Kronecker.jl

This is a Julia package to efficiently work with Kronecker products. It combines lazy evaluation and algebraic tricks such that it can implicitely work with huge matrices. It allows to work with large Kronecker systems both much faster and using much less memory than the naive implementation of the Kronecker product.

Features

Given two matrices (subtype of AbstractArray) A and B, one can construct an instance of the KroneckerProduct type as K = A ⊗ B (typed using \otimes TAB). Equivalently, the function kronecker function can be used as well. Several functions are implemented.

collect(K) computes the Kronecker product (not recommended!)
tr, det, size, eltype, inv, ... are efficient functions to work with Kronecker products. Either the result is a numeric value or a new KroneckerProduct type is returned
Multiplying with a vector v is efficient using the vec trick: K * v
Solving systems of the form A ⊗ B + cI
Working with incomplete systems using the sampled vec trick
Basic compatibility with higher-order Kronecker systems, e.g. A ⊗ B ⊗ C or kronecker(A, 4) which is equivalent with A ⊗ A ⊗ A ⊗ A
A KroneckerSum can be constructed with A ⊕ B (typed using \oplus TAB) or kroneckersum(A,B)
- Multiplication with vectors uses a specialized version of the vec trick
- Higher-order sums are supported, e.g. A ⊕ B ⊕ C or kroneckersum(A,4)
[in progress] GPU compatibility!
[in progress] Autodiff for machine learning models!

Read the documentation to get the specifics.

Benchmarks

Below is a comparision between the native kron function and Kronecker.jl for several operations.

Example

using Kronecker

A = randn(100, 100);
B = rand(50, 50);

v = rand(5000);

K = A ⊗ B

collect(K)  # equivalent with kron(A, B)

K[78, 43]

tr(K)
inv(K)  # yields another lazy Kronecker instance

K * v  # equivalent with vec(B * reshape(v, 50, 100) * A')