why Accordion?

Real-world results:

How it works

PascalIndex

Uses the Combinatorial Number System to compute the index of any monomial in O(k) time using a precomputed Pascal table. No dictionary needed.

index = PascalIndex(num_vars=20, max_degree=6)
print(index)  # PascalIndex(vars=20, deg=6, monomials=60,459, pascal=336 bytes)
# 60K monomials indexed with 336 bytes of memory

AccordionBuilder

Builds sparse matrices in batches using array.array (C-native, 4-8 bytes/element) instead of Python lists (28 bytes/element). Each batch is converted to CSR immediately and raw arrays are freed.

solve_chunks

LSQR solver that operates on a LinearOperator built from column chunks. The matvec/rmatvec operations iterate over chunks sequentially, never needing the full matrix in memory.

Requirements

• Python >= 3.8
• NumPy >= 1.20
• SciPy >= 1.7

License

MIT License - Carmen Esteban, 2025-2026